text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1508.01870v2 [math.GR] 13 Aug 2016
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
Abstract. We say that permutations π1 , . . . , πr ∈ Sn invariably generate Sn if, no matter
how one chooses conjugates π1′ , . . . , πr′ of these permutations, π1′ , . . . , πr′ generate Sn . We
show that if π1 , π2 , π3 are chosen randomly from Sn then, with probability tending to 1 as
n → ∞, they do not invariably generate Sn . By contrast it was shown recently by Pemantle,
Peres and Rivin that four random elements do invariably generate Sn with probability
bounded away from zero. We include a proof of this statement which, while sharing many
features with their argument, is short and completely combinatorial.
1. Introduction
Albeit by Dixon’s theorem [3] two random elements π1 , π2 of the symmetric group Sn
generate at least the whole alternating group An with high probability1 as n → ∞, it is less
clear how large the group generated by π1′ , π2′ must be when π1′ and π2′ are allowed to be
arbitrary conjugates of π1 and π2 . Following Dixon [4] we say that a list π1 , . . . , πr ∈ Sn
has a property P invariably if π1′ , . . . , πr′ has property P whenever πi′ is conjugate to πi for
every i. How many random elements of Sn must we take before we expect them to invariably
generate Sn ?
Several authors [2, 4, 7, 9, 11, 12] have already considered this question, owing to its
connection with computational Galois theory. To briefly explain this connection, suppose
we are given a polynomial f ∈ Z[x] of degree n with no repeated factors. Information about
the Galois group can be gained by reducing f modulo various primes p and factorizing the
reduced polynomial f¯ over Z/pZ. By classical Galois theory, if f¯ has irreducible factors of
degrees n1 , . . . , nr then the Galois group G of f over Q has an element with cycle lengths
n1 , . . . , nr . Moreover by Frobenius’s density theorem, if G = Sn then the frequency with
which a given cycle type arises is equal to the proportion of elements in Sn with that cycle
type. Thus if we suspect that G = Sn then the number of times we expect to have to iterate
this procedure before proving that G = Sn is controlled by the expected number of random
elements required to invariably generate Sn .
BG is supported by ERC Starting Grant 279438 “Approximate Algebraic Structure” and a Simons Investigator Grant.
KF is supported by National Science Foundation grants DMS-1201442 and DMS-1501982.
1We adopt the convention that for a sequence of events E in finite probability spaces depending on some
n
parameter n, “En occurs with high probability” means with P(En ) → 1 as n → ∞.
1
2
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
Luczak and Pyber [9] were the first to prove the existence of a constant C such that
C random permutations π1 , . . . , πC ∈ Sn invariably generate Sn with probability bounded
away from zero. Their method does not directly yield a reasonable value of C, but recently
Pemantle, Peres, and Rivin [12] proved that we may take C = 4.
Theorem 1.1 (Pemantle–Peres–Rivin [12]). If π1 , π2 , π3 , π4 ∈ Sn are chosen uniformly at
random then the probability that π1 , π2 , π3 , π4 invariably generate Sn is bounded away from
zero.
Incidentally, Pemantle, Peres, and Rivin only prove that π1 , π2 , π3 , π4 invariably generate
a transitive subgroup of Sn , but it is little more work to prove the theorem as stated above.
We give a somewhat simplified proof of this theorem in Section 2. Our main contribution
however is the lower bound C > 3, which will be proved in Section 3. Thus C can be taken
as small as 4, but no smaller.
Theorem 1.2. If π1 , π2 , π3 ∈ Sn are chosen uniformly at random then the probability that
π1 , π2 , π3 invariably generate a transitive subgroup (or, in particular, all of Sn ) tends to zero
as n → ∞. Equivalently, with probability tending to 1 there is a positive integer k < n such
that π1 , π2 , π3 each have a fixed set of size k.
As in our recent paper [5], our main tool is the following model for the small-cycle structure
of a random permutation; see for example Arratia and Tavaré [1].
Lemma 1.3. Let X = (X1 , X2 , . . . ) be a sequence of independent Poisson random variables,
where Xj has parameter 1/j. If cj is the number of cycles of length j in a random permutation
π ∈ Sn , then if k is fixed and n → ∞ the distribution of (c1 , . . . , ck ) converges to that of
(X1 , . . . , Xk ).
The set of fixed-set sizes of a random permutation is thus modeled by the random sumset
(
)
X
L (X) =
jxj : 0 6 xj 6 Xj .
(1.1)
j>1
Thus unsurprisingly the main task in proving Theorem 1.1 is to show that
P L (X) ∩ L (X′) ∩ L (X′′ ) ∩ L (X′′′ ) = {0} > 0,
′
′′
′′′
(1.2)
where X , X , X are independent copies of X. Similarly, Theorem 1.2 follows almost immediately from
P L (X) ∩ L (X′ ) ∩ L (X′′ ) = {0} = 0.
(1.3)
Ultimately, these assertions come down to the inequalities log 2 < 43 and 32 < log 2 respectively, as we shall see in the course of the proofs.
These questions about permutations have analogues in number theory. Our proof of
Theorem 1.2 is modeled after that of the well-known theorem of Maier and Tenenbaum [10]
on the propinquity of divisors: a random integer n (selected from {1, . . . , x} for large x) has
two distinct divisors d, d′ with d < d′ 6 2d with high probability (as x → ∞). In particular,
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
3
P
we make heavy use of Riesz products, a device closely related to the sums d|n diθ that one
sees frequently in the propinquity literature.
The Maier–Tenenbaum theorem itself corresponds more perfectly with the assertion that
with high probability a random permutation π has, for some k with 0 < k < n, two different
fixed sets of size k (a true statement, but not one we establish here).
The number-theoretic analogue of Theorem 1.1 is a statement of the following kind:
if x is large and if we select four random integers n1 , n2 , n3 , n4 independently at random
from {1, . . . , x} then, with probability bounded below by an absolute constant, any divisors
d1 |n1 , d2 |n2 , d3|n3 , d4 |n4 should have max di > 2 min di .
The number-theoretic analogue of Theorem 1.2 is a statement of the following kind: if
x is large and if we select three random integers n1 , n2 , n3 independently at random from
{1, . . . , x} then, with probability tending to 1 as x → ∞, there exist divisors d1 |n1 , d2 |n2 , d3 |n3
with max di < 2 min di .
Both of these number-theoretical statements were established nearly 20 years ago by Raouj
and Stef [13]. In fact, rather more precise statements are established in that paper. We thank
Gérald Tenenbaum for bringing this paper to our attention.
The analogous problem of determining the expected number of random elements required
to invariable generatean arbitrary finite group has been considered in several recent papers
[6, 7, 8].
Notation. Throughout we use standard O(·) and o(·) notation, as well as the Vinogradov
notation X ≪ Y to mean X = O(Y ).
2. Four generators are enough
The principal result needed for the proof of Theorem 1.1 is the following proposition.
Proposition 2.1. The following is true uniformly for integers k, n with 1 6 k 6 n/2. If
π1 , π2 , π3 , π4 ∈ Sn are chosen uniformly at random, then the probability that there is some
ℓ ∈ (k/2, k] such that π1 , π2 , π3 , π4 each fix a set of size ℓ is O(k −c) for some c > 0.
We begin with a tool for counting permutations with a given number of cycles of length at
most k. By the Poisson model mentioned in the Introduction (Lemma 1.3), if k is fixed and
n → ∞, this statistic has distribution approaching that of X1 + · · · + Xk , a Poisson variable
with parameter hk = 1 + 21 + · · · + k1 . The next result tells us that the distribution is still
approximately Poisson uniformly over all choices of parameters k and n.
Lemma 2.2. Let n, k, ℓ be integers with n > k > 1 and ℓ > 0. Select π ∈ Sn at random.
Then
ℓ
e (1 + log k)ℓ
1+
.
P(π has exactly ℓ cycles with length 6 k) 6
k
ℓ!
1 + log k
4
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
In particular if ℓ ≪ log k then this is O (1 + log k)ℓ /kℓ! , while if ℓ ≫ log k then this is
O (1 + log k)ℓ−1 /k(ℓ − 1)! .
Proof. Denote by Sn (k, ℓ) the set of π ∈ Sn containing exactly ℓ cycles of length at most k.
Evidently
X
X
|σ|.
n|Sn (k, ℓ)| =
π∈Sn (k,ℓ)
σ|π
σ a cycle
Here the inner sum is over cycles σ which are factors of (i.e., contained in) π, and |σ| denotes
the length of σ. Write π = σπ ′ , and observe that π ′ has either ℓ − 1 or ℓ cycles of length at
most k, depending on whether |σ| 6 k or not. Thus π ′ ∈ Sn−|σ| (k, m), where m = ℓ − 1 or
m = ℓ, so
n
ℓ
X
X
n|Sn (k, ℓ)| 6
j=1 m=ℓ−1
X
π ′ ∈S
ℓ
n
X
X
=
n−j (k,m)
X
j=1 m=ℓ−1 π ′ ∈Sn−j (k,m)
X
j
σ∈Sn ,|σ|=j
σ a cycle
n!
.
(n − j)!
Now rearrange the sum according the cycle type (c1 , . . . , cn ) of the permutation π ′ , i.e., π ′
has ci cycles of length i for 1 6 i 6 n, and c1 + 2c2 + · · · + ncn = n − j if π ′ ∈ Sn−j . The
well known
formula states that the number of π ′ ∈ Sn−j with a given cycle type is
Q Cauchy
ci
(n − j)!/ i ci !i . It follows that
n|Sn (k, ℓ)| 6 n!
n
X
j=1
X
c1 ,...,cn >0
c1 +2c2 +···+ncn =n−j
c1 +···+ck ∈{ℓ−1,ℓ}
X
6 n!
c1 ,...,cn >0
c1 +···+ck ∈{ℓ−1,ℓ}
= n!
= n!
X
c1 ,...,ck >0
c1 +···+ck =ℓ−1
1
ci
i ci !i
Q
1
ci
i ci !i
Q
Q
hkℓ−1
hℓ
+ k
(ℓ − 1)!
ℓ!
1
+
ci
i ci !i
Y
k<i6n
X
c1 ,...,ck >0
c1 +···+ck =ℓ
e1/i ,
1
X
ci
i ci !i
c
,...,c
Q
k+1
1
n >0
Qn
ci
i=k+1 ci !i
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
5
where in the last line we used the multinomial theorem. The claimed bound now follows
using the inequalities hk 6 1 + log k and
X 1
= hn − hk 6 log n − log k + 1.
i
k<i6n
In our paper [5] we showed that the probability of a random permutation π ∈ Sn fixing
log 2
some set of size k is k −δ+o(1) , where δ = 1 − 1+log
≈ 0.086. As noted in that paper,
log 2
the main contribution to this estimate comes from rather exceptional permutations with an
unexpectedly large number, ≈ log k/ log 2, of cycles of length 6 k. By contrast a typical
permutation has ≈ log k cycles of length 6 k. By restricting to this “quenched” regime2 we
can establish a much stronger bound.
Lemma 2.3. Suppose that k, n are integers with 1 6 k 6 n/2, 0 < ε 6 1/2, and choose
π ∈ Sn uniformly at random. Then the probability that π fixes a set of size k and has at
most (1 + ε) log k cycles of length at most k is at most O(k log 2−1+2ε ).
Proof. Fix ℓ 6 (1 + ε) log k and consider permutations π with exactly ℓ cycles of length at
most k. If π fixes some set X, |X| = k, write π1 = π|X and π2 = π|[n]\X for the induced
permutations on X and its complement. Then π1 has ℓ1 cycles of length 6 k, and π2 has ℓ2
cycles of length 6 k, where ℓ1 + ℓ2 = ℓ. By Lemma 2.2 the number of such π, for a given
choice of X and ℓ1 , ℓ2 , is bounded by a constant times
(1 + log k)ℓ1
(1 + log k)ℓ2
k! ·
(n − k)!,
kℓ1 !
kℓ2 !
which means that the probability we are interested in is bounded by a constant times
X 1 (1 + log k)ℓ
2ℓ (1 + log k)ℓ
=
.
2
2 ℓ!
k
ℓ
k
1 !ℓ2 !
ℓ +ℓ =ℓ
1
2
By summing over all ℓ 6 ℓ0 = ⌊(1 + ε) log k⌋ we get the bound
1
k2
X
ℓ6(1+ε) log k
2ℓ (1 + log k)ℓ
1 2ℓ0 (1 + log k)ℓ0
≪ 2
ℓ!
k
ℓ0 !
1
(2e/(1 + ε))(1+ε) log k
k2
1
≪ 1−log 2−2ε .
k
≪
Proof of Proposition 2.1. Let ε > 0 be small and fixed. First we will use Lemma 2.2 to
bound the probability that one of π1 , π2 , π3 , π4 has more than ℓ0 = ⌊(1 + ε) log k⌋ cycles of
2The
terminology is from [12] and apparently comes from statistical physics.
6
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
length at most k. By that lemma, for each ℓ > ℓ0 , the probability that π1 has ℓ cycles of
length at most k is bounded by
(1 + log k)ℓ−1
,
O
k(ℓ − 1)!
so the probability that π1 has more than ℓ0 cycles is bounded by a constant times
ℓ −1
(1+ε) log k
X (1 + log k)ℓ−1
e
(1 + log k)ℓ0 −1
1 e(1 + log k) 0
1
≪
≪
≪
.
k(ℓ − 1)!
k(ℓ0 − 1)!
k
ℓ0 − 1
k 1+ε
ℓ>ℓ
0
2
Now by a Taylor expansion of −1 + (1 + ε) log(e/(1 + ε)), this is bounded by O(k −ε /3 ) if
ε 6 21 . Thus the probability that one of π1 , π2 , π3 , π4 has more than ℓ0 cycles of length at
2
most k is also bounded by O(k −ε /3 ).
On the other hand, by Lemma 2.3, for each ℓ ∈ (k/2, k] the probability that πi has at most
(1 + ε) log k cycles of length at most k and fixes a set of size ℓ is at most k log 2−1+2ε . Thus
the probability that π1 , π2 , π3 , π4 each have at most (1 + ε) log k cycles of length at most k
and each fix a set of the same size ℓ for some ℓ ∈ (k/2, k] is at most k 1+4(log 2−1+2ε) . Since
1 + 4(log 2 − 1) < 0, we have 1 + 4(log 2 − 1 + 2ε) < 0 if ε is small enough (ε = 1/40 works),
and so the theorem holds with
c = min(ε2 /3, −1 − 4(log 2 − 1 + 2ε)).
An immediate corollary of Proposition 2.1 is obtained by fixing k, letting n → ∞, and
recalling the Poisson model (Lemma 1.3) and the definition (1.1) of L (X) .
Corollary 2.4. For any k > 2, the probability that L (X) ∩ L (X′ ) ∩ L (X′′ ) ∩ L (X′′′ )
contains an integer ℓ ∈ (k/2, k] is O(k −c ), for some c > 0.
Remark. If one wished to prove only this, we could substitute Lemma 2.2 with a corresponding bound for P(X1 + · · · + Xk 6 ℓ), which follows very quickly from the fact that
X1 + · · · + Xk is Poisson with parameter hk .
Corollary 2.5. L (X) ∩ L (X′ ) ∩ L (X′′ ) ∩ L (X′′′ ) is almost surely finite, and equal to {0}
with positive probability.
Proof. Let Fk be the event that
L (X) ∩ L (X′) ∩ L (X′′ ) ∩ L (X′′′ ) ∩ (k, ∞)
is nonempty. By applying Corollary 2.4 with k replaced by 2j k, j ∈ N, and summing
the
T
geometric series, we obtain P(Fk ) ≪ k −c for k > 1. In particular P(Fk ) → 0, so P ( Fk ) = 0,
so the first part of the corollary holds. For the second part, fix k0 such that P(Fk0 ) < 1.
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
7
Then
P (L (X) ∩ L (X′ ) ∩ L (X′′ ) ∩ L (X′′′ ) = {0}) > P(Xj = 0 for all j 6 k0 , and Fkc0 )
> P(Xj = 0 for all j 6 k0 ) P(Fkc0 )
> 0.
The second inequality here is a simple case of the FKG inequality [14, Theorem 1.19]. To see
the inequality directly, define X∗ = (X1∗ , X2∗ , . . . ) by putting Xj∗ = Xj if j > k0 and Xj∗ = 0
if j 6 k0 , and let Fk∗0 be the event that
L (X∗ ) ∩ L (X′ ) ∩ L (X′′ ) ∩ L (X′′′ ) ∩ (k0 , ∞)
is nonempty. Clearly Fk∗0 implies Fk0 , so
P(Xj = 0 for all j 6 k0 , and Fkc0 ) = P(Xj = 0 for all j 6 k0 , and Fk∗c0 )
= P(Xj = 0 for all j 6 k0 ) P(Fk∗c0 )
> P(Xj = 0 for all j 6 k0 ) P(Fkc0 ).
Shortly we will complete the proof of Theorem 1.1. In the proof, we will need a trick
to deal with the possibility that π1 , π2 , π3 , π4 ∈ An . The following lemma is helpful in this
regard. It shows that random even and random odd permutations have the same small-cycle
structure as random permutations with unconstrained parity (Lemma 1.3).
Lemma 2.6. Let π ∈ Sn be a random even permutation, and let cj (π) be the number of cycles
of length j. Fix k ∈ N. Then as n → ∞ the distribution of (c1 (π), . . . , ck (π)) converges to
that of (X1 , . . . , Xk ). The same is true if π is a random odd permutation.
Proof. Choose π ∈ Sn uniformly at random, and define σ by putting σ = 1 if π is even
and σ = (12) if π is odd. Then πσ is uniformly distributed over An . By Lemma 1.3, as
n → ∞, the number of cycles in π of length at most 2k approaches a Poisson distribution
with parameter h2k 6 1 + log 2k. Thus, with high probability (as n → ∞) the total number
of points in cycles of π of length at most 2k is at most 2k log n, so with high probability each
of these cycles is disjoint from (12). That is, the points 1 and 2 are both contained in cycles
of π of length at least 2k + 1 with high probability. Now consider the probability that 1
and 2 are both contained in the same cycle and are close together. For each ℓ > 2k + 1, the
number of cycles of length ℓ containing both 1 and 2, which are a distance 6 k from each
other, equals n−2
2k(ℓ − 2)!. Hence, the number of permutations π containing such a cycle
ℓ−2
is at most
X
2k(n − 2)! 6 2k(n − 1)!.
2k+16ℓ6n
Hence, with high probability, if 1 and 2 are in the same cycle they are a distance at least k +1
from each other. Thus, with high probability, cj (πσ) = cj (π) for each j 6 k. Similarly πσ(12)
is uniformly distributed over odd permutations, and with high probability cj (πσ(12)) =
cj (π).
8
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
Proof of Theorem 1.1. Let π1 , π2 , π3 , π4 ∈ Sn be random permutations with π1 odd. Let En,k
be the event that π1 , π2 , π3 , π4 each fix a set of size ℓ for some ℓ in the range 1 6 ℓ 6 k, and let
Fn,k be the event that π1 , π2 , π3 , π4 each fix a set of size ℓ for some ℓ in the range k < ℓ 6 n/2.
By Proposition 2.1 (and summing a geometric series as in the proof of Corollary 2.5) we have
P(Fn,k ) ≪ k −c uniformly for 1 6 k 6 n/2, while by Corollary 2.5 and Lemma 2.6 we have
limn→∞ P(En,k ) 6 1 − δ for all k, for some constant δ > 0. Fix k0 such that P(Fn,k0 ) 6 δ/3
for all n > 2k0. Then P(En,k0 ) + P(Fn,k0 ) 6 1 − δ/3 for all sufficiently large n, so we deduce
that with probability bounded away from zero π1 , π2 , π3 , π4 do not fix sets of the same size
ℓ for any ℓ ∈ [1, n/2].
Thus with probability bounded away from zero π1 , π2 , π3 , π4 invariably generate a transitive
subgroup of Sn . However by the Luczak–Pyber theorem [9], π1 is, with high probability, not
contained in any transitive subgroup smaller than An . Since π1 6∈ An , with probability
bounded away from zero π1 , π2 , π3 , π4 invariably generate Sn .
3. Three generators are not enough
Theorem 1.2 follows immediately from the following more specific proposition.
Proposition 3.1. For every ε > 0 there exists k0 = k0 (ε) and n0 = n0 (ε) such that if n > n0
then with probability at least 1 − ε there is some ℓ 6 k0 such that π1 , π2 , π3 each fix a set of
size ℓ.
Let X be defined as before, and let Y and Z be independent copies of X. For I an interval
in N let
(
)
X
L (I, X) =
jxj : 0 6 xj 6 Xj ,
j∈I
and define L (I, Y) and L (I, Z) analogously.
Lemma 3.2. Let I = {1, . . . , k} and let ε > 0. Then with probability at least 1 − ε we have
L (I, X), L (I, Y), L (I, Z) ⊂ [0, 3ε−1k].
P
P
Proof. Since E j∈I jXj = |I| = k, by Markov’s inequality we have j∈I jXj 6 3ε−1 k with
P
P
probability at least 1 − ε/3. Similarly j∈I jYj 6 3ε−1 k and j∈I jZj 6 3ε−1 k each with
probability at least 1 − ε/3, and the lemma follows.
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
9
Lemma 3.3. Fix ε, 0 < ε < 1/2. There is a constant C(ε) so that with probability at least
1 − ε we have
X
Xj > 0.99 log(k/m) − C(ε),
m<j6k
X
Yj > 0.99 log(k/m) − C(ε), and
m<j6k
X
Zj > 0.99 log(k/m) − C(ε).
m<j6k
for every nonnegative integer m 6 k.
Proof. Let C = C(ε) be a constant whose properties will be specified later. It suffices to
show that the first inequality holds for all m 6 k with probability at least 1 − ε/3. There
is nothing to prove if m > e−C k, so we may suppose m 6 e−C k. We may also suppose that
C > 1.
Let E be the event that
X
Xj > 0.99 log(k/m) − 1
m<j6k
−C
for all m 6 e
k. Suppose E fails, say
X
Xj < 0.99 log(k/m) − 1
m<j6k
for some m 6 e−C k. Writing m′ for the smallest power of 2 with m′ > m, we thus have
X
X
Xj 6
Xj 6 0.99 log(k/m) − 1 6 0.99 log(k/m′ ).
m<j6k
m′ <j6k
Thus
X
1E c 6
0.99
P
m′ <j6k
Xj −0.99 log(k/m′ )
.
m′ 62e−C k
dyadic
Whenever
P is Poisson of parameter λPand a > 0 we have EaP = e(a−1)λ , and the sum
P
′
m′ <j6k Xj is Poisson with parameter
m′ <j6k 1/j = log(k/m ) + O(1), so
X
exp ((0.99 − 1 − 0.99 log(0.99)) log(k/m′ ))
P(E c ) ≪
m′ 62e−C k
dyadic
6
X
(k/m′ )−0.00005
m′ 62e−C k
dyadic
≪ e−0.00005C .
Therefore, P(E c ) 6 ε/3 if C is taken large enough.
10
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
We need a standard estimate for the partial sums of the Fourier series
− log |2 sin(πθ)|. Denote by kxk the distance from x to Z.
P∞
j=1
cos(2πjθ)
j
=
Lemma 3.4.
X cos(2πjθ)
j
j6m
= log min
1
, m + O(1) for kθk > 0.
kθk
Proof. We may assume that 0 < θ 6 21 . Using the bound cos(2πjθ) = 1 + O(j 2 θ2 ), we get
X
j6min(m,1/θ)
cos(2πjθ)
= log min(m, 1/θ) + O(1).
j
This proves the lemma if kθk 6 1/m. Suppose, then, that kθk > 1/m. Set
Sj =
j
X
e2πinθ ,
n=0
and note that by summing the geometric series we have
Sj =
e2πijθ − 1
1
≪ .
2πiθ
e
−1
θ
(3.1)
Thus (by “Abel summation”),
X
1/θ<j6m
X
cos(2πjθ)
=ℜ
j
1/θ<j6m
=ℜ
X
X
e2πijθ
=ℜ
j
1/θ<j6m−1
1/θ<j6m
Sj − Sj−1
j
Sj
Sm S⌈1/θ⌉−1
+
−
.
j(j + 1)
m
⌈1/θ⌉
The latter two terms here are O(1) by the trivial bound |Sj | 6 j, while from (3.1) the sum
is bounded by a constant times
1 X 1
≪ 1.
θ
j2
j>1/θ
Let T = R/Z be the unit torus, and denote e(z) = e2πiz . Given I, X, Y, Z define F : T2 →
C by
Y 1 + e(jθ1 ) Xj 1 + e(jθ2 ) Yj 1 + e(j(−θ1 − θ2 )) Zj
F (θ) =
.
2
2
2
j∈I
By expanding the product we see that F̂ : Z2 → C is supported on the set
S(I, X, Y, Z) = {(n1 − n3 , n2 − n3 ) : n1 ∈ L (I, X), n2 ∈ L (I, Y), n3 ∈ L (I, Z)}.
(3.2)
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
Since
11
P
F̂ (a) = F (0) = 1, by Cauchy–Schwarz we have
2 X X
X
X
|F̂ (a)|2 .
F̂ (a) 6
|F̂ (a)|2 6 |S(I, X, Y, Z)|
1=
1
a∈Z2
a∈Z2
a∈Z2
a∈Z2
a:F̂ (a)6=0
Applying Parseval, we get
|S(I, X, Y, Z)| >
X
|F̂ (a)|
a∈Z2
2
!−1
=
Z
2
|F (θ)| dθ
T2
−1
.
(3.3)
Lemma 3.5. Let
2
− 0.02 ≈ 0.0182,
3 log 2
and let I = (k β , k]. Fix ε ∈ (0, 1/2), and let E = E(ε) be the event from Lemma 3.3. Then
both of the bounds
E1E |F (θ)|2 ≪ε (kkθ1 k1/3 kθ2 k1/3 kθ3 k1/3 )−2.02
(3.4)
β =1−
and
E1E |F (θ)|2 ≪ε (kkθi k1/2 kθj k1/2 )−1.3
({i, j} ⊂ {1, 2, 3})
(3.5)
2
hold uniformly for θ ∈ T , where θ3 = −θ1 − θ2 . The expectation is over X, Y, Z.
Proof. Define, for i ∈ {1, 2, 3},
if kθi k > k −β ;
kβ
1/kθi k if 1/k < kθi k < k −β ;
ki =
k
if kθi k 6 1/k.
It is useful to note the (slightly crude) bound
ki 6
kβ
,
kθi k1−β
(3.6)
which follows by an analysis of the three cases in the definition of ki . If E holds then
X
X
X
Zj > 0.99 log(k 3 /(k1 k2 k3 )) − C(ε),
Yj +
Xj +
k1 <j6k
k3 <j6k
k2 <j6k
so
1E |F (θ)|2 ≪ε (k 3 /k1 k2 k3 )−0.99 log 2 |F (θ)|2 2
P
k1 <j6k
Xj +
P
k2 <j6k
Yj +
P
k3 <j6k
Zj
.
From (3.6) and the inequality 3 × 0.99 log 2 × (1 − β) > 2.02, we deduce that
1E |F (θ)|2 ≪ε (kkθ1 k1/3 kθ2 k1/3 kθ3 k1/3 )−2.02 |F (θ)|2 2
P
k1 <j6k
Xj +
P
k2 <j6k
Yj +
P
k3 <j6k
Zj
.
Thus (3.4) will follow if we can prove
E|F (θ)|2 2
P
k1 <j6k
Xj +
P
k2 <j6k
Yj +
P
k3 <j6k
Zj
≪ 1.
(3.7)
12
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
Similarly, from (3.6) for i = 1, 2 and the trivial bound k3 6 k, and using 2 × 0.99 log 2 ×
(1 − β) > 1.3, we deduce that
1E |F (θ)|2 ≪ε (kkθ1 k1/2 kθ2 k1/2 )−1.3 |F (θ)|2 2
P
k1 <j6k
Xj +
P
k2 <j6k
Yj +
P
k3 <j6k
Zj
,
and similarly for other permutations of the indices 1, 2, 3, so (3.5) will also follow from (3.7).
It remains only to prove (3.7). We have a factorization
E|F (θ)|2 2
P
k1 <j6k
=
Xj +
Yj +
P
k3 <j6k
1 + e(jθ1 )
E
2
2Xj
Y
1 + e(jθ2 )
E
2
2Yj
Y
1 + e(jθ3 )
E
2
2Zj
k β <j6k2
×
k2 <j6k
Y
k β <j6k1
×
P
k β <j6k3
Zj
1 + e(jθ1 )
E 2
2
<j6k
2
1 + e(jθ2 )
E 2
2
<j6k
2
1 + e(jθ3 )
E 2
2
<j6k
2
Y
k1
Y
k2
!Yj
Y
k3
!Xj
!Zj
.
By using again the calculation EaP = e(a−1)λ for P Poisson with parameter λ, we get
E|F (θ)|2 2
P
k1 <j6k
Xj +
P
k2 <j6k
Yj +
P
k3 <j6k
Zj
!
X 1 1 + e(jθi ) 2
X 1
1 + e(jθi )
= exp
−1 +
2
j
2
j
2
i=1
ki <j6k
k β <j6ki
3
X cos(2πjθi ) − 1
X
X cos(2πjθi )
= exp
+
2j
j
β
i=1
k <j6k
3
X
k <j6ki
= exp
3
X
i=1
2
!
−1
i
min(ki , 1/kθi k)
1
ki
min(k, 1/kθi k)
1
log
− log β + log
+ O(1)
β
2
min(k , 1/kθi k) 2
k
min(ki , 1/kθi k)
by Lemma 3.4. Checking the three cases in the definition of ki separately, it can be confirmed
that this is always O(1).
Corollary 3.6. With notation as in Lemma 3.5, we have
Z
E1E |F (θ)|2 dθ ≪ε k −2 .
T2
Proof. Divide T2 into three regions R1 , R2 , R3 as follows:
R1 = {θ ∈ T2 : kθi k > 1/k for all three i ∈ {1, 2, 3}},
R2 = {θ ∈ T2 : kθi k > 1/k for exactly two i ∈ {1, 2, 3}},
R3 = {θ ∈ T2 : kθi k > 1/k for at most one i ∈ {1, 2, 3}}.
(3.8)
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
13
We will bound the integral differently in each region.
Further subdivide R1 according to which of kθ1 k, kθ2 k, kθ3 k is largest. In the subregion
′
R1 in which say kθ1 k is largest we have kθ1 k > kθ2 k1/2 kθ3 k1/2 , so by (3.4) we have
Z
Z
2
(kkθ2 k1/2 kθ3 k1/2 )−2.02 dθ
E1E |F (θ)| dθ ≪ε
R′1
R1
=
Z
(kkθk)
kθk>1/k
−2
≍k .
−1.01
dθ
2
We can bound the integral over the other subregions in the same way, so the integral over
R1 is indeed ≪ε k −2 .
Similarly, subdivide R2 according to the relative order of kθ1 k, kθ2 k, kθ3 k, and focus for
the moment on the subregion R2′ in which kθ1 k 6 kθ2 k 6 kθ3 k. This implies in particular
that kθ1 k 6 1/k while kθ2 k > 1/k. Thus by (3.5) with i = 2 and j = 3 we have
Z
Z
2
E1E |F (θ)| dθ ≪ε
(kkθ2 k)−1.3 dθ ≍ k −2 .
R′2
R′2
Again we can bound the integral over the other subregions in the same way, so the integral
over R2 is also ≪ε k −2 .
Finally, in the region R3 note that because θ1 + θ2 + θ3 = 0 we must have kθi k < 2/k for
each i. Thus from the trivial bound |F (θ)| 6 1 we have
Z
Z
2
1 ≍ k −2 .
E1E |F (θ)| dθ 6
R3
R3
Recall the definition of S(I, X, Y, Z), given in (3.2).
Proposition 3.7. Let I = (k β , k]. There is a constant c > 0 such that with probability at
least 1/2 we have S(I, X, Y, Z) ⊂ [−10k, 10k]2 and |S(I, X, Y, Z)| > ck 2 .
Proof. Apply Lemma 3.3 with ε = 0.01, and let E be the resulting event. By Corollary 3.6
(and interchanging the order of integration and expectation) we have
Z
E1E
|F (θ)|2 dθ ≪ k −2 .
T2
R
Thus by Markov’s inequality there is a constant C such that 1ER T2 |F (θ)|2 dθ 6 Ck −2 with
probability at least 0.99. Since P(E) > 0.99 we deduce that T2 |F (θ)|2 dθ 6 Ck −2 with
probability at least 0.98. Applying (3.3), we have |S(I, X, Y, Z)| > C −1 k 2 with probability
at least 0.98.
On the other hand, by Lemma 3.2 with ε = 1/3 we have S(I, X, Y, Z) ⊂ [−10k, 10k]2
with probability at least 2/3, so we must have both S(I, X, Y, Z) ⊂ [−10k, 10k]2 and
|S(I, X, Y, Z)| > C −1 k 2 with probability at least 1 − 1/3 − 0.02 > 1/2.
14
SEAN EBERHARD, KEVIN FORD, AND BEN GREEN
Proposition 3.8. Let I = (k β , 60k]. Then with probability bounded away from zero we can
find (xj )j∈I , (yj )j∈I , (zj )j∈I not all zero such that 0 6 xj 6 Xj , 0 6 yj 6 Yj , and 0 6 zj 6 Zj
for each j ∈ I and
X
X
X
jxj =
jyj =
jzj .
j∈I
j∈I
j∈I
Proof. Let I ′ = (k β , k]. By Proposition 3.7, with probability at least 1/2 we have
S(I ′ , X, Y, Z) ⊂ [−10k, 10k]2
and |S(I ′, X, Y, Z)| ≫ k 2 . This event depends only on Xj , Yj , Zj for j 6 k, so independently
with probability at least 1/2 we can find j3 ∈ (20k, 50k] such that Zj3 > 0, as
Y
P(Zj = 0 for all j ∈ (20k, 50k]) =
e−1/j 6 1/2.
(3.9)
j∈(20k,50k]
Given such a j3 the set T of pairs of integers (j1 , j2 ) such that 10k < j1 , j2 6 60k and for
which
j1 (1, 0) + j2 (0, 1) − j3 (1, 1) ∈ −S(I ′ , X, Y, Z)
has size |T | ≫ k 2 . In particular there is a set T1 of integers j1 in the range 10k < j1 6 60k
of size |T1 | ≫ k such that for each j1 ∈ T1 there are ≫ k integers j2 in the same range
10k < j2 6 60k such that (j1 , j2 ) ∈ T . Thus by two further computations along the lines
of (3.9), independently with probability ≫ 1 we can find j1 ∈ T1 such that Xj1 > 0, and
then j2 such that (j1 , j2 ) ∈ T and such that Yj2 > 0.
But then by definition of S(I ′ , X, Y, Z) we can find (xj )j∈I ′ , (yj )j∈I ′ , (zj )j∈I ′ such that
0 6 xj 6 Xj , 0 6 yj 6 Yj , and 0 6 zj 6 Zj for all j ∈ I ′ and such that
X
X
X
j1 +
jxj = j2 +
jyj = j3 +
jzj .
j∈I ′
j∈I ′
j∈I ′
Thus the proposition follows from putting xj1 = yj2 = zj3 = 1, and putting all other xj , yj , zj
with j > k equal to 0.
Corollary 3.9. L (X) ∩ L (Y) ∩ L (Z) is almost surely infinite.
Proof. Define k1 to be sufficiently large, and thereafter ki+1 = (60ki )1/β . Then the intervals
Ii = (kiβ , 60ki ] are pairwise disjoint and by the proposition for each the probability that we
can find (xj )j∈Ii , (yj )j∈Ii , (zj )j∈Ii not all zero such that 0 6 xj 6 Xj , 0 6 yj 6 Yj , and
0 6 zj 6 Zj for each j ∈ Ii and
X
X
X
jxj =
jyj =
jzj
j∈Ii
j∈Ii
j∈Ii
is bounded away from zero. Since these events are independent for different values of i the
corollary follows.
INVARIABLE GENERATION OF THE SYMMETRIC GROUP
15
Proof of Proposition 3.1. By Corollary 3.9 there is some k0 = k0 (ε) such that if L (X) ∩
L (Y) ∩ L (Z) ∩ [1, k0 ] is nonempty with probability at least 1 − ε/2. Thus by Lemma 1.3
there is some n0 = n0 (ε) such that if n > n0 then with probability at least 1 − ε there is
some ℓ 6 k0 such that π1 , π2 , π3 each fix a set of size ℓ.
Acknowledgments. The authors thank the referees for carefully reading the paper and
for several helpful suggestions.
References
[1] R. Arratia and S. Tavaré, The cycle structure of random permutations, Ann. Probab. 20 (1992), no. 3,
1567–1591.
[2] J. H. Davenport and G. C. Smith, Fast recognition of alternating and symmetric Galois groups, J. Pure
Appl. Algebra 153 (2000), no. 1, 17–25.
[3] J. Dixon, The probability of generating the symmetric group, Math. Z. 110 (1969), 199–205.
[4] J. Dixon, Random sets which invariably generate the symmetric group, Discrete Math. 105 (1992),
25–39.
[5] S. Eberhard, K. Ford and B. Green, Permutations fixing a k-set, http://arxiv.org/abs/1507.04465.
[6] W. M. Kantor, A. Lubotzky and A. Shalev: Invariable generation and the Chebotarev invariant of a
finite group, J. Algebra 348 (2011), 302–314.
[7] E. Kowalski and D. Zywina, The Chebotarev invariant of a finite group, Exp. Math. 21 (2012), no. 1,
38–56.
[8] A. Lucchini, The Chebotarev invariant of a finte group: a conjecture of Kowalski and Zywina, preprint.
ArXiv:1509.05859
[9] T. Luczak and L. Pyber, On random generation of the symmetric group, Combin. Probab. Comput. 2
(1993), no. 4, 505–512.
[10] H. Maier and G. Tenenbaum, On the set of divisors of an integer, Invent. Math. 76 (1984), no. 1,
121–128.
[11] D. R. Musser. On the efficiency of a polynomial irreducibility test, J. Assoc. Comput. Mach. 25 (1978),
no. 2, 271–282.
[12] R. Pemantle, Y. Peres and I. Rivin, Four random permutations conjugated by an adversary generate Sn
with high probability, http://arxiv.org/abs/1412.3781.
[13] A. Raouj and A. Stef, Sur la proximité des diviseurs des entiers, J. Number Theory 76 (1999), no. 1,
66–93.
[14] T. Tao and V. Vu. Additive Combinatorics. Cambridge University Press, 2006.
Mathematical Institute, Radcliffe Observatory Quarter, Woodstock Road, Oxford OX2
6GG, England
E-mail address: [email protected]
Department of Mathematics, 1409 West Green Street, University of Illinois at UrbanaChampaign, Urbana, IL 61801, USA
E-mail address: [email protected]
Mathematical Institute, Radcliffe Observatory Quarter, Woodstock Road, Oxford OX2
6GG, England
E-mail address: [email protected]
| 4math.GR
|
A STENCIL SCALING APPROACH FOR ACCELERATING MATRIX-FREE
FINITE ELEMENT IMPLEMENTATIONS∗
arXiv:1709.06793v1 [cs.NA] 20 Sep 2017
S. BAUER† , D. DRZISGA‡ , M. MOHR† , U. RÜDE§ , C. WALUGA‡ , AND B. WOHLMUTH‡
Abstract. We present a novel approach to fast on-the-fly low order finite element assembly for scalar elliptic
partial differential equations of Darcy type with variable coefficients optimized for matrix-free implementations.
Our approach introduces a new operator that is obtained by appropriately scaling the reference stiffness matrix
from the constant coefficient case. Assuming sufficient regularity, an a priori analysis shows that solutions obtained
by this approach are unique and have asymptotically optimal order convergence in the H 1 - and the L2 -norm on
hierarchical hybrid grids. For the pre-asymptotic regime, we present a local modification that guarantees uniform
ellipticity of the operator. Cost considerations show that our novel approach requires roughly one third of the
floating-point operations compared to a classical finite element assembly scheme employing nodal integration. Our
theoretical considerations are illustrated by numerical tests that confirm the expectations with respect to accuracy
and run-time. A large scale application with more than a hundred billion (1.6 · 1011 ) degrees of freedom executed
on 14 310 compute cores demonstrates the efficiency of the new scaling approach.
Key words. matrix-free, finite-elements, variable coefficients, stencil scaling, variational crime analysis, optimal order a priori estimates
AMS subject classifications. 65N15, 65N30, 65Y20
1. Introduction. Traditional finite element implementations are based on computing local
element stiffness matrices, followed by a local-to-global assembly step, resulting in a sparse matrix.
However, the cost of storing the global stiffness matrix is significant. Even for scalar equations and
low order 3D tetrahedral elements, the stiffness matrix has, on average, fifteen entries per row, and
thus a standard sparse matrix format will require thirty times as much storage for the matrix as for
the solution vector. This limits the size of the problems that can be tackled and becomes the dominating cost factor since the sparse matrix must be re-read from memory repeatedly when iterative
solvers are applied. On all current and future computing systems memory throughput and memory access latency can determine the run-time more critically than the floating-point operations
executed. Furthermore, energy consumption has been identified as one of the fundamental roadblocks in exa-scale computing. In this cost metric, memory access is again more expensive than
computation. Against the backdrop of this technological development, it has become mandatory
to develop numerical techniques that reduce memory traffic. In the context of partial differential
equations this is leading to a re-newed interest in so-called matrix-free techniques and – in some
sense – to a revival of techniques that are well-known in the context of finite difference methods.
Matrix-free techniques are motivated from the observation that many iterative solvers, e.g.,
Richardson iteration or Krylov subspace methods, require only the result of multiplying the global
system matrix with a vector, but not the matrix itself. The former can be computed by local
operations in each element, avoiding to set up, store, and load the global stiffness matrix. One of
the first papers in this direction is [9], which describes the so called element-by-element approach
(EBE), in which the global matrix-vector-product (MVP) is assembled from the contributions
of MVPs of element matrices with local vectors of unknowns. The element matrices are either
pre-computed and stored or recomputed on-the-fly. Note that storing all element matrices, even
for low-order elements, has a higher memory footprint than the global matrix1 itself.
∗ Submitted
to the editors September 21, 2017.
Funding: This work was partly supported by the German Research Foundation through the Priority Programme 1648 ”Software for Exascale Computing” (SPPEXA) and by WO671/11-1.
† Dept. of Earth and Environmental Sciences, Ludwig-Maximilians-Universität München
‡ Institute for Numerical Mathematics (M2), Technische Universität München
§ Dept. of Computer Science 10, Friedrich-Alexander-Universität Erlangen-Nürnberg
1 However, it requires less memory than storage schemes for sparse direct solvers which reserve space for fill-in,
the original competing scenario in [9].
1
2
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
Consequently, the traditional EBE has not found wide application in unstructured mesh approaches. However, it has been successfully applied in cases where the discretization is based on
undeformed hexahedral elements with tri-linear trial functions, see e.g. [29, 1, 7, 13]. In such a
setting, the element matrix is the same for all elements, which significantly reduces the storage requirements, and variable material parameters can be introduced by weighting local matrix entries
in the same on-the-fly fashion as will be developed in this paper.
Matrix-free approaches for higher-order elements, as described in e.g., [19, 25, 22, 23], differ
from the classic EBE approach in that they do not setup the element matrix and consecutively
multiply it with the local vector. Instead, they fuse the two steps by going back to numerical
integration of the weak form itself. The process is accelerated by pre-computing and storing
certain quantities, such as e.g., derivatives of basis functions at quadrature points within a reference
element. These techniques work for arbitrarily shaped elements and orders, although a significant
reduction of complexity can be achieved for tensor-product elements.
However, these matrix-free approaches have also disadvantages. The element-based approach
for evaluating the global MVP can lead to significant extra cost caused by redundant operations
and redundant accesses to the solution vector. Also, by their design they do not provide direct
access to individual matrix rows or the diagonal entries of the matrix. This information is required
in classical relaxation methods and in many preconditioners. Additionally, while the low-order
settings [29, 1, 7, 13] require structured hexahedral meshes, modern techniques for unstructured
meshes only pay off for elements with polynomial orders of at least two; see [19, 22].
In this paper we will present a novel matrix-free approach for low-order finite elements designed
for the hierarchical hybrid grids framework (HHG); see e.g. [4, 3, 5]. HHG offers significantly more
geometric flexibility than undeformed hexahedral meshes. It is based on two interleaved ideas.
The first one is a special discretization of the problem domain. In HHG, the computational grid
is created by way of a uniform refinement following the rules of [6], starting from a possibly
unstructured simplicial macro mesh. The elements of this macro mesh are called macro-elements
and the resulting sub-triangulation reflects a uniform structure within these macro-elements. The
second aspect is based on the fact that each row of the global finite element stiffness matrix can
be considered as a difference stencil. This notion and point of view is classical on structured
grids and recently has found re-newed interest in the context of finite elements too; see e.g., [12].
In combination with the HHG grid construction this implies that for linear simplicial elements
one obtains stencils with identical structure for each inner node of a macro primitive, i.e., vertex,
edge, face, and element of the macro mesh. If additionally the coefficients of the PDE are constant
per macro-element, then also the stencil entries are exactly the same. Consequently, only a few
different stencils (one per primitive) can occur and need to be stored. This leads to extremely
efficient matrix-free techniques, as has been demonstrated e.g., in [3, 16].
Let us now consider the setting of an elliptic PDE with piecewise smooth variable coefficients,
assuming that the macro mesh resolves jumps in the coefficients. In this case, a standard finite
element formulation is based on quadrature formulas and introduces a variational crime. According
to [10, 27], there is flexibility how the integrals are approximated without degenerating the order
of convergence. This has recently been exploited in [2] with a method that approximates these
integral values on-the-fly using suitably surrogate polynomials with respect to the macro mesh.
The resulting two-scale method is able to preserve the convergence order if the coarse and the fine
scale are related properly. Here we propose an alternative which is based on the fine scale.
For this article, we restrict ourselves to the lowest order case of conforming finite elements
on simplicial meshes. Then the most popular quadrature formula is the one point Gauss rule
which in the simplest case of div(k∇u) as PDE operator just weights the element based reference
stiffness matrix of the Laplacian by the factor of k(xT ) where xT is the barycenter of the element
T . Alternatively, one can selectPa purely vertex-based quadrature formula. Here, the weighting of
d+1
the element matrix is given by i=1 k(xiT )/(d + 1), where d is the space dimension and xiT are the
vertices of element T . Using a vertex-based quadrature formula saves function evaluations and is,
thus, attractive whenever the evaluation of the coefficient function is expensive and it pays off to
STENCIL SCALING FOR FINITE ELEMENTS
3
reuse once computed values in several element stiffness matrices. Note that reusing barycentric
data on general unstructured meshes will require nontrivial storage schemes.
In the case of variable coefficient functions, stencil entries can vary from one mesh node to
1 d`
2 +
another. The number of possibly different stencils within each macro-element becomes d!
O(2(d−1)` ), where ` is the number of uniform refinement steps for HHG. Now we can resort to
two options: Either these stencils are computed once and then saved, effectively creating a sparse
matrix data structure, or they are computed on-the-fly each time when they are needed. Neither
of these techniques is ideal for extreme scale computations. While for the first option O(2d` )
extra memory is consumed and extensive memory traffic occurs, the second option requires recomputation of O(2d` ) local contributions.
The efficiency of a numerical PDE solver can be analyzed following the textbook paradigm [8]
that defines a work unit (WU) to be the cost of one application of the discrete operator for a
given problem. With this definition, the analysis of iterative solvers can be conducted in terms of
WU. Classical multigrid textbook efficiency is achieved when the solution is obtained in less than
10 WU. For devising an efficient method it is, however, equally critical to design algorithms that
reduce the cost of a WU without sacrificing accuracy. Clearly, the real life cost of a WU depends
on the computer hardware and the efficiency of the implementation, as e.g., analyzed for parallel
supercomputers in [16]. On the other side, matrix-free techniques, as the one proposed in this
article, seek opportunities to reduce the cost of a WU by a clever rearrangement of the algorithms
or by exploiting approximations where this is possible; see e.g., also [2].
These preliminary considerations motivate our novel approach to reduce the cost of a WU
by recomputing the surrogate stencil entries for a matrix-free solver more efficiently. We find
that these values can be assembled from a reference stencil of the constant coefficient case which
is scaled appropriately using nodal values of the coefficient function. We will show that under
suitable conditions, this technique does not sacrifice accuracy. However, we also demonstrate that
the new method can reduce the cost of a WU considerably and in consequence helps to reduce the
time-to-solution.
The rest of this paper is structured as follows: In section 2, we define our new scaling approach.
The variational crime is a analyzed in Section 3 where optimal order a priori results for the L2 and H 1 -norm are obtained. In section 4, we consider modifications in the pre-asymptotic regime
to guarantee uniform ellipticity. Section 5 is devoted to the reproduction property and the primitive concept which allows for a fast on-the-fly reassembling in a matrix free software framework.
In section 6, we discuss the cost compared to a standard nodal based element-wise assembling.
Finally, in section 7 we perform numerically an accuracy study and a run-time comparison to
illustrate the performance gain of the new scaling approach.
2. Problem setting and definition of the scaling approach. We consider a scalar elliptic
partial differential equation of Darcy type, i.e.,
−div K∇u = f,
in Ω,
tr u = 0
on ∂Ω
where tr stands for the boundary trace operator and f ∈ L2 (Ω). Here Ω ⊂ Rd , d = 2, 3, is a
bounded polygonal/polyhedral domain, and K denotes a uniformly positive and symmetric tensor
specified by the coefficient functions km , m = 1, . . . , M , where M ≤ 3 in 2D and M ≤ 6 in 3D
due to symmetry.
For the Darcy operator with a scalar uniform positive permeability, i.e., − div(k∇u), we can
set M = 1 and k1 := k. The above setting also covers blending finite elements approaches [17].
Here K is related to the Jacobian of the blending function. For example, if the standard Laplacian
model problem is considered on the physical domain Ωphy but the actual assembly is carried out
4
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
on a reference domain Ω := Φ(Ωphy ), we have
Z
Z
(DΦ)(DΦ)>
(1)
a(v, w) =
∇w dx,
∇vphy · ∇wphy dxphy =
∇v ·
|det DΦ|
Ωphy
Ω
where DΦ is the Jacobian of the mapping Φ, and vphy := v ◦ Φ, wphy := w ◦ Φ.
2.1. Definition of our scaling approach. The weak form associated
with the partial
R
differential equation is defined in terms of the bilinear form a(v, w) := Ω ∇v · K∇w dx, and the
weak solution u ∈ V0 := H01 (Ω) satisfies:
a(u, v) = (f, v),
v ∈ V0 .
This bilinear form can be affinely decomposed as
(2)
a(v, w) :=
M
X
Z
am (v, w),
am (v, w) :=
km (x)(Dm v, Dm w) dx,
v, w ∈ V := H 1 (Ω)
Ω
m=1
where Dm is a first order partial differential operator and (·, ·) stands for some suitable inner
product. In the case of a scalar permeability we find D1 := ∇ and (·, ·) stands for the scalar
product in Rd .
Let TH , H > 0 fixed, be a possibly unstructured simplicial triangulation resolving Ω. We
call TH also macro-triangulation and denote its elements by T . Using uniform mesh refinement,
we obtain Th/2 from Th by decomposing each element into 2d sub-elements, h ∈ {H/2, H/4, . . .};
see [6] for the 3D case. The elements of Th are denoted by t. The macro-triangulation is then
decomposed into the following geometrical primitives: elements, faces, edges, and vertices. Each of
these geometric primitives acts as a container for a subset of unknowns associated with the refined
triangulations. These sets of unknowns can be stored in array-like data structures, resulting in a
contiguous memory layout that conforms inherently to the refinement hierarchy; see [5, 16]. In
particular, the unknowns can be accessed without indirect addressing such that the overhead is
reduced significantly when compared to conventional sparse matrix data structures. Associated
with Th is the space Vh ⊂ V of piecewise linear finite elements. In Vh , we do not include the
homogeneous boundary conditions. We denote by φi ∈ Vh the nodal
to
P basis functions associated
P
the i-th mesh node. Node i is located at the vertex xi . For vh := i νi φi and wh := j χj φj , we
define our scaled discrete bilinear forms ah (·, ·) and ahm (·, ·) by
(3a)
ah (vh , wh ) :=
M
X
ahm (vh , wh ),
m=1
(3b)
ahm (vh , wh ) :=
Z
1 X X
(km |T (xi ) + km |T (xj ))(νi − νj )(χj − χi ) (Dm φi , Dm φj ) dx.
4
T
i,j
T ∈TH
This definition is motivated by the fact that am (vh , wh ) can be written as
Z
1 X X
(4)
am (vh , wh ) =
(νi − νj )(χj − χi )
km (x)(Dm φi , Dm φj ) dx.
2
T
i,j
T ∈TH
Here we have exploited symmetry and the row sum property. It is obvious that if km is a constant
restricted to T , we do obtain ahm (vh , wh ) = am (vh , wh ). In general however, the definition of
ah (·, ·) introduces a variational crime and it does not even correspond to an element-wise local
assembling based on a quadrature formula. We note that each node on ∂T is redundantly existent
in the data structure and that we can easily account for jumps in the coefficient function when
resolved by the macro-mesh elements T .
Similar scaling techniques have been used in [30] for a generalized Stokes problem from Geodynamics with coupled velocity components. However, for vectorial equations such a simple scaling
STENCIL SCALING FOR FINITE ELEMENTS
5
does asymptotically not result in a physically correct solution. For the computation of integrals
on triangles containing derivatives in the form of (2), cubature formulas of the form (3b) in combination with Euler-MacLaurin type asymptotic expansions have been applied [24, Table 1].
Remark 1. At first glance the Definition (3b) might not be more attractive than (4) regarding
the computational cost. In a matrix free approach, however, where we have to reassemble the
entries in each matrix call, (3b) turns out to be much more favorable. In order see this, we have
to recall that we work with hybrid hierarchical meshes. This means that for each inner node i in
T , we find the same entries in the sense that
Z
Z
(Dm φi , Dm φj ) dx = (Dm φl , Dm φxj +δx ) dx.
T
T
Here we have identified the index notation with the vertex notation, and the vertex xl is obtained
from the vertex xi by a shift of δx, i.e., xl = xi + δx. Consequently, the values of
R
(D
φ
m i , Dm φj ) dx do not have to be re-computed but can be efficiently stored.
T
For simplicity of notation, we shall restrict ourselves in the following to the case of the Darcy
equation with a scalar uniformly positive definite permeability; i.e., M = 1 and drop the index
m. However, the proof can be generalized to conceptually the same type of results for M > 1. In
Subsection 7.3, we also show numerical results for M = 6 in 3D.
2.2. Stencil structure. We exploit the hierarchical grid structure to save a significant
amount of memory compared to classical sparse matrix formats. Any direct neighbor xj ∈ NT (xi )
can be described through a direction vector wj such that xj = xi + wj . The regularity of the grid
in the interior of a macro-element T implies that these vectors remain the same, when we move
from one node to another node. Additionally, for each neighbor xj ∈ NT (xi ) \ {xi } there is a
mirrowed neighbor x0j of xi reachable by wj = −wj 0 ; see Fig. 1.
Let ni = |NT (xi )| denote the stencil size at mesh node xi . We define the stencil ŝTxi ∈ Rni
associated to the i-th mesh node xi restricted on T as
Z
(ŝTxi )j := (∇φxi +wj , ∇φxi ) dx .
T
The symmetry of the bilinear form yields
(ŝTxi )j = (ŝTxi +wj )j 0 .
We recall that for each mesh node xi we have ni ≤ 7 in 2D and ni ≤ 15 in 3D. Out of these entries
only 3 in 2D and 7 in 3D have to bePcomputed since the remaining ones follow from symmetry
arguments and the observation that j (ŝTxi )j = 0.
Due to the hierarchical hybrid grid structure, two stencils ŝTxi1 and ŝTxi2 are exactly the same
if xi1 and xi2 are two nodes belonging to the same primitive; i.e., we find only 15 different kinds
of stencils per macro-element in 3D (one for each vertex, edge, face, and volume primitive) and 7
in 2D. This observation allows for an extremely fast and memory-efficient on-the-fly (re)assembly
of the entries of the stiffness matrix in stencil form. For each node xi in the data structure, we
save the nodal values of the coefficient function k. With these considerations in mind, the bilinear
form (3b) can be evaluated very efficiently and requires only a suitable scaling of the reference
entries; see Sec. 6 for detailed cost considerations.
3. Variational crime framework and a priori analysis. In order to obtain order h
and h2 a priori estimates of the modified finite element approximation in the H 1 - and L2 -norm,
respectively, we analyze the discrete bilinear form. From now on, we assume that k|T ∈ W 2,∞ (T )
for each T ∈ TH . Moreover, we denote by k · k0 the L2 -norm on Ω and k · k∞ := supT ∈TH k · k∞;T
defines a broken L∞ -norm. We recall that the coefficient function is only assumed to be elementwise smooth with respect to the macro triangulation. Existence and uniqueness of a finite element
6
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
5
4
6
7
3
2
10
0
1
20
30
70
60
40
50
Fig. 1: From left to right: Exemplary local indices j and their corresponding direction vectors of
a 15 point stencil in 3D; Six elements attached to one edge; Four elements attached to one edge
solution uh ∈ Vh ∩ V0 of
ah (uh , vh ) = f (vh ),
vh ∈ Vh ∩ V0
is given provided that the following assumption (A1) holds true:
(A1) ah (·, ·) is uniformly coercive on Vh ∩ V0
(A2) |a(vh , wh ) − ah (vh , wh )| . hk∇kk∞ k∇vh k0 k∇wh k0 ,
vh , wh ∈ Vh
Here and in the following, the notation . is used as abbreviation for ≤ C, where C < ∞ is
independent of the mesh-size h. The assumption (A2), if combined with Strang’s first lemma,
yields that the finite element solution results in O(h) a priori estimates with respect to the H 1 norm; see, e.g., [10, 27]. We note that for h small enough, the uniform coercivity (A1) follows
from the consistency assumption (A2), since for vh ∈ Vh
ah (vh , vh ) ≥ a(vh , vh ) − |ah (vh , vh ) − a(vh , vh )| ≥ C(1 − ch)k∇vh k20 .
Remark 2. As it is commonly done in the finite element analysis in unweighted Sobolev norms,
we allow the generic constant C to be dependent on supΩ k/ inf Ω k, knowing that for large contrast
applications this estimate is overly pessimistic and can be improved; cf. [26] and the references
therein. However, since we are mainly interested in showing alternative assembly techniques to
the standard finite element method and in comparing them to the well-established approaches in
standard norms, we shall not work out these subtleties here.
3.1. Abstract framework for L2 -norm estimates. Since (A2) does not automatically
guarantee optimal order L2 -estimates, we employ duality arguments. To get a better feeling on
the required accuracy of ah (·, ·), we briefly recall the basic steps occurring in the proof of the upper
bound. As it is standard, we assume H 2 -regularity of the primal and the dual problem. Restricting
ourselves to the case of homogeneous Dirichlet boundaries, the dual PDE and boundary operators
coincide with the primal ones. Let us denote by Ph u the standard Galerkin approximation of u,
i.e., the finite element solution obtained as the solution of a discrete problem using the bilinear
form a(·, ·). It is well-known that under the given assumptions ku − Ph uk0 = O(h2 ). Now, to
obtain an L2 -estimate for uh , we consider the dual problem with uh − Ph u on the right-hand side.
Let w ∈ V0 be the solution of a(v, w) = (uh − Ph u, v)0 for v ∈ V0 . Due to the standard Galerkin
orthogonality, we obtain
(5)
kuh − Ph uk20 = a(uh − Ph u, w) = a(uh − Ph u, Ph w) = a(uh , Ph w) − ah (uh , Ph w).
This straightforward consideration shows us that compared to (A2), we need to make stronger
assumptions on the mesh-dependent bilinear form ah (·, ·). We define (A3) by
7
STENCIL SCALING FOR FINITE ELEMENTS
(A3) |a(vh , wh ) − ah (vh , wh )| . h2 kHkk∞ k∇vh k0 k∇wh k0 + hk∇kk∞ k∇vh k0;Sh k∇wh k0;Sh ,
where Hk denotes the Hessian of k and Sh := ∪T ∈TH Sh (T ) with Sh (T ) := {t ∈ Th ; ∂t ∩ ∂T 6= ∅};
see Fig. 2 for a 2D illustration. The semi-norm k · k0;Sh stands for the L2 -norm restricted to Sh .
Fig. 2: Elements in Sh (T ) (left) and Sh (right) for d = 2
Lemma 3. Let the problem under consideration be H 2 -regular, h be sufficiently small and (A3)
be satisfied. Then we obtain a unique solution and optimal order convergence in the H 1 - and the
L2 -norm, i.e.,
(6)
kuh − uk0 + hk∇(uh − u)k0 . h2 (kHuk0 + kHkk∞ k∇uk0 + k∇kk∞ (k∇uk0 + kHuk0 )).
Proof. Given that h is small enough, (A1) follows from (A3). In terms of (5) and (A3), we
get
(7)
kuh − Ph uk20 . h2 kHkk∞ k∇uh k0 k∇Ph wk0 + hk∇kk∞ k∇uh k0;Sh k∇Ph wk0;Sh .
The stability of the standard conforming Galerkin formulation yields k∇Ph wk0 . kuh − Ph uk0 .
By Definition (3), we find that the discrete bilinear form ah (·, ·) is uniformly continuous for a
coefficient function in L∞ (Ω), and thus k∇uh k0 . k∇uk0 . To bound the two terms involving
k∇ · k0;Sh , we use the 1D Sobolev embedding H 1 ((0, 1)) ⊂ L∞ ((0, 1)) [28]. More precisely, for an
element in H s (Ω), s > 0.5, we have kvk0;Sh . kvkH s (Ω) ; see [21]. Here we use s = 1 and in terms
of the H 2 -regularity assumption, we obtain
√
k∇uh k0;Sh ≤ k∇(u − uh )k0;Sh + k∇uk0;Sh . k∇(u − uh )k0;Sh + h(k∇uk0 + kHuk0 ),
√
√
k∇Ph wk0;Sh ≤ k∇(w − Ph w)k0;Sh + hkuh − Ph uk0 . hkuh − Ph uk0 .
Using the above estimates, (7) reduces to
3
kuh − Ph uk0 . h2 (kHkk∞ k∇uk0 + k∇kk∞ (k∇uk0 + kHuk0 )) + h 2 k∇kk∞ k∇(u − uh )k0 .
Applying the triangle inequality and using the approximation properties of the Galerkin finite
element solution results in the extra term h2 kHuk0 in the upper bound for ku − uh k0 . The
bound for hk∇(uh − u)k0 follows by a standard inverse estimate for elements
in Vh and the best
√
approximation property of Vh ∩ V0 . Since for h small enough it holds c1 hk∇kk∞ ≤ 1/2 where
c1 < ∞ is a suitably fixed positive constant, the upper bound (6) follows.
3.2. Verification of the assumptions. It is well-known [10] that assumptions (A1)-(A3)
are satisfied for the bilinear form
(8)
ãh (uh , vh ) :=
X
t∈Th
d+1
X Z
|t| X
t
t
t
k|t (xi )∇uh |t (xi ) · ∇vh |t (xi ) =
k̄t ∇uh · ∇vh dx,
d + 1 i=1
t
t∈Th
Here xti denotes the vertices of the d-dimensional simplex t, i.e., we approximate the integral by a
Pd+1
nodal quadrature rule and k̄t := ( i=1 k|t (xti ))/(d + 1). Thus, to verify the assumptions also for
8
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
ah (·, ·), it is sufficient to consider ah (vh , wh ) − ãh (vh , wh ) in more detail with ah (·, ·) given by (3).
Let
t
t
t
t
t
t
a1,1
at1,2
Ât :=
...
at1,d+1
a1,2
at2,2
..
.
at2,d+1
...
...
...
a1,d+1
at2,d+1
,
..
.
t
ad+1,d+1
k1,1
t
k1,2
Kt := ..
.
t
k1,d+1
k1,2
t
k2,2
..
.
t
k2,d+1
...
...
...
k1,d+1
t
k2,d+1
,
..
.
t
kd+1,d+1
R
be the local stiffness matrix associated with the nodal basis functions φti , i.e., ati,j := t ∇φti ·∇φtj dx
t
and the local coefficient function with ki,j
:= 21 (k|t (xti ) + k|t (xtj )), i 6= j. The diagonal entries of
Kt are defined differently as
t
ki,i
:=
(9)
−1 X t t
ki,j ai,j .
ati,i
j6=i
We introduce the component-wise Hadamard product between two matrices as (B ◦C)ij := Bij Cij
and define the rank one matrix K̃t by (K̃t )ij := k̄t .
Due to the symmetry of Ât and the fact that the row sum of Ât is equal to zero, we can
rewrite the discrete bilinear forms. With vh , wh ∈ Vh , we associate locally elements vt , wt ∈ Rd+1
with (vt )i := vh (xti ), (wt )i := wh (xti ). We recall that if vh , wh ∈ Vh ∩ V0 and xti is a boundary
node, then vh (xti ) = 0 = wh (xti ).
Lemma 4. The bilinear forms given by (3) and (8) have the algebraic form
(10)
ah (vh , wh ) =
d+1
1 X X
((vt )i − (vt )j ) (Kt ◦ Ât )ij ((wt )j − (wt )i ) ,
2
i,j=1
vh , wh ∈ Vh ,
d+1
1 X X
((vt )i − (vt )j ) (K̃t ◦ Ât )ij ((wt )j − (wt )i ) ,
2
i,j=1
vh , wh ∈ Vh .
t∈Th
(11)
ãh (vh , wh ) =
t∈Th
Proof. We note that (9) yields that the row sum of Kt ◦ Ât is equal to zero. Moreover, Kt ◦ Ât
is by construction symmetric. Introducing for T ∈ TH the set Thi,j;T of all elements T ⊃ t ∈ Th
sharing the global nodes i and j, we identify by it the local index of the node i associated with
the element t and by jt the local index of the global node j. Then the standard local to global
assembling process yields that the right hand side in (10) reads
X
1 X X
kitt jt atit jt .
(νi − νj )(χj − χi )
2
i,j;T
i,j
T ∈TH
t∈Th
Comparing this result with the definition (3b), we find equality since the coefficient function is
assumed to be smooth within each T . The proof of (11) follows with exactly the same arguments
as the one for (10).
Although the proof of the previous lemma is straightforward, the implication of it for large scale
simulations cannot be underestimated. In 2D, the number of different edge types per macroelement is three while in 3D it is seven assuming uniform refinement. All edges in 3D in the
interior of a macro-element T share only four or six elements; see Fig. 1. We have three edge
types that have four elements attached to them and four edge types with six adjacent elements.
The algebraic formulations (11) and (10) allow us to estimate the effects of the variational
crime introduced by the stencil scaling approach.
Lemma 5. Assumptions (A2) and (A3) hold true.
Proof. The required O(h) bound for (A2) is straightforward and also holds true for any unstructured mesh refinement strategy. Recalling that Kt = K̃t if the coefficient function k restricted
STENCIL SCALING FOR FINITE ELEMENTS
9
to t is a constant, (11) and (10) yield
d+1
1 X X
((vt )i − (vt )j ) (Kt − K̃t ) ◦ Ât
((wt )j − (wt )i )
2
ij
t∈Th i,j=1
. h2−d max λmax (Kt − K̃t )λmax (Ât ) k∇vh k0 k∇wh k0
t∈Th
. max λmax (Kt − K̃t ) k∇vh k0 k∇wh k0 . hk∇kkL∞ k∇vh k0 k∇wh k0 ,
|ah (vh , wh ) − ãh (vh , wh )| =
t∈Th
where λmax (·) denotes the maximal eigenvalue of its argument.
To show (A3), we have to exploit the structure of the mesh Th . Let EhT be the set of all edges
in T̄ and The;T the subset of elements t ⊂ T which share the edge e having the two global nodes i
and j as endpoints. As before, we identify the local indices of these endpoints by it and jt . We
note that the two sets EhT and EhT̂ , T 6= T̂ are not necessarily disjoint. Observing that each element
t ⊂ T is exactly contained in 21 d(d + 1) elements of The;T , we find
ah (vh , wh )−ãh (vh , wh ) =
X X X
1
((vt )it − (vt )jt ) (Kt − K̃t ) ◦ Ât
((wt )jt − (wt )it ),
d(d + 1)
it j t
T
e;T
T ∈TH e∈Eh t∈T
h
and thus it is sufficient to focus on the contributions resulting from t ∈ The;T . We consider two
cases separately: First, we consider the case that the edge e is part of ∂T for at least one T , then
we directly find the upper bound
hk∇kkL∞ (t) k∇vh kt k∇wh kt .
Second, we consider the case that e is in the interior of one T . Then for each element t ∈ The;T
there exists exactly one tm ∈ The;T such that tm is obtained by point reflection at the midpoint of
the edge e; see Fig. 3. In the following, we exploit that the midpoint of the edge is the barycenter
of t ∪ tm . Here the local indices it and jt are associated with the global nodes i and j, respectively.
m
Without loss of generality, we assume a local renumbering such that it = 1, jt = 2 and that xtim
t
is the point reflected vertex of xtit ; see also Fig. 3.
j
3
2 1
t
tm
1
3
2
i
(a) Local numbering along an inner edge e
in 2D
(b) Local numbering along an inner edge e
in 3D
Fig. 3: Local numbering in element t and its point reflected element tm in 2D and 3D
10
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
Let us next focus on
t
Eij
(vh , wh ) := ((vt )1 − (vt )2 )((Kt − K̃t ) ◦ Ât )1,2 ((wt )2 − (wt )1 )
+((vtm )2 − (vtm )1 )((Ktm − K̃tm ) ◦ Atm )2,1 ((wtm )1 − (wtm )2 ) .
t
(·, ·) by the local H 1 -seminorms
Exploiting the fact that (Ât )1,2 = (Âtm )2,1 , we can bound Eij
d+1
t
Eij
(vh , wh ) .k∇vh k0;t∪tm k∇wh k0;t∪tm (k|T (xi ) + k|T (xj )) −
m
1 X
(k|T (xtl ) + k|T (xtl )) .
d+1
l=1
m
A Taylor expansion of k in 12 (xt1 + xt1 ) = 12 (xi + xj ) guarantees that the terms of zeroth and first
order cancel out and only second order derivatives of k scaled with h2 remain, i.e.,
t
Eij
(vh , wh ) .h2 kHkkL∞ (t∪tm ) k∇vh k0;t∪tm k∇wh k0;t∪tm .
Then the summation over all macro elements, all edges, and all elements in the subsets Tle;T in
combination with a finite covering argument [14] yields the upper bound of (A3).
4. Guaranteed uniform coercivity. While for our hierarchical hybrid mesh framework the
assumptions (A2) and (A3) are satisfied and thus asymptotically, i.e., for h sufficiently small, also
(A1) is satisfied, (A1) is not necessarily guaranteed for any given mesh Th .
Lemma 6. If the matrix representation of the discrete Laplace operator is an M-matrix, then
the scaled bilinear form ah (·, ·) is positive semi-definite on Vh × Vh for k globally smooth.
Proof. Let Âh be the matrix representation of the discrete Laplace operator, P
then the Mmatrix property guarantees that (Âh )ij ≤ 0 for i 6= j. Taking into account that T ∈TH ŝTij =
(Âh )ij , definition (3) for the special case M = 1 and k globally smooth yields
ah (vh , vh ) =
−1 X
(k(xi ) + k(xj ))(νi − νj )2 (Âh )ij ≥ 0.
4 i,j
Remark 7. In 2D it is well-known [11] that if all elements of the macro mesh have no obtuse
angle, then Âh is an M-matrix and we are in the setting of Lemma 6.
4.1. Pre-asymptotic modification in 2D based on (A2). Here we work out the technical
details of a modification in 2D that guarantees uniform ellipticity assuming that at least one macroelement T has an obtuse angle. Our modification yields a linear condition on the local mesh-size
depending on the discrete gradient of k. It only applies to selected stencil directions. In 2D our
7-point stencil associated with an interior fine grid node has exactly two positive off-center entries
if the associated macro-element has an obtuse angle. We call the edges associated with a positive
reference stencil entry to be of gray type. With each macro-element T , we associate the reference
stiffness matrix ÂT . Without loss of generality, we assume that the local enumeration is done
in such a way that the largest interior angle of the macro-element T is located at the local node
3, i.e., if T has an obtuse angle then aT1,2 > 0, aT1,3 < 0, and aT2,3 < 0 and otherwise aTi,j ≤ 0,
1 ≤ i < j ≤ 3. By λTmin we denote the smallest non-degenerated eigenvalue of the generalized
eigenvalue problem
T
a1,1 aT1,2 aT1,3
2 −1 −1
2 −1 x, x ∈ R3 .
ÂT x := aT1,2 aT2,2 aT2,3 x = λ −1
T
T
T
−1 −1
2
a1,3 a2,3 a3,3
We note that both matrices in the eigenvalue problem are symmetric, positive semi-definite and
have the same one dimensional kernel and thus λTmin > 0.
11
STENCIL SCALING FOR FINITE ELEMENTS
Let e be a gray type edge. For each such edge e, we possibly adapt our approach locally. We
denote by ωe;T the element patch of all elements t ∈ Th , such that t ⊂ T and e ⊂ ∂t. Then we
define
1
ke;min := mine kẽ , kẽ := (k|T (xẽ1 ) + k|T (xẽ2 ))
ẽ∈Eh
2
where Ehe is the set of all edges being in ω̄e;T , and xẽ1 and xẽ2 are the two endpoints of ẽ. In the
pre-asymptotic regime, i.e., if
(ke − ke;min )aT1,2 > ke;min λTmin
(MA2)
we replace the scaling factor ke = 12 (k|T (xe1 ) + k|T (xe2 )) in definition (3) by
!
λTmin
mod
ke := ke;min 1 + T
.
a1,2
Then it is obvious that ke;min < kemod < ke . We note that 2aT1,2 is the value of the 7-point stencil
associated with a gray edge and thus trivial to access.
Lemma 8. Let the bilinear form be modified according to (MA2), then it is uniformly positive
definite on Vh ∩ V0 × Vh ∩ V0 for all simplicial hierarchical meshes.
Proof. As it holds true for standard bilinear forms also our modified one can be decomposed
into element contributions. The local stiffness At matrix for t ⊂ T associated with (3) reads At =
t
t
t
t
t
Kt ◦ Ât = Kt ◦ ÂT and can be rewritten in terms of kmin
:= min1≤i<j≤3 ki,j
and δki,j
:= ki,j
− kmin
as
(12)
−aT1,2
t
t T
At = kmin ÂT + δk1,2 a1,2
0
aT1,2
−aT1,2
0
0
−aT1,3
t
0 + δk1,3 0
aT1,3
0
0
0
0
0
aT1,3
t
0
0 + δk2,3
0
−aT1,3
0
−aT2,3
aT2,3
0
aT2,3
−aT2,3
t
where we use a consistent local node enumeration. We note that each δki,j
≥ 0. Now we consider
two cases separately.
First, let t ⊂ T and T be a macro-element having no obtuse angle then we find that all four
t
matrices on the right of (12) are positive semi-definite. Thus, we obtain At ≥ kmin
ÂT and no
modification is required.
Second, let t ⊂ T and T be a macro-element having one obtuse angle. Then the second matrix
on the right of (12) is negative semi-definite while the three other ones are positive semi-definite.
Now we find
T
−a1,2 aT1,2 0
t
t T
a1,2 −aT1,2 0 x
x> At x ≥ kmin
x> ÂT x + x> δk1,2
0
0
0
t
t
t
≥ λTmin kmin
((x1 − x3 )2 + (x2 − x3 )2 ) + (λTmin kmin
− aT1,2 δk1,2
)(x1 − x2 )2
t
≥ λTmin ke;min ((x1 − x3 )2 + (x2 − x3 )2 ) + (λTmin ke;min − aT1,2 δk1,2
)(x1 − x2 )2 ,
where e is the edge associated with the two local nodes 1 and 2. Provided (MA2) is not satisfied,
we have x> At x ≥ λTmin ke;min ((x1 − x3 )2 + (x2 − x3 )2 ) ≥ cke;min x> ÂT x. If (MA2) is satisfied, we
t
do not work with the bilinear form (3) but replace k1,2
with kemod . With this modification it is
now obvious that the newly defined bilinear form is positive semi-definite on Vh × Vh and moreover
positive definite on Vh ∩V0 ×Vh ∩V0 . The coercivity constant depends only on the shape regularity
of the macro-mesh, min k, and the Poincaré–Friedrichs constant.
Remark 9. From the proof it is obvious that any other positive scaling factor less or equal to
kemod also preserves the uniform ellipticity.
12
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
Remark 10. We can replace in the modification criterion the local condition (MA2) by
lT 2−(`+1) k∇kkL∞ (ωe;T ) aT1,2 ≥ inf k(x)λTmin ,
x∈ω̄e;T
where Th is obtained by ` uniform refinement steps from TH , and lT is the length of the second
longest edge in T . Both these criteria allow a local marking of gray type edges which have to
be modified. To avoid computation of ke;min each time it is needed, and thus further reducing
computational cost, we set the scaling factor for all gray type edges in a marked T to
(13)
1
(kmin (xe1 ) + kmin (xe2 )),
2
kmin (xi ) := min {k(xj ) | xj ∈ NT (xi )}
where NT (xi ) denotes the set of all mesh nodes that are connected to node xi via an edge and
belonging to the macro element T̄ including xi itself. The quantity kmin can be pre-computed for
each node xi once at the beginning and stored as a node based vector such as k is. For non-linear
problems where k depends on the solution itself, kmin can be updated directly after the update of
k.
The presented pre-asymptotic modification based on (A2) yields a condition on the local meshsize and only affects edge types associated with a positive stencil entry. However, the proof of (A3)
shows that a condition on the square of the mesh-size is basically sufficient to guarantee (A1).
This observation allows us to design an alternative modification yielding a condition on the square
of the local mesh-size and involving the Hessian of k, except for the elements which are in Sh .
However, in contrast to the option discussed before, we possibly also have to alter entries which
are associated with negative reference stencil entries, and therefore we do not discuss this case in
detail. It is obvious that for piecewise smooth k there exists an `0 such that for all refinement levels
` ≥ `0 no local modification has to be applied. This holds true for both types of modifications.
Thus, all the a priori estimates also hold true for our modified versions. Since we are interested
in piecewise moderate variations of k and large scale computations, i.e., large `, we assume that
we are already in the asymptotic regime, i.e., that no modification has to be applied for our 3D
numerical test cases, and we do not work out the technical details for the modifications in 3D.
4.2. Numerical counter example in the pre-asymptotic regime. In the case that we
are outside the setting of Lemma 6, it is easy to come up with an example where the scaled
bilinear form is not positive semi-definite, even for a globally smooth k. For a given mesh size h,
one can always construct a k with a variation large enough such that the scaled stiffness matrix
has negative eigenvalues.
Here we consider a 2D setting on the unit square with an initial mesh which is not Delaunay;
see left part of Fig. 4. For the coefficient function k, we use a sigmoid function defined as
k(x, y; m, η) =
η
+ 1,
1 + exp(−m(y − x − 0.2))
with m = 50 which yields a steep gradient. Further, we vary the magnitude of k by setting
η ∈ {1, 10, 100, 1000}.
Now, we assemble the global stiffness matrix and report in Tab. 1 for the different choices
of η its minimal and maximal eigenvalue over a sequence of uniform refinement steps. We show
the eigenvalues for our scaling approach (3) with and without modification (MA2) and for the
standard nodal integration assembly (8). We find that for η > 100 the stiffness matrix of the
scaling approach has negative eigenvalues in the pre-asymptotic regime. In these cases three or
four refinement steps are required, respectively, to enter the asymptotic regime. However, the
positive definiteness of the matrix can be recovered on coarser resolutions if (MA2) is applied.
Asymptotically, the minimal or maximal eigenvalues of all three approaches tend to the same
values.
In Fig. 4, we illustrate the action of the modification and show how the region of elements where
it has to be applied is getting smaller and finally vanishes with increasing number of refinement
13
STENCIL SCALING FOR FINITE ELEMENTS
Table 1: Minimal and maximal eigenvalues of the global stiffness matrix A. Italic entries highlight
application of (MA2).
η\ `
1
10
100
1000
0
2.7
2.8
-8.9
-128
1
0.72
0.89
-10.6
-148
2
0.18
0.24
-5.5
-94.5
η\ `
1
10
100
1000
0
2.7
3.8
4.0
4.4
1
0.72
1.1
1.3
1.3
2
0.18
0.26
0.30
0.33
η\ `
1
10
100
1000
0
2.8
5.7
27.7
247
1
0.72
1.3
1.8
1.9
2
0.18
0.27
0.34
0.38
scaling approach
λA
min
3
4
5
0
1
0.046
0.011
0.0029
9.1
13.3
0.063
0.016
0.0039
40.0
69.3
0.066
0.018
0.0045
353
633
-3.0
0.019
0.0049
3486
6265
scaling approach modified with (MA2)
λA
min
3
4
5
0
1
0.046
0.011
0.0029
9.1
13.3
0.063
0.016
0.0039
40.8
69.4
0.072
0.018
0.0045
364
634
0.078
0.019
0.0049 3598
6278
standard nodal integration
λA
min
3
4
5
0
1
0.046
0.011
0.0029
9.0
12.8
0.065
0.016
0.0040
36.6
64.9
0.076
0.018
0.0045
316
590
0.084
0.020
0.0049
3116
5840
λA
max
2
3
15.1
17.4
80.7
93.2
739
853
7317
8450
4
19.8
108
985
9759
5
21.3
116
1066
10562
λA
max
2
3
15.1
17.4
80.7
93.2
739
853
7321
8453
4
19.8
108
985
9759
5
21.3
116
1066
10562
λA
max
2
3
14.7
17.1
78.4
90.7
717
829
7104
8211
4
19.7
107
977
9682
5
21.2
116
1064
10549
steps. To further illustrate how the region that is affected by (MA2) changes, we show a second
example where all initial elements have an obtuse angle. The underlying color bar represents the
coefficient function k with η = 1000 for the first (left), and η = 100 for the second (right) example.
The location of the steepest gradient of k is marked by a dashed line. Note that the gradient of
k is constant along lines parallel to the dashed one. For all edges that are modified by (MA2)
the adjacent elements are shown. The color intensity refers to the refinement level and goes from
bright (initial mesh) to dark.
Fig. 4: Application of (MA2): Mesh with two critical macro elements used for Tab. 1 (left) and
mesh where all macro elements have an obtuse angle (right).
4.3. The sign of the stencil entries in 3D. In contrast to the 2D setting, a macro-mesh
with no obtuse angle does not yield that Âh is an M-matrix. Here the uniform refinement rule
yields that for each macro element three sub-classes of tetrahedra (gray, blue, green) exist. To each
of these we associate one interior edge type (gray, blue, green) defined by not being parallel to any
of the six edges of the respective tetrahedron type. Fig. 5 shows the sub-classes and as example
the gray edge type. The coloring of the sub-classes is up to now arbitrary. We always associate
the gray color with the macro-element and call the associated interior edge, a gray type edge. The
14
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
interior edges associated with the blue and green elements are called blue and green type edges,
respectively. All other remaining edges are by notation red type edges. If the macro-element
Fig. 5: Uniform refinement of one macro-element T (left) into three subclasses (middle); gray edge
between blue and green sub-tetrahedra (right).
T has no obtuse angle between two faces, then it follows from [20, 18] that the reference stencil
entries, i.e., the entries associated with the Laplace operator, associated with gray type edges have
a positive sign; see Fig. 6. However, it can be shown by some simple geometrical considerations
that not both, green and blue type edges, can have a positive sign. If one of these has a positive
sign, we call this the blue- and the other one the green-type edge. Conversely, if both have a
negative sign, the coloring is arbitrary. In case the macro-element T has no obtuse angle, the sign
associated with all red type edges is automatically not positive. Thus, we find for such elements in
our 15-point stencil either two or four positive off-diagonal entries. Then a modification similar to
the one proposed for the 2D case can be now applied to the edges of gray and possibly blue type.
The situation can be drastically different if the macro-element T has an obtuse angle between two
faces. In that case, we find up to four edge directions which carry a positive sign in the stencil.
5. Reproduction property and primitive concept. As already mentioned, our goal is
to reduce the cost of a WU and the run-times while preserving discretization errors that are
qualitatively and quantitatively on the same level as for standard conforming finite elements.
Here we focus on the 3D case, but similar results can be obtained for the 2D setting. The a priori
bounds of Lemma 3 do not necessarily guarantee that for an affine coefficient function k and affine
solution u the error is equal to zero. In the upper bound (6) a term of the form k∇kk∞ k∇uk0
remains. A closer look at the proof reveals that this non-trivial contribution can be traced back to
the terms associated with nodes on the boundary of a macro-element. This observation motivates
us to introduce a modification of our stencil scaling approach. As already mentioned, all nodes
are grouped into primitives and we have easy access to the elements of these primitives. Recall
that ãh (·, ·) defined as in (8) is associated with the standard finite element approach with nodal
quadrature. Let us by WV , WE , and WF denote the set of all nodes associated with the vertex,
edge, and face primitives, respectively. Now we introduce a modified stencil scaling approach,
cf. (3) and (8),
ãh (vh , φi ),
i∈I
(14)
ah;I (vh , φi ) :=
.
ah (vh , φi )
i 6∈ I
Replacing ah (·, ·) by ah;I (·, ·) with I ⊂ W := WV ∪ WE ∪ WF still yields that all node stencils
associated with a node in a volume primitive are cheap to assemble. The number of node stencils
which have to be more expensively assembled grow only at most with 4` while the total number
of nodes grows with 8` .
Remark 11. The modification (14) introduces an asymmetry in the definition of the stiffness
matrix to which multigrid solvers are not sensitive. Moreover, the asymmetry tends asymptotically
15
STENCIL SCALING FOR FINITE ELEMENTS
+8.3e-02
-5.9e-02
-3.3e-01
+8.3e-02
-8.3e-02
-8.3e-02
-1.7e-01
-2.9e-01
-3.3e-01
-8.3e-02
+8.3e-02
+2.4e+00
-2.9e-01
-2.9e-01
-2.9e-01
+1.2e-01
+8.3e-02
(a) Unit reference tetrahedron:
x1 = (0, 0, 0), x2 = (1, 0, 0)
x3 = (0, 1, 0), x4 = (0, 0, 1)
-5.9e-02
-5.9e-02
-8.3e-02
-3.3e-01
+1.2e-01
-2.9e-01
-1.7e-01
+1.7e+00
-3.3e-01
-2.9e-01
-2.9e-01
-2.9e-01
-5.9e-02
(b) Regular tetrahedron:
√
√
x1 = (1, 0, −1/ 2) , x2 = (−1, 0, −1/ 2)
√
√
1
1
x3 = (0, 1, + / 2) , x4 = (0, −1, + / 2)
Fig. 6: Stencil entries colored by their sign at an inner node of two times refined tetrahedra,
each without any obtuse angles between faces. On the left (A), the stencil of the unit reference
tetrahedron and on the right (B), the stencil of a regular tetrahedron is depicted. The gray edge
corresponds to the interior edge through 21 (x1 + x3 ) and 21 (x2 + x4 ). The green edge is the one
in direction between x1 and x3 and the blue one in direction between x2 and x4 . All other edge
directions are marked in red.
to zero with O(h2 ).
Lemma 12. Let I = W or I = WV ∪ WE , then an affine solution can be reproduced if k is
affine, i.e., uh = u.
Proof. For i 6∈ W, associated with a node in the macro-element T and k affine, we have
kij (ŝTi )j = (sTi )j , and thus the bilinear form ah;I (·, ·) is identical with ãh (·, ·) for I = W. Since
then no variational crime occurs, we find uh = u for any affine solution u.
The case I = WV ∪ WE is more involved. Here the bilinear forms ah;I (·, ·) and ãh (·, ·) are not
identical. However, it can be shown that for any affine function vaff we find
ah;I (vaff , ·) = ãh (vaff , ·)
on Vh ∩ V0 . To see this, we follow similar arguments as in the proof of Lemma 5. This time we
do not use a point reflected element tm but a shifted element ts ; see Fig. 7. We note that ∇vaff
is constant and
s
1
1
(k(xi ) + k(xi + wj )) − (k(xi ) + k(xi + wj 0 )) = k(xtc ) − k(xtc ).
2
2
(15)
s
Here xtc and xtc stand for the barycenter of the elements t and ts , respectively. Two of the four
vertices of element t are the nodes xi and xi + wj . The shifted element ts is obtained from t by a
shift of −wj .
This observation yields for a node i 6∈ WV ∪ WE that the sum over the nodes can be split into
16
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
Used bilinear form to
define FE approximation
Discretization error
in discrete L2 -norm
ãh (·, ·) standard FE
ah;I (·, ·), I = W
ah;I (·, ·), I = WV ∪ WE
ah;I (·, ·), I = WV ∪ WF
ah;I (·, ·), I = WE ∪ WF
ah;I (·, ·), I = WV
ah;I (·, ·), I = WE
ah (·, ·) stencil scaling
3.85104e-15
2.10913e-15
1.75099e-15
6.82596e-04
9.19145e-07
6.82032e-04
9.17523e-07
6.81942e-04
(a) Element t and shifted one ts
(b) Reproduction property
Fig. 7: Illustration of the effect of the hierarchical mesh structure.
two parts
ah;I (vaff , φi ) = ah (vaff , φi ) =
1 X
(k(xj ) + k(xi ))(vaff (xj ) − vaff (xi ))(ŝTxi )j
2
j∈W
1 X
(k(xj ) + k(xi ))(vaff (xj ) − vaff (xi ))(ŝTxi )j .
+
2
j6∈W
For the second sum on the right, we have already shown equality to the corresponding term in the
bilinear form ãh (vaff , φi ). We recall that each node j in W such that xi and xj form an edge has a
point mirrored node xj 0 , i.e., xi = 0.5(xj + xj 0 ). For the first term on the right, we find that for a
node j ∈ W it holds (ŝTxi )j = (ŝTxi )j 0 . In terms of vaff (xi + wj ) − vaff (xi ) = vaff (xi ) − vaff (xi + wj 0 ),
the first summand on the right can thus be further simplified to
ni
1
1X
(k(xi + wj ) + k(xi ) − (k(xi + wj 0 ) + k(xi )))(vaff (xi + wj ) − vaff (xi )) (ŝTxi )j .
2 j=1 2
Together with (15), this yields the stated equality.
To illustrate Lemma 12, we consider a 3D example on the unit cube Ω = (0, 1)3 . We use as
solution u(x, y, z) = −7x + y + 3z and k(x, y, z) = 2x + 3y + 5z + 1 for the coefficient function. The
standard finite element solution reproduces the exact solution up to machine precision. We test
the influence of the stencil scaling approach on different sets of primitives. In the right of Fig. 7,
we report the discretization error in the discrete L2 -norm for different combinations. Note that the
macro triangulation of the unit cube consists of 12 tetrahedral elements and of one non-boundary
macro vertex at (0.5, 0.5, 0.5). These considerations show that the choice I = WV ∪ WE is quite
attractive. The number of stencils which have to be expensively evaluated grows only with 2`
while we still can guarantee the reproduction property for an affine solution in the case of an
affine coefficient function.
6. Cost of a work unit. The stencil scaling has been introduced as means to reduce the
cost of a WU and thus help to design less expensive and thus more efficient PDE solvers. We will
now resort to a cost metric based on operation count. While we are aware that real run-times
will be influenced by many additional factors, including e.g. the quality of the implementation,
compiler settings, and various hardware details, the classic measure still provides useful insight.
A hardware-aware performance analysis, as conducted in [16] to evaluate the efficiency on real
17
STENCIL SCALING FOR FINITE ELEMENTS
computers, is beyond our current scope. Since asymptotically the contributions from the element
primitives dominate the cost, we restrict ourselves to the study of stencils for nodes located in the
interior of a macro element T .
6.1. Cost for stencil scaling. Let ŝTi be the 15-point stencil associated with the Laplacian
at an inner node i of T which is independent of the node location within T . Recall that the scaled
stencil is given by
sTij :=
1
κ(xi ) + κ(xj ) ŝTij ,
2
∀j ∈ NT (i) \ {i} ,
X
sTii := −
sTij
j∈NT (i)\{i}
where κ(xi ) represents either the value of the parameter k at node xi or a modified version of it
as suggested in (13). Note that this definition is the direct translation of the approach described
in Sec. 2.1 from the bilinear form to the pure stencil. Assuming that the constant 1/2 factor is
incorporated into the stencil ŝTi at setup and taking into account the number of edges emanating
from xi , being six in 2D and 14 in 3D, we can compute the non-central stencil entries with 12 and
28 operations, respectively. The computation of the central entry via the zero row-sum property
takes 5 and 13 additions, respectively. This is summarized in Tab. 2.
We next compare the cost for the stencil scaling approach to an on-the-fly computation based
on classic FEM techniques which, however, exploits the advantages of hierarchical hybrid grids,
see also [15].
Remark 13. The cost to approximate the stencil coefficients with the two-scale interpolation
method of [2] amounts to q operations per stencil entry when the polynomial degree is chosen as
q. When the coefficients are locally smooth, q = 2 delivers good results in [2]. The cost of this
method becomes 2 × 7 = 14 operations in 2D, and 2 × 15 = 30 operations in 3D, respectively.
This is slightly cheaper than the stencil scaling of this paper, but it neglects the setup cost that
is necessary to construct the polynomials and completely ignores the fact that for a fixed macro
mesh size no asymptotic optimality can be achieved.
6.2. Cost of the on-the-fly computation for the classical FEM. Employing the bilinear
form (8) we can write the stencil coefficients as
(16)
s̃Tij = ãh (φj , φi ) =
d+1
d+1
X
X
X X
k(xt` )
∇φjt · ∇φit |t| =
k(xt` )
(E t )it ,jt
(d
+
1)
e;T
e;T
t∈Th
`=1
t∈Th
`=1
where
t
(E )it ,jt
1
:=
(d + 1)
Z
∇φjt · ∇φit dx
t
denotes the local stiffness matrix including the volume averaging factor 1/(d+1). The regularity
of the mesh inside a macro element implies that there exist only a fixed number of differently
shaped elements t. Thus, we can compute stencil entries on-the-fly from one (2D) or three (3D)
precomputed stiffness matrices and the node-based coefficient values2 . In 2D, always six elements
t are attached to an interior node. Summing nodal values of k for each element first would require
12 operations. Together with the fact that two elements are attached to each edge, computing all
six non-central stencil entries via (16) would then require a total of 30 operations. This number
can be reduced by eliminating common sub-expressions. We pre-compute the values
k(xi ) + k(j2 ) ,
k(xi ) + k(j4 ) ,
k(xi ) + k(j6 ) ,
2 Note that in our HHG implementation, we use six element-matrices as this is advantageous with respect to
local to global indexing; see [3] for details.
18
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
j5
j4
j6
j3
j1
j2
Fig. 8: Neighborhood of node i and edges [red/dashed] selected for common sub-expression elimination.
i.e.,
k for the vertices at the ends of the edges marked red in Fig. 8. Then, we obtain
P3 we sum
t
`=1 k(x` ) by adding the value of k at the third vertex of t to the pre-computed expression. In
this fashion, we can compute all six sums with 9 operations and obtain a total of 27 operations
for the non-central entries.
The situation in 3D is more complicated. This stems mainly from the fact that the number
of elements sharing the edge from node i to node j is no longer a single value as in 2D. Instead
we have in 3D that |The;T | = 4 for the gray, blue, and green type edges that emanate from node i,
while |The;T | = 6 for the remaining 8 red type edges. Thus, for the computation of the non-central
stencil entries we obtain the cost
6(4 + 3) + 8(6 + 5) + 40 = 170 .
The factor 40 represents the operations required for summing the nodal k values employing again
elimination of common sub-expressions; see [15] for details. Assembling the central entry involves
contributions from 6 elements in 2D and 24 in 3D. Thus, executing this via (16) cannot require
less operations than using the row-sum property.
6.3. Cost of stencil application. The cost of the application of the assembled stencil is
independent of the approach to compute the stencil. Using the first expression in
X
X
sTi v =
sTij v(xj ) =
sTij (v(xj ) − v(xi ))
j∈NT (i)
j∈NT (i)\{i}
results in 13 operations in 2D and 29 in 3D. Exploiting the row-sum property avoids computing
the central entry, but increases the cost for the stencil application, cf. second expression, so that
in total only a single operation is saved. Tab. 2 summarizes the results of this section. We observe
that assembling the stencil and applying it once in the scaling approach requires only about 1/3 of
the operations compared to a classical on-the-fly variant.
approach
dimension
non-central entries
central entry
assembly + application
stencil scaling
2D
3D
6 add / 6 mult
14 add / 14 mult
5 add / 0 mult
13 add / 0 mult
17 add / 12 mult
41 add / 28 mult
on-the-fly FEM
2D
3D
15 add / 12 mult
98 add / 72 mult
5 add / 0 mult
13 add / 0 mult
26 add / 18 mult
125 add / 86 mult
Table 2: Comparison of operation count for assembling and applying the stencil using either stencil
scaling or on-the-fly assembly.
In 3D, the stencil scaling approach results in a total of 69 operations per node, and is thus only
STENCIL SCALING FOR FINITE ELEMENTS
2.4 times more expensive than
save a factor of 3 compared to
mized on-the-fly assembly and
variants of on-the-fly-assembly
19
the constant-coefficient case treated in e.g. [3]. The 69 operations
the 211 operations per node for the conventional, yet highly optiroughly an order of magnitude when compared with unoptimized
techniques.
To which extent a factor three savings in floating point operations will be reflected in run-time
(or other practical cost metrics, such as e.g., energy consumption), depends on many details of the
hardware and system software. For instance, we note that in the above algorithms, the addition
and multiplication operations are not ideally balanced and they do not always occur such that
the fused multiply-add operations of a modern processor architecture can be used. Furthermore,
advanced optimizing compilers will restructure the loops and will attempt to find a scheduling
of the instruction stream that avoids dependencies. Thus, in effect, the number of operations
executed on the processor may not be identical to the number of operations calculated from the
abstract algorithm or the high level source code. Nevertheless, though the operation count does
not permit a precise prediction of the run-times, we will see in the following that our effort to
reduce the number of operations pays off in terms of accelerated execution.
7. Numerical accuracy study and run-time comparison. In this section, we provide
different numerical results which illustrate the accuracy and run-time of the new scaling approach
in comparison to the element-wise finite element assembling based on nodal integration within a
matrix free framework. We consider different cases such as scalar and tensorial coefficient functions
k and the scenario of a geometry mapping. Throughout this section, we denote the time-to-solution
by tts and by the relative tts we always mean the ratio of the time-to-solution of the stencil scaling
approaches with respect to the nodal integration. From our theoretical considerations for one
stencil application from Tab. 2, we expect a relative tts of roughly one third. Further, we denote
the estimated order of convergence by eoc and the asymptotic convergence rates of the multigrid
1/i∗ −5
∗
solver by ρ defined as ρ = r(i ) /r(5)
where r(i) is the L2 residual at iteration i, and i∗ the
final iteration of the solver. Each of the following 3D computations was conducted on SuperMUC
Phase 2 using the Intel 17.0 compiler together with the Intel 2017 MPI library. For all runs, we
specify the compiler flags -O3 -march=native -xHost. Note that the serial runs using only a
single compute core are not limited to run on large machines like SuperMUC but can also be run
on usual modern desktop workstations with enough memory. The peak memory usage by our
largest serial run was at about 4.46 GiB.
7.1. A quantitative comparison in 2D for a scalar permeability. In this example, we
consider as domain the unit-square Ω = (0, 1)2 and use a non-polynomial manufactured solution
u(x, y) =
x4 y
.
xy + 1
We consider as coefficient function k(x, y; m) = 2 + sin(mπx) sin(mπy) with m ∈ {2, 4, 8}. The
right hand side is computed by inserting the above definitions into the equation. This construction
has the advantage that we can study the effect of the magnitude of ∇k in a systematic fashion by
adjusting m. We perform a study on a regular triangular mesh comparing errors of the discrete
solutions obtained by two standard Galerkin finite element approaches and our proposed scaling
approach. Results are listed in Tab. 3. Here, unh denotes the approximation obtained by employing
a nodal quadrature rule resulting in bilinear form (8), while ubh uses a quadrature rule that evaluates
k at the triangle’s barycenter.
7.2. A quantitative comparison in 3D for a scalar permeability. As a second test,
we consider a non-linear solution similar to the 2D tests in the previous section
u=
x3 y + z 2
,
xyz + 1
20
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
midpoint integration
L ku − ub
eoc |u − ub
h k0
h |1
eoc
0
1
2
3
4
3.75e-03
9.61e-04
2.42e-04
6.06e-05
1.51e-05
0.00
1.96
1.99
2.00
2.00
9.13e-02
4.60e-02
2.31e-02
1.15e-02
5.77e-03
0.00
0.99
1.00
1.00
1.00
0
1
2
3
4
3.61e-03
9.14e-04
2.30e-04
5.77e-05
1.44e-05
0.00
1.98
1.99
2.00
2.00
9.25e-02
4.62e-02
2.31e-02
1.15e-02
5.78e-03
0.00
1.00
1.00
1.00
1.00
l
0
1
2
3
4
ku − ub
h k0
4.06e-03
9.16e-04
2.32e-04
5.97e-05
1.50e-05
eoc |u − ub
h |1
0.00 1.01e-01
2.15 4.87e-02
1.98 2.35e-02
1.96 1.16e-02
1.99 5.78e-03
eoc
0.00
1.06
1.05
1.02
1.01
nodal integration
ku − un
k
eoc |u − un
h 0
h |1
m=2
3.25e-03 0.00 9.15e-02
8.59e-04 1.92 4.61e-02
2.18e-04 1.97 2.31e-02
5.50e-05 1.99 1.15e-02
1.37e-05 2.00 5.78e-03
m=4
3.64e-03 0.00 9.32e-02
1.02e-03 1.83 4.70e-02
2.81e-04 1.87 2.33e-02
7.27e-05 1.95 1.15e-02
1.83e-05 1.99 5.78e-03
m=8
ku − un
eoc |u − un
h k0
h |1
4.71e-03 0.00 1.07e-01
1.14e-03 2.04 5.09e-02
4.02e-04 1.51 2.55e-02
1.21e-04 1.73 1.20e-02
3.20e-05 1.92 5.84e-03
eoc
stencil-approach
ku − ush k0 eoc |u − ush |1
eoc
0.00
0.99
1.00
1.00
1.00
3.34e-03
8.65e-04
2.18e-04
5.48e-05
1.37e-05
0.00
1.95
1.98
2.00
2.00
9.14e-02
4.61e-02
2.31e-02
1.15e-02
5.78e-03
0.00
0.99
1.00
1.00
1.00
0.00
0.99
1.02
1.01
1.00
3.54e-03
9.40e-04
2.43e-04
6.13e-05
1.53e-05
0.00
1.91
1.95
1.98
2.00
9.36e-02
4.67e-02
2.31e-02
1.15e-02
5.78e-03
0.00
1.00
1.01
1.00
1.00
eoc |u − ush |1
0.00 1.03e-01
2.04 5.23e-02
1.73 2.44e-02
1.90 1.17e-02
1.97 5.80e-03
eoc
0.00
0.98
1.10
1.06
1.02
eoc ku − ush k0
0.00 4.47e-03
1.08 1.09e-03
1.00 3.29e-04
1.08 8.80e-05
1.04 2.24e-05
Table 3: Results for a regular triangular mesh. Here, L denotes the (uniform) refinement level
and eoc the estimated order of convergence.
with a parameter dependent scalar coefficient function k(x, y, z; m) = cos(mπxyz) + 2 on the unitcube Ω = (0, 1)3 discretized by six tetrahedra. We employ a multigrid solver with a V(3,3) cycle
on a single compute core and stop after 10 multigrid iterations, i.e., i∗ = 10, which is enough to
reach the asymptotic regime. In Tab. 4, we report convergence of the discretization error for the
three approaches, i.e., classical FE with nodal integration, stencil scaling on volumes and faces, but
classical FE assembly on edges and vertices, and stencil scaling on all primitives. The refinement
is given by L where L = −2 denotes the macro mesh. We observe quadratic convergence of the
discrete L2 error for all three approaches and a relative tts of about 32% on level L = 6.
nodal integration
scale Vol+Face
L
DOF
error
eoc
ρ
1
2
3
4
5
6
3.43e+02
3.38e+03
2.98e+04
2.50e+05
2.05e+06
1.66e+07
2.46e-03
7.06e-04
1.80e-04
4.46e-05
1.11e-05
2.75e-06
–
1.80
1.97
2.01
2.01
2.01
0.07
0.13
0.16
0.18
0.17
0.16
1
2
3
4
5
6
3.43e+02
3.38e+03
2.98e+04
2.50e+05
2.05e+06
1.66e+07
3.17e-03
1.50e-03
4.83e-04
1.31e-04
3.32e-05
8.30e-06
–
1.08
1.63
1.89
1.98
2.00
0.07
0.15
0.19
0.19
0.20
0.21
error
eoc
m=3
2.42e-03
–
5.97e-04 2.02
1.46e-04 2.03
3.59e-05 2.02
8.88e-06 2.01
2.21e-06 2.01
m=8
5.43e-03
–
1.54e-03 1.82
4.04e-04 1.93
1.01e-04 1.99
2.52e-05 2.01
6.29e-06 2.01
scale all
ρ
rel.
tts
error
eoc
ρ
rel.
tts
0.06
0.12
0.15
0.15
0.14
0.13
1.13
0.45
0.34
0.30
0.31
0.32
2.51e-03
6.05e-04
1.47e-04
3.59e-05
8.88e-06
2.21e-06
–
2.05
2.05
2.03
2.02
2.01
0.07
0.12
0.15
0.15
0.14
0.13
0.70
0.45
0.28
0.29
0.32
0.32
0.09
0.15
0.17
0.16
0.16
0.16
1.33
0.53
0.31
0.30
0.31
0.33
5.21e-03
1.56e-03
4.06e-04
1.02e-04
2.53e-05
6.29e-06
–
1.74
1.94
2.00
2.01
2.01
0.08
0.15
0.16
0.17
0.17
0.16
0.70
0.45
0.36
0.29
0.32
0.32
Table 4: 3D results in the case of a scalar coefficient function with errors measured in the discrete
L2 -norm.
7.3. A quantitative comparison in 3D for a permeability tensor. As a third test, we
consider a full symmetric and positive definite permeability tensor K with non-linear components.
21
STENCIL SCALING FOR FINITE ELEMENTS
L
1
2
3
4
5
6
DOF
8.55e+02
7.47e+03
6.26e+04
5.12e+05
4.15e+06
3.34e+07
nodal integration
error
eoc
ρ
1.92e-03
–
0.07
4.40e-04 2.13 0.16
1.06e-04 2.05 0.24
2.60e-05 2.02 0.30
6.47e-06 2.01 0.35
1.61e-06 2.00 0.37
scale Vol+Face
error
eoc
ρ
2.21e-03
–
0.07
5.19e-04 2.09 0.16
1.27e-04 2.03 0.24
3.15e-05 2.01 0.30
7.86e-06 2.00 0.35
1.96e-06 2.00 0.37
rel.
tts
0.60
0.35
0.25
0.26
0.30
0.31
Table 5: 3D results in the case of a tensorial permeability with discretization errors measured in
the discrete L2 -norm.
The off-diagonal components are negative or zero:
2
x + 2y 2 + 3z 2 + 1
−y 2
2
2
−y
2x + 3y 2 + z 2 + 1
K=
2
−z
−x2
−z 2
−x2
2
2
2
3x + y + 2z + 1
The manufactured solution is set to
u=
x4 y + 2z
,
xyz + 1
and we consider as domain the unit-cube discretized by twelve tetrahedra. We employ the same
multigrid solver as in the previous subsection on a single compute core, but stop the iterations if
the residual is reduced by a factor of 10−9 . We denote the final iteration by i∗ . The results for
the classical FE approach and the stencil scaling approach on volumes and faces with classical FE
assembly on edges and vertices are reported in Tab. 5. We observe quadratic convergence of the
discrete L2 error for both approaches and a relative tts of 31% on our finest level L = 6.
7.4. Application to a blending setting and large scale results. To demonstrate the
advantages of our novel scaling approach also for a more realistic scenario, we consider an example
using a blending function, as mentioned in Sec. 2. To this end, we consider a half cylinder mantle
with inner radius r1 = 0.8 and outer radius r2 = 1.0, height z1 = 4.0 and with an angular
coordinate between 0 and π as our physical domain Ωphy . The cylinder mantle is additionally
warped inwards by w(z) = 0.2 sin (zπ/z1 ) in axial direction. The mapping Φ : Ωphy → Ω is given
by
p
x2 + y 2 + w(z)
√
Φ(x, y, z) = arccos (x/ x2 +y2 )
z
with the reference domain Ω = (r1 , r2 ) × (0, π) × (0, z1 ). Using (1), it follows for the mapping
tensor K
0 2
w (z) + 1
0
w0 (z)
>
p
(DΦ)(DΦ)
1/x2 +y 2
0
0 .
K=
= x2 + y 2
|det DΦ|
w0 (z)
0
1
Obviously, this tensor is symmetric and positive definite. In addition to the geometry blending,
we use a variable material parameter a(x, y, z) = 1 + z. On the reference domain Ω this yields the
PDE − div aK∇u = f . As analytic solution on the reference domain we set
x̂ − r1
π cos (4ŷ) exp (ẑ/2) .
u(x̂, ŷ, ẑ) = sin
r2 − r1
The analytic solution mapped to the physical domain is illustrated in the right part of Fig. 9.
22
S. BAUER, D. DRZISGA, M. MOHR, U. RÜDE, C. WALUGA, AND B. WOHLMUTH
Fig. 9: Macro mesh of the reference domain Ω (left) and analytic solution u (right) mapped to
the physical domain Ωphy .
For our numerical experiments, we employ a macro mesh composed of 9540 hexahedral blocks,
where each block is further split into six tetrahedral elements; see Fig. 9 (left). The resulting
system is solved using 14 310 compute cores, i.e., we assign four macro elements per core. For
the largest run we have a system with O 1011 DOF. We employ a multigrid solver with a V(3,3)
cycle. The iteration is stopped when the residual has been reduced by a factor of 10−8 . In Tab. 6,
we report the resulting discretization error, the asymptotic multigrid convergence order ρ, and the
time-to-solution.
L
1
2
3
4
5
6
DOF
4.7e+06
3.8e+07
3.1e+08
2.5e+09
2.0e+10
1.6e+11
error
2.43e-04
6.00e-05
1.49e-05
3.72e-06
9.28e-07
2.32e-07
nodal integration
eoc
ρ
tts
0.522
2.5
2.02 0.536
4.2
2.01 0.539
12.0
2.00 0.538
53.9
2.00 0.536
307.2
2.00 0.534 1822.2
error
2.38e-04
5.86e-05
1.46e-05
3.63e-06
9.06e-07
2.26e-07
scale Vol+Face
eoc
ρ
0.522
2.02 0.536
2.01 0.539
2.00 0.538
2.00 0.536
2.00 0.534
tts
2.0
2.6
4.5
15.3
88.9
589.6
rel.
tts
0.80
0.61
0.37
0.28
0.29
0.32
Table 6: Results for large scale 3D application with errors measured in the discrete L2 -norm.
These results demonstrate that the new scaling approach maintains the discretization error,
as is expected on structured grids from our variational crime analysis, as well as the multigrid
convergence rate. For small L we observe that the influence of vertex and edge primitives is
more pronounced as the improvement in time-to-solution is only small. But for increasing L this
influence decreases and for L > 4 run-time is reduced to about 30%.
Acknowledgments. The authors gratefully acknowledge the Gauss Centre for Supercomputing (GCS) for providing computing time on the supercomputer SuperMUC at Leibniz-Rechenzentrum (LRZ).
REFERENCES
[1] P. Arbenz, G. H. van Lenthe, U. Mennel, R. Müller, and M. Sala. A scalable multi-level preconditioner
for matrix-free µ-finite element analysis of human bone structures. International Journal for Numerical
Methods in Engineering, 73(7):927–947, 2008.
[2] S. Bauer, M. Mohr, U. Rde, J. Weismller, M. Wittmann, and B. Wohlmuth. A two-scale approach for efficient
on-the-fly operator assembly in massively parallel high performance multigrid codes. Applied Numerical
Mathematics, 122:14 – 38, 2017.
[3] B. Bergen. Hierarchical Hybrid Grids: Data Structures and Core Algorithms for Efficient Finite Element
Simulations on Supercomputers. PhD thesis, Technische Fakultät der Friedrich-Alexander-Universität
Erlangen-Nürnberg, 2006. SCS Publishing House.
STENCIL SCALING FOR FINITE ELEMENTS
23
[4] B. Bergen and F. Hülsemann. Hierarchical hybrid grids: Data structures and core algorithms for multigrid.
Numer. Linear Algebra Appl., 11:279–291, 2004.
[5] B. Bergen, G. Wellein, F. Hülsemann, and U. Rüde. Hierarchical hybrid grids: achieving TERAFLOP
performance on large scale finite element simulations. International Journal of Parallel, Emergent and
Distributed Systems, 22(4):311–329, 2007.
[6] J. Bey. Tetrahedral grid refinement. Computing, 55(4):355–378, 1995.
[7] J. Bielak, O. Ghattas, and E.-J. Kim. Parallel octree-based finite element method for large-scale earthquake
ground motion simulation. Computer Modeling in Engineering & Sciences, 10(2):99–112, 2005.
[8] A. Brandt. Barriers to achieving textbook multigrid efficiency (TME) in CFD. Institute for Computer
Applications in Science and Engineering, NASA saidLangley Research Center, 1998.
[9] G. F. Carey and B.-N. Jiang. Element-by-element linear and nonlinear solution schemes. Communications
in Applied Numerical Methods, 2(2):145–153, 1986.
[10] P. G. Ciarlet. The finite element method for elliptic problems. Studies in Mathematics and its Applications.
North-Holland Publishing Co., Amsterdam-New York-Oxford, 1978.
[11] P. G. Ciarlet and P.-A. Raviart. Maximum principle and uniform convergence for the finite element method.
Computer Methods in Applied Mechanics and Engineering, 2(1):17–31, 1973.
[12] C. Engwer, R. D. Falgout, and U. M. Yang. Stencil computations for pde-based applications with examples
from DUNE and Hypre. Concurrency and Computation: Practice and Experience, page 13, 2017.
[13] C. Flaig and P. Arbenz. A highly scalable matrix-free multigrid solver for µfe analysis based on a pointerless octree. In I. Lirkov, S. Margenov, and J. Waśniewski, editors, Large-Scale Scientific Computing:
8th International Conference, LSSC 2011, Sozopol, Bulgaria, June 6-10, 2011, Revised Selected Papers,
pages 498–506. Springer Berlin Heidelberg, 2012.
[14] L. E.-R. Gariepy and L. C. Evans. Measure theory and fine properties of functions, 1992.
[15] B. Gmeiner. Design and Analysis of Hierarchical Hybrid Multigrid Methods for Peta-Scale Systems and
Beyond. PhD thesis, Technische Fakultät der Friedrich-Alexander-Universität Erlangen-Nürnberg, 2013.
[16] B. Gmeiner, U. Rüde, H. Stengel, C. Waluga, and B. Wohlmuth. Towards textbook efficiency for parallel
multigrid. Numer. Math. Theor. Meth. Appl., 8(01):22–46, 2015.
[17] W. J. Gordon and C. A. Hall. Transfinite element methods: blending-function interpolation over arbitrary
curved element domains. Numer. Math., 21:109–129, 1973/74.
[18] S. Korotov, M. Křı́žek, and P. Neittaanmäki. Weakened acute type condition for tetrahedral triangulations
and the discrete maximum principle. Math. Comp., 70(233):107–119, 2001.
[19] M. Kronbichler and K. Kormann. A generic interface for parallel cell-based finite element operator application.
Computers and Fluids, 63:135–147, 2012.
[20] M. Křı́žek and L. Qun. On diagonal dominance of stiffness matrices in 3d. East–West J. Numer. Math.,
3:59–69, 1995.
[21] J. Li, J. M. Melenk, B. Wohlmuth, and J. Zou. Optimal a priori estimates for higher order finite elements for
elliptic interface problems. Appl. Numer. Math., 2010.
[22] K. Ljungkvist. Matrix-free finite-element computations on graphics processors with adaptively refined unstructured meshes. In Proceedings of the 25th High Performance Computing Symposium, HPC ’17, pages
1:1–1:12. Society for Computer Simulation International, 2017.
[23] K. Ljungkvist and M. Kronbichler. Multigrid for matrix-free finite element computations on graphics processors. Technical Report 2017-006, Department of Information Technology, Uppsala University, 2017.
[24] J. Lyness and U. Rüde. Cubature of integrands containing derivatives. Numerische Mathematik, 78(3):439–
461, 1998.
[25] D. A. May, J. Brown, and L. L. Pourhiet. A scalable, matrix-free multigrid preconditioner for finite element
discretizations of heterogeneous stokes flowc. Computer methods in applied mechanies and engineering,
290:496–523, 2015.
[26] C. Pechstein and R. Scheichl. Weighted Poincaré inequalities. IMA Journal of Numerical Analysis, 33(2):652,
2013.
[27] G. Strang and G. Fix. An analysis of the finite element method. Wellesley-Cambridge Press, Wellesley, MA,
second edition, 2008.
[28] H. Triebel. Interpolation theory, function spaces, differential operators. Second edition. Johann Ambrosius
Barth, Heidelberg, 1995.
[29] B. van Rietbergen, H. Weinans, R. Huiskes, and B. Polman. Computational strategies for iterative solutions of
large fem applications employing voxel data. International Journal for Numerical Methods in Engineering,
39(16):2743–2767, 1996.
[30] W.-S. Yang. Variable viscosity thermal convection at infinite Prandtl number in a thick spherical shell. PhD
thesis, University of Illinois at Urbana-Champaign, 1997.
| 5cs.CE
|
arXiv:1711.02326v1 [cs.AI] 7 Nov 2017
S PARSE ATTENTIVE BACKTRACKING :
L ONG -R ANGE C REDIT A SSIGNMENT
N ETWORKS
IN
R ECURRENT
Nan Rosemary Ke12 , Anirudh Goyal1 , Olexa Bilaniuk1 , Jonathan Binas1
Laurent Charlin13 , Chris Pal12 , Yoshua Bengio1†
1
MILA, Université de Montréal
2
Polytechnique Montréal
3
HEC Montréal
†
CIFAR Senior Fellow
A BSTRACT
A major drawback of backpropagation through time (BPTT) is the difficulty of
learning long-term dependencies, coming from having to propagate credit information backwards through every single step of the forward computation. This
makes BPTT both computationally impractical and biologically implausible. For
this reason, full backpropagation through time is rarely used on long sequences,
and truncated backpropagation through time is used as a heuristic. However, this
usually leads to biased estimates of the gradient in which longer term dependencies are ignored. Addressing this issue, we propose an alternative algorithm,
Sparse Attentive Backtracking, which might also be related to principles used by
brains to learn long-term dependencies. Sparse Attentive Backtracking learns an
attention mechanism over the hidden states of the past and selectively backpropagates through paths with high attention weights. This allows the model to learn
long term dependencies while only backtracking for a small number of time steps,
not just from the recent past but also from attended relevant past states.
1
I NTRODUCTION
Recurrent Neural Networks (RNNs) are state-of-the-art for many machine learning sequence processing tasks. Examples where models based on RNNs shine include speech recognition (Miao
et al., 2015; Chan et al., 2016), image captioning (Xu et al., 2015; Vinyals et al., 2015; Lu et al.,
2016), machine translation (Bahdanau et al., 2014; Sutskever et al., 2014; Luong et al., 2015), and
speech synthesis (Mehri et al., 2016). It is common practice to train these models using backpropagation through time (BPTT), wherein the network states are unrolled in time and gradients are
backpropagated through the unrolled graph. Since the parameters of an RNN are shared across the
different time steps, BPTT is more prone to vanishing and exploding gradients (Hochreiter, 1991;
Bengio et al., 1994; Hochreiter, 1998) than equivalent deep feedforward networks with as many
stages. This makes credit assignment particularly difficult for events that have occurred many time
steps in the past, and thus makes it challenging in practice to capture long-term dependencies in the
data (Hochreiter, 1991; Bengio et al., 1994). Having to wait for the end of the sequence in order to
compute gradients is neither practical for machines nor for animals (should we wait for the end of
our life to learn?), when the dependencies extend for very long times. It considerably slows down
training (the rate of convergence depends crucially on how often we can make updates of the parameters). For instance, consider a student taking a test, the outcome of which depends on their decision
several months back in the past: how much to study. If the student is using BPTT, the training signal
(the derivative of the test score with respect to the student’s state) would need to travel through many
time steps in which the student was not actively thinking about the test. However, a human would
probably learn to change that early decision when, after failing the test, he remembers it. This is the
inspiration for the proposed approach.
In practice, proper long-term credit assignment in RNNs is very inconvenient, and it is common
practice to employ truncated versions of BPTT for long sequences (Sak et al., 2014; Saon et al.,
1
2014). In truncated BPTT (TBPTT), gradients are backpropagated only for a fixed and limited
number of time steps and parameters are updated after each such subsequence. Truncation is often
motivated by computational concerns: memory, computation time and the advantage of faster learning obtained when making more frequent updates of the parameters rather than having to wait for
the end of the sequence. However, it makes capturing correlations across distant states even harder.
In the student example from before, if the student was learning using TBPTT, then no gradient signal
would flow through to the hidden states in which the student had made the relevant decision: how
much to study for the test.
Regular RNNs are parametric: their hidden state vector has a fixed size. We believe that this is a
critical element in the classical analysis of the difficulty of learning long-term dependencies (Bengio
et al., 1994). Indeed, the fixed state dimension becomes a bottleneck through which information has
to flow, both forward and backward.
We thus propose a semi-parametric RNN, where the next state is potentially conditioned on all the
previous states of the RNN, making it possible—thanks to attention—to jump through any distance
through time. We distinguish three types of states in our proposed semi-parametric RNN:
• The fixed-size hidden state h(t) , the conventional state of an RNN model at time t;
• The monotonically-growing macrostate M = {m(1) , . . . , m(s) }, the array of all past microstates, which plays the role of a random-access memory;
• And the fixed-size microstate m(i) , which is the ith hidden state (one of the h(t) ) that was
chosen for inclusion within the macrostate M.
There are as many hidden states as there are timesteps in the sequence being analyzed by the RNN.
A subset of them will become microstates, and this subset is called the macrostate.
The computation of the next hidden state h(t+1) is based on the whole macrostate M, in addition to
the external input x(t) . The macrostate being variable-length, we must devise a special mechanism
to read from this ever-growing array. As a key component of our model, we propose to use an
attention mechanism over the microstate elements of the macrostate.
The attention mechanism in the above setting may be regarded as providing adaptive, dynamic
skip connections: any past microstate can be linked, via a dynamic decision, to the current hidden
state. Skip connections allow information to propagate over very long sequences. Such architectures
should naturally make it easier to learn long-term dependencies. We name our algorithm sparse
attentive backtracking (SAB). SAB is especially well-suited to sequences in which two parts of a
task are closely related yet occur very far apart in time.
Inference in SAB involves examining the macrostate and selecting some of its microstates. Ideally, SAB will not select all microstates, instead attending only to the most salient or relevant ones
(e.g., emotionally loaded, in animals). The attention mechanism will select a number of relevant
microstates to be incorporated into the hidden state. During training, local backpropagation of gradients happens in a short window of time around the selected microstates only. This allows for the
updates to be asynchronous with respect to the time steps we attend to, and credit assignment takes
place more globally in the proposed algorithm.
With the proposed framework for SAB, we present the following contributions:
• A principled way of doing sparse credit assignment, based on a semi-parametric RNN.
• A novel way of mitigating exploding and vanishing gradients, based on reducing the number of steps that need to be backtracked through temporal skip connections.
• Competitive results compared to full backpropagation through time (BPTT), and much
better results as compared to Truncated Backpropagation through time, with significantly
shorter truncation windows in our model.
Mechanisms such as SAB may also be biologically plausible. Imagine having taken a wrong turn on
a roadtrip and finding out about it several miles later. Our mental focus would most likely shift directly to the location in time and space where we had made the wrong decision, without replaying in
reverse the detailed sequence of experienced traffic and landscape impressions. Neurophysiological
findings support the existence of such attention mechanisms and their involvement in credit assignment and learning in biological systems. In particular, hippocampal recordings in rats indicate that
brief sequences of prior experience are replayed both in the awake resting state and during sleep,
both of which conditions are linked to memory consolidation and learning (Foster & Wilson, 2006;
2
Davidson et al., 2009; Gupta et al., 2010). Moreover, it has been observed that these replay events
are modulated by the reward an animal does or does not receive at the end of a task in the sense that
they are more pronounced in the presence of a reward signal and less pronounced or absent in the
absence of a reward signal (Ambrose et al., 2016). Thus, the mental look back into the past seems
to occur exactly when credit assignment is to be performed.
2
2.1
R ELATED W ORK
T RUNCATED BACKPROPAGATION T HROUGH T IME
When training on very long sequences, full backpropagation through time becomes computationally expensive and considerably slows down training by forcing the learner to wait for the end of
each (possibly very long sequence) before making a parameter update. A common heuristic is to
backpropagate the loss of a particular time step through only a limited number of time steps, and
hence truncate the backpropagation computation graph (Williams & Peng, 1990). While truncated
backpropagation through time is heavily used in practice, its inability to perform credit assignment
over longer sequences is a limiting factor for this algorithm, resulting in failure cases even in simple
tasks, such as the Copying Memory and Adding task in (Hochreiter & Schmidhuber, 1997).
2.2
D ECOUPLED N EURAL I NTERFACES
The Decoupled Neural Interfaces method (Jaderberg et al., 2016) replaces full backpropagation
through time with synthetic gradients, which are essentially small networks, mapping the hidden
unit values of each layer to an estimator of the gradient of the final loss with respect to that layer.
While training the synthetic gradient module requires backpropagation, each layer can make approximate gradient updates for its own parameters in an asynchronous way by using its synthetic
gradient module. Thus, the network learns how to do credit assignment for a particular layer from
a few examples of the gradients from backpropagation, reducing the total number of times that
backpropagation needs to be performed.
2.3
A PPROXIMATE F ORWARD -M ODE O NLINE RNN S
Online credit assignment in RNNs without backtracking remains an open research problem. One
approach (Ollivier et al., 2015) attempts to solve this problem by estimating gradients using an approximation to forward mode automatic differentiation instead of backpropagation. Forward mode
automatic differentiation allows for computing unbiased gradient estimates in an online fashion,
however it normally requires storage of the gradient of the current hidden state values with respect
to the parameters, which is O(N 3 ) where N is the number of hidden units. The Unbiased Online
Recurrent Optimization (UORO) (Tallec & Ollivier, 2017) method gets around this by updating a
rank-1 approximation to this gradient tensor, which is shown to keep the estimate of the gradient
unbiased, but potentially at the risk of increasing the variance of the gradient estimator.
2.4
S KIP - CONNECTIONS AND G RADIENT F LOW
Neural architectures such as Residual Networks (He et al., 2016) and Dense Networks (Huang et al.,
2016) allow information to skip over convolutional processing blocks of an underlying convolutional
network architecture. In the case of Residual Networks identity connections are used to skip over
convolutional processing blocks and this information is recombined using addition. This construction provably mitigates the vanishing gradient problem by allowing the gradient at any given layer
to be bounded. Densely-connected convolutional networks alleviate the vanishing gradient problem
by allowing a direct path from any point in the network to the output. In contrast, here we propose
and explore what one might regard as a form of dynamic skip connection, modulated by an attention
mechanism.
3 S PARSE ATTENTIVE BACKTRACKING
We now introduce the idea of Sparse Attentive Backtracking (SAB). Classical RNN models such as
those based on LSTMs or GRUs only use the previous hidden state in the computation of the next
one, and therefore struggle with extremely long-range dependencies. SAB sidesteps this limitation
by additionally allowing the model to select and use (a subset of) any of the past microstates in the
computation of the next hidden state. In doing so the model may potentially reference microstates
computed arbitrarily long ago in time.
3
Figure 1: This figure illustrates the forward pass in SAB for the configuration ktop = 3,
katt = 1, ktrunc = 2. This involves Microstate Selection (§ 3.2), Summarization of microstates (§ 3.3), and incorporation
into next microstate
(§ 3.4). Red
h
i
(t)
arrows depict how attention weights at{−8,−7,−6,−5,−4,−3,−2,−1} are evaluated, first
by broadcasting the current provisional hidden state ĥ(t) against the macrostate
(which, in the presented case of katt = 1, consists of all past hidden states),
concatenating, then passing the result to an affine transformation. The attention
weights are then run through the sparsifier which selects the ktop = 3 attention weights, while the others are zeroed out. Black arrows show the microstates
(t)
(t)
(t)
corresponding to the non-zero sparse attention weights {ãt−6 , ãt−3 , ãt−2 }, namely
(t−6)
(t−6)
(t−3)
(t−3)
(t−2)
(t−2)
{m
=h
,m
=h
,m
=h
}, being weighted, summed, then
incorporated into ĥ(t) to compute the current final hidden state h(t) .
Since the classic RNN models do not support such operations on their past, we make a few architectural additions. On the forward pass of a training step, a mechanism is introduced that selects
microstates from the macrostate, summarizes them, then incorporates this summary into the next
hidden state. The hidden state may or may not become a microstate. On the backward pass, the
gradient is allowed to flow not only through the (truncated) master chain linking consecutive hidden
states, but also to the microstates which are selected in the forward pass.
In the forward pass, the microstate selection process can be denser or sparser, and the summarization
and incorporation can be more or less sophisticated. In the backward pass, the gating of gradient
flow from a hidden state to its ancestor microstates can also be denser or sparser, although it can be
no denser than the forward pass was.
For instance, it is possible for the forward pass to be dense, incorporating a summary of all microstates, but for the backward pass to be sparse, only allowing gradient flow to some of the microstate contributors to the hidden state (Dense Forward, Sparse Backward). Another possibility
is for the forward pass to be sparse, making only a few, hard, microstate selections for the summary. In this case, the backward pass will necessarily also be sparse, since few microstates will have
contributed to the hidden state, and therefore to the loss (Sparse Forward, Sparse Backward).
Noteworthy is that not all hidden states need be eligible to become microstates. In practice, we have
found that restricting the pool of eligible hidden states to only every katt ’th one still works well, while
reducing both memory and computation expense. Such an increase in the granularity of microstate
selection can also improve performance, by preventing the model from attending exclusively to the
most recent hidden states and temporally spreading microstates out from each other.
3.1
U NDERLYING RNN A RCHITECTURE
The SAB algorithm is widely applicable, and is compatible with numerous RNN architectures, including vanilla, GRU and LSTM models. However, since it necessarily requires altering the hidden4
Figure 2: This figure illustrates the backward pass in SAB for the configuration ktop = 3, katt = 1,
ktrunc = 2. The gradients are passed to the microstates selected in the forward pass and
a local truncated backprop is performed around those microstates. Blue arrows show the
gradient flow in the backward pass. Red crosses indicate TBPTT truncation points, where
the gradient stops being backpropagated.
to-hidden transition function substantially, it’s currently incompatible with the accelerated RNN
kernels offered by e.g. NVIDIA on its GPU devices through cuDNN library (Chetlur et al., 2014).
For vanilla and GRU-inspired RNN architectures, SAB’s selection and incorporation mechanisms
operate over the (hidden) state. For the LSTM architecture, which we adopt for our experiments,
they operate over the hidden state but not the cell state.
3.2
M ICROSTATE S ELECTION
The microstate selection mechanism determines which microstate subset of the macrostate will be
selected for summarization on the forward pass of the RNN, and which subset of that subset will
receive gradient on the backward pass during training. This makes it the core of the attention mechanism of a SAB implementation.
While the selection mechanism may use hard-coded attention heuristics, there is no reason why the
microstate selection mechanism could not itself be a (deep) neural network trained alongside the
RNN model over which it operates.
In the models we use here, the selection mechanism is chosen to be a linear transformation that
computes a scalar attention weight ai for each eligible microstate vector m(i) , and a sparsifier that
masks out all but the ktop greatest attention weights, producing the sparse attention weights ãi . We
empirically demonstrate that even this simple mechanism learns to focus on past time steps relevant
to the current one, thus successfully performing credit assignment. The use of a higher complexity
model here would be an interesting avenue for future research.
3.3
S UMMARIZATION OF M ICROSTATES
The selected microstates must be somehow combined into a fixed-size summary for incorporation
into the next hidden state. While many options exist for doing so, we choose to simply perform a
summation of the microstates, weighted by their sparsified attention weight ãi .
3.4
I NCORPORATION I NTO N EXT M ICROSTATE
Lastly, the summary must be incorporated into the hidden state. Again, multiple options exist, such
as addition (as done in ResNets) or concatenation (as done in DenseNets).
For our purposes we choose to simply sum the summary into the provisional hidden state output
ĥ(t) computed by the LSTM cell to produce the final hidden state h(t) that will be conditioned upon
in the next timestep.
5
3.5
I MPLEMENTATION D ETAILS
We now give the equations for the specific SAB-augmented LSTM model we use in our experiments.
At time t, the underlying LSTM receives a vector of hidden states h(t−1) , a vector of cell states
c(t−1) and an input x(t) , and computes a provisional hidden state vector ĥ(t) that also serves as a
provisional output.
We next use an attention mechanism that is similar to (Bahdanau et al., 2014), but modified to produce sparse discrete attention decisions. First, the provisional hidden state vector ĥ(t) is concatenated to each microstate vector m(i) . Then, an affine transformation maps each such concatenated
(t)
vector to an attention weight ai representing the salience of the microstate i at the current time t.
This can be equivalently expressed as:
(t)
ai = w1> m(i) + w2> ĥ(t) + bi
(1)
where the weights matrices w1 and w2 and the bias vector b are learned parameters.
Following this, we apply a piece-wise linear function that sparsifies the attention while making
discrete decisions. (This is different from typical attention mechanisms that normalize attention
(t)
weights using a Softmax function (Bahdanau et al., 2014), whose output is never sparse). Let aktop
be the ktop th greatest-valued attention weight at time t; then the sparsified attention weights are
computed as
(t)
(t)
(t)
(2)
ãi = ReLU ai − aktop
(t)
This has the effect of zeroing all attention weights less than aktop , thus masking out all but the ktop
most salient microstates in M. The few selected microstates receive gradient information, while no
gradient flows to the rest.
A summary vector s(t) is then obtained using a weighted sum over the macrostate, employing the
sparse attention weights:
X
(t)
s(t) =
ãi m(i)
(3)
m(i) ∈M
Given that this sum is very sparse, the summary operation is very fast.
To incorporate the summary into the final hidden state at timestep t, we simply sum the summary
and the provisional hidden state:
h(t) = ĥ(t) + s(t)
(4)
Lastly, to compute the output at the time step t, we concatenate h(t) and the sparse attention weights
ã(t) , then apply an affine output transform to compute the output. This can be equivalently expressed
as:
y (t) = V1> h(t) + V2> ã(t) + b
(5)
where the weights matrices V1 and V2 and bias vector b are learned parameters.
In summary, for a given time step t, a hidden state h(i) selected by the hard-attention mechanism
has two paths contributing to the hidden states h(t) in the forward pass. One path is the regular
sequential forward path in an RNN; the other path is through the dynamic skip connections in the
attention mechanism. When we perform backpropagation through the skip connections, gradient
only flows from h(t) to microstates m(i) selected by the attention mechanism (those for which
(t)
ãi > 0).
4
E XPERIMENTS
We now report and discuss the results of an empirical study that analyses the performance of SAB
using five different tasks. We first study synthetic tasks—the copying and adding problems (Hochreiter & Schmidhuber, 1997) designed to measure models’ abilities to learn long-term dependencies—
meant to confirm that SAB can successfully perform credit assignment for events that have occurred
many time steps in the past. We then study more realistic tasks and larger datasets.
6
Baselines We compare the quantitative performance of our model against two LSTM baselines
(Hochreiter & Schmidhuber, 1997). The first is trained with backpropagation through time (BPTT)
and the second is trained using truncated backpropagation through time (TBTPP). Both methods are
trained using teacher forcing (Williams & Zipser, 1989). We also used gradient clipping (that is,
we clip the gradients to 1 to avoid exploding gradients). Hyperparameters that are task-specific are
discussed in the tasks’ respective subsections, other hyperparameters that are also used by SAB and
that we set to the same value are discussed below.
Compared to standard RNNs, our model has two additional hyperparameters:
• ktop , the number of most-salient microstates to select at each time step for passing gradients
in the backward pass
• katt , the granularity of attention. Every katt th hidden state is chosen to be a microstate. The
special case katt = 1 corresponds to choosing all hidden states to be microstates as well.
In addition, we also study the impact of the TBPTT truncation length, which we denote as ktrunc .
This determines how many timesteps backwards to propagate gradients through in the backward
pass. This effect of this hyperparameter will also be studies for the LSTM with TBTPP baseline.
For all experiments we used a learning rate of 0.001 with the Adam (Kingma & Ba, 2014) optimizer unless otherwise stated. For SAB, we attend to every second hidden states, i.e. katt =2, unless
otherwise stated.
Our main findings are:
1. SAB performs almost optimally and significantly outperforms both full backpropagtion
through time (BPTT), and truncated backpropagation through time (TBPTT) on the synthetic copying task.
2. For the synthetic adding, two language modelling task (using PennTree Bank and Text8),
and sequential MNIST classification tasks, SAB reaches the performance of BPTT and
outperforms TBPTT. In addition, for the adding task, SAB outperforms TBPTT using much
shorter truncation lengths.
4.1
T HE COPYING MEMORY PROBLEM
The copying memory task tests the model’s ability to memorize salient information for long time
periods. We follow the setup of the copying memory problem from Hochreiter & Schmidhuber
(1997). In details, the network is given a sequence of T + 20 inputs consisting of: a) 10 (randomly
generated) digits (digits 1 to 8) followed by; b) T blank inputs followed by; c) a special end-ofsequence character followed by; d) 10 additional blank inputs. After the end-of-sequence character
the network must output a copy of the initial 10 digits.
Tables 1, 2, and 3 report both accuracy and cross-entropy (CE) of the models’ predictions on unseen
sequences. We note that SAB is able to learn this copy task almost perfectly for all sequence-lengths
T . Further, SAB outperforms all baselines. This is particularly noticeable for longer sequences, for
example, when T is 300 the best baseline achieves 35.9% accuracy versus SAB’s 98.9%.
To better understand the learning process of SAB, we visualized the attention weights while learning
the copying task (T = 200, ktrunc = 10, ktop = 10). Figure 3 (appendix) shows the attention
weights (averaged over a single mini-batch) at three different learning stages of training, all within
the first epoch. We note that the attention quickly (and correctly) focuses on the first ten timesteps
which contain the input digits.
4.2 T HE A DDING TASK
The adding task requires the model to sum two specific entries in a sequence of T (input) entries
(Hochreiter & Schmidhuber, 1997). In the spirit of the copying task, larger values of T will require
the model to keep track of longer-term dependencies. The exact setup is as follows. Each example in
the task consists of 2 input vectors of length T . The first, is a vector of uniformly generated values
between 0 and 1. The second vector encodes binary a mask that indicates which 2 entries in the
first input to sum (it consists of T − 2 zeros and 2 ones). The mask is randomly generated with the
constraint that masked-in entries must be from different halves of the first input vector.
7
Method
LSTM (full BPTT)
LSTM (TBPTT, ktrunc = 5)
LSTM (TBPTT, ktrunc = 10)
LSTM (TBPTT, ktrunc = 20)
SAB (ktrunc =1, ktop =1)
SAB (ktrunc =1, ktop =10)
SAB (ktrunc =5, ktop =5)
SAB (ktrunc =10, ktop =10)
Copy Length (T)
Accuracy (%)
CE (last 10 chars)
Cross Entropy
100
100
100
100
100
100
100
100
98.8
31.0
29.6
30.5
17.6
62.7
99.6
100.0
0.030
1.737
1.772
1.714
2.044
0.964
0.017
0.000
0.002
0.145
0.148
0.143
0.170
0.080
0.001
0.001
Table 1: Test accuracy and cross-entropy loss performance on the copying task with sequence
lengths of T = 100. Models that use TBPTT cannot solve this task while SAB and BPTT
can both achieve optimal performance.
Method
LSTM (full BPTT)
LSTM (TBPTT, ktrunc = 5)
LSTM (TBPTT, ktrunc = 10)
LSTM (TBPTT, ktrunc = 20)
LSTM(ktrunc =150)
SAB (ktrunc =1, ktop =1)
SAB (ktrunc =5, ktop =5)
SAB (ktrunc =10, ktop =10)
Copy Length (T)
Accuracy (%)
CE (last 10 chars)
Cross Entropy
200
200
200
200
200
200
200
200
56.0
17.1
20.2
35.8
35.0
29.8
99.3
98.7
1.07
2.03
1.98
1.61
1.596
1.818
0.032
0.049
0.046
0.092
0.090
0.073
0.073
0.083
0.001
0.002
Table 2: Test accuracy and cross-entropy loss performance on the copying task with sequence
lengths of T = 200. Different configurations of SAB all reach near optimal performance.
Method
LSTM (full BPTT)
LSTM (TBTT, ktrunc = 1)
LSTM (TBPTT, ktrunc = 20)
LSTM (TBPTT, ktrunc = 150)
SAB (ktrunc =1, ktop =1)
SAB (ktrunc =1, ktop =5)
SAB (ktrunc =5, ktop =5)
Copy Length (T)
Accuracy (%)
CE (last 10 chars)
Cross Entropy
300
300
300
300
300
300
300
35.9
14.0
25.7
24.4
26.7
46.5
98.9
0.197
2.077
1.848
1.857
0.27
1.340
0.048
0.047
0.065
0.197
0.058
0.059
0.042
0.002
Table 3: Test accuracy and cross-entropy loss performance on copying task with sequence lengths
of T = 300. On these long sequences SAB’s performance can still be very close to optimal.
Tables 4 and 5 report the cross-entropy (CE) of the model’s predictions on unseen sequences. When
T = 200, SAB’s performance is similar to the best performance of both baselines. With even longer
sequences (T = 400), SAB outperforms the TBPTT but is outperformed by BPTT.
Method
Adding Length (T)
Cross Entropy
200
200
200
200
200
0.0000
0.0011
0.0003
0.0004
0.0001
LSTM (full BPTT)
LSTM (TBPTT, ktrunc = 20)
LSTM (TBPTT, ktrunc =50)
SAB (ktrunc =5, ktop =10)
SAB (ktrunc =10, ktop =10)
Table 4: Performance on unseen sequences of the T = 200 adding task. We note that all methods
have configurations that allow them to perform near optimally.
8
Method
Adding Length (T)
Cross Entropy
400
400
400
400
0.00000
0.00068
0.00037
0.00014
LSTM (full BPTT)
LSTM (TBPTT, ktrunc =100)
SAB (ktrunc =5, ktop =10, katt =5)
SAB (ktrunc =10, ktop =10, katt =10)
Table 5: Performance on unseen sequences of the T = 400 adding task. BPTT slightly outperforms
SAB which outperforms TBPTT.
4.3
C HARACTER L EVEL P ENN T REE BANK (PTB)
We evaluate our model on language modelling task using the Penn TreeBank dataset (Marcus et al.,
1993). Our LSTM baselines use 1000 hidden units and a learning rate of 0.002. We used nonoverlapping sequences of 100 in the batches of 32. We trained SAB for 100 epochs.
We evaluate the performance of our model using the bits-per-character (BPC) metric. As shown in
Table 6, we perform slightly worse than BPTT, but better than TBPTT.
Method
Valid BPC
Test BPC
1.48
1.57
1.54
1.52
1.51
1.49
1.38
1.47
1.44
1.43
1.42
1.40
LSTM (full BPTT)
LSTM (TBPTT, ktrunc =1)
LSTM (TBPTT, ktrunc =5)
LSTM (TBPTT, ktrunc =20)
SAB (TBPTT, ktrunc =5, ktop =5, katt = 2)
SAB (TBPTT, ktrunc =5, ktop =10, katt = 2)
Table 6: BPC evaluation on the validation set of the character-level PTB (lower is better).
4.4
T EXT 8
This dataset is derived from the text of Wikipedia and consists of a sequence of a total of 100M characters (non-alphabetical and non-space characters were removed). We follow the setup of Mikolov
et al. (2012); use the first 90M characters for training, the next 5M for validation and the final 5M
characters for testing. We train on non-overlapping sequences of length 180. Due to computational
constraints, all baselines use 1000 hidden units. We trained all models using a batch size of 64. We
trained SAB for a maximum of 30 epochs. We have not done any hyperparameter search for our
model as it’s computationally expensive.
Table 7 reports BPC of the model’s predictions on the validation and test sets. Note that SAB’s
performance closely matches BPTT, and also significantly outperforms TBPTT.
Method
LSTM (full BPTT)
LSTM (TBPTT, ktrunc =5)
SAB (ktrunc =5, ktop =5, katt =5)
Valid BPC
Test BPC
1.54
1.64
1.56
1.51
1.60
1.53
Table 7: Bit-per-character (BPC) Results on the validation and test set for Text8 (lower is better).
4.5
P IXEL - BY- PIXEL MNIST
Our last task is a sequential version of the MNIST classification dataset. The task involves predicting
the label of the image after being given the image pixel by pixel (in scanline order). All models use
an LSTM with 128 hidden units. The prediction is produced by passing the final hidden state of
the network into a softmax. We used a learning rate of 0.001. We trained our model for about 100
epochs, and did early stopping based on the validation set. Table 8 shows that SAB performs about
as well as BPTT.
9
Method
Valid Accuracy (%)
Test Accuracy (%)
LSTM (full BPTT)
91.2
90.3
SAB (ktrunc =14, ktop =5, katt = 14)
90.6
89.8
SAB (ktrunc =14, ktop =10, katt = 14)
92.2
90.9
SAB (ktrunc =10, ktop =10, katt = 10)
92.2
91.1
Table 8: Test and validation accuracy for the sequential MNIST classification task. The performance
of all methods is similar on this task.
5
F UTURE W ORK
An interesting direction for future development of the Sparse Attentive Backtracking method from
the machine learning standpoint would be improving the computational efficiency when the sequences in question are very long. Since the Sparse Attentive Backtracking method uses selfattention on every step, the memory requirement grows linearly in the length of the sequence and
computing the attention mechanism requires computing a scalar between the current hidden states
and all previous hidden states (to determine where to attend). It might be possible to reduce the
memory requirement by using a hierarchical model as done by Chandar et al. (2016), and then recomputing the states for the lower levels of the hierarchy only when our attention mechanism looks
at the corresponding higher level of the hierarchy. It might also be possible to reduce the computational cost of the attention mechanism by considering a maximum inner product search algorithm
(Shrivastava & Li, 2014), instead of naively computing the inner product with all hidden states
values in the past.
6 C ONCLUSION
Improving the modeling of long-term dependencies is a central challenge in sequence modeling,
and the exact gradient computation by BPTT is not biologically plausible as well as inconvenient
computationally for realistic applications. Because of this, the most widely used algorithm for
training recurrent neural networks on long sequences is truncated backpropagation through time,
which is known to produced biased estimates of the gradient (Tallec & Ollivier, 2017), focusing
on short-term dependencies. We have proposed Sparse Attentive Backtracking, a new biologically
motivated algorithm which aims to combine the strengths of full backpropagation through time and
truncated backpropagation through time. It does so by only backpropagating gradients through paths
selected by its attention mechanism. This allows the RNN to learn long-term dependencies, as with
full backpropagation through time, while still allowing it to only backtrack for a few steps, as with
truncated backpropagation through time, thus making it possible to update weights as frequently as
needed rather than having to wait for the end of very long sequences.
ACKNOWLEDGMENTS
The authors would like to thank Hugo Larochelle, Walter Senn, Alex Lamb, Remi Le Priol, Matthieu
Courbariaux, Gaetan Marceau Caron for the useful discussions, as well as NSERC, CIFAR, Google,
Samsung, Nuance, IBM and Canada Research Chairs for funding, and Compute Canada and
NVIDIA for computing resources. The authors would also like to thank Alex Lamb for code review. The authors would also like to express debt of gratitude towards those who contributed to
Theano over the years (now that it is being sunset), for making it such a great tool.
10
R EFERENCES
R. Ellen Ambrose, Brad E. Pfeiffer, and David J. Foster. Reverse replay of hippocampal place cells
is uniquely modulated by changing reward. Neuron, 91(5):1124 – 1136, 2016.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.
Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient
descent is difficult. IEEE transactions on neural networks, 5(2):157–166, 1994.
William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. Listen, attend and spell: A neural
network for large vocabulary conversational speech recognition. In Acoustics, Speech and Signal
Processing (ICASSP), 2016 IEEE International Conference on, pp. 4960–4964. IEEE, 2016.
Sarath Chandar, Sungjin Ahn, Hugo Larochelle, Pascal Vincent, Gerald Tesauro, and Yoshua Bengio. Hierarchical memory networks. CoRR, abs/1605.07427, 2016. URL http://arxiv.
org/abs/1605.07427.
Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. cudnn: Efficient primitives for deep learning. arXiv preprint
arXiv:1410.0759, 2014.
Thomas J Davidson, Fabian Kloosterman, and Matthew A Wilson. Hippocampal replay of extended
experience. Neuron, 63(4):497–507, 2009.
David J Foster and Matthew A Wilson. Reverse replay of behavioural sequences in hippocampal
place cells during the awake state. Nature, 440(7084):680–683, 2006.
Anoopum S Gupta, Matthijs AA van der Meer, David S Touretzky, and A David Redish. Hippocampal replay is not a simple function of experience. Neuron, 65(5):695–705, 2010.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
770–778, 2016.
Sepp Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Diploma, Technische Universität München, 91, 1991.
Sepp Hochreiter. The vanishing gradient problem during learning recurrent neural nets and problem
solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 6(02):
107–116, 1998.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
Gao Huang, Zhuang Liu, Kilian Q Weinberger, and Laurens van der Maaten. Densely connected
convolutional networks. arXiv preprint arXiv:1608.06993, 2016.
Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, and
Koray Kavukcuoglu. Decoupled neural interfaces using synthetic gradients. arXiv preprint
arXiv:1608.05343, 2016.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive
attention via a visual sentinel for image captioning. arXiv preprint arXiv:1612.01887, 2016.
Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attentionbased neural machine translation. arXiv preprint arXiv:1508.04025, 2015.
Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated
corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.
11
Soroush Mehri, Kundan Kumar, Ishaan Gulrajani, Rithesh Kumar, Shubham Jain, Jose Sotelo,
Aaron Courville, and Yoshua Bengio. Samplernn: An unconditional end-to-end neural audio
generation model. arXiv preprint arXiv:1612.07837, 2016.
Yajie Miao, Mohammad Gowayyed, and Florian Metze. Eesen: End-to-end speech recognition using
deep rnn models and wfst-based decoding. In Automatic Speech Recognition and Understanding
(ASRU), 2015 IEEE Workshop on, pp. 167–174. IEEE, 2015.
Yann Ollivier, Corentin Tallec, and Guillaume Charpiat. Training recurrent networks online without
backtracking. arXiv preprint arXiv:1507.07680, 2015.
Haşim Sak, Andrew Senior, and Françoise Beaufays. Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition. arXiv preprint
arXiv:1402.1128, 2014.
George Saon, Hagen Soltau, Ahmad Emami, and Michael Picheny. Unfolded recurrent neural networks for speech recognition. In Fifteenth Annual Conference of the International Speech Communication Association, 2014.
Anshumali Shrivastava and Ping Li. Asymmetric lsh (alsh) for sublinear time maximum inner product search (mips). In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp. 2321–2329. Curran
Associates, Inc., 2014.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks.
In Advances in neural information processing systems, pp. 3104–3112, 2014.
C. Tallec and Y. Ollivier. Unbiasing Truncated Backpropagation Through Time. ArXiv e-prints,
May 2017.
Corentin Tallec and Yann Ollivier.
arXiv:1702.05043, 2017.
Unbiased online recurrent optimization.
arXiv preprint
Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural
image caption generator. In Proceedings of the IEEE conference on computer vision and pattern
recognition, pp. 3156–3164, 2015.
Ronald J Williams and Jing Peng. An efficient gradient-based algorithm for on-line training of
recurrent network trajectories. Neural computation, 2(4):490–501, 1990.
Ronald J Williams and David Zipser. A learning algorithm for continually running fully recurrent
neural networks. Neural computation, 1(2):270–280, 1989.
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich
Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual
attention. In International Conference on Machine Learning, pp. 2048–2057, 2015.
12
7
7.1
A PPENDIX
C OMPUTATIONAL C OMPLEXITY OF SAB
The time complexity of the forward pass of both training and inference in SAB is O(tn2 ), with t
the number of timesteps and n the size of the hidden state, although our current implementation
scales as O(t2 n2 ). The space complexity of the forward pass of training is unchanged at O(tn),
but the space complexity of inference in SAB is now O(tn) rather than O(n). However, the time
cost of the backward pass of training cost is very difficult to formulate. Hidden states depend on
a sparse subset of past microstates, but each of those past microstates may itself depend on several
other, even earlier microstates. The web of active connections is, therefore, akin to a directed acyclic
graph, and it is quite possible in the worst case for a backpropagation starting at the last hidden state
to touch all past microstates several times. However, if the number of microstates truly relevant to
a task is low, the attention mechanism will repeatedly focus on them to the exclusion of all others,
and pathological runtimes will not be encountered.
7.2
G RADIENT F LOW
Our method approximates the true gradient but in a sense it’s no different than the kind of approximation made with truncated gradient, except that instead of truncating to the last ktrunc time steps,
we truncate to one skip-step in the past, which can be arbitrarily far in the past. This provides a way
of combating exploding and vanishing gradient problems by learning long-term dependencies. To
verify the fact, we ran our model on all the datasets (Text8, Pixel-By-Pixel MNIST, char level PTB)
with and without gradient clipping. We empirically found, that we need to use gradient clipping only
for text8 dataset, for all the other datasets we observed little or no difference with gradient clipping.
7.3
ATTENTION WEIGHT PLOTS
We visualize how the attention weights changes during training for the Copying Memory Task in
section 4.1. The attention weights are averaged over the batch. The salient information in a copying
task are in the first 10 steps. The figure shows how the attention learns to move towards and concentrate on the beginning of the sequence as training procedes. Note these all happened with the first
epoch of training, such that the model learns in a reasonable amount of time.
Time step attending
210
A
220
210
B
220
210
C
220
0
40
80
120
Time step attended
160
200
220
Figure 3: This figure shows how attention weights change over time for the Copying Task of copy
length 200. The vertical axis is the time step attending from timestep 210 to timestep
220. The horizontal axis is the time step being attended. The top most subfigure A is the
attention plot for iteration 400 of epoch 0, subfigure B is for iteration 800 of epoch 0 and
subfigure C is for iteration 3000 of epoch 0.
13
| 2cs.AI
|
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY
STABLE IDEALS
arXiv:1406.2188v1 [math.AC] 9 Jun 2014
GABRIEL SOSA
Abstract. Generalizing techniques that prove that Veronese subrings are Koszul, we show that Rees and
multi-Rees algebras of certain types of principal strongly stable ideals are Koszul. We provide explicit
Gröbner basis for the defining ideals of these multi-Rees algebras with squarefree initial monomials, to show
that they are also normal Cohen-Macaulay domains.
Introduction
L
Let K be a field and R a standard graded K-algebra, i.e. R = i≥0 Ri , (where each
Ri is a K-vector space), R0 = K, dimK (R1 ) = s ∈ N and Ri Rj = Ri+j , for all i, j.
By letting {b1 , . . . , bs } be a K-basis for R1 , we can consider a graded surjective map ϕ :
K[T1 , T2 , . . . , Ts ] → R, given by the rule ϕ(Ti ) = bi , and obtain that R ∼
= K[T1 , . . . , Ts ]/Kerϕ.
We call the graded ideal I = Kerϕ, the defining ideal of R. We say that R is Koszul if
its residue field K = R/m, where m is the graded maximal ideal of R, has a linear free
R-resolution (i.e. the nonzero entries of the matrices describing the maps in the free Rresolution of K are linear forms). Koszul Algebras were introduced by Priddy, in [Pri70].
Fröberg, proved in [Frö99] that if the defining ideal of R is generated by quadratic monomials,
then R is Koszul. Fröberg’s result coupled with a deformation argument provides a sufficient,
yet not necessary, condition for the Koszulness of R: the existence of a monomial order on
K[T1 , . . . , Ts ] such that I has a quadratic Gröbner basis, in this case, it is said that I satisfies
the G-quadratic condition (the G-quadratic condition is not necessary for the Koszulness of
R as evidenced by examples produced by Hibi and Ohsugi [HO99] and Caviglia [Cav09]).
Showing that the defining ideal satisfies the G-quadratic condition remains a useful technique to determine the Koszulness of standard graded algebras. In their survey paper,
Conca, De Negri and Rossi [CDR13] present a proof that the dth Veronese subring, the
subring of K[X1 , . . . , Xn ] generated by all monomials of degree d, is Koszul (the Gröbner basis for the defining ideal is defined using a concept that we will refer to as ordering). In his book, Sturmfels [Stu95], proves that Veronese type algebras, K[N ], with
N = {X1β1 X2β2 · · · Xnβn : β1 + β2 + · · · + βn = d and βi ≤ αi , for 1 ≤ i ≤ n} and d, α1 , . . . , αn
fixed nonnegative integers, are Koszul, by finding squarefree quadratic Gröbner bases for
their defining ideals and introducing the concept of sorting. A squarefree Gröbner basis
for the defining ideal of K[N ], when N is an affine semigroup, guarantees that K[N ] is
also a Cohen-Macaulay normal domain because of results by Sturmfels [Stu95] and Hochster
[Hoc72]. De Negri [DeN99] proved that the defining ideals for subalgebras of Veronese type
2010 Mathematics Subject Classification. 13A30, 13B25, 13D02, 13P10, 13P20.
Key words and phrases. Koszul Algebra, Gröbner bases, multi-Rees Algebra, Cohen-Macaulay, strongly stable ideals.
1
2
GABRIEL SOSA
generated by principal strongly stable sets, and lex-segments, also possess a squarefree Gröbner basis according to a different monomial ordering.
Strongly stable ideals play a pivotal role on computational commutative algebra because
they are in correspondance with Borel-fixed ideals when char(K) = 0, and this fact links their
study to that of generic initial ideals and their extremal properties as seen in Galligo [Gal74],
Bayer and Stillman [BS87], [BS87b], Eliahou and Kervaire [EK90] and Conca [Con04]. In
their paper, Francisco, Mermin and Schweig [FMS11] study many properties of strongly
stable ideals, amongst them a combinatorial interpretation of the Betti numbers of principal
strongly stable ideals in char(K) = 0 .
Recently, examples of Rees, and multi-Rees, algebras that are Koszul, Cohen-Macaulay
normal domains, because of the G-quadratic condition, have surfaced: Rees Algebras of Hibi
ideals, by Ene, Herzog and Mohammadi [EHM11] and the multi-Rees algebra of the direct
sum of powers of the maximal ideal by Lin and Polini [LP13].
In Section 1, we introduce the concept of monomial sets closed under comparability and
use it to establish a monomial order that will allow ideals, generated by differences between
incomparable products of monomials and their corresponding comparable product, to have
squarefree quadratic Gröbner bases.
In Section 2, we prove that these squarefree quadratic Gröbner bases are the defining
ideals for certain types of Rees and multi-Rees algebras, characterize sets closed under comparability in terms of principal strongly stable ideals and present examples of algebras that
are Koszul Cohen-Macaulay normal domains that had not been previously considered.
I would like to thank my advisor, Giulio Caviglia, for his guidance on the material presented
and his suggestions regarding the structure of this paper.
1. Preliminaries
Let S = K[X1 , X2 , . . . , Xn ] be a polynomial ring over a field K. We consider the standard
L
grading of S as a K-algebra, (i.e. S = i≥0 Si , with a K-basis for Si given by all monomials
in S of degree i, denoted Mon(Si )). Additionally, we establish the following order for the
elements of N2 : given a, b ∈ N2 , we say that a <lex b if the first nonzero entry of b − a is
positive, and the following order for the variables of S, (X1 > X2 > · · · > Xn ).
For a monomial u with deg(u) = d, we will use the standard factorization of u =
(u) (u)
(u)
(u)
(u)
X1 X2 · · · Xd , where Xs(u) ∈ {X1 , . . . , Xn }, for all 1 ≤ s ≤ d, and X1 ≥ X2 ≥
(u)
(u)
(u)
· · · ≥ Xd . We will define max(u) = X1 , and min(u) = Xd , which makes the fact that
max(u) ≥ Xs(u) ≥ min(u), for all 1 ≤ s ≤ d, natural.
Definition 1.1. Let p ≤ q, u ∈ Mon(Sp ) and v ∈ Mon(Sq ). We define:
- If p = q, the sorting of u, v, denoted sort(u, v) as (u, v) =
Q
(uv) Qq
(uv)
q+p
, j=1 Xj
j=q+1 Xj
Q
(uv) Qp
(uv)
p
j=1 X2j−1 , j=1 X2j
- The ordering of u, v, denoted ord(u, v), as the pair (û, v̂) =
Notice that deg(u) = deg(û) = deg(u), deg(v) = deg(v̂) = deg(v) and uv = ûv̂ = uv.
Additionally max(û) ≤ min(v̂) and u >rev v. Recall that u >rev v if either (a) deg(u) >
(u)
(v)
deg(v), or, (b) if deg(u) = deg(v) = d and there is 1 ≤ k ≤ d such that Xk > Xk and
(u)
(v)
Xk0 = Xk0 for all k 0 > k.
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY STABLE IDEALS
3
As an example, for u = X1 X1 X3 and v = X2 X3 X4 , ord(u, v) = (X3 X3 X4 , X1 X1 X2 ) and
sort(u, v) = (X1 X2 X3 , X1 X3 X4 ).
The new examples of Koszul, Cohen-Macaulay normal domains introduced in this paper
mainly deal with Rees and multi-Rees algebras of ideals of S. For this reason we will recall
their structure: Given ideals I1 , I2 , ..., Is ⊂ S, consider M = I1 ⊕ I2 ⊕ · · · ⊕ Is , the multi-Rees
algebra of I1 , . . . , Is over S, denoted R(M ), is the multigraded algebra:
R(M ) =
M
I1a1 ta11 · · · Isas tar s ⊂ S[t1 , . . . , ts ] ∼
= K[X1 , . . . , Xn , t1 , . . . , ts ].
(a1 ,...,as )∈Ns
We will restrict ourselves to the case where each Ii is a homogeneous monomial ideal with
minimal generators of a fixed degree di , with the requirement d1 ≤ d2 ≤ · · · ≤ ds . We
will index the minimal homogeneous generators of Ii in the following way ui1 >rev ui2 >rev
· · · >rev uini . And to ease notation in the latter parts of this paper we will let n0 = n, and
declare u0j = Xj for all 1 ≤ j ≤ n0 .
`
Consider the following sets M = si=0 {uij : 1 ≤ j ≤ ni }, N 0 = {uij ti : 1 ≤ i ≤ s and 1 ≤
j ≤ ni }, and N = N 0 ∪ {u0j : 1 ≤ j ≤ n} . With this notation, we are able to define the
following surjective map:
Ψ : K[Tij : uij ∈ M] −→ K[N ] ∼
= S[N 0 ] = R(M ) ⊂ S[t1 , . . . , ts ] ∼
= K[X1 , . . . , Xn , t1 , . . . , ts ].
Given by Ψ(Tij ) = uij ti for all uij ∈ M, with 0 ≤ i ≤ s and 1 ≤ j ≤ ni and letting t0 = 1.
This implies that R(M ) ∼
= K[Tij : uij ∈ M]/KerΨ.
Our first goal is to define properties on M so that KerΨ has a squarefree quadratic Gröbner
basis.
Definition 1.2. Let uij , ui0 j 0 ∈ M, with (i, j) <lex (i0 , j 0 ). We say that uij , ui0 j 0 are comparable, and denote it uij ≺ ui0 j 0 , if:
(a) i = i0 and (uij , ui0 j 0 ) = (uij , ui0 j 0 ), or (b) i < i0 and (uˆij , uiˆ0 j 0 ) = (uij , ui0 j 0 ).
By contrast, we will say that u, v ∈ M are incomparable if u ⊀ v and v ⊀ u.
We will say that M is closed under comparability if for all uij , ui0 j 0 ∈ M, with
(i, j) <lex (i0 , j 0 ), there are elements uil , ui0 l0 ∈ M such that uij ui0 j 0 = uil ui0 l0 and uil ≺ ui0 l0 .
An equivalent statement would be:
M is closed under comparability if for any uij , ui0 j 0 ∈ M, with (i, j) <lex (i0 , j 0 ), we have
that uij , ui0 j 0 ∈ M, if i = i0 , and uˆij , uiˆ0 j 0 ∈ M, if i < i0 .
(u )
Additionally, to ease notation later, from this point on we will denote Xk ij by Yijk .
An example of a set that is closed under comparability is presented now:
Example 1.3. Let n = 4. If u ∈ Mon(Sd ), then the principal strongly stable set of
X
X
u is the set B(u) = v ∈ Mon(S) : v = Xji1 . . . Xjit u, ik < jk with 1 ≤ k ≤ t . A monomial
1
t
ideal I is principal strongly stable if its minimal monomial generating set is the principal
strongly stable set for some u ∈ I. Consider the following principal strongly stable ideals
I1 = (B(X3 X4 )), I2 = (B(X2 X2 X3 )), I3 = (B(X1 X2 X2 )), I4 = (X15 ).
We can determine the surjective map onto the multi-Rees algebra of I1 , I2 , I3 , I4 over S, by
defining the elements of M in the following manner: d0 = 1, d1 = 2, d2 = 3, d3 = 3, d4 = 4,
4
GABRIEL SOSA
u01
u02
u03
u04
= X1
= X2
= X3
= X4
u11
u12
u13
u14
u15
= X1 X1
= X 1 X2
= X 2 X2
= X 1 X3
= X 2 X3
u16
u17
u18
u19
= X3 X3
= X1 X4
= X2 X4
= X3 X4
u21
u22
u23
u24
u25
= X1 X1 X1 u26 = X1 X2 X3 u31 = X1 X1 X1
= X1 X1 X2 u27 = X2 X2 X3 u32 = X1 X1 X2
= X1 X2 X2
u33 = X1 X2 X2
= X2 X2 X2
= X1 X1 X3
u41 = X1 X1 X1 X1
The following Lemma is the central result of this section, the remainder of the section will
be devoted to providing definitions and remarks that will allow for its proof.
Lemma 1.4. Let M be closed under comparability. Consider the polynomial ring S 0 =
K[Tij : uij ∈ M]. There is a monomial order <τ , such that the set
n
G = Tij Ti0 j 0 − Til Ti0 l0 | (i, j) <lex (i0 , j 0 ), uij ui0 j 0 = uil ui0 l0 , uij ⊀ ui0 j 0 , uil ≺ ui0 l0
o
is a Gröbner basis for I = (G), and the underlined terms are initial terms according to <τ .
It is clear that (G) ⊆ KerΨ, since Ψ(Tij Ti0 j 0 − Til Ti0 l0 ) = uij ui0 j 0 ti ti0 − uil ui0 l0 ti ti0 and
uij ui0 j 0 = uil ui0 l0 for Tij Ti0 j 0 − Til Ti0 l0 ∈ G. In the next section we will prove that (G) = KerΨ.
We need to guarantee the existence of a monomial order, <τ , such that the underlined
terms above are initial according to <τ first. For this purpose we will fix our attention on
the concepts in the following definition.
Definition 1.5. Let F = {f1 , f2 , . . . , fp } be a set of polynomials in S 0 such that there is a
distinguished monomial mi ∈ supp(fi ) for all 1P≤ i ≤ p. To simplify notation we will assume
that the coefficient of mi in fi is 1. Let f = m∈supp(f ) bm m, we say that f ∈ S 0 reduces
to f 0 , in one step, modulo F, denoted f →F f 0 , if there are 1 ≤ j ≤ p and m0 ∈ supp(f )
bm0 m0
such that mj | m0 and f 0 = f −
(fj ). We say that →F is a Noetherian reduction
mj
relation if given any f ∈ S only a finite number of one step reductions modulo F can be
carried out starting at f .
It is clear that if G is a Gröbner basis, then →G is a Noetherian reduction relation.
Sturmfels wrote a partial converse for this statement using coherent markings.
Let F = {f1 , f2 , . . . , fp } be a set of polynomials such that there is a distinguished monomial
mi ∈ supp(fi ). We say that the set F is marked coherently if there is a monomial ordering
<τ such that mi = in<τ (fi ).
Theorem 1.6. A finite set F is marked coherently if and only if →F is a Noetherian reduction relation.
The proof of this statement appears in [Stu95, Chapter 3].
To prove that →G , with G as in Lemma 1.4, is a Noetherian reduction relation, we
need to create a way to encode information for monomials in S 0 . Let m ∈ Mon(S 0 ), then m
can be expressed as:
m=
s
Y
ri
Y
j=1
ri
Y
i=0
where, as usual,
Tiσim (j) with uiσim (j) ∈ M, and 1 ≤ σim (j) ≤ ni
j=1
Tiσim (j) = 1 if ri = 0.
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY STABLE IDEALS
5
i
Tiσim (j) is the ith level of m, and it can be represented as a ri × di We say that rj=1
matrix, Ai , with entries given by (Ai )jk = Yiσim (j)k , with the understanding that if ri = 0
then the matrix is empty. If we consider the set M in Example 1.3,and the monomial
X1 X1 X2
m = T32 T32 T33 , we obtain that A0 , A1 , A2 , A4 are empty and that A3 = X1 X1 X2 ; but
X1 X2 X2
X1 X1 X2
if we had expressed m as T32 T33 T32 , then A3 =
X1 X2 X2 .
X1 X1 X2
This indicates that the representation of the ith -level of m as a matrix is not unique, (it
depends on the way that the product is expressed). To explain our choice of representation,
we need to define the number of inversions of a c × d-matrix A, denoted eA , as
Q
eA =
c
X
d
X
i=1
|{(i0 , j 0 ) : (j, i) <lex (j 0 , i0 ) and Ai0 j 0 > Aij }| .
j=1
In other words, for every entry Aij , let eA,ij be the number of entries in A that are larger
than Aij , and are located either to the right of Aij , or, in the same column (j-th column) as
Aij , but below it; eA is the addition of the eA,ij over all the entries in A.
We say that a c × d-matrix B is inversion minimal, if Bjk ≤ Bjk0 , whenever 1 ≤ k <
k 0 ≤ d, (for all 1 ≤ j ≤ c), and eB ≤ eB 0 for all B 0 obtained by permuting the rows of B.
We will denote the set of all inversion minimal matrices with entries in S, and size c × di ,
for some c ∈ N, by Tdi .
With these stipulations in place, we realize a monomial m in S 0 as a (s + 1)-tuple of
inversion minimal matrices in Tds × Tds−1 × · · · × Td1 × Td0 , in the following manner:
(m)
(m)
(m)
(m)
m −→ (Bs(m) , Bs−1 , . . . , B1 , B0 ), where Bi
is an inversion minimal ri × di -matrix,
representing the ith level of m.
We can now define the number of inversions at the ith level of m, denoted em,i , as
(m)
em,i = eB (m) , (i.e. the number of inversions of an inversion minimal matrix Bi for the ith
i
level of m).
P
The total number of inversions of m will be denoted em := si=0 em,i .
We define the comparability number of m, denoted cm , as
cm =
s
X
ri
X
i=0
di n
X
j=1
k=1
o
(i0 , σim0 (j 0 ), k 0 ) : i0 > i and Yi0 σim0 (j 0 )k0 < Yiσim (j)k .
We can realize what cm means in terms of the entries in any (s + 1)-tuple of matrices
(As , . . . , A1 , A0 ), with the property that Ai is a matrix representation for the ith -level of m, (in
(m)
particular, for a representation of m in terms of inversion minimal matrices (Bs(m) , . . . , B0 )):
0
For every entry (Ai )jk , let cm,Ai ,jk be the number of entries in matrices Ai0 , with i > i, that
are smaller than (Ai )jk ; cm is the addition of the cm,Ai ,jk over all the entries in the (s+1)-tuple
(Ar , . . . , A1 , A0 ).
The pair (cm , em ) will be defined as the level of reduction of m. If cm = em = 0 we say
that m is completely reduced.
6
GABRIEL SOSA
X X1 X1
X1 X 1 X 2 1
If we consider the set M in Example 1.3, then
X2 X4 X1 is a
X2 X 2 X 3
X3 X3 X4
2
representation for the monomial m = T01 T04 T11 T16 T18 T22 T27 in terms of inversion minimal
matrices. The bolded variables make contributions to the comparability number of m,
while the underlined variables make contributions to the total number of inversions of
m. The level of reduction in this case is (25, 4).
h ih i
"
#
Remark 1.7. Let M be a set closed under comparability.
(i) If uij ⊀ ui0 j 0 with i < i0 , and, Tij Ti0 j 0 divides m; then cm > cm0 where m0 = TTijil TTi00 l00 m,
i j
and (uil , ui0 l0 ) = ord(uij , ui0 j 0 ). This implies that (cm0 , em0 ) <lex (cm , em ).
(ii) If A is a matrix such that there are entries Aj1 k1 < Aj2 k2 , with (k1 , j1 ) <lex (k2 , j2 ),
and we consider the matrix A0 such that A0jk = Ajk whenever (j, k) 6= (j1 , k1 ), (j2 , k2 ),
A0j1 k1 = Aj2 k2 and A0j2 k2 = Aj1 k1 , then eA0 < eA .
(iii) If uij ⊀ uij 0 and uij 0 ⊀ uij , and, Tij Tij 0 divides m; then cm = cm0 , where m0 = TTijil TTil00 m,
ij
and (uil , uil0 ) = sort(uij , uij 0 ). Additionally em0 ,i < em,i and em0 ,i0 = em,i0 , for all i0 6= i,
which implies em0 < em , and then logically (cm0 , em0 ) <lex (cm , em ).
(iv) A monomial m ∈ S 0 is completely reduced if and only if uij ≺ ui0 j 0 or ui0 j 0 ≺ uij ,
whenever Tij Ti0 j 0 divides m.
Proof. (i) Since uij ⊀ ui0 j 0 , it must happen that Yij1 = max(uij ) > min(ui0 j 0 ) = Yi0 j 0 di0 .
Let p = max{k : 1 ≤ k ≤ di and Yijk > Yi0 j 0 (di0 −k+1) }. Consider:
u=
di
Y
k=p+1
Yijk
p
Y
k=1
di0 −p
!
Yi0 j 0 (di0 −k+1)
and v =
Y
k=1
Yi0 j 0 k
p
Y
!
Yijk
k=1
Realize that max(u) = max(Yij(p+1) , Yi0 j 0 (di0 −p+1) ), while min(v) = min(Yijp , Yi0 j 0 (di0 −p) ).
The definition of p implies that Yi0 j 0 (di0 −p) ≥ Yij(p+1) and Yijp > Yi0 j 0 (di0 −p+1) . The former
statement coupled with the fact that Yijp ≥ Yij(p+1) and Yi0 j 0 (di0 −p) ≥ Yi0 j 0 (di0 −p+1) , proves
that min(v) ≥ max(u). Hence u = uil and v = ui0 l0 .
Since cm and cm0 are independent of whether their ith -levels are represented by inversion
minimal matrices or not, we will consider representations (Ar , . . . , A1 , A0 ) and (A0r , . . . , A01 , A00 ),
for m and m0 respectively, such that Aî = A0î for î 6= i, i0 , (Ai )ĵ k̂ = (A0i )ĵ k̂ and (Ai0 )ĵ k̂ = (A0i0 )ĵ k̂
whenever ĵ > 1, and, (Ai )1k = Yijk , (Ai0 )1k = Yi0 j 0 k , (A0i )1k = Yilk and (A0i0 )1k = Yi0 l0 k .
Notice that (Ai0 )1k = Yi0 j 0 k ≤ Yi0 l0 k = (A0i0 )1k for all 1 ≤ k ≤ di0 , since (Yi0 l0 1 , Yi0 l0 2 , . . . , Yi0 l0 di0 )
is a rearrangement of (Yi0 j 0 1 , . . . , Yi0 j 0 (di0 −p) , Yij1 , . . . , Yijp ), with entries written in decreasing
order, and Yijk ≥ Yijp > Yi0 j 0 (di0 −p+1) ≥ Yi0 j 0 (di0 −p+k) for 1 ≤ k ≤ p.
Furthermore, for p + 1 ≤ k ≤ di , there exists a π(k) such that (Ai )1k = (A0i )1π(k) ; for
1 ≤ k ≤ di0 − p, there exists a π 0 (k) such that (Ai0 )1k = (A0i0 )1π0 (k) ; and, for 1 ≤ k ≤ p,
there exist γ(k) and γ 0 (k) such that (Ai )1k = (A0i0 )1γ 0 (k) and (Ai0 )1(di0 −k+1) = (A0i )1γ(k) , and
{π(k)} ∩ {γ(k)} = {π 0 (k)} ∩ {γ 0 (k)} = ø .
We will proceed to compare the contributions that the entries in (As , . . . , A1 , A0 ) make to
the comparability number of m against the contributions that the entries in the arrangement
(A0s , . . . , A01 , A00 ) make to the comparability number of m0 .
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY STABLE IDEALS
7
(a) cm0 ,A0 ,ĵ k̂ = cm,A ,ĵ k̂ whenever î < i or î > i0 . Since (Aî )ĵ k̂ = (A0î )ĵ k̂ , and the enî
î
tries in the arrangement (A0s , . . . , A0î+2 , A0î+1 ) are the same entries as in the arrangement
(As , . . . , Aî+2 , Aî+1 ) after a permutation.
(b) cm0 ,A0 ,ĵ k̂ ≤ cm,A ,ĵ k̂ whenever i < î < i0 , or, î = i, i0 and ĵ > 1. Since Aî = A0î , for î 6= i, i0 ;
î
î
(Aî )ĵ k̂ = (A0î )ĵ k̂ , for î = i, i0 and ĵ > 1, and (Ai0 )1k ≤ (A0i0 )1k for all 1 ≤ k ≤ di0 .
(c) cm0 ,A0i0 ,1π0 (k) = cm,Ai0 ,1k , whenever 1 ≤ k ≤ di0 − p, by the same argument as in (a).
(d) cm0 ,A0i ,1π(k) ≤ cm,Ai ,1k , whenever p + 1 ≤ k ≤ di , by the same argument as in (b).
(e) cm0 ,A0i ,1γ 0 (k) + cm0 ,A0i0 ,1γ(k) < cm,Ai ,1k + cm,Ai0 ,1(di0 −k+1) , whenever 1 ≤ k ≤ p. We will break
down the explanation:
- If (A0î )ĵ k̂ < (A0i0 )1γ 0 (k) with î > i0 , then (Aî )ĵ k̂ < (Ai )1k , since A0î = Aî (whenever î 6= i0 , i),
(A0i0 )1γ 0 (k) = (Ai )1k and if î > i0 then î > i. Analogously, if (A0î )ĵ k̂ < (A0i )1γ(k) with î > i0 ,
then (Aî )ĵ k̂ < (Ai0 )1(di0 −k+1) .
This proves that if (A0î )ĵ k̂ contributes to the growth of cm0 ,A0i ,1γ 0 (k) + cm0 ,A0i0 ,1γ(k) , when î > i0 ,
then (Aî )ĵ k̂ contributes to the growth of cm,Ai ,1k + cm,Ai0 ,1(di0 −k+1) .
- If (A0î )ĵ k̂ < (A0i )1γ(k) with i0 > î > i, or î = i0 and ĵ > 1, then (Aî )ĵ k̂ < (Ai )1k . Since
(A0î )ĵ k̂ = (Aî )ĵ k̂ (whenever î 6= i0 , i, or, î = i0 and ĵ > 1), and (A0i )1γ(k) = (Ai0 )1(di0 −k+1) =
Yi0 j 0 (di0 −k+1) < Yijk = (Ai )1k .
This proves that if (A0î )ĵ k̂ contributes to the growth of cm0 ,A0i ,1γ 0 (k) + cm0 ,A0i0 ,1γ(k) , when i0 >
î > i, or î = i0 and ĵ > 1, then (Aî )ĵ k̂ contributes to the growth of cm,Ai ,1k + cm,Ai0 ,1(di0 −k+1) .
- Finally, we have that (A0i0 )1k̂ = Yi0 l0 k̂ ≥ min(ui0 l0 ) ≥ max(uil ) ≥ Yilk0 = (A0i )1k0 , for all
1 ≤ k 0 ≤ di and all 1 ≤ k̂ ≤ di0 ; which implies that (A0î )1k̂ makes no contribution to the
growth cm0 ,A0i ,1τ 0 (k) + cm0 ,A0i0 ,1τ (k) , with î = i0 .
On the other hand, for di0 − p + 1 ≤ k̂ ≤ di0 , we have (Ai0 )1k̂ = Yi0 j 0 k̂ < Yijk = (Ai )1k , for all
1 ≤ k ≤ di , which means that (Aî )1k̂ contributes to the growth of cm,Ai ,1k + cm,Ai0 ,1(di0 −k+1) ,
when î = i0 and di0 − p + 1 ≤ k̂ ≤ di0 .
Putting all the contributions together cm0 ,A0i ,1γ 0 (k) + cm0 ,A0i0 ,1γ(k) < cm,Ai ,1k + cm,Ai0 ,1(di0 −k+1) as
we wanted .
Adding all the inequalities and equalities from (a), (b), (c), (d) and (e) together we obtain
that cm0 < cm .
(ii) We will proceed to compare the contributions that the entries of A make to the number
of inversions of A against the contributions that the entries of A0 make to the number of
inversions of A0 .
(a) eA0 ,ĵ k̂ = eA,ĵ k̂ whenever (k̂, ĵ) <lex (k1 , j1 ) or (k2 , j2 ) <lex (k̂, ĵ). Since after a permutation
the entries Ajk with (k̂, ĵ) <lex (k, j) become the entries A0jk with (k̂, ĵ) <lex (k, j).
(b) eA0 ,ĵ k̂ ≤ eA,ĵ k̂ for (k1 , j1 ) <lex (k̂, ĵ) <lex (k2 , j2 ). Since Ajk ≥ A0jk for (k, j) >lex (k1 , j1 ).
(c) eA0 ,j1 k1 + eA0 ,j2 k2 < eA,j1 k1 + eA,j2 k2 . We will break down the explanation:
8
GABRIEL SOSA
- If A0ĵ k̂ > A0j2 k2 , with (k2 , j2 ) <lex (k̂, ĵ), then Aĵ k̂ = A0ĵ k̂ > A0j2 k2 = Aj1 k1 , and (k1 , j1 ) <lex
(k̂, ĵ). Also if A0ĵ k̂ > A0j1 k1 , with (k2 , j2 ) <lex (k̂, ĵ), then Aĵ k̂ = A0ĵ k̂ > A0j1 k1 = Aj2 k2 .
This proves that if A0ĵ k̂ contributes to the growth of eA0 ,j1 k1 + eA0 ,j2 k2 then Aĵ k̂ contributes
to the growth of eA,j1 k1 + eA,j2 k2 when (k̂, ĵ) >lex (k2 , j2 ).
- If A0ĵ k̂ > A0j1 k1 , with (k1 , j1 ) <lex (k̂, ĵ) <lex (k2 , j2 ) , then Aĵ k̂ = A0ĵ k̂ > A0j1 k1 = Aj2 k2 >
Aj1 k1 .
This proves that if A0ĵ k̂ contributes to the growth of eA0 ,j1 k1 + eA0 ,j2 k2 then Aĵ k̂ contributes
to the growth of eA,j1 k1 + eA,j2 k2 , when (k1 , j1 ) <lex (k̂, ĵ) <lex (k2 , j2 ).
- Finally, notice that A0j2 k2 ≯ A0j1 k1 , while Aj2 k2 > Aj1 k1 . Hence A0j2 k0 does not contribute to
2
the growth of eA0 ,j1 k1 + eA0 ,j2 k2 , but Aj2 k2 contributes to the growth of eA,j1 k1 + eA,j2 k2 .
Putting all the contributions together we obtain eA0 ,j1 k1 + eA0 ,j2 k2 < eA,j1 k1 + eA,j2 k2 as we
wanted.
Adding all the inequalities and equalities described in (a), (b) and (c) together we obtain
eA0 < eA .
(m)
(m)
(iii) Let (Bs(m) , . . . , B1 , B0 ) be a representation of m in terms of inversion minimal ma0
(m0 )
(m0 )
trices, and (Bs(m ) , . . . , B1 , B0 ) be a representation of m0 in terms of inversion minimal
(m)
(m0 )
(m0 )
matrices, then Bi0 = Bi0 for i0 6= i, and the entries of Bi
are the same as the entries
(m)
0
of Bi after a permutation. This proves that cm = cm and that em0 ,i0 = em,i0 for i0 6= i.
To prove that em0 ,i < em.i , (and thus obtain that em0 < em ): Let h and h0 be such that
(m)
(m)
(Bi )hk = Yijk and (Bi )h0 k = Yij 0 k . We will assume without loss of generality that h < h0 .
(m)
Let A0i be the matrix such that (A0i )ĵk = (Bi )ĵk whenever ĵ 6= h, h0 , and (A0i )hk = Yilk ,
(m0 )
(A0i )h0 k = Yil0 k , then it is clear that eB (m0 ) ≤ eA0i by the minimality of Bi
i
to prove that eA0i < eB (m) .
. So we only need
i
Since uij ⊀ uij 0 and uij 0 ⊀ uij , then there are (h1 , k1 ) = min{(h̃, k̃) : h̃ = h or h̃ =
(m)
(m)
0
h and ∃(ĥ, k̂), with (k̃, h̃) <lex (k̂, ĥ), ĥ = h or ĥ = h0 , and (Bi )h̃k̃ < (Bi )ĥk̂ }, and
(m)
(m)
(h01 , k10 ) = min{(h̃, k̃) : h̃ = h or h̃ = h0 , (k1 , h1 ) <lex (k̃, h̃) and (Bi )h1 k1 < (Bi )ĥk̂ }.
(m)
(1)
(1)
(1)
Replace Bi by Bi , a matrix with the property that (Bi )h1 k1 = (Bi )h2 k2 , (Bi )h2 k2 =
(1)
(m)
(Bi )h1 k1 and (Bi )ĵk = (Bi )ĵk , whenever (ĵ, k) 6= (h1 , k1 ), (h2 , k2 ).
(1)
If Bi 6= A0i , we can apply the exchange described above again, and after a finite number
(η)
of exchanges we must obtain that Bi = A0i , and then by Remark 1.7(ii), we have that
eA0i = eB (η) < eB (η−1) < · · · < eB (1) < eB (m) , as we wanted.
i
i
i
i
(iv) (⇒) We will proceed by contradiction. Assume m is completely reduced and that there
are Tij and Ti0 j 0 that divide m such that uij ⊀ ui0 j 0 and ui0 j 0 ⊀ uij . Then, by either Remark
1.7(i) or Remark 1.7(iii), there is m0 such that (cm0 , em0 ) <lex (cm , em ), which implies that
(cm , em ) 6= (0, 0) contradicting the fact that m is completely reduced.
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY STABLE IDEALS
9
( ⇐) If uij ≺ ui0 j 0 or ui0 j 0 ≺ uij whenever Tij Ti0 j 0 divides m, then we can represent m as:
m=
s
Y
ri
Y
i=0
j=1
Tiσim (j) with (?) uiσim (j) ≺ ui0 σim0 (j 0 ) whenever (i, j) <lex (i0 , j 0 )
Then the arrangement (Ar , . . . , A1 , A0 ), with (Ai )jk = Yiσim (j)k is a representation of m
by matrices and because of (?), we get cm = 0 and eAi = 0, for all i. This means that
Ai is inversion minimal, implying (cm , em ) = (0, 0), or equivalently, that m is completely
reduced.
We will use the fact that the set {Ψ(m) : m ∈ Mon(S 0 ), m completely reduced} is linearly
independent over K. This is a direct consequence of the following lemma.
Lemma 1.8. Let m, m0 ∈ Mon(S 0 ), with m, m0 completely reduced. If Ψ(m) = Ψ(m0 ), then
m = m0 .
(m)
(m0 )
0
(m)
(m0 )
Proof. Let (Bs(m) , . . . , B1 , B0 ) and (Bs(m ) , . . . , B1 , B0 ) be representations of m and
m0 by (s + 1)-tuples of inversion minimal matrices, respectively, and let Ψ(m) = Ψ(m0 ) =
(m)
(m0 )
X1α1 · X2α2 · · · · · Xnαn · tβ1 1 · tβ2 2 · · · · · tsβs . If we prove that Bi = Bi
for all 0 ≤ i ≤ s then
0
we can conclude that
m
=
m
.
Q
Qri
Qs
Qri
T
Ψ(T
)
Since m = si=0
,
then
Ψ(m)
=
(m)
(m)
j=1 iσ
j=1
i=0
(j)
iσ
(j)
i
Ψ(m) =
r0
Y
j=1
Y0σ(m) (j) ·
0
s
Y
i
ri
Y
i=1
j=1
uiσ(m) (j) ti =
i
r0
Y
j=1
(m)
Y0σ(m) (j) ·
0
s
Y
ri
Y
i=1
j=1
uiσ(m) (j) ·
i
s
Y
ri
ti
!
.
i=1
(m)
This implies that the size of Bi is βi × di for 1 ≤ i ≤ s, and the size of B0 is β0 × 1,
P
P
(m0 )
is
where β0 = ni=0 αi − ri=0 βi · di . An analogous argument shows that the size of Bi
Pr
Pn
(m0 )
βi × di for 1 ≤ i ≤ s, and the size of B0 is β0 × 1, where β0 = i=0 αi − i=0 βi · di .
(m)
(m0 )
We will proceed by contradiction and assume m 6= m0 . Then (Bi )jk 6= (Bi )jk , for
(m)
(m0 )
some (i, j, k). Let i0 = max{i : ∃(j, k) such that (Bi )jk 6= (Bi )jk } and let (j 0 , k 0 ) be
(m)
(m0 )
(m)
(m0 )
such that (Bi0 )j 0 k0 6= (Bi0 )j 0 k0 and (Bi0 )jk = (Bi0 )jk for all (k, j) <lex (k 0 , j 0 ). Without
(m0 )
(m)
loss of generality we can assume that (Bi0 )j 0 k0 < (Bi0 )j 0 k0 = Xĩ . Since m is completely
(m)
(m)
reduced we know that (Bi )jk ≤ (Bi0 )j 0 k0 < Xĩ for all i < i0 and all (k, j) >lex (k 0 , j 0 )
when i = i0 . This implies that the highest power of Xĩ that divides Ψ(m0 ) is higher than the
highest power of Xĩ that divides Ψ(m), hence Ψ(m0 ) 6= Ψ(m), which is a contradiction.
We are now ready to prove Lemma 1.4.
Proof of Lemma 1.4. To guarantee the existence of a monomial order, <τ , such that the
underlined terms in G are initial according to <τ , it is enough to prove that the reduction
relation modulo G is Noetherian, byPTheorem 1.6.
P
Let f ∈ S 0 , we will define cf = m∈supp(f ) cm and ef = m∈supp(f ) em and the level of
reduction of f as (cf , ef ).
Let f 0 be a one step reduction of f modulo G, then supp(f 0 ) = (supp(f ) − {m0 }) ∪ {m00 },
for some m0 ∈ supp(f ) such that m00 = TTijil TTi00 l00 m0 , with i ≤ i0 , uij ⊀ ui0 j 0 and ui0 j 0 ⊀ uij ,
i j
10
GABRIEL SOSA
uij ui0 j 0 = uil ui0 l0 and uil ≺ ui0 l0 . Then, by either Remark 1.7(i) or Remark 1.7(iii), we get
(cm00 , em00 ) <lex (cm0 , em0 ).
Notice that:
(cf 0 , ef 0 ) ≤lex (cm00 , em00 ) +
X
(cm , em ) <lex
m∈(supp(f )−{m0 })
X
(cm , em ) = (cf , ef )
m∈supp(f )
If the reduction relation modulo G is not Noetherian, then there is an infinite set of
polynomials {f, f (1) , f (2) , . . . } such that (cf , ef ) >lex (cf (1) , ef (1) ) >lex (cf (2) , ef (2) ) >lex . . . ,
but this contradicts one of the corollaries of Dickson’s Lemma, which establishes that there
are not infinitely decreasing sequences of terms in any monomial order, (see [HH11, 2.1.7]).
Hence →G is indeed Noetherian and the existence of <τ is guaranteed. Furthermore, by
Remark 1.7(iv), if fˆ is a remainder of f when dividing by G, then m ∈ supp(fˆ) if and only
if m is completely reduced.
To prove that G is Gröbner basis with respect to <τ , index the elements of G as g1 , g2 , . . . , gκ ,
we will denote the S 0 -polynomial of gi , gj by gi,j .
P
Let gˆi,j be a remainder of gi,j when divided by G. It is known that gi,j = ( κk=1 fk gk )+ gˆi,j .
Applying Ψ on both sides of the equation, along with the fact that G ⊆ KerΨ, we obtain
0 = Ψ(gˆi,j ). And this implies gˆi,j = 0; for if gˆi,j 6= 0, then gˆi,j is a linear combination of
completely reduced monomials, and by Lemma 1.8 their images are linearly independent
over K which contradicts the fact that 0 = Ψ(gˆi,j ).
We have proved that all S 0 -polynomials of G reduce to 0 when divided by G, so Buchberger’s criterion (see [CLO92]) asserts that G is a Gröbner basis according to <τ .
2. Main Result and Examples
We will establish again the setup for our main result. Let 1 = d0 ≤ d1 ≤ · · · ≤ ds , I0 the
maximal ideal of S and Ii homogeneous monomial ideals with minimal generators of fixed
degree di for 1 ≤ i ≤ s. Let the minimal monomial generators of Ii be uij , for 0 ≤ j ≤ ni ,
with the additional condition ui1 >rev ui2 >rev · · · >rev uini . Let M = I1 ⊕ I2 ⊕ · · · ⊕ Is ,
`
M = si=0 {uij : 1 ≤ j ≤ ni } and N = {uij ti : 0 ≤ i ≤ s, 1 ≤ j ≤ ni } with the convention
that t0 = 1. Consider the surjective map Ψ : K[Tij : uij ∈ M] −→ K[N ] ∼
= R(M ), given by
Ψ(Tij ) = uij ti for all uij ∈ M.
Lemma 2.1. If M is closed under comparability KerΨ has a squarefree quadratic Gröbner
basis.
Proof. Consider G as in Lemma 1.4. We already proved that (G) ⊆ KerΨ.
To prove the other inclusion: Let f ∈ S 0 , and assume f does not reduce to 0 with respect
to G, (i.e f ∈
/ (G)). Let f 0 be the remainder of f when divided by G.
Hence Ψ(f ) = Ψ(f 0 ) 6= 0, since f 0 is a linear combination of completely reduced monomials
and, by Lemma 1.8, their images are linearly independent over K,.
We conclude that if f does not reduce to 0 with respect to G, (equivalently f ∈
/ (G)), then
f∈
/ KerΨ. Proving that KerΨ = (G) has a Gröbner basis of quadrics given by G.
To prove that KerΨ = (G) is squarefree, notice that the initial term of Tij Ti0 j 0 − Til Ti0 l0
is a square only if (i, j) = (i0 , j 0 ). Since (uil , ui0 l0 ) = sort(uij , uij ) = (uij , uij ), we conclude
Tij Ti0 j 0 − Til Ti0 l0 = 0.
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY STABLE IDEALS
11
Theorem 2.2. If M is closed under comparability then K[Tij : uij ∈ M]/KerΨ ∼
= K[N ] ∼
=
R(M ) is Koszul and a normal Cohen-Macaulay domain.
Proof. The fact that KerΨ = (G), implies that the defining ideal of R(M ) has a quadratic
Gröbner basis that is square free. Because KerΨ satisfies the G-quadratic condition, we
conclude that R(M ) is Koszul.
Additionally, the fact that the Gröbner basis is squarefree implies that K[N ] is a normal
domain, due to a result by Sturmfels [Stu95]. Another result states that K[N ] is a normal
domain if and only if N is an affine normal semigroup. Finally, a result by Hochster [Hoc72]
proves that if N is an affine normal semigroup then R(M ) ∼
= K[N ] is Cohen-Macaulay.
We will proceed to present a thorough characterization of sets that are closed under
comparability, and correspond to Rees and multi-Rees algebras, to obtain corresponding
Koszul and Cohen-Macaulay normal domains via Theorem 2.2. We will denote by Mi the
set of minimal monomial generators for Ii . Our first step is to prove that if M is closed
under comparability, then Mi has to be a principal strongly stable set. We will start by
recalling a set of equivalences given in a Lemma due to De Negri.
Lemma 2.3 (A description of principal strongly stable sets). Let w1 = X1α1 X2α2 . . . Xnαn and
w2 = X1β1 X2β2 . . . Xnβn be two monomials of the same degree. Then w2 ∈ B(w1 ) if and only if
Pn
Pn
i=k αi for 2 ≤ k ≤ n.
i=k βi ≤
The proof of this lemma can be found in [DeN99, Lemma 1.3].
Next, we will show that the minimal generating set of a principal strongly stable ideal
is closed under sorting. For this we will first describe the sorting of two monomials in a
different way.
Let u1 = X1α1 X2α2 . . . Xnαn and u2 = X1β1 X2β2 . . . Xnβn be two monomials of the same degree.
Let J = {i : αi +βi is odd }, then the cardinality of J is an even number, 2t, and its elements
can be indexed as i1 < i2 < · · · < i2t .
0
0
α0
α0
β0
β0
It is easy to see that sort(u1 , u2 ) = (X1 1 X2 2 . . . Xnαn , X1 1 X2 2 . . . Xnβn ), where αi0 = βi0 =
r+1
r
αi +βi
if i ∈
/ J and αi0 r = αir +βir +(−1)
and βi0r = αir +βir2+(−1) .
2
2
P
P
P
k
Furthermore, nk=i αk0 = nk=i βk0 = nk=i αk +β
if i > i2t , i ≤ i1 , or , i2r < i ≤ i2r+1 .
P2n
P
P
k
On the other hand, if i2r−1 < i ≤ i2r then k=i αk0 = − 21 + nk=i αk +β
and nk=i βk0 =
2
P
1
k
+ nk=i αk +β
.
2
2
Lemma 2.4. Let u1 , u2 ∈ B(w), then u1 , u2 ∈ B(w), with (u1 , u2 ) = sort(u1 , u2 ).
β1 β2
γ1 γ2
βn
γn
Proof. Let u1 = X1α1 X2α2 . .P
. Xnαn , uP
2 = X1 X2 . . . Xn and w = X1 X2 . . . Xn . Since
P
u1 , u2 ∈ B(w), we get that nk=i αk , nk=i βk ≤ nk=i γk for i = 2, . . . , n, because of Lemma
P
P
k
2.3. Thus nk=i αk +β
≤ nk=i γk . P
2
P
This implies immediately that nk=i αk0 ≤ nk=i γk for all i = 2, . . . , n, by the previous
observation.
Pn
(α + βk ) is an odd
number when i2s−1
< i ≤ i2s , so the fact that
Realize
now
that
Pn
Pn k=i k
Pn
Pn
+ βk ) ≤ 2 k=i γk implies that 1 + k=i (αk + βk ) ≤ 2 k=i γk , and this in turn shows
k=i (α
Pk
P
that nk=i βk0 ≤ nk=i γk for all i = 2, . . . , n.
And we get that u1 , u2 ∈ B(w) because of Lemma 2.3.
12
GABRIEL SOSA
The following theorem gives a complete characterization of sets closed under comparability.
Theorem 2.5. Consider a sequence of natural numbers 1 ≤ d1 ≤ d2 ≤ · · · ≤ ds , and
monomial homogeneous ideals Ii generated in fixed degree di .
Let M0 = {X1 , X2 , . . . , Xn } and let Mi be the minimal monomial generating set of Ii .
Index the elements of Mi as ui1 >rev ui2 >rev · · · >rev uini .
`
Then M = ri=0 Mi is closed under comparability, if and only if, (i) Mi = B(uini ) and
(ii) max(uini ) ≤ min(u(i+1)ni+1 ) for all 1 ≤ i ≤ s − 1.
Proof.
(⇒) We will first prove that Mi ⊆ B(uini ) by contradiction.
Let N = {j : uij ∈
/ B(uini )}, if Mi * B(uini ) then N 6= ∅. Let j 0 = max(N ).
Consider uini = X1α1 X2α2 . . . Xnαn and uij 0 = X1β1 X2β2 . . . Xnβn , and let p = max{j : αj 6= βj },
then αp > βp since uini <rev uij 0 . Additionally, since uij 0 ∈
/ B(uini ), the set N 0 = {k :
Pn
Pn
6 ∅, because of Lemma 2.3. Let q = max(N 0 ).
i=k αi <
i=k βi } =
Let (u, v) = sort(uini , uij 0 ), then v ∈ Mi , since M is closed under comparability. So v =
P
P
0
β0
k
uij 00 for some j 00 ≤ ni . Let v = uij 00 = X1 1 . . . Xnβn . Notice that either nk=q βk0 = nk=q αk +β
2
P
Pn
P
P
k
or k=q βk0 = 21 + nk=q αk +β
. Regardless of the case, nk=q βk0 > nk=q αk , which implies that
2
uij 00 ∈
/ B(uini ) by Lemma 2.3.
p
On the other hand, βi0 = βi when i > p, and either βi0 = αp +β2 p +1 or βi0 = αp +β
, because
2
00
0
of the maximality of p. Regardless, v = uij 00 <rev uij 0 , which implies j > j . Contradicting
the maximality of j 0 .
Now we will prove that B(uini ) ⊆ Mi .
γ∗
Q
Q
γ
Let u = pj=1 Xj j ∈ Mi with γp 6= 0, we will define u∗ = pj=1 Xj j with γj∗ = γj for
∗
= γp−1 + 1 and γp∗ = γp − 1. Notice that u <rev u∗ . The fact that M is closed
j < p − 1, γp−1
under comparability implies that u∗ ∈ Mi , since ord(Xp−1 , u) = (Xp , u∗ ). The repeated
application of this fact proves that ui1 = X1di , and it also proves that if u = X1di −k X2k ∈ Mi ,
then B(u) ⊂ Mi .
Let k = max{j ≤ ni : B(uij ) ⊂ Mi }, notice that k ≥ 1, since ui1 = X1di . We will prove
that k = ni by contradiction.
Q
β
If k < ni , let ui(k+1) = pj=1 Xj j with βp 6= 0. Notice p ≥ 3, since B(ui(k+1) ) * Mi .
Q
γ
Additionally, the set N = {v ∈
/ Mi : v ∈ B(ui(k+1) )} =
6 ∅, so we can denote w = pj=1 Xj j =
max<rev N . Remember that ui(k+1) <rev (ui(k+1 )∗ ∈ Mi , which implies that (ui(k+1 )∗ = uil
for some l < k + 1.
P
P
Notice that γp = βp . Otherwise γp ≤ βp − 1 = βp∗ , and then nj=s γj = 0 = nj=s βj∗ when
P
P
P
P
P
s > p, nj=p γj = γp ≤ βp∗ = nj=p βj0 and nj=s γj ≤ nj=s βj = nj=s βj∗ when s < p, which by
Lemma 2.3 implies that w ∈ B(uil ) ⊂ Mi , which is false.
Q
βj
βp−1 +1 βp −1
A similar argument shows that w∗ = ( p−2
Xp
∈ B(uil ) ⊂ Mi . Let
j=1 Xj )Xp−1
Q
β
β
+1
β
−1
p−3
j
p−2
p−1
w0 = ( j=1 Xj )Xp−2 Xp−1 Xpβp , then w0 ∈ B(ui(k+1) ) and w <rev w0 , which implies that
w0 ∈ Mi , by the maximality of w.
Q
βj
βp−2 +1 βp−1 βp −1
So sort(w0 , w∗ ) = (( p−3
Xp−1 Xp , w), and this implies that w ∈ Mi , since
j=1 Xj )Xp−2
M is closed under comparability. But this is a contradiction, and we obtain that k = ni .
ON THE KOSZULNESS OF MULTI-REES ALGEBRAS OF CERTAIN STRONGLY STABLE IDEALS
13
Finally, if max(uini ) > min(u(i+1)ni+1 ), let bi and si be such that max(uini ) = Xbi and
Q
α
min(uini ) = Xsi . Then bi < si+1 and uini = nj=bi Xj j , with αbi 6= 0, while u(i+1)ni+1 =
β
Xj j , with βsi+1P6= 0. Let ord(uini , u(i+1)ni+1 )
j=1 P
that nj=t δj = di = nj=bi αj implies that t ≤ si+1 ,
Qsi+1
δ
δ0
= ( nj=t Xj j , tj=1 Xj j ), then the fact
furthermore since αbi 6= 0 6= βsi+1 and
Q
Qn
δ
δ
bi < si+1 , then δsi+1 > αsi+1 , which means that if j=t Xj j ∈ Mi then uini >rev nj=t Xj j ,
which is a contradiction to the minimality of uini .
Hence, if M is closed under comparability then max(uini ) ≤ min(u(i+1)ni+1 ).
Q
Q
(⇐) If Mi = B(uini ), then M is closed under sorting by Lemma 2.4.
So we only have to prove that M isQclosed under ordering. Let l > k, so dl ≥ dk , and
α
l
Xj j , and αbl 6= 0 6= αsl and u ∈ Mk = B(uknk )
consider v ∈ Ml = B(ulnl ) with ulnl = sj=b
l
Qk
β
with uknk = sj=b
Xj j and βbk 6= 0 6= βsk . Since min(ulnl ) ≥ max(uknk ), it must happen
k
that sl ≤ bk .
α0
β0
Ql
Qk
We can express v and u in the following way v = sj=1
Xj j and u = sj=1
Xj j . We want
to prove that û ∈ Mk = B(uknk ) and that v̂ ∈ Ml = B(ulnl ), with (û, v̂) = ord(u, v) =
δ
δ0
l
k
αj0 = dl = tj=1 δj0 and that δj0 = αj0 + βj0
( sj=t
Xj j , tj=1 Xj j ). Notice that t ≤ sl , since sj=1
if j < t, δt0 + δt = αt0 + βt0 , δj = αj0 + βj0 if t < j ≤ sl and δj = βj0 if sl < j ≤ sk .
Applying
Lemma
2.3 after the
following observations proves that û ∈ Mk = B(uknk ):
Pk
Pk
Pk
δj = dk = sj=b
βj if t ≤ r ≤ sl . Remember sl ≤ bk .
δj ≤ sj=t
- sj=r
k
P
Psk
Psk
0
- j=r δj = j=s βj ≤ j=max(s,bk ) βj if sl < r ≤ sk .
Pk
Pk
δj =
βj0 = dk = sj=t
To prove that v̂ ∈ Ml = B(ulnl ), we need the following fact sj=1
P
P
P
Psl
sl
sk
t−1 0
0
0
0
0
0
and
then
apply
Lemma
2.3
to
β
−
=
α
,
or
equivalently
δ
−
δ
+
β
α
j=t j
j=t j
j=t j
j=1 j
t
t
the P
following observations:
Pl
P
αj if r ≤ bl . Remember t ≤ sl .
- tj=r δj0 ≤ tj=1 δj0 = dl = sj=b
l
Psl
Psl
Pl
P
Pt−1 0
Pt
0
0
0
0
0
αj0 − r−1
- j=r δj = j=r (αj + βj ) + δt = sj=r
j=r αj ≤
j=s αj if bl < r.
j=1 βj ≤
Q
Q
P
P
We can now present examples:
Example 2.6 (Rees Algebras of principal strongly stable Ideals). Let s = 1, consider I =
(B(u)), the principal strongly stable ideal generated by the monomial u. Then Theorems 2.2
and 2.5 guarantee that the Rees Algebra R(I) is a Koszul, Cohen-Macaulay normal domain.
Example 2.7 (Multi-Rees Algebras of particular principal strongly stable ideals). Consider
Ii = (B(ui )), with the constraints max(ui ) ≤ min(ui+1 ) and deg(ui+1 ) ≥ deg(ui ).
Then Theorems
2.2 and 2.5 guarantee that the multi-Rees algebra of the module M =
L L
L
I1 I2 · · · Is is a Koszul, Cohen-Macaulay normal domain. In particular, by letting
ui = Xndi , we obtain that the multi-Rees algebra generated by the direct sum of powers of
the maximal ideal of S is a Koszul Cohen-Macaulay normal domain, a fact proved by Lin
and Polini in [LP13].
We conclude with the following observation that produces examples of Koszul CohenMacaulay normal domains for certain semigroup rings, and with the appropriate notation,
for the special fiber ring of R(M ), F(M ) = K[N 0 ] = R(M )/(X1 , . . . , Xn )R(M ), with N 0
defined as in page 3.
14
GABRIEL SOSA
Let 1 ≤ d1 ≤ d2 ≤ · · · ≤ ds < m and let S ∗ = K[X1 , . . . , Xn , Xn+1 , . . . , Xn+s ]. For each
1 ≤ i ≤ s let M∗i ⊆ Mon(Sd∗i ), with |M∗i | = ni , and such that if u ∈ M∗i then Xj - u for all
m−di
j > n. Index the elements of M∗i`as ui1 >rev ui2 >rev · · · >rev uini and let N ∗ = {uij Xn+i
:
s
∗
∗
1 ≤ i ≤ s, 1 ≤ j ≤ ni }. If M = i=1 Mi is closed under comparability and we consider the
m−di
surjective map Ψ∗ : K[Tij : uij ∈ M∗ ] → K[N ∗ ] given by Ψ∗ (Tij ) = uij Xn+i
then Lemma
1.4, Remark 1.7, Lemma 1.8 and Lemma 2.1 still hold and their proofs are similar to ones
presented under the appropriate modifications. An equivalent to Theorem 2.2 can be stated
for this case as
K[N ∗ ] ∼
= K[Tij : uij ∈ M∗ ]/KerΨ∗ is a Koszul, normal Cohen-Macaulay domain.
Additionally we get the following modification for Theorem 2.5:
(1) M∗ is closed under comparability if M∗i = B(uini ) and max(uini ) ≤ min(u(i+1)ni+1 )
(2) If M∗ is closed under comparability then M∗i ⊆ B(uini ) and max(uini ) ≤ min(u(i+1)ni+1 )
And its proof is analogous to that of the applicable parts of Theorem 2.5.
The fact that the condition M∗i = B(uini ) and max(uini ) ≤ min(u(i+1)ni+1 ) is sufficient
but not necessary is evident by considering the example with d1 = 2, d2 = 3, m = 4, M∗1 =
{X32 , X3 X4 , X3 X5 , X4 X5 }, M∗2 = {X13 , X12 X3 }, which makes M∗ to be closed under comparability, while M∗1 6= B(X4 X5 ) and M∗2 6= B(X1 X32 ), yet they are both generating sets of
Veronese type, and K[X13 X7 , X12 X3 X7 , X32 X62 , X3 X4 X62 , X3 X5 X62 , X4 X5 X62 ] a Koszul, CohenMacaulay, normal domain.
References
[BS87a] D. Bayer and M. Stillman, A criterion for detecting m-regularity, Invent. Math. 87 (1987), 1–11. ↑2
[BS87b]
, A theorem on refining division orders by the reverse lexicographical order, Duke Math. J. 55 (1987), 321–328.
↑2
[Cav09] G. Caviglia, The pinched Veronese is Koszul, J. Algebraic Combin. 4 (2009), 539–548. ↑1
[CDR13] A. Conca, E. DeNegri, and M. E. Rossi, Koszul algebras and regularity, Commutative Algebra: Expository Papers
Dedicated to David Eisenbud on the Occasion of his 65th Birthday. Springer (2013), 285–315. ↑1
[CLO92] D. Cox, J. Little, and D. O’Shea, Ideals, varieties and algorithms, Springer, New York, NY, 1992. ↑10
[Con04] A. Conca, Koszul homology and extremal properties of Gin and Lex, Trans. Amer. Math. Soc. 356 (2004), 2945–2961.
↑2
[DeN99] E. DeNegri, Toric rings generated by special stable sets of monomials, Mathematische Nachrichten 203 (1999), 31–45.
↑1, 11
[EHM11] V. Ene, J. Herzog, and F. Mohammadi, Monomial ideals and toric rings of Hibi type arising from a finite poset, Eur.
J. Comb. 32 (2011), 404–421. ↑2
[EK90] S. Eliahou and M. Kervaire, Minimal resolutions of some monomial ideals, J. Algebra 129 (1990), 1–25. ↑2
[FMS11] C. A. Francisco, J. Mermin, and J. Schweig, Borel generators, J. Algebra 332 (2011), 522–542. ↑2
[Frö99] R. Fröberg, Koszul algebras, Advances in Commutative Ring Theory. Proceedings of the 3rd International Conference,
Fez, Lect Notes Pure Appl. Math. Marcel Dekker. New York 205 (1999), 337–350. ↑1
[Gal74] A. Galligo, Ápropos du théorem de préparation de Weierstrass. In:Fonctions de plusiers variables complexes., Lect.
Notes. Math 409 (1974), Springer. ↑2
[HH11] J. Herzong and T. Hibi, Monomial ideals, Springer, Graduate Texts in Mathematics, 2011. ↑10
[HO99] T. Hibi and H. Ohsugi, Toric ideals generated by quadratic binomials, J. Algebra 218 (1999), 509–527. ↑1
[Hoc72] M. Hochster, Rings of invariants of tori, Cohen-Macaulay rings generated by monomials, and polytopes, Ann. of
Mathematics 96 (1972), 228–235. ↑1, 11
[LP13] K-N Lin and C. Pollini, Rees algebras of truncations of complete intersections, 2013. arXiv:1308.5285, To appear in
Journal of Commutative Algebra. ↑2, 13
[Pri70] S. B. Priddy, Koszul resolutions, Trans. Amer. Math. Soc. 152 (1970), 39–60. ↑1
[Stu95] B. Sturmfels, Gröbner bases and convex polytopes, American Mathematical Society, Providence, RI, 1995. ↑1, 4, 11
Department of Mathematics, Purdue University 150 N. University Street, West Lafayette, IN 47907-2067
E-mail address: [email protected]
| 0math.AC
|
arXiv:1711.00137v2 [cs.AI] 27 Feb 2018
pomegranate: fast and flexible probabilistic modeling
in python
Jacob Schreiber
Paul G. Allen School of Computer Science
University of Washington
Seattle, WA 98195
[email protected]
Abstract
We present pomegranate, an open source machine learning package for probabilistic modeling in Python. Probabilistic modeling encompasses a wide range of
methods that explicitly describe uncertainty using probability distributions. Three
widely used probabilistic models implemented in pomegranate are general mixture models, hidden Markov models, and Bayesian networks. A primary focus
of pomegranate is to abstract away the complexities of training models from their
definition. This allows users to focus on specifying the correct model for their
application instead of being limited by their understanding of the underlying algorithms. An aspect of this focus involves the collection of additive sufficient
statistics from data sets as a strategy for training models. This approach trivially
enables many useful learning strategies, such as out-of-core learning, minibatch
learning, and semi-supervised learning, without requiring the user to consider
how to partition data or modify the algorithms to handle these tasks themselves.
pomegranate is written in Cython to speed up calculations and releases the global
interpreter lock to allow for built-in multithreaded parallelism, making it competitive with—or outperform—other implementations of similar algorithms. This paper presents an overview of the design choices in pomegranate, and how they have
enabled complex features to be supported by simple code. The code is available
at https://github.com/jmschrei/pomegranate
Keywords: probabilistic modeling, Python, Cython, machine learning, big data
1 Introduction
The Python ecosystem is becoming increasingly popular for the processing and analysis of data.
This popularity is in part due to easy-to-use libraries such as numpy [van der Walt et al., 2011], scipy
[Jones et al., 2001], and matplotlib [Hunter, 2007] that aim to provide fast general purpose functionality. However, equally important are the libraries that are built on top of these to provide higher level
functionality, such as pandas [McKinney, 2010] for data analysis, scikit-image [van der Walt et al.,
2014] for computer vision, Theano [Theano Development Team, 2016] for efficient evaluation of
mathematical expressions, gensim [Řehůřek and Sojka, 2010] for topic modeling in natural language processing, and countless others. Naturally, many machine learning packages have also been
developed for Python, including those that implement classic machine learning algorithms, such as
scikit-learn [Pedregosa et al., 2011], mlpy [Albanese et al., 2012], shogun [Sonnenburg et al., 2017],
and xgboost [Chen and Guestrin, 2016].
pomegranate fills a gap in the Python ecosystem that encompasses building probabilistic machine
learning models that utilize maximum likelihood estimates for parameter updates. There are several
packages that implement certain probabilistic models in this style individually, such as hmmlearn for
hidden Markov models, libpgm for Bayesian networks, and scikit-learn for Gaussian mixture models
and naive Bayes models. However, pomegranate implements a wider range of probabilistic models
and does so in a more modular fashion than these other packages, having two main effects. The first
is that the addition of a new probability distribution in pomegranate allows for all models to be built
using that distribution immediately. The second is that improvements to one aspect of pomegranate
immediately propagate to all models that would use that aspect. For example, when GPU support
was added to multivariate Gaussian distributions, this immediately meant that all models with multivariate Gaussian emissions could be GPU accelerated without any additional code. pomegranate
currently includes a library of basic probability distributions, naive Bayes classifiers, Bayes classifiers, general mixture models, hidden Markov models, Bayesian networks, Markov chains, as well
as implementations of factor graphs and k-means++/|| that can be used individually but primarily
serve as helpers to the primary models.
There are several already existing Python libraries that implement Bayesian methods for probabilistic modeling. These include, but are not limited to, PyMC3 [Salvatier et al., 2016], PyStan [Stan Development Team, 2016], Edward [Tran et al., 2016], pyro [Inc., 2017], and emcee
[Foreman-Mackey et al., 2013]. Bayesian approaches typically represent each model parameter as
its own probability distribution, inherently capturing the uncertainty in that parameter, whereas maximum likelihood approaches typically represent each model parameter as a single value. An example
of this distinction is that a mixture model can either be represented as a set of probability distributions and a vector of prior probabilities, or as a set of probability distributions that themselves have
probability distributions over their respective parameters (such as the mean and standard deviation,
should these distributions be normal distributions) and as a dirichlet distribution representing the
prior probabilities. The first representation typically specifies models that are faster to both train and
perform inference with, while the second is illustrative of the type of models one could build with
packages that implement Bayesian methods, such as PyMC3. Both representations have strengths
and weaknesses, but pomegranate implements models falling solely in the first representation.
pomegranate was designed to be easy to use while not sacrificing on computational efficiency. Models can either be specified by writing out each of the components individually if known beforehand,
or learned directly from data if not. Key features, such as out-of-core learning and parallelization,
can be toggled for each model independently of the definition or method calls, typically by simply
passing in an optional parameter. The core computational bottlenecks are written in Cython and
release the global interpreter lock (GIL), enabling multi-threaded parallelism that typically Python
modules cannot take advantage of. Lastly, linear algebra operations such as matrix-matrix multiplications are implemented using BLAS with the ability to toggle a GPU if present.
All comparisons were run on a computational server with 24 Intel Xeon CPU E5-2650 cores with
a clock speed of 2.2 GHz, a Tesla K40c GPU, and 256 GB of RAM running CentOS 6.9. The
software used was pomegranate v0.8.1 and scikit-learn v0.19.0. pomegranate can be installed using pip install pomegranate or conda install pomegranate on all platforms. Pre-built
wheels are available for Windows builds, removing the sometimes difficult requirement of a working
compiler.
2 The API
pomegranate provides a simple and consistent API for all implemented models that mirrors the
scikit-learn API as closely as possible. The most important methods are fit , from_samples ,
predict and probability . The fit method will use the given data and optional weights to
update the parameters of an already initialized model, using either maximum-likelihood estimates
(MLE) or expectation-maximization (EM) as appropriate. In contrast, the from_samples method
will create a model directly from data in a manner similar to scikit-learn’s fit method. For simple
models like single distributions this corresponds only to MLE on the input data, but for most other
models this corresponds to an initialization step plus a call to fit . This initialization can range
from using k-means for mixture models to structure learning for Bayesian networks. The predict
method returns the posterior estimate argmaxM P (M |D), identifying the most likely component
of the model for each sample. The probability method returns the likelihood of the data given
2
the model P (D|M ). The other methods include predict_proba which returns the probability
of each component for each sample P (M |D), predict_log_proba which returns the log of
the previous value, and summarize and from_summaries that jointly implement the learning
strategies detailed below.
3 Key Features
pomegranate supports many learning strategies that can be employed during training, including outof-core learning for massive datasets, semi-supervised learning for datasets with a mixture of labeled
and unlabeled data, and minibatch learning. In addition, one can employ multithreaded parallelism
or a GPU for data-parallel speedups. These features are made possible by separating out the collection of sufficient statistics from a data set (using the summarize method) from the actual parameter
update step (using the from_summaries method).
Sufficient statistics are the smallest set of numbers needed to calculate some statistic on a dataset.
As an example, fitting a normal distribution to data involves the calculation of the mean and the
variance. The sufficient
for the mean and the variance
are the sumof the weights of the
n statistics
n
P
P
points seen so far
wi Xi , and the sum of the
wi , the sum of the weighted samples
i=1
i=1
n
P
weighted samples squared
wi Xi2 . The mean and variance can then be directly calculated
i=1
from these three numbers using the following two equations:
µ=
n
P
wi Xi
i=1
n
P
i=1
σ2 =
wi
n
P
wi Xi2
i=1
n
P
i=1
wi
P
n
wi Xi
i=1
−
n
P
i=1
wi
2
(1)
Out-of-core Learning: The additive nature of the sufficient statistics means that if one were to
summarize two batches of data successively and then add the sufficient statistics together, they would
get the same sufficient statistics as if they were calculated from the full data set. This presents an
intuitive way to handle data sets that are too large to fit in memory, by chunking the dataset into
batches that do fit in memory and summarizing them successively, adding the calculated sufficient
statistics together afterwards. This can be done by passing in a batch_size parameter to your
training method, for example model.fit(X, batch_size=10000) would train a pre-initialized
model on more data than can fit in memory by successively summarizing batches of size 10,000
until the full data set has been seen. The summarize and from_summaries methods can also be
used independently to implement custom out-of-core strategies.
Minibatch Learning: A natural extension of the out-of-core strategy is minibatch learning, where
a parameter update is done after one or a few batches, instead of the full data set. This is in contrast
to batch methods that calculate an update using the entire dataset, and stochastic methods that typically update using only a single sample. Minibatching can be specified by passing values to both
batch_size and batches_per_epoch parameters when using fit or from_summaries ,
where the batches_per_epoch is the number of batches to consider before making an update.
Semi-supervised Learning: Semi-supervised learning is the task of fitting a model to a mixture
of both labeled and unlabeled data. Typically this arises in situations where labeled data is sparse,
but unlabeled data is plentiful, and one would like to make use of both to learn an informed model.
pomegranate supports semi-supervised learning for HiddenMarkovModel , BayesClassifier ,
and NaiveBayes models as a combination of EM and MLE. Models are initialized using MLE on
the labeled data. Next, a version of EM is used that combines the sufficient statistics calculated from
the labeled data using MLE with the sufficient statistics calculated from the unlabeled data using
3
EM at each iteration until convergence. This is automatically toggled whenever -1 is present in the
label set, following scikit-learn conventions.
This EM-based approach compares favorably to scikit-learn. To demonstrate, we generate a dataset
of 100k samples in 10 dimensions from 2 overlapping Gaussian ellipses with means of 0 and 1
respectively and standard deviations of 2. It took pomegranate ∼0.04s to learn a Gaussian naive
Bayes model with 10 iterations of EM, ∼0.2s to learn a multivariate Gaussian Bayes classifier with
a full covariance matrix with 10 iterations of EM, whereas the scikit-learn label propagation model
with a RBF kernel did not converge after ∼220s and 1000 iterations, and took ∼2s with a knn kernel
with 7 neighbors. Both pomegranate models achieved validation accuracies over 0.75, whereas the
scikit-learn models did no better than chance.
Parallelism: Another benefit of the use of additive sufficient statistics is that it presents a clear
data-parallel way to parallelize model fitting. Simply, one would divide the data into several batches
and calculate the sufficient statistics for each batch locally. These sufficient statistics can then be
added together back on the main job and all parameters updated accordingly. This is implemented by
dividing the data into batches and running summarize on each of them using separate threads and
then running from_summaries after all threads finish. Typically, the global interpreter lock (GIL)
in Python prevents multiple threads from running in parallel in the same python process. However,
since the computationally intensive aspects are written in Cython the GIL can be released, allowing
for multiple threads to run at once. On a synthetic data set with 3M samples with 1K dimensions
it takes ∼65 seconds to train a Gaussian naive Bayes classifier using pomegranate with 1 thread,
but only ∼17 seconds with 8 threads. For comparison, it takes ∼53 seconds to train a Gaussian
naive Bayes classifier using scikit-learn. On another synthetic data set with 2M samples and 150
dimensions it takes pomegranate ∼470s to learn a Gaussian mixture model with a full covariance
matrix with 1 thread, ∼135s with 4 threads, ∼57s with 16 threads, and ∼200s using a GPU. Lastly,
we compared the speed at which pomegranate and hmmlearn could train a 10 state dense Gaussian
hidden Markov model with diagonal covariance matrices. On a synthetic data set of 100 sequences,
each containing 1,000 10 dimensional observations, it took hmmlearn ∼25s to run five iterations of
Baum-Welch training, while it only took pomegranate ∼13s with 1 thread, ∼4s with 4 threads, and
∼2s with 16 threads.
4 Discussion
pomegranate aims to fill a niche in the Python ecosystem that exists between classic machine learning methods and Bayesian methods by serving as an implementation of flexible probabilistic models.
The design choices that were made early on while building pomegranate allowed for a great number
of useful features to be added later on without significant effort.
A clear area of improvement in the future is the handling of missing values, because many probabilistic models can intuitively modify the EM algorithm to infer these missing values. For example,
when trying to learn a Bayesian network over a dataset with missing values, one can identify the best
structure over the incomplete dataset, infer the missing values, and relearn the structure, iterating
until convergence. Given the prevalence of missing data in the real world, extending pomegranate
to handle missing data efficiently is a priority.
Acknowledgments
We would like to first acknowledge all of the contributors and users of pomegranate, whom without
this project would not be possible. We would also like to acknowledge Adam Novak, who wrote
the first iteration of the hidden Markov model code. Lastly, we would also like to acknowledge Dr.
William Noble for suggestions and guidance during development. This work was partially supported
by NSF IGERT grant DGE-1258485.
References
Davide Albanese, Roberto Visintainer, Stefano Merler, Samantha Riccadonna, Giuseppe Jurman,
and Cesare Furlanello. mlpy: Machine learning python, 2012.
4
Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. CoRR, abs/1603.02754,
2016. URL http://arxiv.org/abs/1603.02754.
D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman. emcee: The mcmc hammer. PASP,
125:306–312, 2013. doi: 10.1086/670067.
John D. Hunter. Matplotlib: A 2d graphics environment. Computing in Science and Engineering,
pages 90–95, 2007.
Uber Technologies Inc. pyro. https://github.com/uber/pyro, 2017.
Eric Jones, Travis Oliphant, Pearu Peterson, et al. SciPy: Open source scientific tools for Python,
2001. URL http://www.scipy.org/.
Wes McKinney. Data structures for statistical computing in python. Proceedings of the 9th Python
in Science Conference, pages 51–56, 2010.
Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Olivier Grisel
Bertrand Thirion, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard
Duchesna. Scikit-learn: Machine learning in python. Journal of Machine Learning Research,
pages 2825–2830, 2011.
Radim Řehůřek and Petr Sojka. Software Framework for Topic Modelling with Large Corpora. In
Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks, pages 45–50,
Valletta, Malta, May 2010. ELRA. http://is.muni.cz/publication/884893/en.
John Salvatier, Thomas V. Wiecki, and Christopher Fonnesbeck. Probabilistic programming in
python using pymc3. PeerJ Computer Science, 2:e55, April 2016. ISSN 2376-5992. doi: 10.
7717/peerj-cs.55. URL https://doi.org/10.7717/peerj-cs.55.
Soeren Sonnenburg, Heiko Strathmann, Sergey Lisitsyn, Viktor Gal, Fernando J. Iglesias García,
Wu Lin, Soumyajit De, Chiyuan Zhang, frx, tklein23, Evgeniy Andreev, JonasBehr, sploving, Parijat Mazumdar, Christian Widmer, Pan Deng / Zora, Saurabh Mahindre, Abhijeet
Kislay, Kevin Hughes, Roman Votyakov, khalednasr, Sanuj Sharma, Alesis Novik, Abinash
Panda, Evangelos Anagnostopoulos, Liang Pang, Alex Binder, serialhex, Esben Sørig, and
Björn Esser. shogun-toolbox/shogun: Shogun 6.0.0 - Baba Nobuharu, April 2017. URL
https://doi.org/10.5281/zenodo.556748.
Stan Development Team. Pystan: the python interface to stan, 2016.
Theano Development Team.
Theano:
A Python framework for fast computation
of mathematical expressions.
arXiv e-prints, abs/1605.02688, May 2016.
URL
http://arxiv.org/abs/1605.02688.
Dustin Tran, Alp Kucukelbir, Adji B. Dieng, Maja Rudolph, Dawen Liang, and David M.
Blei. Edward: A library for probabilistic modeling, inference, and criticism. arXiv preprint
arXiv:1610.09787, 2016.
Stéfan van der Walt, Chris Colbert, and Gaël Varoquaux. The numpy array: A structure for efficient
numerical computation. Journal of Machine Learning Research, pages 22–30, 2011.
Stéfan van der Walt, Johannes L. Schönberger, Juan Nunez-Iglesias, François Boulogne, Joshua D.
Warner, Neil Yager, Emmanuelle Gouillart, Tony Yu, and the scikit-image contributors. scikitimage: Image processing in python. PeerJ, 2014.
5
| 2cs.AI
|
Liveness-Driven Random Program Generation
Gergö Barany
arXiv:1709.04421v1 [cs.PL] 13 Sep 2017
Inria Paris, France
[email protected]
Abstract. Randomly generated programs are popular for testing compilers and program analysis tools, with hundreds of bugs in real-world C
compilers found by random testing. However, existing random program
generators may generate large amounts of dead code (computations whose
result is never used). This leaves relatively little code to exercise a target
compiler’s more complex optimizations.
To address this shortcoming, we introduce liveness-driven random program generation. In this approach the random program is constructed
bottom-up, guided by a simultaneous structural data-flow analysis to
ensure that the generator never generates dead code.
The algorithm is implemented as a plugin for the Frama-C framework.
We evaluate it in comparison to Csmith, the standard random C program
generator. Our tool generates programs that compile to more machine
code with a more complex instruction mix.
Keywords: code generation, random testing, data-flow analysis, program optimization
1
Motivation
Optimizing compilers for real-world programming languages are complex pieces
of software. Compiler bugs may manifest in several ways: As compiler crashes,
missed optimizations, or as silent miscompilations. The third category is especially serious as it may introduce bugs in correct programs. Such wrong-code
bugs may invalidate all correctness guarantees provided by source-level verification of safety-critical (and other) software systems.
Two main avenues of work address these problems: compiler verification and
compiler testing. Compiler verification has seen much research [3], with CompCert as a prominent example [8]. However, such compilers have not entered
the mainstream yet: Compiler verification is difficult and time-consuming, and
verified compilers therefore perform fewer optimizations and target fewer CPU
architectures than others.
A different approach is to test compilers in a way that instills confidence.
Standard compiler test suites exist for exercising C compilers, in particular for
testing their conformance to various details of the standard [11, 12]. In addition,
randomized differential testing has gained prominence in recent years. Compiling
many random programs with various compilers and comparing the behaviors of
1
the generated binaries can uncover input programs that cause compiler crashes
or miscompilations. The best-known example of this approach is the work of
Yang et al. on Csmith [16], a generator of random C programs. Csmith generates programs that are fully self-contained (including all their inputs in initialized
global variables) and conform to the C standard by construction. If two compilers produce code that behaves differently for a Csmith-generated program, one
of the compilers must contain a miscompilation bug. Testing of C compilers with
Csmith has uncovered hundreds of bugs in total, including crashes and miscompilations in every compiler under test. This included bugs in (unverified parts
of) the CompCert verified C compiler [16].
This article describes a random generator of C programs developed for a
project on finding missed optimizations in C compilers. Inspired by the successes
of Csmith, in this project we generate random C programs, compile them using
various compilers, then use custom tooling to search for possible optimizations
in the resulting binaries. (The details will be described in a separate paper.)
Starting with Csmith as our program generator, we found early on that it was
not an optimal fit for our intended use case: Despite generating realistic-looking
programs with complex arithmetic expressions, accesses to global and local variables including through pointers, structures, and arrays, as well as nested loops
and branches, it produces large amounts of dead code whose results are never
used. (See our experiments in Section 4.) Dead code elimination, a standard part
of every optimizing compiler, can thus remove large parts of the code generated
by Csmith, leaving very little relevant code for the remaining more interesting
optimizations. Csmith often generate functions of several hundred lines of code
that are compiled to ten machine instructions, representing only a small fraction
of the computations present on the source code level.
In this paper we address this problem with our new liveness-driven random
generator ldrgen. Our tool uses liveness analysis during program generation to
avoid generating dead code. In the following sections we describe liveness-driven
program generation; the implementation of our tool as a Frama-C plugin; and its
experimental evaluation, showing that ldrgen generates programs that compile
to a larger amount of code and a more complex instruction mix than programs
generated by Csmith.
2
Fully live programs
In this section we briefly recall the basics of liveness analysis and then present
our novel formulation as a set of structural inference rules.
2.1
Principles of liveness analysis
A variable is called live at a program point if the value it holds at that point
may be read in the future, without an intervening redefinition; otherwise, it is
called dead. For example, in a code snippet like x = a + b; x = 0; return x;,
the variable x is dead after the first assignment but live after the second one
(because it is used in the return statement). We can extend the notion of liveness
from variables to the assignment statements defining them: An assignment v = e
is live iff the variable v is live just after this statement. The first assignment
to x above is dead, the second one is live. (Unfortunately, some authors use the
term dead code to refer to unreachable code, as in if (false) x = y. These
concepts are not the same; our use of the terms live and dead does not refer to
reachability.)
Dead assignments without other side effects are useless and can be removed
from the program. Even mildly optimizing compilers implement a dead code
elimination pass that would completely remove the addition from the first program fragment above. Our goal is to generate only live code, i. e., only code that
does not contain any such opportunities for dead code elimination.
Liveness analysis is one of the classical data-flow analyses [10]. It is traditionally performed as backward fixed-point iteration over a program’s control-flow
graph. A statement S in the control-flow graph has a live-in set S • and a live-out
set S ◦ which capture the sets of live variables before and after execution of S.
Every statement S also has a transfer function fS relating these sets. The important case are assignments of the form v = e for some variable v and expression e.
We have the following transfer function for such assignments, where F V (e) denotes the set of all variables occurring in e:
S • = fv=e (S ◦ ) = (S ◦ \ {v}) ∪ F V (e)
The function is said to kill the set {v} and generate the set F V (e). Unless v
occurs in e, it will be dead before this assignment, since the value it holds at
that point will be overwritten by the assignment.
Evaluation of expressions without overwriting a variable, as in a branch condition, only generates new live variables. Individual transfer functions are connected in a system of equations such that each statement’s
live-out set is the
S
union of the live-in sets of all of its successors: S ◦ = Si ∈succ(S) Si•
If the program contains loops, their live-out sets will depend recursively on
their own live-in sets. The desired solution is the least fixed point of the system
of equations, which can be found efficiently by propagating data-flow information backward, i. e., starting at the program’s end and proceeding against the
direction of control flow.
2.2
Recognizing fully live programs
We will call a program fully live if all of the assignment statements it contains are
live. This section develops an inference system characterizing fully live programs.
Figure 1 shows the abstract syntax of our programming language of interest,
a subset of C function bodies without declarations. The language contains variables, constants, and all side-effect-free arithmetic and bitwise operators of C.
Statements are assignments (the only source of side effects), return statements,
if statements and general while loops. In contrast to C’s concrete syntax, we
view the semicolon ; as a statement separator, not a terminator. For now, the
language does not include for loops, nor any structures, arrays, or pointers. All
variables are considered local.
v ::=
n ::=
e ::=
S ::=
|
|
|
|
|
a | b | ... | x | y | ...
0 | 1 | ...
v | n | e + e | e - e | . . . | e << e | . . . | -e | !e | . . .
v = e
return v
{}
S; S
if (e) S else S
while (e) S
variables
constant literals
expressions
assignment statement
return statement
empty block
sequencing
conditional branch
loop
Fig. 1. Abstract syntax of a C-like programming language
Figure 2 shows a system of inference rules that characterize fully live programs. In these rules we use a notation similar to Hoare triples. A liveness triple
hL• i S hL◦ i
means that the variables in the set L• are live immediately before the statement S
(live in), and the variables in L◦ are live immediately after S (live out ). A
program S is fully live iff there is a set of variables L• such that the liveness
triple hL• i S h∅i is derivable in the system.
Intuitively, the system of inference rules encodes two things. First, the rules
are an alternative presentation of the transfer functions of liveness analysis. A
triple hL• i S hL◦ i that appears in a valid derivation corresponds to a dataflow equation L• = fS (L◦ ) where fS is the transfer function for the statement S. For example, the transfer function fv=e for an assignment v = e is
just fv=e (L◦ ) = (L◦ \ {v}) ∪ F V (e), as in the side condition of the Assign rule.
Similarly, the Sequence rule encodes the composition of transfer functions, and
the If rule encodes the split and join of data-flow information along different
program paths.
Second, the other side conditions add a system of constraints to ensure full
liveness: Any assignment statement appearing in a fully live program S (i. e., for
which a derivation of hL• i S h∅i for some L• exists) is itself live. This follows
directly from the Assign rule’s side condition v ∈ L◦ . For example, a triple of
the form
hL• i x = a; x = b hL◦ i
can never be derived in the system because the first of the two assignments is
dead. The Sequence rule says that to derive this triple, there must be some
intermediate set L′ of variables such that x ∈ L′ due to Assign on x = a while
at the same time L′ = (L \ {x}) ∪ {b} due to Assign on x = b. This is a
contradiction, so the derivation attempt must fail.
Return
h{v}i return v h∅i
Assign
Sequence
If
v ∈ L◦
L 6= ∅
hLi {} hLi
L• = (L◦ \ {v}) ∪ F V (e)
hL• i v = e hL◦ i
hL•1 i S1 hL•2 i
hL•1 i S1 hL◦ i
Skip
hL•1 i
hL•2 i S2 hL◦2 i
S1 ;
hL•2 i S2 hL◦ i
L•2 6= ∅
S2 hL◦2 i
L• = L•1 ∪ L•2 ∪ F V (c)
•
hL i if (c) S1 else S2 hL◦ i
While
hB • i S hB ◦ i
B ◦ = L◦ ∪ B • ∪ F V (c) (minimal)
L• = B • ∪ L◦
L◦ 6= ∅
hL• i while (c) S hL◦ i
Fig. 2. System of inference rules for fully live programs
While the other rules are straight-forward, the While rule deserves some
discussion. Unlike the two branches of the if statement, the whole loop’s liveout set L◦ is not identical to the loop body’s live-out set B ◦ : Typically there are
loop-carried dependences, i. e., cases where a variable is set on one iteration of
the loop and its value is read on a later iteration. Such variables are live out of
(and live into) the loop body, but if they are no longer used once the loop has
terminated, they are not live out of the loop. When performing a derivation in
the system, we must guess or calculate the set of these additional variables.
Let fS denote the liveness transfer function corresponding to the loop body
statement S. Then from the liveness triple hB • i S hB ◦ i we have B • = fS (B ◦ ),
and the equation B ◦ = L◦ ∪ B • ∪ F V (c) means that B ◦ is a fixed point of the
function λ B. (L◦ ∪ fS (B) ∪ F V (c)). The minimality side condition additionally
specifies that we are interested in the least fixed point of this function. This least
fixed point exists and is unique [10].
In Figure 3 we illustrate the use of the inference system to prove full liveness of
a program taking an input variable n (assumed to be non-negative) and returning
the n-th Fibonacci number. We omit some details to focus on the analysis of the
loop. Note that only the return variable a is live after the loop. However, the
live-out set of the loop’s body is {a, b, n}. This includes the return variable a and
the variable n that is used in the loop condition. It also includes the variable b,
which is the element computed by fixed point iteration: The value of b at the
end of the loop body will be used on the next loop iteration, if any. Conversely,
if b were not live at some point in the loop body, our inference system would not
h{a, b, n}i n = n - 1 h{a, b, n}i
··
h{a, n, t}i b = t h{a, b, n}i
·
h{b, n, t}i a = b h{a, n, t}i
··
·
··
·
h{a, b, n}i t = a + b; a = b; b = t; n = n - 1 h{a, b, n}i
h{a, b, n}i t = a + b h{b, n, t}i
h{a, b, n}i while (n > 0) { t = a + b; a = b; b = t; n = n - 1 } h{a}i
..
.
h{n}i a = 0; b = 1; while (n > 0) { t = a + b; a = b; b = t; n = n - 1 }; return a h∅i
Fig. 3. Example derivation proving full liveness.
allow derivation of a triple for the assignment b = t. Indeed, all assignments in
the loop body satisfy the condition that they define variables that are live after
the assignment. That is, this program is fully live by our definition.
2.3
Limitations of the system
Note that fully live programs may still contain opportunities for simple optimizations that remove code that does not have interesting effects. For example,
programs accepted by the inference system above may contain fragments like
if (0) { ... } else { ... } where one of the branches of the if statement
is unreachable and thus irrelevant; or assignments like y = ...; x = y - y;
where the computation for the value of y is irrelevant for x’s final value of 0.
Our inference rules do not consider the semantics of the code in enough detail
to exclude such cases.
Our claims with regards to full liveness are relative to a purely syntactic
notion of liveness that does not consider such semantic issues. In particular, we
cannot guarantee that the liveness analysis embedded in these rules is equivalent
to liveness analysis as performed by any given compiler. Any other analyses or
transformations performed by the compiler before liveness analysis may influence
the results, typically making the compiler’s results more precise than ours.
As our experimental results in Section 4 show, our generator performs well
nonetheless, so we can leave refinements of the system for future work.
2.4
Generating fully live programs
The inference rules can be translated almost directly into an executable random
(or exhaustive) generator of fully live programs. Like traditional liveness analysis,
generation proceeds backwards, i. e., in the direction opposite control flow.
The side conditions of the inference rules ensure that a fully live program
always ends in a return statement, as no other statement may have an empty
live-out set. The generator can thus start by picking a random program variable v
and generating a statement return v with live-in set L = {v}. It then iteratively
prepends random statements S to the current program fragment and updates the
live-in set according to fS (L). The possibilities for the generation of S are guided
by L. In particular, if the generator decides to generate an assignment statement,
the target variable v must be an element of L at that point. Conversely, if L
ever becomes empty, generation of the current block of code must stop at that
point: Any code preceding that point would be dead. Figure 4 shows pseudocode
of such a generator in an OCaml-like functional language.
Every statement generation function takes a live variable set L and returns a
pair of a newly generated statement and an updated live variable set according
to the statement’s transfer function. We iterate statement generation and collect
a list of statements forming a block. In this presentation we omit helper functions
such as the ones for generating random variables and expressions.
As before, the handling of loops merits more discussion. Just as the inference
system needs a minimal set B • containing new variables that are live into and out
of the loop body, the random generator must synthesize such a set of variables.
But here the problem is more difficult: During inference, we can start with an
initial live-out set and derive the eventual live-in set by fixed-point iteration.
During code generation this is not possible since we cannot analyze the loop
body before we have constructed it. Instead, we first generate a random set of
newly live variables and let this choice guide generation of the loop’s code.
This solution relies on the following observation: The new variables we are
interested in are ones that are defined before the loop, may be defined on some
loop iteration, and then used on some later iteration. In the example of Figure 3,
this is the case for variable b, which stores the next Fibonacci number for assignment to a on the subsequent iteration. Using the names in the While rule
of Figure 2, the set of these ‘new’ variables is B ′ = B ◦ \ (L◦ ∪ F V (c)). It follows
that B ′ ⊆ B • , i. e., every b ∈ B ′ is live into the loop.
To generate a fully live loop, we choose a random set B ′ of new variables
and generate a loop body in a way that ensures that B ◦ = L◦ ∪ B ′ ∪ F V (c) is a
least fixed point of the loop. For this we add B ′ to the live variable set L before
generating a loop body block (along with the live variables F V (c) generated
by the loop condition). The generated loop body may or may not define or
use variables in B ′ , and thus these variables may or may not be live into the
generated block. However, we must force them to be used in the loop body and
be live into the body: If there were some b in the generated B ′ that is not live
into the loop body, we would violate the condition B ′ ⊆ B • established above.
On the other hand, if b were live into the loop body but did not have a use
anywhere in the body, then B • would not be minimal and hence B ◦ would not
be a least fixed point of the constraint system.
To ensure a correct, minimal solution, we therefore find the set V of all b ∈ B ′
that are not in L′ or that are in L′ but have no use in the generated loop body.
let random_statements L code =
if L 6= ∅ then
let (S, L′ ) = random_statement L;
random_statements L′ (S :: code)
else
(code, L)
let random_statement L =
let statement_generator = random_select [assignment; branch; loop];
statement_generator L
let assignment L =
let v = random_select L;
let e = random_expression ();
("v = e", (L \ {v}) ∪ F V (e))
let branch L =
let (t, L1 ) = random_statements L [];
let (f, L2 ) = random_statements L [];
let c = random_expression ();
("if (c) t else f ", L1 ∪ L2 ∪ F V (c))
let loop L =
(∗ See main text for explanation. ∗)
let c = random_expression ();
let B ′ = random_variable_set ();
let (code, L’) = random_statements (L ∪ B ′ ∪ F V (c)) [];
let V = {b ∈ B ′ | b ∈
/ L′ or b not used in S};
if V 6= ∅ then
let e = random_expression_on_variables V;
let v = random_select L′ ;
let code’ = "v = e" :: code;
("while (c) code′ ", (L′ \ {v}) ∪ V ∪ L)
else
("while (c) code", L′ ∪ L)
(∗ Start generation with the terminating return statement. ∗)
let v = random_variable ()
let (code, L) = random_statements {v} ["return v"]
Fig. 4. Pseudocode of a liveness-driven random program generator.
We pick a random live variable v ∈ L′ and prepend an assignment v = e to the
generated loop body, where e is an expression containing all the variables in V .
This final loop body ensures that all variables in B ′ are live into it and used in
it, hence ensuring that B ◦ = L◦ ∪ B ′ ∪ F V (c) is a least fixed point of the loop’s
liveness constraint system.
A small detail not illustrated in the pseudocode is the case when L′ is empty
at the beginning of the generated loop body. This can only be the case if the
first statement in the body is an assignment of a constant expression (i. e., not
using any variables) to some variable v, since such assignments are the only
statements that can remove variables from the live variable set without adding
any new ones. In this case, we replace this assignment’s right-hand side with the
expression e generated as above.
3
Implementation
We have implemented the random program generation algorithm sketched above
as a plugin for Frama-C [6]. Frama-C is a general, extensible framework for
source-level analysis and transformation of C programs. It is written in OCaml
and can be extended with plugins written in that language. For this work, we
do not need any advanced Frama-C features but benefit from its AST (abstract
syntax tree) type definitions, utilities for managing variables and constructing
AST fragments, and its pretty-printer for outputting the generated AST as C
source code. As these general parts are provided by Frama-C, ldrgen itself can
be quite small: It consists of only about 600 lines of generator code, plus some
utilities and configuration.
ldrgen is free software, available at https://github.com/gergo-/ldrgen.
3.1
Random generation
The core of the generator has the same structure as the pseudocode in Figure 4.
After generating an empty function definition and a return statement, it fills in
the function’s body by generating a fully live sequence of statements as in the
pseudocode. Statements are represented by AST fragments; we never need to
worry about generating actual C syntax. The current version of ldrgen always
generates a single function.
Random expressions are generated by choosing an operator among the arithmetic operators available in C and recursively generating the appropriate number of operand expressions. At this point, C’s type system becomes relevant; if
needed, we insert type casts to ensure all operands of an operator have the same
type. Type casts are also needed in some other cases: Bitwise operators and the
modulo operator cannot be applied to floating-point numbers in C, so we insert
conversions to integer types in such cases.
Many C operators may invoke undefined behavior when applied to inappropriate values. Two examples are division by zero and signed integer arithmetic
overflow. Unlike Csmith in its default mode, ldrgen does not try to guard against
such undefined operations, except for two cases that compilers have repeatedly
warned us about: We clamp the right-hand-side operands of bit-shift operations
to the bit size of the expression on the left-hand-side, and we always generate
division and modulo operations of the form e1 /(e2 +c) for some constant c instead of just e1 /e2 . The idea behind this is that e2 +c is less likely to evaluate
to zero than a random expression in general. This approach is primitive, but we
have found it to work well in practice.
Leaves of expressions are constants or variables. For constant literals we
simply generate a random number. For a variable use we either use a previously
used variable or generate a new one. Variables generated in this way may be
local variables or function parameters. Both can be used in expressions, but we
only generate assignments to locals, not to parameters.
Some of the generator’s choices are weighted by manually chosen parameters
to ensure generation of somewhat more realistic-looking programs. For example,
we prefer generation of basic arithmetic operations to bitwise operators. We
also ensure that loop and branch conditions are not constant expressions, i. e.,
that they contain at least one variable. In order to avoid trivial non-termination
issues, we also ensure that every loop body’s final statement is an assignment to
some variable that occurs in the loop condition. If there were no modification
of any of these variables at all, a loop once entered could never terminate. Even
so, termination is not at all guaranteed.
Bottom-up generation of the function’s body may stop if there are no more
live variables, or if a user defined limit is reached. In this latter case, there may
remain live local variables at the start of the function’s body. Their liveness
means that they may be used without being assigned to, so we must ensure that
they are initialized. We therefore finalize the function definition by initializing
all such live-in variables to constants or to the values of function parameters.
3.2
Configuration
The generator’s behavior may be tuned using command-line arguments. These
may specify features of the sub-language of C that is used. For example, the
user may request the generation of code that only uses integer types, or only
floating-point types. They may also specify that no bitwise operations or no
divisions should be generated, and whether loops may be generated. Other flags
specify structural properties: The maximal number of statements per block, and
the maximal nesting depths of statements and expressions.
ldrgen’s random generation uses OCaml’s standard pseudorandom number
generator, which can be seeded with a random seed or with a seed value specified
as a command line argument. Invoking a given version of ldrgen with a fixed
set of arguments and a fixed seed thus always gives the same reproducible result.
3.3
Extensions to the basic model
We describe two extensions to the core language of Figure 1 that are already
implemented in ldrgen: very limited uses of pointers and for loops over arrays.
First, in addition to the arithmetic types used so far, we can generate function
parameters of type T * (pointer to T) for some arithmetic type T. A parameter p
of such type can be used in generated code as *p. We currently do not generate
assignments to such dereferenced pointers, nor any pointer arithmetic.
Second, we want to generate arrays and restricted forms of loops over them
in order to exercise loop optimizations such as unrolling or vectorization. For
this we generate pointer arguments T *arr which are only used in for loops of
the following form:
v = . . .;
for (unsigned int i = 0; i < N; i++) {
v = v ◦ f (arr[i]);
}
Here, N is a global variable considered to hold the array’s size, f (arr[i]) is
a random expression involving arr[i], and ◦ is a randomly chosen binary arithmetic operator. This loop pattern implements a map-reduce operation, mapping
some function f over the array and reducing (folding) the result with ◦. It is
currently the only kind of for loop we implement, but this would be easy to generalize. Similar forms of loops are already generated by Csmith, but there their
results are virtually never used. In ldrgen, we choose a loop result variable v
that is live after the loop to ensure that it is used.
3.4
Future extensions
In the future, we are planning to extend ldrgen to generate structure types and
allow the use of their members.
In the longer term, ldrgen will also be extended to support programs consisting of several random functions which may call each other. We are not planning
to support non-structured control flow using goto. The more structured break
and continue statements might eventually be supported, but this is not a priority as they complicate the structural liveness analysis.
4
Evaluation
The design goal of ldrgen was to have a random program generator that exposes
as much interesting code as possible to all passes of the compiler under test; recall
that we found Csmith-generated code to contain much dead code which is never
seen by many parts of the compiler because it can be optimized away early on.
We will have achieved our goal if, for comparable amounts of generated C code,
ldrgen’s output results in more, and ideally more varied, assembly code than
Csmith’s output. We therefore compare the two generators along these lines. We
do not claim superiority to Csmith in any other regard, especially not concerning
its power to find subtle miscompilation bugs.
Csmith is designed to run complete, self-contained applications consisting of
several functions, driven by a main function. In contrast, ldrgen only generates
individual functions without a driver. However, Csmith’s many configuration
options allow us to ask it to generate files consisting only a single function
without main.1
Table 1 presents our experimental results for 1000 programs each generated
by Csmith and ldrgen. We investigate three characteristics of the generated
programs: lines of C code, number of instructions in the generated code, and
number of unique opcodes in the generated code. In all cases, the C code was
compiled to x86-64 machine code using GCC 5.4.0 with optimization setting -O3.
For each characteristic, the table shows the total over the 1000 files as well as
the minimum, median, and maximum values. (In cases where the median is not
unique, we chose the arithmetic mean of the two closest values.)
Table 1. Comparison of code generated by Csmith and ldrgen in 1000 runs each.
min
25
12
median
368.5
411.5
max
2953
1003
total
459021
389939
Csmith
ldrgen
1
1
15.0
952.5
1006
4420
45606
1063503
unique opcodes Csmith
ldrgen
1
1
8
95
74
124
146
204
lines of code
instructions
generator
Csmith
ldrgen
Our command line flags for Csmith were chosen in order to generate comparable numbers of lines of C code to ldrgen. In fact Table 1 shows that it
generates somewhat more, but these numbers are difficult to compare precisely
because Csmith-generated code tends to contain many initializers for global variables; ldrgen does not generate any global variables at all. We believe that the
settings we chose allow a fair comparison of the generators.
Next we compare the number of instructions (executable code only, excluding
static data, assembler directives etc.) emitted by the compiler for the generated
source files. ldrgen was designed to increase this number compared to Csmith,
and the table shows that we have succeeded: While on average Csmith’s output compiles to a single machine instruction per ten lines of code, ldrgen’s
output has almost three instructions per single line of source code. Overall,
ldrgen-generated programs compile to about 20 times as much machine code
as Csmith-generated programs of comparable size. We can also see that the distribution for Csmith is highly skewed: The median shows that at least half of
the functions generated by Csmith compile to 15 instructions or fewer. This
also confirms our initial, more informal observation that Csmith-generated code
tends to contain large amounts of dead code. ldrgen manages to generate code
with a less skewed distribution, and in particular with generally higher numbers
of emitted instructions.
1
The concrete flags we used were –nomain –float –max-funcs 1 –no-safe-math
–max-block-size 8 –concise.
On a side note, we remark that both Csmith and ldrgen sometimes generate functions that compile to a single machine instruction. Inspection showed
that this happens in cases where the compiler recognizes that a function ends
up in an infinite loop without externally visible side effects. Such functions are
then compiled into a single unconditional jump instruction looping back to itself. Many other functions compile to two instructions, typically some simple
operation on a function argument or a constant followed by a return. It would
be difficult to completely avoid generating infinite loops, but comparatively easy
(at least within ldrgen) to avoid generating functions that return after a single
operation. For both Csmith and ldrgen, about 10 % of all cases fall into one of
these trivial categories (with Csmith producing fewer infinite loops).
We analyze the coverage of the instruction set in the generated code by
looking at the number of different opcodes generated. Here, too, we see that individual functions generated by ldrgen have a more varied instruction mix than
functions generated by Csmith: Even the median for ldrgen is higher than the
maximum for Csmith. Totaling over all the machine code in 1000 functions, we
see that Csmith-generated code compiles to a mix of 146 different opcodes, while
ldrgen-generated code contains 204 different opcodes, an increase in instruction
set coverage of 40 %. Inspection of the sets of opcodes shows that this difference
is almost entirely due to various vector (SIMD) arithmetic instructions generated
for ldrgen’s code. Compiling to such instructions was the goal of adding for
loops over arrays to ldrgen. Manual inspection of some cases shows that such
loops are indeed the origin of these instructions. Disabling generation of for
loops in ldrgen brings its total number of unique instructions down to 147,
comparable to Csmith.
One of the few opcodes emitted for Csmith-generated code but not for ldrgen
are call instructions to memcpy which are sometimes generated by compilers for
structure copies. ldrgen currently does not generate structures at all.
Finally we compare the speed of the two generators. Generating the 1000
files each analyzed above took 871 seconds with Csmith and 124 seconds with
ldrgen (Csmith backtracks if it finds that it has generated unsafe code). Csmith
generates about 527 lines of C code per second, with ldrgen generating 3140
(about 6 × more). With respect to final machine code, Csmith-generated code
compiles to about 52 instructions per second of generation time, whereas ldrgen
produces 8563 (about 160 ×).
5
Related work
The best-known random program generator is Csmith [16], based on an earlier
system called randprog [4, 15]. Csmith generates complete, self-contained programs that take all their input from initialized global variables and compute an
output consisting of a hash over the values of all global variables at the end
of execution. The generator is designed to only generate programs with welldefined semantics: Operations that may be undefined in C, such as overflowing
signed integer arithmetic, are guarded by conditionals that exclude undefined
cases (these guards can be disabled, and we disabled them for the experiments
reported above). Like ldrgen, Csmith performs data-flow analysis during generation, although the details differ due to the differing design goals. Csmith’s
forward analysis computes points-to facts and uses them for safety checks. If
the checks fail, Csmith backtracks, deleting code it generated until a safe state
is reached again. In contrast, ldrgen’s data-flow analysis only deals with liveness, and ldrgen never backtracks: Full liveness of variables in loops is ensured
by construction. Csmith generates a larger subset of C than current or currently planned versions of ldrgen, including unstructured control flow and less
restricted use of pointers.
Csmith has been used to find hundreds of bugs in C compilers when compiling
the programs it generates [16]. It has also been used as the basis of mutationbased systems, where Csmith’s output was modified using other tools to provoke
compiler bugs [7]. The CLsmith tool derived from Csmith has been used to find
many bugs in OpenCL compilers [9]. Another notable generator is jsfunfuzz for
JavaScript, which has found thousands of bugs in JavaScript engines [13].
The JTT program generator [17] is aimed directly at testing compiler optimizations. It uses a model-based approach, where generation is guided using test
scripts. These scripts contain code templates and temporal logic specifications
of the optimizations to be tested. For example, the authors specify opportunities
for dead code elimination as cases where a variable is assigned, then assigned
again before being used. The test script contains a temporal logic formula expressing this pattern and the test condition that the compiler should eliminate
the first assignment. Using this script, JTT generates test programs containing
this pattern. JTT was used successfully to find bugs and increase the test suite’s
statement coverage for an industrial C compiler.
Other work specifically aimed at testing and comparing program verification
tools generates code from randomly generated LTL formulae [14]. The generated
code is guaranteed to satisfy the specified temporal properties.
Our formulation of liveness analysis as set of structural inference rules is
inspired by formulations of interval-based data-flow analysis where reducible
programs are decomposed into components called intervals, and analysis data is
efficiently propagated among the intervals [1, 2, 5].
6
Conclusions
We presented ldrgen, a new generator of random C programs designed for testing C compilers. In contrast to Csmith, the dominant player in this field, ldrgen
is driven by liveness analysis to avoid generating dead code. We designed an inference system to capture our liveness analysis and implemented its rules as an
excutable program generation system.
ldrgen is implemented as a plugin for the Frama-C framework. Our evaluation of ldrgen in comparison to Csmith shows that we have achieved our goal of
generating C code that compiles to larger amounts of machine code with a more
interesting instruction mix, including many SIMD instructions. We are actively
using ldrgen in a project on finding missed optimizations in compilers. Because
it is able to exercise loop optimizations not usually addressed by Csmith, it may
also be useful for finding correctness bugs in these optimizations.
Acknowledgments
The author would like to thank the anonymous reviewers and John Regehr for
insightful comments on earlier versions of this paper. This research was partially
supported by ITEA 3 project no. 14014, ASSUME.
References
1. Allen, F.E.: Control flow analysis. SIGPLAN Not. 5(7), 1–19 (Jul 1970),
http://doi.acm.org/10.1145/390013.808479
2. Cocke, J.: Global common subexpression elimination. SIGPLAN Not. 5(7), 20–24
(Jul 1970), http://doi.acm.org/10.1145/390013.808480
3. Dave, M.A.: Compiler verification: A bibliography. SIGSOFT Softw. Eng. Notes
28(6) (Nov 2003), http://doi.acm.org/10.1145/966221.966235
4. Eide, E., Regehr, J.: Volatiles are miscompiled, and what to do about it. In: EMSOFT ’08. ACM (2008), http://doi.acm.org/10.1145/1450058.1450093
5. Graham, S.L., Wegman, M.: A fast and usually linear algorithm for global flow
analysis. J. ACM 23(1), 172–202 (Jan 1976)
6. Kirchner, F., Kosmatov, N., Prevosto, V., Signoles, J., Yakobowski, B.: Frama-C:
A software analysis perspective. Formal Aspects of Comp. 27(3), 573–609 (2015)
7. Le, V., Afshari, M., Su, Z.: Compiler validation via equivalence modulo inputs. In:
PLDI ’14. pp. 216–226. ACM (2014), http://doi.acm.org/10.1145/2594291.2594334
8. Leroy, X.: Formal verification of a realistic compiler. Commun. ACM 52(7), 107–
115 (Jul 2009), http://doi.acm.org/10.1145/1538788.1538814
9. Lidbury, C., Lascu, A., Chong, N., Donaldson, A.F.: Many-core compiler fuzzing.
In: PLDI ’15. pp. 65–76. ACM (2015)
10. Nielson, F., Nielson, H.R., Hankin, C.: Principles of Program Analysis. SpringerVerlag New York, Inc., Secaucus, NJ, USA (1999)
11. Perennial,
Inc.:
ACVS
ANSI/ISO/FIPS-160
C
validation
suite,
http://www.peren.com/pages/acvs_set.htm
12. Plum
Hall,
Inc.:
The
Plum
Hall
validation
suite
for
C,
http://www.plumhall.com/stec.html
13. Ruderman, J.: jsfunfuzz (2015), https://github.com/MozillaSecurity/funfuzz/tree/
master/js/jsfunfuzz
14. Steffen, B., Isberner, M., Naujokat, S., Margaria, T., Geske, M.: Property-driven
benchmark generation: synthesizing programs of realistic structure. International
Journal on Software Tools for Technology Transfer 16(5), 465–479 (Oct 2014)
15. Turner, B.: Random C program generator (2007), https://sites.google.com/site/
brturn2/randomcprogramgenerator
16. Yang, X., Chen, Y., Eide, E., Regehr, J.: Finding and understanding bugs in C
compilers. In: PLDI ’11. pp. 283–294. ACM (2011)
17. Zhao, C., Xue, Y., Tao, Q., Guo, L., Wang, Z.: Automated test program generation
for an industrial optimizing compiler. In: 2009 ICSE Workshop on Automation of
Software Test. pp. 36–43 (May 2009)
| 6cs.PL
|
Visual Representation of 3D Language
Constructs Specified by Generic Depictions
Jan Wolter
University of Paderborn
Department of Computer Science
Fürstenallee 11, 33102 Paderborn, Germany
[email protected]
arXiv:1311.5126v1 [cs.PL] 20 Nov 2013
Abstract
Several modeling domains make use of three-dimensional representations,
e.g., the “ball-and-stick” models of molecules. Our generator framework
DEViL3D supports the design and implementation of visual 3D languages for
such modeling purposes. The front-end of a language implementation generated
by DEViL3D is a dedicated 3D graphical structure editor, which is used to construct programs in that domain. DEViL3D supports the language designer to
describe the visual appearance of the constructs of the particular language in
terms of generic 3D depictions. Their parameters specify where substructures
are embedded, and how the graphic adapts to space requirements of nested constructs. The 3D editor used for such specifications is generated by DEViL3D,
too. In this paper, we briefly introduce the research field of 3D visual languages and report about our generator framework and the role that generic
depictions play in the specification process for 3D languages. Our results show
that our approach is suitable for a wide range of 3D languages. We emphasize
this suitability by presenting requirements on the visual appearance for different
languages.
Key words. three-dimensional depictions, visual languages, visual programming, automated generation, 3D interaction techniques.
Contents
1 Introduction
2
2 DEViL3D
3
3 Generic depictions
3.1 Application of Generic Depictions . . . . . . . . . . . . . . . . . . . . .
3.2 The 3D Editor for Generic Depictions . . . . . . . . . . . . . . . . . .
3.3 Code-generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
7
8
11
4 Range of application
12
5 Related Work
15
6 Conclusion
15
This paper is a extended version of my paper Specifying Generic Depictions of Language Constructs for 3D Visual Languages, which I presented at the VL/HCC 2013: http://dx.doi.org/10.
1109/VLHCC.2013.6645258
1
1
Introduction
Visual languages are particularly beneficial for domain-specific applications, since
they support graphical metaphors of their domain. Up to now the majority of visual languages are two-dimensional. Examples are LabVIEW [lab]—used in industrial automation and instrument control—and the well-known UML, which is used
to model several aspects of object-oriented software systems. Both languages use
two-dimensional representations, e.g., boxes and lines connecting them, in order to
visualize dataflow or dependences.
For some domains, using the third dimension is advantageous or necessary: Inherently three-dimensional languages that make use of real-world objects as architecturelike modeling domains can be represented without loss of information only in 3D. The
“ball-and-stick” models of molecules visualize atoms as balls and bonds between them
as sticks. The arrangement of the atoms in 3D space is the result of the electron cloud
repulsion and therefore the arrangement of language constructs to one another is inherently 3D.
Another argument in favor of 3D languages is to assign a semantic meaning to
each dimension. A good example is the web-based 3D editor ToneCraft [Din11] that
lets the user build music in 3D by composing boxes into a matrix-like area: the y-axis
represents the pitch of a tone, the x-axis represents the time and the z-axis makes
it possible to layer sounds. Even Petri Nets can benefit from a three-dimensional
representation, for example, Petri Nets modeling different aspects—such as control
flow and data flow—that have connections to one another. The representation of such
Petri Nets in 2D is confusingly complex due to crossing edges and confound aspects.
This can be solved by laying out the Petri Net in three dimensions, where each aspect
is represented on a different plane, which can be stacked together to one Petri Net
[R0̈7].
Moreover, the third dimension can be used to overcome limitations in 2D arrangements. For example, in some cases the 2D representation of UML diagrams is not
efficient enough and can be extended to 3D, e.g., to overcome the problem of intersecting edges in sequence diagrams. Alternatively, the third dimension can be used
to focus on specific classes of interest in class diagrams [GRR99].
Some further systems in context of visual programming make use of threedimensionality. Alice [KP07] is a 3D programming environment that employs a story
telling metaphor and teaches children to learn fundamental programming concepts
by programming the behavior of 3D objects, e.g., people or animals. This system’s
effectiveness has been proven by various user studies. Another programming environment that uses three-dimensional objects is AgentCubes [IRW09], the successor of
AgentSheets, which allows children to create interactive 3D games. The key challenges
of AgentCubes are intuitive mechanisms to create 3D objects incrementally, including
subsequent programming and animation aspects.
The above mentioned 3D languages and programming environments consist of
objects with different 3D shapes. Some languages—as molecular models, ToneCraft,
or Petri Nets—use relatively simple shapes like cubes, spheres, or cylinders. But
the 3D scenes composed with Alice or AgentCubes consist of more complex shapes,
visualizing real world objects such as people, buildings, or cars. There is an additional
challenging task: Because of the fact that language constructs can be nested inside
each other, their depiction needs mechanisms to adopt the size of interior constructs.
Our approach relies on a tool that makes the development of 3D language editors as simple as possible. The development of a language specific implementation is
2
justified, only if the effort is appropriately small. Therefore, effective generator systems are useful. We are developing the generator framework DEViL3D (Development
Environment for Visual Languages in 3D) that accomplishes this task. One central
part of developing visual languages (either two- or three-dimensional) is the definition
of the visual appearance of language constructs. For such a task, DEViL3D provides
a 3D editor to specify generic depictions for language constructs. They may consist of
a large set of three-dimensional geometric shapes and their parameters specify where
substructures are embedded, and how the graphic adapts to space requirements of
nested constructs.
Figure 1: A language construct before and after stretch.
Figure 1 shows an exemplary language construct consisting of a blue box and a
green sphere. The green sphere contains a text label and the blue box is able to embed
further language constructs visualized as red boxes. These language constructs are
laid out as a list, which grows along the x-axis. When new constructs are inserted,
the blue box has to adapt its space requirements to the needs of the nested list and
the green sphere has to move right.
The remainder of this paper is structured as follows. In Section 2, we give a brief
overview of DEViL3D. Since this paper focuses on generic depictions of 3D languages,
Section 3 introduces these and presents an interactive 3D language editor to compose
them and illustrates a method instantiated depictions perform to adapt their size
according to embedded constructs. In Section 4 we show that specifying generic
depictions according to our approach is applicable for a large set of languages. We
will give a survey over developing generic depictions for different languages. Then we
discuss related work and compare it to our approach. Section 6 concludes the paper.
2
DEViL3D
The generator system DEViL3D [Wol12] allows to generate 3D structure editors,
supporting the direct manipulation paradigm [Shn83] and, therefore, prevent the user
from constructing syntactically incorrect programs. DEViL3D combines approved
concepts of the predecessor system DEViL [SKC06, SCK07, SK03] and new aspects
necessary to construct three-dimensional programs. This section gives a brief overview
on the steps elementary to generate a structure editor with DEViL3D. It includes an
overview on several specification parts and, additionally, a brief outline on 3D-specific
aspects that are applicable in all generated editors.
Figure 2 visualizes the specification process. It is layered into three parts. The
upper area shows specification parts a language designer has to formulate: the abstract structure, visual representations, and code generators. The area in the middle
illustrates our generator framework DEViL3D. As input, DEViL3D gets a language
specification and generates a language processor which has a dedicated 3D graphi-
3
Language designer
Abstract structure
Visual representations
Analysis
&
code generation
attribute evaluator
generic depictions
Generator developer
DEViL3D
visual patterns
editor for generic depictions
consistency checks
synchronization
3D engine
Domain expert
Chemical
formula:
C9 H8 O4
...
generated editor
transformation
Figure 2: Specification process with DEViL3D.
cal structure editor as its front-end. Domain experts use such editors to construct
three-dimensional programs of their domain, e.g., molecular models.
The abstract structure describes the language constructs and how they are connected, without defining a concrete representation. For such purpose, a specifically
tailored textual domain specific language is used, which is strongly related to object
oriented programming languages. It is based on well known concepts like classes,
inheritance, attributes, and references.
The specification of the visual representation is based on attribute grammars,
which are translated into computations of the graphical representation and arrange
objects in 3D space. The attribute grammars are based on a context-free grammar
that is generated from the abstract structure. The arrows in Figure 2 indicate this
dependency. DEViL3D provides a library of visual patterns that encapsulate common
representation arrangements like three-dimensional sets, lists, line connections, or
cone-trees. They are defined as visual roles, which can be assigned to symbols of
the grammar in a declarative way. The visual patterns automatically contribute
layout and interaction properties. The generic depictions, which define the visual
representation of a language construct, are referenced by visual patterns. For the
generation of a molecular editor with DEViL3D, amongst others, generic depictions
for kinds of atoms and for bonds are needed. They play the role of building blocks
from which the visual program will be composed.
4
Further to the visual representation, the language designer can define a set of code
generators based on attribute grammars that transform the 3D program into different
textual representations.
To ensure that an editor generated by DEViL3D is as easy as possible to use,
we have analyzed established three-dimensional editors from different domains, e.g.,
Autodesks 3ds Max [Aut], which allows to create 3D models and animations. From
this and other systems, we have adopted basic techniques to make the interaction
with the 3D scene as simple as possible. Such challenge will become clear, if we
compare editors that allow the creation of 2D and 3D representations. They can be
distinguished by the degree of freedom (DOF) that describes the possibilities of object
placement in space. The 2D space has three DOFs: translation along the two axes
and rotation around the neutral point. In contrast, there are twice as many DOFs in
three-dimensional space, namely, translation and rotation along all three axes.
In order to cope with such an increase in complexity, editors generated by
DEViL3D comprise techniques for navigation, interaction, and layout. Our aim is
that the editors are usable with a classical mouse, which can not directly capture all
six DOFs. To support this, dedicated widgets are provided that perform interaction
tasks like translating, scaling, or rotating. In complex 3D scenes objects can occlude
each other, but a first-person-view camera lets the user navigate inside the scene and
explore it.
The layout and interaction tasks are encapsulated in visual patterns. The language designer does not need to implement such functionality. The assignment of
visual patterns to language constructs is generally sufficient. Interaction tasks are
automatically tailored to the needs of the representation of a visual pattern: For example, elements of a linearly ordered list can only be moved along the direction of
its arrangement. The task of inserting language constructs is triggered by so called
insertion contexts that are also tailored to the needs of the visual pattern.
The implementation of 3D-specific aspects makes use of the underlying jMonkeyEngine [jme], which in turn uses OpenGL to render 3D scenes. To organize the
3D scene efficiently, the provided scene graph—a data structure to organize the 3D
scene—is used extensively.
3
Generic depictions
Generic depictions describe the visual appearance of constructs of a particular language in terms of generic 3D graphics. From a formal point of view generic depictions
are an abstract concept that can be described by a quadruple of graphical primitives,
representation properties, containers, and stretch intervals: D = (P, R, C, I). These
parts carry out the following tasks:
• A collection of graphical primitives P determine the graphical representation
of a language construct. The following types of primitives are available: P
determine the graphical representation of a language construct: P = Box ∪˙
Sphere ∪˙ Cone ∪˙ Cylinder ∪˙ Arrow ∪˙ Line ∪˙ Quad ∪˙ T orus ∪˙ 3DM odel ∪˙ T ext.
3D models allow to integrate objects with more complex shapes.
• The representation properties R describe materials, e.g., color or texture definitions, that can be mapped to graphical primitives P.
• A set of containers C is responsible to embed nested objects of arbitrary size,
when the generic depiction is instantiated. Each container needs a unique name.
5
• The specification of layout properties is managed by a set of stretch intervals
I. Such intervals determine, which part of a container grows, when the size of
nested objects exceeds the container’s size.
Figure 3: 3D editor for generic depictions.
The best way to build visual constructs is to do it visually. Such an approach
ensures a close mapping between the domain world and the notation (according to
the cognitive dimension closeness of mapping [GP96]). Figure 3 shows a screenshot of
the generic depiction editor that provides a 3D canvas in which the language designer
visually composes a set of generic depictions for some language constructs. The
depiction shown in the figure consists of two graphical primitives in form of a box and
a sphere, two containers named c1 and c2, and four stretch intervals. One container is
located inside the box, the other one in the sphere. Each stretch interval is responsible
for one dimension and is located on light red colored coordinate axes. The abstract
visual program in Figure 1 uses the generic depiction specified in Figure 3.
Containers constitute the interface of generic depictions. If a language designer
wants to change the visual representation of a language construct, two generic depictions can replace each other, if they coincide according to the number and the naming
of their containers.
The following requirements restrict a semantically correct generic depiction: Each
container must be covered by at least one stretch interval in each dimension. Otherwise, it is not clear how to respond to increasing size requirements of nested constructs.
6
Any two stretch intervals must not overlap, so that in all cases a uniquely determined
interval is responsible to stretch the container.
The following subsection will convey an impression, how the stretch algorithm for
instantiated depictions proceeds. Part 3.2 presents how generic depictions can be
constructed with our 3D editor. Afterwards, we explain the code generation process
that transforms the 3D program into Java code. We have developed the 3D editor
for generic depictions by pursuing a bootstrapping approach. The editor is specified
with DEViL3D. Hence, we have acted as a language designer and developed a set of
high-level specifications as presented in Figure 2, which describes the 3D language for
generic depictions. Of course, visual representations of language constructs used in
this editor have initially been implemented on a lower level, i.e. textual descriptions
and Java code. This specification part is comparable with the code the editor for
generic depictions generates within DEViL3D.
3.1
Application of Generic Depictions
As seen in Figure 1 language constructs of 3D visual programs—which support embedded substructures—have to adopt their size according to the requirements of these
substructures. Such nested structures require the specification of containers and
stretch intervals as seen in Figure 3. In general, the embedded constructs can be
laid out according to any visual pattern. In the example of Figure 1 the list pattern
is used for the constructs inside the blue box.
By inserting more constructs in the list, the container’s size could become insufficient. To adapt the size, an algorithm automatically stretches the container linearly.
The algorithm operates on one depiction and iterates over all containers for each spatial dimension. If the preferred size, determined by the nested constructs, exceeds the
actual container size, all stretch intervals that intersect the container are computed.
The intervals of a container that are covered by a stretch interval will be stretched
linearly. For this purpose, each container must be covered by at least one stretch
interval.
c1
c2
c1
a
d
c1.actSize = a
c1.prefSize = a + b
c2
a
b
d
c1.actSize = a + b
Figure 4: Behavior of the stretch algorithm.
Figure 4 shows a schematic sketch of the generic depiction of Figure 3 reduced to
the x-axis to illustrate the stretch algorithm for one dimension. Initially the actual size
of container c1 is a, but the embedded constructs need more space, so the preferred
size is a + b. Hence, container c1 must be stretched to reach the new size. The stretch
process behaves as if the containers and the primitives were printed onto elastic rubber
and the start and end position of the stretch intervals were handles. To stretch the
container and the primitives, the algorithm “pulls” the margins of the interval. Areas
that are not covered by stretch intervals are not be stretched. In particular, the
distance of two such points must remain the same after the transformation. Hence,
7
the distance d between the box and the sphere is the same after application of the
stretch algorithm and the sphere with container c2 inside has simply to move right.
This algorithm is encapsulated in DEViL3D and always comes into operation
when new elements are inserted into a container. From an editor user’s point of view,
the algorithm provides a natural behavior when new language elements are inserted.
Language designers do not need to care about the dynamic adaptation of generic
depictions, as the generator system automatically provides it.
3.2
The 3D Editor for Generic Depictions
The editor for generic depictions provides a multi-document interface that shows the
three-dimensional view, where a depiction is created (see Figure 5(a)). In the middle
of the canvas the generic depiction can be constructed. The buttons on the lefthand side represents components of generic depictions (see item list at the beginning
of Section 3) that can be inserted into the 3D view. When a button is clicked,
all appropriate three-dimensional insertion contexts appear in the 3D canvas and
highlight valid positions, where such an object can be inserted. The nature of an
insertion context is determined by the way the language object is organized in the
scene.
(a) Cube-shaped insertion context to insert containers and primitives.
(b) Plane-shaped red insertion context to insert stretch intervals.
Figure 5: Different kinds of insertion contexts.
8
Since graphical primitives and containers can be placed anywhere in 3D space, the
insertion context is cube-shaped and includes a plane to indicate a three-dimensional
position (see Figure 5(a)). This plane can be moved along the z-axis. The 3D position
of the object is then determined by a mouse click on the plane. The stretch intervals
can be positioned onto coordinate axes, which are located in the 3D scene. To insert
such intervals, a plane-shaped insertion context appears along the chosen coordinate
axis (see Figure 5(b)), whereby the user can determine a position on the axis.
After the insertion of a generic depiction’s component, the language designer is
able to modify it. In order to translate, scale, or rotate the object, a suitable widget
attached to the object in question (see Figure 6). Such widgets are automatically
adjusted to the requirements of the object according to the degree of freedom the
object has. The graphical primitives and containers can be positioned everywhere
inside the yellow insertion context and can therefore be manipulated along all three
spatial dimensions. The widget to translate an object provides three arrows and three
planes between them. The user may translate it along one dimension separately (via
an arrow) or along two dimensions simultaneously (via a plane). However, stretch
intervals can only be reshaped along one spatial dimension, as depicted in Figure
6(b).
(a) Translate 3D
(b) Translate 1D
(c) Scale
(d) Rotate
Figure 6: Widgets to manipulate objects.
These manipulation techniques are generally available in all structure editors generated with DEViL3D. From a language designers point of view, the automatic adaptation of widgets is achieved by the application of visual patterns. Since the editor
for generic depiction is bootstrapped, graphical primitives and containers are laid out
according to the 3D set pattern and stretch intervals according to the 1D set pattern.
The latter determines the position along two dimensions and lets the user only change
the position along one continuous dimension.
All interaction and manipulation tasks can be performed by using a classical 2D
mouse. To do so, the editor supports a ray casting metaphor. If the user wants to
interact with an object, a ray starting at the mouse cursor is shot into the 3D scene
to determine the first object that is intersected by the ray. This technique supports
the direct manipulation approach in 3D and is used for all mentioned tasks: the interaction with insertion contexts to insert objects, the selection, and manipulation of
language constructs with widgets. In some cases it is desirable to select multiple objects at the same time. The structure editor provides different methods to accomplish
this task. As in many editors—2D or 3D—the user may select multiple objects one
after the other by picking them while a special control key is pressed. There are 3D
scenes where this approach is not applicable: Objects that are placed far away from
9
the actual camera position appear much smaller than near located objects, which
makes a precise selection difficult. To overcome this problem, we have developed a
cylinder metaphor inspired by Tavanti et al. [TDL04]. The user of the editor sees
only the circle-shaped cylinder cover on the 2D monitor screen (see Figure 7(a)). The
cylinder expands into the 3D scene and selects all objects that are enclosed by it. Of
course, the editor user can adjust the size of the lateral surface. But again, even this
method is not distinct enough, if the objects intended for selection cannot be captured
by a cylinder. A lasso metaphor may be better suited: An arbitrary polygon can be
created, which expands into the 3D scene as well to select the objects (see Figure
7(b)).
(a) Cylinder metaphor
(b) Lasso metaphor
Figure 7: Two ways to select multiple objects.
Particularly important for the visual representation of generic depictions is, apart
of the shape of the objects, their color. In the context of 3D applications the term
material has been established, which includes everything that influences what the
surface of a 3D object looks like: the color, texture, shininess, and transparency. The
generic depictions allow to define three different types of materials. Color materials
that define RGB and transparency values. A texture material that refers to a normal
image file to realize a special impression (e.g. the box in the top-right corner in Figure
7(b) that refers to an image file of the DEViL3D logo). And a custom material that
refers to a shader specification, which defines sets of instructions that are executed on
the GPU to realize special effects (see the box on the right-hand side in Figure 7(a)).
The language designer and user of the generic depictions editor may create arbitrary
many material objects, which are referred by almost all graphical primitives. The
only exception are 3D models, which model very complex shapes and were created by
external 3D modeling frameworks like Blender [ble].
Generally, 3D scenes can be more complex than 2D scenes and therefore objects
may be hard to locate when they are far away. To overcome such problems the user
needs mechanism to explore the 3D scene. To do so, a camera is used that allows to
explore the scene without an explicit target, or to search a particular target. The editor offers different ways to control the camera. The ViewSphere located on the upper
10
right corner allows to rotate the camera around any language object to be selected by
the user (see Figure 5(a)). Furthermore, the ViewSphere provides around its sphere
different segments labeled with a cardinal point to switch the viewing position rapidly
and show the scene from eight distinct positions. To indicate the actual position, the
editor provides a coordinate system in the lower right corner, which rotates according
to the camera and minimizes the lost-in-space problem. Besides these possibilities,
advanced users can control the camera by using mouse and keyboard. We have experienced to control the camera with 3D specific devices, e.g., a 3D mouse. To see the
scene simultaneously from different perspectives the user can switch to three additional lateral views by using the button in the upper left corner. Moreover, the lower
left corner can be expanded to see a larger overview of the scene (see. Figure 3).
These general navigation facilities make especially the orientation inside the 3D
scene and furthermore the construction of generic depictions much easier. For example, the user can navigate the camera closer to an appropriate insertion context
to precisely locate an insertion context. This is specifically important for the insertion of stretch intervals, which must be positioned in a way so that a container is
covered by at least one interval in each dimension. To satisfy the requirement, this
box must intersect the container. But to realize this requirement, the intervals must
be located on a position onto the coordinate axes indicating the projection of the
container onto the axes. Because of distortion effects, this position is often hard to
meet. To overcome it, we realized a depth cue that exactly shows the projection of
the container on the axes by a wired box with a blue bar at the back (see see Figure
5(b) and Figure 3). This facilitates the insertion of the intervals, because the user
can orient himself by the position of the bars. In Figure 3 the stretch interval located
on the y-axis coincides exactly with the position of the wired box, which indicates
the projection of container “c1”. The second stretch interval on the x-axis is smaller
than the projection of “c2” and therefore only covers the middle of “c2”.
The editor for generic depictions offers a way to indicate violations of requirements
(containers have to be covered by at least one stretch interval in each direction and
do not overlap each other) made on generic depictions. To ensure that only correct
depictions are created, the requirements will be checked before generating code from
the visual specification of the generic depictions. During the creation of generic depictions the user can check the program at any time by pressing a button that opens a
view that lists all violations. From a language developer’s point of view, requirement
checks are easy to realize: For each language construct so called check functions can
be defined. Inside these functions the whole abstract structure can be accessed by
using so called path expressions.
3.3
Code-generation
Using the editor for generic depictions the language designer constructs threedimensional programs to describe objects of his language. That visual specification
is transformed to Java code, to be integrated in the generated language, where it is
executed to draw program elements. Listing 1 shows the code generated from the
generic depiction shown in Figure 3. For each generic depiction a class will be generated, which inherits from the AbstractDepiction class that contributes features,
which are common for each depiction. The generated class inherits various functions
to add primitives, containers, or stretch intervals. To add a container, three pieces
of information are necessary: the position, the size, and the name (compare the parameters in lines 5 and 6). Stretch intervals need the declaration for which dimension
11
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
class RootDepiction extends A b s t r a c t D e p i c t i o n {
public RootDepiction () {
/* --- container --- */
addContainer ( new Vector3f (2 f , 2f , 2 f ) , new Vector3f (10 f , 10 f , 10 f ) , " c1
") ;
addContainer ( new Vector3f (23 f , 3f , 3 f ) , new Vector3f (6 f , 6f , 6 f ) , " c2 ") ;
/* --- stretch intervals --- */
a d d S t r e t c h I n t e r v a l ( ’x ’ , 2f , 9 f ) ;
a d d S t r e t c h I n t e r v a l ( ’x ’ , 25 f , 28 f ) ;
a d d S t r e t c h I n t e r v a l ( ’y ’ , 2f , 12 f ) ;
a d d S t r e t c h I n t e r v a l ( ’z ’ , 2f , 12 f ) ;
/* --- primitives --- */
addBox ( new Vector3f (0 f , 0f , 0 f ) , new Quaternion (0 f , 0f , 0f , 1 f ) , new
Vector3f (16 f , 16 f , 16 f ) , Materials . blue ) ;
addSphere ( new Vector3f (20 f , 0f , 0 f ) , new Quaternion (0 f , 0f , 0f , 1 f ) , new
Vector3f (14 f , 14 f , 14 f ) , Materials . green ) ;
}
}
Listing 1: Generated code for generic depiction.
they are responsible and furthermore the start and the end position (lines 8–11). The
graphical primitives need the information about their position and size, a rotation
value (represented by a Quaternion object) and the type of material, which together
determine their visual representation. The positions of the components of the generic
depictions is not extracted one-to-one from the 3D canvas but normalized to a neutral
point. This neutral point is specified as the left most point for each dimension where
a primitive, container, or stretch interval is located. The container inside the blue
box is two pixels further inside, which can be seen on the position of the container
and box in lines 5 and 13.
The code generator of the generic depictions editor were be specified on the basis
of the language’s abstract structure. The attributes store code fragments and analysis
results that are stepwise combined in order to construct the final target code. For
such purpose libraries and domain-specific languages of the well-known Eli system
[KPJ98] are used.
4
Range of application
In Figure 2 the reader has already seen exemplary specifications for generic depictions
for molecular models. Such depictions—basically representing atoms and bonds—play
the role of building blocks from which molecules are be constructed. The construction
of molecules is supported by a dedicated editor, which is generated by DEViL3D. To
demonstrate that our approach of specifying generic depictions is feasible for a wide
range of three-dimensional languages, we present further four 3D languages and focus
on the specification of generic depictions for these languages.
Figure 8 shows a specification of generic depictions for the SAM (Solid Agents
in Motion) language on the left-hand side and its application in a generated editor
on the right-hand side. The SAM language [GMR98] is a parallel, synchronous and
state-oriented 3D language that is based on the well known 2D language Pictorical
Janus. A SAM program describes a set of agents that synchronously communicate by
exchanging messages. We have specified a set of generic depictions of language constructs occurring in SAM programs. The agent is depicted by a large transparent blue
sphere, which provides a container to embed rules. Furthermore, input and output
12
ports are visualized as cones of different colors. They will be positioned in the instantiated program (right-hand side in the figure) on the surface of the agent’s sphere.
Messages—represented by a transparent box with two cones at its outside—are connected to the agent, too. The generic depiction for messages provides a container
inside the box that can embed values of different types (for example a string value
represented by a sphere) in the generated editor.
Figure 8: Generic depictions for the SAM language.
Figure 9: Visual representation of Petri Nets specified by generic depictions.
As mentioned in the introduction, Petri Nets can benefit from a three-dimensional
representation. Rölke [R0̈7] presented in his paper only the idea of 3D Petri Nets
without indicating a 3D editor that allows to construct effectively in 3D. By using
DEViL3D, we have specified such an editor. To realize the visual representations for
Petri Nets, five generic depictions are needed. One depiction consists of a container
that embeds the whole Petri Net and adjusts its size according to the stretch algorithm
when new objects are inserted. Then there are generic depictions for transitions
(visualized as boxes), places, tokens (both visualized as spheres), and arrows. The
depiction for places consists of a container that is located inside the sphere and can
accommodate tokens. Using the generated Petri Net editor (see right-hand side in
Figure 9), we have constructed a Petri Net, which has been taken from Rölke’s paper.
The Petri Net is partitioned in two planes: On the lower plane the control flow of
13
the Petri Net is located and the place on the upper plane is responsible for the dataflow. The first transition writes data to this place that other transitions read. Up to
now, the structure editor for 3D Petri Nets allows the creation of static Petri Nets. To
simulate the execution of Petri Nets, DEViL3D would have be extended by simulation
support, as the predecessor framework DEViL [CK09].
Figure 10: Generic depictions for an exemplary real-world editor.
Figure 10 demonstrate the specification of 3D languages, which based on real-world
objects. Such objects are characterized by complex shapes, which can be modeled
with tools like Blender [ble] or downloaded from libraries in the Internet containing
predefined models. Such models are referenced in generic depictions. The left-hand
side in Figure 10 represents generic depictions for street vehicles as bus, car, and
motor cycle. The fourth depiction shows a terrain that can be modeled with tools
provided by the jMonkeyEngine [jme]. In the generated editor the terrain models a
base ground and the vehicles can be positioned on top of it. The positioning of the
vehicles is determined by a visual pattern that arranges a set of objects on a common
plane shaped surface.
Figure 11: Music in Space editor.
Inspired by the web-based 3D editor ToneCraft, we have specified a Music in
Space editor with DEViL3D. The generated editor (see Figure 11) lets the user insert
distinctively colored boxes representing music instruments in a matrix like area. From
14
such a 3D program a music string according to the JFugue music programming API
[Koe] can be generated and played. The generic depictions for this editor are very
simple: For each instrument a depiction representing a colored box is needed. Another
depiction, which provides a container that embeds the whole program, which may be
constructed according to the matrix pattern.
5
Related Work
The basic idea of containers and stretch intervals—as used in our three-dimensional
generic depictions—goes back to the VPE system [Gra98] that generates 2D visual
language editors. Furthermore, the idea is successfully used in the generator system
DEViL [SKC06], which generates 2D language editors, too.
The idea for three-dimensional languages goes back to a publication of Glinert
[Gli87]. Najork [Naj96] developed the first 3D language Cube, which is semantically
similar to Prolog and makes use of the data-flow paradigm. In this language all
language constructs are represented by a cube. The hierarchical nesting of constructs
is an inherent concept of the Cube language. In the context of generator systems for
visual languages, Minas has supervised an exploration [Voß09] of 3D languages in the
context of his DiaGen/DiaMeta [Min02, Min06] frameworks.
The work of Chung et al. [CHM99] exactly addresses the topic of our paper.
They have developed a tool called 3DComposer that is related to our editor for
generic depictions. 3DComposer is a tool for specifying so called 3Dvixels, which can
be used as building blocks for 3D applications. Such 3Dvixels can be generally used
in 3D applications, which include 3D languages and also visualization tools. The
usage in different applications is possible by generating reusable software components
in form of JavaBeans. The construction of exemplary 3D programs is being done
directly in 3DComposer by the end user. 3DComposer is not part of a generator
framework, which would distinguish between language designer and language user.
Hence, 3DComposer does not need concepts such as containers or stretch intervals.
The AgentCubes [IRW09] system uses a mechanism to construct language object
representations. For such purpose a so called Inflatable Icons Editor is provided. It
allows users to quickly draft 3D objects by drawing 2D images and turn them step
by step into 3D models by using a diffusion-based inflation technique. In 3D games,
which are specified with AgentCubes, these models are mostly located on the ground
plane. Hence, the flat bottom side resulting from the inflation approach is not a
concern.
6
Conclusion
In this paper we described the process of specifying generic depictions for 3D visual
languages with the generator system DEViL3D. For such purpose DEViL3D provides
the editor, which allows the language designer to specify generic depictions. This
editor was also generated with DEViL3D in a boostrapping approach. Hence, the
interaction and navigation tasks, we have reported about, are available in all editors
generated with DEViL3D. For specifying generic depictions, the possibility to define
containers that can embed nested constructs is particularly important. We have
presented an algorithm that stretches the containers when their nested elements need
more room. The language designer does not need to care about many techniques—as
15
the stretch algorithm or the adaptation of interaction techniques to the requirements
of a particular representation—because they are automatically provided by DEViL3D.
The generic depictions editor is able to specify depictions for a wide range of 3D
languages covering languages as Petri Nets or molecular models with rather simple
visual representations but also languages, which consist of real-world objects that
have more advanced visual representations.
Acknowledgments
The author gratefully acknowledges the support of the German Research Foundation
(Deutsche Forschungsgemeinschaft – DFG); contract no. KA 537/6-1. Furthermore,
the author would like to thank Marius Dransfeld for the implementation of the first
prototype of the generic depictions editor as well as Elena Rybka and Johann Rybka
who developed interaction techniques mentioned in this paper.
References
[Aut]
Autodesk. 3ds Max. http://www.autodesk.com/3dsmax. [Online; accessed
28-February-2013].
[ble]
Blender – 3D creation for everyone. http://www.blender.org/. [Online;
accessed 23-February-2013].
[CHM99] Vincent Chung, John Hosking, and Rick Mugridge. Visual Specification of
3D Notations using 3DComposer. In Proceedings of the IEEE Symposium
on Visual Languages, pages 198–199, September 1999.
[CK09]
Bastian Cramer and Uwe Kastens. Animation automatically generated
from simulation specifications. In Proceedings of the IEEE Symposium on
Visual Languages and Human-Centric Computing (VL/HCC), pages 157–
164, September 2009.
[Din11]
Dinahmoe. ToneCraft, 2011. http://labs.dinahmoe.com/ToneCraft/.
[Online; accessed 5-February-2013].
[Gli87]
Ephraim P. Glinert. Out of Flatland: Towards 3-D Visual Programming. In
Proceedings of the Fall Joint Computer Conference on Exploring technology:
today and tomorrow, pages 292–299, 1987.
[GMR98] Christian Geiger, Wolfgang Müller, and Waldemar Rosenbach. SAM –
An Animated 3D Programming Language. In Proceedings of the IEEE
Symposium on Visual Languages, pages 228–235, September 1998.
[GP96]
T. R. G. Green and Marian Petre. Usability Analysis of Visual Programming Environments: A ’Cognitive Dimensions’ Framework. Journal of Visual Languages and Computing, 7(2):131–174, 1996.
[Gra98]
Calum A.M. Grant. Visual Language Editing Using a Grammar-Based
Visual Structure Editor. Journal of Visual Languages and Computing,
9(4):351–374, 1998.
16
[GRR99] Martin Gogolla, Oliver Radfelder, and Mark Richters. Towards ThreeDimensional Representation and Animation of UML Diagrams. In Proceedings of the International Conference on The Unified Modeling Language:
Beyond the Standard, pages 489–502, September 1999.
[IRW09] Andri Ioannidou, Alexander Repenning, and David C. Webb. AgentCubes:
Incremental 3D end-user development. Journal of Visual Languages and
Computing, 20(4):236–251, 2009.
[jme]
jMonkeyEngine 3 – Documentation. http://jmonkeyengine.org/wiki/
doku.php/jme3. [Online; accessed 3-April-2012].
[Koe]
David Koelle. JFugue – Java API for Music Programming. http://www.
jfugue.org/. [Online; accessed 12-March-2013].
[KP07]
Caitlin Kelleher and Randy Pausch. Using storytelling to motivate programming. Communications of the ACM, 50(7):58–64, July 2007.
[KPJ98] Uwe Kastens, Peter Pfahler, and Matthias Jung. The Eli System. In
Compiler Construction, volume 1383 of Lecture Notes in Computer Science,
pages 294–297. Springer-Verlag Berlin Heidelberg, 1998.
[lab]
National Instruments LabVIEW. http://www.ni.com/labview. [Online;
accessed 23-April-2012].
[Min02]
Mark Minas. Concepts and realization of a diagram editor generator
based on hypergraph transformation. Science of Computer Programming,
44(2):157–180, 2002.
[Min06]
Mark Minas. Generating Meta-Model-Based Freehand Editors. In Albert
Zündorf and Dániel Varró, editors, Proceedings of the 3rd International
Workshop on Graph Based Tools (GraBaTs’06), Natal (Brazil), Satellite
event of the 3rd International Conference on Graph Transformation, volume 1 of Electronic Communications of the EASST, September 2006.
[Naj96]
Marc A. Najork. Programming in Three Dimensions. Journal of Visual
Languages and Computing, 7(2):219–242, 1996.
[R0̈7]
Heike Rölke. 3-D Petri nets – Making use of 3 Dimensions in Executable
Petri Net Modelling. Petri Net Newsletter, 72:3–9, 2007.
[SCK07] Carsten Schmidt, Bastian Cramer, and Uwe Kastens. Usability Evaluation
of a System for Implementation of Visual Languages. In Proceedings of
the IEEE Symposium on Visual Languages and Human-Centric Computing
(VL/HCC), pages 231–238, September 2007.
[Shn83]
Ben Shneiderman. Direct Manipulation: A Step Beyond Programming
Languages. IEEE Computer, 16(8):57–69, August 1983.
[SK03]
Carsten Schmidt and Uwe Kastens. Implementation of visual languages
using pattern-based specifications. Software – Practice and Experience,
33(15):1471–1505, 2003.
17
[SKC06] Carsten Schmidt, Uwe Kastens, and Bastian Cramer. Using DEViL for
Implementation of Domain-Specific Visual Languages. In Proceedings of
the Workshop on Domain-Specific Program Development, July 2006.
[TDL04] Monica Tavanti, Nguyen-Thong Dang, and Hong-Ha Le. Usability Inspection of a 3D Interaction Metaphor. In Proceedings of the Conference Internationale Associant Chercheurs Vietnamiens et Francophones en Informatique, pages 41–46, February 2004.
[Voß09]
Volker Voß. Dreidimensionale Darstellung zweidimensionaler visueller
Sprachen, June 2009. Diplomarbeit, Universität der Bunderwehr München.
[Wol12]
Jan Wolter. DEViL3D – A Generator Framework for Three-Dimensional
Visual Languages. In Proceedings of the International Workshop on Visual Languages and Computing (VLC) in conjunction with the 18th International Conference on Distributed Multimedia Systems (DMS), pages
171–176, August 2012.
18
| 6cs.PL
|
What Is The Collision Probability And How To Compute It
arXiv:1711.07060v2 [cs.SY] 14 Jan 2018
Richard Altendorfer and Christoph Wilkmann
Abstract—We revisit the computation of probability of collision
in the context of automotive collision avoidance (the estimation
of a potential collision is also referred to as conflict detection
in other contexts). After reviewing existing approaches to the
definition and computation of collision probability we argue that
the fundamental quantity that is required to answer a question
like ”What is the probability of collision within the next three
seconds?” is a collision probability rate. We derive a general
expression for the distribution of the collision probability rate and
demonstrate that it exactly reproduces distributions obtained by
large-scale Monte-Carlo simulations. This expression is valid for
arbitrary prediction models including process noise. We derive
an approximation for the distribution of the collision probability
rate that can be computed on an embedded platform. In order to
efficiently sample this probability rate distribution for determination of its characteristic shape an adaptive method to obtain the
sampling points is proposed. The probability of collision is then
obtained by one-dimensional numerical integration over the time
period of interest. We also argue that temporal collision measures
such as time-to-collision should not be calculated as separate or
even prerequisite quantities but that they are properties of the
distribution of the collision probability rate.
we derive a general expression valid for arbitrary prediction
models including process noise. Inclusion of process noise
is crucial for collision avoidance systems since it allows to
encode the uncertainty in the relative motion of the host and
the colliding vehicle. This uncertainty is particularly relevant
for predictions over several seconds where it is unknown
whether the colliding vehicle keeps its motion, accelerates or
slows down, or whether the host vehicle driver perceives the
risk and slows down, for example.
The basis of our derivations are the distributions
pt (x, y, ẋ, ẏ, ẍ, ÿ), t ∈ ∆T of the predicted relative state ξ − (t)
of the colliding vehicle. The predicted state ξ − (t) is a nonstationary stochastic process. In the remainder of this paper the
time dependence of ξ − (t) and its elements will be suppressed,
however the temporal dependence of probability distributions
will be indicated by p → pt where appropriate.
In the next section Monte-Carlo simulations of two collision
scenarios are performed and it is shown that the result is
naturally represented by a collision probability rate.
I. INTRODUCTION
II. T HE COLLISION PROBABILITY GROUND TRUTH :
LARGE - SCALE M ONTE -C ARLO SIMULATIONS
The implementation of a collision mitigation or collision
avoidance system requires in general the computation of a
collision probability in order to assess the criticality of the
current traffic situation as well as its evolution in the shortterm future. Sometimes, in a separate calculation, see e. g.
[1],[2] an estimate of the time-to-collision (TTC) is computed.
There are two fundamentally different approaches to computing a collision probability that are known to the authors:
1) probability of the spatial overlap of the host vehicle with
the colliding vehicle’s probability distribution, see e. g.
[3], [4], and
2) probability of penetrating a boundary around the host
vehicle, see e. g. [5],[6].
There is currently no satisfying way to compute the collision
probability over a time period: there is a heuristic proposal to
pick the maximal collision probability over that period as the
collision probability for that time period [5],[1], and there is
a rather involved calculation valid only for piecewise linear
trajectories penetrating a line interval to directly compute the
collision probability over a time period while ignoring process
noise [6].
In the following we will derive an expression of the
probability of penetrating a boundary around the host vehicle
in a time period ∆T = [t1 , t2 ]. This will be the result of
the temporal integration of the probability rate for which
R. Altendorfer is with Driver Assistance Systems, ZF TRW, Germany
[email protected]
C.
Wilkmann
is
with
RWTH
[email protected]
Aachen,
Germany
In order to obtain ground truth data for the collision
probability Monte-Carlo simulations are performed. The target
vehicle on a possibly colliding path with the host vehicle is
modeled by the state vector ξ = (x y ẋ ẏ ẍ ÿ)> and the
dynamical system as specified in appendix B. Accelerations
are deliberately included in the state in order to demonstrate
that accelerated motions can be handled. The target vehicle
is chosen to be detected by a radar sensor mounted at the
middle of the front bumper of the host vehicle with standard
radar measurements also specified in app. B.
The starting point for an individual simulation is a sample
point in state space ξi− where the target vehicle is some
distance away from the host vehicle - either directly in front
or coming from the right side, see fig. 1. This sample point is
drawn from a multivariate Gaussian distribution characterized
by its mean vector and covariance matrix. The covariance
matrix is taken from steady state1 at this mean vector using
the discrete algebraic Riccati equation (25) as detailed in app.
B. An instance ξi− of an initial state of the target vehicle is
−
drawn as a sample of N (ξ − ; µ−
ξ , P∞ ). This state is predicted
using the stochastic state equation (22) until it crosses the host
vehicle boundary or a certain time limit is exceeded. The time
until the crossing is recorded and a new simulation with a new
sample of initial conditions is started. Examples of colliding
1 Strictly speaking there is no steady state at those points since the system
is non-linear and the relative speed is not zero. Nevertheless the solution of
the Riccati equation is still representative if the filter settles within a smaller
time period than the time period in which the state changes significantly.
trajectories starting from an initial position in front of the host
vehicle are depicted in fig. 1.
We have performed simulations of Ntraj = 1 · 107 trajectories for the two starting points. The result is represented by
a histogram of the number of collisions that occur within a
histogram bin, i. e. time interval, with respect to time.
Hence simulating colliding trajectories naturally leads to a
collision probability rate.
An example is given in fig. 2 where the bins are normalized
by the total number of trajectories Ntraj and the chosen
bin width of dt = 0.05s to obtain a collision probability
rate. In addition, the collision probability rate integrated by
simple midpoint quadrature from 0 to time t is shown. In this
example the probability of collision with the target vehicle
reaches a threshold of 40%, for example, within the first 4.88s.
The asymptotic value of the collision probability as t → ∞
indicates the overall probability of collision over all times.
The main contribution of this paper will be to derive
formulae that reproduce this collision probability and collision
probability rate.
In the following two sections we will review existing
approaches to computing a collision probability.
III. C OLLISION PROBABILITY FROM 2D SPATIAL OVERLAP
This is the probability of the spatial overlap between the
host vehicle and the colliding vehicle as proposed in [3], [4].
In [3] the variables of the relevant probability distribution
are either defined by the relative two-dimensional position
and relative orientation (x, y, ψ), or by the distribution of the
difference of independent probability distributions of global
two-dimensional position and orientation as in [4]. It is not explained how higher order derivatives necessary for prediction
such as ẋ, ẏ, ... have been dealt with (e. g. by marginalization).
Then in [3] the collision probability is obtained as the
integral over the pdf of relative position and orientation over
a collision volume D
ZZZ
PC (t) =
p (x, y, ψ) dxdydψ
(1)
x,y,ψ∈D
which in turn is approximated as the convolution of independent probability distributions of global two-dimensional
position of both host and colliding vehicle - orientation is
ignored. Note that even this two-dimensional integral, i. e. the
cumulative distribution function of a bivariate Gaussian, cannot be solved in closed form; however, numerical approximation schemes exist [7]. By further ignoring the x−y-covariance
the multivariate Gaussian decouples so that integration can be
factored into one-dimensional Gaussian distributions [3]. In
[4], the collision probability is computed as the integral of the
product of the two global distributions.
A problem of deriving a collision probability from 2D
spatial overlap is that this approach directly yields a collision
probability for a specific time, see fig. 3. Hence it does not
allow to answer the question ”What is the probability of
collision within the next three seconds?” because integration
of the collision probability over time does not yield a collision
probability over a certain time period as already pointed out
in [6]. In particular, time is not a random variable that can be
marginalized
over and the integral over a time interval ∆T :
R
P (t)dt as in [8] has dimension of time and does not
∆T C
constitute a probability. A rather heuristic proposal to solve
this problem has been to pick the maximal collision probability
over a time period as the collision probability for that period
[5],[1].
Another issue is that the collision probability is determined
by the presence of the colliding object anywhere inside the
entire area of the host vehicle. It could, for example be in
the center of the host vehicle, which would be a catastrophic
event and much too late for collision avoidance. What we are
actually interested in is the probability of the colliding object
touching and/or penetrating the boundary of the host vehicle.
This requires a different approach than integration over state
space as in eq. (1). Some existing approaches that consider a
boundary instead of a state space volume for the computation
of a collision probability are reviewed in the next section.
IV. C OLLISION PROBABILITY AT BOUNDARY
A probabilistic approach to computing the probability of
penetrating a boundary - instead of the probability of a spatial
overlap - has been proposed in [6]. Their method is based
on the probability density of a so-called time-to-go which is
the time to cross a straight, axis-aligned boundary assuming
a constant velocity model. The derived collision probability
refers to a time period and not just a time instant. It is
only applicable to straight paths or combinations of piecewise
straight paths and does not take into account more complex
geometries such as a rectangle. Another limitation is that
process noise is not considered. This method is an extension
of the approach discussed [5] in the sense that the uncertainty
of the initial condition is included.
A different, semi-probabilistic approach is taken in [2]
where the time-to-collision of a 2D constant velocity model
with respect to the host vehicle’s front boundary at x = 0
is computed. This TTC is then inserted into the prediction
equation to arrive at
x + tẋ
0
x
y + tẏ
→ y − ẋ ẏ
(2)
ẋ
ẋ
ẏ
ẏ
Then the second component of this vector is singled out and
interpreted as a probabilistic expression. Its distribution as
a function of the initial condition is determined by MonteCarlo simulation and integrated over the host vehicle’s front
boundary to obtain a collision probability at the TTC. As in
the previous approaches more complex geometries as well as
process noise are not considered.
The approaches discussed above all rely on a time-togo or TTC as a prerequisite quantity - either probabilistic
or non-probabilistic. As we will show in the next section,
such a temporal collision measure is not necessary for the
computation of a collision probability. Instead, we show that
the fundamental quantity to compute the collision probability
is the collision probability rate.
10
10
shape of initial
x-y covariance
sample collision
trajectories
x [m]
5
x [m]
5
shape of initial
x-y covariance
sample collision
trajectories
0
0
-5
-5
-10
-5
0
5
10
-10
-5
0
y [m]
5
10
y [m]
(a) The target is coming from the front (µx , µy ) = (10, 0) m
(b) The target is coming from the front right (µx , µy ) = (10, 10) m
0.09
0.5
collision probability rate
collision probability
0.45
P(t) from 2D overlap
TTC x
0.08
0.4
0.07
0.35
0.06
0.3
P(t) []
collision probability rate [s -1 ], collision probability []
Fig. 1. Samples of simulated colliding trajectories for vehicles initially coming from the front (a) and from the front right (b) side. The number of drawn
trajectories colliding with the edges of the host vehicle indicate for the ratio of colliding the respective edge compared to the total number of collisions.
0.25
0.05
0.04
0.2
0.03
0.15
0.1
0.02
0.05
0.01
0
0
1
2
3
4
5
6
7
8
0
2
3
4
t [s]
Fig. 2. Collision probability rate as a function of time for (µx , µy ) =
(10, 0) m based upon Ntraj = 1·107 trajectories. Also shown is the collision
probability obtained by integrating over time.
V. C OLLISION PROBABILITY RATE AT BOUNDARY
5
6
We have seen that simulating colliding trajectories naturally
gives us a probability rate and that a collision probability
rate allows us to perform temporal integration to arrive at
a collision probability for an extended period of time. The
expression for the collision probability rate will be derived on
the basis of a theorem on boundary crossings of stochastic
vector processes. For sake of lucidity of arguments we restrict ourselves to one of the four straight boundaries of the
host vehicle, see fig. 4; extension to the other boundaries is
straightforward.
We start with the prediction of the pdf of the state vector
(see eq. (22) in app. B) of a colliding object from an initial
condition at t = 0 to a future time t where process noise
8
Fig. 3. Example of collision probability over time.
(parametrized by its covariance matrix Q) is explicitly incorporated:
t,Q
A. Derivation of collision probability rate
7
t [s]
prediction : p0 (x, y, ẋ, ẏ, ẍ, ÿ) 7−→ pt (x, y, ẋ, ẏ, ẍ, ÿ)
(3)
In order to cast the following expressions into a more readable
format we define a probability distribution with fewer variables
by marginalization (see app. A for marginalization of Gaussian
densities) over the accelerations:
Z Z
pt (x, y, ẋ, ẏ) :=
pt (x, y, ẋ, ẏ, ẍ, ÿ)dẍdÿ.
(4)
ẍ∈R ÿ∈R
What we are looking for is an expression for
dPC+
(Γf ront , t)
dt
dP +
(5)
i. e. the collision probability rate dtC with dimension [s−1 ]
at time t for the front boundary Γf ront . The superscript + is
used to denote that this probability rate is referring to boundary
crossings from outside to inside.
We start with the probability of the colliding object being
inside an infinitesimally thin strip at the boundary Γf ront (see
fig. 4)
Z Z Z
dPC+ (Γf ront , t) =
pt (x0 , y, ẋ, ẏ) dxdydẋdẏ
y∈Iy ẋ≤0 ẏ∈R
Here, since we are only interested in colliding trajectories, i.
e. trajectories that cross the boundary from outside to inside,
we do not fully marginalize over ẋ but restrict the x−velocity
to negative values at the boundary.
A collision probability rate can now be obtained by dividing
the unintegrated differential dx by dt; in that way the “flow”
of the target vehicle through the host vehicle boundary is
described at x0 with velocity ẋ ≤ 0:
Z Z Z
dPC+
(Γf ront , t) = −
pt (x0 , y, ẋ, ẏ) ẋ dydẋdẏ (6)
dt
y∈Iy ẋ≤0 ẏ∈R
Here, since the velocity is restricted to negative values a minus
sign is required to obtain a positive rate.
This intuitive derivation can be corroborated as well as
generalized in a mathematically rigorous way by invoking a
result on crossings of a surface element by a stochastic vector
process described in the following box:
x
Front
boundary: front
yL
yR
dx
Infinitesimally
thin strip
x0
Iy
left
y
right
Host vehicle
rear
Fig. 4. Horizontal view of the host vehicle rectangle with local Cartesian
coordinate system and coordinate origin at the middle of the front boundary
characterized by x = 0 and y ∈ [yL , yR ] = Iy .
In order to apply eq. (7) to the front boundary Γf ront as in
fig. 4 we need to perform the following identifications:2
ζ(t) = (x, y)
>
Γf ront = {(x, y) : x − x0 = 0 ∧ y ∈ Iy }
g(x) = x − x0
>
nΓf ront (x) = (−1, 0)
ds(x) = dy
(9)
Probability rate and probability for boundary crossings
Hence we obtain for the intermediate expectation operator
of stochastic vector processes
Here two expressions are given for the probability rate and
n
o
the probability for crossings of a surface by a stochastic
E hnΓf ront (x), ζ̇(t)i+ ζ(t) = x =
vector process that were stated in [9] and generalized in
Z Z
[10].
−
ẋ pt (ẋ, ẏ|x, y) dẋdẏ (10)
Let ζ(t) be a continuously differentiable n−dimensional
ẋ≤0 ẏ∈R
vector stochastic process with values x ∈ Rn . The probn
ability densities pt (x) and pt (ẋ, x) exist where ẋ ∈ R
dP +
are the values of ζ̇(t). Let the region S ∈ Rn be bounded
and the collision entry rate dtC as in eq. (7) becomes
by the smooth surface ∂S defined by the smooth function
g as ∂S = {x : g(x) = 0} and let Γ ⊆ ∂S be a subset
Z Z Z
+
dPC
of that surface. Let nΓ (x) be the surface normal at x
(Γf ront , t) = −
ẋ pt (ẋ, ẏ|x0 , y)dẋdẏpt (x0 , y)dy
dt
directed towards the interior of the region.
+
−
y∈Iy ẋ≤0 ẏ∈R
Then the entry rate dPdt and the exit rate dPdt through Γ
Z Z Z
at time t is given by
=−
ẋ pt (x0 , y, ẋ, ẏ) dydẋdẏ
(11)
Z
n
o
±
dP
y∈Iy ẋ≤0 ẏ∈R
(Γ, t) = E hnΓ (x), ζ̇(t)i± ζ(t) = x pt (x)ds(x)
dt
x∈Γ
which is identical to eq. (6). Because of the simple projection
(7)
of the velocity onto the surface normal hnΓf ront (x), ζ̇(t)i+ =
where h·, ·i is the scalar product, ds(x) is an infinitesimal
ẋ for the front boundary this expression could be further
surface element of Γ at x and (·)+ := max(·, 0) and
simplified by marginalization over ẏ.
−
(·) := − min(·, 0).
The computation above applies to the front boundary of
The probability of entries and exits through the surface
the
host vehicle. In order to cover all four boundaries of the
subset Γ within a time interval ∆T = [t1 , t2 ] is given by
host vehicle the collision probabilities of the four boundaries
Zt2
can be added since the crossing events through the different
dP ±
±
P (Γ, t1 , t2 ) =
(Γ, t) dt
(8)
dt
t1
These results hold for general non-stationary processes.
2 Unlike in the box above we now do not distinguish anymore between a
stochastic process and its values.
boundaries are mutually exclusive. Hence the total collision
probability rate is given by
dPC+
(Γhost
dt
vehicle , t)
=
dPC+
dPC+
(Γf ront , t) +
(Γright , t)+
dt
dt
dPC+
dPC+
(Γlef t , t) +
(Γrear , t) (12)
dt
dt
The collision probability over a certain time period can then
be obtained by one-dimensional numerical integration over
time as in eq. (8).
The derivation above relied on a couple of assumptions.
Here we want to comment on whether these assumptions pose
any restrictions on the application of the derived collision
probability rate on more general scenarios.
a) Rectangular boundary: The boundary of the host
vehicle is modeled as a rectangle, however by the theorem
stated in the box above the formula can be applied to any
subsets of smooth surfaces, including higher dimensional ones
for three-dimensional objects, for example.
b) Point distribution: The relative state of the colliding
object is modeled by a point distribution. In many ADAS
applications the object state is indeed modeled by a single
reference point and possibly additional attributes such as width
and length. This is partly due to the fact that the estimation
of extended objects is still an active area of research, for an
overview see [11]. On the other hand, our approach can be
executed in parallel for distributions of points representative
of the colliding vehicle’s geometry [12] or for individual point
distributions of a Gaussian mixture model, for example.
c) State vector: the state vector ξ of the colliding object
needs to contain 2D relative position (x y)> and 2D relative
velocity (ẋ ẏ)> . In many ADAS applications the target vehicle
dynamics is modeled directly in relative coordinates. For state
vectors that do not contain the 2D relative velocity but other
quantities such as the velocity over ground (see e. g. [13]),
a probabilistic transformation to relative velocities must be
performed first.
B. Implementation
For further computations - especially in the Gaussian case
- it will be convenient to marginalize over ẏ and rewrite eq.
(11) in terms of a conditional probability:
Z Z
dPC+
(Γf ront , t) = −pt (x0 )
ẋ pt (ẋ, y|x0 ) dẋdy (13)
dt
ẋ≤0 y∈Iy
For general distribution functions the integral in eq. (13)
cannot be computed in closed form and numerical integration
methods must be used. Even in the bivariate Gaussian case
there is no explicit solution known to the authors. However,
by a Taylor-expansion with respect to the off-diagonal element
of the inverse covariance matrix of p(y, ẋ|x0 ) as detailed in
app. C, the integral can be factorized into one-dimensional
Gaussians and solved in terms of the standard normal onedimensional cumulative distribution function Φ. To zeroth
order the integration yields:
dPC+
(Γf ront , t) = − N (x0 ; µx , σx )
dt
−µẋ|x0
·
µẋ|x0 Φ
σ̃ẋ|x0
2
N (0; µẋ|x0 , σ̃ẋ|x0 )
σ̃ẋ|x
0
·
−
yR − µy|x0
yL − µy|x0
· Φ
−Φ
σ̃y|x0
σ̃y|x0
!
+ O Σ−1
(14)
12
Here, if Σ ∈ q
R2×2 is the covariance
q matrix of p(ẋ, y|x0 ),
|Σ|
|Σ|
then σ̃ẋ|x0 =
Σyy and σ̃y|x0 =
Σẋẋ , see app. C where
the integration has also been carried out to first order in Σ−1
12 .
Expression (14) can be computed on an embedded platform
using the complementary error function available in the C math
library.3
In the next section an extensive numerical study using the
above formulae and Monte-Carlo simulations is presented.
VI. N UMERICAL S TUDY
A numerical study has been carried out to address the
following questions:
• Is the expression for calculating the collision probability
rate from eq. (11) consistent with the results from large
scale Monte-Carlo simulations?
• How does the approximation (14) perform in comparison
with the numerical integration of the derived expression
for the collision probability rate?
• Can the computational effort be reduced by increasing ∆t
and still accurately calculating the collision probability
rate?
A. Is the expression of collision probability rate corroborated
by Monte-Carlo simulation?
In order to address the first question, large scale MonteCarlo simulations as described in sec. II have been performed.
Collision probability rates were calculated based on 1 · 107
sample trajectories for each of the two initial conditions
−
−
−
Ni (ξ − ; µ−
ξ i , P∞ ), where µξ i is shown in table I, and P∞ is
calculated using eq. (25) with the matrices defined in appendix
B. The two initial conditions (i ∈ {f, f r}) describe a starting
point directly in front of the host vehicle, and in front to the
right at an angle of 45 degrees with respect to the host vehicle.
Table II shows the number of collisions divided into the
respective boundaries of the host vehicle where the impact or
boundary crossing occurred for the two different simulations.
The resulting histograms of the collision probability rates
are shown in fig. 5 together with the numerical integration of
3 Φ is related to the error function erf and complementary error function
√
√
erfc by Φ(x) = 21 erfc −x
= 12 − 12 erf −x
.
2
2
TABLE I
M EAN OF INITIAL CONDITIONS FOR M ONTE -C ARLO SIMULATIONS
i
HH f
µ−
ξ
µ−
10
x [m]
µ−
0
y [m]
m
µ−
−2
ẋ s
m
µ−
0.1
ẏ h s i
− m
µẍ s2
−0.2
h i
m
µ−
0.0
ÿ
s2
HH
fr
10
10
−2
−1.6
−0.001
−0.01
TABLE II
N UMBER OF COLLISIONS AT HOST VEHICLE BOUNDARIES FOR 1 · 107
SIMULATED TRAJECTORIES WITH DIFFERENT INITIAL CONDITIONS
PP
PP i
Γj
P
P
f ront
right
lef t
rear
P
f
fr
106
4.46 ·
36874
6196
0
4.50 · 106
1.40 · 106
2.88 · 106
2
0
4.28 · 106
the collision probability rate as well as the difference between
the simulation and the calculation. The difference is calculated
by evaluating the collision probability rate at the same time
as the mid points of the histogram bins.
As can be seen in fig. 5, the numerical integration of the
collision probability rate accurately reproduces the collision
probability rate from the Monte-Carlo simulations. In order to
illustrate the increase in accuracy as a function of the number
of simulated trajectories, fig. 6 shows the differences between
simulation and numerical integration with increasing amount
of simulated trajectories for collisions at the right side of the
host vehicle in the front scenario.
B. Does the approximation by Taylor-expansion accurately
reproduce the exact result?
In order to be able to compute the collision probability rate
efficiently on an embedded platform, an approximation of the
exact expression (eq. (11)) was derived in eq. (14). Fig. 7
shows the differences between this approximation as well as a
higher-order approximation where the pdf is Taylor-expanded
to linear order with respect to the off-diagonal element of
the inverse covariance matrix around 0 (see app. C) and
the numerical integration. As can be seen, the higher-order
approximation reduces the error to a large extent while it can
be still calculated efficiently on an embedded platform using
the complementary error function.
C. An adaptive method to sample the collision probability rate
over ∆T
Above, the approximations of the exact expression for
calculating the collision probability rate were evaluated at
small time increments of ∆t = 0.05s. Thus, the calculation
over the entire time period of interest (e.g 8s as used above)
and for every relevant object could induce a substantial computational burden. In order to reduce this effort, we propose
an adaptive method to sample the collision probability rate
function with variable – in general larger – time increments
∆t over the time period of interest while still capturing the
characteristics of this function, in particular its shape around
the maximum. The sampling starting point is based upon the
non-probabilistic TTCs for single, straight boundaries using
a one-dimensional constant acceleration model. Those TTCs
for penetrating the front, left, and right boundaries can then be
used as initial condition for the start of the sampling iteration
of the collision probability rate.4 To reproduce the collision
probability rate without substantial loss of information but
with lower computational effort, the following algorithm is
proposed:
• Calculate the times of penetrating the front, left and
right boundaries based upon the non-probabilistic TTCs
described above.
• Calculate the collision probability rate for each time. Pick
the time with the maximum collision probability rate as
a starting point.
• Move left and right from this starting point with equally
spaced ∆t1 > ∆t and calculate the collision probability
rate at these time points. Stop on each side if the collision
dP +
probability rate has reached a lower threshold of dtC low .
• While moving left and right, check if the slope of the
collision probability rate has changed its sign.
• On every slope sign change, calculate the collision probability rate around this time interval with decreased
∆t2 < ∆t1 .
Examples of this implementation can be found in fig. 8 for the
front and front-right scenarios. It can be seen that the collision
probability rate as well as the integrated collision probability
over a certain time period can be determined with considerably
fewer sampling points while still capturing the shape of the
functions to be approximated.
VII. W HAT IS THE TTC?
There are various approaches to computing the TTC. In
[2], the TTC is computed as the mean of the time distribution
of reaching the x0 boundary of the car as a function of the
initial conditions assuming a constant speed model; process
noise is not considered. This is also presented in [1]; in
addition the time distribution for reaching the x0 boundary
as a function of the initial conditions assuming a constant
acceleration model is calculated by Monte-Carlo-simulation
and its mean values depending upon the initial condition setup
is given - again, process noise for this motion model is not
considered. However these simple temporal quantities do not
fully capture the characteristics of horizontal plane collision
scenarios as will be shown below.
In fig. 9 the collision probability rate is plotted together
with the constant acceleration motion TTCs for the x- and ydirections for an initial position at the front, right side. Both
4 Due to the low probability of penetration the non-probabilistic TTC for
the rear boundary is not considered for the determination of the sampling
starting point.
0.5
0.5
Monte-Carlo histogram
numerical integration of exact expression
0.45
0.4
collision probability rate [s -1 ]
0.4
collision probability rate [s -1 ]
Monte-Carlo histogram
numerical integration of exact expression
0.45
0.35
0.3
0.25
0.2
0.15
0.35
0.3
0.25
0.2
0.15
0.1
0.1
0.05
0.05
0
0
2
3
4
5
6
7
8
2
3
4
5
t [s]
(a) Front scenario total collision probability rate
8
10-3
2
numerical integration - Monte-Carlo histogram
numerical integration - Monte-Carlo histogram
1.5
collision probability rate [s -1 ]
1.5
collision probability rate [s -1 ]
7
(b) Front-Right scenario total collision probability rate
10-3
2
6
t [s]
1
0.5
0
-0.5
1
0.5
0
-0.5
-1
-1
-1.5
-1.5
2
3
4
5
6
7
8
2
3
4
5
t [s]
6
7
8
t [s]
(c) Front scenario total collision probability rate difference
(d) Front-Right scenario total collision probability rate
difference
Fig. 5. The histogram resulting from Monte-Carlo simulation is shown together with the numerical integration of the expression for the collision probability rate
for front (a) and front-right (b) scenario. The differences between simulation and numerical integration are calculated by evaluating the numerical integration
at the same time as the mid points of the histogram bins and shown in (c) and (d).
10-3
5
4.5
2
1.5
4
collision probability rate [s-1 ]
collision probability rate [s-1 ]
3
2.5
3.5
3
2.5
2
1.5
3.5
3
2.5
2
1.5
1
1
1
0.5
0.5
0.5
0
0
1
2
3
4
5
6
7
8
0
1
2
3
t [s]
(a) Simulation based on 1 · 105 trajectories.
Monte-Carlo histogram
numerical integration of exact expression
4.5
4
3.5
10-3
5
Monte-Carlo histogram
numerical integration of exact expression
4.5
4
collision probability rate [s-1 ]
10-3
5
Monte-Carlo histogram
numerical integration of exact expression
4
5
6
7
8
t [s]
(b) Simulation based on 1 · 106 trajectories.
1
2
3
4
5
6
7
8
t [s]
(c) Simulation based on 1 · 107 trajectories.
Fig. 6. The collision probability rate for the right side of the host vehicle for the front scenario is shown comparing the results from Monte-Carlo simulation
with increasing amount of simulated trajectories (a)-(c) with the calculated collision probability rate of the numerical integration of the expression for the
collision probability rate eq. (11). The process noise PSD for both coordinates is q̃x = q̃y = 1.0125m2 s−5 .
TTCs in x- and y-direction are significantly different from the
time of the maximum of the collision probability rate. Since
the bulk of the colliding trajectories go through two sides front and right (see also fig. 1b) - only a collision model that
takes into account the full geometry of the host vehicle can
yield accurate results.
In fig. 10 the collision probability rate is plotted together
with the constant acceleration motion TTCs (i. e. their mean
values) for the x-direction for an initial position that is straight
in front of the vehicle hence almost all trajectories pass
through the front boundary. Nevertheless the maximum of
the probability rate occurs before the mean of the TTC. This
difference increases as the process noise increases as can be
seen in fig. 11. This is due to the fact that the time of the
maximum is strongly influenced by the factor pt (x0 ) in eq.
(13); an increased level of process noise leads to a faster
10-7
approx. - num. integration
higher order approx. - num. integration
4
approx. - num. integration
higher order approx. - num. integration
10
2
collision probability rate [s -1 ]
collision probability rate [s -1 ]
10-4
12
0
-2
-4
-6
8
6
4
2
0
-8
-2
2
3
4
5
6
7
8
2
3
4
5
t [s]
6
7
8
t [s]
(a) Front scenario
(b) Front-Right scenario
Fig. 7. Differences between numerical integration and two approximations of the exact expression of the collision probability rate (eq. (11)). The first
approximation (solid line) ignores the non-diagonal entries of the covariance matrix while the second approximation (dashed line) Taylor-expands the pdf to
linear order with respect to the off-diagonal elements of the inverse covariance matrix around 0.
0.35
starting point
with increment of 0.5s
with increment of 0.2s around max
linear-order approximation
linear-order approximation with reduced samples
0.4
starting point
with increment of 0.5s
with increment of 0.2s around max
linear-order approximation
linear-order approximation with reduced samples
0.3
collision probability rate [s -1 ]
collision probability rate [s -1 ]
0.5
0.3
0.2
0.25
0.2
0.15
0.1
0.1
0.05
0
0
2
3
4
5
6
7
8
2
3
4
t [s]
5
6
7
8
t [s]
(a) Front scenario collision probability rate
(b) Front-Right scenario collision probability
0.6
0.5
Monte-Carlo simulation
linear-order approximation
linear-order approximation with reduced samples
0.5
Monte-Carlo simulation
linear-order approximation
linear-order approximation with reduced samples
0.45
collision probability []
collision probability []
0.4
0.4
0.3
0.2
0.35
0.3
0.25
0.2
0.15
0.1
0.1
0.05
0
0
2
3
4
5
6
7
8
t [s]
(c) Front scenario collision probability rate
2
3
4
5
6
7
8
t [s]
(d) Front-Right scenario collision probability
Fig. 8. Examples for reducing the number of calculations to determine the collision probability rate and the integrated collision probability. (a) and (c) show
the results for the front scenario and (b) and (d) for the front-right scenario. The used parameters in these examples are ∆t1 = 0.5s, ∆t2 = 0.2s and
+
dPC
dt low
= 0.01. In doing so the number of calculations for the collision probability rate could be reduced from 120 (using a fixed sampling increment of
∆t = 0.05s) to 13 for the front scenario and to 16 for the front-right scenario, respectively.
spreading of pt (x0 ) and hence the maximum is reached earlier.
The above discussion shows that temporal collision characteristics are encoded by the distribution of the collision
probability rate which incorporates the full geometry of the
host vehicle as well as process noise during prediction.
A quantity called TTC could then be defined as one of the
0.18
collision probability rate
TTC x
collision probability rate [s -1 ]
0.16
TTC y
0.14
0.12
0.1
0.14
collision probability rate
TTC x
0.12
collision probability rate [s -1 ]
characteristic properties of this distribution such as the mode
or the mean or the median or as a property of the integrated
collision probability rate, e. g. the time when the collision
probability exceeds a certain threshold.
0.1
0.08
0.06
0.04
0.02
0.08
0
2
0.06
3
4
5
6
7
8
t [s]
0.04
0.02
0
2
3
4
5
6
7
8
Fig. 11. Collision probability rate and mean of constant acceleration TTC for
an initial condition almost in front of the vehicle: (x, y) = (10, 0)m. The
process noise PSD for both coordinates has been increased to q̃x = q̃y =
0.1013m2 s−5 .
t [s]
Fig. 9. Collision probability rate and mean of constant acceleration TTC for
an initial condition at the front, right side of the vehicle: (x, y) = (10, 10)m.
The process noise PSD for both coordinates is q̃x = q̃y = 0.0405m2 s−5 .
0.4
collision probability rate
TTC
collision probability rate [s -1 ]
0.35
x
0.3
0.25
0.2
0.15
0.1
geometries such as a single line segment do not properly
characterize the true collision probability rate distributions.
The ground truth collision probability rate distribution has
been obtained by Monte-Carlo simulations and reproduced by
our derived collision probability rate expression. We have also
implemented an approximation of the collision probability rate
expression that can be computed in closed form on an embedded platform. This approximate formula provided collision
probability rate distributions that are almost indistinguishable
from distributions obtained by numerical integration for the
scenarios considered in this paper. In order to efficiently
sample this probability rate distribution for determination of its
characteristic shape we have worked out an adaptive method
to obtain the sampling points.
IX. ACKNOWLEDGEMENTS
Helpful clarifications by Prof. Georg Lindgren are gratefully
acknowledged.
0.05
0
2
3
4
5
6
7
8
t [s]
Fig. 10. Collision probability rate and mean of constant acceleration TTC for
an initial condition almost in front of the vehicle: (x, y) = (10, 0)m. The
process noise PSD for both coordinates is q̃x = q̃y = 0.0101m2 s−5 .
VIII. C ONCLUSIONS
As detailed in our literature review a common approach
to compute a collision probability is via temporal collision
measures such as time-to-collision or time-to-go. In this paper,
however, we have pursued a different approach, namely the
derivation of a collision probability rate without temporal
collision measures as an intermediate or prerequisite quantity.
A collision probability rate then affords the provision of a collision probability over an extended period of time by temporal
integration. We have also shown that computations of TTCs
using assumptions based on the shape of trajectories or the
prediction dynamics (disregard of process noise), or on simple
A PPENDIX
A. Partitioned Gaussian densities
In many calculations in stochastic estimation there is a need
to marginalize over certain elements of a state vector or to
obtain lower dimensional distributions by conditioning with
respect to certain elements. For these calculations the original
state vector ξ can be rearranged or partitioned such that xr
denotes the remaining state vector and xm denotes the states
to be marginalized over or which are used for conditioning.
xr
ξ=
(15)
xm
Hence the mean vector µ and covariance matrix Σ can be
partitioned into
Σrr Σrm
µr
µ=
, Σ=
(16)
µm
Σ>
rm Σmm
The following two well-known results on multivariate Gaussians are used in this paper:
a) Marginalization: The probability density of ξ
marginalized with respect to xm is
Z
p (xr ) =
p (ξ) dxm = N (xr ; µr , Σrr )
(17)
xm
b) Conditioning: The probability density of ξ conditioned on xm is
covariance matrix R(tk ). The measurement function h is given
by typical radar measurements (r, φ, ṙ), i. e.
p
x2 + y 2
y
h(ξ) = arctan x
x
ẋ+y
ẏ
√ 2 2
x +y
(24)
∂h
∂ξ
p (ξ|xm ) = p (xr |xm )
= N xr ; µr|m , Σr|m
(18)
with
µr|m = µr + Σrm Σ−1
mm (xm − µm )
(19)
>
Σr|m = Σrr − Σrm Σ−1
mm Σrm
(20)
and H =
is its linearization.
−
The covariance matrix P∞
at steady state can be determined
by the the discrete algebraic Riccati equation using the matrices defined above:
−
− >
P∞
= F P∞
F
− >
− >
−F P∞
H HP∞
H +R
−1
− >
HP∞
F
+Q
B. Dynamical system
The vehicle kinematics is characterized by a sixdimensional state vector
>
ξ = x y ẋ ẏ ẍ ÿ
(21)
(25)
C. Evaluation of the 2D integral for the collision probability
rate
The integral
Z Z
ẋ pt (y, ẋ|x0 ) dydẋ
(26)
The dynamical model is a discrete-time equivalent white noise
jerk model (see e. g. [14])
ẋ≤0 y∈Iy
in eq. (13) for the collision probability rate cannot be computed in closed form if the covariance matrix of pt (y, ẋ|x0 )
is not diagonal. Instead of a full numerical 2D integration
where
scheme we Taylor-expand the 2D pdf with respect to the off
∆t2k
diagonal element of the inverse covariance matrix around 0 to
1 0 ∆tk
0
0
2
2
a certain order and then integrate. For a general 2D Gaussian
∆t
k
0 1
0
∆tk
0
2
pdf p(x1 , x2 ) = N (ξ; µ, Σ) with ξ = (x1 , x2 )> and mean µ
1
0
∆tk
0
F (∆tk ) = 0 0
and covariance matrix Σ the Taylor-expansion to linear order
0 0
0
1
0
∆tk
with respect to Σ−1
12 reads
0 0
0
0
1
0
q
q
0 0
0
0
0
1
N (ξ; µ, Σ) = N x1 ; µ1 , Σ̃11 N x2 ; µ2 , Σ̃22
q
with ∆tk = tk+1 −tk . The corresponding predicted covariance
−
−1
matrix is denoted by P (tk+1 ). The process noise ν(tk )
·
−Σ12 (x1 − µ1 )N x1 ; µ1 , Σ̃11
is modeled by a white, mean-free Gaussian process with
q
covariance matrix Q
· (x2 − µ2 )N x2 ; µ2 , Σ̃22
5
∆t4
∆t3
∆tk
k
k
q̃
0
q̃x
0
q̃x
0
20 x
8
6
2
∆t5
∆t4
∆t3
+O (Σ−1
k
k
k
0
12 )
q̃
0
q̃
0
q̃
ξ − (tk+1 ) = F (∆tk )ξ(tk ) + ν(tk )
4
∆tk
q̃
Q(∆tk ) = 8 x
0
3
∆tk
6 q̃x
0
20
y
∆t3
k
q̃x
3
0
∆t4
k
8
q̃y
0
∆t3
k
q̃y
6
0
∆t2
k
q̃x
2
0
(22)
y
8
∆t2
k
q̃x
2
0
∆t3
k
3
q̃y
0
0
∆tk q̃x
∆t2
k
q̃y
2
0
6
0
y
q̃
y
2
0
∆tk q̃y
∆t2
k
where q̃x , q̃y are the jerk power spectral densities (PSD) in xand y-direction of the underlying continuous noise process.
Discrete-time equivalent noise – as opposed to discrete-time
counterpart noise – has been deliberately chosen because the
power spectral density as characterizing noise parameter is
independent of ∆tk . Hence ∆tk in prediction can be changed
arbitrarily without adjusting the power spectral density.
The measurement model is
z(tk ) = h(ξ − (tk )) + r(tk )
(23)
where z(tk ) is the measurement and the measurement noise
r(tk ) is modeled by a white, mean-free Gaussian process with
Σ12
with Σ−1
12 = − |Σ| and Σ̃11 =
the following integral
|Σ|
Σ22 , Σ̃22
=
|Σ|
Σ11 .
This leads to
Zx1u Zx2u
x1 p(x1 , x2 )dx1 dx2 =
x1l x2l
!
#x1u
q
x1 − µ1
p
− Σ̃11 N x1 ; µ1 , Σ̃11
·
2Σ̃11
x1l
"
!#x2u
1
x2 − µ2
·
erf p
2
2Σ̃22
x2l
"
!#x1u
q
Σ̃11
x1 − µ1
−1
−Σ12 x1 Σ̃11 N x1 ; µ1 , Σ̃11 −
erf p
·
2
2Σ̃11
x1l
x2u
q
2
· Σ̃22 N x2 ; µ2 , Σ̃22
+ O (Σ−1
12 )
"
µ1
erf
2
x2l
We expand with respect to the off-diagonal entry of the
inverse covariance since this is the element that prevents factorization into one-dimensional Gaussians. We have checked
that Taylor-expansion with respect to the off-diagonal entry
of the covariance matrix leads to less accurate results to first
order.
R EFERENCES
[1] J. Jansson and F. Gustafsson, “A framework and automotive application
of collision avoidance decision making,” Automatica, vol. 44, no. 9, pp.
2347–2351, 2008.
[2] M. M. Muntzinger, S. Zuther, and K. Dietmayer, “Probability estimation
for an automotive pre-crash application with short filter settling times,”
in Proceedings of IEEE Intelligent Vehicles Symposium, 2009, pp. 411–
416.
[3] J. Jansson, “Collision avoidance theory: With application to automotive
collision mitigation,” Ph.D. dissertation, Linköping University Electronic
Press, 2005.
[4] A. Lambert, D. Gruyer, and G. Saint Pierre, “A fast monte carlo
algorithm for collision probability estimation,” in Control, Automation,
Robotics and Vision, 2008. ICARCV 2008. 10th International Conference
on. IEEE, 2008, pp. 406–411.
[5] M. Prandini, J. Hu, J. Lygeros, and S. Sastry, “A probabilistic approach
to aircraft conflict detection,” IEEE Transactions on intelligent transportation systems, vol. 1, no. 4, pp. 199–220, 2000.
[6] P.-J. Nordlund and F. Gustafsson, Probabilistic conflict detection for
piecewise straight paths. Linköping University Electronic Press, 2008.
[7] A. Genz, “Numerical computation of rectangular bivariate and trivariate
normal and t probabilities,” Statistics and Computing, vol. 14, no. 3, pp.
251–260, 2004.
[8] A. Eidehall and L. Petersson, “Statistical threat assessment for general
road scenes using monte carlo sampling,” IEEE Transactions on intelligent transportation systems, vol. 9, no. 1, pp. 137–147, 2008.
[9] Y. K. Belyaev, “On the number of exits across the boundary of a region
by a vector stochastic process,” Theory of Probability & Its Applications,
vol. 13, no. 2, pp. 320–324, 1968.
[10] G. Lindgren, “Model processes in nonlinear prediction with application
to detection and alarm,” The Annals of Probability, pp. 775–792, 1980.
[11] L. Mihaylova, A. Y. Carmi, F. Septier, A. Gning, S. K. Pang, and
S. Godsill, “Overview of bayesian sequential monte carlo methods for
group and extended object tracking,” Digital Signal Processing, vol. 25,
pp. 1–16, 2014.
[12] P. Broßeit, M. Rapp, N. Appenrodt, and J. Dickmann, “Probabilistic
rectangular-shape estimation for extended object tracking,” in Intelligent
Vehicles Symposium (IV), 2016 IEEE. IEEE, 2016, pp. 279–285.
[13] R. Altendorfer, “Observable dynamics and coordinate systems for automotive target tracking,” in Proceedings of IEEE Intelligent Vehicles
Symposium, 2009, pp. 741–746.
[14] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with Applications to Tracking and Navigation. Wiley, 2001.
| 3cs.SY
|
arXiv:1802.01098v1 [math.AG] 4 Feb 2018
GEOMETRIC EQUIVALENCE of π-TORSION-FREE
NILPOTENT GROUPS.
R. Lipyanski
This article is dedicated to my teacher Prof. B. Plotkin on his 90th anniversary.
Abstract. In this paper we study the property of geometric equivalence of
groups introduced by B. Plotkin [P1, P2]. Sufficient and necessary conditions
are presented for π-torsion-free nilpotent group to be geometrically equivalent to its π-completion. We prove that a relatively free nilpotent π-torsionfree group and its π-completion define the same quasi-variety. Examples of
π-torsion-free nilpotent groups that are geometrically equivalent to their πcompletions are given.
1. Introduction
I attended Professor Plotkin’s lectures about 45 years ago as a student at the
University of Latvia. Over the years, I participated in Plotkin’s seminars on group
theory in Latvia and Israel. I am grateful to Boris Isaakovich Plotkin for revealing
to me the world of group theory in all its diversity. The problems posed by Plotkin
at seminars and in his numerous papers have always been interesting to me. I would
like to wish my teacher further creative success in Algebra in the next 30 years.
The foundation of universal algebraic geometry for different algebraic systems
were laid by B. Plotkin [P1, P2, P3]. Within the framework of universal algebraic
geometry (UAG), the important notion of geometric equivalence of universal algebras was introduced in [P1, P2]. Geometric equivalence of two algebras means, in
some sense, equal possibilities for solving systems of equations in these algebras.
Algebraic geometry over groups was considered together with B.Plotkin by G.
Baumslag, A. Miasnikov, V. Remeslennikov [B1, B2] and others.
In [P3] B. Plotkin posed a question: Under what conditions is a nilpotent
torsion-free group geometrically equivalent to its Mal’tsev’s completion? Sufficient
conditions for this equivalence were presented in [T]. A criterion for a nilpotent
torsion-free group to be geometrically equivalent to its Mal’tsev’s completion was
given by V. Bludov and B. Guskov [BG]. Based on this criterion, they also provided
an example of a torsion-free three-step nilpotent group that is not geometrically
equivalent to its Mal’tsev’s completion.
1991 Mathematics Subject Classification. Primary 54C40, 14E20; Secondary 46E25, 20C20.
Key words and phrases. Geometric equivalence, Nilpotent groups, Qπ -powered groups.
1
2
R. LIPYANSKI
There exists a natural generalization of the concept of a torsion-free group (see
[L1]). Let π be a (non empty) set of primes. An integer is said to be a π-number
if it is a product of powers of primes from π. A group G is π-divisible if any
equation of the form xn = a, where a ∈ G and n is a π-number, has a solution
in G. Group G is π-torsion-free if any equation of the form xn = 1, where n is a
π-number, has only the trivial solution in G. By definition, a π-divisible π-torsionfree group is called π-complete. It is known that every π-torsion-free nilpotent
group G can be embedded as a subgroup into a π-complete nilpotent group (see
[Kh1]). The minimal π-complete nilpotent group containing the group G is called
the π-completion of G. If π consists of all primes, the π-completion of group G
coincides with its Mal’tsev’s completion.
Here we study the problem related to geometric equivalence of a nilpotent πtorsion-free group to its π-completion. We give necessary and sufficient conditions
for this geometric equivalence. The proof of this criterion uses the theory of πisolators of groups, as well as the approach proposed by V. Bludov and B. Guskov,
to study the problem of geometric equivalence of a torsion-free nilpotent group to
its Mal’tsev’s completion. Based on this criterion we prove that a relatively free
nilpotent π-torsion-free group is geometrically equivalent to its π-completion. If π
is the set of all primes, this result has already been proven by A. Tsurkov in [T].
In [BG] it was proven that every torsion-free two-step nilpotent group is geometrically equivalent to its Mal’tsev’s completion. This result corresponds to the
case when π is the set of all prime numbers. If π is a set of primes, the situation becomes more complicated and requires additional consideration. We prove that if G
is a finitely generated π-torsion-free two-step nilpotent group such that the group
G/Iπ (G′ ) is torsion free, then G is geometrically equivalent to its π-completion.
This assertion implies the above result on the geometric equivalence of an arbitrary
two-step nilpotent torsion-free group to its Mal’tsev’s completion.
The question of the geometric equivalence of an arbitrary π-torsion-free twostep nilpotent group to its π-completion remains open.
2. Preliminaries
2.1. Geometric equivalence of groups. Let us recall the main definitions
from [P1].
Let F = F (X) be a free group generated by a finite set X and G be a group. The
set Hom(F, G) of all homomorphisms from F to G can be treated as an affine space.
It is possible to define a Galois correspondence ′ between subsets in Hom(F, G) and
subsets in F . For a subset T ⊆ F define TG′ , G-closure of T
TG′ = {ϕ ∈ Hom(F, G) | T ⊆ Kerϕ}
On the other hand, for any subset A ⊆ Hom(F, G) define A′G , G-closure of A
\
A′G =
Kerµ
µ∈A
The set T ⊆ F can be treated as a system of equations that we solve in G. Then
TG′ is the set of all solutions in G of this system. Similarly, the set A can be treated
as an affine space whose points are homomorphisms. Then A′G is the system of all
equations written in F such that A is the set of all solutions of this system.
GEOMETRIC EQUIVALENCE OF π-TORSION-FREE NILPOTENT GROUPS.
3
Definition 2.1 ([P1],[P4]). A subset T ⊆ F (a subset A ⊆ Hom(F, G)) is
called G-closed if TG = TG′′ (respectively, AG = A′′G ).
Definition 2.2 ([P1],[P2]). Two groups G1 and G2 are called geometrically
equivalent if for any free group F = F (X) of finite rank and for any subset T ⊆ F
TG′′1 = TG′′2
In other words, the groups G1 and G2 are geometrically equivalent if and only
if every G1 -closed subset of the group F is G2 -closed and vice versa. Note that
the collection of all G-closed sets of the group F defines the Zariski topology on it.
Therefore, the groups G1 and G2 are geometrically equivalent if and only if they
define the same Zariski topology on F .
Proposition 2.3 ([P1]). Groups G1 and G2 are geometrically equivalent if and
only if every finitely generated subgroup of G1 can be approximated by subgroups of
G2 and vice versa.
Definition 2.4 ([BG],[T]). Let H and G be groups and F = F (X) be any
free group finitely generated by X. If every H-closed subset of F is G-closed, we
say that the topology defined by H is weaker than the topology defined by G and
write H G.
The relation is a preorder on the class of groups. This preorder defines an
order relation on classes of geometrically equivalent groups.
Proposition 2.5 ([BG],[T]). H G if and only if any finitely generated
subgroup of the group H can be approximated by subgroups of G.
Proposition 2.6 ([BG],[T]). H G for any subgroup H of G.
It is clear that if for groups G1 and G2 the conditions G1 G2 and G2 G1
hold, then they are geometrically equivalent.
2.2. π-completion of a π-torsion-free nilpotent group. Let π be a (non
empty) set of primes. An integer is said to be a π-number if it is a product of
powers of primes from π. For a subgroup H of a group G, the π-isolator Iπ (H) of
H in G is defined as follows:
Iπ (H) = {g ∈ G|g n ∈ H for some π − number n = n(g)}
A subgroup A is said to be π-isolated if Iπ (A) = A.
If π is the set of all prime numbers, the π-isolator of H in G is denoted by I(H)
and is called the isolator of H in G.
Proposition 2.7 (see [Kh1], [LR]). If H is a subgroup of a nilpotent group
G and π is a set of prime numbers, then Iπ (H) is a subgroup.
A group G is π-divisible if for every π-number n, every element g ∈ G has
an nth root of g in G, i.e., there exists an element h ∈ G such that hn = g. A
group is π-torsion-free if it has no non-identity element whose order is π-number,
i.e., Iπ (1) = 1. A π-divisible π-torsion-free group is said to be π-complete.
Let Qπ be the ring of all rational numbers whose denominators are π-numbers
and G be a π-divisible π-torsion-free group. An unary operation on G can be
defined by taking powers in Qπ satisfying the laws (xr )s = xrs for all r, s ∈ Qπ .
The group G with this additional unary operation is called a Qπ -powered group.
4
R. LIPYANSKI
Theorem 2.8 ([Kh1]). (a) Every π-torsion-free nilpotent group G of class c
b π of the same
can be embedded as a subgroup in the Qπ -powered nilpotent group G
bπ = Iπ (G), i.e., for any element g ∈ G
b π there exists a π-number
class c such that G
m such that g m ∈ G.
b π is unique up to isomorphism; moreover every isomorphism
(b) The group G
′
bπ onto G
c′ π .
ϕ : G → G extends to an isomorphism of G
b π is the minimal π-complete group containing G and it is called
The group G
the π-completion of G.
bπ as a π-torsion-free π-divisible
From here on we consider Qπ -powered group G
abstract group, i.e., as the group with respect to only the usual group operations.
In the special case, where π consists of all primes, then Qπ = Q, the property of
being π-complete coincides with the property of being complete. In this case, the
bπ of the group G coincides with the Mal’tsev completion G
b of G
π-completion G
(see [M]).
Definition 2.9. Let G be a group and H ≤ G be a subgroup of G. Denote
by H m the subgroup of G generated by mth powers of elements from H.
The subgroup H m is a verbal subgroup of G defined by the word xm .
We also need the following lemma, given by A. Mal’tsev [M]. The initial proof
of this lemma is based on the theory of Lie groups. The proof of this lemma in
the framework of group theory (without Lie context) was given by A. Klyachko
(Lemma 2.2, in [K]). See also the proof of Lemma 1.2 in [BG].
Lemma 2.10 ([M]). Let G be a nilpotent group of nilpotency class k. Fix a
k
natural number n and denote by Gn a subgroup of G generated by all elements of
k
the form g n , g ∈ G. Then the equation xn = h has a solution in the group G for
k
each h ∈ Gn .
3. π-torsion-free nilpotent groups
3.1. Geometric equivalence of π-torsion-free nilpotent groups to their
π-completions. Let π be a set of primes. Now we consider some examples of πtorsion-free nilpotent groups for which we can describe their π-completions.
Example 3.1. Let H be a finitely generated π-torsion-free abelian group. We
b π.
wish to describe the structure of the group H
It is well known (see, e.g., [Kr], Ch.6, Sec.20) that the group H is the finite
direct product of cyclic π-torsion-free groups, i.e., there exist integers s, k ≥ 0 and
prime numbers p1 ≤ · · · ≤ pk together with positive integers α1 , . . . αr such that
α
(3.1)
H∼
= Zs × Z/pα1 Z × · · · × Z/p k Z,
1
k
where pi , i = 1, . . . , k, are not π-numbers.
Denote by Q+
π the additive group of the ring Qπ . It is clear, that π-completion
b π of Z is equal to Q+ . If G is a p-group, where p ∈
Z
/ π, then for every element
π
g ∈ G and m ∈ π there exists an mth root of g. In fact, the mapping g → g m is an
i
/ π,
automorphism of the cyclic subgroup < g >. Therefore, every group Z/pα
i Z, pi ∈
is π-complete. It is easy to check that if a group A is π-torsion-free nilpotent such
c2 π . As a consequence, we obtain
bπ = A
c1 π × A
that A = A1 × A2 , then A
α
bπ ∼
H
= (Q+ )s × Z/pα1 Z × · · · × Z/p k Z
π
1
k
GEOMETRIC EQUIVALENCE OF π-TORSION-FREE NILPOTENT GROUPS.
5
b π.
By Proposition 2.3, that H is geometrically equivalent to H
Denote by T the torsion group of H and by Tbπ its π-completion. It is clear
that Tbπ = T , i.e., the group T is π-complete.
Now we wish to select another class of π-complete nilpotent groups.
Example 3.2. Let G be a π-torsion-free nilpotent group such that the quotient
group G/G′ is finite. We will show that the group G is π-complete.
Since the group G is nilpotent and G/G′ is finite, G is finite (see [W], Sec.3,
p.9). Therefore, the group G is the direct product of its Sylow subgroups:
G = G(p1 ) × G(p2 ) × · · · × G(pk ).
Since G(pi ) is a π-torsion-free nilpotent group, pi ∈
/ π. As a consequence, the
group G(pi ) is π-complete. Hence, the group G, as the direct product of π-complete
groups, is π-complete too.
Our goal is to provide a criterion for nilpotent π-torsion-free groups to be
geometrically equivalent to their π-completions. For that purpose we need the
result from ([LR] Ch. 2, Sec. 3) that was established by A. Mal’tsev ([M]) in the
case when π is the set of all prime numbers.
Lemma 3.3. Let G be a finitely generated nilpotent group, g1 , . . . , gn be a set of
generators of G, and H be a subgroup of G. Suppose that giri ∈ H for some positive
π-numbers ri , i = 1, . . . , n. Then each element of G has a positive π-power in H
and | G : H | is a finite π-number.
We shell need the following lemma
bπ be its π-completion.
Lemma 3.4. Let G be a nilpotent π-torsion-free group and G
π
b
Then for any finitely generated subgroup H of G there exists a π-number r = r(G)
such that H r ≤ G.
Proof. Let the group H be generated by h1 , . . . , hn . Then there are πnumbers ri such that hri i ∈ G. Denote by D the subgroup of H generated by
hr11 , . . . , hrnn . The group D is a subnormal subgroup of H, i.e., there is a normal
series
D = D1 ≤ D2 ≤ · · · ≤ Dn−1 ≤ Dn = H
from D to H. By Lemma 3.3, the index | H : D | is a π-number. Therefore, each
factor in this series has an index that is a π-number. If rn−1 is the order of the
quotient group H/Dn−1 , then H rn−1 ≤ Dn−1 . By induction, it is easy to prove
that there exists a sequence of π-numbers r1 , . . . , rn−1 such that H r ≤ D, where
r = r1 · · · rn−1 . Since D is a subgroup of G, this completes the proof.
Now we derive a criterion of geometric equivalence of a π-torsion-free nilpotent
group to its π-completion.
bπ be its πTheorem 3.5. Let G be a π-torsion-free nilpotent group and G
b π are geometrically equivalent if and only if the
completion. Then groups G and G
group G is geometrically equivalent to subgroups Gm for every π-number m.
Proof. Our argument goes along the lines of Theorem 3.1 in [BG] but we
add new assertions related to π-torsion in nilpotent groups.
bπ be geometrically equivalent. Let us show that then
Let the groups G and G
m
groups G and G are also geometrically equivalent for any π-number m. The group
6
R. LIPYANSKI
Gm is a subgroup of G. By Proposition 2.6, we have Gm G. Let us prove that
G Gm . Consider a finitely generated subgroup H =< h1 , . . . , hs > of G. Denote
1
1
1
bπ . Denote by H m1 =< h m , . . . , hsm >
by him the root of the equation xm = hi in G
1
1
bπ generated by h m , i = 1, . . . , s. It is clear that H m1 G
bπ .
the subgroup of G
i
Q
1
π
b
Therefore, there exists an embedding ϕ : H m → i∈I Gi , where Gi , i ∈ I, are
bπ
isomorphic copies of G. Since the groups
Q G and G are geometrically equivalent,
1
m
there exists an embedding ψ : H → i∈I Gi . Consequently, H is embedded into
a cartesian product of the group Gm . Therefore, H Gm . As a consequence, we
obtain G Gm . Therefore, G and Gm are geometrically equivalent.
Conversely, let the groups G and Gm be geometrically equivalent for every πb π are geometrically equivalent. Since G is a
number m. We will prove that G and G
π
π
b
b
bπ G. Let H be a subgroup
subgroup of G , we have G G . Let us prove that G
π
b
of G generated by elements h1 , . . . , hs . By Lemma 3.4, there exists a π-number
r such that H r ≤ G. By Proposition 2.6, H r G. If k is the nilpotency class of
k
k
G, then G is geometrically equivalent to Gr . Therefore, H r Gr . Hence, there
Q
k
exists an embedding ϕ : H r → i∈I Gri , where Gi , i ∈ I, are isomorphic copies of
the group G. Let ϕi , i ∈ I, are the coordinate functions of ϕ, i.e., ϕ = (ϕi )i∈I .
π
d
rk = G
b π and G
b π . By Theorem 2.8, the mapping ϕ can
cr π = H
It is clear that H
bπ → G
bπ . Then ϕ̄(hr ) = ϕ(hr ) = (ϕi (hr ))i∈I .
be extended to the embedding ϕ̄ : H
j
j
j
r
By Lemma 2.10, the equation x = ϕi (hj r ) has a solution in Gi . Since Gi is a
π-torsion-free group, ϕi (hj ) ∈ Gi . Therefore, the mapping ϕ̄ is an embedding of
bπ G. As a consequence, the
H into a cartesian power of the group G. Hence, G
π
b and G are geometrically equivalent.
groups G
Corollary 3.6. Let G be a π-torsion-free nilpotent group. If every finitely
b π , then
generated subgroup H of G is geometrically equivalent to its π-completion H
π
b
the groups G and G are geometrically equivalent.
b π are geometrically equivalent. Therefore, by
Proof. The groups H and H
m
Theorem 3.5, H and H are geometrically equivalent for every π-number m. As
aQconsequence, H H m . By Proposition 2.5, there exists an embedding µ : H →
m
of the group H. The mapping
i∈I Hi , where Hi , i ∈ I, are isomorphic copies Q
µ can be extended to the embedding µ̄ : H → i∈I Gm
i , where Gi , i ∈ I, are
isomorphic copies of the group G. Therefore, G Gm . Since Gm G, the
groups G and Gm are geometrically equivalent. By Theorem 3.5, the group G is
bπ .
geometrically equivalent to its π-completion G
Corollary 3.7. Let G be a π-torsion-free abelian group. Then G is geometbπ .
rically equivalent to its π-completion G
Proof. Let H be a finitely generated subgroup of G. According to Example
b π and H are geometrically equivalent. By Corollary 3.6, the group
3.1, the groups H
G is geometrically equivalent to its π-completion.
Now we give other examples of π-torsion-free nilpotent groups which are geobπ .
metrically equivalent to their π-completions G
In what follows, we use the following Dick’s Theorem (cf. [Kr], Ch.5, Sec.18)
GEOMETRIC EQUIVALENCE OF π-TORSION-FREE NILPOTENT GROUPS.
7
Theorem 3.8. Let a group G be represented as follows
G =< xi | rj (xj1 , . . . xjs >= 1; i, jk ∈ I, j ∈ J >
Suppose that a group D contains a set {di , i ∈ I} such that
rj (dj1 , . . . , djs ) = 1
holds in D for every j ∈ J. Then the map xi → di , i ∈ I extends to a homomorphism of G onto D.
We also need the following claim
Proposition 3.9 (see [Kh2]). Let H be a π-torsion-free nilpotent group. If
xn · y m = y m · xn , where x, y ∈ H and m, n are π-numbers, then x · y = y · x.
Now we are ready to prove
Proposition 3.10. Let G be a finitely generated π-torsion-free two step nilpotent group such that the group G/Iπ (G′ ) is torsion-free. Then G is geometrically
bπ .
equivalent to its π-completion G
Proof. First, let us show that Iπ (G′ ) is an abelian group. If d1 , d2 ∈ Iπ (G′ ),
then there are π-numbers k1 , k2 such that dk11 , dk22 ∈ G′ . Since G is a two step
nilpotent group, dk11 · dk22 = dk22 · dk11 . By Proposition 3.9, d1 · d2 = d2 · d1 . Since
G is a finitely generated nilpotent group, the subgroup Iπ (G′ ) is an abelian finitely
generated group.
Let us choose a basis ḡ1 , . . . , ḡs in the abelian group G/Iπ (G′ ). Denote by
g1 , . . . , gs the preimages of these elements in the group G. Now choose a basis
c1 , . . . , ck in the abelian group Iπ (G′ ). By definition, there are π-numbers pi , i =
1, . . . k, such that cpi i ∈ G′ .
The group G is generated by elements g1 , . . . , gs , c1 , . . . , ck . Let
(3.2)
R(g1 , . . . , gs , c1 , . . . , ck ) = 1
be a relation in the group G. Using the collection process in the group G (see ,e.g.,
[LR], Ch.2, Sec.2), the relation (3.2) can be rewritten as follows
(3.3)
g1r1 · . . . · gsrs · ct11 . . . · ctkk · W1 ([gi , gj ]) · W2 ([cl , gd ]) = 1, ri ≥ 0, ti ≥ 0,
where W1 ([hi , hj ]) is a word in the commutators [hi , hj ] and W2 ([cl , gd ]) is a word
in the commutators [cl , gd ]. In the group G/Iπ (G′ ) the relation (3.3) has the form
ḡ1r1 · . . . · ḡsrs = 1
Since the group G/Iπ (G′ ) is free abelian, r1 = · · · = rs = 0. Therefore, the relation
(3.3) is equivalent to
(3.4)
ct11 . . . · ctkk · W1 ([gi , gj ]) · W2 ([cl , gd ]) = 1
Denote by p = max pi . Raising the relation (3.4) to the power p, we have
i
(3.5)
ptk
p
p
1
cpt
1 . . . · ck · W1 ([gi , gj ] ) · W2 ([cl , gd ] ) = 1
Since G is a π-torsion-free group, the relation (3.5) is equivalent to (3.4). Since
[cl , gd ]p = [cpl , gd ] = 1, we obtain W2 ([cl , gd ]p ) = 1. As a consequence, the relation
(3.5) is equivalent to
(3.6)
ptk
p
1
cpt
1 . . . · ck · W1 ([gi , gj ] ) = 1
i
from (3.6) belong to G′ .
Note, that elements cpt
i
8
R. LIPYANSKI
For a π-number m consider the map ϕ : G → Gm given by the rules:
2
ϕ(gi ) = gim , ϕ(ci ) = cm
i .
2
Since [gim , gjm ] = [gi , gj ]m , the relation (3.6) has the following form
(3.7)
cm
1
2
pt1
. . . · cm
k
2
ptk
2
· (W1 ([gi , gj ]p ))m = 1
Since the group G is π-torsion-free, the relation (3.7) is equivalent to the relation
(3.6), and, as a consequence, to the relation (3.2). By Dick’s Theorem, the mapping
ϕ can be extended to a homomorphism of G into Gm . Since G is a π-torsion-free
group, this homomorphism is a monomorphism G into Gm . Therefore, G Gm .
bπ .
By Theorem 3.5, G is geometrically equivalent to G
Corollary 3.11 ([BG]). Let G be a torsion-free two-step nilpotent group.
b
Then G is geometrically equivalent to its Mal’tsev’s completion G.
Proof. Let H be a finitely generated subgroup of G. The group H/I(H ′ ) is
torsion-free. By Proposition 3.10, the group G is geometrically equivalent to its
b
Mal’tsev’s completion G.
Example 3.12. Let H3 (Z) be the Heisenberg group generated by elements
x, y, z with the defining relations zx = xz, yz = zy, [x, y] = z. Elements of H3 (Z)
are all formal expressions of the form xl y s z n , where l, m, n ∈ Z. The group H3 (Z) is
torsion-free two-step nilpotent. The commutator H3 (Z)′ is generated by the element
z. The group H3 (Z)/H3 (Z)′ is a free abelian one generated by the elements x and
y. By Corollary 3.11, the group H3 (Z) is geometrically equivalent to its Mal’tsev’s
\
completion H
3 (Z).
Example 3.13. Let G be a group generated by elements a, b, c with the defining relations ca = ac, cb = bc, [a, b] = c2 , c4 = 1. Elements of G are all formal
expressions of the form al bs cn , where l, m, ∈ Z and n ∈ Z3 . The group G is two
step nilpotent π-torsion-free, where π is the set of all prime numbers without 2.
The commutator G′ =< c2 > and Iπ (G′ ) =< c >. The group G/Iπ (G′ ) is a free
abelian one generated by the elements a and b. By Proposition 3.10, the group G
bπ .
is geometrically equivalent to G
The following question arises: Is it true that every π-torsion-free two step
nilpotent group is geometrically equivalent to its π-completion?
3.2. Geometric equivalence of a relatively free nilpotent π-torsionfree group to its π-completion. It was proved by A Tsurkov in [T] that a
torsion-free nilpotent relatively free group is geometrically equivalent to its Mal’tsev’s
completion. His proof is based on the Mal’tsev correspondence technique between
nilpotent Q-powered groups and nilpotent Lie algebras over Q as well as on the
Lazard correspondence between groups with a central filtration and Lie rings.
Our goal is to extend Tsurkov’s result to the class of relatively free nilpotent πtorsion-free groups. We prove that a relatively free nilpotent π-torsion-free group is
geometrically equivalent to its π-completion. Note that, there is no good correspondence between nilpotent Qπ -powered groups and Lie Qπ -algebras (see [Kh1]). For
this reason, here we use the technique related to the theory of isolators in groups.
In the proof of this result we also apply the criterion for a π-torsion-free nilpotent
group to be geometrically equivalent to its π-completion (see Theorem 3.5).
GEOMETRIC EQUIVALENCE OF π-TORSION-FREE NILPOTENT GROUPS.
9
The method of associating a Lie ring to a group was studied in detail by M.
Lazard [L2] (see also [B]).
Let G be a group. Consider the lower central series of G:
G ≥ γ1 (G) ≥ · · · ≥ γi (G) ≥ γi+1 (G) ≥ · · ·
Let π be a set of primes. Denote by Gi = Iπ (γi (G)) the π-isolator the group γi (G)
in G. Note that the group Gi = Iπ (γi (G)) is a subgroup of G, since it coincides
with the preimage of the π-isolator of the identity subgroup of the nilpotent groups
G/γi (G) (see Theorem 2.7).
Therefore, we have a normal series:
(3.8)
G ≥ G1 ≥ · · · ≥ Gi ≥ Gi+1 ≥ · · ·
Lemma 3.14 (see [Kh1]). For any group G and any natural numbers i, j the
following holds
[Gi , Gj ] ≤ Gi+j ,
that is, {Gi |i = 1, . . . } is a central filtration of G.
Now consider the additive abelian group
∞
Lπ (G) = ⊕ Gi /Gi+1 (direct sum)
i=1
Since {Gi |i = 1, . . . } is a central filtration of G, then we can define the structure of
the Lie ring on the additive group Lπ (G): for any āi ∈ Gi /Gi+1 and āj ∈ Gj /Gj+1 ,
by definition, [āi , āj ] = [ai , aj ]Gi+j+1 ∈ Gi+j /Gi+j+1 . Then this bracket multipli∞
cation is extended to the direct sum Lπ (G) = ⊕ Gi /Gi+1 by the distributive laws.
i=1
We obtain the Lie ring Lπ (G) of the group G corresponding to the central filtration
(3.8). It is clear that the additive group of the ring Lπ (G) is π-torsion-free.
Let Fn = Fn (X) be a relatively free n-step nilpotent group generated by X. A
group G = Fn /H, where H is a π-isolated verbal subgroup of Fn , is relatively free
n-step nilpotent π-torsion-free. Note that if T is a verbal subgroup of Fn , then the
π-isolator Iπ (T ) is a π-isolated verbal subgroup of Fn .
The main aim of this section is to prove the following
Theorem 3.15. Let G be a relatively free nilpotent π-torsion-free group. Then
bπ .
G is geometrically equivalent to its π-completion G
Proof. By Theorem 3.5, it suffices to prove that the group G is geometrically
equivalent to the subgroups Gm for every π-number m. Let (xi )i∈I , I ⊆ N, be free
generators of the group G. Consider the map ϕ : G → G such that ϕ(xi ) = xm
i ,i ∈
I, where m is a π-number. Since G is a relatively free group, the map ϕ can be
extended to the endomorphism of G. This endomorphism is also denoted by ϕ. Let
us show that Ker ϕ = 0.
n−1
To this end we consider the Lie ring Lπ (G) = ⊕ Gi /Gi+1 , where n is the
i=1
nilpotency class of G. Since ϕ(γi (G)) ≤ γi (G), we have ϕ(Gi ) ≤ Gi . Therefore, the
endomorphism ϕ : G → G induces the endomorphism ϕ̄ : Lπ (G) → Lπ (G) of the
ring Lπ (G) such that ϕ̄(gi Gi+1 ) = ϕ(gi )Gi+1 for gi ∈ Gi . Let us write ϕ̄ = (ϕ̄i )i∈J ,
where ϕ̄i , i ∈ J = {1, . . . , n − 1}, are the coordinate functions of ϕ̄. We first prove
that Ker ϕ̄ = 0.
10
R. LIPYANSKI
Let ḡ = (ḡi )i∈J , where ḡi = gi Gi+1 , belongs to Ker ϕ̄. Since gi ∈ Gi =
Iπ (γi (G)), there exists π-number si such that gisi ∈ γi (G). In Lπ (G) we obtain
X
si g i =
vik mod γi+1 (G),
k
where vik = vik (xt1 , . . . , xtr ) are simple commutators in xt1 , . . . , xtr of the length
i. Without loss of generality, we assume that si gi ∈
/ Gi+1 . We obtain in Lπ (G)
X
si g i =
vik mod Gi+1 ,
i.e., si ḡi =
P
k
v̄ik , where v̄ik = vik Gi+1 . Thus we have
k
ϕ̄i (si ḡi ) =
X
i
m
v̄ik (xm
t1 , . . . , xtr ) = m
k
X
v̄ik (xt1 , . . . , xtr ) = mi si ḡi
k
i
Therefore, si ϕ̄j (ḡi ) = si m ḡi . Since si and m are π-numbers and the group Gi /Gi+1
is π-torsion-free, ḡi = 0. Hence, Ker ϕ̄ = 0.
Now we prove that Ker ϕ = 0. Let ϕ(h) = 0 for some h ∈ Gi \ Gi+1 . Since
ϕ(h) = 0, we have ϕ̄(h̄) = 0, i.e., h ∈ Gi+1 . This contradiction proves desired.
As a consequence, we obtain that ϕ is a monomorphism G into Gm . Therefore,
G Gm for every π-number m. It is clear that Gm G. Hence, G is geometrically equivalent to Gm for each π-number m. By Theorem 3.5, the group G is
bπ .
geometrically equivalent to its π-completion G
In the special case, that π consists of all primes, we obtain
Corollary 3.16 ([T]). A relatively free nilpotent torsion-free group is geometrically equivalent to its Mal’tsev’s completion.
It is known (see [P4],[MR]) that two nilpotent groups of the same nilpotency
class are geometrically equivalent if and only if they have the same quasiidentities.
By Theorem 3.15, we obtain
Corollary 3.17. A relatively free nilpotent π-torsion-free group and its πcompletion define the same quasi-variety.
4. Acknowledgments
I would like to express my appreciation to Prof. B. I. Plotkin for many inspiring discussions. I am also very pleased to thank Prof. G. I. Mashevitzky whose
suggestions helped make the paper more readable.
References
[B] Yu. A. Bahturin, Identical Relations in Lie Algebras, Utrecht:VNU, Science Press, 1987
[B1] G. Baumslag, A. Miasnikov, V. Remeslennikov, Algebraic geometry over groups I: Algebraic
sets an ideal theory, J. Algebra, vol. 219, 1999, pp. 16-19.
[B2] G. Baumslag, A. Miasnikov, V. Remeslennikov, Two theorem about equationnaly Noetherian
groups, J. Algebra, vol.194, 1994, pp. 654-664.
[BG] V.V. Bludov, B. V. Gusev, Geometric equivalence of groups, Proc. Steklov Inst. Math.,
vol.257, 2007, pp. 561582.
[K] A.A. Klyachko Group Theory, Preprint, http://halgebra.math.msu.su/staff/klyachko/lect4.pdf,
2007
[Kr] A.G. Kurosh, The theory of groups, Nauka, Moskow, 1952.
GEOMETRIC EQUIVALENCE OF π-TORSION-FREE NILPOTENT GROUPS.
11
[Kh1] E.I. Khukhro, p-atomorphisms of finite p-groups, LMS, Lecture Note Series, vol. 246, Cambridge, 1998.
[Kh2] E.I. Khukhro, Nilpotent groups and their automorphisms, Berlin,;Nework: de Gruyter, 1993
[L1] M. Lazard, Problemes d’extension concernant les N-groups; inversion dela formule de Hausdorff, Ann. Sci. Comptes Rendus, Paris, 237, 1953, pp. 1377-1379.
[L2] M. Lazard, Sur les groups nilpotents et les anneaux de Lie, Ann. Sci. Ecole Norm. Sup., 71,
1954, pp. 101-190.
[M] A.I. Mal’tsev, Nilpotent torsion-free groups. (Russian), Izvestiya Akad. Nauk. SSSR. Ser.
Mat. vol.13 1949, pp. 201-212.
[MR] A. Miasnikov, V. Remeslennikov, Algebraic geometry over groups I: Algebraic geometry
over group II: Logical fondation, J. Algebra, vol. 234, 2000, pp. 225-276.
[P1] B.I. Plotkin, Varieties of algebras and algebraic varieties. Categories of algebraic varieties,
Siberien Adv. Math., vol. 2, 1997, pp. 64-97.
[P2] B.I. Plotkin, Algebraic logic, varieties of algebras and algebric variety, Proc. Int. Alg. Conf.
St. Petersburgs, New York, London, 1998.
[P3] B.I. Plotkin, E. Plotkin, A Tsurkov, Geometricakl equivalence of groups, Commun. Algebra,
vol.8, 1999, pp. 4015-4025
[P4] B.I. Plotkin, Seven lectures on Universal Algebraic Geometry, Preprint of Inst. Math., Hebrew Univ., 2000, Jerusalem.
[LR] John C. Lennox, Derek J. Robinson The theory of infinite soluble groups Oxford Mathematical Monographs. The Clarendon Press, 2004.
[T] A. Tsurkov, Geometric equivalence of nilpotent groups, Journal of Mathematical Sciences,
vol.110. (5) 2007 pp. 407–409.
[W] Robert B. Warfield, Jr., Nilpotent groups, Springer Verlag, New York, 1976.
Department of Mathematics, Ben Gurion University, Beer Sheva 84105, Israel
Current address: Department of Mathematics, Ben Gurion University, Beer Sheva 84105,
Israel
E-mail address: [email protected]
| 4math.GR
|
Extremal dependence of random scale constructions
Sebastian Engelke1 , Thomas Opitz2 , and Jennifer Wadsworth3
1
Research Center for Statistics, University of Geneva, Boulevard du Pont d’Arve 40, 1205 Geneva, Switzerland,
[email protected]
2
Biostatistics and Spatial Processes, INRA, 84914, Avignon, France, [email protected]
3
Department of Mathematics and Statistics, Fylde College, Lancaster University, LA1 4YF, UK,
[email protected]
arXiv:1803.04221v1 [math.PR] 12 Mar 2018
March 13, 2018
Abstract
A bivariate random vector can exhibit either asymptotic independence or dependence between the largest
values of its components. When used as a statistical model for risk assessment in fields such as finance, insurance or meteorology, it is crucial to understand which of the two asymptotic regimes occurs. Motivated
by their ubiquity and flexibility, we consider the extremal dependence properties of vectors with a random
scale construction (X1 , X2 ) = R(W1 , W2 ), with non-degenerate R > 0 independent of (W1 , W2 ). Focusing
on the presence and strength of asymptotic tail dependence as expressed through commonly-used summary
parameters, broad factors that affect the results are: the heaviness of the tails of R and (W1 , W2 ), the shape
of the support of (W1 , W2 ), and dependence between (W1 , W2 ). When R is distinctly lighter tailed than
(W1 , W2 ), the extremal dependence of (X1 , X2 ) is typically the same as that of (W1 , W2 ), whereas similar
or heavier tails for R compared to (W1 , W2 ) typically result in increased extremal dependence. Similar tail
heavinesses represent the most interesting and technical cases, and we find both asymptotic dependence
and asymptotic independence of (X1 , X2 ) possible in such cases when (W1 , W2 ) exhibit asymptotic independence. Our results unify and extend many existing examples, and we use them to propose new models that
encompass both dependence classes.
Keywords: copula, extreme value theory, residual tail dependence, tail dependence.
MSC2010: 60G70, 60E05, 62H20
1
Introduction
A rich variety of bivariate dependence models have a pseudo-polar representation
(X1 , X2 ) = R(W1 , W2 ),
R > 0, independent of (W1 , W2 ) ∈ W ⊆ R2 ,
(1)
where we term R the radial variable, assumed to have a non-degenerate distribution, and (W1 , W2 ) the angular
variables. Indeed, many well-known copula families, including the elliptical, Archimedean, Liouville, and multivariate Pareto families have such a representation. In this work, our focus is on the upper tail dependence
of such constructions. In particular, we examine whether a given (X1 , X2 ) displays asymptotic dependence or
asymptotic independence, and the strength of dependence within these classes. Our results are particularly
useful for constructing new models with properties that reflect the challenges of real data in, for instance, finance, meteorology and hydrology. Specifically, it is often ambiguous whether data should be modeled using an
asymptotically dependent or asymptotically independent distribution, and most families of distributions only
exhibit one type of dependence. Through our results we are able to present new models that smoothly bridge
these two classes. A vector (X1 , X2 ) with Xj ∼ FXj is said to display asymptotic dependence if the limit
−1
−1
χX = lim P{X1 ≥ FX
(q), X2 ≥ FX
(q)}/(1 − q)
1
2
q→1
(2)
−1
exists and is positive; a limit of zero defines asymptotic independence. In (2) and throughout, FX
denotes the
j
(generalized) inverse of the distribution function FXj . The parameter χX is termed the (upper) tail dependence
coefficient, and the value of χX ∈ (0, 1] summarizes the strength of the dependence within the class of asymptotically dependent variables. Under asymptotic independence, a more useful summary is the rate at which the
convergence to zero in equation (2) occurs, and a widely satisfied assumption (Ledford and Tawn, 1996, 1997)
is
−1
−1
P{X1 ≥ FX
(q), X2 ≥ FX
(q)} = `(1 − q)(1 − q)1/ηX ,
1
2
1
ηX ∈ [0, 1],
(3)
where ` : [0, 1] → R+ is slowly varying at zero, i.e., lims→0 `(sx)/`(s) = 1, x > 0. The parameter ηX
is termed the residual tail dependence coefficient; positive and negative extremal association are indicated
respectively by ηX ∈ (1/2, 1] and ηX ∈ [0, 1/2), whilst asymptotically dependent variables have ηX = 1 and
χX = limq→1 `(1 − q). A value of ηX = 0 means that the left-hand side of (3) decays faster than any power of
1 − q, whilst if the left-hand side is exactly zero for some q < 1, we say that ηX is not defined.
Our particular interest in the extremal dependence of constructions of the form (1) stems not so much from
their novelty, but from their ubiquity and flexibility. As mentioned, (1) already encompasses many well-known
families, and moreover these families display different types of extremal dependence, which may be determined
by the distribution of R, the distribution of (W1 , W2 ), or its support W. There is a large body of literature that
treats either individual constructions of the form (1), or a particular subset of these constructions where R or
(W1 , W2 ) have certain specified properties; this literature will be reviewed in Section 4. Our aim is to bring this
scattered treatment together and more systematically characterize how the extremal dependence of (X1 , X2 ) is
determined by the properties of R and (W1 , W2 ). By understanding which facets of the construction lead to
different dependence properties, we are able to determine dependence models that can capture both types of
extremal dependence within a single parametric family; the recent proposals in Wadsworth et al. (2017), Huser
et al. (2017) and Huser and Wadsworth (2018) are specific examples of this.
A broad split in representations of type (1) is the dimension of W, the support of (W1 , W2 ). The most
common case in the literature is that W is a one-dimensional subset of R2 , such as the unit sphere defined by
some norm or other homogeneous function. Examples include the Mahalanobis norm (elliptical distributions),
L1 norm (Archimedean and Liouville distributions), or L∞ norm (multivariate Pareto distributions). On top
of the support W, to obtain distributions within a particular family, R or (W1 , W2 ) may be specified to have
a certain distribution. Where W is two-dimensional, it may sometimes be reduced to the one-dimensional case
by redefining R, such as in the Gaussian scale mixtures of Huser et al. (2017); other times, such as for the scale
mixtures of log-Gaussian variables in Krupskii et al. (2016), or the model presented in Huser and Wadsworth
(2018), this cannot be done. Where W is two-dimensional, the possible constructions stemming from (1) form
an especially large class, since (W1 , W2 ) can itself have any copula. In this case, we particularly focus on how the
multiplication by R changes the extremal dependence of (W1 , W2 ), summarized by the coefficients (χW , ηW ),
to obtain the extremal dependence of the modified vector (X1 , X2 ) in terms of its coefficients (χX , ηX ). The
marginal distributions of (W1 , W2 ) and R will play a crucial role, since, intuitively, the heavier the tail of R the
more additional dependence is introduced in the vector (X1 , X2 ).
In this work we are focused on the upper tail of (X1 , X2 ), and as such we henceforth assume (W1 , W2 ) ∈ R2+ ;
observe that by the invariance of copulas to monotonic marginal transformations, we also therefore cover random
location constructions of the form (Y1 , Y2 ) = S + (V1 , V2 ), S ∈ R, (V1 , V2 ) ∈ V ⊆ R2 . For simplicity of
presentation, we will often make the restriction that W1 and W2 have the same distribution, with comments on
relaxations of this assumption given in Section 6. Furthermore, whilst our focus on the bivariate case permits
simpler notation, the results are directly applicable to the bivariate margins of multivariate and spatial models,
whose extremal dependence is typically analyzed in terms of the coefficients (2) and (3). Examples will be given
in Section 4.
No widely recognized standard for ordering univariate tail decay rates in a stepless manner from the slowest
to the fastest rate has emerged in the literature so far. A broad, but coarse, characterization is given by the three
domains of attraction of the maximum. Specifically, we say that the random variable R is in the max-domain
of attraction (MDA) of a generalized extreme value distribution if there exists a function b(t) > 0 such that as
t → r? = sup{r : P(R ≤ r) < 1},
−1/ξ
P(R ≥ t + r/b(t))/P(R ≥ t) → (1 + ξr)+
,
r ≥ 0,
(4)
for some ξ ∈ R, where a+ = max(a, 0). The cases ξ > 0, ξ = 0, ξ < 0 define respectively the Fréchet, Gumbel
and negative Weibull domains of attraction; the tail heaviness of R increases with ξ. However, the Gumbel limit
in particular attracts distributions with highly diverse tail behavior such as finite upper bounds or heavy tails.
Overall, this classification is therefore too coarse for our requirements, and it excludes important classes such as
superheavy-tailed distributions defined through the property of heavy-tailed log-transformed random variables,
which yield relevant constructions with random location. In addition to the maximum domains of attraction,
we will utilize various commonly used tail classes, which are defined in Section 1.1; see also Appendix A.
We begin in Section 2 by presenting results concerning the tail dependence of construction (1) according to
the tail behavior of R and the shape of W, in the case where it is a one-dimensional support defined through a
norm. We then characterize various cases where W is two-dimensional, according to the behavior of both R and
(W1 , W2 ), in Section 3. Section 4 is devoted to literature review and framing a large number of existing examples
in terms of our general results, whilst Section 5 illustrates the properties of some new examples inspired by
the developments in the manuscript. In Section 6 we comment on generalizations and conclude. Proofs are
presented in Section 7.
2
1.1
Terminology and notation
For a random variable Q, we define its survival function F Q (q) = P(Q ≥ q), and distribution function FQ (q) =
1 − F Q (q). If Q represents a bivariate random vector Q = (Q1 , Q2 ), we denote the minimum of its margins
by Q∧ = Q1 ∧ Q2 . For two functions f and g with g(x) 6= 0 for values x above some threshold value x0 , we
write f ∼ g if f (x)/g(x) → 1, where the limit is considered for x → ∞ if not stated otherwise. Similarly, we
write f (x) = o(g(x)) to indicate that f (x)/g(x) → 0. The convolution of X ∼ FX and Y ∼ FY is denoted
FX ? FY = FX+Y .
We recall definitions of upper tail behavior classes for a random variable X with distribution F . Abbreviations for distribution classes are chosen to be as self-explanatory as possible. Important tail parameters for
these classes may be given as subscript, such as in ETα to refer to exponential-tailed distributions with rate α,
but we may omit the subscript if the specific value of the parameter is not of interest. Relationships between
distribution classes and membership of common distribution families are detailed in Appendix A.
Definition 1 (Heavy-tailed and superheavy-tailed distributions (HT, SHT)). The distribution F is heavy-tailed
if exp(−λx)F (x) → ∞ as x → ∞, for any λ > 0. Further, F is superheavy-tailed if F (exp(·)) is heavy-tailed.
If F is not heavy-tailed, it is light-tailed.
Definition 2 (Regularly varying functions and distributions (RV0α and RV∞
α )). A function g is regularly varying
at infinity or at zero with index α ∈ R if g(tx)/g(t) → xα as t → ∞ or t → 0 respectively for any x > 0. We
0
write g ∈ RV∞
α or g ∈ RVα respectively. If α = 0, then g is said to be slowly varying. A probability distribution
?
F with upper endpoint x? = ∞ is regularly varying with index α ≥ 0 if F ∈ RV∞
−α . If x < ∞, then F is
0
?
?
regularly varying at x with index α if F (x − ·) ∈ RVα .
Definition 3 (Exponential-tailed distributions (ETα and ETα,β )). The distribution F with upper endpoint
x? = ∞ is exponential-tailed with rate α > 0 if
F (t + x)/F (t) → exp(−αx)
(5)
as t → ∞, for any x > 0. If (5) holds with α = 0, we call F long-tailed, and we write F ∈ ET0 . If α > 0 and
F (x) = r(x) exp(−αx),
r ∈ RV∞
β ,
we write F ∈ ETα,β .
By definition, F ∈ ETα with α ≥ 0 if and only if F (log(·)) ∈ RV∞
−α . The class ETα,β with β > −1 is also
referred to as gamma-tailed distributions. Another important subclass of ETα are the convolution-equivalent
distributions.
Definition 4 (Convolution-equivalent distributions
R ∞ (CEα )). A distribution F is convolution equivalent with
index α ≥ 0 if F ∈ ETα and F ? F (x)/F (x) → 2 −∞ exp(αx)F (dx) < ∞. We write F ∈ CEα .
We refer to the class CE0 as subexponential distributions.
Definition 5 (Weibull-tailed distributions (WTβ )). The distribution F is Weibull-tailed with index β > 0 if
β
there exist α > 0 and r ∈ RV∞
γ such that F (x) ∼ r(x) exp(−αx ).
Definition 6 (Log-Weibull-tailed distribution (LWTβ )). The distribution F is log-Weibull-tailed with index
β > 0 if F (exp(·)) ∈ WTβ .
2
Constrained angular variables
In this section we focus on the case where W is defined by a norm ν; specifically we let W = {(w1 , w2 ) ∈
R2+ : ν(w1 , w2 ) = 1}. Other types of constrained spaces may occasionally be of interest, however norm spheres
are the most common restriction, and this focus allows greater generality in other aspects. In particular, all
components of the vector are bounded in absolute value when the value of the norm is fixed. We then examine
the extremal dependence based on the heaviness of the tail of R. Because the (W1 , W2 ) are bounded above,
and subject to additional mild assumptions, we can and do classify R according to its MDA in this section.
The case where R belongs to the Fréchet MDA is the least delicate one, since as long as R has a much heavier
tail than each of (W1 , W2 ), results do not depend strongly on other considerations. No equality in distribution
is assumed between W1 , W2 in this case. When R is in the Gumbel or negative Weibull MDA, the shape of the
norm ν becomes important, and some minor additional regularity conditions will be assumed, to be detailed in
Section 2.2.
3
2.1
Radial variable in Fréchet MDA
Many of the most familiar results in the literature on extremal dependence concern the case when R is in the
Fréchet MDA; this is equivalent to regular variation of the tail of R, namely F R ∈ RV∞
−1/ξ , ξ > 0, where α = 1/ξ
is called the tail index. A classical example of this is the multivariate Pareto copula, which can be constructed
as in equation (1) with R standard Pareto, and W = {(w1 , w2 ) ∈ R2+ : max(w1 , w2 ) = 1} (Ferreira and de Haan,
2014). Multivariate Pareto copulas can be identified with so-called extreme value copulas, which arise as the
limiting copulas of suitably normalized componentwise maxima; see e.g. Rootzén et al. (2018). The next result
provides the general form of the tail dependence coefficient for these models.
α+ε
)<
Proposition 1 (R in Fréchet MDA). Let F R ∈ RV∞
−α , α ≥ 0, P(W1 > 0) = P(W2 > 0) = 1, and E(Wj
∞, j = 1, 2, for some ε > 0. Then ηX = 1, and
W2α
W1α
,
.
(6)
χX = E min
E(W1α ) E(W2α )
Remark 1. When F R (r) ∼ Cr−α for some C > 0, then the condition E(Wjα+ε ) < ∞ can be replaced by
E(Wjα ) < ∞, by Lemma 2.3 of Davis and Mikosch (2008).
Remark 2. The condition E(Wjα+ε ) < ∞ is guaranteed when W is the unit sphere of a norm ν; Proposition 1
notably also covers the case where (W1 , W2 ) ∈ R2+ .
Remark 3. The result includes the case α = 0, although the tail of such an R is too heavy to be in any domain
of attraction. In this case, χX = 1, representing perfect upper tail dependence. This case is discussed further
in Section 3.1.
From (6), we observe that asymptotic dependence arises since P{min(W1α /E(W1α ), W2α /E(W2α )) > 0} = 1.
If the conditions of the Proposition were relaxed to P(W1 > 0), P(W2 > 0) > 0, then it is possible that for
α > 0, P{min(W1α /E(W1α ), W2α /E(W2α )) > 0} = 0 which would yield asymptotic independence. In this latter
case, ηX would not be defined. The Fréchet case with one-dimensional W is therefore very restricted in its
capacity to represent varied asymptotically independent behaviors.
A more complete description of asymptotically dependent random vectors is given by the exponent function,
defined as
−1
−1
VX (x1 , x2 ) = lim t[1 − P(X1 ≤ FX
{1 − 1/(tx1 )}, X2 ≤ FX
{1 − 1/(tx2 )}],
1
2
t→∞
x1 , x2 > 0.
Small modifications to Proposition 1 yield
VX (x1 , x2 ) = E max
W1α
W2α
,
E(W1α )x1 E(W2α )x2
.
(7)
The link between χX and VX (1, 1) can be obtained simply by inclusion-exclusion arguments; in particular since
min(a, b) = a + b − max(a, b), χX = 2 − VX (1, 1).
2
Example 1. Suppose that F R ∈ RV∞
−1 , i.e., α = 1, and (W1 , W2 ) ∈ W = {(w1 , w2 ) ∈ [0, 1] : w1 + w2 = 1}.
Letting (W1 , W2 ) = (W, 1 − W ) then W ∈ [0, 1], with E(W ) = 1/2, is the random variable described by the L1
spectral measure, discussed for instance in Coles and Tawn (1991). A simple non-trivial example is the Gumbel
or logistic spectral measure, which has Lebesgue density
h(w) =
11−θ
{w(1 − w)}1/θ−2 {w1/θ + (1 − w)1/θ }θ−2
2 θ
−1/θ
χX = 2 − 2θ and VX (x1 , x2 ) = (x1
2.2
θ ∈ (0, 1),
(8)
−1/θ θ
+ x2
) .
Radial variable in Gumbel MDA
Suppose that R is in the Gumbel domain of attraction, with upper endpoint r? i.e.,
lim F R (t + r/b(t))/F R (t) = e−r ,
t→r ?
where b(t) is termed the auxiliary function. One representation for distributions with this property is
Z r
F R (r) = c(r) exp −
b(t)dt ,
(9)
z
where z < r < r? , c(r) → c > 0 as r → r? , and a = 1/b is absolutely continuous with density a0 satisfying
limt→r? a0 (t) = 0 (e.g Embrechts et al., 2013, Chapter 3.3). If r? = ∞, we also have that
lim (rb(r))ρ F R (λr)/F R (r) = 0
r→∞
(10)
for any λ > 1, ρ ∈ R (Hashorva, 2012). Several distributions in this domain have mass on the negative half line,
but we suppose here that R is conditioned to be positive, which does not affect the tail behavior.
4
1.0
0.8
ζ
0.0
0.2
0.4
0.6
1
0
0.0
0.2
0.4
0.6
0.8
0
1.0
b1
b2
1
Figure 1: Left: the unit sphere for a particular norm ν; Right: illustration of τ (z) (solid line) and τ (1 − z)
(dashed line) for the same ν.
Notation and assumptions for (W1 , W2 )
d
d
Suppose that W1 = W2 = W ∈ [0, 1] and ν(W1 , W2 ) = 1. To this end, we assume that ν is a symmetric norm,
i.e., ν(x, y) = ν(y, x), and scaled to satisfy ν(x, y) ≥ max(x, y), such that the unit sphere of ν is contained in
that of max, with ν(b, 1 − b) = b for some b ≥ 1/2. Let τ (z) = z/ν(z, 1 − z) = 1/ν(1, 1/z − 1). For a random
variable Z ∈ [0, 1], with distribution symmetric about 1/2, we can express
(W1 , W2 ) =
(Z, 1 − Z)
= (τ (Z), τ (1 − Z)).
ν(Z, 1 − Z)
(11)
Define Iν = [b1 , b2 ] ⊆ [1/2, 1] as the interval such that τ (z) = 1 for all z ∈ Iν , and τ (z) < 1 for z 6∈ Iν , and write
τ1 (z), z ∈ [0, b1 ],
τ (z) = 1,
z ∈ [b1 , b2 ],
τ2 (z), z ∈ [b2 , 1],
with τ1 strictly increasing and τ2 strictly decreasing. Figure 1 illustrates τ for a particular ν; further illustrations
are given in Appendix B. We assume further that
(Z1): Z has a Lebesgue density, fZ , positive everywhere on [0, 1], and that its survival function is regularly
varying at one, with F Z (1 − ·) ∈ RV0αZ , αZ > 0,
and make the following mild regularity assumptions on the norm, ν, or equivalently τ .
(N1): The function τ is twice (piecewise) continuously differentiable except for finitely many points, at which
we only require existence of left and right derivatives of first and second order.
(N2): τ is regularly varying as it approaches 1 from either side, i.e., 1 − τ1 (b1 − ·) ∈ RV01/γ1 , and, if b2 < 1,
1 − τ2 (b2 + ·) ∈ RV01/γ2 , γj ∈ (0, 1], j = 1, 2. We label γ = min(γ1 , γ2 ) with γ = γ1 if b2 = 1.
In practice, (N1) and (N2) are satisfied by a wide variety of commonly used norms; furthermore the upper
limit of γ1 , γ2 ≤ 1 in (N2) is a consequence of convexity of the norm ν; see Lemma 3 in Section 7. Note that
τ (z) ≶ τ (1 − z) ⇔ z ≶ 1 − z, so that
(
τ (z),
z ∈ [0, 1/2],
min(τ (z), τ (1 − z)) =
τ (1 − z), z ∈ [1/2, 1].
Finally, denote ζ = τ (1/2) ∈ [1/2, 1], so that W∧ = min(τ (Z), τ (1 − Z)) ∈ [0, ζ].
Proposition 2 (R in Gumbel MDA). Assume F R satisfies (9) and that (N1), (N2) and (Z1) hold. Then:
1. If ζ < 1, χX = 0 and ηX = limx→r? log F R (x)/ log F R (x/ζ), which is defined only for r? = ∞.
2. If ζ = 1, then b1 = 1/2 and
χX =
0
if b2 > 1/2, i.e., P(W = 1) > 0,
2τ20 (1/2+ )
0
τ2 (1/2+ ) − τ10 (1/2− )
and ηX = 1.
5
otherwise,
We observe that asymptotic independence arises for ζ < 1, with the residual tail dependence coefficient
determined by the properties of F R . The following corollary covers an important subclass of distributions in
the Gumbel MDA.
δ
Corollary 1. If ζ < 1 and − log F R ∈ RV∞
δ , δ ≥ 0, then ηX = ζ .
We note that if δ = 0, such as in the case of log-normal R, ηX = 1. Other possibilities in the Gumbel domain
include − log F R (x) ∼ exp(x), as in the reverse Gumbel distribution, for which ηX = 0. If r? < ∞ then since
the upper endpoint of X∧ is less than that of X1 , ηX is not defined.
If ζ = 1, then one has asymptotic independence only if P(W = 1) = P(Z ∈ Iν ) > 0, which is equivalent
to ν(x, y) behaving locally like the L∞ norm around the point x = y. If ζ = 1 and P(W = 1) = 0, then we
must have b1 = b2 = 1/2, and the “pointy” shape of such norms induces asymptotic dependence. The following
example provides an illustration of this case.
Example 2. Let ν(x, y) = θ max(x, y) + (1 − θ) min(x, y), θ > 1; see Figure 2 for an illustration. We have
ζ = 1 and b1 = b2 = 1/2, so we can calculate χX > 0 using Proposition 2, by evaluating τ10 (1/2+ ) = 4θ and
τ20 (1/2− ) = −4(θ − 1). This yields χX = 2(θ − 1)/(2θ − 1), which is an increasing function of θ; in particular
χX → 0+ as θ → 1+ and χX → 1− as θ → +∞.
θ=5
1.0
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
θ = 1.2
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Figure 2: Illustration of the unit sphere of ν(x, y) = θ max(x, y) + (1 − θ) min(x, y) for two different values of θ.
Example 3. Let ν(x, y) = (xp + y p )1/p , p ≥ 1, so that ζ = 2−1/p < 1. In this case b1 = b2 = 1 and
1 − τ −1 (1 − s) = ps1/p [1 + o(1)], i.e., γ = 1/p. For any Z satisfying the regular variation condition on F Z ,
then the conditions of Proposition 2 are satisfied, and ηX = limx→r? log F R (x)/ log F R (21/p x). As a concrete
example, if F R (x) = exp(−xδ ), then ηX = 2−δ/p by Corollary 1.
2.3
Radial variable in negative Weibull MDA
Suppose that R > 0 is in the negative Weibull domain of attraction with upper endpoint r? > 0, i.e,
` ∈ RV00 , αR > 0;
F R (r? − s) = `(s)sαR ,
(12)
equivalently F R (r? − ·) ∈ RV0αR . Note that the distribution of R cannot have a point mass on r? . The general
assumptions for (W1 , W2 ) are the same as in Section 2.2.
Proposition 3 (R in negative Weibull MDA). Assume F R (r? − ·) ∈ RV0αR and that (N1), (N2) and (Z1) hold.
Then:
1. If ζ < 1, χX = 0 and ηX is not defined.
2. If ζ = 1, then b1 = 1/2 and
χX =
0
if b2 > 1/2, i.e., P(W = 1) > 0,
2τ20 (1/2+ )
0
τ2 (1/2+ ) − τ10 (1/2− )
otherwise.
and
(
ηX =
αR /(1 + αR )
1
if P(W = 1) > 0,
otherwise.
Example 4. Let ν(x, y) = max(x, y), so that ζ = 1, b1 = 1/2, b2 = 1 and P(W = 1) = 1/2. If, for example,
−1/λ
F R = (1 + λr)+ , λ < 0 with r? = −1/λ, this represents (part of) a model given in Wadsworth et al. (2017),
with ηX = (1 − λ)−1 . This follows directly from Proposition 3, noting αR = −1/λ.
Table 1 summarizes the tail dependence for (X1 , X2 ) using the norms from Examples 2–4, under different
tail behaviors for R.
6
Radial variable R
Regularly varying
F R ∈ RV∞
−α , α > 0
Log-normal
− log F R (r) ∼ k(log r)2 , k > 0
Weibull-like
− log F R (r) ∼ krδ , k, δ > 0
a) exponential (δ = 1)
b) normal (δ = 2)
Log of exponential
− log F R (r) ∼ k exp(r), k > 0
Exponential behavior at r? < ∞
F R (r? − 1/r) ∼ k exp(−r), k > 0
Negative Weibull
F R (r? − ·) ∈ RV0α , α > 0
a) uniform (r? = 1, α = 1)
ν = Lp norm, p ≥ 1
χX
ηX
ν = L∞ norm
χX
ηX
ν = θ max +(1 − θ) min, θ ≥ 1
χX
ηX
eq. (6)
1
eq. (6)
1
eq. (6)
1
0
1
0
1
2(θ − 1)/(2θ − 1)
1
0
0
0
2−δ/p
2−1/p
2−2/p
0
0
0
1
1
1
2(θ − 1)/(2θ − 1)
2(θ − 1)/(2θ − 1)
2(θ − 1)/(2θ − 1)
1
1
1
0
0
0
1
2(θ − 1)/(2θ − 1)
1
0
ND
0
1
2(θ − 1)/(2θ − 1)
1
0
0
ND
ND
0
0
α/(1 + α)
1/2
2(θ − 1)/(2θ − 1)
2(θ − 1)/(2θ − 1)
1
1
Table 1: The values of χX and ηX for (X1 , X2 ) = R(W1 , W2 ) with different tail decay rates of the radial variable
R and angular variables defined on Lp , L∞ norms, and the norm of Example 2. ND = not defined.
3
Unconstrained angular variables
We now treat the case where the support W is two-dimensional. As noted in Section 1, there are cases where
(W1 , W2 ) itself might have a random scale representation, and by redefining the scaling variable we get back to
the situation of one-dimensional W. We thus focus on constructions where this is not necessarily the case. To
avoid additional complications we assume throughout this section that W1 and W2 share the common marginal
distribution FW of the angular variable W . We also generally assume that the tail dependence coefficient χW
and the residual tail dependence coefficient ηW of (W1 , W2 ) exist, although some results may still be obtained
with the latter undefined.
In Section 2, the constraints imposed by W being a unit sphere gave bounded marginal distributions for
Wj , j = 1, 2, and deterministic dependence between (W1 , W2 ). For two-dimensional W, the variety of marginal
and dependence behaviors possible for (W1 , W2 ) means that systematic characterization according only to the
MDA of R is more difficult. In fact, we need to consider different tail decays of both the radial variable R
and the angular variable W since the combination of the two is crucial to classify the extremal dependence
of (X1 , X2 ) = R(W1 , W2 ). We focus on some interesting sub-classes that still incorporate a wide variety of
structures and cover most of the parametric univariate distributions available for R and W .
This section is structured according to the tail heaviness assumed for R, W , or both of them. In decreasing order we consider distributions with superheavy tails, regularly varying distributions, distributions
of log-Weibull and Weibull type, and finally distributions with finite upper endpoint in the negative Weibull
domain of attraction. Table 2 summarizes the general results developed in the following, and Table 3 contains
the extremal dependence coefficients for all combinations of tail decays of R and W for the specific, yet interesting example where W1 and W2 are independent. Although we abstract away from marginal distributions in
(X1 , X2 ), the interpretation of results from this section may sometimes be more natural in terms of S = log R
and V = log W with random location structure, especially when S or V are heavy-tailed.
3.1
Superheavy-tailed variables
Suppose that R or W is superheavy-tailed, i.e., log R or log W is heavy-tailed, respectively. This case naturally
arises when considering random location constructions log R + (log W1 , log W2 ), and we therefore additionally
assume W > 0 so that log Wj , j = 1, 2, are well defined. We will derive general tail dependence characterizations
using classical results where convolution equivalence, and more specifically subexponentiality, is useful.
Proposition 4 (Superheavy-tailed variables).
1. If Flog R ∈ CE0 and F W (x)/F R (x) → c ≥ 0 as x → ∞, then ηX = 1 and
χX =
1 + c χW
> 0.
1+c
7
(13)
Radius R
SHT
a) F W (x)/F R (x) → c
b) F R = o(F W )
RV∞
−αR
a) E(W αR +ε ) < ∞
b) F W ∈ RV∞
−αW
(i) αR > αW
(ii) αR = αW
LWTβR >1
WTβR
Gumbel
Neg. Weibull
additional assumptions
χX
ηX
Prop. 4
χW > 0
χW = 0, F R (x) ≤ CF W∧ (x)
χW = 0, F W∧ = o(F R )
1+cχW
1+c
χW
0
0
1
1
ηW
(14)
P(W > 0) = 1
(6)
1
χW
Prop. 6
(19)
Prop. 8
χW
(15)
1
(20)
Prop. 8
1
χW
0
ηW
(25)
FW
Prop. 6
FW , FW∧ ∈ LWTβR
∈ WTβW , FW∧ ∈ WTβW∧
Prop. 9(1)
Prop. 9(2)
Prop. 9(3)
Table 2: Tail dependence summaries χX and ηX for (X1 , X2 ) = R(W1 , W2 ) with different tail decay rates of
d
the radial variable R > 0 and unconstrained variables W1 = W2 .
Angle W
Radius R
Super-heavy
Reg. varying
Super-heavy
1
1+c
ηX : Prop. 4(2)
*
χX =
Reg. varying
log-Weibull (βW > 1)
Weibull
Neg. Weibull
χX = 1
χX = 1
χX = 1
χX = 1
χX = (6) > 0
χX = (6) > 0
χX = (6) > 0
βR = βW : ηX = (20)
unknown
log-Weibull
(βR > 1)
Weibull
*
αR < αW : χX = (6) > 0
αR = αW : Prop. 6
αW < αR < 2αW :
ηX = αW /αR
αR > 2αW : ηW = 1/2
*
*
*
unknown
ηX = 2−βR /(βR +βW )
Neg. Weibull
*
*
*
*
d
ηX
χX = 0
ηX = 1
χX = 0
ηX = 1
αW + αR
=
2αW + αR
Table 3: The values of χX and ηX for (X1 , X2 ) = R(W1 , W2 ) with W1 , W2 = W independent, with different
tail decay rates of the radial and angular variables. The *’s indicate that multiplication with R does not change
the tail dependence of (W1 , W2 ), i.e., χX = χW = 0 and ηX = ηW = 1/2. The combinations of Weibull and
log-Weibull tails remain open problems.
8
2. If Flog W ∈ CE0 and F R = o(F W ), then χX = χW for χW ≥ 0. If further Flog R ∈ CE0 and
(a) Flog W∧ ∈ CE0 with F R (x)/F W∧ (x) ≤ C for a constant C > 0 as x → ∞, then ηX = ηW ;
(b) F W∧ = o(F R ), then
ηX = lim
x→∞
log F W (x)
,
log F R (x)
(14)
provided the limit exists.
Remark 4. The case c = 0 in point 1 of Proposition 4 is also covered by Remark 3 following Proposition 1,
yielding χX = 1. Proposition 4 extends this case to allow (W1 , W2 ) also to have superheavy tails.
Example 5 (Independence model). In order to illustrate the results of this section, we consider the example
where R, W1 and W2 are independent. In this case F W∧ = (F W )2 , χW = 0 and ηW = 1/2. If Flog R ∈ CE0
and F W (x)/F R (x) → c ≥ 0 as x → ∞, then Proposition 4(1) yields asymptotic dependence in (X1 , X2 ) with
χX = (1 + c)−1 . If Flog R ∈ CE0 and W has a comparable tail, then χX ∈ (0, 1), whilst if W has tail lighter than
superheavy, then χX = 1. On the other hand, if W is superheavy-tailed with Flog W ∈ CE0 , then W∧ is also
superheavy-tailed. If R has lighter tail than W∧ , then F R (x) ≤ CF W∧ (x) for large x with some C > 0, and by
Proposition 4(2a) we have χX = 0 and ηX = 1/2. The case ηX 6= ηW may arise when the tail of W dominates
the tail of R and the tail of R dominates the tail of W∧ . For a concrete example, consider log-Weibull tails
in W and R with F W (exp(x)) ∼ exp(−xβ ) and F R (exp(x)) ∼ exp(−(1 + c)xβ ), where 0 < β, c < 1. Then,
ηX = 1/(1 + c) according to Proposition 4(2b). The first row and column of Table 3 summarize these results.
3.2
Regularly varying variables
In this section we consider the case where R, W or both of them are regularly varying. When R is regularly
varying with index αR > 0 and E(W αR +ε ) < ∞ for some ε > 0, then the tail dependence coefficient χX is as
described in Proposition 1 in Section 2.1. We firstly consider the case where W is regularly varying with index
αW > 0 and R is lighter tailed, i.e., either also regularly varying with αR > αW or even lighter tailed such as
distributions in the Gumbel or negative Weibull domain of attraction. Secondly, we study the case where both
R and W are regularly varying with the same index αW = αR , which turns out to be particularly involved and
which requires additional assumptions.
Proposition 5 (Regularly varying variables with W heavier than R). Let F W ∈ RV∞
−αW , αW ≥ 0, and suppose
∞
that either F R ∈ RV−αR with αR > αW , or R is in the Gumbel or negative Weibull domain of attraction; denote
the latter case by αR = +∞. Then χX = χW and
(
αW /αR , if αR < αW /ηW , ηW = 0 or ηW not defined,
(15)
ηX =
ηW ,
if αR > αW /ηW or αR = +∞.
The case where R and W are regularly varying with the same index α > 0 leads to various scenarios for
the extremal dependence in (X1 , X2 ). Since F R , F W ∈ RV∞
−α is equivalent to Flog R , Flog W ∈ ETα , and ETα
is closed under convolutions, we have that Flog X ∈ ETα (e.g., Lemma 2.5, Watanabe, 2008). Regarding the
following results in Proposition 6, recall that the convolution equivalence property Flog Y ∈ CEα implies finite
moment E(Y α ) < ∞ for a random variable Y .
Proposition 6 (Regularly varying R and W with the same index). Let F R , F W ∈ RV∞
−α with α > 0. Then:
1. ηX = 1.
2. If Flog R ∈ CEα , and if F W (x)/F R (x) → c ≥ 0 as x → ∞, then
χX =
E(W∧α ) + c χW E(Rα )
.
E(W α ) + c E(Rα )
3. If Flog W ∈ CEα and F R = o(F W ), then χX = χW .
4. Let Flog R ∈ ETα,βR with βR ≥ −1 and E(Rα ) = ∞ if βR = −1, and let Flog W ∈ ETα,βW .
(a) If χW > 0 and if either βW > −1 or βW = −1 and E(W α ) = ∞, then χX = χW .
(b) If χW ≥ 0 and if either βW < −1 or βW = −1 < βR and E(W α ) < ∞, then
χX = E(W∧α )/E(W α ).
9
(16)
(c) If βR > −1, βW > −1 and E(W∧α+ε ) < ∞ for some ε > 0, then χX = 0.
Remark 5. Proposition 6 contains certain results of Proposition 4 as a special case when allowing for α = 0.
Proposition 6(2,3) treats the case of convolution-equivalent tails in log R or log W , which are relatively light since
the expectation E(Rα ) or E(W α ) respectively is finite; notice that ETα,β with β < −1 is an important subclass
of CEα , see Lemma 2.3 of Pakes (2004). The tail of R is not dominated by that of W in Proposition 6(2), while
it is dominated in Proposition 6(3). Proposition 6(4) shifts focus to relatively heavy tails in R with E(Rα ) = ∞,
such as the gamma tails of ETα,β with β > −1.
Example 6 (Independence model). We continue Example 5, where now W1 and W2 are independent and
regularly varying with index αW , and R is regularly varying with index αR . We encounter different tail behavior
of (X1 , X2 ) depending on the relation between αR and αW . If αR < αW , then we have asymptotic dependence
with χX given in (6). The same is true in general when W has a lighter tail than R that is not necessarily
regularly varying. By Proposition 5, if αW < αR < 2αW , then (X1 , X2 ) is asymptotically independent with
ηX = αW /αR , and if αR > 2αW , then ηX = 1/2. In general, if R is even lighter tailed, it does not affect the
coefficients χX and ηX . If αR = αW = α, then ηX = 1 and different scenarios for χX can arise depending on
the distributions of R and W , see Proposition 6. We consider α > 0. Since F W∧ ∈ RV∞
−2α in this case, we
have F X∧ ∼ E(W∧α )F R , and then χX = E(W∧α )/c > 0 if F X ∼ cF R with some constant c > 0. In particular,
c = E(W α ) if FW ∈ ETα,βW and FR ∈ ETα,βR with βW < βR and βW < −1. This fills the second row and
column of Table 3.
3.3
Log-Weibull-type variables
In this and the following section we concentrate on radial and angular variables in the Gumbel domain of
attraction. Due to the large variety of distributions in this domain we consider subsets that include the most
commonly used distribution families. First, in this section, we study the case where both R and W are
log-Weibull-tailed; equivalently, log R and log W are Weibull-tailed. We recall that a random variable Y is
log-Weibull-tailed if
(17)
F Y (y) = `(log y)(log y)γ exp −α(log y)β ,
` ∈ RV∞
0 , γ ∈ R, α, β > 0,
and we write FY ∈ LWTβ . The parameter β has the predominant influence on the tail decay rate, with β = 1 if
∞
and only if F Y ∈ RV∞
−α , while β < 1 gives FY ∈ SHT, and β > 1 yields rapid variation of Y , i.e., F Y ∈ RV−∞ ;
these results can be shown by using a binomial series expansion for F Y (tx)/F Y (t). In the following, we denote
the β-parameters of R and W by βR and βW , respectively. The superheavy-tailed case, βR < 1 or βW < 1,
is already covered by Section 3.1, and the case of regularly varying tails with βR = 1 or βW = 1 is treated in
Section 3.2.
In this section we study the remaining case βR > 1 and βW > 1, which encompasses important radial and
angular distributions such as the log-Gaussian ones. Similarly to Section 3.1, it makes sense to log-transform the
random scale construction R(W1 , W2 ) to obtain the random location construction log R + log(W1 , W2 ) where we
can apply convolution-based results. When independent heavy-tailed summands are involved in the convolution
typically only one of the values of summands has a dominant contribution to a high values of the sum, resulting
in relatively simple formulas; see Section 3.1. On the contrary, in the light-tailed setup all summands may
contribute significantly when high values arise in the sum, rendering the tail analysis more intricate. Only
relatively few general results on convolutions with tails lighter than exponential are available in the literature.
The following lemma will be useful for this and the next section.
Lemma 1. Let (W1 , W2 ) be a bivariate random vector with marginal distribution FW such that both FW ∈
WTβW and FW∧ ∈ WTβW∧ , with respective parameters being indexed by W and W∧ .
1. If βW∧ = βW , αW∧ = αW , then
`W∧ (x)xγW∧
,
x→∞ `W (x)xγW
χW = lim
(18)
provided the limit exists.
2. If βW∧ = βW , then ηW = αW /αW∧ .
3. If βW∧ > βW , then log F W = o(log F W∧ ), and ηW is not defined.
Remark 6. The proof of the above lemma is straightforward from (17). It also covers the case where (W1 , W2 )
is a bivariate random vector with W and W∧ are log-Weibull-tailed, since the tail dependence coefficient and
the residual tail dependence coefficient are invariant under monotonic marginal transformations.
10
We consider the set-up where the components R, W and W∧ are log-Weibull-tailed with the same coefficient
β > 1 and a simplified form of the slowly varying function ` by assuming that it is asymptotically constant, i.e.,
`(x) ∼ c > 0.
Proposition 7 (Light-tailed random location with FR ∈ LWTβ , β > 1). Suppose that FR , FW , FW∧ ∈ LWTβ
with possibly different parameters α, γ indexed by the corresponding R, W and W∧ , but where β = βR = βW =
βW∧ > 1. Assume that the slowly varying functions ` behave asymptotically like positive constants. If χW > 0,
then
χX = χW > 0.
(19)
If χW = 0, then χX = 0 and
ηX = ηW ×
1/(β−1)
+ αR
1/(β−1)
1/(β−1)
+ αR
αW∧
!β−1
1/(β−1)
αW
,
(20)
where ηW = αW /αW∧ , and ηX = ηW if αW = αW∧ .
Example 7 (Gaussian factor model). Suppose that log R is univariate standard Gaussian and that log(W1 , W2 )
is bivariate standard Gaussian, independent of R and with Gaussian correlation ρW ∈ (−1, 1]. Then we have
log-Weibull tails with parameters βR = βW = βW∧ = 2, αR = αW = 1/2 and αW∧ = 1/(1 + ρW ) (see
Example 9). Applying (20) gives ηX = ηW × (3 + ρW )/(2(1 + ρW )) = (3 + ρW )/4.
Example 8 (Independence model). As in Examples 5 and 6 we let R, W1 and W2 be independent, and we now
assume that they are log-Weibull-tailed with equal β parameter. Using the independence assumption we observe
that FW∧ ∈ LWTβ with αW∧ = 2αW . Proposition 7 implies that (X1 , X2 ) is asymptotically independent, and
the residual tail dependence coefficient ηX can be calculated by formula (20).
3.4
Weibull-type variables
In this section we consider the case where R and W follow a Weibull-type distribution, a rich class in the
Gumbel domain of attraction. Recall that a variable Y is of Weibull-type, FY ∈ WTβ , if
(21)
F Y (y) = `(y)y γ exp −αy β ,
` ∈ RV∞
0 , γ ∈ R, α, β > 0.
Well-known examples of Weibull-tailed distributions are the Gaussian with β = 2, the gamma with β = 1 or,
more generally, the Weibull where β is called the Weibull index.
For developing useful results, we further assume that, in addition to R and W , W∧ also has a Weibull-type
tail. As previously, we index the corresponding ` functions and the parameters α, γ in (21) by the variable
name. We also recall Lemma 1 concerning the dependence coefficients of the vector (W1 , W2 ).
Proposition 8 (Weibull-type variables). Suppose that FR ∈ WTβR , FW ∈ WTβW and FW∧ ∈ WTβW∧ . We
have the following hierarchy of dependence structures:
1. If βW∧ = βW , αW∧ = αW , γW∧ = γW , then
χX = χW = lim
x→∞
`W∧ (x)
,
`W (x)
(22)
if the limit exists, and ηX = ηW = 1.
2. If βW∧ = βW , αW∧ = αW , γW∧ < γW , then χX = 0 and ηX = ηW = 1.
3. If βW∧ = βW , αW∧ > αW , then χX = 0 and
β /(βR +βW )
ηX = ηWR
=
αW
αW∧
βR /(βR +βW )
.
(23)
4. If βW∧ > βW , then χX = 0 and ηX = ηW = 0.
In all of the cases encompassed by Proposition 8, (X1 , X2 ) and (W1 , W2 ) have the same tail dependence
coefficient χ, which can be positive only in case 1. In all other cases the variables are asymptotically independent,
and only in case 3 the residual tail dependence coefficient η changes under the multiplication of the radial variable
R. Since βR /(βR +βW ) ∈ (0, 1), this multiplication always leads to an increase in dependence, that is, ηX > ηW .
11
Example 9 (Gaussian scale mixtures). To illustrate the most interesting case 3 in the above proposition we
consider (W1 , W2 ) following a bivariate normal distribution with standardized margins and correlation ρW . We
have that
F W (x) ∼ rW (x) exp(−x2 /2),
F W∧ (x) ∼ rW∧ (x) exp{−x2 /(1 + ρW )},
where the tail distribution of the minimum follows from bounds on the multivariate Mills ratio, see e.g. Hashorva
and Hüsler (2003), and rW and rW∧ regularly varying functions of order −1 and −2 respectively. Therefore,
ηW = (1 + ρW )/2 and Proposition 8 yields
ηX =
β /(β +2)
ηWR R
=
1 + ρW
2
βR /(βR +2)
,
confirming Theorem 2 in Huser et al. (2017).
Example 10 (Independence model). We continue the example where R, W1 and W2 are independent, and they
are now assumed to be Weibull-tailed, where the latter two have the same distribution as W . By independence,
W∧ is also Weibull-tailed with βW∧ = βW and αW∧ = 2αW . Therefore, the third part of Proposition 8 entails
that ηX = 2−βR /(βR +βW ) . Note that this expression tends to 1/2 if βR /βW tends to infinity such that the tail
of W dominates strongly with respect to that of R; if βR /βW tends to 0, then ηX tends to 1.
3.5
Variables in the negative Weibull domain of attraction
The remaining cases are those where R, W or both of them have finite upper endpoint and are in the negative
Weibull domain of attraction. We recall that a variable Y is in the negative Weibull domain of attraction if
F Y (y ? − s) = `(s)sα ,
` ∈ RV00 , α > 0.
(24)
The case where R is superheavy-tailed or regularly varying and W satisfies (24) has been covered in part 1 of
Proposition 4 and Proposition 1, respectively. On the other hand, the case where the tail of R satisfies (24) and
W is superheavy-tailed or regularly varying is treated by part 2 of Proposition 4 and Proposition 5, respectively.
It remains to study the situation where one of R or W is of form (24), and the other is in the Gumbel domain
of attraction as defined in (9). In this section we focus on the case where W∧ has the same upper endpoint as
W ; for a more detailed study where W∧ can have a smaller upper endpoint, and W may have a point mass on
its upper endpoint, see Section 2.
Proposition 9 (Variables in the negative Weibull domain of attraction).
1. Suppose that R is in the Gumbel MDA with upper endpoint r? and that W and W∧ satisfy (24) with parameters αW and αW∧ , respectively. Then χX = χW and ηX = 1.
2. Let R satisfy (24) and let W and W∧ be in the Gumbel MDA with equal upper endpoint w? ∈ (0, ∞] and
auxiliary functions bW and bW∧ , such that limx→w? bW (x)/bW∧ (x) exists. Then χX = χW and ηX = ηW .
3. Let R, W and W∧ all satisfy (24) with endpoints r? , w? , w? and parameters αR , αW and αW∧ , respectively.
If αW∧ = αW then χX = χW and ηX = 1. If αW∧ > αW then χX = χW = 0 and
ηX =
αW + αR
αW
>
= ηW .
αW∧ + αR
αW∧
(25)
d
Example 11 (Independence model). Let us consider again the simple model where R, W1 = W2 are independent, with FR , FW satisfying (24) with parameters αR , αW . Clearly, FW∧ also satisfies (24), with parameter
αW∧ = 2αW . The third part of Proposition 9 shows that
ηX =
αW + αR
∈ (1/2, 1),
2αW + αR
hence by varying the parameters αR , αW > 0 we can attain the whole range of residual tail dependence coefficients related to positive association.
4
Literature review and examples
Here we present an overview of related literature, detailing how existing examples and results fit into the
framework of this paper.
12
Elliptical copulas Let Σ be a positive-definite covariance matrix with Cholesky decomposition Σ = AAT ,
and (U1 , U2 ) be uniformly distributed on the L2 sphere {(w1 , w2 ) : (w1 , w2 )T (w1 , w2 ) = 1}. Then
(X1 , X2 ) = RA(U1 , U2 )T
has an elliptical distribution for any R > 0 called the generator. Therefore (W1 , W2 )T = A(U1 , U2 )T lies on the
Mahalanobis sphere W = {(w1 , w2 ) : (w1 , w2 )T Σ−1 (w1 , w2 ) = 1}, and the extremal dependence in the upper
right orthant is unchanged by taking (Wj )+ = max(Wj , 0). It is well known that (X1 , X2 ) is asymptotically
dependent if and only if R is in the Fréchet MDA (Theorem 4.3 Hult and Lindskog, 2002). In that case,
the tail dependence coefficient χX is given by (6), with Wj replaced by (Wj )+ ; see also Opitz (2013). For R
in the Gumbel MDA, the scaling condition on ν such that τ (w) ∈ [0, 1] yields Σ with diagonal elements 1,
off-diagonal elements ρ ∈ (−1, 1), and residual tail dependence coefficient is given by Proposition 2(1) with
ζ = τ (1/2) = {(1 + ρ)/2}1/2 . Hashorva (2010) details calculation of ηX assuming R to be in the Gumbel MDA,
providing an alternative perspective on the derivation. The spatial model of Huser et al. (2017) is elliptical,
and hence is also covered by this case.
Example (Gaussian). The Gaussian distribution arises when F R (r) = e−r
ζ 2 = (1 + ρ)/2.
2
/2
, so that by Corollary 1, ηX =
Archimedean and Liouville copulas Archimedean (respectively Liouville) copulas arise as the survival
copula when (W1 , W2 ) is uniformly (respectively Dirichlet) distributed on the positive part of the L1 sphere
W = {(w1 , w2 ) ∈ [0, 1]2 : w1 +w2 = 1}, and R > 0. That is, (X1 , X2 ) = R(W1 , W2 ) has an inverted Archimedean
or Liouville copula, whilst the Archimedean or Liouville copula itself is that of (t(X1 ), t(X2 )), for a monotonic
strictly decreasing transformation t. By taking t(x) = 1/x, we obtain 1/(X1 , X2 ) = (X̃1 , X̃2 ) = R̃(W̃1 , W̃2 ), so
Archimedean copulas have a random scale representation with (W̃1 , W̃2 ) constrained by functional dependence
that is not represented by a norm.
Archimedean copulas are typically defined in terms of a non-increasing continuous generator function ψ :
[0, ∞) → [0, 1], such that C(u1 , u2 ) = ψ(ψ −1 (u1 ) + ψ −1 (u2 )). The link between ψ and the variable R ∼ FR is
given in equation (3.3) of McNeil and Nešlehová (2009); for d = 2 this is
F R (r) = ψ(r) − rψ 0 (r+ ),
r > 0,
(26)
where ψ 0 (r+ ) denotes the right-hand derivative of ψ.
Archimedean copulas can be viewed as a special case of Liouville copulas, whose dependence properties are
studied in Belzile and Nešlehová (2018). For (X1 , X2 ), their Theorem 1 states that R in the Fréchet MDA leads
to asymptotic dependence, whilst the Gumbel and negative Weibull maximum domains of attraction lead to
asymptotic independence. The exponent function given in their Theorem 1 matches equation (7). In their Theorem 2, Belzile and Nešlehová (2018) consider the extremal dependence properties of 1/(X1 , X2 ) = R̃(W̃1 , W̃2 ),
i.e., the Liouville copula itself. Since the reciprocal of Dirichlet random variables have regularly varying tails,
this links with Proposition 5 which states that asymptotic independence arises if (W̃1 , W̃2 ) themselves are
asymptotically independent and heavier-tailed than R. Proposition 6(4c) is relevant if R̃ and W̃ are regularly
varying with the same index.
Example (Gumbel and inverted Gumbel copulas). The Gumbel (also known as the logistic) Archimedean
θ
θ
copula arises when ψ(x) = e−x , θ ∈ (0, 1]. By (26), F R (r) = ψ(r) = e−r [1 + θrθ ] ∈ WTθ . The copula
of (X1 , X2 ) is the asymptotically independent inverted Gumbel copula (Ledford and Tawn, 1997). We have
ζ = τ (1/2) = 1/2 and so ηX = 2−θ by Corollary 1.
−θ
The Gumbel copula is that of 1/(X1 , X2 ) = R̃(W̃1 , W̃2 ), with FR̃ (r) = e−r [1 + θr−θ ], so F R̃ ∈ RV∞
−θ . The
dependence structure follows from Proposition 1 for θ < 1 since E(W̃ θ ) < ∞. Noting that W̃∧ is a bounded
random variable, χX = 0 for θ = 1, as given by Proposition 6(4c). In fact, for θ = 1, the copula is the
independence copula.
Multivariate (ρ-)Pareto copulas Let ρ : (0, ∞)2 → (0, ∞) be a positive homogeneous function. Multivariate ρ-Pareto copulas arise when F R (r) = r−1 , i.e. standard Pareto, and the random vector (W1 , W2 ) is
concentrated on W = {(w1 , w2 ) ∈ R2+ : ρ(w1 , w2 ) = 1} with marginals satisfying E(W ) < ∞ (Dombry and
Ribatet, 2015). The case of ρ(x1 , x2 ) = max(x1 , x2 ) leads to the multivariate Pareto copula associated to
multivariate generalized Pareto distributions (Rootzén and Tajvidi, 2006; Ferreira and de Haan, 2014; Rootzén
et al., 2018). Such copulas are asymptotically dependent, except for the case outlined in Section 2.1, with χX
given by (6). Although we have focused on norms and ρ need not be convex, there is nothing in Proposition 1
requiring this.
13
Example (Bivariate Pareto copula associated to the Gumbel copula). Since the Gumbel copula is a max-stable
distribution, it has an associated Pareto copula. If Z has density fZ (z) = h(z) max(z, 1 − z)21−θ , where h is
given by (8), then (X1 , X2 ) = R(W1 , W2 ) = R(Z, 1−Z)/ max(Z, 1−Z) with F R (r) = r−1 leads to the associated
bivariate Pareto copula. The distribution function of (X1 , X2 ) is
P(X1 ≤ x1 , X2 ≤ x2 ) =
−1/θ
where V (x1 , x2 ) = (x1
V (min(x1 , 1), min(x2 , 1)) − V (x1 , x2 )
,
V (1, 1)
−1/θ θ
+ x2
) is the exponent function for the logistic / Gumbel distribution.
Model of Wadsworth et al. (2017) They consider the copula induced by taking R to be generalized
−1/λ
Pareto, F R (r) = (1 + λr)+ , and W = {(w1 , w2 ) ∈ [0, 1]2 : k(w1 , w2 )k∗ = 1} where k · k∗ is a symmetric norm
subject to certain restrictions. These restrictions mean that λ ≤ 0 corresponds to asymptotic independence; the
residual tail dependence coefficient ηX is as given in Proposition 2 for λ = 0 with ζ = τ (1/2) = k(1, 1)k−1
∗ , and
Proposition 3 for λ < 0. Here we note that if the norm ν has certain shapes that were excluded in Wadsworth
et al. (2017), asymptotic dependence is possible for λ ≤ 0. When R is in the Fréchet MDA (λ > 0) then
asymptotic dependence holds with χX given by (6).
Model of Krupskii et al. (2016) They consider location mixtures of Gaussian distributions, corresponding
to scale mixtures of log-Gaussian distributions. According to their Proposition 1, asymptotic dependence occurs
when the location variable is of exponential type, i.e. the scale is of Pareto type; the given χX can then be
obtained via (6). When the location is Weibull-tailed but with shape in (0, 1), the scale is superheavy-tailed, with
F R ∈ RV∞
0 , and perfect extremal dependence (χX = 1) arises, as noted in Remark 3 following Proposition 1.
When the random location is Weibull-tailed with shape in (1, ∞) then the random scale R is in the Gumbel
MDA and asymptotic independence arises. If F log R ∈ WT2 has the same Weibull coefficient 2 as the standard
Gaussian log W and as the minimum of log W1 and log W2 (provided that their linear correlation is in (−1, 1]),
then we can apply Proposition 7 to calculate the value of ηX given as
ηX = ηW
αW∧ + αR
1 + ρ (1 + ρ)−1 + αR
1 + (1 + ρ)αR
=
=
,
αW + αR
2
1/2 + αR
1 + 2αR
which extends the results of Krupskii et al. (2016). Specifically, with standard Gaussian log R we get ηX =
(3 + ρ)/4, see Example 7.
Model of Huser and Wadsworth (2018) They consider scale mixtures of asymptotically independent
vectors where both R and W have Pareto margins with different shape parameters. Asymptotic dependence
arises when R is heavier tailed; χX is then given by (6), whilst asymptotic independence arises when W is
heavier tailed and ηX is given by (15). When R and W have the same shape parameter, the assumption ηW < 1
implies that E(W∧α+ε ) < ∞ for some ε > 0, hence Proposition 6 (4c) gives asymptotic independence.
Various other articles also focus on polar or scale-mixture representations. Hashorva (2012) examines the
extremal behavior of scale mixtures when R is in the Gumbel MDA. He considers both one- and two-dimensional
W, both with similarities and differences to the set-up herein. For one-dimensional W, he supposes a functional
constraint of the form (W1 , W2 ) = (W, ρW + z ∗ (W )) for measurable z ∗ : [0, 1] → (0, ∞); ρ ∈ (−1, 1). (In fact,
the specification also allows for negative components, but we focus here on the positive part.) Constraints to
certain norm spheres, such as the Mahalanobis or Lp norm, could be written in this way, however examples
such as the L∞ norm do not admit this representation. Where the representations overlap, our results coincide
(e.g., Section 4.3 of Hashorva (2012)). In the case of two-dimensional W, W1 is assumed bounded, whilst W∧
is in the negative Weibull MDA. Although the framework of Hashorva (2012) does not assume the symmetry of
our framework, there are nonetheless some connections between the results in our Section 3.5 and that paper.
Nolde (2014) provides an interpretation of extremal dependence in terms of a gauge function (see also
Balkema and Nolde (2010)), which, loosely speaking, corresponds to level sets of the density in light-tailed
margins. Their main result (Theorem 2.1) is presented in terms of Weibull-type margins, such that − log F X ∈
∞
RV∞
δ , δ > 0; in terms of Section 2, this corresponds to − log F R ∈ RVδ . By noting that where the density of
R, fR , exists, the joint density of (X1 , X2 ) is
f(X1 ,X2 ) (x1 , x2 ) = fR (ν(x1 , x2 ))fW (x1 /(x1 + x2 ))
ν(x1 , x2 )
,
(x1 + x2 )2
the gauge function of (X1 , X2 ) is obtained as ν when − log F R ∈ RV∞
δ , using Proposition 3.1 therein. We found
that ηX = ζ δ , with ζ = ν(1, 1)−1 , precisely as given in Nolde (2014).
14
Various papers focus on extremal dependence arising from certain types of polar representation, but from a
conditional extremes perspective (Heffernan and Tawn, 2004; Heffernan and Resnick, 2007). This is different to
our focus; here we examine the extremal dependence as both variables grow at the same rate. In the conditional
approach, different rates of growth may be required in the different components. Articles taking this perspective
include Abdous et al. (2005) in the context of elliptical copulas, whilst Fougères and Soulier (2010) and Seifert
(2014) consider the constrained W case, with R in the Gumbel MDA.
5
New examples
We present two new constructions that have the desirable property of smoothly interpolating between asymptotic dependence and asymptotic independence, whilst yielding non-trivial structures within each class. By
smoothness, we mean that the transition between classes occurs at an interior point, θ0 , of the parameter
space Θ, and, assuming increasing dependence with θ, limθ→θ0+ χX = 0, limθ→θ0− ηX = 1. To our knowledge, the only other models in the literature with this behavior are (i) that of Wadsworth et al. (2017), where
−1/λ
ν(x, y) = max(x, y) and F R (x) = (1 + λx)+ , λ ∈ R, and (ii) that of Huser and Wadsworth (2018) where
F R (x) = x−1/δ , F W (x) = x−1/(1−δ) , δ ∈ (0, 1), and ηW < 1. The first example is constructed using constrained
(W1 , W2 ) (Section 2), where the required ingredients are F R , ν, and FZ , whilst the second uses unconstrained
(W1 , W2 ) (Section 3) with ingredients F R , FW and the dependence structure of (W1 , W2 ).
5.1
Model 1
In Propositions 2 and 3, it was demonstrated how the shape of ν affects the tail dependence of (X1 , X2 ) when R
is in the Gumbel or negative Weibull MDA. In Example 2, a particular norm that yields asymptotic dependence
was given; here we extend the parameterization of this norm and use our results to present a new asymptotically
(in)dependent copula. Since the cases where R has finite upper endpoint r? < ∞ lead to undefined ηX , we
focus on r? = ∞.
Proposition 10. Let R be in the Gumbel MDA with r? = ∞ and let ν(x, y) = θ max(x, y) + (1 − θ) min(x, y),
θ ≥ 1/2. Then for (W1 , W2 ) defined through (11) and Z satisfying the conditions in Section 2.2,
χX = max(2(θ − 1)/(2θ − 1), 0)
ηX = lim log F R (x)/ log F R (x/ min(θ, 1)).
x→∞
We note that if − log F R ∈ RVδ , for example, then we have a continuous parametric family exhibiting
asymptotic dependence for θ > 1 with χX = 2(θ − 1)/(2θ − 1), and asymptotic independence for θ ≤ 1 with
ηX = θ δ .
To make things concrete, we propose the following model.
Model 1.
F R (r) = e−r
δ
ν(x, y) = θ max(x, y) + (1 − θ) min(x, y), θ ≥ 1/2
Z ∼ Beta(α, α).
The set of models defined in Proposition 10, exemplified by Model 1, has some rather interesting behavior
in the extremes. Whilst the limiting quantities χX , ηX are given in Proposition 10, the subasymptotic behavior
of (X1 , X2 ), in particular the behavior of the slowly varying function ` in (2), is not prescribed by any of the
propositions in this paper. Combining equations (2) and (3), define
−1
−1
χX (q) = Pr(X1 ≥ FX
(q), X1 ≥ FX
(q))/(1 − q) = `(1 − q)(1 − q)1/ηX −1 ,
1
1
so that for ηX = 1, χX (q) = `(1 − q). For Model 1 we find that χX (q) is not necessarily monotonic, and
may decrease before increasing to a positive limit value. Figure 5.1 shows χX (q) plotted against q for various
parameterizations of the model. This non-monotonic behavior is seemingly uncommon; to our knowledge there
are no well-known theoretical examples of this.
5.2
Model 2
The following proposition collates results from Propositions 1 and 9, and provides a general principle for constructing new dependence models permitting both asymptotic dependence and asymptotic independence.
Proposition 11. Let R be in the MDA of a generalized extreme value distribution with shape parameter ξ ∈ R,
d
d
and let (W1 , W2 ) with W1 = W2 = W have χW = 0, well-defined ηW ∈ (0, 1), and F W (w? − ·) ∈ RV0αW ,
αW > 0. Then
15
θ=1,α=1
θ = 1.25 , α = 1
10−3
10−4
10−5
10−6
10−7
0.5
χX(q)
0.2
10−1
10−2
10−3
10−4
10−5
10−6
10−7
10−1
10−5
θ = 0.75 , δ = 1
θ=1,δ=1
θ = 1.25 , δ = 1
10−4
10−5
10−6
10−7
10−6
10−7
0.5
10−7
0.3
χX(q)
0.2
0.0
0.1
0.1
0.0
10−3
10−6
0.4
0.4
0.3
χX(q)
0.2
0.4
0.3
0.2
0.1
χX(q)
10−4
1−q
0.0
10−2
10−3
1−q
α = 0.5
α=1
α = 1.5
α=2
α=3
10−1
10−2
1−q
0.5
10−2
0.0
0.1
0.1
0.0
0.0
10−1
0.3
0.4
0.4
0.3
χX(q)
χX(q)
0.1
0.2
0.3
δ = 0.5
δ=1
δ = 1.5
δ=2
δ=3
0.2
0.4
0.5
θ = 0.75 , α = 1
10−1
10−2
10−3
1−q
10−4
10−5
10−6
10−7
10−1
1−q
10−2
10−3
10−4
10−5
1−q
Figure 3: Theoretical χX (q) for Model 1 plotted against 1 − q on a logarithmic scale for q ∈ [1 − 10−1 , 1 − 10−7 ].
Different columns show different values of θ; thick horizontal lines show the true limiting values of χX =
(0, 0, 1/3) (left-right). Top row: δ varies within a panel; bottom row: α varies within a panel.
h
i
1/ξ
1. For ξ > 0, χX = E W∧ /E(W 1/ξ ), ηX = 1,
2. For ξ = 0, χX = 0, ηX = 1,
3. For ξ < 0, χX = 0, ηX =
1 − ξαW
.
1 − ξαW /ηW
The model construction opportunities from Proposition 11 are quite varied; specifically taking F R that
permits all three tail behaviors produces a flexible range of models spanning the two dependence classes. We
therefore propose the following concrete model, based on our running independence example.
Model 2.
−1/ξ
F R (r) = (1 + ξx)+
, ξ∈R
W1 ⊥⊥ W2
W ∼ Beta(α, α).
If we specifically take α = 1, so that W ∼ Unif(0, 1), then one can explicitly calculate χX and VX as well as
ηX . By Proposition 11, for ξ < 0, ηX = (1 − ξ)/(1 − 2ξ) with limξ→0− ηX = 1, and χX = 0; for ξ = 0, χX = 0,
ηX = 1, and for ξ > 0, χX = 2ξ/(2ξ + 1), whilst
−1
VX (x1 , x2 ) = min(x1 , x2 )
1
+
2ξ + 1
min(x1 , x2 )
max(x1 , x2 )
ξ
max(x1 , x2 )−1 .
(27)
−1
−1
The limits of (27) as ξ → 0 and ξ → ∞ are x−1
, corresponding to independence and
1 + x2 and min(x1 , x2 )
perfect dependence (e.g. Beirlant et al., 2004, Ch. 8). Figure 5.2 displays the function χX (q) for Model 2 across
a range of different α and ξ values.
6
Discussion
The paper studies the extremal properties of copulas (X1 , X2 ) = R(W1 , W2 ) and determines the coefficients χX
for asymptotic dependence and ηX for asymptotic independence, respectively.
In the case of constrained (W1 , W2 ) defined on the sphere corresponding to some norm ν, in Section 2,
the extremal dependence is characterized in terms of the tail heaviness of R. Classical results on multivariate
16
10−1
10−2
10−3
10−4
1−q
10−5
10−6
10−7
0.7
0.6
χX(q)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.5
0.4
χX(q)
0.1
0.0
0.0
0.1
0.2
0.3
0.4
0.3
χX(q)
0.5
0.6
ξ = 0.2
ξ = 0.1
ξ=0
ξ = − 0.1
ξ = − 0.2
0.2
α=2
0.7
α=1
0.7
α = 0.5
10−1
10−2
10−3
10−4
10−5
10−6
10−7
10−1
10−2
10−3
1−q
10−4
10−5
10−6
10−7
1−q
Figure 4: Theoretical χX (q) for Model 2 plotted against 1 − q on a logarithmic scale for q ∈ [1 − 10−1 , 1 − 10−7 ].
Different columns show different values of α; dashed horizontal lines show the true limiting values of χX , which
depends on ξ (and is zero for ξ ≤ 0).
Pareto copulas are recovered for regularly varying R, whereas new structures are obtained for distributions of
R with light tails or finite upper endpoint. For the Gumbel domain of attraction, in particular, we get a large
variety of behaviors for asymptotically independent (X1 , X2 ) that strongly depend on the auxiliary function b
of R and the shape of the ν-sphere. This extends the results of Wadsworth et al. (2017) who considered only
the exponential distribution in this class.
For unconstrained distributions of both R and W , Section 3 formalizes the general intuition that heavier
tails of R introduce more additional dependence in (X1 , X2 ). The results summarized in Table 3 for the special
case of the independence model allow for several conclusions. The most interesting (and involved) situations
figure along the main diagonal where R and W have similar tail behavior. Above this main diagonal, R is so
heavy that it mostly dominates the extremal dependence in (X1 , X2 ). On the other hand, below the diagonal,
R is too light tailed, relatively to W , to have an impact on the tail dependence coefficients χX and ηX . Similar
observations hold true for the more general case of arbitrary dependence in (W1 , W2 ) summarized in Table 2.
We note that there is a clear overlap between the results obtained in Sections 2 and 3. If one considers χW as
derived from the shape of W, then many results in Section 2 are obtained from Section 3, just as Proposition 1
is relevant in both sections. However, the separate treatment seems justified on the grounds of the importance
of such constructions, and the additional insight gained in focusing on the shape of W.
The above results provide a general and unifying framework to analyze bivariate extremal dependence, and
Section 4 shows that they cover many of the existing examples in the copula and the extreme value literature.
Most importantly, combining the insights from different sections enables the construction of numerous new
statistical models that smoothly interpolate between asymptotic dependence and independence; see Section 5
for two instances. Although our focus was on dependence, knowledge on how the marginal scales of R and W
and the dependence properties of (W1 , W2 ) influence the dependence of (X1 , X2 ) makes it easier to construct
models (X1 , X2 ) that naturally accommodate both marginal distributions and dependence of multivariate data.
Such modeling avoids what may be construed as the artificial separation of modeling of margins and dependence
known as copula modeling, which has been strongly criticized by some authors since it may lack interpretation
with respect to the data generating mechanism (Mikosch, 2006). For example, in factor constructions based on
independent random variables, such as the ones with independent W1 and W2 discussed throughout, our results
give guidance on the relative tail heaviness of R with respect to (W1 , W2 ) necessary to transition from asymptotic
independence to asymptotic dependence in (X1 , X2 ), and both heavy- or light-tailed marginal distributions are
possible by considering the distribution of either (X1 , X2 ) or log(X1 , X2 ) as a model for data.
d
d
In Sections 2 and 3, we often considered the simplification W1 = W2 , yielding X1 = X2 , which allows the
coefficients χX and ηX to be calculated without reference to marginal quantile functions. A weaker sufficient
condition for this is F X1 (x) ∼ F X2 (x) as x → x? , with x? a common upper end point. To see this sufficiency,
−1
−1
−1
−1
(q), FX
(q)}, and note that
define xq = min{FX
(q), FX
(q)}, xq = max{FX
1
2
1
2
−1
−1
P(X1 ≥ FX
(q), X2 ≥ FX
(q))
P(X1 ≥ xq , X2 ≥ xq )
P(X1 ≥ xq , X2 ≥ xq )
1
2
≤
≤
,
1−q
max{F X1 (xq ), F X2 (xq )}
min{F X1 (xq ), F X2 (xq )}
(28)
where max{F X1 (xq ), F X2 (xq )} = min{F X1 (xq ), F X2 (xq )} = 1−q. Consequently, the tail dependence coefficient
χX of (X1 , X2 ), if it exists, is bounded between the limit superior of the left-hand side and the limit inferior of
the right-hand side in (28), respectively, for q → 1. Whilst these bounds hold in general for common upper end
point, they deliver the precise coefficient χX only if F X1 (x) ∼ F X2 (x), x → x? or both limits are zero. Similar
arguments apply to the residual tail dependence coefficient ηX , where the corresponding bounds determine ηX
17
under the weaker requirement log F X1 (x) ∼ log F X2 (x), x → x? .
Whilst our focus has been on the coefficients χX and ηX , we note that there are important aspects of the
dependence structure that are not described by these coefficients. For example, in Section 2, we found that
when R was in the Gumbel or negative Weibull MDA, χX and ηX depended only on the shape of ν and the
distribution of R, but not at all on the distribution of Z. Nonetheless, the latter plays an important role in the
behavior of the slowly varying function ` in (3), which was exemplified in Figure 5.1.
7
Proofs
We recall Breiman’s lemma, which will be useful in several contexts.
Lemma 2 (Breiman’s lemma, see Breiman (1965); Cline and Samorodnitsky (1994) and Pakes (2004), Lemma 2.1).
Suppose X ∼ F , Y ∼ G are independent random variables. If F ∈ RV∞
−α with α ≥ 0 and Y ≥ 0 with
E(Y α+ε ) < ∞ for some ε > 0, then
F XY (x) ∼ E (Y α ) F (x),
x → ∞.
Equivalently, if F ∈ ETα and E(e(α+)Y ) < ∞, then F X+Y (x) = F ? G(x) ∼ E eαY
F (x).
We also note that in the case of common marginal distributions FX with upper endpoint x? , the tail dependence
coefficient in (2) is χX = limx→x? F X∧ (x)/F X (x), whilst the residual tail dependence coefficient in (3) is given
by ηX = limx→x? log F X (x)/ log F X∧ (x).
7.1
Proofs for Section 2
Proof of Proposition 1. Since E(Wjα+ε ) < ∞, j = 1, 2, Lemma 2 gives
F Xj (x) ∼ E(Wjα )F R (x),
x → ∞,
(29)
−1
−1
so that F Xj ∈ RV∞
−α . Now consider the quantile functions of Xj and R; denote these by FXj (q), FR (q).
Suppose firstly that α > 0. By taking the reciprocal of relation (29), and using Proposition 2.6 (vi) of Resnick
(2007), we have
−1
FX
(q) ∼ FR−1 (q)E(Wjα )1/α ,
j
q → 1.
(30)
Consider now
P(X1 ≥
−1
FX
(q), X2
1
≥
−1
FX
(q))
2
= P R min
W1
W2
[1 + o(1)],
[1 + o(1)]
α
1/α
E(W1 )
E(W2α )1/α
≥
FR−1 (q)
.
Since E(Wjα+ε ) < ∞, j = 1, 2, and min(W1 /E(W1α )1/α , W2 /E(W2α )1/α )α+ε ≤ [Wj /E(Wjα )1/α ]α+ε for j = 1, 2
we have E[min(W1 /E(W1α )1/α , W2 /E(W2α )1/α )α+ε ] < ∞. By dominated convergence we therefore also have
α
W2
W1
[1 + o(1)],
[1 + o(1)]
→ E[min(W1α /E(W1α ), W2α /E(W2α ))],
E min
E(W1α )1/α
E(W2α )1/α
and so as q → 1
W1
W2
W1α
W2α
−1
,
F R (FR−1 (q)),
P R min
[1
+
o(1)],
[1
+
o(1)]
≥
F
(q)
∼
E
min
R
E(W1α ) E(W2α )
E(W1α )1/α
E(W2α )1/α
from which the result follows. For α = 0 we have F Xj (x) ∼ F R (x) ∈ RV∞
0 , as well as F X∧ (x) ∼ F R (x). Using
the bounds in (28) and taking limits, we get χX = 1. As noted after the proposition, the conditions ensure
χX > 0 and hence ηX = 1.
Before proceeding to the proofs of Propositions 2 and 3, Lemma 3 provides further detail about the function
τ , whilst Lemma 4 details the behavior of W and W∧ .
Lemma 3. If 1 − τ1 (b1 − ·) ∈ RV01/γ1 , for γ1 > 0, then b1 − τ1−1 (1 − ·) ∈ RV01/γ1 , and γ1 ≤ 1. Similarly, if
1 − τ2 (b2 + ·) ∈ RV01/γ2 , for γ2 > 0, then τ2−1 (1 − ·) − b2 ∈ RV0γ2 and γ2 ≤ 1.
18
Proof. The argument is similar for both cases, so we focus on the first one. Let g(t) = 1 − τ1 (b1 − t), which is
decreasing as t → 0 and invertible, with g −1 (s) = b1 − τ1−1 (1 − s). Since g ∈ RV01/γ1 , then g −1 ∈ RV0γ1 (Resnick
(2007, Proposition 2.6(v)), adapting to regular variation at zero). Now make the left-sided Taylor expansion
τ1 (b1 − t) = τ1 (b1 ) − tτ10 (b1− ) + O(t2 ),
and note that τ1 (b1 ) = 1. Consequently, 1 − τ1 (b1 − t) = tτ10 (b1− ) + O(t2 ), and so finiteness of τ10 (b1− ) will
imply that the index of regular variation of g is at least 1. Define the convex function µ : (0, ∞) → (0, ∞) by
µ(x) = ν(1, x), so that τ (x) = 1/µ(1/x − 1). Since τ1 is increasing on (0, b1 ), µ is increasing on ((1 − b1 )/b1 , ∞).
We have
τ 0 (x− ) = x−2 µ0 ((1/x − 1)+ )/µ(1/x − 1)2 ,
0
2
and so τ10 (b1− ) = b−2
1 µ ([(1 − b1 )/b1 ]+ )/µ((1 − b1 )/b1 ) . For h ∈ (0, 1), convexity entails
µ((1 − b1 )/b1 + h) ≤ hµ((1 − b1 )/b1 + 1) + (1 − h)µ((1 − b1 )/b1 ),
and so
µ((1 − b1 )/b1 + h) − µ((1 − b1 )/b1 )
≤ µ((1 − b1 )/b1 + 1) − µ((1 − b1 )/b1 ) < ∞.
h
Hence µ0 ([(1 − b1 )/b1 ]+ ) < ∞, giving τ10 (b1− ) < ∞. Thus the index of regular variation of g is at least 1.
0≤
Lemma 4. Assume (Z1), (N1) and (N2).
1. If b1 = b2 = 1 then F W (1 − ·) ∈ RV0αW with αW = αZ γ.
2. If b1 < 1 then
a) F W (1 − ·) − P(W = 1) ∈ RV0αW with αW = γ.
b) In particular, if b1 < 1, τ10 (b1− ), τ20 (b2+ ) 6= 0, then F W (1 − s) − P(W = 1) = s`(s), with ` satisfying
lim `(s) = fZ (b1 )/τ10 (b1− ) − fZ (b2 )/τ20 (b2+ ) ∈ (0, ∞),
s→0
(31)
where for b2 = 1 we put τ20 (b2+ ) = ∞. In this case, F W (1−·)−P(W = 1) ∈ RV0αW with αW = γ = 1.
3. F W∧ (ζ(1 − ·)) ∈ RV01 , with slowly varying function `∧ satisfying
lim `∧ (s) = 2fZ (1/2)ζ/τ10 (1/2− ).
s→0
(32)
Proof of Lemma 4. 1. If b1 = b2 = 1 then P(W = 1) = 0 and F W (1 − s) = F Z (τ −1 (1 − s)). By assumption,
F Z (τ −1 (1 − s)) = `Z (1 − τ −1 (1 − s))(1 − τ −1 (1 − s))αZ , where `Z ∈ RV00 . Since 1 − τ −1 (1 − s) ∈ RV0γ (Lemma 3)
with limit zero, results on composition of regularly varying functions (Resnick, 2007, Proposition 2.6 (iv)) implies the result, with αW = αZ γ.
2.a) If b1 < 1 then
F W (1 − s) − P(W = 1) = P(Z ∈ [τ1−1 (1 − s), b1 )) + P(Z ∈ (b2 , τ2−1 (1 − s)]),
(33)
and since Z has a Lebesgue density,
P(Z ∈ [τ1−1 (1 − s), b1 )) + P(Z ∈ (b2 , τ2−1 (1 − s))) ∼ fZ (b1 )(b1 − τ1−1 (1 − s)) + fZ (b2 )(τ2−1 (1 − s) − b2 ) ∈ RV0γ ,
(34)
again using Lemma 3.
2.b) A left-sided Taylor expansion of τ1−1 about 1 gives
τ1−1 (1 − s) = b1 − (τ1−1 )0 (1− )s + O(s2 ),
where O(s2 )/s uniformly tends to 0 as s → 0, and similarly we can make a right-sided expansion for τ2−1 (1 −
s). Hence, using (33) and (34), F W (1 − s) − P(W = 1) = s`(s) with lims→0 `(s) = fZ (b1 )(τ1−1 )0 (1− ) −
fZ (b2 )(τ2−1 )0 (1− ). Noting the link 1/τ10 (b1− ) = (τ1−1 )0 (1− ), similarly for 1/τ20 (b2+ ), gives (31).
3. For the final part, we have
F W∧ (ζ(1 − s)) = F Z (τ1−1 (ζ(1 − s))) − F Z (1 − τ1−1 (ζ(1 − s))) ∼ fZ (1/2) 1 − 2τ1−1 (ζ(1 − s)) .
19
Again by left-sided Taylor expansion of τ1−1 about ζ = τ (1/2), we have
τ1−1 (ζ(1 − s)) = 1/2 − (τ1−1 )0 (ζ− )ζs + O(s2 ),
and so we obtain F Z (τ1−1 (ζ(1 − s))) − F Z (1 − τ1−1 (ζ(1 − s))) = s`∧ (s) with lims→0 `∧ (s) = 2fZ (1/2)ζ(τ1−1 )0 (ζ− ).
Noting again that 1/τ10 (1/2− ) = (τ1−1 )0 (ζ− ), we arrive at (32).
For the proofs of Propositions 2 and 3 we will repeatedly use Theorem 3.1 of Hashorva et al. (2010), and so
restate the relevant components here as a lemma.
Lemma 5 (Hashorva et al. (2010)). Let Y = RS, where R ∈ (0, r? ), r? ∈ (0, ∞], S ∈ (0, 1) and F S (1 − s) =
`S (s)sαS , `S ∈ RV00 , αS > 0. Then
• If R is in the Gumbel MDA with auxiliary function b,
F Y (x) ∼ Γ(1 + αS )F S (1 − (xb(x))−1 )F R (x),
x → r? .
(35)
• If R is in the negative Weibull MDA,
F Y (r? − s) ∼
Γ(1 + αS )Γ(1 + αR )
F S (1 − s/r? )F R (r? − s),
Γ(1 + αS + αR )
s → 0.
(36)
Remark 7. Lemma 5 is easily extended to the case where S has upper endpoint s? ∈ (0, ∞), by writing
RS = s? R × S/s? , and noting that: (i) if R is in the Gumbel MDA, with upper endpoint r? and auxiliary
function b(t), then s? R is in the Gumbel MDA with upper endpoint s? r? and auxiliary function b(t/s? )/s? ; (ii)
if R is in the negative Weibull MDA, so is s? R, with F s? R (s? r? − s) = F R (r? − s/s? ).
Proof of Proposition 2. For the marginal tail F X of X1 and X2 , we have
Z
Z
F X (x) =
F R (x/v)dFW (v) = F R (x)P(W = 1) +
[0,1]
F R (x/v)dFW (v),
[0,1)
with P(W = 1) = P(Z ∈ Iν ). Letting S = W |W < 1, F S (1 − s) = `(s)sαW , s → 0, where αW is given by the
relevant part of Lemma 4, and ` ∈ RV00 . By Lemma 5, we have
F X (x) − F R (x)P(W = 1) ∼ Γ(1 + αW )`((xb(x))−1 )(xb(x))−αW F R (x),
x → r? .
(37)
?
We always have xb(x) → ∞ as x → r , (Hashorva et al., 2010; Resnick, 1987, Lemma 1.2), hence if P(W =
1) = 0, F X (x) = o(F R (x)), x → r? , whilst if P(W = 1) > 0, F X (x) ∼ P(W = 1)F R (x), x → r? .
For the joint distribution, F X∧ (x) = P(RW∧ ≥ x), with W∧ ∈ [0, ζ] for ζ ∈ [1/2, 1]. By Lemma 4,
F W∧ (ζ(1 − s)) = `∧ (s)s, s → 0, whilst by Lemma 5 and Remark 7,
F X∧ (x) ∼ Γ(2)`∧ (((x/ζ)b(x/ζ))−1 )((x/ζ)b(x/ζ))−1 F R (x/ζ),
x → ζr? .
(38)
Therefore,
χX = lim?
x→r
F X∧ (x)
Γ(2)`∧ (((x/ζ)b(x/ζ))−1 )((x/ζ)b(x/ζ))−1 F R (x/ζ)
= lim?
,
x→r F R (x)P(W = 1) + Γ(1 + αW )`((xb(x))−1 )(xb(x))−αW F R (x)
F X (x)
(39)
whilst for ζr? = r? , i.e., ζ = 1 or r? = ∞,
log F R (x)P(W = 1) + Γ(1 + αW )`((xb(x))−1 )(xb(x))−αW F R (x)
log F X (x)
ηX = lim?
= lim?
x→r log F X (x)
x→r
log Γ(2)`∧ (((x/ζ)b(x/ζ))−1 )((x/ζ)b(x/ζ))−1 F R (x/ζ)
∧
= lim?
x→r
log F R (x)
.
log F R (x/ζ)
(40)
The latter equality follows by definition (9), and l’Hôpital’s rule providing limx→r? log(xb(x))/ log F R (x) = 0.
Now if ζ < 1 and r? = ∞, then χX = 0 by (10), with ηX given by (40). If r? < ∞, X∧ has upper bound
?
ζr < r? and so χX = 0 also, and ηX is not defined.
If ζ = 1 then χX = 0 if P(W = 1) > 0 by (39), whilst ηX = 1 by (40). If P(W = 1) = 0 then b1 = b2 = 1/2,
with Lemma 4 giving αW = 1 and
χX = lim?
x→r
`∧ ((xb(x))−1 )
2fZ (1/2)/τ10 (1/2− )
=
.
`((xb(x))−1 )
fZ (1/2)/τ10 (1/2− ) − fZ (1/2)/τ20 (1/2+ )
The function τ is not differentiable at 1/2 when ζ = 1 so τ10 (1/2− ) 6= τ20 (1/2+ ) and rearrangement gives the
result.
20
Proof of Proposition 3. Following a similar line to Proposition 2, the marginal distribution satisfies
Γ(1 + αW )Γ(1 + αR )
`(s/r? )(s/r? )αW F R (r? − s)
Γ(1 + αW + αR )
αW +αR
˜
∼ `(s)s
,
s → 0,
F X (r? − s) − F R (r? − s)P(W = 1) ∼
˜ whilst for the joint distribution
for a new slowly varying function `,
Γ(2)Γ(1 + αR )
`∧ (s/(ζr? ))(s/(ζr? ))F R (r? − s/ζ)
Γ(2 + αR )
∼ `˜∧ (s)s1+αR ,
s → 0.
F X∧ (ζr? − s) ∼
If ζ < 1, then X∧ has upper bound ζr? < r? and so χX = 0, with ηX not defined. If ζ = 1 and P(W = 1) > 0,
then χX = 0, and
log F R (r? − s)
log F X (r? − s)
=
lim
= αR /(1 + αR ).
s→0 log{sF R (r ? − s)}
s→0 log F X (r ? − s)
∧
ηX = lim
Otherwise, if ζ = 1 and P(W = 1) = 0, then since αW = 1
F X∧ (r? − s)
`∧ (s/r? )
∼
,
?
`(s/r? )
F X (r − s)
s → 0,
with the behavior of ` and `∧ as given in (31) and (32).
7.2
Proofs for Section 3
For simplicity of notation, we often write S = log R and V = log W and V∧ = log W∧ in the following. We
denote the common marginal distribution of X1 and X2 by FX . We start by recalling an important result from
the literature on the convolution with a convolution-equivalent distribution.
Lemma 6 (Convolution with a distribution in CEα , see Theorem 1 of Cline (1986) and Lemma 5.1 of Pakes
(2004)). Let Y1 , Y2 be two random variables with distributions F1 , F2 respectively. If F1 ∈ CEα with α ≥ 0 and
F 2 (x)
→ c ≥ 0,
F 1 (x)
x → ∞,
then
F1 ? F2 (x)
→ E eαY2 + cE eαY1 ,
F 1 (x)
x → ∞.
(41)
The following lemma establishes the link between convolutions of nonnegative exponential-tailed distributions and the case where negative values may arise; it will be useful in several contexts.
Lemma 7 (Convolutions of exponential-tailed distributions with negative values). For i = 1, 2 and probability
distributions Fi ∈ ETα defined over R with α > 0, denote pi = F i (0) ∈ [0, 1] the probability of nonnegative
values. Using the convention 0/0 = 0, let Fi+ with 1 − Fi+ (x) = F i (x)/pi , x ≥ 0, denote the conditional distribution of Fi over nonnegative values, and let Fi− (x) = Fi (x)/(1
R ∞ − pi ), x < 0, denote the conditional distribution
of Fi over negative values. We use the notation MH (α) = −∞ exp(αy)H(dy) for a given distribution H. Then:
1. If Fi+ (x)/F1+ ? F2+ (x) → 0 when x → ∞ for i = 1, 2, then
F1 ? F2 ∼ p1 p2 F1+ ? F2+ .
(42)
2. If F1+ ? F2+ ∼ c1 F1+ + c2 F2+ with constants 0 ≤ c1 , c2 < ∞, then
F1 ? F2 (x) ∼ F 1 (x) p2 c1 + (1 − p2 )MF − (α) + F 2 (x) p1 c2 + (1 − p1 )MF − (α) .
2
1
Specifically, if c1 = MF + (α) and c2 = MF + (α), then
2
1
F1 ? F2 (x) ∼ MF2 (α)F 1 (x) + MF1 (α)F 2 (x),
and if c1 = MF + (α) and c2 = 0, then
2
F1 ? F2 (x) ∼ MF2 (α)F 1 (x) + F 2 (x)(1 − p1 )mF − (α).
1
21
(43)
Proof. We start with the mixture representation
F1 ? F2 (x) = p1 p2 F1+ ? F2+ (x) + p1 (1 − p2 )F1+ ? F2− (x) + (1 − p1 )p2 F1− ? F2+ (x),
x ≥ 0.
We can then use the equation pi Fi+ (x) = Fi (x), x ≥ 0, and the following inequalities for {i1 , i2 } = {1, 2},
0 = Fi−1 ? Fi−2 (x) = Fi−1 (x) ≤ Fi+1 ? Fi−2 (x) ≤ Fi+1 (x) ≤ Fi+1 ? Fi+2 (x),
x ≥ 0.
Moreover, Lemma 2 can be applied for mixed terms, yielding Fi+1 ? Fi−2 ∼ MF − (α)Fi+1 , and then (42) and (43)
i2
follow from straightforward calculations. To determine the behavior for special cases of c1 and c2 , observe that
(1 − pi )MF − (α) + pi MF + (α) = MFi (α).
i
i
Proof of Proposition 4.
1. The assumption F W (x)/F R (x) → c, x → ∞, is equivalent to F V (x)/F S (x) → c. We then apply (41)
and obtain F X ∼ (1 + c)F R . Since F V∧ (x)/F V (x) → χW , x → ∞, we get F X∧ (x) = F S+V∧ (log x) ∼
(1 + cχW )F S (log x) = (1 + cχW )F R (x), and the value χX in (13) follows. Since χX > 0 in all cases, we have
ηX = 1.
2. Let FV ∈ CE0 and F R = o(F W ). Then, F S = o(F V ), and (41) gives F X ∼ F W . If χW > 0, then
F S = o(F V∧ ), and (41) yields the tail equivalence F X∧ (x) ∼ F W∧ (x), which entails χX = χW . In the case
where χW = 0, we have F V∧ = o(F V ) and F S = o(F V ), such that Foss et al. (2009, Corollary5(ii)) establishes
F X∧ ∼ F R + F W∧ + o(F W ), and χX = χW = 0 follows.
2.a) When further FS , FV∧ ∈ CE0 and F R (x)/F W∧ (x) ≤ C with C > 0 as x → ∞, we consider the two
boundary cases F R = o(F W∧ ) and F R /F W∧ ∼ C. Using (41), we get
log F X (x)
log F W (x)
log F W (x)
∼
=
log F X∧ (x)
log(F R (x) + F W∧ (x))
log F W∧ (x) + log 1 +
∼
log F W (x)
∼ ηW ,
log F W∧ (x)
F R (x)
F W∧ (x)
x → ∞,
in both cases, which proves ηX = ηW .
2.b) When further FS ∈ CE0 and F W∧ = o(F R ), then we get F X∧ (x) ∼ F R (x) from (41), such that
log F X (x)
log F W (x)
∼
,
log F X∧ (x)
log F R (x)
x → ∞,
and the limit is ηX if it exists.
Proof of Proposition 5. Since E(RαW +ε ) < ∞ for a small ε > 0, by Lemma 2, the marginal distributions satisfy
F X (x) ∼ E(RαW )F W (x),
x → ∞,
(44)
so that X is also regularly varying with index αW . The coefficient ηW is not defined if W∧ has a finite upper
endpoint, in which case E(W∧αR +ε ) < ∞. Otherwise, we have that W∧ is regularly varying with index αW /ηW ,
or if ηW = 0, F W∧ decays faster than any power. Again, by Lemma 2, we obtain that
(
E(W∧αR )F R (x),
if αR < αW /ηW , ηW = 0 or ηW not defined,
F X∧ (x) ∼
αW /ηW
E(R
)F W∧ (x), if αR > αW /ηW or αR = +∞.
If χW > 0 then ηW = 1, so we conclude that
χX = lim F X∧ (x)/F X (x) = χW .
x→∞
For the coefficient of residual tail dependence between X1 and X2 we obtain
1/ηX = lim log F X∧ (x)/ log F X (x)
x→∞
(
limx→∞ log F R (x)/ log F W (x) = αR /αW , if αR < αW /ηW , or ηW not defined,
=
limx→∞ log F W∧ (x)/ log F W (x) = 1/ηW , if αR > αW /ηW or αR = +∞.
22
Before proceeding to the proof of Proposition 6, we recall several lemmas from the literature.
Lemma 8 (Convolution of distributions in ETα,β>−1 , see Theorem 4(v) of Cline (1986)). For two distributions
Fi ∈ ETα,βi , i = 1, 2, possessing gamma-type tail
F i (x) ∼ `i (x)xβi exp(−αx),
x→∞
α > 0, βi > −1,
i = 1, 2,
(45)
with slowly varying `i (x), we get
F1 ? F2 (x) ∼ α
Γ(β1 + 1)Γ(β2 + 1)
`1 (x)`2 (x)xβ1 +β2 +1 exp(−αx),
Γ(β1 + β2 + 2)
x → ∞.
(46)
Proof. The result for nonnegative S and V is found in Theorem 4(v) of Cline (1986). The extension to negative
values then follows from Lemma 7(1).
Lemma 9 (Ratio of convolutions with a distribution in ETα ). Let F ∈ ETα with α > 0, and let G1 , G2 be
Rdistributions satisfying G2 (x)/G1 (x) → c, x → ∞, with c ≥ 0. Given a distribution H, we write MH (α) =
exp(αx)H(dx) ∈ (0, ∞].
1. If G1 ∈ ETα , c > 0 and MF (α) = MG1 (α) = ∞, then
F ? G2
∼ c,
F ? G1
x → ∞.
(47)
2. If MG1 (α) < ∞ and F ? G1 ∼ MG1 (α)F , then
F ? G2
MG2 (α)
,
∼
MG1 (α)
F ? G1
x → ∞.
(48)
Proof. The result is given in Theorem 6(ii,iii) of Cline (1986) for F (0) = 1 and G1 (0) = 1. For point 1, the
extension to negative values in F and G1 follows from observing that Theorem 6(iii) of Cline (1986) implies
F ? G 1 ∼ F + ? G+
1,
x → ∞,
(49)
where F + is obtained from F by setting F + (0) = F (0) and F + (0) = 1, and the same construction is taken for
+
G+
and G+
1 ; i.e., F
1 arise from projecting negative values to 0. For point 2, the extension to negative values can
be shown using Lemma 7(2). Indeed, the same limit MG1 (α) arises for F ? G1 (x)/F (x) if we project negative
values in F and G1 to 0 or not.
Proof of Proposition 6.
1. To show ηX = 1, we exploit the closure of ETα under convolution (Lemma 2.5, Watanabe, 2008); equivalently,
RVα is closed under product convolutions. Applying Resnick (2007, Proposition 2.6(i)) yields
log F S+V (x)/ log F S (x) ∼ αx/(αx) = 1.
(50)
It remains to show that the limit in (50) does not change when we substitute log F S+V∧ (x) for log F S (x) in the
denominator, where no additional assumption on the distribution of V∧ is made. If p+
∧ = F V∧ (0) > 0, we get
F S+V∧ (x) ≥ p+
F
(x)
for
x
>
0,
and
then
∧ S
log F S+V (x)/ log F S+V∧ (x) ≥ log F S+V (x)/(log F S (x) + log p+
∧ ) ∼ 1,
α
such that ηX = 1. If p+
∧ = 0, then Lemma 2 gives F S+V∧ ∼ E(W∧ )F S , and ηX = 1 follows by analogy with (50).
2. In the case where FS ∈ CEα and F V (x)/F S (x) → c ≥ 0, x → ∞, we can use Lemma 6 with F1 = FS and
F2 = FV , which yields
F S+V (x)/F S (x) ∼ E eαV + cE eαS = E (W α ) + cE (Rα )
and by setting F2 = FV∧ , we get
F S+V∧ (x)/F S (x) ∼ E eαV∧ + cχW E eαS = E (W∧α ) + cχW E (Rα ) .
Combining these two results yields the value of χX .
23
3. If FV ∈ CEα and F S (x)/F V (x) → 0, x → ∞, we use Lemma 6 to show F S+V (x)/F V (x) ∼ E(eαS ) = E(Rα ).
If χW > 0, then also FV∧ ∈ CEα and we have F S+V∧ (x)/F V∧ (x) ∼ E(Rα ) by analogy. By combining these two
results, we get F S+V∧ (x)/F S+V (x) ∼ F V∧ (x)/F V (x) ∼ χW , such that χX = χW .
If χW = 0, then F V∧ (x)/F V (x) → 0, x → ∞. Consider the copula (V1 , V2 ) defined as the mixture of (V1 , V2 )
and (V, V ) with probabilities (1 − ) and , respectively, for some 0 < < 1. The marginal distribution of
(V1 , V2 ) is still FV , and the induced V∧ satisfies
F V∧ (x) = (1 − )F V∧ (x) + F V (x) ≥ F V∧ (x).
Therefore, F V∧ (x)/F V (x) → > 0 and F S (x)/F V∧ (x) → 0 as x → ∞. Since V∧ is stochastically larger than
V∧ , this means that also Y + V∧ is stochastically larger than Y + V∧ for any random variable Y , by a coupling
argument. Thus,
F S+V∧ (x)
F V∧ (x)
F S+V∧ (x)
χX = lim
≤ lim
= lim
= ,
x→∞ F S+V (x)
x→∞ F S+V (x)
x→∞ F V (x)
where the last but one equation follows from the former case where χW > 0 (with V∧ taking the role of V∧ ).
Since > 0 is arbitrary, the result follows.
4. The assumptions in this case are FS ∈ ETα,βR with either βR > −1 or βR = −1 and E (Rα ) = ∞, and
FV ∈ ETα,βW .
4.a) We consider the case χW > 0 where βW > −1 or βW = −1 and E(W α ) = ∞. We can apply Lemma 9(1),
which shows χX = χW .
4.b) We consider the case χW ≥ 0 where βW < −1 or βW = −1 < βR and E (W α ) < ∞. Then, Theorem 4(iv)
of Cline (1986) yields F S+V (x)/F S (x) ∼ E (W α ). Therefore, we can apply Lemma 9(2) with c = χW , such that
χX = E (W∧α ) /E (W α ) .
4.c) Here, βR , βW > −1 so log R, log W are gamma-tailed. Since we assume E W∧α+ε < ∞, Lemma 2 provides
F S+V∧ (x) ∼ E (W∧α ) F S (x), whilst by Lemma 8,
F S+V (x) ∼ α
Γ(βW + 1)Γ(βR + 1)
`W (x)xβW +1 F S (x),
Γ(βW + βR + 2)
so as x → ∞,
χX ∼
E (W∧α ) Γ(βR + βW + 2)
→ 0.
αΓ(βW + 1)Γ(βR + 1)`W (x)xβW +1
Proof of Proposition 7. If χW > 0, then necessarily αW = αW∧ , γW = γW∧ and χW = cW∧ /cW where `W ∼ cW
and `W∧ ∼ cW∧ . Using Asmussen et al. (2017, Theorem 3.1 and Theorem 4.1(iii)), see also earlier results of
Balkema et al. (1993), we therefore get χX = kX∧ /kX , where k. are the constants made explicit in the cited paper
for the sums of the Weibull-tailed random variables S + V and S + V∧ respectively. By simplifying the resulting
expression of χX , which follows from the equalities αW = αW∧ , γW = γW∧ , we obtain χX = cW∧ /cW = χW .
When χW = 0, we again use Asmussen et al. (2017, Theorem 3.1 and Theorem 4.1(iii)) and get
β/(β−1)
β/(β−1)
αR α
+ αW αR
F log X (x) ∼ cX xγX exp −αX xβ , γX = γR + γW + β/2, αX = W
β ,
1/(β−1)
1/(β−1)
αR
+ αW
(51)
with a constant cX > 0 that can be made explicit, see Asmussen et al. (2017, Theorem 3.1). Moreover, replacing
the symbol W in (51) with W∧ yields the tail approximation of log X∧ = S + V∧ .
Applying Lemma 1 to the random vector (X1 , X2 ), we get
β
β/(β−1)
β/(β−1)
1/(β−1)
1/(β−1)
α
α
+
α
α
α
+
α
R
W
W
R
R
W∧
αX
ηX =
=
β
αX∧
β/(β−1)
β/(β−1)
1/(β−1)
1/(β−1)
αR αW∧
+ αW∧ αR
αR
+ αW
1/(β−1) β
αW∧
1/(β−1)
1/(β−1)
1
+
αR
α
+ αR
αW
=
× W
1/(β−1) .
1/(β−1)
1/(β−1)
αW∧
αW
αW∧
+ αR
1 + αR
By substituting αW /αW∧ = ηW , and simplifying, equation (20) follows.
If χW = 0 and ηW < 1, then ηX < 1, implying χX = 0. If ηW = 1 but χW = 0, then equation (18) with the
assumption of asymptotically constant slowly varying functions implies γW∧ < γW . Combining (51) and (18),
γX∧ < γX and χX = 0 also.
24
For the proof of Proposition 8, we need the following Lemma.
Lemma 10 (Tail decay of products of Weibull-type variables, see Theorem 2.1(b) of Dȩbicki et al. (2018)). If
two independent random variables Y1 ≥ 0 and Y2 ≥ 0 are Weibull-tailed such that
F Yj (x) ∼ rj (x) exp(−αj xβj ),
x → ∞,
j = 1, 2,
(52)
with rj ∈ RV∞
γj . Then, as x → ∞,
β1 β2
β1
β1 β2
β2
F Y1 Y2 (x) ∼ C0 x 2(β1 +β2 ) r1 C1−1 x β1 +β2 r2 C1 x β1 +β2 exp −C2 x β1 +β2 ,
(53)
with constants C1 = (α1 β1 /(α2 β2 ))1/(β1 +β2 ) , C0 = (2πα2 β2 /(β1 + β2 )) C1 2 and C2 = α1 C1−β1 + α2 C1β2 .
For β1 = β2 = β this simplifies to
√
1/4
√
(54)
F Y1 Y2 (x) ∼ π α1 α2 xβ
r1 (α2 /α1 )1/(2β) x1/2 r2 (α1 /α2 )1/(2β) x1/2 exp −2 α1 α2 xβ/2 .
1/2
β /2
Consequently F Y1 Y2 is of the form (52) with r12 ∈ RV∞
γ12 , with γ12 = (β1 β2 + 2γ1 β2 + 2γ2 β1 )/{2(β1 + β2 )}.
Proof of Proposition 8. Consider the functions hj (x) = `j (x)xγj exp(−αj xβj ) with `j ∈ RV∞
0 , j = 1, 2. One
easily shows that
h1 (x)/h2 (x) → 0,
x → ∞,
(55)
if β1 > β2 , or if β1 = β2 and α1 > α2 , or if β1 = β2 and α1 = α2 and γ1 < γ2 .
Since F W∧ (x)/F W (x) ≤ 1, we always have βW ≤ βW∧ ; moreover, βW = βW∧ implies αW ≤ αW∧ , and
βW = βW∧ , αW = αW∧ implies γW ≤ γW∧ .
Using Lemma 10 for X = RW , we find that FX ∈ WTβX with parameters βX = βR βW /(βR + βW ),
αX = αR C1−βR + αW C1βW , where C1 = {αR βR /(αW βW )}1/(βR +βW ) ,
γX = (βR βW + 2γR βW + 2γW βR )/{2(βR + βW )},
(56)
and slowly varying function `X > 0. The same result applies to X∧ = RW∧ with constants βX∧ , αX∧ , γX∧ ,
and slowly varying `X∧ > 0.
1. In this case, it follows from Lemma 10 that also βX∧ = βX , αX∧ = αX , and γX∧ = γX , and thus
χX = lim F X∧ (x)/F X (x) = lim `W∧ (x)/`W (x) = χW ∈ [0, 1]
x→∞
x→∞
(57)
since all other dominating terms of higher order cancel out, and the further results follow straightforwardly.
2. Since γW∧ < γW , equation (56) implies γX∧ < γX , whilst βX = βX∧ and αX = αX∧ . Similarly to (57), we
therefore obtain
`X∧ (x)xγX∧
= 0 = χW ,
x→∞ `X (x)xγX
χX = lim
On the other hand, we have
log F X (x)
αX xβX
αW xβW
= lim
= 1 = lim
= ηW .
β
X
x→∞ log F X (x)
x→∞ αX∧ x ∧
x→∞ αW∧ xβW∧
∧
ηX = lim
3. Since βW∧ = βW and αW∧ > αW , it follows from (55) that χW = 0, and that ηW = αW /αW∧ . We clearly
have that βX = βX∧ , and
αX
=
αX∧
αR
αR βR
αW βW
−βR /(βR +βW )
+ αW
αR βR
αW βW
βW /(βR +βW )
−βR /(βR +βW∧ )
βW∧ /(βR +βW∧ )
αR βR
αR βR
αR αW
+
α
W
∧
αW∧ βW∧
∧ βW∧
βR /(βR +βW )
αW
=
∈ (0, 1),
αW∧
after some algebra and using that βW∧ = βW . Thus, αX∧ > αX , and by (55) we conclude that χX = 0, and
further
βR /(βR +βW )
αX xβX
αW
ηX = lim
=
6= ηW .
x→∞ αX∧ xβX∧
αW∧
25
4. For the case βW < βW∧ , we also have the inequality βX = βR βW /(βW + βR ) < βR βW∧ /(βW∧ + βR ) = βX∧ .
It therefore follows from (55) that χX = χW = 0 and ηX = ηW = 0.
Proof of Proposition 9. 1. Applying Lemma 5 to compute F X (x) and F X∧ (x) gives
χX = lim?
x→r
Γ(1 + αW∧ )`W∧ ((xbR (x))−1 )(xbR (x))−αW∧ F R (x)
Γ(1 + αW∧ )`W∧ (s)sαW∧
= χW ,
=
lim
s→0 Γ(1 + αW )`W (s)sαW
Γ(1 + αW )`W ((xbR (x))−1 )(xbR (x))−αW F R (x)
since xbR (x) → ∞ for x → r? , αW∧ ≥ αW , and if χW > 0 then necessarily αW∧ = αW . Similarly,
ηX = lim? log F X (x)/ log F X∧ (x) = 1,
x→r
since, by l’Hôpital’s rule limx→r? log(xbR (x))/ log F R (x) = 0, and therefore the term log F R (x) dominates both
the numerator and the denominator.
2. We observe that for some z < w? and K > 0,
χW
Z x
F W∧ (x)
bW∧ (t) [1 − bW (t)/bW∧ (t)] dt .
= lim? K exp −
= lim?
x→w
x→w F W (x)
z
If χW > 0 then we must have limx→w? bW (x)/bW∧ (x) = 1, since
χX = lim?
x→w
R w?
z
bW∧ (t)dt = ∞. Therefore, by Lemma 5,
`R ((xbW∧ (x))−1 )(xbW∧ (x))−αR F W∧ (x)
= χW .
`R ((xbW (x))−1 )(xbW (x))−αR F W (x)
(58)
On the other hand, if χW = 0, then
ηX = lim? log F W (x)/ log F W∧ (x) = ηW .
x→w
3. The upper endpoint of both X and X∧ is x? = r? w? . We use Lemma 5 and Remark 7 to see that
F X (x? − s) ∼
Γ(1 + αR )Γ(1 + αW )
Γ(1 + αR )Γ(1 + αW )
F W (w? − s/r? )F R (r? − s/w? ) ∼
`R (s)`W (s)sαR +αW ,
Γ(1 + αR + αW )
Γ(1 + αR + αW )
as s → 0, and similarly for F X∧ (x? − s). Therefore, if αW∧ = αW , then
`W∧ (s) αW∧ −αW
F X∧ (x? − s)
F W∧ (w? − s)
∼
∼
s
,
?
`W (s)
F X (x − s)
F W (w? − s)
and consequently, χX = χW . If αW∧ > αW , then χX = χW = 0. For the residual tail dependence coefficient
we compute
log F X (x? − s)
αW + αR
=
,
ηX = lim
s→0 log F X (x? − s)
αW∧ + αR
∧
since we only need to keep the dominating terms.
7.3
Proofs for Section 5
Proof of Proposition 10. For θ > 1, the norm ν is as in Example 2, and the result for χX is derived using
Proposition 2. For θ < 1, ν requires rescaling to meet our requirement that ν(x, y) ≥ max(x, y) with equality
somewhere. Note that scaling by any constant K ∈ (0, ∞) does not affect the dependence, i.e., (X1 , X2 ) =
R(τ (Z), τ (1 − Z)) has the same copula as (X1 , X2 ) = R(τ (Z), τ (1 − Z))/K. For θ ∈ [1/2, 1] we therefore define
ν ∗ (x, y) = θ−1 ν(x, y) which satisfies the required scaling, so that ζ = 1/ν ∗ (1, 1) = θ. The result for ηX is then
also given by Proposition 2.
Proof of Proposition 11. The first claim follows directly from Proposition 1, since E(W 1/ξ+ε ) < ∞. Since
ηW ∈ (0, 1) is defined, W∧ has the same upper endpoint as W , and F W∧ (w? − s) = `W∧ (F W (s))F W (s)1/ηW ,
implying W∧ is also in the negative Weibull MDA with αW∧ = αW /ηW . The second and third claims then
follow from parts 1 and 3 of Proposition 9.
26
RV0α
normal
log-normal
exponential
gamma
inverse normal
logistic
log-logistic
Gumbel
Weibull
t
Pareto
Fréchet
stable
F (shp1, shp2)
uniform
Beta(shp1, shp2)
triangular
GEV
WTβ
ETα
HT
LWTβ
3
β=2
RV∞
α
β=2
β=1
β=1
β=1
β=1
β=1
β=1
β = shp
α = scl
α = scl
shp
α = 2 mean
2
α = scl
3
α = scl
shp = 1, α = scl
β = shp = 2
α=1
α = shp2
α=2
α = 1/shp < 0
shp = 0, β = 1
shp = 0, α = scl
α = shp
shp < 1
3
3
3
shp < 2
3
β=1
β=1
β=1
shp < 2,β = 1
α = shp
α = shp
α = shp
α = 1/shp > 1/2
α = 2/shp2
shp > 0
shp > 0, β = 1
α = 1/shp
Table 4: Membership in tail classes (columns) for distribution families (rows). The column RV0α refers to the
behavior of F (x? − ·) when x? < ∞. All heavy-tailed distributions in this table are also subexponential. All
distributions in ETα listed in this table are in ETα,β except for the inverse normal; the inverse normal is in
CEα . The parameter shp of the stable distributions is here chosen as their stability parameter.
A
Tail classes and examples
Definitions of tail classes are given in Section 1.1. The following lemma summarizes important relationships
between such tail classes.
Lemma 11 (Relationships between tail classes). The following relationships between distribution classes hold:
1. RV∞
α ⊂ CE0 for α > 0,
2. ET0 ( HT.
3. For ETα with α > 0, we have:
• F (exp(·)) ∈ ETα ⇔ F ∈ RV∞
α ,
• CEα ⊂ ETα ,
• ETα,β>−1 ∩ CE = ∅.
4. For WTβ , we have:
S
• WT1 ⊂ α>0 ETα ,
• WTβ ⊂ CE0 for β < 1,
• LWTβ ⊂ SHT for β < 1.
5. By denoting F1 ≺ F2 if F 1 (x)/F 2 (x) → 0 for x → ∞, we have:
∞
• If α̃ < α, then WTβ>1 ≺ ETα ≺ ETα̃ ≺ WTβ<1 ≺ LWTβ>1 ≺ RV∞
α>0 ≺ RVα̃ ≺ SHT.
• CEα>0 ≺ ETα̃,β for α̃ ≤ α and any β > 0.
We recall the membership in tail classes for well-known parametric distribution families in Table 4, see
Johnson et al. (1994, 1995) for reference about parameters. Here we abstract away from the usual parameter
symbols of these distributions to avoid conflicting notations with general tail parameters. We refer parameters
as scl and loc if scl × X + loc has scale scl and location loc, where X has scale 1 and location 0. Another
parameter shp may be related to shape for some distributions.
27
B
Additional illustrations
Figure 5 illustrates further examples of norms ν and related functions τ (z) and τ (1 − z), as defined in Section 2.
0.2
0.4
0.6
0.8
1.0
ζ=1
0.0
0.0
0.2
0.4
0.6
0.8
1.0
ζ=1
0
0.4
0.6
0.8
0
1.0
b1
b2
1
0.2
0.4
0.6
0.8
0
1.0
b1 = b2
1
1
0.8
1
0.8
0
0.0
1.0
0.2
1.0
0.0
0.6
0.2
0.4
ζ
0.0
0.0
0.2
0.4
0.6
ζ
0
0.0
0.2
0.4
0.6
0.8
1.0
0
b1 = b2
1
0
0.0
0.2
0.4
0.6
0.8
1.0
0
b1 = b2 = 1
Figure 5: Further illustration of different norms ν and their related functions τ (z) (solid line) and τ (1 − z)
(dashed line).
Acknowledgments
Financial support from the Swiss National Science Foundation grant 200021-166274 (Sebastian Engelke) and
UK EPSRC grant EP/P002838/1 (Jennifer Wadsworth) is gratefully acknowledged.
References
Abdous, B., Fougères, A.-L., and Ghoudi, K. (2005). Extreme behaviour for bivariate elliptical distributions.
Canadian Journal of Statistics, 33(3):317–334.
Asmussen, S., Hashorva, E., Laub, P. J., and Taimre, T. (2017). Tail asymptotics of light-tailed Weibull-like
sums. arXiv preprint arXiv:1712.04070.
Balkema, A. A., Klüppelberg, C., and Resnick, S. I. (1993). Densities with Gaussian tails. Proceedings of the
London Mathematical Society, 3(3):568–588.
Balkema, G. and Nolde, N. (2010). Asymptotic independence for unimodal densities. Advances in Applied
Probability, 42(2):411–432.
Beirlant, J., Goegebeur, Y., Segers, J., and Teugels, J. L. (2004). Statistics of Extremes. Wiley.
Belzile, L. and Nešlehová, J. (2018). Extremal attractors of Liouville copulas. Journal of Multivariate Analysis,
160:68–92.
Breiman, L. (1965). On some limit theorems similar to the arc-sin law. Theory of Probability & Its Applications,
10(2):323–331.
Cline, D. B. (1986). Convolution tails, product tails and domains of attraction. Probability Theory and Related
Fields, 72(4):529–557.
Cline, D. B. and Samorodnitsky, G. (1994). Subexponentiality of the product of independent random variables.
Stochastic Processes and their Applications, 49(1):75–98.
Coles, S. G. and Tawn, J. A. (1991). Modelling extreme multivariate events. Journal of the Royal Statistical
Society: Series B (Statistical Methodology), pages 377–392.
28
Davis, R. A. and Mikosch, T. (2008). Extreme value theory for space–time processes with heavy-tailed distributions. Stochastic Processes and their Applications, 118(4):560–584.
Dȩbicki, K., Farkas, J., and Hashorva, E. (2018). Extremes of randomly scaled Gumbel risks. Journal of
Mathematical Analysis and Applications, 458(1):30–42.
Dombry, C. and Ribatet, M. (2015). Functional regular variations, Pareto processes and peaks over threshold.
Statistics and its Interface, 8:9–17.
Embrechts, P., Klüppelberg, C., and Mikosch, T. (2013). Modelling extremal events: for insurance and finance,
volume 33. Springer Science & Business Media.
Ferreira, A. and de Haan, L. (2014). The generalized Pareto process; with a view towards application and
simulation. Bernoulli, 20(4):1717–1737.
Foss, S., Korshunov, D., and Zachary, S. (2009). Convolutions of long-tailed and subexponential distributions.
Journal of Applied Probability, 46(3):756–767.
Fougères, A.-L. and Soulier, P. (2010). Limit conditional distributions for bivariate vectors with polar representation. Stochastic models, 26(1):54–77.
Hashorva, E. (2010). On the residual dependence index of elliptical distributions. Statistics & Probability
Letters, 80(13):1070–1078.
Hashorva, E. (2012). Exact tail asymptotics in bivariate scale mixture models. Extremes, 15(1):109–128.
Hashorva, E. and Hüsler, J. (2003). On multivariate Gaussian tails. Annals of the Institute of Statistical
Mathematics, 55(3):507–522.
Hashorva, E., Pakes, A. G., and Tang, Q. (2010). Asymptotics of random contractions. Insurance: Mathematics
and Economics, 47(3):405–414.
Heffernan, J. E. and Resnick, S. I. (2007). Limit laws for random vectors with an extreme component. The
Annals of Applied Probability, 17(2):537–571.
Heffernan, J. E. and Tawn, J. A. (2004). A conditional approach for multivariate extreme values (with discussion). Journal of the Royal Statistical Society: Series B (Statistical Methodology), 66(3):497–546.
Hult, H. and Lindskog, F. (2002). Multivariate extremes, aggregation and dependence in elliptical distributions.
Advances in Applied Probability, 34(03):587–608.
Huser, R. G., Opitz, T., and Thibaud, E. (2017). Bridging asymptotic independence and dependence in spatial
extremes using Gaussian scale mixtures. Spatial Statistics, 21(A):166–186.
Huser, R. G. and Wadsworth, J. L. (2018). Modeling spatial processes with unknown extremal dependence
class. Journal of the American Statistical Association. To appear. DOI: 10.1080/01621459.2017.1411813.
Johnson, N. L., Kotz, S., and Balakrishnan, N. (1994). Continuous univariate distributions. Vol. 1. 2nd ed.
Chichester: Wiley, 2nd ed. edition.
Johnson, N. L., Kotz, S., and Balakrishnan, N. (1995). Continuous univariate distributions. Vol. 2. 2nd ed.
New York, NY: Wiley, 2nd ed. edition.
Krupskii, P., Huser, R., and Genton, M. (2016). Factor copula models for replicated spatial data. Journal of
the American Statistical Association, To appear. DOI: 10.1080/01621459.2016.1261712.
Ledford, A. W. and Tawn, J. A. (1996). Statistics for near independence in multivariate extreme values.
Biometrika, 83(1):169–187.
Ledford, A. W. and Tawn, J. A. (1997). Modelling dependence within joint tail regions. Journal of the Royal
Statistical Society: Series B (Statistical Methodology), 59(2):475–499.
McNeil, A. J. and Nešlehová, J. (2009). Multivariate Archimedean copulas, d-monotone functions and `1 −norm
symmetric distributions. The Annals of Statistics, 37(5B):3059–3097.
Mikosch, T. (2006). Copulas: Tales and facts—rejoinder. Extremes, 9(1):55–62.
Nolde, N. (2014). Geometric interpretation of the residual dependence coefficient. Journal of Multivariate
Analysis, 123:85–95.
29
Opitz, T. (2013). Extremal t processes: Elliptical domain of attraction and a spectral representation. Journal
of Multivariate Analysis, 122:409–413.
Pakes, A. G. (2004). Convolution equivalence and infinite divisibility. Journal of Applied Probability, 41(02):407–
424.
Resnick, S. I. (1987). Extreme values, regular variation and point processes. Springer.
Resnick, S. I. (2007). Heavy-tail phenomena: probabilistic and statistical modeling. Springer Science & Business
Media.
Rootzén, H., Segers, J., and Wadsworth, J. L. (2018). Multivariate peaks over thresholds models. Extremes,
21:115–145.
Rootzén, H. and Tajvidi, N. (2006). Multivariate generalized Pareto distributions. Bernoulli, pages 917–930.
Seifert, M. I. (2014). On conditional extreme values of random vectors with polar representation. Extremes,
17(2):193.
Wadsworth, J. L., Tawn, J. A., Davison, A. C., and Elton, D. M. (2017). Modelling across extremal dependence
classes. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 79(1):149–175.
Watanabe, T. (2008). Convolution equivalence and distributions of random sums. Probability Theory and
Related Fields, 142(3):367–397.
30
| 10math.ST
|
CT sinogram-consistency learning for metal-induced beam
hardening correction
arXiv:1708.00607v2 [physics.med-ph] 12 Jan 2018
Hyoung Suk Park
Division of Integrated Mathematics,
National Institute for Mathematical Sciences, Daejeon, 34047, Korea.
Sung Min Lee, Hwa Pyung Kim,Jin Keun Seo∗
Department of Computational Science and Engineering,
Yonsei University, Seoul, 120-749, South Korea
Yong Eun Chung
Department of Radiology, Yonsei University College of Medicine, Seoul, 03722, South Korea
1
Abstract
Purpose: This paper proposes a sinogram-consistency learning method to deal with beamhardening related artifacts in polychromatic computerized tomography (CT). The presence of
highly attenuating materials in the scan field causes an inconsistent sinogram, that does not match
the range space of the Radon transform. When the mismatched data are entered into the range
space during CT reconstruction, streaking and shading artifacts are generated owing to the inherent
nature of the inverse Radon transform
Methods: The proposed learning method aims to repair inconsistent sinogram by removing
the primary metal-induced beam-hardening factors along the metal trace in the sinogram. Taking
account of the fundamental difficulty in obtaining sufficient training data in a medical environment,
the learning method is designed to use simulated training data and a patient-type specific learning
model is used to simplify the learning process.
Results: The feasibility of the proposed method is investigated using a dataset, consisting of
real CT scan of pelvises containing simulated hip prostheses. The anatomical areas in training
and test data are different, in order to demonstrate that the proposed method extracts the beam
hardening features, selectively. The results show that our method successfully corrects sinogram
inconsistency by extracting beam-hardening sources by means of deep learning.
Conclusion: This paper proposed a deep learning method of sinogram correction for beam
hardening reduction in CT for the first time. Conventional methods for beam hardening reduction
are based on regularizations, and have the fundamental drawback of being not easily able to use
manifold CT images, while a deep learning approach has the potential to do so.
Keywords: Tomographic image reconstruction, Computerized tomography, Metal artifact reduction, Deep
learning
2
I.
INTRODUCTION
In computerized tomography (CT), the presence of highly attenuating materials such
as metal, concentrated iodinated contrast media or bone complicates reconstruction1 by
violating an assumption of the forward model: that sinogram data are the Radon transform
of an image. The increasing use of metallic implants in a generally aging population makes
metal-induced artifacts a major impediment to CT diagnosis. The mismatched projection
data due to these effects cause severe streaking and shading artifacts in the reconstructed
CT images, as shown in Fig. 1. Metal artifacts are caused by the effects of beam-hardening
of polychromatic X-ray photon beams and the various complicated metal-tissue interactions
such as scattering, nonlinear partial volume effects and electric noise. Although extensive
research efforts have sought to improve CT reconstruction methods, tackling metal-related
artifacts is a very challenging problem because the inconsistent data induced by metal depend
non-linearly on the geometries and placements of the metal objects.
FIG. 1. CT images for patients in the presence of high-attenuating materials (hip prosthesis).
This paper uses a deep learning approach to repair sinogram inconsistencies, so as to
minimize artifacts induced by highly attenuating material. The proposed learning method
aims to correct only the main sources of beam-hardening artifacts, while leaving fine data
structures intact. We corrects the primary beam-hardening factor of metals along a sinusoidal metal trace. To simplify the learning process, we consider a patient-type specific
learning model (e.g., learning beam hardening artifacts caused by hip prosthesis). This paper explores the feasibility of the proposed method using simulated training dataset, due
to difficulty in collecting training data from patients (i.e., pairs of artifact-free and artifactcontaminated CT images). The anatomical areas in training and test data are different,
3
in order to demonstrate that the proposed method extracts the beam hardening features,
selectively; for example, the test data comes from sinograms corresponding to CT images
of the pelvis, whereas the training sinograms are obtained from the CT images of abdomen
and chest. (See Fig. 5 in the results section.) The beam-hardening factors due to metals
are expected to lie on a very-low-dimensional manifold, thus allowing extraction of beamhardening features in a supervised way. We also apply the proposed method to the projection
data acquired in different acquisition environments from those used for training; for example,
data acquired with different energy spectrum or data that includes (approximated) scatter.
We employ a U-net20 for learning the nonlinear beam-hardening features.
To be precise, let x(ϕ, s) (input data for deep learning) represent the projection data at
projection angle ϕ ∈ [0, 2π) and detector position s ∈ R. For simplicity, we shall restrict
to considering two-dimensional parallel-beam CT. According to the Lambert-Beer’s law2,10 ,
the CT sinogram data x can be expressed as
Z
η(E) exp − [R(µE )](ϕ, s) dE ,
x(ϕ, s) = − ln
(1)
where µE denotes the attenuation coefficient distribution at photon energy level E, η(E)
represents fractional energy at E 8,18 , and R(µE ) is the Radon transform of µE . See Fig.
2. Let Y denote the range space of the Radon transform R. Then, Y can be viewed as a
subspace of the sinogram space X.
The goal is to learn the sinogram correction function f : X → Y such that BFBP (f (x)) is
artifact-removed image, where BFBP represents the filtered backprojection (FBP) operator4 .
Noting that the orthogonal projection of x onto Y is Px = R(BFBP (x)), we have BFBP (x) =
BFBP (Px) and therefore f = RBFBP is not appropriate for MAR.
The proposed method uses a U-Net to learn the sinogram-correction function f : x 7→ y
from a training data {(x(i) , y (i) ) : i = 1, · · · , N } where each x(i) is an inconsistent sinogram
in the presence of a highly attenuating material and y (i) is the corresponding sinogram in
the absence of the highly attenuating material. Roughly speaking, the f can be achieved
from
N
1 X
kf (x(i) ) − y (i) k2
f = argmin
f ∈Unet N i=1
(2)
where Unet is a deep convolutional neural network with some domain knowledge.
Various numerical experiments show that the proposed deep learning method provides the
proper correction of the inconsistent sinogram with preserving its fine-structure. To confirm
4
the goodness of the sinogram repair, we use the above-mentioned method for evaluation
of the output f (x). The results suggest that the proposed deep learning method using
simulated training data can appropriately correct an inconsistent sinogram for MAR.
f (
)=
uncorrected
∈
/ MCT
B FBP
η(E)
Beer-Lamber Law
0.07
Weighted
average
0.04
0
Energy
10
60
R(µE )
100keV
60keV
20keV
µE
100
BFBP (
ction
kPro je
ed Bac
Filter
corrected
x
f
)=
BFBP (f (x)) ∈ MCT
BFBP (f (
y
))=
Px = R(BFBP (x))
CT manifold MCT
Y =Range space of R
FIG. 2. Schematic diagram of data inconsistency correction in polychromatic X-ray CT. (Left
bottom) Images of µE over 20keV≤ E ≤ 150 keV and the corresponding Randon transforms. (Left)
The sinogram data is approximately a weighted sum of the Radon transforms. Data inconsistencies
are caused by the energy-dependent behavior of the attenuation distribution. (Middle-bottom)
Filtered backprojections BFBP of the uncorrected x and corrected data f (x). (Right) Inconsistency
artifact is related to the degree of discrepancy between x and the range space Y of Radon transform.
We need to extract the source of artifacts in the data x by learning the features considering CT
image information (i.e. CT manifold MCT ).
II.
METHOD
A.
Learning objectives and U-net
Let x ∈ X represent a sinogram data in two-dimensional parallel-beam CT, where
X = L2 ([0, 2π) × R) is the space of square integrable functions with its norm kxk =
qR R
2π
0
2
R |x(ϕ, s)|
dsdϕ. Let Y be the range space of Radon transform R, a subspace of
X. To avoid notational complexity and for ease of explanation, we use the same notations
(x, y, X, Y ) to represent their discrete versions, used in the practice.
In the presence of the high attenuation materials such as metallic objects, x may not fit
5
with Y and the inconsistency of x can be quantified by
dx :=
min kR(µ) − xk.
(3)
µ∈L2 (R2 )
The mismatch dx is somehow related to the degree of artifacts in the image reconstructed
by the FBP method. Denoting FBP operator by BFBP , the reconstructed image BFBP (x)
can be viewed as the least-square fitting solution, in the sense that
BFBP (x) := argmin kR(µ) − xk
(4)
µ∈L2 (R2 )
where ‘argmin’ stands for the argument of the minimum. Hence, the projection operator
from X onto Y is expressed as P = RBFBP .
Υx
=B
FBP
B
(x)
FBP
Υx
(y) =
Υy
U-net
y
x
64
64
64
Conv 3 × 3, ReLU
128
128
128
Max pooling 2 × 2
256
128
128
128
Avg unpooling 2 × 2
128
64
64 64
Copy and Concat
64
64
Conv 1 × 1
FIG. 3. Deep learning architecture relying on operations of convolutional and pooling layers, and
rectified linear uints (ReLU). To address the difficulties of collecting training data in a CT clinical
environment, we use a patient-type specific learning model (e.g. hip replacements) and simplify
the learning process.
The goal is to find a suitable correction map f : x 7→ y for metal artifact reduction in such
a way that y ≈ R(µE∗ ) with µE∗ being the attenuation distribution at a fixed energy E∗ .
Imagine that MCT is a manifold near which various monochromatic CT images are residing.
Here, it is assumed that the data of monochromatic CT image-generating distribution is
concentrated near MCT embedded in the space R256×256 . The dimensionality of MCT can
be much lower than 256 × 256, because the probability of generating a CT-like image by
choosing pixel intensities independently of each other is close to zero3 . This manifold MCT
should not include images having serious metal artifacts, as shown in Fig. 2.
6
The map f must take account of this CT manifold MCT , because MCT is regarded as
prior information of CT images. Hence, f (x) should be determined by not only x but also
the manifold MCT . Due to highly nonlinear and complicated structure of MCT , it could be
very difficult to find f without using machine learning techniques.
We adopt U-net to learn f . Denoting by feed the feed forward map corresponding to f ,
the relation between an input vector x and an output vector is expressed as
y = f (x) = feed (x; w1 , · · · , wL )
(5)
where wl is the vector including the weight and bias for lth layer. We collect a set of training
data {(x(i) , y (i) )}N
i=1 and determine the vectors w 1 , · · · , w L by solving
N
1 X
kfeed (x(i) ; w1 , · · · , wL ) − y (i) k2
argmin
N
(w1 ,··· ,wL )
i=1
(6)
The most fundamental challenge is to collect enough medical image data to be well-labeled
{(x(i) , y (i) )}. This problem seems to be very difficult to solve. To address this fundamental disadvantages, we develop a patient-type specific learning method, which simplify the
way for learning f from a simulated data. The proposed method fixes only major beamhardening sources along the metal trace and does not touch relatively minor artifact sources.
Noting that the nonlinear beam-hardening factor is determined mainly by the geometry and
the arrangement of the metal objects, it is possible to learn the sinorgam correction from
simulated data. Fig. 3 depicts a rough schematic of this approach.
The network consists of a contracting path and an expansive path. The contracting
path contains convolutions with a size of 3 × 3, each followed by a rectified linear unit
(ReLU), along with 2 × 2 max pooling for downsampling. The expansive path used 2 ×
2 average unpooling22 instead of max pooling for upsampling. It was concatenated with
the correspondingly cropped feature from the contracting path. At the last layer a 1 × 1
convolution is used to reduce the dimensions.
B.
Sinogram inconsistency effects
The data consistency condition for sinogram x involves the following moment condiR
tion: for any integer k ≥ 0, the function R sk x(ϕ, s)ds with respect to (cos ϕ, sin ϕ) is a
homogeneous polynomial of order k 14 .
7
For simplicity, we only consider the zeroth order moment consistency condition. Define
Z
Υx(ϕ) :=
x(ϕ, s)ds
(7)
R
This Υx can be divided into its low frequency part ΥxL and its high frequency part ΥxH :
Υx(ϕ) = ΥxL (ϕ) + ΥxH (ϕ),
Note that
d
Υx(ϕ)
dϕ
ϕ ∈ (0, 2π].
(8)
= 0 for all ϕ ∈ [0, 2π) when the residual x − Px is zero. Hence, the
standard deviation of Υx, denoted by std (Υx), quantifies the amount of variation of the
inconsistency. The crucial observations are the followings:
• High frequency mismatch std (ΥxH ) (associated with abrupt change of Υx causes
bright and dark streaking artifacts in the CT image.
• Low frequency mismatch std (ΥxL ) (associated with a smoothly varying part of Υx
with respect to ϕ) generates shading artifacts in the CT image.
BFBP (x)
graph of Υx(ϕ), Υ00 x(ϕ)
x(ϕ, s)
Υx
D1
D2
10000
A
titanium
Υ00 x
s
-5000
ϕ
7π
18
A,1
11π
18
π
0
7π
18
A,1
11π
18
π
FIG. 4. Left figure shows the CT image µCT for numerical phantom containing two disk-shaped
metallic objects (titanium) occupying the region D1 and D2 . The attenuation coefficient of background image (marked by “YONSEI”) is assumed to be the constant with energy E. Middle and
right figures show projection data x and graph of Υx(ϕ) and
d2
Υx(ϕ)
dϕ2
of the phantom in the
left figure, respectively. Here, CT image is simulated at energy range between 0-150 keV (C=-250
HU/W=2500 HU).
Fig. 4 illustrates the behavior of Υx(ϕ) and
d2
Υx(ϕ)
dϕ2
using a numerical phantom con-
taining two disk-shaped metallic objects (titanium) occupying the region D1 and D2 . In this
simulation, we only consider the beam hardening artifacts for simplicity. The x is consistent
8
in all the regions except the diamond shaped area A := {(ϕ, s) : RχD1 (ϕ, s)RχD2 (ϕ, s) 6=
0}, where χD is the characteristic function, that is, χD = 1 on D and zero otherwise. This
local inconsistency of x on A corresponds to variation of Υx on the interval A,1 := {ϕ :
(ϕ, s) ∈ A }. In the case of a single disk-shaped metallic object, the corresponding projection data x lies in the range space of the Radon transform16 . However, BFBP (x) contains
cupping artifact inside the disk. Note that cupping artefacts only disturb the image in the
regions of problematic objects, whereas streaking artefacts corrupt the tomographic image
outside the problematic region15 .
From the inherent nature of the pseudo inverse of the Radon transform R, the local
inconsistency of x on A generates severe global artifacts in BFBP (x), which appear as
streaking and shading artifacts. Denoting the back-projection by R∗ (the dual of R), its
mathematical structure can be explained as follows:
BFBP (x)
x
R∗R(
) = R∗(
) + R∗(
)
=0
= R∗(
−
corrected sinogram
= R∗R(
) + R∗(
) + R∗R(
+
)
artifact sources
)
As illustrated in Fig. 4, the abrupt changes in Υx(ϕ) between projection views occur at
ϕ = π/2 and ϕ ≈ 7π/18, 11π/18. These abrupt discrepancies are mapped to the bright
and dark streaking artifacts between the boundaries of two metallic objects in the BFBP (x).
Moreover, the gradually varying inconsistency in the A,1 = [7π/18, 11π/18] interval causes
shading artifacts between or near the two metallic objects.
III.
RESULTS
This section demonstrates the feasibility of the proposed method through realistic computer simulations. In our simulations, patient-type specific learning model was adopted, owing to difficulty in handling all possible metal geometries for training. This section focuses
9
Simulated training data
Test data
input
pelvis
input
label
abdomen
label
chest
FIG. 5. The anatomical areas in training and test data are different. The test data comes from a
CT image of the pelvis, whereas the training sinograms are made from the CT images of abdomen
and chest. We intentionally use these unrealistic training data to demonstrate that the proposed
method effectively remove artifacts without deteriorating underlying morphological information.
10
×103
Training
Validation
cost
8
6
4
2
0
0
100
200
epoch
300
400
500
FIG. 6. Convergence plot of the cost function (6) for training and validation set with respect to
training epochs.
on learning beam hardening artifacts caused by hip prosthesis. We inserted two simulated
hip prostheses (made of iron) into CT images. Projection data x containing Poisson noise
was generated using the attenuation coefficient given in9 and energy spectrum with 2.5 mm
Al filtration at tube voltage of 150 kVp19 . Here, we assumed that µE for pelvic CT images
10
FIG. 7. Reconstruction results for pelvic CT images containing two simulated hip prostheses (made
of iron). First and second columns show the reference and uncorrected image, respectively. The
third and fourth columns represent the results with NMAR and the proposed method, respectively.
First, second, third and fourth rows show results of variations of metal geometries with different
tissue backgrounds. (C=-250 HU/W=2500 HU.)
as a tissue. Other causes of metal artifacts, such as scattering and nonlinear partial volume
effects were not considered.
For training, we intentionally did not use CT images of pelvis (test data’s anatomical
area), in order to demonstrate that the proposed method effectively extracts the beamhardening factors without affecting artifact-free data. The input data is the projection data
from CT images of chest and abdomen in the presence of hip prosthesis and the output is
the corresponding forward projection data in the absence of hip prosthesis (See Fig. 5). To
11
Reference
Uncorrected
Proposed
FIG. 8. Results of applying the proposed method to the projection data acquired in different
environments from those used for training: the data generated by using the energy spectrum with
1.5 mm Al filtration at tube voltage of 150 kVp (first row) and data that includes the scattered
photons (second row).
correct the main beam-hardening source along the metal trace, the method needs to apply
exclusively on the metal trace. The metal trace is obtained by the forward projection of the
metal region segmented from the uncorrected CT image.
To train the network, we generated 3780 CT sinogram data with a size of 368 × 180 for
CT images by changing position and size of the two metallic objects (iron). 3700 sinogram
dataset were used for training and 80 sinogram dataset were used for evaluation. The cost
function (6) was minimized using the RMSPropOptimizer21 with a learning rate of 0.001,
weight decay of 0.9, and mini-batch size of 32 at each epoch. Fig. 6 shows the convergence
plot for the cost function (6) with respect to training epochs. We used a trained network
with 500 training epochs. Training was implemented by Tensorflow7 on a CPU (Intel(R)
Core(TM) i7-6850K, 3.60GHz) and four GPU (NVIDIA GTX-1080, 8GB) system. The
network required approximately 16 hours for training.
Fig. 7 shows the comparison between the proposed method and normalized metal artifact
12
TABLE I. NRMSD of the reconstructed images with the NMAR method and the proposed method
Phantom
Uncorrected
NMAR
Proposed
Pelvis 1
38.3734
22.8523
8.1213
NRMSD
Pelvis 2
47.8539
18.8835
10.0256
(%)
Pelvis 3
47.4931
23.9704
8.7837
Pelvis 4
46.8481
20.8512
9.3886
reduction (NMAR) method13 for the simulated bi-lateral hip prostheses. As shown in the
second column of Fig. 7, the uncorrected images suffer from severe beam hardening artifacts,
which appear as streaking and shading artifacts between or near the two metallic objects. In
addition, bright streaking artifacts occur in direction of metallic materials. This is because
of noise due to insufficient photon reaching the detector after passing though the metallic
objects1 . The proposed and NMAR methods effectively reduce the streaking and shading
artifacts arising from metallic objects. However, as shown in the third column of Fig. 7, the
NMAR method does not recover the bone or tissue structures near the metallic object, since
NMAR fills the corrupted metal trace from its surrounding information. Compared with
NMAR method, the proposed method considerably recovers the morphological structure
near the metallic objects (see fourth column of Fig. 7). This is because the proposed
method tries to repair inconsistent sinograms by removing the primary metal-induced beamhardening factors along the metal trace in the sinogram.
In order to estimate the quantitative errors of the uncorrected and corrected images, the
normalized root mean square difference (NRMSD)11 (with the label image) were computed
on the outside of metals. The NRMSD (%) for each result is listed in Table I. In terms
of NRMSD, overall errors of NMAR and the proposed methods are significantly reduced
compared with that of uncorrected image. Compared with NMAR, the proposed method
achieves a lower NRMSD.
Fig. 8 shows the results of applying the proposed method to the projection data acquired
in different acquisition environments from those used for training: the data generated by
using the energy spectrum with 1.5 mm Al filtration at tube voltage of 150 kVp (first row)
and data that includes the scattered photons (second row). Specifically, the projection data
13
x considering scatter is generated by
Z
x̃(ϕ, s) = − ln
η(E) exp − [R(µE )](ϕ, s) + κE (ϕ, s)dE ,
(9)
where κE denotes the contribution of scatters. In this simulation, we assumed that κE is a
constant with respect to (ϕ, s)5,12 . As shown in third column of Fig. 8, the proposed method
reduces the metal induced artifacts while preserving the information near the metallic objects
even for both data with slightly different energy spectrum and data with scatters. However,
some streaking artifacts are still visible due to the inaccurate data along the metal trace.
IV.
DISCUSSION AND CONCLUSION
This paper proposed a deep learning method of sinogram correction for beam hardening
reduction in CT for the first time. Conventional methods for beam hardening reduction are
based on regularizations, and have the fundamental drawback of being not easily able to
use manifold CT images, while a deep learning approach has the potential to do so. The
fundamental challenge is to collect enough labeled training data, a necessary step for a deep
learning process. We circumvent this issue by using simulated training data as the basis for
learning sinogram repair. We use the fact that the geometry of a metal trace region in a
sinogram is non-linearly related to the beam-hardening feature. We use a U-net to learn
the map from the geometry of the metal trace region to the corresponding beam-hardening
factor. Results demonstrate that the proposed method can effectively remove metal-induced
inconsistencies.
Let us briefly discuss conventional regularization techniques such as compressed sensing
using `1 -regularization. These approaches can effectively prevent inadmissible solutions, but
they discard detailed information (e.g., they remove small anomalies containing clinically
useful information for diagnosis), and hence they have limited applicability to computed
medical imaging. Deep learning has the potential to learn non-linear regression for various
artifact sources, because it effectively uses complicated prior knowledge of CT images and
artifacts.
There is much room for improvement, and further research is necessary to deal with
various metal and bone-related artifacts. This paper does not consider metal artifacts that
arise from scattering, nonlinear partial volume effects or electric noise. The performance
14
of the learning-based sinogram correction method could be improved by enhancing the forward model, which accurately represents various realistic artifacts. Our future research will
include further investigation into learning approaches for sinogram correction and clinical
research with patients.
ACKNOWLEDGEMENTS
This work was supported by Samsung Science & Technology Foundation (No. SSTFBA1402-01). The first author was partially supported by the National Research Foundation
of Korea(NRF) grant funded by the Korea government(Ministry of Science, ICT & Future
Planning) (No. NRF-2016R1C1B2008098) and the National Institute for Mathematical
Sciences (NIMS) grant funded by the Korean government (No. A21300000).
∗
Corresponding author: J.K. Seo (email: [email protected]).
1
J. F. Barrett and N. Keat, “Artifacts in CT: recognition and avoidance,” Radiographics, vol.
24, pp. 1679-1691, 2004.
2
A. Beer, “Bestimmung der absorption des rothen lichts in farbigen flussigkeiten,” Ann. Physik.,
vol. 162, pp. 78-88, 1852.
3
Y. Bengio, I. Goodfellow, and A. Courville, “ Deep Learning,” Book in preparation for MIT
Press, Available: http://www.deeplearningbook.org/version-2015-10-03/, 2015
4
R. N. Bracewell and A. C. Riddle, “Inversion of fan-beam scans in radio astronomy,” Astrophys.
J., vol. 150, pp. 427, 1967.
5
B. De Man, J. Nuyts, P. Dupont, G. Marchal, P. Suetens, “Metal streak artifacts in X-ray
computed tomography : A simulation study,” IEEE Trans. Nucl. Sci. vol. 4, pp. 1860-1865,
1999.
6
D. Garcia, “Robust smoothing of gridded data in one and higher dimensions with missing
values,” Comput. Stat. Data Anal., vol. 54, pp. 1167-1178, 2010.
7
Google, TensorFlow:
Large-scale machine learning on heterogeneous systems, Available:
http://tensorflow.org, 2015
15
8
G. T. Herman and S. S. Trivedi, “A comparative study of two postreconstruction beamhardening correction methods,” IEEE Trans. Med. Imag., vol. 2, pp. 128-135, 1983.
9
J. H. Hubbell and S. M. Seltzer, “Tables of X-ray mass attenuation coefficients and mass energyabsorption coefficients from 1 keV to 20 MeV for Elements Z = 1 to 92 and 48 Additional
Substances of Dosimetric Interest,” NIST, 1996.
10
J. H. Lambert and E. Anding, Lamberts Photometrie: Photometria, Sive De Mensura et
Gradibus Luminis, Colorum et Umbrae, W. Engelmann, pp. 1728-1777, 1892.
11
A. Mehranian, M. R. Ay, A. Rahmim and H. Zaidi, “X-ray CT metal artifact reduction using
wavelet domain sparse regularization,” IEEE Trans. Med. Imag., vol. 32, pp. 1702-1722, 2013.
12
A. Mehranian, M. R. Ay, A. Rahmim and H. Zaidi, “3D Prior Image Constrained Projection
Completion for X-ray CT Metal Artifact Reduction,” IEEE Trans. Nucl. Sci., vol. 60, pp.
3318-3332, 2013.
13
E. Meyer, R. Raupach, M. Lell, B. Schmidt, and M. Kachelrieß, “Normalized metal artifact
reduction (NMAR) in computed tomography,” Med. Phys., vol. 37, pp. 5482-5493, 2010.
14
F. Natterer, The Mathematics of Computerized Tomography, SIAM, 1986.
15
H. S. Park, Y. E. Chung and J. K. Seo, “Computed tomographic beam hardening
artefacts:
mathematical characterization and analysis,” Philos Trans. R. Soc. A, DOI:
10.1098/rsta.2014.0388, 2015.
16
H. S. Park, J. K. Choi, J. K. Seo, “Characterization of Metal Artifacts in X-Ray Computed
Tomography,” Comm. Pure Appl. Math., DOI: 10.1002/cpa.21680, 2017.
17
H.S. Park, D. Hwang and J.K. Seo, “Metal artifact reduction for polychromatic X-ray CT based
on a beam hardening corrector”, IEEE Tran. Med. Imag. vol.35, pp. 480-487, 2016.
18
G. Poludniowski, G. Landry, F. DeBlois, P. M. Evans and F. Verhaegen “Spekcalc: a program
to calculate photon spectra from tungsten anode x-ray tubes,” Phys. Med. Biol., vol. 54, N433,
2009.
19
J. Punnoose, J. Xu, A. Sisniega, W. Zbijewski, J. H. Siewerdsen, “Technical Note: Spektr 3.0-A
computational tool for X-ray spectrum modeling and analysis”, Med. Phys., 43, pp. 4711– 4717,
2016.
20
O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image
segmentation,” in Proc. Med. Imag. Comput. Comput.-Assist. Intervention, Springer, pp. 234241, 2015.
16
21
T.
Tieleman
running
and
average
of
G.
its
Hinton,“Lecture
recent
6.5-rmsprop:
magnitude,”
Available:
Divide
the
gradient
by
http://cs.toronto.edu/
a
tij-
men/csc321/slides/lecture slides lec6.pdf, 2012
22
M. D. Zeiler, and R.Fergus, “Visualizing and understanding convolutional networks,” in Proc.
Eur. Conf. Comput. Vis., Springer, pp. 818-833, 2014.
17
| 1cs.CV
|
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
arXiv:1111.0442v4 [math.AC] 28 Aug 2013
A BSTRACT. We give a numerical characterization of the possible extremal Betti numbers (values as well
as positions) of any homogeneous ideal in a polynomial ring over a field of characteristic 0.
I NTRODUCTION
The main purpose of this note is to characterize the possible extremal Betti numbers (values as well as
positions) of any homogeneous ideal in a polynomial ring over a field of characteristic 0. These special
graded Betti numbers were introduced by Bayer, Charalambous and Popescu in [BCP]: One reason
why they naturally arise is that they are equal to the dimensions of certain cohomology vector spaces
of the projective scheme associated to the ideal. We also investigate on the possible Betti tables of a
componentwise linear ideal: Such a problem seems to be very hard, indeed we could solve it just in
some special cases. We provide some examples illustrating the main obstructions to the issue.
After a preliminary section, in Section 2 we study the possible Betti tables of componentwise linear
ideals, introduced by Herzog and Hibi in [HH1]. This issue is equivalent to a characterization of the
graded Betti numbers of strongly stable ideals. We denote by I j the jth graded component of a strongly
stable ideal I, set µi j (I) equal to the number of the monomials in I j ∩ K[x1 , . . . , xi ] divisible by xi , and
define the matrix M (I) = (µi j (I)), which we call the matrix of generators of I. As explained in the
beginning of Section 2, the matrix M (I) and the graded Betti numbers of I determine each other. Thus
we are led to characterize the integer matrices (µi j ) for which there exists a strongly stable ideal I such
that M (I) = (µi j ). From a result obtained by Murai in [Mu], which yields the characterization of the
possible Betti numbers of ideals with linear resolution (see Proposition 2.3), one can deduce some necessary conditions for (µi j ) being the matrix of generators for some strongly stable ideal, Proposition 2.6.
Unfortunately these conditions are not sufficient to describe the matrices of generators of strongly stable
ideals, as shown in Example 2.7. The difficulty of the task of characterizing Betti tables of componentwise linear ideals is also shown by Example 2.8, where we exhibit a noncomponentwise linear ideal
(in three variables) with the same Betti table of a componentwise linear ideal, answering negatively a
question raised by Nagel and Römer in [NR]. After discussing the main obstruction to constructing
strongly stable ideals with a prescribed matrix of generators, we give sufficient conditions for a matrix
to be of the form M (I) where I is strongly stable in Proposition 2.9. As a consequence it is shown in
Corollary 2.10 that the necessary conditions given in Proposition 2.6 are also sufficient when dealing
with strongly stable ideals in three variables. Another instance for which the matrix of generators of a
particular class of strongly stable ideals can be described is given in Proposition 2.13, which gives the
possible matrices of generators of lexsegment ideals, and therefore, exploiting a result in [HH1], of Gotzmann ideals. Though a complete characterization of the possible Betti numbers of a strongly stable ideal
seems to be quite difficult, we succeed in Section 3 to characterize all possible extremal Betti numbers of
any homogeneous ideal I ⊂ S = K[x1 , . . . , xn ], provided that K has characteristic 0. According to [BCP],
a Betti number βi,i+ j 6= 0 of I is called extremal if βk,k+l = 0 for all pairs (k, l) 6= (i, j) with k ≥ i and
l ≥ j. It is shown in [BCP] that the positions as well as the values of the extremal Betti numbers of a
graded ideal are preserved under taking the generic initial ideal with respect to the reverse lexicographical order. Thus, since K has characteristic 0, we may restrict our attention to characterize the extremal
Betti numbers of strongly stable ideals. More precisely, let i1 < i2 < · · · < ik < n, j1 > j2 > · · · > jk
and b1 , . . . , bk be sequences of positive integers. In Theorem 3.7 we give numerical conditions which are
Date: February 5, 2018.
2000 Mathematics Subject Classification. 13D02.
Key words and phrases. Betti tables; componentwise linear ideals; extremal Betti numbers; strongly stable monomial ideals.
1
2
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
equivalent to the property that there exists a homogeneous ideal I ⊂ S whose extremal Betti numbers are
precisely βi p ,i p + j p (I) = b p for p = 1, . . . , k.
We are very grateful to the anonymous referee for suggesting us the point (iv) of Theorem 3.7 and the
last statement in Lemma 3.3.
1. P RELIMINARIES
Let n be a positive integer. We will essentially work with the polynomial ring
S = K[x1 , . . . , xn ],
where the xi ’s are variables over a field K. Given a monomial u ∈ S, we set:
m(u) = max{e ∈ {1, . . . , n} : xe divides u}.
(1)
A monomial space V ⊂ S is called stable if for any monomial u ∈ V , then (u/xm(u) ) · xi ∈ V for all
i < m(u). It is called strongly stable if for any monomial u ∈ V and for each j ∈ {1, . . . , n} such that x j
divides u, then (u/x j ) · xi ∈ V for all i < j. Obviously a strongly stable monomial space is stable. By
a stable (strongly stable) monomial ideal I ⊆ S we mean that the underlying monomial space is stable
(strongly stable); or equivalently, that the monomial space hG(I)i, where G(I) is the unique minimal set
of monomial generators of I, is stable (strongly stable).
On the monomials, unless we explicitly say differently, we use a degree lexicographical order with
respect to the ordering of the variables x1 > x2 > . . . > xn . A monomial space V ⊂ S is called lexsegment
if, for all d ∈ N, there exists a monomial u ∈ Sd (the degree d part of S) such that
V ∩ Sd = hv ∈ Sd : v ≥ ui.
We will sometimes denote by:
L≥u = {v ∈ Sd : v ≥ u}.
Clearly, a lexsegment monomial space is strongly stable. The celebrated theorem of Macaulay explains
when a lexsegment monomial space is an ideal. We remind that given a natural number a and a positive
integer d, the dth Macaulay representation of a is the unique writing:
d
k(i)
a=∑
such that k(d) > k(d − 1) > . . . > k(1) ≥ 0,
i
i=1
see [BH, Lemma 4.2.6]. Then:
hdi
a
d
=∑
i=1
k(i) + 1
.
i+1
hdi
A numerical sequence (hi )i∈N is called an O-sequence if h0 = 1 and hd+1 ≤ hd for all d ≥ 1. (The reader
should be careful because the definition of O-sequence depends on the numbering: A vector (m1 , . . . , mn )
hi−1i
will be an O-sequence if m1 = 1 and and mi+1 ≤ mi
for all i ≥ 2). The theorem of Macaulay (for
example see [BH, Theorem 4.2.10]) says that, given a numerical sequence (hi )i∈N , the following are
equivalent:
(i) (hi )i∈N is an O-sequence with h1 ≤ n.
(ii) There is a homogeneous ideal I ⊂ S such that (hi )i∈N is the Hilbert function of S/I.
(iii) The lexsegment monomial space L ⊂ S such that L ∩ Sd consists in the biggest n+d−1
− hd
d
monomials, is an ideal.
For any Z-graded finitely generated S-module M, there is a minimal graded free resolution:
0→
M
j∈Z
S(− j)β p, j (M) →
M
j∈Z
S(− j)β p−1, j (M) → . . . →
M
S(− j)β0, j (M) → M → 0,
j∈Z
where S(k) denotes the S-module S supplied with the new grading S(k)i = Sk+i . The celebrated Hilbert’s
Syzygy theorem (for example see [BH, Corollary 2.2.14 (a)]) guarantees p ≤ n. The natural numbers
βi, j = βi, j (M) are numerical invariants of M, and they are called the graded Betti numbers of M. The
coarser invariants βi = βi (M) = ∑ j∈Z βi, j are called the (total) Betti numbers of M. A graded Betti
number βi,i+d is said to be extremal if βi,i+d 6= 0 and for all (p, q) 6= (i, d) such that p ≥ i and q ≥ d,
β p,p+q = 0. We will refer to the matrix (βi, j ) as the Betti table of M. Actually, in the situations we will
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
3
consider in this paper M = I is a homogeneous ideal of S. In this case βi, j = 0 whenever i ≥ n or j ≤ i
(unless I = S). We will present the Betti table of I as follows:
β0,1 β1,2 β2,3 · · · · · · βn−1,n
β0,2 β1,3 β2,4 · · · · · · βn−1,n+1
β0,3 β1,4 β2,5 · · · · · · βn−1,n+2 .
..
..
..
..
.
.
··· ···
.
.
Also if the definition of the Betti table of M predicts infinite many rows, only a finite number of entries
are not zero (because M is finitely generated). Consequently, only a finite number of rows are significant,
and in the examples we will present throughout the paper we will draw just the significant rows. Notice
that a graded Betti number is extremal if and only if it is the nonzero top left “corner” in a block of zeroes
of the Betti table.
For an integer d, the S-module M is said to have a d-linear resolution if βi, j (M) = 0 for every i =
0, . . . , p and j 6= i + d; equivalently, if βi (M) = βi,i+d (M) for any i = 0, . . . , p. Notice that if M has dlinear resolution, then it is generated in degree d. The S-module M is said componentwise linear if Mhdi
has d-linear resolution for all d ∈ Z, where Mhdi means the S-submodule of M generated by the elements
of degree d of M. It is not difficult to show that if M has a linear resolution, then it is componentwise
linear.
Let I be a stable monomial ideal. For all i ∈ {1, . . . , n} and d ∈ N we set:
(2)
mi,d (I) = |{u monomials in G(I) ∩ Sd : m(u) = i}|
mi (I) = |{u monomials in G(I) : m(u) = i}|.
By the Eliahou-Kervaire formula [EK] (see also [HH2, Corollary 7.2.3]) we have:
n
k−1
βi,i+d (I) = ∑
mk,d (I).
(3)
i
k=i
It is convenient to introduce the analog of the mi,d ’s for all Z-graded finitely generated S-module M: To
this aim, for all i ∈ {1, . . . , n + 1} and d ∈ Z we set:
n
k
k−i+1
βk,k+d (M).
(4)
mi,d (M) = ∑ (−1)
i−1
k=0
The following lemma shows that knowing the mi,d (M)’s is equivalent to knowing the Betti table of M,
and that the two definitions agree when M = I is a stable ideal.
Lemma 1.1. Let M be a Z-graded finitely generated S-module. Then:
n+1
k−1
mk,d (M).
βi,i+d (M) = ∑
(5)
i
k=i
Proof. Set mk,d = mk,d (M) and βi, j = βi, j (M). By the definition of the mk,d ’s we have the following
identity in Z[t]:
n+1
n
k=1
i=0
∑ mk,d t k−1 = ∑ βi,i+d (t − 1)i .
Replacing t by s + 1, we get the identity of Z[s]
n+1
n
k=1
i=0
∑ mk,d (s + 1)k−1 = ∑ βi,i+d si ,
that implies the lemma.
Let us define also the coarser invariants:
(6)
mi (M) =
∑ mi,d (M)
d∈Z
∀ i = 1, . . . , n + 1.
4
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
2. G RADED B ETTI
NUMBERS OF COMPONENTWISE LINEAR IDEALS
In this section we want to discuss the problem of characterizing the graded Betti numbers of a componentwise linear ideal I ⊂ S = K[x1 , . . . , xn ]. This is a difficult task, in fact we are not going to solve the
problem, rather we are going to explain why it is problematic. Such an issue is equivalent to characterize
the possible graded Betti numbers of a strongly stable monomial ideal of S. In fact, in characteristic 0 this
is true because the generic initial ideal of any ideal I is strongly stable [Ei, Theorem 15.23]. Moreover, if
I is componentwise linear and the term order is degree reverse lexicographic, then the graded Betti numbers of I are the same of those of Gin(I) by a result of Aramova, Herzog and Hibi in [AHH]. In positive
characteristic it is still true that for a degree reverse lexicographic order the graded Betti numbers of I are
the same of those of Gin(I), provided that I is componentwise linear. But in this case Gin(I) might be
not strongly stable. However, it is known that, if we start from a componentwise linear ideal, it is stable
(see Conca, Herzog and Hibi [CHH, Lemma 1.4]). The graded Betti numbers of a stable ideal do not
depend from the characteristic, because of the Eliahou-Kervaire formula (3). So to compute the graded
Betti numbers of Gin(I) we can consider it in characteristic 0. Let us call J the ideal Gin(I) viewed in
characteristic 0. The ideal J, being stable, is componentwise linear, so we are done by what said above.
Summarizing, we showed:
Proposition 2.1. The following sets coincide:
(1) {Betti tables (βi, j (I)) where I ⊂ S is componentwise linear};
(2) {Betti tables (βi, j (I)) where I ⊂ S is strongly stable};
Before beginning the discussion on graded Betti numbers, it is worthwhile to notice that to characterize
the total Betti numbers of a componentwise linear ideal is an easy task. Indeed, together with Proposition
2.1, the following remark of Murai yields the answer:
Remark 2.2. (Murai). Let (m1 , . . . , mn ) be a sequence of natural numbers. The following are equivalent:
(i) m1 = 1 and mi+1 = 0 whenever mi = 0.
(ii) There exists a strongly stable ideal I ⊂ S such that mi (I) = mi for any i = 1, . . . , n.
That (ii) =⇒ (i) is very easy to show. For the reverse implication, given a sequence (m1 , . . . , mn )
satisfying (i), set k = maxℓ {mℓ 6= 0}. By assumption we have mi ≥ 1 for all i = 1, . . . , k, therefore it
makes sense to define the following monomial spaces for each j = 1, . . . , k − 1:
j−2
j−2
j−2
m −1 m +m −1
m −1 m j −2 m j+1 +1
m −1 m j −1 m j+1
xj
.
, . . . , ∏ xi i+1 · x j j j+1
x j , ∏ xi i+1 · x j−1
V j = ∏ xi i+1 · x j−1
i=1
i=1
i=1
We also define:
Vk =
k−2
∏
i=1
m −1 mk −1
xk ,
xi i+1 · xk−1
k−2
∏
m −1 mk −2 2
xk ,
xi i+1 · xk−1
k2
...,
∏
m −1
k
xi i+1 · xm
k
i=1
i=1
.
Clearly, for all j = 1, . . . , k, we have wi (V j ) = m j if i = j and wi (V j ) = 0 otherwise. Set:
k
M
Vk ⊂ S.
I=
j=1
It is easy to see that I is a strongly stable monomial ideal and that hG(I)i =
Lk
j=1 Vk ,
so we get (ii).
Actually, also the possible Betti numbers of an ideal with linear resolution are known. Proposition 2.1
and [Mu, Proposition 3.8] yield the following:
Proposition 2.3. Let m1 , . . . , mn be a sequence of natural numbers. Then the following are equivalent:
(1) There exists a homogeneous ideal I ⊂ S with d-linear resolution such that mk (I) = mk for all
k = 1, . . . , n;
(2) (m1 , . . . , mn ) is an O-sequence such that m2 ≤ d, that is:
(a) m1 = 1;
(b) m2 ≤ d;
hi−1i
(c) mi+1 ≤ mi
for any i = 2, . . . , n − 1.
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
5
Discussion 2.4. In [Mu, Proposition 3.8] is shown the equivalence between point (2) of Proposition
2.3 and the existence of a strongly stable ideal I ⊂ S generated in degree d such that mk (I) = mk for all
k = 1, . . . , n. Looking carefully at the proof, one can see that, given an O-sequence (m1 , . . . , mn ), a special
strongly stable ideal I ⊂ S such that mk (I) = mk shows up. This special strongly stable ideal, which we
are going to define more explicitly, will play a crucial role throughout the paper.
Given an O-sequence (m1 , . . . , mn ) such that m2 ≤ d, set:
Vi = {biggest mi monomials u ∈ Sd such that m(u) = i}
for all i = 1, . . . , n. Then the discussed ideal is:
I=(
n
[
Vi ) ⊂ S.
i=1
We will refer to I as the piecewise lexsegment monomial ideal (of type (d, (m1 , . . . , mn ))). From what
said above, it follows that the piecewise lexsegment of type (d, (m1 , . . . , mn )) is strongly stable if and
only if (m1 , . . . , mn ) is a O-sequence such that m2 ≤ d.
Let I ⊂ S be a strongly stable monomial ideal. Notice that both Ih ji and mI, where j is a natural number
and m = (x1 , . . . , xn ) is the graded maximal ideal of S, are strongly stable. For all j ∈ N and i = 1, . . . , n,
we define:
µi, j (I) = mi (Ih ji ).
To know the matrix (mi, j (I)) or (µi, j (I)) are equivalent issues: Indeed, if J ⊂ S is a strongly stable
monomial ideal, then for all i = 1, . . . , n:
i
mi (mJ) =
∑ mq(J).
q=1
Therefore we have the formula:
i
(7)
mi, j (I) = mi (Ih ji ) − mi(mIh j−1i ) = µi, j (I) − ∑ µq, j−1 (I)
q=1
that implies that we can pass from the µi, j ’s to the mi, j ’s. It also follows from this formula that we can do
the converse path by induction on j, because µi,d (I) = mi,d (I) if d is the smallest degree in which I is not
zero. Therefore, using Proposition 2.1, to characterize the possible Betti tables of the componentwise
linear ideals is equivalent to answer the following question:
Question 2.5. What are the possible matrices M (I) = (µi, j (I)) where I ⊂ S is a strongly stable ideal?
We will refer to M = M (I) as the matrix of generators of the strongly stable ideal I. We will feature
M as follows:
µ1,1 µ2,1 µ3,1 · · · · · · µn,1
µ1,2 µ2,2 µ3,2 · · · · · · µn,2
M = µ1,3 µ2,3 µ3,3 · · · · · · µn,3
..
..
..
..
.
.
.
··· ···
.
We can immediately state the following:
Proposition 2.6. Let M = (µi, j ) be the matrix of generators of a strongly stable monomial ideal I ⊂ S.
Then the following conditions hold:
(i) Each non-zero row vector (µ1, j , µ2, j , . . . , µn, j ) of M is an O-sequence such that µ2, j ≤ j.
(ii) For all i and j one has µi, j ≥ ∑iq=1 µq, j−1 .
Proof. Condition (i) follows from Proposition 2.3 since Ih ji has a j-linear resolution for all j greater than
or equal to the lower degree in which I is not zero. Condition (ii) follows from (7).
Notice that the Noetherianity of S (or if you prefer conditions (i) and (ii) of Proposition 2.6) implies
that there exists m ∈ N such that µi, j (I) = ∑iq=1 µq, j−1 (I) for all j > m and i ∈ {1, . . . , n}. So, though M
has infinitely many rows, the relevant ones are just a finite number, and in the examples we will write
just them.
6
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
One may expect that the conditions described in Proposition 2.6 are sufficient. But this is not the case
at all:
Example 2.7. One obstruction is illustrated already by Remark 2.2: Consider the matrix
0
0
0
0
..
..
..
..
.
.
.
.
M = 0
0
0
0
1
d
0
0
1 d +1 d+1 k
where the first nonzero row from the top is the dth and d + 1 < k ≤ (d + 1)h2i . Such a matrix clearly
satisfies the necessary conditions of Proposition 2.6. However, if there existed a strongly stable ideal
I ⊂ K[x1 , x2 , x3 , x4 ] with matrix of generators M , then it would satisfy m1 (I) = 1, m2 (I) = d, m3 (I) = 0
and m4 (I) = k − d − 1 > 0, a contradiction to Remark 2.2. The first matrix of this kind is:
0 0 0 0
M = 1 2 0 0 .
1 3 3 4
The explained obstruction gives rise to a class of counterexamples. However, such a class does not fill
the gap between the existence of a strongly stable ideal with matrix of generators M and the necessary
conditions of Proposition 2.6. Let us look at the following matrix.
0 0 0 0
0 0 0 0
0 0 0 0
.
M =
0 0 0 0
1 3 2 2
1 4 6 9
One can check that the necessary conditions described in Proposition 2.6 hold. However one can show
that there is no strongly stable monomial ideal I ⊂ K[x1 , . . . , x4 ] with M as matrix of generators. Notice
that such an ideal would have m1 (I) = 1, m2 (I) = 3, m3 (I) = 2 and m4 (I) = 3, which does not contradict
Remark 2.2.
Example 2.8. Obviously the property of having linear resolution can be detected looking at the graded
Betti numbers. In the following example we show that this is not anymore true for componentwise linear
ideals, and this strengthens the impression that to give a complete characterization of the possible graded
Betti numbers of a componentwise linear ideal is probably a hard task. More precisely, we are going
to exhibit two ideals I and J, one componentwise linear and one not, with the same Betti tables. This
answers negatively a question raised in [NR, Question 1.1].
Consider the ideals of K[x1 , x2 , x3 ]:
I = (x41 , x31 x2 , x21 x22 , x1 x32 , x42 , x31 x3 , x21 x2 x23 , x21 x33 , x1 x22 x23 )
and
J = (x41 , x31 x2 , x21 x22 , x31 x3 , x1 x22 x3 , x1 x2 x23 , x1 x42 , x21 x33 , x42 x3 ).
Notice that I and J are generated in degrees 4 and 5. By CoCoA [Co] one can check that I and J have
the same Betti table, namely:
0 0 0
0 0 0
0 0 0 .
6 6 1
3 6 3
One can easily check that I is strongly stable, so in particular it is componentwise linear. On the contrary
J is not componentwise linear, since Jh4i = (x41 , x31 x2 , x21 x22 , x31 x3 , x1 x22 x3 , x1 x2 x23 ), as one can check by
CoCoA, has not a 4-linear resolution.
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
7
We are going to explain some reasons why the conditions of Proposition 2.6 for a matrix M are
in general not sufficient to have a strongly stable ideal corresponding to it. By the discussion after
Proposition 2.3, we know that for a given sequence (m1 , . . . , mn ) of integers, there exists a strongly stable
monomial ideal J generated in degree d such that mi (J) = mi if and only if the piecewise lexsegment
ideal of type (d, (m1 , . . . , mn )) is strongly stable. Unfortunately, even if J is a piecewise lexsegment, the
ideal mJ is not necessarily a piecewise lexsegment. For instance, keeping in mind the last matrix in
Example 2.7, the piecewise lexsegment ideal of type (5, (1, 3, 2, 2)) is:
J = (x51 , x41 x2 , x31 x22 , x21 x32 , x41 x3 , x31 x2 x3 , x41 x4 , x31 x2 x4 ).
However u = x31 x33 ∈
/ mJ, whereas v = x21 x32 x3 ∈ mJ. Since v is lexicographically smaller than u and
m(u) = m(v) = 3, mJ is not a piecewise lexsegment ideal. This fact does not make any troubles when the
number of variables is at most three, as we can see later. To see this, we need the following more general
proposition.
Proposition 2.9. Let M = (µi, j ) be a matrix. Then M is the matrix of generators of a strongly stable
monomial ideal I ⊂ S, provided that the following conditions hold:
(1) Each nonzero row vector (µ1, j , . . . , µn, j ) of M is an O-sequence with µ2, j ≤ j.
(2) For all i ∈ {1, . . . , n} one has µi, j ≥ ∑iq=1 µq, j−1 .
(3) If m ∈ N is the least number such that µi,h = ∑iq=1 µq,h−1 for each h > m, then for each j ≤ m
there exists a strongly stable ideal B( j) ⊂ S generated in degree j such that mi (B( j)) = µi, j for
all i = 1, . . . , n and B( j) ∩ K[x1 , . . . , xn−1 ] j+1 is a piecewise lexsegment monomial space.
Proof. Let d be the least natural number such that the row vector (µ1,d , . . . , µn,d ) is nonzero. We claim
to have built a strongly stable ideal I( j) ⊂ S such that µi,k (I( j)) = µi,k for any k ≤ j and I( j)h ji ∩
K[x1 , . . . , xn−1 ] = B( j) ∩ K[x1 , . . . , xn−1 ]. If j = m, then the desired ideal is I = I(m). If not, however we
can assume j ≥ d (I(d) = B(d)). We set
i
mi, j+1 = µi, j+1 − ∑ µq, j .
q=1
Let L( j + 1) be the ideal generated by the biggest mi, j+1 (i = 1, . . . , n) monomials u ∈ S j+1 \ I( j) such
that m(u) = i (they exist thanks to condition (1)). Set I( j + 1) = I( j) + L( j + 1). Clearly the first j + 1
rows of the matrix of generators of I( j + 1) coincide with the ones of M . Then notice that I( j + 1)h j+1i ∩
K[x1 , . . . , xn−1 ] is the piecewise lexsegment of type ( j + 1, (µ1, j+1 , · · · , µn−1, j+1 )) by construction. Because B( j + 1) ∩ K[x1 , . . . , xn−1 ] j+2 is a piecewise lexsegment monomial space, B( j + 1) ∩ K[x1 , . . . , xn−1 ]
is forced to be the piecewise lexsegment of type ( j + 1, (µ1, j+1 , · · · , µn−1, j+1 )). So we get the equality:
I( j + 1)h j+1i ∩ K[x1 , . . . , xn−1 ] = B( j + 1) ∩ K[x1 , . . . , xn−1 ].
To conclude the proof, we have to show that I( j + 1) is strongly stable. This reduces to show that, if
u ∈ I( j + 1) is a monomial of degree j + 1 with m(u) = n, then (u/xi )xk belongs to I( j + 1) for all
1 ≤ k < i ≤ n such that xi | u. We consider two cases. If u ∈ I( j) we are done, because I( j) ⊂ I( j + 1) is
strongly stable. If u ∈ L( j + 1), then we consider the monomial ideal
T ( j + 1) = (v ∈ G(I( j + 1)h j+1i ) : m(v) < n or v ≥ u) ⊂ I( j + 1).
Observe that T ( j + 1) is a piecewise lexsegment ideal of type ( j + 1, (µ1, j+1 , . . . , µn−1, j+1 , a)), where
a ≤ µn, j+1 . Since (µ1, j+1 , . . . , µn−1, j+1 , a) is an O-sequence with µ2, j+1 ≤ j + 1, it follows that T ( j + 1)
is strongly stable by the discussion after Proposition 2.3. Thus for each 1 ≤ k < i ≤ n such that xi |u, we
have that (u/xi )xk ∈ T ( j + 1) ⊂ I( j + 1).
Corollary 2.10. Let M = (µi, j ) be a matrix with 3 columns. Then M is the matrix of generators of a
strongly stable monomial ideal I ⊂ K[x1 , x2 , x3 ] if and only if the following conditions hold:
(1) Each non-zero row vector (µ1, j , µ2, j , µ3, j ) of M is an O-sequence with µ2, j ≤ j.
(2) For all j ∈ N one has µ2, j ≥ µ1, j−1 + µ2, j−1 and µ3, j ≥ µ1, j−1 + µ2, j−1 + µ3, j−1 .
Proof. The conditions are necessary from Proposition 2.6. Furthermore, since an ideal I ⊂ K[x1 , x2 ]
generated in one degree is piecewise lexsegment if and only if it is strongly stable, we automatically
have condition (3) of Proposition 2.9.
8
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
Although the complete characterization of the matrix of generators of an arbitrary strongly stable ideal
seems to be very complicated, based on the fact that the lexsegment property of an ideal is preserved under multiplication by the maximal ideal m, one may expect a characterization for the matrix of generators
of lexsegment ideals. For answering this question, first we define the concept of a d-lex sequence.
Definition 2.11. A sequence of non-negative integers m1 , · · · , mn is called a d-lex sequence, if there
exists a lexsegment ideal L ⊂ S generated in degree d such that mi (L) = mi for all i.
Because if I ⊂ S is a lexsegment ideal, then mI is still a lexsegment ideal, we clearly have that M =
(µi, j ) is the matrix of generators of a lexsegment ideal if and only if the following conditions hold:
(1) Each non-zero row vector (µ1, j , µ2, j , . . . , µn, j ) of M is a j-lex sequence.
(2) For all i and j one has µi, j ≥ ∑iq=1 µq, j−1 .
Therefore to characterize the matrix of generators of lexsegment ideals we need to characterize arbitrary
d-lex sequences. To do this, we have to recall the definition of the natural decomposition of the complement set of monomials belonging to a lexsegment ideal generated in a fixed degree. In what follows we
denote by [xt , . . . , xn ]r (1 ≤ t ≤ n) the set of all monomials of degree r in the variables xt , . . . , xn .
Definition 2.12. Let u = x j(1) . . . x j(d) ∈ Sd (1 ≤ j(1) ≤ · · · ≤ j(d) ≤ n) be a monomial and set L<u =
{v ∈ Sd | v < u}. Following the method described in [BH, page 159] (where L<u is denoted by Lu ) we
can partition the set L<u as:
L<u =
d
[
[x j(i)+1 , . . . , xn ]d−i+1 · x j(1) · · · x j(i−1) ,
i=1
which is called the natural decomposition of L<u .
Before proving the next result, notice that the powers of the maximal ideal are lexsegment ideals, and
the following formula holds for their d-lex sequences:
i+d−2
d
(8)
mi (m ) =
.
d −1
Proposition 2.13. Let m1 , . . . , mn be a sequence of natural numbers and let µ = ∑ni=1 mi . Suppose that
d
n+d −1
k( j)
ℓ=
−µ = ∑
d
j
j=1
is the d-th Macaulay representation of ℓ. Then m1 , · · · , mn is a d-lex sequence, if and only if
d
i+d−2
k( j) − n + i − 1
mi =
−∑
.
d −1
j−1
j=1
Proof. The sequence m1 , . . . , mn is a d-lex sequence if and only if Iu = (L≥u ) satisfies mi (Iu ) = mi for all
i = 1, . . . , n, where u is the µ th biggest monomial of degree d. Let us write u = x j(1) · · · x j(d) , 1 ≤ j(1) ≤
· · · ≤ j(d) ≤ n. By the natural decomposition of L<u we have:
d
d
n − j(p) + d − p
ℓ = |L<u | = ∑ dimK [x j(i)+1 , . . . , xn ]d−p+1 = ∑
.
d − p+1
p=1
p=1
k(t)
d
Setting t = d − i + 1 and k(t) = n − j(d − t + 1) + t − 1, we have that ∑t=1
is the dth Macaulay
t
representation of ℓ. The natural decomposition of L<u and (8) show that
d
d
d
i − j(d − t + 1) + t − 2
k(t) − n + i − 1
t
mi ((L<u )) = ∑ mi (x j(d−t+1)+1 , . . . , xn ) = ∑
=∑
.
t −1
t −1
t=1
t=1
t=1
Because
mi (Iu ) = mi (md ) − mi ((L<u )),
we get the conclusion thanks to (8).
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
9
We recall that a homogeneous ideal I ⊂ S is said to be Gotzmann if the number of minimal generators
of mIh ji is the smallest possible for every j ∈ N, namely is equal to:
h ji
n+ j
n+ j−1
−
,
− µj
j+1
j
where µ j is the number of minimal generators of Ih ji . The graded Betti numbers of a Gotzmann ideal
coincide with its associated lexsegment ideal, see [HH1]. Therefore Proposition 2.13 characterizes also
the graded Betti numbers of Gotzmann ideals.
3. T HE
B ETTI NUMBERS OF A GRADED IDEAL
For a fixed ℓ ∈ {1, . . . , n}, d ∈ N and k ≤ ℓ+d−2
ℓ−1 , we denote by u(ℓ, k, d) the kth biggest monomial
u ∈ Sd such that m(u) = ℓ. Or, equivalently, xℓ times the kth biggest monomial in K[x1 , . . . , xℓ ]d−1 . By
U (ℓ, k, d) we denote the ideal of S generated by the set L≥u(ℓ,k,d) ∩ K[x1 , . . . , xℓ ]. Notice that U (ℓ, k, d)
is not a lexsegment in S. However, it is the extension of a lexsegment in K[x1 , . . . , xℓ ]. Furthermore,
U (ℓ, k, d) is obviously a piecewise lexsegment in S. In this section we need to introduce the following
definition: A monomial ideal I ⊂ S generated in one degree is called piecewise lexsegment up to ℓ if
I ∩ K[x1 , . . . , xℓ ] ⊂ K[x1 , . . . , xℓ ] is piecewise lexsegment.
POSSIBLE EXTREMAL
Remark 3.1. Notice that, for all q ∈ N, denoting by m ⊂ S the maximal irrelevant ideal, mqU (ℓ, k, d) ∩
K[x1 , . . . , xℓ ] is equal to U (ℓ, mℓ (mqU (ℓ, k, d)), d + q)) ∩ K[x1 , . . . , xℓ ]. In particular, mqU (ℓ, k, d) is a
piecewise lexsegment up to ℓ.
Lemma 3.2. The ideal U (ℓ, k, d) ⊂ S is the smallest strongly stable ideal containing the biggest k monomials ui ∈ Sd such that m(ui ) = ℓ for all i = 1, . . . , k.
Proof. Let J ⊂ S be the smallest strongly stable ideal containing the biggest k monomials ui ∈ Sd such
that m(ui ) = ℓ for all i = 1, . . . , k. Being the extension of a lexsegment, U (ℓ, k, d) is strongly stable, so
that J ⊂ U (ℓ, k, d). Therefore, let us show the inclusion U (ℓ, k, d) ⊂ J. Let u be a minimal monomial
generator of U (ℓ, k, d). So u has degree d and m(u) ≤ ℓ. Actually, we can assume m(u) < ℓ, otherwise
there is nothing to prove. So let us write:
a
ℓ−1
u = xa11 . . . xℓ−1
.
By definition u > u(ℓ, k, d) = xb11 . . . xbℓ ℓ . Set F = {i : ai > bi }. Because u > u(ℓ, k, d), we have F 6= 0/
and a j = b j for all j < i0 = min{i : i ∈ F}. If |F| = 1, then ai = bi for all i0 < i < ℓ and bℓ = ai0 − bi0 ,
ai −bi
ai −bi
so that u = xi0 0 0 · (u(ℓ, k, d)/xℓ 0 0 ) ∈ J. If |F| > 1, take j > i0 such that a j > b j . The monomial
u′ = xℓ · (u/x j ) is such that u′ > u(ℓ, k, d) and m(u′ ) = ℓ. Therefore u′ ∈ J, so that u = x j · (u′ /xℓ ) belongs
to J too.
The above lemma allows us to characterize the possible extremal Betti numbers of a homogeneous
ideal in a polynomial ring. To this aim, we start with a discussion. To U (ℓ, k, d) we can associate
the numerical sequence (m1 , . . . , mℓ ) where mi = mi (U (ℓ, k, d)). Notice that mℓ = k. By the discussion
done after Proposition 2.3, if V is a strongly stable monomial ideal generated in degree d such that
mℓ (V ) = k, then there must exist a strongly stable piecewise lexsegment ideal U such that mi (U ) = mi (V )
and containing the k biggest monomials u ∈ Sd such that m(u) = ℓ. By Lemma 3.2 U (ℓ, k, d) ⊂ U , so
that mi ≤ mi (V ) for all i. It is possible to characterize the possible numerical sequences like these. To
this purpose, we need to introduce a notion. Given a natural
number a and a positive integer d, consider
the dth Macaulay representation of a, say a = ∑di=1 k(i)
.
For
all integer numbers j, we set:
i
d
k(i) + j
hd, ji
a
=∑
,
i+ j
i=1
where we put qp = 0 whenever p or q are negative, and 00 = 1. Notice that ahd,0i = a and ahd,1i = ahdi .
Lemma 3.3. If k ≤ ℓ+d−2
ℓ−1 , then:
mi (U (ℓ, k, d)) = khℓ−1,i−ℓi ∀ i = 1, . . . , ℓ.
10
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
Furthermore, if i ≥ 2, then khℓ−1,i−ℓi = min{a : k ≤ ahi−1,ℓ−ii }.
Proof. First we will show that, if i ≥ 2, then:
khℓ−1,i−ℓi = min{a : k ≤ ahi−1,ℓ−ii }.
Let us consider the (ℓ − 1)th Macaulay representation of k, namely k = ∑ℓ−1
j=1
ℓ−1
k( j) + i − ℓ
b = khℓ−1,i−ℓi = ∑
.
j+i−ℓ
j=ℓ−i
k( j)
j .
So
If max{ j : k( j) < j} ≥ ℓ−i, then the above one is the (i−1)th Macaulay representation of b: Therefore
bhi−1,ℓ−ii = k, so the statement is obvious in this case.
So we can assume that max{ j : k( j) < j} < ℓ − i. In particular, k(ℓ − i) ≥ ℓ − i, so that the (i − 1)th
Macaulay representation of b − 1 is
ℓ−1
k( j) + i − ℓ
b−1 = ∑
.
j+i−ℓ
j=ℓ−i+1
k( j)
hi−1,ℓ−ii }.
Thus (b− 1)hi−1,ℓ−ii = ∑ℓ−1
j=ℓ−i+1
j , which in this case is smaller than k. So b ≤ min{a : k ≤ a
h( j)
On the other hand, let us consider the (i − 1)th Macaulay representation of b, namely b = ∑i−1
j=1
j . By
[BH, Lemma 4.2.7], we infer the inequality
(h(i − 1), . . . , h(1)) > (k(ℓ − 1) + i − ℓ, . . ., k(ℓ − i + 1) + i − ℓ)
in the lexicographical order. Of course the inequality keeps to be true when shifting of ℓ − i, namely
(h(i − 1) + ℓ − i, . . . , h(1) + ℓ − i) > (k(ℓ − 1), . . . , k(ℓ − i + 1))
in the lexicographical order. Again using [BH, Lemma 4.2.7], we deduce that bhi−1,ℓ−ii > k. So b ≥
min{a : k ≤ ahi−1,ℓ−ii }, that lets us conclude this part.
Let us prove that
mi (U (ℓ, k, d)) = khℓ−1,i−ℓi ∀ i = 1, . . . , ℓ.
The condition k ≤ ℓ+d−1
assures that we can construct V = U (ℓ, k, d). The equality is true for i = 1,
ℓ
hℓ−1,1−ℓi
because k
= 1. From Proposition 2.3 we have, for all i = 2, . . . , ℓ:
mi+1 (V ) ≤ mi (V )hi−1i , mi+2 (V ) ≤ mi+1 (V )hii , . . . , k = mℓ (V ) ≤ mℓ−1 (V )hℓ−2i .
Putting together the above inequalities, we get:
k ≤ mi (V )hi−1,ℓ−ii .
From this and what proved above we deduce that:
mi (V ) ≥ khℓ−1,i−ℓi .
From the above argument and the discussion following Proposition 2.3, it is clear that a piecewise lexsegment monomial space W ⊂ Sd with mi (W ) = khℓ−1,i−ℓi ∀ i = 1, . . . , ℓ must exist. We have V ⊂ W by
Lemma 3.2, so we get also the inequality:
mi (V ) ≤ khℓ−1,i−ℓi .
We introduce the function T : Nr → Nr such that T(v) = (v1 , v1 + v2 , . . . , v1 + v2 + . . . + vr ), where
v = (v1 , . . . , vr ). Furthermore, we define Sq (v) as the last entry of Tq (v).
Remark 3.4. The significance of the above definition is the following: Let I ⊂ S = K[x1 , . . . , xn ] be a
stable ideal generated in one degree. One can easily show that, for all q ∈ N and i ∈ {1, . . . , n},
Sq ((m1 (I), m2 (I), . . . , mi (I))) = mi (mq I).
Notice that we can also rephrase the second condition of Proposition 2.6 as
µi, j ≥ S1 ((µ1, j−1 , µ2, j−1 , . . . , µi, j−1 )).
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
11
Example 3.5. In the next theorem the functions Sq will play a crucial role. Especially, using Remark
3.4, Lemma 3.3 and Remark 3.1, one has:
Sq ((khℓ−1,1−ℓi , khℓ−1,2−ℓi , . . . , khℓ−1,i−ℓi )) = Sq ((m1 (U (ℓ, k, d)), m2 (U (ℓ, k, d)), . . . , mi (U (ℓ, k, d))))
= mi (mqU (ℓ, k, d)))
= mi (U (ℓ, mℓ (mqU (ℓ, k, d)), d + q)))
= Sq ((khℓ−1,1−ℓi , khℓ−1,2−ℓi , . . . , k))hℓ−1,i−ℓi .
Notice that the first time Sq is applied to a vector in Ni , whereas the last time to a vector in Nℓ .
Let I be a homogeneous ideal of S and βi, j = βi, j (I) its graded Betti numbers. Let the extremal Betti
numbers of I be
βi1 ,i1 + j1 , βi2 ,i2 + j2 , . . . , βik ,ik + jk .
Notice that k < n, and up to a reordering, we can assume 0 < i1 < i2 < . . . < ik < n and j1 > j2 > . . . >
jk ≥ 0. If I is a stable ideal then, exploiting the Eliahou-Kervaire formula, one can check that βi,i+ j (I) is
extremal if and only if mi+1, j (I) 6= 0 and m p+1,q (I) = 0 for all (p, q) 6= (i, j) such that p ≥ i and q ≥ j.
In this case, moreover, we have βi,i+ j (I) = mi+1, j (I). Before showing the main result of the paper, we
introduce the following concept.
Definition 3.6. Let i = (i1 , . . . , ik ) and j = ( j1 , . . . , jk ) be such that 0 < i1 < i2 < . . . < ik < n, j1 > j2 >
. . . > jk > 0. We say that I ⊂ S is a (i, j)-lex ideal if I = ∑kp=1 (L p ), where L p is a lexsegment ideal
generated in degree j p in K[x1 , . . . , xi p +1 ].
Theorem 3.7. Let i = (i1 , . . . , ik ) and j = ( j1 , . . . , jk ) be such that 0 < i1 < i2 < . . . < ik < n and j1 >
j2 > . . . > jk > 0, and let b1 , . . . , bk be positive integers. For all p = 1, . . . , k let :
hi ,−i p i
v p = (b p p
hi ,1−i p i
, bp p
hi ,i p−1 −i p i
, . . . , bp p
) ∈ Ni p−1 +1 .
If K has characteristic 0, then the following are equivalent:
(i) There is a homogeneous ideal I ⊂ S with extremal Betti numbers βi p ,i p + j p (I) = b p for all p =
1, . . . , k.
(ii) There is a strongly stable ideal I ⊂ S with extremal Betti numbers βi p ,i p + j p (I) = b p for all p =
1, . . . , k.
(iii) bk ≤ ik +ijkk −1 and S j p − j p+1 (v p+1 ) + b p ≤ i p +ijpp −1 for all p = 1, . . . , k − 1.
(iv) There is an (i, j)-lex ideal I ⊂ S with extremal Betti numbers βi p ,i p + j p (I) = b p for all p = 1, . . . , k.
Proof. (i) ⇐⇒ (ii) follows by by [BCP, Theorem 1.6]. (iv) =⇒ (i) is obvious.
(ii) =⇒ (iii). By what said before the theorem, we can replace βi p ,i p + j p (I) by mi p +1, j p (I) with
mr+1,s (I) = 0 for all (r, s) 6= (i p , j p ) such that r ≥ i p and s ≥ j p . Since mik +1, jk (I) = bk , we have
ik + jk − 1
bk ≤
.
ik
We must have that:
mik−1 +1 m jk−1 − jk (Ih jk i ) + bk−1 = |{monomials u ∈ Ih jk i ∩ S jk−1 with m(u) = ik−1 + 1}|
+|{monomials u ∈ I jk−1 \ Ih jk−1 −1i with m(u) = ik−1 + 1}|
≤ |{monomials u ∈ S jk−1 with m(u) = ik−1 + 1}|
ik−1 + jk−1 − 1
=
ik−1
From the discussion before the theorem, we also have:
hi ,i−i −1i
mi Ih jk i ≥ bk k k
∀ i ≤ ik .
We eventually get:
mik−1 +1 m jk−1 − jk (Ih jk i ) ≥ S jk−1 − jk (vk ).
12
JÜRGEN HERZOG, LEILA SHARIFAN, AND MATTEO VARBARO
Putting together the above inequalities we obtain, for p = k − 1,
ip + jp − 1
p+1
S j p − j p+1 (v ) + b p ≤
,
ip
and we can go on in the same way to show this for all p = 1, . . . , k − 1.
(iii) =⇒ (iv). If bk ≤ ik +ijkk −1 , then we can form U (ik + 1, bk , jk ). Let us call k I = U (ik + 1, bk , jk ).
We have that:
mik−1 +1 (k I)h jk−1 i
= S jk−1 − jk (vk ).
From Remark 3.1, we deduce that
(k I)h jk−1 i ∩ K[x1 , . . . , xik−1 +1 ] = U (ik−1 + 1, S jk−1 − jk (vk ), jk−1 ) ∩ K[x1 , . . . , xik−1 +1 ].
By the assumed numerical conditions, U (ik−1 + 1, S jk−1 − jk (vk ) + bk−1 , jk−1 ) exists and contains exactly
bk−1 new monomials u such that m(u) = ik−1 + 1. Therefore set:
k−1 ′
I = (U (ik−1 + 1, S jk−1 − jk (vk ) + bk−1 , jk−1 )).
and
k−1
I = k I + k−1 I ′ .
By construction k−1 I is a ((ik−1 , ik ), ( jk−1 , jk ))-lex ideal with extremal Betti numbers βik−1 ,ik−1 + jk−1 (k−1 I) =
bk−1 and βik ,ik + jk (k−1 I) = bk . Keeping on with the recursion we will end up with the desired (i, j)-lex
ideal I = 1 I.
Remark 3.8. A different characterization of the extremal Betti numbers was also given by Crupi and
Utano in [CU].
Remark 3.9. For the reader who likes more the language of algebraic geometry, Theorem 3.7 can be
used in the following setting: Let X ⊂ Pn−1 be a projective scheme over a field of characteristic 0 and
IX its ideal sheaf. Then, L
by the graded version of the Grothendieck’s local duality, βi,i+d is an extremal
Betti number of the ideal m∈N Γ(X , IX (m)) ⊂ S if and only if, setting p = n − i − 1 and q = d − 1:
(1) p ≥ 1;
(2) dimK (H p (X , IX (q − p))) = βi,i+d 6= 0.
(3) H r (X , IX (s − r)) = 0 for all (r, s) 6= (p, q) with 1 ≤ r ≤ p and s ≥ q.
Example 3.10. Let us consider the following Betti table:
∗ ∗ ∗ ∗ ∗ ∗
∗ ∗ ∗ a 0 0
∗ ∗ ∗ 0 0 0
∗ ∗ b 0 0 0
0 0 0 0 0 0
···
···
···
···
···
.
Theorem 3.7 implies that there exists a homogeneous ideal in a polynomial ring (of characteristic 0)
whose Betti table looks like the above one (where a and b are extremal) if and only if we are in one of
the following cases:
(i) a = 2 and b = 1, 2;
(ii) a = 1 and b = 1, 2, 3, 4.
In fact, we have b = β2,6 and a = β3,5 . Theorem 3.7 implies a ≤ 4.
If a = 2, then the vector v2 ∈ N3 is:
v2 = (1, 2, 2).
Therefore S2 (v2 ) = 8, and Theorem 3.7 gives 8 + b ≤ 10. So we get b = 1, 2 as desired.
If a = 1, then the vector v2 ∈ N3 is:
v2 = (1, 1, 1).
So S2 (v2 ) = 6, and Theorem 3.7 yields b = 1, 2, 3, 4 as desired.
Eventually, if a > 2, a positive integer b satisfying the conditions of Theorem 3.7 does not exist.
THE POSSIBLE EXTREMAL BETTI NUMBERS OF A HOMOGENEOUS IDEAL
13
R EFERENCES
[AHH] A. Aramova, J. Herzog, T. Hibi, Ideals with stable Betti numbers, Adv. Math. 152 (2000), no. 1, pp. 72–77.
[BCP] D. Bayer, H. Charalambous, S. Popescu, Extremal Betti numbers and applications to monomial ideals, J. Algebra 221
(1999), no. 2, pp. 497–512.
[BH] W. Bruns, J. Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics 39, 1993.
[Co] CoCoA Team, CoCoA: a system for doing Computations in Commutative Algebra, available at
http://cocoa.dima.unige.it.
[CHH] A. Conca, J. Herzog, T. Hibi, Rigid resolutions and big Betti numbers, Comment. Math. Helv. 79 (2004), no. 4, pp.
826–839.
[CU] M. Crupi, R. Utano, Extremal Betti numbers of graded ideals, Result. Math. 43 (2003), pp. 235-244.
[ER] J.A. Eagon, V. Reiner, Resolutions of Stanley Reisner ideals and Alexander duality, J. Pure and Appl. Algebra 130
(1988), no. 3, pp. 265–275.
[Ei]
D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, Graduate Text in Mathematics 150, 1995.
[EK] S. Eliahou, M. Kervaire, Minimal resolutions of some monomial ideals, J. Algebra 129 (1990), no. 1, pp. 1–25.
[HH1] J. Herzog, T. Hibi, Componentwise linear ideals, Nagoya Mathematical Journal 153, (1999), pp. 141–153.
[HH2] J. Herzog, T. Hibi, Monomial ideals, Graduate Texts in Mathematics 260, 2011.
[Mu] S. Murai, Hilbert functions of d-regular ideals, J. Algebra 317 (2007), no. 2, pp. 658-690.
[NR] U. Nagel, T. Römer, Criteria for componentwise linearity, available at arXiv:1108.3921v2, 2011.
FACHBEREICH M ATHEMATIK , U NIVERSIT ÄT D UISBURG -E SSEN , C AMPUS E SSEN , 45117 E SSEN , G ERMANY
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , H AKIM S ABZEVARI U NIVERSITY, S ABZEVAR , I RAN
E-mail address: [email protected]
D IPARTIMENTO DI M ATEMATICA , U NIVERSIT À DEGLI S TUDI DI G ENOVA , I TALY
E-mail address: [email protected]
| 0math.AC
|
F -SIGNATURE AND HILBERT-KUNZ MULTIPICITY:
A COMBINED APPROACH AND COMPARISON
arXiv:1608.02678v2 [math.AC] 28 Oct 2016
THOMAS POLSTRA AND KEVIN TUCKER
Abstract. We present a unified approach to the study of F -signature, Hilbert-Kunz multiplicity, and related limits governed by Frobenius and Cartier linear actions in positive characteristic commutative algebra. We introduce general techniques that give vastly simplified
proofs of existence, semicontinuity, and positivity. Furthermore, we give an affirmative answer to a question of Watanabe and Yoshida allowing the F -signature to be viewed as the
infimum of relative differences in the Hilbert-Kunz multiplicites of the cofinite ideals in a
local ring.
1. Introduction
Throughout this paper, we shall assume all rings R are commutative and Noetherian with
prime characteristic p > 0. Central to the study of such rings is the use of the Frobenius
or p-th power endomorphism F : R → R defined by r 7→ r p for all r ∈ R. Following the
result of Kunz [Kun69] characterizing regularity by the flatness of Frobenius, it has long been
understood that the behavior of Frobenius governs the singularities of such rings. In this
article, we will be primarily concerned with two important numerical invariants that measure
the failure of flatness for the iterated Frobenius: the Hilbert-Kunz multiplicity [Mon83] and
the F -signature [SVdB97, HL02]. Our aim is to revisit a number of core results about these
invariants – existence [Tuc12], semicontinuity [Smi16, Pol15], positivity [HH94, AL03] – and
provide vastly simplified proofs, which in turn yield new and important results. In particular,
we confirm the suspicion of Watanabe and Yoshida allowing the F -signature to be viewed
as the infimum of relative differences in the Hilbert-Kunz multiplicites of the cofinite ideals
in a local ring [WY04, Question 1.10].
For the sake of simplicity in introducing Hilbert-Kunz multiplicity and F -signature, assume
that (R, m, k) is a complete local domain of dimension d and k = k 1/p is perfect. If I ⊆ R
is an ideal with finite colength ℓR (R/I) < ∞, the Hilbert-Kunz multiplicity along I is
e
e
defined by eHK (I) = lime→∞ p1ed ℓR (R/I [p ] ) where I [p ] = (F e (I)) is the expansion of I along
the e-iterated Frobenius. Note that, unlike the case for the Hilbert-Samuel multiplicity, the
e
function e 7→ ℓR (R/I [p ] ) is far from polynomial in pe [HM93]. The existence of Hilbert-Kunz
limits was shown by Monsky [Mon83], and it has recently been shown by Brenner that there
exist irrational Hilbert-Kunz multiplicities [Bre13]. Of particular interest is the Hilbert-Kunz
multiplicity along the maximal ideal m denoted by eHK (R) = eHK (m), where it is known that
eHK (R) ≥ 1 with equality if and only if R is regular [WY00, Theorem 1.5]. More generally,
if the Hilbert-Kunz multiplicity of R is sufficiently small, then R is Gorenstein and strongly
F-regular [BE04, Proposition 2.5] [AE08, Corollary 3.6]. Recently, it has been shown that
The second author is grateful to the NSF for partial support under Grants DMS #1419448 and #1602070,
and for a fellowship from the Sloan Foundation.
1
the Hilbert-Kunz multiplicity determines an upper semicontinuous R-valued function on ring
spectra [Smi16].
e
Another useful perspective comes from viewing the Hilbert-Kunz function ℓR (R/m[p ] ) =
e
µR (R1/p ) as measuring the minimal number of generators of the finitely generated R-module
e
R1/p , the ring of pe -th roots of R inside an algebraic closure of its fraction field. As the
e
e
rank rk R (R1/p ) of this torsion free R-module is ped , we see immediately that R1/p is free –
e
and hence F e is flat and R is regular – if and only if µR (R1/p ) = ped . Simliarly, letting ae =
e
frk(R1/p ) denote the largest rank of a free summand appearing in an R-module direct sum
e
e
decomposition of R1/p , we have that R1/p is free if and only if ae = ped . The limit s(R) =
e
lime→∞ p1ed frk(R1/p ) was first studied in [SVdB97] and revisited in [HL02], where it was
coined the F -signature and shown to exist for Gorenstein rings; existence in full generality
was first shown by the second author in [Tuc12]. Aberbach and Leuschke [AL03] have
shown that the positivity of the F -signature characterizes the notion of strong F -regularity
introduced by Hochster and Huneke [HH94] in their celebrated study of tight closure [HH90].
Recently, it has been shown by the first author that the F -signature determines a lower
semicontinuous R-valued function on ring spectra [Pol15]; an unpublished and independent
proof was simultaneously found and shown to experts by the second author, and has been
incorporated into this article.
At the heart of this work is an elementary argument simultaneously proving the existence
of Hilbert-Kunz multiplicity and F -signature (Theorem 3.2), derived from basic properties
of the functions µR ( ) and frkR ( ). Building upon the philosophy introduced in [Tuc12],
we further attempt to carefully track the uniform constants controlling convergence in the
proof and throughout this article. To that end, we rely heavily on the uniform bounds for
Hilbert-Kunz functions over ring spectra (Theorem 3.3) shown in the work of the first author
[Pol15, Theorem 4.3]. In particular, this also allows us to give vastly simplified proofs of the
upper semicontinuity of Hilbert-Kunz multiplicity (Theorem 3.4) and lower semicontinuity
of the F -signature (Theorem 3.8).
The F -signature has also long been known to be closely related to the relative differences
eHK (I) − eHK (J) in Hilbert-Kunz multiplicities for ideals I ⊆ J with finite colength [HL02,
Theorem 15] (see Lemma 6.1). The infimum among these differences was studied independently by Watanabe and Yoshida [WY04], and the connection to F -signature was made by
Yao [Yao06]. Moreover, Watanabe and Yoshida expected and formally questioned if the
infimum of the relative Hilbert-Kunz differences was equal to the F -signature in [WY04,
Question 1.10]. We confirm their suspicions in Section 6.
Theorem A (Corollary 6.5). If (R, m, k) is an F-finite local ring, then
s(R) =
eHK (I) − eHK (J)
=
inf
eHK (I) − eHK ((I, x)).
I⊆J⊆R, ℓR (R/I)<∞
I⊆R, ℓR (R/I)<∞
ℓR (J/I)
inf
I6=J, ℓR (R/J)<∞
x∈R, (I:x)=m
Much interest in relative differences of Hilbert-Kunz multiplicities stems from the result
of Hochster and Huneke that such differences can be used to detect instances of tight closure
[HH90, Theorem 8.17]. Together with the characterization of strong F -regularity in terms
of the positivity of the F -signature [AL03], these two results are at the core of the use of
asymptotic Frobenius techniques to measure singularities. Building off of previous work of
the second author [BST12, ST14, ST15], we present new and highly simplified proofs of these
2
results. In particular, the first proof of Theorem 5.1 gives a readily computable lower bound
for the F -signature of a strongly F -regular ring.
Using standard reduction to characteristic p > 0 techniques, the singularities governed by
Frobenius in positive characteristic commutative algebra have been closely related to those
appearing in complex algebraic geometry. This connection has motivated several important
generalized settings for tight closure. Hara and Watanabe have defined tight closure for
divisor pairs [HW02] (cf. [Tak04]), and Hara and Yoshida have done the same for ideals with
a real coefficient [HY03]; both of these works build on the important results of Smith [Smi97],
Hara [Har98], and Mehta and Srinivas [MS97]. Previous work of the second author [BST12,
BST13] has extended the notion of F -signature to these settings and beyond, incorporating
the notion of a Cartier subalgebra (see [BS13]). We work to extend all of our results as far
as possible to these settings, including new and simplified proofs of existence (Theorem 4.7),
semicontinuity (Theorems 4.9, 4.11, 4.12), and positivity (Corollaries 5.12, 5.7 (ii.), 5.8 (ii.));
many of these results have not appeared previously. In particular, while Hilbert-Kunz theory
of divisor and ideal pairs has not been introduced, we are able to to give an analogue of
relative Hilbert-Kunz differences in these settings (Corollaries 5.7 (i.), 5.8 (i.)). Once more,
the F -signatures can be viewed as a minimum among the relative Hilbert-Kunz differences
(Corollaries 6.9, 6.10).
e
To give an overview of our methods, for each e ∈ N let IeHK = m[p ] and IeF−sig = (r ∈ R |
e
e
ψ(r 1/p ) ∈ m for all ψ ∈ Hom R (R1/p , R)). As k is perfect, straightforward computations
e
e
show that p1ed µ(R1/p ) = p1ed ℓ(R/IeHK ) and p1ed frk(R1/p ) = p1ed ℓ(R/IeF−sig ), so that the HilbertKunz multiplicity and the F-signature can be understood by studying properties of sequences
of ideals {Ie }e∈N satisfying properties similar to those of {IeHK }e∈N and {IeF−sig }e∈N .
Theorem B (Theorems 4.5, 5.5). Let (R, m, k) be an F-finite local domain of dimension d,
e
and consider a sequence of ideals {Ie }e∈N such that m[p ] ⊆ Ie for each e. Then the limit
η = lime→∞ p1ed ℓ(R/Ie ) exists provided one of the following conditions holds.
[p]
(i.) There exists 0 6= c ∈ R so that cIe ⊆ Ie+1 for all e ∈ N.
1/p
(ii.) There exists a nonzero ψ ∈ Hom R (R1/p , R) so that ψ(Ie+1 ) ⊆ Ie for all e ∈ N.
T
Moreover, in case (ii.), we have η > 0 if and only if e∈N Ie = 0.
It is easy to check that the sequences {IeHK }e∈N or {IeF−sig }e∈N satisfy both conditions (i.)
and (ii.) above (with any choice of c and ψ). Our results on positivity come from the stated
criterion for sequences of ideals satisfying (ii.), while the relative Hilbert-Kunz statements
largely come through careful tracking of uniform constants bounding the growth rates for
sequences satisfying (i.).
The article is organized as follows. Section 2 recalls a number of preliminary statements
about rings in positive characteristic; in particular, we review properties of F -finite rings
which will be the primary setting of this article. We also discuss elementary properties of
the maximal free rank (see Lemma 2.1) which will be used in the unified proof of the existence
of the Hilbert-Kunz multiplicity and the F -signature provided in Section 3. The remainder
of this Section 3 is largely devoted to the new proofs of the semicontinuity. Section 4 is
the technical heart of the paper, and generalizes the methods of the previous sections for
Hilbert-Kunz multiplicity and F -signature alone to the various sequences of ideals satisfying
the conditions in Theorem B above.
3
Having dispatched with the results on existence and semicontinuity, we turn in Section 5
to a discussion of positivity statements. In particular, two simple proofs of the positivity of
the F -signature for strongly F -regular rings appear at the beginning of this section. The
relative Hilbert-Kunz criteria for tight closure along a divisor or ideal pair appear at the end
of this section, together with a discussion of the positivity of the F -splitting ratio. Section 6
is devoted to relating the F-signature with minimal relative Hilbert-Kunz differences giving
a proof of Theorem A, and we conclude in Section 7 with a discussion of a number of related
open questions.
2. Preliminaries
Throughout this paper, we shall assume all rings R are commutative with a unit, Noetherian, and have prime characteristic p > 0. If p ∈ Spec (R), we let k(p) = Rp /pRp denote the
corresponding residue field. A local ring is a triple (R, m, k) where m is the unique maximal
ideal of the ring R and k = k(m) = R/m.
2.1. Maximal Free Rank. For any ring R and R-module M, ℓR (M) denotes the length
of M, and µR (M) denotes the minimal number of generators of M. When ambiguity is
unlikely, we freely omit the subscript R from these and similar notations. We define the
maximal free rank frkR (M) of M to be the maximal rank of a free R-module quotient of M.
As a surjection onto a free module necessarily admits a section, frkR (M) is also the maximal
rank of a free direct summand of M. Equivalently, M admits a direct sum decomposition
M = R⊕ frkR (M ) ⊕ N where N has no free direct summands – a property characterized by
φ(N) 6= R for all φ ∈ Hom R (N, R). It is immediate that frkR (M) ≤ µR (M) with equality if
and only if M is a free R-module.
When R is a domain, we will use rk R (M) to denote the torsion-free rank of M. It is easily
checked that
(1)
frkR (M) ≤ rk R (M) ≤ µR (M)
and both inequalities are strict when M is not free. Notice that, whereas µ( ) is sub-additive
on short exact sequences over local rings, this is not the case for frk( ) and motivates the
following lemma.
Lemma 2.1. Let (R, m, k) be a local ring.
(i.) If M1 and M2 are R-modules, then frk(M1 ⊕ M2 ) = frk(M1 ) + frk(M2 ).
(ii.) If M is an R-module with M ′ ⊆ M a submodule and M ′′ = M/M ′ , then
frk(M ′′ ) ≤ frk(M) ≤ frk(M ′ ) + µ(M ′′ ).
Proof. (i.) If M1 = R⊕ frkR (M1 ) ⊕ N1 and M2 = R⊕ frkR (M2 ) ⊕ N2 where N1 , N2 have no free
direct summands, then M1 ⊕M2 = R⊕(frkR (M1 )+frkR (M2 )) ⊕(N1 ⊕N2 ). If φ ∈ Hom R (N1 ⊕N2 , R),
then φ(N1 ), φ(N2 ) ⊆ m as N1 , N2 have no free direct summands. It follows that φ(N1 ⊕N2 ) =
φ(N1 ) + φ(N2 ) ⊆ m as well, showing N1 ⊕ N2 has no free direct summands as desired.
′′
(ii.) For the first inequality, a surjection M ′′ → R⊕ frk(M ) induces another M → M/M ′ =
′′
M ′′ → R⊕ frk(M ) , showing frk(M ′′ ) ≤ frk(M). To show the final inequality, let n be the
maximal rank of a mutual free direct summand of M and M ′ . In other words, simultaneously
decompose M = R⊕n ⊕ N and M ′ = R⊕n ⊕ N ′ where M ′ ⊆ M is given by equality on
R⊕n and an inclusion N ′ ⊆ N, and we have φ(N ′ ) ⊆ m for every φ : N → R. Taking a
4
surjection Φ : N → R⊕ frk(N ) , it follows Φ(N ′ ) ⊆ m⊕ frk(N ) . Thus, Φ induces a surjection
M ′′ = N/N ′ → k ⊕ frk(N ) and hence also M ′′ /mM ′′ → k ⊕ frk(N ) which shows µ(M ′′ ) ≥ frk(N).
This gives frk(M) = n + frk(N) ≤ frk(M ′ ) + µ(M ′′ ) as desired.
For X any topological space, recall that a function f : X → R is lower semicontinuous if
and only if for any δ ∈ R the set {x ∈ X | f (x) > δ} is open. Similarly, f : X → R is upper
semicontinuous if and only if for any δ ∈ R the set {x ∈ X | f (x) < δ} is open.
Lemma 2.2. If M is a finitely generated R-module, the function Spec (R) → R given by
p 7→ frkRp (Mp ) is lower semicontinuous. Similarly, the function Spec (R) → R given by
p 7→ µRp (Mp ) is upper semicontinuous.
Proof. For p ∈ Spec (R), let k(p) = Rp /pRp denote the corresponding residue field. If
frkRp (Mp ) = n, we can find a surjection Mp → Rp⊕n . Without loss of generality, since
M is finitely generated, we may assume this surjection is the localization of an R-module
homomorphism M → R⊕n . As it becomes surjective when localized at p, there is some
g ∈ R \ p so that Mg → Rg⊕n is surjective. Localizing at any q ∈ Spec (R) with g 6∈ q yields
a surjection Mq → Rq⊕n , implying frkRq (Mq ) ≥ n = frkRp (Mp ) for any q in the Zariski open
neighborhood D(g) = {q ∈ Spec (R) | g 6∈ q} of p in Spec (R). If δ ∈ R and frkRp (Mp ) > δ,
then so also frkRq (Mq ) > δ for any q ∈ D(g). This shows the function Spec (R) → R given by
p 7→ frkR (Mp ) is lower semicontinuous. That the function Spec (R) → R given by p 7→ µ(Mp )
is upper semicontinuous proceeds in an analogous manner.
2.2. F-finite rings. The Frobenius or p-th power endomorphism F : R → R is defined by
e
r 7→ r p for all r ∈ R. Similarly, for e ∈ N, we have F e : R → R given by r 7→ r p . The
e
expansion of an ideal I over F e is denoted I [p ] = (F e (I)). In case R is a domain, we let
e
R1/p denote the subring of pe -th roots of R inside a fixed algebraic closure of the fraction
e
field of R. By taking pe -th roots, R1/p is abstractly isomorphic to R, and the Frobenius
e
map is identified with the ring extension R ⊆ R1/p . More generally, for any R and any
e
R-module M, we will write M 1/p for the R-module given by restriction of scalars for F e ;
this is an exact functor on the category of R modules. We say that R is F -finite when
e
R1/p is a finitely generated R-module, which further implies that R is excellent, [Kun76,
Theorem 2.5]. A complete local ring (R, m, k) is F -finite if and only if its residue field k is
e
F -finite. When R is F -finite and M is a finitely generated R-module, M 1/p is also a finitely
generated R-module and
e
e
e
e
ℓR (M 1/p ) = [k 1/p : k] · ℓR1/pe (M 1/p ) = [k 1/p : k] · ℓR (M).
In particular, note that
(2)
e
e
e
e
e
e
µR (M 1/p ) = ℓR (M 1/p ⊗R k) = ℓR ((M/m[p ] M)1/p ) = [k 1/p : k] · ℓR (M/m[p ] M).
Lemma 2.3. Suppose (R, m, k) is a complete local F -finite domain with coefficient field
k and system of parameters x1 , . . . , xd chosen so that R is module finite and generically
separable over the regular subring A = k[[x1 , . . . , xd ]]. Then there exists 0 6= c ∈ A so that
e
e
e
c · R1/p ⊆ R[A1/p ] = R ⊗A A1/p for all e ∈ N.
Proof. Note first that, by the Cohen-Gabber Structure Theorem, every complete local ring
admits such a coefficient field and system of parameters; see [KS15] for an elementary proof.
Let K = Frac(A) ⊆ L = Frac(R) be the corresponding extension of fraction fields. The
5
e
e
separability assumption implies L and K 1/p are linearly disjoint over K, so that L⊗K K 1/p =
e
e
e
LK 1/p is reduced. Since A1/p is a free A-module, R ⊗A A1/p is a free R-module and injects
e
e
e
into its localization L ⊗A A1/p = L ⊗K K 1/p . Thus R ⊗A A1/p is also reduced, which gives
e
e
the equality R ⊗A A1/p = R[A1/p ]. If e1 , . . . , es ∈ R generate R as an A-module, it is easy
e
e
to see c = (det (tr(ei ej )))2 satisfies c · R1/p ⊆ R[A1/p ] where tr is the trace map; see [HH00,
Lemma 4.3] for a proof and [HH90, Section 6] for further discussion.
The Corollaries below follow immediately; see [Kun76] for details.
e
Corollary 2.4. If (R, m, k) is an F -finite local domain of dimension d, then rk R (R1/p ) =
e
[k 1/p : k] · ped .
Corollary 2.5. If R is F -finite, for any two prime ideals p ⊆ q of R it follows
e
e
[k(p)1/p : k(p)] = [k(q)1/p : k(q)] · pe dim(Rq /pRq ) .
If additionally R is locally equidimensional, we have
e
e
[k(p)1/p : k(p)] · pe dim(Rp ) = [k(q)1/p : k(q)] · pe dim(Rq )
e
and the function Spec (R) → N given by p 7→ [k(p)1/p : k(p)] · pe dim(Rp ) is constant on the
connected components of Spec (R).
Corollary 2.6. If R is a locally equidimensional F -finite reduced ring with connected spectrum and pγ = [k(p)1/p : k(p)] · pdim(Rp ) for all p ∈ Spec (R), there exist short exact sequences
(3)
0
(4)
0
R⊕p
/
/
γ
R1/p
/
R1/p
/
R⊕p
M
/
γ
/
N
/
/
0
0
so that dim(M) < d and dim(N) < d.
3. F-signature and Hilbert-Kunz Multiplicity Revisited
The first goal of this section is to provide an elementary proof that simultaneously shows
the existence of the F -signature and Hilbert-Kunz multiplicity for an F -finite local domain
(R, m, k). At the heart of all known proofs of existence is a basic observation on the growth
rate of a Hilbert-Kunz function.
Lemma 3.1. [Mon83, Lemma 1.1] Let (R, m, k) be a local ring and M a finitely generated
R-module. Then there exists a positive constant C(M, m) ∈ R so that
e
ℓ(M/m[p ] M) ≤ C(M, m) · pe dim(M )
for each e ∈ N.
e
e
e
Proof. If m is generated by t elements, then mtp ⊆ m[p ] for each e ∈ N. Thus ℓ(M/m[p ] M) ≤
e
ℓ(M/mtp M), which is eventually polynomial of degree dim(M) in the entry pe .
Theorem 3.2. [Mon83, Theorem 1.8] [Tuc12, Theorem 4.9] Suppose that (R, m, k) is an
F -finite local domain and dim(R) = d. Then the limits
e
e
frkR (R1/p )
s(R) = lim 1/pe
e→∞ [k
: k] · ped
µR (R1/p )
eHK (R) = lim 1/pe
e→∞ [k
: k] · ped
exist.
6
e
Proof. Let pγ = [k 1/p : k] · pd = rk R (R1/p ) and ν( ) denote either µ( ) or frk( ). Set
e
ηe = p1eγ νR (R1/p ), and observe using Lemma 3.1 with (1) and (2) that the sequence {ηe }e∈N
is bounded above. Put η + = lim supe→∞ ηe and η − = lim inf e→∞ ηe .
As in Lemma 2.6 (2), fix a short exact sequence
0
R⊕p
/
γ
R1/p
/
/
M
/
0
e
where dim(M) < d. Applying the exact functors ( )1/p gives the short exact sequences
0
/
e
(R1/p )⊕p
γ
/
R1/p
e+1
/
M 1/p
e
/
0
and Lemma 2.1 implies
ν(R1/p
e+1
e
e
) ≤ pγ · ν(R1/p ) + µ(M 1/p )
for each e ∈ N. By Lemma 3.1 and using (2), there exists a positive constant C(M, m) ∈
e
R with µ(M 1/p ) ≤ C(M, m)pe dim(M ) ≤ C(M, m)pe(d−1) . Dividing through by p(e+1)d and
setting C = C(M,m)
∈ R yields ηe+1 ≤ ηe + pCe . Iterating this inequality gives
pd
1
2C
1
C
ηe+e′ ≤ ηe + e 1 + + · · · + e′ −1 ≤ ηe + e
p
p
p
p
for all e, e′ ∈ N. For each e, taking lim supe′ →∞ then gives η + ≤ ηe +
lim inf e→∞ gives η + ≤ η − as desired.
2C
.
pe
Now taking
Our next goal is to give a direct proof of the upper semicontinuity of Hilbert-Kunz multiplicity. The key observation is that the constant appearing in Lemma 3.1 can be taken
uniformly on ring spectra.
Theorem 3.3. [Pol15, Theorem 4.3] For any F -finite ring R and finitely generated module
M, there exists a positive constant C(M) such that for all p ∈ Spec (R)
e
ℓRp (Mp /p[p ] Mp ) ≤ C(M)pe dim(Mp ) .
From the proof of Theorem 3.2 above, which is valid without change for a local equidimensional reduced ring, it follows that there is a positive constant C ∈ R such that
1/pe
1/pe
frkRp (Rp )
2C
+ e
s(Rp ) ≤
e
e
dim(R
)
1/p
p
: k(p)] · p
[k(p)
p
µRp (Rp )
2C
+ e
eHK (Rp ) ≤
e
e
dim(R
)
1/p
p
: k(p)] · p
[k(p)
p
for all p ∈ Spec (R) and all e ∈ N.
Theorem 3.4. [Smi16, Main Result] For any F -finite locally equidimensional reduced ring
R, the function eHK : Spec (R) → R given by p 7→ eHK (Rp ) is upper-semicontinuous.
Proof. Restricting to a connected component, we may assume without loss of generality that
Spec (R) is connected and hence pγ = [k(p)1/p : k(p)] · pdim(Rp ) is constant for all p ∈ Spec (R)
by Corollary 2.5. If δ ∈ R and eHK (Rq ) < δ for some q ∈ Spec (R), then
1/pe
µRq (Rq )
< δ−ǫ
1/p
[k(q) e : k(q)] · pe dim(Rq )
7
for some 0 < ǫ ≪ 1 and some e ∈ N with
p in a neighborhood of q. This yields
2C
pe
< ǫ. By Lemma 2.2, the same holds true for all
1/pe
µRp (Rp )
2C
2C
eHK (Rp ) ≤
+ e < δ−ǫ+ e < δ
e
1/p
e
dim(R
)
p
[k(p)
: k(p)] · p
p
p
as desired and completes the proof.
To get a similar argument for lower semicontinuity of F -signature, we need to reverse
the estimates arising in the proof of existence. To that end, we first record the following
elementary lemma.
Lemma 3.5. Let p a prime number, d ∈ N, and {λe }e∈N be a sequence of real numbers so
that { p1ed λe }e∈N is a bounded.
1
(i.) If there exists a positive constant C ∈ R so that p(e+1)d
λe+1 ≤ p1ed λe + pCe for all e ∈ N,
for all e ∈ N.
then the limit λ = lime→∞ p1ed λe exists and λ − p1ed λe ≤ 2C
pe
1
1
λe+1 + pCe for all e ∈ N,
(ii.) If there exists a positive constant C ∈ R so that ped λe ≤ p(e+1)d
for all e ∈ N.
then the limit λ = lime→∞ p1ed λe exists and p1ed λe − λ ≤ 2C
pe
1
(iii.) If there exists a positive constant C ∈ R so that | p(e+1)d λe+1 − p1ed λe | ≤ pCe for all e ∈ N,
for all e ∈ N. In particular,
then the limit λ = lime→∞ p1ed λe exists and | p1ed λe −λ| ≤ 2C
pe
ed
e(d−1)
λe = λp + O(p
).
Proof. Let λ+ = lim supe→∞ p1ed λe and λ− = lim inf e→∞ p1ed λe , which are finite as { p1ed λe }e∈N
is bounded. Iterating the inequality in (i.) yields
1
1
1
1
C
1
2C
′ ≤
1
+
≤
λ
λ
+
+
·
·
·
+
λ
+
e+e
e
e
′
′
p(e+e )d
ped
pe
p
pe −1
ped
pe
, and then
for all e, e′ ∈ N. Taking lim supe′ →∞ for each fixed e ∈ N gives λ+ ≤ p1ed λe + 2C
pe
1
1
+
−
applying lim inf e→∞ gives λ ≤ λ so that λ = lime→∞ ped λe exists and λ − ped λe ≤ 2C
for
pe
all e ∈ N. Similarly, iterating the inequality in (ii.) yields
1
C
2C
1
1
1
1
λe ≤ (e+e′ )d λe+e′ + e 1 + + · · · + e′ −1 ≤ (e+e′ )d λe+e′ + e
ed
p
p
p
p
p
p
p
, and then
for all e, e′ ∈ N. Taking lim inf e′ →∞ for each fixed e ∈ N gives p1ed λe ≤ λ− + 2C
pe
1
1
+
−
applying lim supe→∞ gives λ ≤ λ so that λ = lime→∞ ped λe exists and ped λe − λ ≤ 2C
for
pe
all e ∈ N. The final statement (iii.) follows immediately from a combination of (i.) and
(ii.).
Theorem 3.6. If R is a locally equidimensional reduced F -finite ring of dimension d, there
is a positive constant C ∈ R so that
1/pe
1/pe
µRp (Rp )
C
− eHK (Rp ) ≤ e
e
1/p
e
dim
R
p
[k(p)
: k(p)] · p
p
frkRp (Rp )
C
− s(Rp ) ≤ e
e
1/p
e
dim
R
p
[k(p)
: k(p)] · p
p
for all e ∈ N and p ∈ Spec (R).
8
Proof. Restricting to a connected component, we may assume without loss of generality that
Spec (R) is connected and hence pγ = [k(p)1/p : k(p)] · pdim(Rp ) is constant for all p ∈ Spec (R)
by Corollary 2.5.
The proofs of Theorems 3.2 and 3.4 above relied on a short exact sequence as in Lemma 2.6
(2) and produced inequalities as in Lemma 3.5 (i.). Repeating those same arguments on a
short exact sequence as in Lemma 2.6 (3) yields inequalities as in Lemma 3.5 (ii.), which
combine to give the desired result.
Corollary 3.7. If (R, m, k) is an F -finite local domain of dimension d,
e
e
µR (R1/p )
= eHK (R)ped + O(pe(d−1) )
[k : k pe ]
frkR (R1/p )
= s(R)ped + O(pe(d−1) ).
[k : k pe ]
Theorem 3.8. [Pol15, Theorem 5.7] For any F -finite locally equidimensional reduced ring
R, the function s : Spec (R) → R given by p 7→ s(Rp ) is lower-semicontinuous.
Proof. From Theorem 3.6, there is a C ∈ R such that
1/pe
frkRp (Rp )
C
≤ s(Rp ) + e
e
1/p
e
dim(R
)
p
[k(p)
: k(p)] · p
p
for all p ∈ Spec (R) and all e ∈ N. The argument in Theorem 3.4 immediately gives the
desired result. Alternatively, using the whole of Theorem 3.6, both semicontinuity statements
follow from the fact that the uniform limit of upper ([Kun76, Corollary 3.4]) or lower ([EY11,
Corollary 2.5]) semicontinuous functions is upper or lower semicontinuous, respectively.
4. Limits via Frobenius and Cartier linear maps
4.1. Background. Suppose that M is an R-module. Recall that an additive map φ ∈
Hom Z (M, M) is said to be Frobenius linear or p-linear if φ(rm) = r p φ(m) for all r ∈ R
e
and m ∈ M. Similarly, we say that φ ∈ Hom Z (M, M) is pe -linear if φ(rm) = r p m for all
r ∈ R and m ∈ M. The set of all pe -linear maps on M is denoted Fe (M) and is naturally
e
both a left and right R-module readily identified with Hom R (M, M 1/p ). The composition
of a pe1 -linear map φ1 and a pe2 -linear map φ2Lyields a pe1 +e2 -linear map φ1 ◦ φ2 , so that the
ring of Frobenius linear operators F (M) =
e≥0 Fe (M) on M forms a non-commutative
N-graded ring with R ⊆ F0 (M) = Hom R (M, M). In the case M = R, we have that every
e
e
pe -linear map φ ∈ Hom R (R, R1/p ) = R1/p is a post-multiple of the Frobenius F e by some
e
c1/p ∈ R.
Dual to p-linear maps are the Cartier linear or p−1 -linear maps; an additive map φ ∈
Hom Z (M, M) is said to be p−1 -linear if φ(r p m) = rφ(m) for all r ∈ R and m ∈ M. Similarly,
e
we say that φ ∈ Hom Z (M, M) is p−e -linear if φ(r p m) = rm for all r ∈ R and m ∈ M. The
set of all p−e -linear maps on M is denoted Ce (M) and is naturally both a left and right
e
R-module readily identified with Hom R (M 1/p , M). The composition of a p−e1 -linear map
−(e1 +e2 )
φ1 and a p−e2 -linear map φ
-linear map φ1 ◦ φ2 , so that the ring of Cartier
2 yields a p
L
linear operators C(M) = e≥0 Ce (M) on M forms a non-commutative N-graded ring with
R ⊆ C0 (M) = Hom R (M, M).
e
e
If we have φ1 ∈ Hom R (M 1/p 1 , M) and φ2 ∈ Hom R (M 1/p 2 , M) we write φ1 · φ2 for the
e
composition φ1 ◦ (φ2 )1/p 1 , which coincides with their product when viewed as elements of
e
C(M). In particular, given φ ∈ Hom R (M 1/p , M) we write φe ∈ Hom R (M 1/p , M) for the
9
e
e
corresponding e-th iterate. If ψ ∈ Hom R (M 1/p , M) we let ψ(r 1/p ·
map
M 1/p
given by pre-multiplying with r
1/pe
e
∈R
r 1/p
e
1/pe
.
/
M 1/p
e
ψ
/
) denote the R-linear
R
Lemma 4.1. Let R be an F-finite domain of dimension d with rk R (R1/p ) = pγ .
(i.) If 0 6= c ∈ R, then there exists a short exact sequence
0
R⊕p
/
γ
Φ
/
R1/p
M
/
/
0
γ
so that dim(M) < R and Φ(R⊕p ) ⊆ (cR)1/p .
(ii.) Let 0 6= ψ ∈ Hom R (R1/p , R) be a nonzero map. There exists a short exact sequence
0
/
R1/p
Ψ
/
R⊕p
γ
/
N
/
0
so that every component function of Ψ is a pre-multiple of ψ. In other words, there
1/p
1/p
exists r1 , ..., rpγ ∈ R so that Ψ = (ψ(r1 · ), ..., ψ(rpγ · )).
γ
Proof. For (i.), start with any inclusion R⊕p ⊆ R1/p with a torsion quotient as in (3) and
post-multiply by c1/p on R1/p . For (ii.), note that rk R1/p (Hom R (R1/p , R)) = 1 so that there
is some non-zero z ∈ R with z 1/p · Hom R (R1/p , R) ⊆ ψ · R1/p . Start with an inclusion
γ
R1/p ⊆ R⊕p with torsion quotient as in (4) and pre-multiply by z 1/p on R1/p to achieve the
desired sequence.
If R is an F -finite normal domain and D is a (Weil) divisor on X = Spec (R), we use R(D)
to denote Γ(X, OX (D)). There is a well known correspondence between p−e -linear maps and
certain effective Q-divisors. Fixing a canonical divisor KR , standard duality arguments for
e
e
finite extensions show that Hom R (R1/p , R) = (R((1 − pe )KR ))1/p for each e ∈ N. Thus,
e
to each 0 6= φ ∈ Hom R (R1/p , R) we can associate a divisor Dφ so that Dφ ∼Z (1 − pe )KR ,
e
and we set ∆φ = pe1−1 Dφ . If φ, ψ ∈ Hom R (R1/p , R), then ∆φ ≥ ∆ψ if and only if φ( ) =
e
ψ(r 1/p · ) for some 0 6= r ∈ R, in which case ∆φ = ∆ψ + pe1−1 divR (r). Moreover, if
e
e
e
φ1 ∈ Hom R (R1/p 1 , R) and φ2 ∈ Hom R (R1/p 2 , R) with φ = φ1 · φ2 = φ1 ◦ (φ2 )1/p 1 , then
∆φ = pe11+e2 ((pe1 − 1)∆φ1 + pe1 (pe2 − 1)∆φ2 ). In particular, ∆φ = ∆φn for all e, n ∈ N and all
e
φ ∈ Hom R (R1/p , R). See [ST12, Section 4] or [ST14, Section 2] for further details.
Lemma 4.2. Let R by an F -finite domain and 0 6= ψ ∈ Hom R (R1/p , R). There exists
e
an element 0 6= z ∈ R so that, for any e ∈ N and any φ ∈ Hom R (R1/p , R) the map
e
e
e
e
e
z · φ( ) = φ(z · ) = φ((z p )1/p · ) = ψ e (r 1/p · ) for some r 1/p ∈ R1/p . In other words,
zφ is a pre-multiple of ψ e .
Proof. Let R be the normalization of R in its fraction field, and take 0 6= c ∈ R inside the
conductor ideal c = Ann R (R/R). One can show [BS13, Exercise 6.14] that every p−e -linear
e
map φ : R1/p → R extends uniquely to a p−e -linear map on R compatible with c, which we
1/pe
→ R. Let 0 6= x ∈ R be such that divR (x) ≥ ∆ψ and put z = cx.
denoted here by φ : R
e
pe
Then ∆xφ = ∆φ + pe −1 divR (x) ≥ ∆ψ = ∆ψe so that φ(x · ) = ψ e (r1/p · ) for some r ∈ R.
e
e
e
e
Letting r 1/p = cr 1/p ∈ R1/p and restricting back to R, this gives φ(z · ) = ψ e (r 1/p · )
as desired.
10
Recall that a Cartier subalgebra on R is a subring D ⊆ C(R) of the ring of Cartier linear
operators on R. Cartier subalgebras can be seen as a natural generalization of a number
of commonly studied settings in positive characteristic commutative algebra. For instance,
if R is an F -finite domain and 0 6= a ⊆ R is an ideal and t ∈ R≥0 , the Cartier subalgebra
L
t
t
C a = e≥0 Cea where
t
Cea
e
e
e
= a⌈t(p −1)⌉/p Hom R (R1/p , R)
e
e
e
e
= {φ(x1/p · ) | x ∈ a⌈t(p −1)⌉/p and φ ∈ Hom R (R1/p , R)}
recovers the framework of [HY03]. Similarly, if R is an F -finite normal domain and ∆ is an
L
(R,∆)
effective Q-divisor on Spec (R), the Cartier subalgebra C (R,∆) = e≥0 Ce
where
(R,∆)
Ce
e
= {φ ∈ Hom R (R1/p , R) | ∆φ ≥ ∆}
e
e
= im Hom R (R(⌈(pe − 1)∆⌉)1/p , R) → Hom R (R1/p , R)
recovers the setting of [HW02, Tak04]. For more information on Cartier subalgebras, see
[BS13]. We will mainly be interested in the generalization of F -signature to a Cartier subalgebra D introduced in [BST12, BST13].
4.2. Existence and semicontinuity.
Theorem 4.3. Let (R, m, k) be an F -finite local domain of dimension d, and {Ie }e∈N a
e
sequence of ideals such that m[p ] ⊆ Ie for all e ∈ N.
[p]
(i.) If there exists 0 6= c ∈ R so that cIe ⊆ Ie+1 for all e ∈ N, then η = lime→∞ p1ed ℓR (R/Ie )
exists. Moreover, there exists a positive constant C(c) depending only on c ∈ R with
η − p1ed ℓR (R/Ie ) ≤ C(c)
for all e ∈ N.
pe
1/p
(ii.) If there exists a non-zero R-linear map ψ : R1/p → R so that ψ(Ie+1 ) ⊆ Ie for all
e ∈ N, then η = lime→∞ p1ed ℓR (R/Ie ) exists. Moreover, there exists a positive constant
C(ψ) ∈ R depending only on ψ such that
1
ℓ (R/Ie )
ped R
−η ≤
C(ψ)
pe
for all e ∈ N.
Remark 4.4. Note that the conditions on ideal sequences {Ie }e∈N in Theorem 4.3 (i.) and
(ii.) are far more symmetric when phrased in terms of p-linear and p−1 -linear maps. In
(i.), the requirement is simply that there exists a p-linear map 0 6= φ ∈ F1 (R) on R such
that φ(Ie ) ⊆ Ie+1 for each e ∈ N. Similarly, for (ii.) the requirement is that there exists a
p−1 -linear map 0 6= φ ∈ C1 (R) such that φ(Ie+1) ⊆ Ie for each e ∈ N.
Proof of Theorem 4.3. For (i.), put rk R (R1/p ) = [k 1/p : k] · pd = pγ and consider a short
exact sequence
(5)
0
/
R⊕p
γ
Φ
/
R1/p
/
M
/
0
⊕pγ
as in Lemma 4.1 (i.) with dim(M) < dim(R) and Φ(R ) ⊆ (Rc)1/p . Note that the
e
condition cI [p ] ⊆ Ie+1 for all e ∈ N can be restated as Ie (Rc)1/p ⊆ (Ie+1 )1/p , so that
γ
γ
Φ(Ie⊕p ) = Φ(Ie (R⊕p )) ⊆ Ie (Rc)1/p ⊆ (Ie+1 )1/p . In particular, Φ induces a quotient map
γ
Φ : (R/Ie )⊕p → (R/Ie+1 )1/p
and it follows that
[k 1/p : k]ℓR (R/Ie+1 ) = ℓR ((R/Ie+1 )1/p ) ≤ pγ ℓR (R/Ie ) + ℓR (coker (Φ)).
11
e+1
e
Since m[p ] ⊆ Ie+1 and coker (Φ) is a quotient of (R/Ie+1)1/p , we have m[p ] ⊆ Ann R (coker (Φ)).
e
But coker (Φ) is also a quotient of coker (Φ) = M and thus M/m[p ] M, so that ℓR (coker (Φ)) ≤
e
ℓR (M/m[p ] M) ≤ C(M)pe(d−1) by Lemma 3.1. Dividing through by [k 1/p : k]p(e+1)d = pγ+ed
yields
1
C(M, m)/pγ
1
ℓ
(R/I
)
≤
ℓ
(R/I
)
+
R
e+1
R
e
p(e+1)d
ped
pe
and the result now follows from Lemma 3.5 with C(c) = 2C(M, m)/pγ independent of the
sequence of ideals {Ie }e∈N satisfying the condition in (i.).
Similarly for (ii.), consider a short exact sequence
0
/
R1/p
Ψ
/
R⊕p
γ
/
N
/
0
as in Lemma 4.1 (ii.) with dim(N) < dim(R) so that every component function of Ψ is a
γ
pre-multiple of ψ. It follows that Ψ((Ie+1)1/p ) ⊆ Ie⊕p , so that Ψ induces a quotient map
Ψ : (R/Ie+1 )1/p → (R/Ie )⊕p
and thus
γ
pγ ℓR (R/Ie ) ≤ ℓR ((R/Ie+1 )1/p ) + ℓR (coker (Ψ))
= [k 1/p : k]ℓR (R/Ie+1 ) + ℓR (coker (Ψ)).
γ
e
e
Since m[p ] ⊆ Ie and coker (Ψ) is a quotient of (R/Ie )⊕p , we have that m[p ] ⊆ Ann R (coker (Ψ)).
e
But coker (Ψ) is also a quotient of coker (Ψ) = N and thus N/m[p ] N, so that ℓR (coker (Ψ)) ≤
e
ℓR (N/m[p ] N) ≤ C(N)pe(d−1) by Lemma 3.1. Dividing through by [k 1/p : k]p(e+1)d = pγ+ed
yields
1
1
C(N, m)/pγ
ℓ
(R/I
)
≤
ℓ
(R/I
)
+
R
e
R
e+1
ped
p(e+1)d
pe
and the result now follows from Lemma 3.5 with C(ψ) = 2C(N, m)/pγ independent of the
sequence of ideals {Ie }e∈N satisfying the condition in (ii.).
Corollary 4.5. Let (R, m, k) be an F -finite local domain of dimension d, and {Ie }e∈N a
e
sequence of ideals such that m[p ] ⊆ Ie for all e ∈ N. Suppose there exists 0 6= c ∈ R so that
[p]
1/p
cIe ⊆ Ie+1 for all e ∈ N, and a non-zero R-linear map ψ : R1/p → R so that ψ(Ie+1 ) ⊆ Ie
for all e ∈ N. Then η = lime→∞ p1ed ℓR (R/Ie ) exists, and there is a positive constant C ∈ R
such that | p1ed ℓR (R/Ie ) − η| ≤ pCe for all e ∈ N and all such sequence of ideals {Ie }e∈N . In
particular, ℓR (R/Ie ) = ηped + O(pe(d−1) ).
Corollary 4.5 gives yet another perspective on the existence proofs for Hilbert-Kunz
multiplicity and F -signature in terms of the properties of certain sequences of ideals. If
e
e
(R, m, k) is a local domain and we set IeHK = m[p ] and IeF−sig = (r ∈ R | φ(r 1/p ) ∈
e
e
e
m for all φ ∈ Hom R (R1/p , R)), it is easy to see that µR (R1/p ) = ℓR ((R/IeHK )1/p ) and
e
e
frkR (R1/p ) = ℓR ((R/IeF−sig )1/p ). It is shown in [Tuc12] that both sequences satisfy the conditions of Theorem 4.5; in fact, any choice of 0 6= c ∈ R and 0 6= ψ ∈ Hom R (R1/p , R) will
suffice.
As in [Tuc12], in settings such as Theorem 4.3, we have opted to consider sequences of
e
ideals {Ie }e∈N with m[p ] ⊆ Ie . In light of Theorem 3.3, uniformity of constants over Spec (R)
would seem more transparent for such sequences. Nonetheless, for any m-primary ideal J,
e
e
one could consider sequences with J [p ] ⊆ Ie . One has m[p 0 ] ⊆ J for some e0 ∈ N, so
12
e+e
e
that m[p 0 ] ⊆ J [p ] ⊆ Ie for such sequences. In particular, our techniques carry over to
this setting after allowing for reindexing, and allows one to recover Monsky’s original limit
existence result.
Corollary 4.6. [Mon83] If (R, m, k) is a local domain of dimension d and J ⊆ R is an
m-primary ideal, then
1
e
eHK (R, J) = lim ed ℓR (R/J [p ] )
e→∞ p
exists and there is a positive constant C ∈ R such that
1
C
e
|eHK (R, J) − ed ℓR (R/J [p ] )| < e
p
p
for all e.
Another useful modification of the setup of Theorem 4.3 proceeds in the following manner.
Fixing e0 ∈ N, one can modify the condition on the sequences of ideals {Ie }e∈N of Theorem 4.3
[pe0 ]
⊆ Ie+e0 for all e ∈ N. Similarly, in Theorem 4.3 (ii.), one can ask
(i.) to require that cIe
e
e
that there exists 0 6= ψ ∈ Hom R (R1/p 0 , R) such that ψ((Ie+e0 )1/p 0 ) ⊆ Ie for all e ∈ N.
In both cases, after reindexing, the same methods apply to yield analogous results. For
example, this last generalization is particularly relevant when working with an arbitrary
Cartier subalgebra D on an local F -finite domain (R, m, k). If ΓD = {e ∈ N | such that De 6=
0} is the semi-group of D and e ∈ ΓD , one defines the e-th F -splitting number aD
e of R along
D to be the maximal number of copies of R with projection maps in De appearing in an
e
e
R-module direct sum decomposition of R1/p . Setting IeD = (r ∈ R | φ(r 1/p ) ∈ m for all φ ∈
D 1/pe
De ), it is again easy to check that aD
), and the method of Theorem 4.3 (ii.)
e = ℓR ((R/Ie )
yields the following result.
Theorem 4.7. Let (R, m, k) be an F -finite local domain of dimension d and D a Cartier
subalgebra on R. Then the F -signature s(R, D) = lime→∞ [k1/pe1:k]·ped aD
e of R along D exists,
e∈ΓD
and there is a positive constant C ∈ R so that
C
1
aD
s(R, D) − 1/pe
e ≤ e
ed
[k
: k] · p
p
for all e ∈ ΓD .
Proof. Let pγ = rk R (R1/p ) = [k 1/p : k] · pd and take a set1 of generators e1 , . . . , es for the
semigroup ΓD . Fixing 0 6= ψi ∈ Dei for each i = 1, . . . , s, we can find a short exact sequence
of R-modules
(6)
0
/
e
R1/p i
Ψi
/
R⊕p
γei
/
Mi
/
0
where dim(Mi ) < d and every component function of Ψi is a pre-multiple of ψi as in
Lemma 4.1 (ii.). Since the Cartier linear maps in D are closed under composition, it fole+e
e+e
D
D
lows readily that ψi ((Ie+e
)1/p i ) ⊆ IeD for any e ∈ ΓD . In particular, Ψi ((Ie+e
)1/p i ) ⊆
i
i
γe
(IeD )⊕p i and proceeding as in Theorem 4.3 (ii.), we see that
1 D C′
1
D
a
≤
a + e
p(e+ei )γ e+ei
peγ e
p
1Every
sub-semigroup of N is finitely generated [Gri01, Proposition 4.1].
13
for any i = 1, . . . , s and e ∈ ΓD where C ′ = max{C(M1 , m)/pe1 γ , . . . , C(Ms , m)/pes γ }. It
1 D
2C ′
follows as in Lemma 3.5 (ii.) that s(R, D) = lime→∞ p1eγ aD
e exists and s(R, D) − peγ ae ≤ pe
e∈ΓD
for all e ∈ ΓD as desired.
Remark 4.8 (cf. [Pol15] condition (1)). Consider a Cartier subalgebra D on an F -finite local
domain (R, m, k) with the following property: if φ ∈ De+1 for some e ∈ N, then φ|R1/pe ∈ De .
D
It is easy to check (IeD )[p] ⊆ Ie+1
for all e ∈ N = ΓD , and Theorem 4.3 (i.) gives that there
is a positive constant C ∈ R so that
C
1
aD
s(R, D) − 1/pe
(7)
e ≤ e
ed
[k
: k] · p
p
for all e ∈ N. Arbitrary Cartier subalgebras will not satisfy such properties, and we know of
no reason to expect (7) to hold in general. However, we will see below in Theorem 4.11 and
Theorem 4.12 that the Cartier subalgebras constructed using ideals and divisors do satisfy
(7) by showing they enjoy a perturbation of the above property.
The proof of Theorem 4.3 made use of positive constants arising from Lemma 3.1. However,
in case R is not local and the p-linear map in (i.) or p−1 -linear map in (ii.) extend to R, one
can make these constants spread uniformly over Spec (R) using Theorem 3.3 instead. In the
case of Theorem 4.7, this immediately yields the following result.
Theorem 4.9. Let R be a locally equidimensional F -finite domain and D a Cartier subalgebra on R. There is a positive constant C ∈ R so that
C
1
p
aD
s(Rp , Dp ) −
e
e ≤ e
1/p
e
ht
(p)
[k(p)
: k(p)] · p
p
for all e ∈ ΓD and all p ∈ Spec (R). Moreover, the function Spec (R) → R given by
p 7→ s(Rp , Dp ) is lower semicontinuous.
One can easily relax the requirements on R above – any F -finite ring will suffice – by using
the fact that the F -signature function is identically zero off of the strongly F -regular locus
of R (and hence s(Rp , Dp ) = 0 for any p ∈ Spec (R) so that Rp is a Cohen-Macaulay normal
domain).
Proof. Following along in the proof of Theorem 4.7, observe that sequences in 6 can be taken
to be global and then localized to each p ∈ Spec (R). By Theorem 3.3, we may then take the
positive constant C = max{2C(M1 )/pe1 γ , . . . , 2C(Ms )/pes γ } independent of p ∈ Spec (R).
For the remainder, first note the argument of Lemma 2.2 readily adapts to show the function
D
Spec (R) → R given by p 7→ ae p is lower semicontinuous. The lower semicontinuity of the
F -signatures thus follows from the argument given in Theorem 3.8.
Remark 4.10. Analogous to the argument above, for an R-valued function on Spec (R) governed locally by sequences of ideals as in Theorem 4.3 (i.), upper semicontinuity passes from
the individual terms in the sequences to the limit as in Theorem 3.4. However, we are
unaware of any such functions that are not directly related Hilbert-Kunz multiplicity itself.
Theorem 4.11. Let R be an F -finite domain, a ⊆ R a non-zero ideal, t ∈ R≥0 , and
p ∈ Spec (R). Then the F -signature
1
t
atp
a
s(Rp , atp ) = s(Rp , C ap ) = lim
e
e
e→∞ [k(p)1/p : k(p)] · pe ht (p)
14
of Rp along atp exists and determines a lower semicontinuous R-valued function on Spec (R).
Moreover, there is a positive constant C ∈ R so that
at
s(Rp , atp )
(8)
−
[k(p)1/pe
aep
C
≤ e
e
ht
(p)
: k(p)] · p
p
at
for all e ∈ N, all p ∈ Spec (R), and all t ∈ R≥0 . In particular,
O(pe(ht (p)−1) ).
aep
e
[k(p)1/p :k(p)]
= s(Rp , atp )pe ht (p) +
Proof. The existence and semicontinuity statements follow immediately from Theorems 4.7
at
and 4.9 above. Let f1 , . . . , fs be a set of generators for a. If t ≥ s, then aep = 0 for all e ∈ N
and any p ∈ Spec (R). Thus, we are free to assume t < s going forward.
For a fixed t ∈ R, one inequality in (8) also follows from Theorem 4.7 above; however, it
remains to show that the positive constant C can be chosen independent of t ∈ R. To that
t
end, fix a choice of 0 6= x ∈ as(p−1) and φ ∈ Hom R (R1/p , R). Then ψ( ) = φ(x1/p · ) ∈ C1a
for any t < s. The desired independence follows from the observation that a short exact
sequence as in Lemma 4.1 (ii.) can be used for (6) in the proof of Theorem 4.7 for any t < s.
e+1
To show the reverse inequality, choose an element 0 6= c ∈ R such that ca⌈t(p −1)⌉ ⊆
e
p
(a⌈t(p −1)⌉ )[p] for all e ∈ N and any t < s; one can check that c = f1 · · · fsp will suffice. We will
at
atp
show that c(Ie p )[p] ⊆ Ie+1
for all t < s and p ∈ Spec (R), after which the result follows by
using a short exact sequence as in Lemma 4.1 (i.) for (5) in the proof of Theorem 4.3 (i.) for
any t < s and p ∈ Spec (R) with C(Mp , pRp ) = C(M) from Theorem 3.3. Supposing that
1/pe+1
at
⌈t(pe −1)⌉
e+1
, Rp ) and y ∈ ap
, we must show φ((yc(Ie p )[p] )1/p ) ⊆ pRp . Write
φ ∈ Hom Rp (Rp
P
1/pe
⌈t(pe −1)⌉
cy = ni=1 gip ri where each ri ∈ Rp and gi ∈ ap
. Let φi ∈ Hom Rp (Rp , Rp ) be the
e
1/pe+1
at
e
e+1
map defined by φi (z 1/p ) = φ(ri
z 1/p ) for all z ∈ Rp . If x ∈ Ie p , then φ((ycxp )1/p ) =
P
Pn
at
e
e
1/pe
1/pe+1
(gi x)1/p ) = ni=1 φi ((gi x)1/p ) ∈ pRp as φi (gi · ) ∈ Ce p for all i = 1, . . . , n.
i=1 φ(ri
Theorem 4.12. Let R be an F -finite normal domain, ∆ an effective Q-divisor on Spec (R),
and p ∈ Spec (R). Then the F -signature
1
p ,∆)
s(Rp , ∆) = s(Rp , C (Rp ,∆) ) = lim
a(R
e
e
e→∞ [k(p)1/p : k(p)] · pe ht (p)
of Rp along ∆ exists and determines a lower semicontinuous R-valued function on Spec (R).
Moreover, if ∆ is a fixed effective integral divisor on Spec (R), there exists a positive constant
C ∈ R so that
(R ,∆)
(9)
s(Rp , ∆) −
ae p
C
≤ e
e
1/p
e
ht
(p)
[k(p)
: k(p)] · p
p
for all e ∈ N, all p ∈ Spec (R), and all effective Q-divisors ∆ with ∆ ≤ ∆. In particular, we
have that
(R ,∆)
ae p
e
[k(p)1/p :k(p)]
= s(Rp , ∆)pe ht (p) + O(pe(d−1) ).
Proof. The existence and semicontinuity statements follow immediately from Theorems 4.7
and 4.9 above. For a fixed ∆, one inequality in (9) also follows from Theorem 4.7 above;
however, it remains to show that the positive constant C can be chosen independent of
15
(R,∆)
(R,∆)
(R,∆)
. Since C1
⊆ C1
, the desired
∆ ≤ ∆. To that end, fix a choice of 0 6= ψ ∈ C1
independence follows from the observation that a short exact sequence as in Lemma 4.1 (ii.)
can be used for (6) in the proof of Theorem 4.7 for any ∆ ≤ ∆.
For the reverse inequality, choose 0 6= c ∈ R so that divR (c) ≥ p∆. We will show that
(R ,∆)
(R ,∆)
c(Ie p )[p] ⊆ Ie+1p
for all ∆ ≤ ∆, after which the result follows by using a short exact
sequence as in Lemma 4.1 (i.) for (5) in the proof of Theorem 4.3 (i.) for any ∆ ≤ ∆
and p ∈ Spec (R) with C(Mp , pRp ) = C(M) from Theorem 3.3. Supposing p ∈ Spec (R)
e+1
(R ,∆)
(R ,∆)
1/pe+1
, Rp )
and φ ∈ Ce+1p , we must show φ((c(Ie p )[p] )1/p ) ⊆ pRp . Let ψ ∈ Hom Rp (Rp
1/pe
1/pe+1
be the map given by ψ( ) = φ(c
· ), and ψe = ψ|R1/pe ∈ Hom Rp (Rp , Rp ) be the
p
1/pe
Rp .
(Rp ,∆)
restriction to
It suffices to show ∆ψe ≥ ∆, so that for any x ∈ Ie
p 1/pe+1
1/pe
φ((cx )
) = ψe (x ) ∈ m.
To that end, consider the following inclusions for each e ∈ N
R1/p
e
R1/p
⊆
we have
e+1
⊆
⊆
(R(⌈(pe − 1)∆)⌉)1/p
e
(R(p⌈(pe − 1)∆⌉))1/p
⊆
e+1
⊆
(R(⌈(pe+1 − 1)∆⌉ + divR (c)))1/p
e+1
which hold because ⌈(pe − 1)∆⌉ ≤ (pe − 1)∆ + ∆ implies
p⌈(pe − 1)∆⌉ ≤ (pe+1 − p)∆ + p∆
≤ (pe+1 − 1)∆ + p∆
≤ ⌈(pe+1 − 1)∆⌉ + divR (c).
1/pe+1
We know that ψ is the restriction to Rp
of a map in
Hom Rp ((Rp (⌈(pe+1 − 1)∆⌉ + divR (c)))1/p
e+1
, Rp ),
and localizing the above inclusions at p we see that ψe = ψ|R1/pe can be extended to a map
e
in Hom Rp ((Rp (⌈(p − 1)∆)⌉)
1/pe
p
, Rp ). It follows immediately that ∆ψe ≥ ∆.
As the above examples demonstrate, Theorem 4.3 can be used to show the existence of
limits in a large number of important settings. Moreover, the following well-known lemma
allows similar techniques to be used more broadly still.
Lemma 4.13. Let (R, m, k) be an F -finite local domain of dimension d and 0 6= c ∈ R.
Suppose {Ie }e∈N and {Je }e∈N are two sequences of ideals in R so that
e
m[p ] ⊆ Ie ⊆ Je ⊆ (Ie : c)
for all e ∈ N. Then there exists a positive constant C ∈ R so that
1
1
C
1
ℓR (R/Ie ) − ed ℓR (R/Je ) = ed ℓR (Je /Ie ) ≤ e
ed
p
p
p
p
for all e ∈ N and so lime→∞ p1ed ℓR (R/Ie ) − p1ed ℓR (R/Je ) = 0. Hence lime→∞
exists if and only if lime→∞
1
ℓ (R/Ie )
ped R
exists, in which case they are equal.
16
1
ℓ (R/Je )
ped R
Proof. Note that as (Ie : c)/Ie is the kernel and R/(Ie , c) the cokernel of multilication by c
on R/Ie , their lengths are equal. Thus
1
ℓ (J /I )
ped R e e
≤
=
≤
1
ℓ ((Ie : x)/Ie )
ped R
1
ℓ (R/(Ie , c))
ped R
e
1
ℓ (R/(m[p ] , c))
ped R
≤
C
pe
by applying Lemma 3.1 with M = R/(c).
Note that, once again, the constant C in the above lemma depends only on the choice of
0 6= c ∈ R and not on any particular sequence of ideals; moreover, in case R is not local, the
constant C can be chosen uniformly over Spec (R) using Theorem 3.3.
5. Positivity
5.1. Positivity of the F-signature. Recall that an F -finite ring R is said to be strongly
e
F -regular if and only if, for all 0 6= x ∈ R there exists some e ∈ N and φ ∈ Hom R (R1/p , R)
e
with φ(x1/p ) = 1. Strongly F -regular rings are always Cohen-Macaulay normal domains,
and remain strongly F -regular after completion. In particular, the completion of a strongly
F -regular local domain remains a domain.
Theorem 5.1. [AL03, Main Result] Suppose (R, m, k) is a local F -finite domain. Then R
is strongly F -regular if and only if s(R) > 0.
We will give two new proofs of Theorem 5.1. The first proof is notable in that it gives a
readily computable lower bound for the F -signature.
First proof of Theorem 5.1. We will assume for simplicity in the exposition that k = k p is
perfect – the proof is easily adapted to arbitrary k. Suppose first that R is not strongly
e
F -regular, so that there exists some 0 6= x ∈ R with φ(x1/p ) ∈ m for all e ∈ N and all
e
e
1/pe
φ ∈ Hom R (R1/p , R). Take a surjection Φ : R1/p → R⊕ frkR (R ) of R-modules. It follows
e
1/pe
e
that Φ((xR)1/p ) ⊆ m⊕ frkR (R ) , so that Φ induces a surjection of R-modules (R/xR)1/p →
e
1/p
e
e
k ⊕ frkR (R ) and hence frkR (R1/p ) ≤ µR ((R/xR)1/p ) ≤ C(R/xR, m)pe(d−1) using (2) and
Lemma 3.1 for some positive constant C(R/xR, m) ∈ R. Dividing through by ped and taking
lime→∞ then gives s(R) = 0. Thus, s(R) > 0 implies that R is strongly F -regular.
Conversely, suppose that R is strongly F -regular. The F -signature remains unchanged
upon completion, so we may assume R is complete and choose a coefficient field k and
system of parameters x1 , . . . , xd so that R is module finite and generically separable over the
e
e
e
regular subring A = k[[x1 , . . . , xd ]]. Take 0 6= c ∈ A so that cR1/p ⊆ R[A1/p ] = R ⊗A A1/p
as in Lemma 2.3 for all e ∈ N. Since R is strongly F -regular, we can find e0 ∈ N and
e
e
φ ∈ Hom R (R1/p 0 , R) with φ(c1/p 0 ) = 1. We will show s(R) ≥ pe10 d > 0.
e
α /pe
α /pe
For any α = (α1 , . . . , αd ) ∈ Zd and e ∈ N, we write xα/p for x1 1 · · · xd d . Let
e
Ie = {α = (α1 , . . . , αd ) ∈ Zd | 0 ≤ αi < pe for i = 1, . . . , d}. The monomials xα/p for
e
α ∈ Ie are a free basis for A1/p over A. As such, for each α ∈ Ie , we can find an A-linear
e
e
e
map πα : A1/p → A with πα (xα/p ) = 1 and πα (xβ/p ) = 0 for all α 6= β ∈ Ie . Applying
e
e
e
R ⊗A gives an R-linear map π̃α : R[A1/p ] → R with π̃α (xα/p ) = 1 and π̃α (xβ/p ) = 0 for
e
e
all α 6= β ∈ Ie . Multiplication by c gives an R-linear map R1/p → R[A1/p ], and composing
e
e
e
with π̃α gives an R-linear map π̃α (c · ) : R1/p → R so that π̃α (c xα/p ) = c π̃α (xα/p ) = c
17
e
e
and π̃α (c xβ/p ) = c π̃α (xβ/p ) = 0 for all α 6= β ∈ Ie (since c ∈ A). Setting ψα ( ) =
e+e
e
e+e
e+e
φ ◦ (π̃α (c · ))1/p 0 : R1/p 0 → R, we have ψα ∈ Hom R (R1/p 0 , R) with ψα (xα/p 0 ) = 1
e+e
ed
e+e
and ψα (xβ/p 0 ) = 0 for all α 6= β ∈ Ie . It follows that Ψ = ⊕α∈Ie ψα : R1/p 0 → R⊕p
e+e
is an R-module surjection and hence frkR (R1/p 0 ) ≥ ped . Dividing through by p(e+e0 )d and
taking lime→∞ yields s(R) ≥ pe10 d > 0 as desired.
For the second proof, we will need the following lemma, which should be compared with
[HH91, Theorem 3.3]. In the next subsection, we will show how to adapt this proof to
arbitrary sequences of ideals as in Theorem 4.3 (ii.).
Lemma 5.2. Let (R, m, k) be a complete local F-finite Cohen-Macaulay domain of dimension
e
d. There exist N ∈ N with the following property: for any e ∈ N and all x ∈ R \ m[p ] , there
e
e
exists a map φ ∈ Hom R (R1/p , R) with φ(x1/p ) 6∈ mN .
Proof. Choosing a coefficient field k and system of parameters x1 , . . . , xd , we have that R is
a finitely generated free module over A = k[[x1 , . . . , xd ]] (see [BH93, Proposition 2.2.11]); let
mA denote the maximal ideal of A. Fix a non-zero map τ ∈ Hom A (R, A). Since Hom A (R, A)
is a rank one torsion free R-module, we can find find 0 6= y ∈ Ann R (Hom A (R, A)/(Rτ ));
replacing by a nonzero multiple, we may further assume 0 6= y ∈ A. If N ′ , N ∈ N are
′
′
N
sufficiently large so that y 6∈ mN
⊆ mN
A and m
A R, we will show N satisfies the desired
property.
e
e
e
Fix e ∈ N and x ∈ R \ m[p ] , so also x 6∈ mpA R. Since R1/p is a free A-module and
e
e
e
x1/p 6∈ mA R1/p , x1/p can be taken as part of a free basis and there exists an A-linear map
e
e
χ : R1/p → A with χ(x1/p ) = 1. If Tr : Hom A (R, A) → A is the A-linear map given by
evaluation at 1, there is a commutative diagram
(10)
e
χ̃
/ Hom A (R, A)
/ Hom A (R1/pe , A)
❨❨❨❨❨❨
❨❨❨❨❨❨
❨❨❨❨❨❨
Tr
❨❨❨❨❨❨
χ
❨❨❨❨❨❨
❨❨❨❨,
R1/p ❨❨❨❨❨❨❨
A
e
·y
Rτ
/
≃
R
/
τ
Tr
·y
/
A
=
/
A
e
where χ̃ is the R1/p -linear map sending 1 7→ χ and Hom A (R1/p , A) → Hom A (R, A) is the
e
e
R-linear map given by restriction to R ⊆ R1/p . Let φ ∈ Hom R (R1/p , R) be the composition
e
e
along the top row of this diagram, so that yχ = τ ◦ φ. We have τ (φ(x1/p )) = yχ(x1/p ) =
′
e
′
′
1/p
N
y 6∈ mN
) 6∈ mN
⊆ mN
A , and thus φ(x
A R as τ is A-linear. In particular, since m
A R, we
e
1/p
N
have φ(x ) 6∈ m as desired.
Second proof of Theorem 5.1. Suppose that R is strongly F -regular, and thus Cohen-Macaulay.
As before, without loss of generality, we may assume R is complete. If IeF−sig = (r ∈ R |
e
e
e′
φ(r 1/p ) ∈ m for all φ ∈ Hom R (R1/p , R)) and ψ ∈ Hom R (R1/p , R), it is easy to check that
F−sig 1/pe′
F−sig
⊆ IeF−sig
ψ((Ie+e
) ⊆ IeF−sig for all e, e′ ∈ N. Since R is F -split, we may conclude Ie+1
′ )
T
F−sig
for all e ∈ N; moreover, by the definition of strong F -regularity, we have e∈N Ie
= 0.
Thus, for N as in Lemma 5.2, Chevalley’s Lemma [Che43, Lemma 7] gives e0 ∈ N with
e
F−sig
⊆ m[p ] for all e ∈ N. Indeed for
IeF−sig
⊆ mN . It follows from Lemma 5.2 that Ie+e
0
0
e
e
e
each x ∈ R \ m[p ] there is some φ ∈ Hom R (R1/p , R) with φ(x1/p ) 6∈ mN and hence
18
F−sig 1/p
F−sig
φ(x1/p ) 6∈ IeF−sig
; as φ((Ie+e
) ) ⊆ IeF−sig
we must have x 6∈ Ie+e
. Thus, we compute
0
0
0
0
1
1
1
1
e
F−sig
s(R) = lim (e+e0 )d ℓR (R/Ie+e
) ≥ lim (e+e0 )d ℓR (R/m[p ] ) = e0 d eHK (R) ≥ e0 d > 0
0
e→∞ p
e→∞ p
p
p
as desired.
e
e
5.2. Positivity via a Cartier linear map. In this section, we generalize the method of the
second proof of Theorem 5.1 to examine the positivity of the limits appearing in Theorem 4.3
(ii.). The essential technique is to exploit the following generalization of Lemma 5.2, which
should again be compared with [HH91, Theorem 3.3].
Proposition 5.3. Let (R, m, k) be a complete local F -finite domain of dimension d and
suppose that 0 6= ψ ∈ Hom R (R1/p , R). There exist N ∈ N and 0 6= c ∈ R so that
e
e
ψ e ((xR)1/p ) 6⊆ mN for any e ∈ N and all x ∈ R \ (m[p ] : c).
Proof. Choose a coefficient field k and system of parameters x1 , . . . , xd so that R is module
finite and generically separable over the regular subring A = k[[x1 , . . . , xd ]]. Take 0 6= c ∈ A
e
e
e
so that cR1/p ⊆ R[A1/p ] = R ⊗A A1/p as in Lemma 2.3 for all e ∈ N.
Fixing a non-zero map τ ∈ Hom A (R, A), we can find 0 6= y ∈ Ann R (Hom A (R, A)/(Rτ ))
as Hom A (R, A) is a rank one torsion free R-module. Let 0 6= z ∈ R be as in Lemma 4.2;
replacing by nonzero multiples, we may further assume both 0 6= y ∈ A and 0 6= z ∈ A. If
′
N
N′
N ′ , N ∈ N are sufficiently large that yz 6∈ mN
A and m ⊆ mA R, we will show the pair of c
and N satisfy the desired property.
e
e
e
[pe ]
Suppose e ∈ N and x ∈ R \ (m[p ] : c). Since xc 6∈ mA R ⊆ m[p ] , we have (xc)1/p 6∈
e
e
e
e
e
mA R1/p and so also (xc)1/p 6∈ mA R[A1/p ]. As R[A1/p ] = R ⊗A A1/p is a free A-module,
e
e
e
there is a χ′ ∈ Hom A (R[A1/p ], A) with χ′ ((xc)1/p ) = 1. Let χ ∈ Hom A (R1/p , A) be the
composition
R1/p
·c
e
/
χ′
e
R[A1/p ]
/
A
e
which satisfies χ(x1/p ) = 1. If Tr : Hom A (R, A) → A is the A-linear map given by evaluation
e
at 1, the same diagram (10) from Lemma 5.2 commutes where χ̃ is the R1/p -linear map
e
sending 1 7→ χ and Hom A (R1/p , A) → Hom A (R, A) is the R-linear map given by restriction
e
e
to R ⊆ R1/p . Let φ ∈ Hom R (R1/p , R) be the composition along the top row of this diagram,
e
so that yχ = τ ◦ φ. By Lemma 4.2, we can find r ∈ R so that zφ( ) = ψ e (r 1/p · ). We
compute
e
e
e
e
τ (ψ e ((rx)1/p )) = τ (zφ(x1/p )) = zτ (φ(x1/p )) = yzχ(x1/p ) = yz 6∈ mN
A
e
′
′
′
N
N
and conclude ψ e ((rx)1/p ) 6∈ mN
A R as τ is A-linear. In particular, since m ⊆ mA R, we have
e
1/pe
N
ψ ((xR) ) 6⊆ m as desired.
In the proof of the main result of this section, we will need to use a well-known result
on the stabilization of the images of an iterated p−1 -linear map. Rooted in the work of
Hartshorne and Speiser [HS77, Proposition 1.11] and later generalized by Lyubeznik [Lyu97],
the following version is due to Gabber [Gab04].
Theorem 5.4. [Gab04, Lemma 13.1] [BB11, Proposition 2.14] Let M be a finitely generated
module over an F -finite ring R and 0 6= ϕ ∈ Hom R (M 1/p , M). Then the descending chain
of R-modules
2
e
M ⊇ φ(M 1/p ) ⊇ φ2 (M 1/p ) ⊇ · · · ⊇ φe (M 1/p ) ⊇ · · ·
19
stabilizes for e ≫ 0.
Theorem 5.5. Let (R, m, k) be a complete local F-finite domain of dimension d and {Ie }e∈N
e
a sequence of ideals so that m[p ] ⊆ Ie for all e ∈ N. Suppose there exists a non-zero ψ ∈
Hom R (R1/p , R) so that ψ((Ie+1 )1/p ) ⊆ Ie for all e ∈ N. Then the limit lime→∞ p1ed ℓR (R/Ie )
T
is positive if and only if e∈N Ie = 0.
T
e
Proof. By Lemma 3.1, if 0 6= cT∈ e∈N Ie , then ℓR (R/Ie ) ≤ ℓR (R/(m[p ] , c)) ≤ C(R/(c), m)pe(d−1)
for all e ∈ N. It follows that e∈N Ie 6= 0 implies lime→∞ p1ed ℓR (R/Ie ) = 0.
T
e
For the converse, suppose e∈N Ie = 0. If 0 6= σ = ψ e (R1/p ) for e ≫ 0 is the stable image
T
T
T
of ψ as in Theorem 5.4, we also have e∈N (Ie : σ) = 0 as σ e∈N (Ie : σ) ⊆ e∈N Ie = 0
and R is a domain. In addition, it follows from ψ(σ 1/p ) = σ that (Ie+1 : σ) ⊆ (Ie : σ) for all
e ∈ N as
1/p
(Ie+1 : σ)σ = (Ie+1 : σ)ψ(σ 1/p ) = ψ(((Ie+1 : σ)[p] σ)1/p ) ⊆ ψ(((Ie+1 : σ)σ)1/p ) ⊆ ψ(Ie+1 ) ⊆ Ie .
Take N ∈ N and 0 6= c ∈ R as in Proposition 5.3. By Chevalley’s Lemma [Che43, Lemma 7],
there is some e0 ∈ N with Ie0 ⊆ (Ie0 : σ) ⊆ mN . It follows from Proposition 5.3 that
e
e
e
Ie+e0 ⊆ (m[p ] : c) for all e ∈ N. Indeed for each x ∈ R \ (mp : c) we have ψ e ((xR)1/p ) 6⊆ mN
e
e
and hence ψ e ((xR)1/p ) 6⊆ Ie0 ; as ψ e ((Ie+e0 )1/p ) ⊆ Ie0 we must have x 6∈ Ie+e0 . Using
Lemma 4.13, we compute
lime→∞
1
ℓ (R/Ie+e0 )
p(e+e0 )d R
≥
=
=
1
lime→∞
pe0 d
1
lime→∞
pe0 d
1
e (R)
pe0 d HK
which concludes the proof.
e
1
ℓ (R/(m[p ]
ped R
e
1
ℓ (R/m[p ] )
ped R
≥ pe10 d > 0,
: c))
Theorem 5.5 is quite powerful and allows one to show positivity of the limits appearing
in Theorem 4.3 (ii.) in a number of different situations. One should view this result not
only as a generalization of the work done by Aberbach and Leuschke in [AL03], but also the
work of Hochster and Huneke [HH90]. If R is a domain and I ⊆ R, recall that x ∈ R is in
e
e
the tight closure of I ∗ of I if there exists an element 0 6= c ∈ R such that cxp ∈ I [p ] for all
e ∈ N. The tight closure I ∗ is an ideal containing I, and I ∗∗ = I. A ring is said to be weakly
F -regular if all ideals I of R are tightly closed, i.e. satisfy I ∗ = I. See [Hun96] or [Hoc] for
further details.
Corollary 5.6. [HH90, Theorem 8.17] (Length Criterion for Tight Closure). Let (R, m, k)
be a complete local F-finite domain of dimension d. Suppose that I ⊆ J is an inclusion of
m-primary ideals in R. Then I ∗ = J ∗ if and only if eHK (I) = eHK (J).
Proof. Applying the criterion to each term of a composition series of J/I, we may assume
e
e
J = (I, x) for some x ∈ R with (I : x) = m. Consider the sequence of ideals (I [p ] : xp ) for
e
e
e+1
e+1
each e ∈ N. We have that (I [p ] : xp )[p] ⊆ (I [p ] : xp ) for all e ≥ 0, so that in particular
e
e
e
m[p ] ⊆ (I [p ] : xp ) for each e ∈ N. Moreover, for any nonzero ψ ∈ Hom R (R1/p , R) it is easy
e+1
e+1
e
e
e
e
e
e
to check ψ((I [p ] : xp )1/p ) ⊆ (I [p ] : xp ). Since (I, x)[p ] /I [p ] ≃ R/(I [p ] : xp ), we see by
Theorem 5.5 that
1
e
e
(11)
eHK (I) − eHK ((I, x)) = lim ed ℓR (R/(I [p ] : xp ))
e→∞ p
20
is zero if and only if
T
e∈N (I
[pe ]
e
: xp ) 6= 0, which is equivalent to x ∈ I ∗ by definition.
Recall that there are also a number of well-known generalizations of tight closure and
strong F -regularity [HY03, HW02, Tak04]. If R is an F -finite local domain, a is a nonzero
t
ideal of R, and t ∈ R≥0 , one can speak of the at -tight closure I ∗a of an ideal I ⊆ R. By
t
e
e
e
definition, if x ∈ R then x ∈ I ∗a if and only if there exists 0 6= c ∈ R with ca⌈t(p −1)⌉ xp ∈ I [p ]
for all e ∈ N. The pair (R, at ) is strongly F -regular provided for any 0 6= x ∈ R there exists
t
e
e ∈ N and φ ∈ Cea with φ(x1/p ) = 1. Similarly, if R is an F -finite normal local domain
and ∆ is an effective Q-divisor on Spec (R), we have the ∆-tight closure I ∗∆ of an ideal
I ⊆ R . By definition, if x ∈ R then x ∈ I ∗∆ if and only if there exists 0 6= c ∈ R with
e
e
cxp ∈ I [p ] (R(⌈(pe − 1)∆⌉)) for all e ∈ N. The pair (R, ∆) is strongly F -regular provided for
e
(R,∆)
any 0 6= x ∈ R there exists e ∈ N and φ ∈ Ce
with φ(x1/p ) = 1. As of yet, there does not
exist a well-formed theory of Hilbert-Kunz multiplicity for an ideal or divisor; nonetheless,
one can use Theorem 5.5 to give analogues of Theorem 5.6 for these operations.
Corollary 5.7 (Length Criterion for at -tight closure). Let (R, m, k) be a complete local
F -finite domain of dimension d, a a nonzero ideal of R, and t ∈ R≥0 .
e
e
e
(i.) For any m-primary ideal I and x ∈ R, lime→∞ p1ed ℓR (R/(I [p ] : a⌈t(p −1)⌉ xp )) exists;
t
moreover, this limit equals zero if and only if x ∈ I ∗a .
t
(ii.) The F -signature s(R, at ) = lime→∞ p1ed ℓR (R/Iea ) of R along at is positive if and only
if (R, at ) is strongly F -regular.
t
Proof. For (i.), take 0 6= ψ ∈ C1a = (a⌈t(p−1)⌉ )1/p ·Hom R (R1/p , R), and consider the sequence of
e
e
e
e+1
e+1
e+1
ideals (I [p ] : a⌈t(p −1)⌉ xp ) for e ∈ N. It is easy to check that ψ((I [p ] : a⌈t(p −1)⌉ xp )1/p ) ⊆
e
e
e
e
(I [p ] : a⌈t(p −1)⌉ xp ) for all e ∈ N. If e0 ∈ N is sufficiently large that m[p 0 ] ⊆ I, then we have
e
e
e
e+e
m[p 0 ] ⊆ (I [p ] : a⌈t(p −1)⌉ xp ) for all e ∈ N. After shifting the index, Theorem 4.3 (ii.) and
Theorem
giving the existence of the limit and showing that it equals zero if and
T 5.5 apply,
e
e
e
t
only if e∈N (I [p ] : a⌈t(p −1)⌉ xp ) 6= 0, which is equivalent to x ∈ I ∗a by definition. To see
t
t
a
(ii.), simply note once more that ψ((Ie+1
)1/p ) ⊆ Iea and apply Theorem 5.5.
Corollary 5.8 (Length Criterion for ∆-tight closure). Let (R, m, k) be a complete local
F -finite normal domain of dimension d, and ∆ an effective Q-divisor on Spec (R).
e
e
(i.) For any m-primary ideal I ⊆ R and x ∈ R, lime→∞ p1ed ℓR (R/(I [p ] R(⌈(pe −1)∆⌉) : xp ))
exists; moreover, this limit equals zero if and only if x ∈ I ∗∆ .
(R,∆)
(ii.) The F -signature s(R, ∆) = lime→∞ p1ed ℓR (R/Ie
) of R along ∆ is positive if and
only if (R, ∆) is strongly F -regular.
Proof. Let 0 6= ψ ∈ C1∆ = im Hom R ((R(⌈(p − 1)∆⌉))1/p , R) → Hom R (R1/p , R) . In other
words, identifying ψ with its unique extension to a p−1 -linear map on K = Frac(R), we have
ψ((R(⌈(p − 1)∆⌉))1/p ) ⊆ R. Twisting by ⌈(pe − 1)∆⌉ and using that p⌈(pe − 1)∆⌉ + ⌈(p −
1)∆⌉ ≥ ⌈(pe+1 − 1)∆⌉, observe that ψ((R(⌈(pe+1 − 1)∆⌉))1/p ) ⊆ R(⌈(pe − 1)∆⌉) for all e ∈ N.
e
e
For (i.), consider the sequence of ideals (I [p ] R(⌈(pe − 1)∆⌉) : xp ) for e ∈ N. It is easy to
e+1
e+1
e
e
check ψ((I [p ] R(⌈(pe+1 − 1)∆⌉) : xp )1/p ) ⊆ (I [p ] R(⌈(pe − 1)∆⌉) : xp ) for all e ∈ N. If
e
e+e
e
e
e0 ∈ N is sufficiently large that m[p 0 ] ⊆ I, then we have m[p 0 ] ⊆ (I [p ] R(⌈(pe − 1)∆⌉) : xp )
apply, giving the
for all e ∈ N. After shifting the index, Theorem 4.3 (ii.) and Theorem
T 5.5
[pe ]
existence of the limit and showing that it equals zero if and only if e∈N (I R(⌈(pe −1)∆⌉) :
21
e
xp ) 6= 0, which is equivalent to x ∈ I ∗∆ by definition. For (ii.), simply note once more that
(R,∆)
(R,∆)
ψ((Ie+1 )1/p ) ⊆ Ie
and apply Theorem 5.5.
Remark 5.9. Note that the analogue of Theorem 5.5 for sequences of ideals satisfying Theorem 4.3 (i.) instead of (ii.) is false. Here is an easy counterexample. Let (R, m, k) be a
⌊e/2⌋ ]
complete regular local ring of dimension d, and consider the sequence of ideals Ie = m[p
T
e
[p]
for e ∈ N. Then m[p ] ⊆ Ie and Ie ⊆ Ie+1 for all e ∈ N and certainly e∈N Ie = 0. However,
1
= 0.
we have lime→∞ p1ed ℓR (R/Ie ) = lime→∞ p(e−⌊e/2⌋)d
To see yet another application of Theorem 5.5, suppose that (R, m, k) is an F -split F -finite
e
e
local ring. If IeF−sig = (r T
∈ R | φ(r 1/p ) ∈ m for all φ ∈ Hom R (R1/p , R)), it is straightforward to check that P = e∈N IeF−sig is a prime ideal (see [Tuc12, Lemma 4.7]), coined the
F -splitting prime by Aberbach and Enescu in [AE05]. Furthermore, they suspected that
e
the dimension of the F -splitting prime governed the growth rate of frkR (R1/p ) when R is
not strongly F -regular; this observation was verified in joint work of the second author with
Blickle and Schwede.
Theorem 5.10. [BST12] Suppose (R, m, k) is an F -split F -finite local ring and P is the
e
F -splitting prime. If n = dim(R/P), the limit rF (R) = lime→∞ [k1/pe1:k]·pen frkR (R1/p ) exists
e
and is positive, called the F -splitting ratio. Moreover, we have that [k1/p1e :k] frkR (R1/p ) =
rF (R)pen + O(pe(n−1) ).
e
Proof. Without loss of generality, we may assume R is complete. If IeF−sig = hr ∈ R | φ(r 1/p ) ∈
e
e
F−sig
m for all φ ∈ Hom R (R1/p , R)i, we have m[p ] ⊆ IeF−sig and (IeF−sig )[p] ⊆ Ie+1
for all e ∈ N.
F−sig 1/p
1/p
F−sig
Fixing a surjective map ψ ∈ Hom R (R , R), we have ψ((Ie+1 ) ) ⊆ Ie
. In particular,
1/p
this implies ψ(P 1/p ) ⊆ P and so ψ induces a map ψ ∈ Hom R (R , R) where R = R/P.
Note that ψ is still surjective and hence nonzero. Thus, passing to the sequence of ideals
[p]
e
Ie = ITeF−sig R, we have m[p ] R ⊆ Ie , Ie ⊆ Ie+1 , and ψ((Ie+1 )1/p ) ⊆ Ie for all e ∈ N. Moreover, e∈N Ie = 0 in R. The result now follows immediately from Corollary 4.5 together with
e
Theorem 5.5, using that
frkR (R1/p )
[k 1/pe :k]
= ℓR (R/IeF−sig ) = ℓR (R/Ie ) for all e ∈ N.
Remark 5.11. It is straightforward to generalize the notions of F -splitting prime and F splitting ratio to arbitrary Cartier subalgebras; see [BST12] for further details. In all cases,
the method of the proof of Theorem 5.10 applies and greatly simplifies the proofs. In
particular, the methods of the proofs of Theorem 4.7 and Theorem 5.5 immediately give an
alternative proof of the following result.
Corollary 5.12. [BST12, Theorem 3.18] Let (R, m, k) be an F -finite local domain of dimension d and D a Cartier subalgebra on R. Then the F -signature s(R, D) = lime→∞ [k1/pe1:k]·ped aD
e
e∈ΓD
of R along D is positive if and only if (R, D) is strongly F -regular.
Remark 5.13. Another straightforward generalization comes modifying the condition in Theorem 4.3 (ii) and Theorem 5.5 to consider sequences of m-primary ideals governed by a nonzero p−e0 -linear map for some e0 ∈ N. Beginning with short exact sequences in Lemma 4.1
e
replacing R1/p with R1/p 0 and tracing through the arguments of previous sections, one readily arrives at the following statement.
22
Corollary 5.14. Let (R, m, k) be a complete local F-finite domain of dimension d and e0 ∈ N.
ne
Suppose {In }n∈N a sequence of ideals so that m[p 0 ] ⊆ In for all n ∈ N. Suppose there exists
e
e
a non-zero ψ ∈ Hom R (R1/p 0 , R) so that ψ((In+1 )1/p 0 ) ⊆TIn for all n ∈ N. Then the limit
limn→∞ pne10 d ℓR (R/In ) exists and is positive if and only if n∈N In = 0.
6. F-signature and Minimal Relative Hilbert-Kunz Multiplicity
Our next aim is to realize the F -signature as the infimum of relitive differences in the
Hilbert-Kunz multiplicities of the cofinite ideals in a local ring (Corollary 6.5). After first
bounding such differences from below by the F -signature (Lemma 6.1), we will make use of
approximately Gorenstein sequences to find differences arbitrarily close to the F -signature.
The crucial step and main technical result is Theorem 6.3, which uses the uniformity of the
constants tracked above swap limits between iterations of Frobenius and progression in an
approximately Gorenstein sequence.
The remainder of the section is reserved for constructions of explicit sequences of relative
Hilbert-Kunz differences that approach the F -signature (see Corollary 6.6); we also analyze
when the infimum is known to be achieved (Corollary 6.8). Generalizations to divisor and
ideal pairs are also given (Corollaries 6.9, 6.10).
e
Lemma 6.1. Suppose that (R, m, k) is an F -finite local ring and IeF−sig = (r ∈ R | φ(r 1/p ) ∈
e
m for all φ ∈ Hom R (R1/p , R)) for e ∈ N. Then
X
X
e
e
e
e
(12)
IeF−sig ⊇
(I [p ] : xp )
(I [p ] : J [p ] ) ⊇
I⊆J⊆R
0<ℓR (J/I)<∞
I⊆R, ℓR (R/I)<∞
x∈R, (I:x)=m
and we have
(13)
s(R) ≤
eHK (I) − eHK (J)
≤
inf
eHK (I) − eHK ((I, x)).
I⊆R, ℓR (R/I)<∞
I⊆J⊆R, ℓR (R/I)<∞
ℓR (J/I)
inf
x∈R, (I:x)=m
I6=J, ℓR (R/J)<∞
e
1/p
Proof. If I ( J is a proper inclusion of ideals with ℓR (J/I) < ∞
, R),
Pand φ ∈ Hom[pRe(R
]
[pe ]
[pe ]
[pe ] 1/pe
(I
:
J
)⊆
we have that φ((I
: J ) ) ⊆ (I : J) ⊆ m. It follows that
I⊆J⊆R
0<ℓR (J/I)<∞
IeF−sig . Since the sum on the right is over a smaller set of proper inclusions, (12) follows
immediately.
e
e
For (13), if I ⊆ R is m-primary and x ∈ R with (I : x) = m, then (I [p ] : xp ) ⊆ IeF−sig for
all e ∈ N implies eHK (I) − eHK ((I, x)) ≥ s(R) using (11). Moreover, if I ( J is a proper
inclusion of m-primary ideals, summing up this inequality for each factor in a composition
series of J/I shows ℓR (J/I)s(R) ≤ (eHK (I) − eHK (J)). The inequalities in (13) now follow
immediately, noting again that the infimum on the right is over a smaller set of proper
inclusions.
Recall that a local ring (R, m, k) is said to be approximately Gorenstein if there exists a
descending chain of irreducible ideals {Jt }t∈N cofinal with powers of the maximal ideal. In
particular, each R/Jt is a zero-dimensional Gorenstein local ring and has a one dimensional
socle. By [Hoc77, Theorem 1.6], a reduced excellent local ring is always approximately
Gorenstein. It is easy to check that equality holds throughout (12) for such rings.
23
Lemma 6.2. Suppose that (R, m, k) is an approximately Gorenstein F -finite local ring, and
{Jt }t∈N is a descending chain of irreducible ideals cofinal with the powers of m. If δt ∈ R
[pe ]
[pe ]
[pe ]
[pe ]
generates the socle of R/Jt , then for all e ∈ N we have (Jt : δt ) ⊆ (Jt+1 : δt+1 ) and
IeF−sig =
=
=
=
e
e
hr ∈ R | φ(r 1/p ) ∈ m for all φ ∈ Hom R (R1/p , R)i
P
[pe ]
[pe ]
(Jt : δt )
t∈N
S
[pe ]
[pe ]
(Jt : δt )
t∈N
[pe ]
[pe ]
(Jte : δte ) for all te ≫ 0 sufficiently large.
Moreover, R is weakly F -regular if and only if Jt∗ = Jt is tightly closed for all t ∈ N.
Proof. Since each R/Jt is an Artinian Gorenstein local ring, we have that Ann E (Jt ) ≃ R/Jt
where E = ER (k) is the injective hull of the residue field of R. Thus, we may view E =
ER (k) = lim R/Jt as the direct limit of inclusions R/Jt → R/Jt+1 mapping (the class of)
−→
e
[pe ]
[pe ]
[pe ]
[pe ]
δt 7→ δt+1 . In particular, after applying ⊗R R1/p , one sees that (Jt : δt ) ⊆ (Jt+1 : δt+1 )
P
S
[pe ]
[pe ]
[pe ]
[pe ]
[pe ]
[pe ]
for all t ∈ N. It follows immediately that t∈N (Jt : δt ) = t∈N (Jt : δt ) = (Jte : δte )
for all te ≫ 0.
An inclusion R → M to a finitely generated R-module M determined by 1 7→ m splits
if and only if E → E ⊗R M remains injective, which is equivalent to δt m 6∈ Jt M for all
t ∈ N. See [Hoc, page 155] for further details. In particular, if x ∈ R, applying this splitting
e
e
criterion to the map R → R1/p with 1 7→ x1/p gives that x ∈ R \ IeF−sig if and only if
e
e
[p ]
[p ]
[pe ]
[pe ]
x ∈ R \ (Jt : δt ) for all t ∈ N, and so we have that IeF−sig = (Jte : δte ) forTte ≫ 0.
Lastly, suppose there is an ideal I ⊆ R that is not tightly closed. Then I = n∈N (I + mn )
is an intersection of m-primary ideals. The arbitrary intersection of tightly closed ideals
is tightly closed, [HH90, Proposition 4.1(b)]. Hence we may replace I with an m-primary
ideal which is not tightly closed and choose x ∈ I ∗ with m = (I : x). Since R/I injects
into a direct sum of copies of E, we can find an R-module homomorphism R/I → E so
that x + I has nonzero image in E and hence must generate the socle k ⊆ E. Using that
E = lim R/Jt , we may assume R/I → R/Jt and x + I 7→ δt + Jt for some t ∈ N. For each
−→
e
e
e
[pe ]
e ∈ N, applying ⊗R R1/p and viewing as a R1/p -module gives R/I [p ] → R/Jt where
e
e
e
e
e
[pe ]
[pe ]
[pe ]
xp + I [p ] 7→ δtp + Jt . In particular, it follows that (I [p ] : xp ) ⊆ (Jt : δt ). Thus
T
T
e
e
[pe ]
[pe ]
: xp ) ⊆ e∈N (Jt : δtp ) is not contained in any minimal prime of R; it follows
e∈N (I
that δt ∈ Jt∗ and hence Jt is not tightly closed (cf. [HH90, Proposition 8.23(f)]).
In the next result, we show how to make use of the uniformity of constants from Theorem 4.3 together with approximately Gorenstein sequences in order to compare F -signature
and Hilbert-Kunz multiplicity. As an application, we answer a question posed by Watanabe
and Yoshida [WY04, Question 1.10].
Theorem 6.3. Let (R, m, k) be a complete local F-finite domain of dimension d, and fix
e
0 6= c ∈ R. Suppose we are given sequences of ideals {It,e }t,e∈N satisfying m[p ] ⊆ It,e ,
[p]
c(It,e ) ⊆ It,e+1 , and It,e ⊆ It+1,e for all t, e ∈ N. Then
1
1
1
lim lim ed ℓR (R/It,e ) = lim ed ℓR (R/Ie ) = lim lim ed ℓR (R/It,e )
e→∞ t→∞ p
e→∞ p
t→∞ e→∞ p
P
where Ie = t∈N It,e .
24
[p]
e
Proof. For each fixed t ∈ N, as m[p ] ⊆ It,e and c(It,e ) ⊆ It,e+1 for e ∈ N, Theorem 4.3 (i) guarantees the existence of ηt = lime→∞ p1ed ℓR (R/It,e ) and provides a uniform positive constant
C ∈ R so that
C
1
(14)
ηt ≤ ed ℓR (R/It,e ) + e
p
p
e
[p]
for all t, e ∈ N. The sequence {Ie }e∈N inherits the properties m[p ] ⊆ Ie and c(Ie ) ⊆ Ie+1 for
e ∈ N, hence lime→∞ p1ed ℓR (R/Ie ) exists as well. Since It,e ⊆ Ie and so ℓR (R/Ie ) ≤ ℓR (R/It,e )
for all t, e ∈ N, applying lime→∞ gives
1
(15)
lim ed ℓR (R/Ie ) ≤ ηt
e→∞ p
for all t ∈ N.
Since It,e ⊆ It+1,e is increasing in t for fixed e, it follows that ηt ≥ ηt+1 ≥ 0 for all t ∈ N and
hence limt→∞ ηt exists. We also have Ie = Ite ,e for te ≫ 0, so that lime→∞ limt→∞ p1ed ℓR (R/It,e ) =
lime→∞ p1ed ℓR (R/Ie ). Applying limt→∞ to (14) and (15) gives
1
1
C
ℓ
(R/I
)
≤
lim
η
≤
ℓ
(R/I
)
+
R
e
t
R
e
e→∞ ped
t→∞
ped
pe
lim
for all e ∈ N. Further taking lime→∞ gives limt→∞ ηt = lime→∞ p1ed ℓR (R/Ie ) and completes
the proof.
Theorem 6.4. Let (R, m, k) be an approximately Gorenstein F -finite local ring of dimension d. Suppose {Jt }t∈N is a descending chain of irreducible ideals cofinal with the powers of
m, and δt ∈ R generates the socle of R/Jt . Then s(R) = limt→∞ eHK (Jt ) − eHK ((Jt , δt )).
Proof. Both invariants are unchanged after completion, so we may assume R is complete.
[pe ]
Suppose first that R is not weakly F -regular, so that δt ∈ Jt∗ for some t ∈ N and (Jt :
[pe ]
[pe ]
[pe ]
[pe ]
[pe ]
δt ) ⊆ (Jt+1 : δt+1 ) ⊆ · · · ⊆ (Jt+t′ : δt+t′ ) ⊆ · · · for all e, t′ ∈ N by Lemma 6.2. If
T
T
e
[pe ]
[pe ]
pe
c ∈ e∈N (Jt
: δtp ) ⊆ e∈N (Jt+t′ : δt+t
′ ) is not in any minimal prime, then we have
e
that 0 ≤ eHK (Jt+t′ ) − eHK ((Jt+t′ , δt+t′ )) ≤ lime→∞ p1ed ℓR (R/(c, m[p ] )) = 0 using (11) from
Corollary 5.6 and applying Lemma 3.1 with M = R/(c). Using Lemma 6.1, we have that
s(R) = limt′ →∞ eHK (Jt+t′ ) − eHK ((Jt+t′ , δt+t′ )) = 0 as desired.
Thus, we assume for the remainder that R is weakly F -regular and hence a domain.
[pe ]
[pe ]
Consider the sequences of ideals It,e = (Jt : δt ) for t, e ∈ N. We check
e
[p]
It,e
[pe ]
e
[pe ]
m[p ] = (Jt : δt )[p ] ⊆ (Jt : δt ) = It,e
e
[pe ]
[pe ]
[pe+1 ]
[pe+1 ]
= (Jt : δt )[p ] ⊆ (Jt : δt )[p] ⊆ (Jt
: δt
) = It,e+1
[pe ]
[pe ]
[pe ]
[pe ]
It,e = (Jt : δt ) ⊆ (Jt+1 : δt+1 ) = It+1,e
so that Theorem 6.3 applies with c = 1. Using Lemma 6.2 and (11) from Corollary 5.6, we
conclude
1
1
s(R) = lim ed ℓR (R/IeF−sig ) =
lim lim ed ℓR (R/It,e )
e→∞ p
e→∞ t→∞ p
1
= lim lim ed ℓR (R/It,e ) = lim eHK (Jt ) − eHK ((Jt , δt ))
t→∞ e→∞ p
t→∞
which completes the proof.
25
Theorem 6.4 provides a positive answer to [WY04, Question 1.10].
Corollary 6.5. If (R, m, k) is an F-finite local ring, then
s(R) =
eHK (I) − eHK (J)
=
inf
eHK (I) − eHK ((I, x)).
I⊆J⊆R, ℓR (R/I)<∞
I⊆R, ℓR (R/I)<∞
ℓR (J/I)
inf
I6=J, ℓR (R/J)<∞
x∈R, (I:x)=m
Proof. If R is not reduced, we can find some 0 6= x ∈ R with xp = 0. For n ≫ 0, we have
e
x 6∈ mn and can find an ideal I with mn ⊆ I ⊆ (mn , x) where (I : x) = m. Since xp = 0
e
e
for all e ∈ N we have I [p ] = (I, x)[p ] , and thus eHK (I) = eHK ((I, x)). It follows from
Lemma 6.1 that s(R) = 0 and equality holds throughout (13). Thus, we may assume R is
reduced. By [Hoc77, Theorem 1.7], R is approximately Gorenstein and Theorem 6.4 implies
equality holds throughout (13) as desired.
Note that, when (R, m, k) is a complete Cohen-Macaulay local F -finite domain of dimension d, one can make Theorem 6.4 more explicit still. Recall that a canonical ideal J ⊆ R is
an ideal such that J is isomorphic to a canonical module ωR , which exists as R is assumed
complete. A canonical ideal J is necessarily unmixed with height 1, and moreover R/J will
be Gorenstein of dimension d − 1, see [BH93, Proposition 3.3.18]. When R is also normal,
fixing a canonical ideal is equivalent to fixing a choice of effective anticanonical divisor.
Corollary 6.6. Suppose that (R, m, k) is a complete Cohen-Macaulay local F -finite domain
of dimension d, and J is a canonical ideal of R. Let x1 ∈ J and x2 , . . . , xd ∈ R be chosen so
that x1 , . . . , xd give a system of parameters for R, and suppose δ ∈ R generates the socle of
R/(J, x2 , . . . , xd ). Then
s(R) = lim eHK ((J, xt2 , . . . , xtd )) − eHK ((J, xt2 , . . . , xtd , (xt−1
· · · xt−1
2
d δ))).
t→∞
t
t
Proof. Note that xt−1
1 J is yet another canonical ideal for any t ∈ N, and x2 , . . . , xd give
t−1
t−1
a system of parameters for R/(x1 J). Thus, the sequence Jt = (x1 J, xt2 , . . . , xtd ) gives
a descending chain of irreducible ideals cofinal with the powers of m. It is easy to check
t−1
that δt = xt−1
· · · xt−1
of R/Jt using that x1 , . . . , xd form a regular
1 x2
d δ generatese the socle
e
e
pe
sequence, which further implies (Jtp : δtp ) = ((J, xt2 , . . . , xtd )[p ] : (xt−1
· · · xt−1
2
d δ) ) for all
t, e ∈ N. Thus, Theorem 6.4 gives
e
e
1
s(R) = lim ed ℓR (R/(Jtp : δtp ))
t→∞ p
t
t
t−1
t
t
t−1 t−1
= lim eHK ((xt−1
· · · xt−1
1 J, x2 , . . . , xd )) − eHK ((x1 J, x2 , . . . , xd , (x1 x2
d δ)))
t→∞
1
t
t [pe ]
t−1
t−1 pe
ℓ
(R/((J,
x
,
.
.
.
,
x
)
:
(x
·
·
·
x
R
2
d
2
d δ) ))
t→∞ ped
= lim eHK ((J, xt2 , . . . , xtd )) − eHK ((J, xt2 , . . . , xtd , (xt−1
· · · xt−1
2
d δ)))
= lim
t→∞
using the relation in (11) once more.
One can push the above analysis further still. In the notation of the previous proof,
when R is normal x2 can be chosen so that Rx2 is Gorenstein and Jx2 is principal. This
allows one to remove the exponent t on x2 in the limit above using (i.) from the subsequent
lemma. Following the methods of [Abe02] (cf. [AE06, Mac96, Yao06]), we present a complete
treatment in Corollary 6.8 below.
26
Lemma 6.7. Suppose that (R, m, k) is a complete Cohen-Macaulay local F -finite normal
domain of dimension d, and D an effective Weil divisor on Spec (R). Put J = R(−D) ⊆ R
so that J (n) = R(−nD) for n ∈ N. Let x1 ∈ J and x2 , . . . , xd ∈ R be chosen so that x1 , . . . , xd
give a system of parameters for R, and fix e ∈ N.
(i.) If x2 J ⊆ a2 R for some a2 ∈ J, there exists b2 ∈ J so that a2 , x2 + b2 , x3 , . . . , xd give a
system of parameters for R. Moreover, for any non-negative integers N2 , . . . , Nd with
N2 ≥ 2, we have that
e
e
e
(N −1)pe
e
2p
3p
((J (p ) , xN
, xN
, . . . , xdNd p ) : x2 2
)
2
3
(N2 −1)pe
Nd p e
N3 p e
N2 p e
[pe ]
= ((J , x2 , x3 , . . . , xd ) : x2
)
2pe
N3 p e
Nd p e
pe
[pe ]
= ((J , x2 , x3 , . . . , xd ) : x2 ).
(ii.) Suppose xnd J (n) ⊆ ad R for some n ∈ N and ad ∈ J (n) . Then there exists bd ∈ J so
that ad , x2 , . . . , xd−1 , xd + bd give a system of parameters for R. Moreover, for any
non-negative integers N2 , . . . , Nd with Nd ≥ 2, we have that
e
e
N
pe
e
(N −1)pe
d−1
2p
((J (p ) , xN
, . . . , xd−1
, xdNd p ) : xd d
2
e
e
e
e
e
Nd−1 p
2p
n p
, . . . , xd−1
, x2p
⊆ ((J (p ) , xN
2
d ) : x1 xd )
)
Proof. For (i.), note first that a2 is a non-zero divisor on R/(x3 , . . . , xd ) as its multiple x1 x2 ∈
x2 J is a non-zero divisor. Hence dim(R/(x3 , . . . , xd , a2 )) = d−1. Since ℓR (R/(J, x2 , . . . , xd )) ≤
ℓR (R/(x1 , . . . , xd )) < ∞, we know that (x2 , J) is not contained in any minimal prime ideal
of (x3 , . . . , xd , a2 ). In particular, using standard prime avoidance arguments, one can find
b2 ∈ J so that x2 + b2 also avoids all of the minimal primes of (x3 , . . . , xd , a2 ); it follows that
the sequence x3 , . . . , xd , a2 , x2 + b2 is a system of parameters for R.
Since R(− div(x2 ) − D) = x2 J ⊆ a2 R = R(− div(a2 )), we have that D + div(x2 ) ≥ div(a2 )
e
e
e
e
e
and hence also pe D + div(xp2 ) ≥ div(ap2 ); it follows that xp2 J (p ) ⊆ ap2 R. The ideals
e
e
3p
inclusions ⊇ in (i.) are clear, so we need only check the reverse. Let I = (xN
, . . . , xdNd p )
3
e
e
(N −1)pe
2p
and suppose that we have cx2 2
∈ (J (p ) , xN
, I) for some c ∈ R. We can find r2 ∈ R
2
e
e
e
(N2 −1)pe
pe
(pe )
with (c − r2 x2 )x2
∈ (J , I). Since b2 ∈ J, we have bp ∈ J (p ) and so (c − r2 xp2 )(x2 +
e
e
e
e
e
e
b2 )(N2 −1)p ∈ (J (p ) , I). Multiplying through by xp2 and using that xp2 J (p ) ⊆ ap2 gives
e
e
e
(N2 −1)pe
∈ (ap2 , I). Using that a2 , x2 + b2 , x3 , . . . , xd give a system
(cxp2 − r2 x2p
2 )(x2 + b2 )
e
e
e
e
e
of parameters for R, we see (cxp2 − r2 x2p
) ∈ (ap2 , I) ⊆ (J [p ] , I) and conclude cxp2 ∈
2
e
e
(J [p ] , x2p
2 , I) as desired.
Statement (ii.) proceeds similarly. We have that ad is a non-zero divisor on R/(x2 , . . . , xd−1 )
as its multiple xnd xn1 is a non-zero divisor, and again (xd , J) is not contained in any minimal
prime of (ad , x2 , . . . , xd−1 ). We can find bd ∈ J so that xd + bd is not in any minimal prime
of (ad , x2 , . . . , xd−1 ) and thus ad , x2 , . . . , xd−1 , xd + bd is a system of parameters for R. From
e
e
⌊pe /n⌋
xnd J (n) ⊆ ad R, we have nD + div(xnd ) ≥ div(ad ) and it follows xpd J (p ) ⊆ ad
R. To see
e
Nd−1 pe
(Nd −1)pe
N2 p e
(pe )
dp
∈ (J , I, xN
) for
the final inclusion, let I = (x2 , . . . , xd−1 ) and suppose cxd
d
(Nd −1)pe
pe
(pe )
some c ∈ R. We can find rd ∈ R so that (c − rd xd )xd
∈ (J , I). As bd ∈ J so
e
pe
pe
(pe )
(Nd −1)pe
(pe )
bd ∈ J , hence also (c − rd xd )(xd + bd )
∈ (J , I). Multiplying through by xpd
e
e
⌊pe /n⌋
(Nd −1)pe
gives (cxpd − rd x2p
∈ (ad
, I). Using that ad , x2 , . . . , xd−1 , xd + bd are
d )(xd + bd )
⌊pe /n⌋
pe
2pe
a system of parameters, we see (cxd − rd xd ) ∈ (ad
, I). Multiplying through by xn1
27
⌊pe /n⌋
e
e
e
e
e
(p )
and using that ad
xn1 ∈ (J (n) )⌊p /n⌋+1 ⊆ J (p ) gives (cxn1 xpd − rd xn1 x2p
, I) and in
d ) ∈ (J
e
e
e
particular cxn1 xpd ∈ (J (p ) , I, x2p
)
as
desired.
d
Corollary 6.8. Suppose that (R, m, k) is a complete Cohen-Macaulay local F -finite domain
of dimension d.
(i.) [HL02] If R is Gorenstein and x1 , . . . , xd are any system of parameters for R and δ ∈ R
e
e
generates the socle of R/(x1 , . . . , xd ), then IeF−sig = ((x1 , . . . , xd )[p ] : δ p ) for all e ∈ N
and s(R) = eHK ((x1 , . . . , xd )) − eHK ((x1 , . . . , xd , δ)).
(ii.) If the punctured spectrum Spec (R) \ {m} is Q-Gorenstein, there exists an m-primary
ideal I and x ∈ R with (I : x) = m so that s(R) = eHK (I) − eHK ((I, x)) and moreover
e
e
IeF−sig ⊆ ((I [p ] )∗ : xp ) for all e ∈ N. In particular, if R is weakly F -regular, IeF−sig =
e
e
(I [p ] : xp ) for all e ∈ N.
Proof. (i.) Since R is Gorenstein, we may take J = (x1 ) to be the canonical ideal above in
e
(t−1)pe
(t−1)pe pe
tpe
Corollary 6.6. For fixed e ∈ N, we have that IeF−sig = ((xtp
δ ))
· · · xd
1 , . . . , xd ) : (x1
for all t ≫ 0. However, since x1 , . . . , xd are a regular sequence on R, it follows that
e
e
e
e
(t−1)pe
(t−1)pe pe
tpe
((xtp
δ )) = ((xp1 , . . . , xpd ) : δ p ) for any t ∈ N. The final
· · · xd
1 , . . . , xd ) : (x1
statement follows again using the relation in (11) once more.
For (ii.), we begin with the following construction. Let J be a canonical ideal of R, and
choose 0 6= x1 ∈ J. Let U2 be the complement of the minimal primes of x1 , which are all
height one contain the set of minimal primes of J. We have that U2−1 J is principal (as U2−1 R
is a semi-local Dedekind domain) and can choose x2 ∈ U2 and a2 ∈ J so that Jx2 = a2 Rx2
and x2 J ⊆ a2 R.
(n)
As Rp is Q-Gorenstein for all p ∈ Spec (R) \ {m}, we may choose n ∈ N so that Jp
is principal for all p ∈ Spec (R) \ {m}. Assuming x1 , . . . , xi−1 have already been chosen,
let Ui be the complement of the set of minimal primes of x1 , . . . , xi−1 . Again, Ui−1 J (n) is
principal (as Ui−1 R is a semilocal domain, every locally principal ideal is principal by [Kap74,
(n)
Theorem 60]) so we can choose xi ∈ Ui and ai ∈ J (n) so that Jxi = ai Rxi and xni J (n) ⊆ ai R.
Continuing in this manner gives a system of parameters x1 , . . . , xd of R with x1 ∈ J, and a
sequence of elements a3 , . . . , ad ∈ J (n) so that xni J (n) ⊆ ai R for each i = 3, . . . , d. For fixed
e ∈ N and t ≫ 0, we have from Corollary 6.6 and repeated application of Lemma 6.7 (ii.)
for xd , . . . , x3 and then Lemma 6.7 (i.) for x2 that
e
(16)
e
e
e
(t−1)pe
e
(d−2)n (t−1)pe
(t−1)pe (t−1)pe
e
tp
tp
IeF−sig = ((J [p ] , xtp
· · · xd−1 xd
δ p ))
2 , . . . , xd−1 , xd ) : (x2
e
e
e
e
e
e
e
(t−1)p
(t−1)p (t−1)p pe
tp
tp
δ ))
⊆ ((J (p ) , xtp
· · · xd−1 xd
2 , . . . , xd−1 , xd ) : (x2
e
(t−1)pe pe pe
tpe
tpe
2pe
(pe )
n (t−1)p
⊆ ((J , x2 , . . . , xd−1 , xd ) : (x1 x2
· · · xd−1 xd δ ))
..
..
..
⊆
.
.
.
e
e
e
e
2p
2p
⊆ ((J (p ) , xtp
x2
xp3 · · · xpd δ p ))
2 , x3 , . . . , xd ) : (x1
e
e
e
e
e
e
e
(d−2)n p p
2p
2p
2p
[p ]
pe
= ((J
x2 x3 · · · xpd δ
))
3 , . . . , xd ) : (x1
, x2 , x
(d−2)n
2pe
2pe
pe
pe pe
[pe ]
=
(J , x2 , . . . , xd ) : (x2 · · · xd δ ) : x1
e
e
so setting I = (J, x22 , . . . , x2d ) and x = x2 · · · xd δ gives s(R) = eHK (I) − eHK ((I, x)) using
e
e
(d−2)n
2pe
pe
pe pe
F−sig
, and c = x1
).
Lemma 4.13 (with Ie = ((J [p ] , x2p
2 , . . . , xd ) : (x2 · · · xd δ )), Je = Ie
28
e′
F−sig
For the final statement, since (IeF−sig )[p ] ⊆ Ie+e
for all e, e′ ∈ N, it follows from (16) that
′
\
′
′
(d−2)n
[pe+e ]
pe F−sig [pe ]
0 6= x1
∈
I
: (x Ie
)
e′ ∈N
pe
[pe ]
e
e
so that x IeF−sig ⊆ (I )∗ or IeF−sig ⊆ ((I [p ] )∗ : xp ) as claimed. In particular, if R is weakly
e
e
F -regular, we have that IeF−sig ⊆ ((I [p ] ) : xp ) and equality follows from Lemma 6.1.
One can also use Theorem 6.3 to show that the length criteria in Corollary 5.7 and Corollary 5.8 also determine the F -signatures in those settings.
Corollary 6.9. Let (R, m, k) be a complete F -finite local domain of dimension d. Let a ⊆ R
be a non-zero ideal and ξ ∈ R≥0 . Suppose {Jt }t∈N is a descending chain of irreducible ideals
cofinal with the powers of m, and δt ∈ R generates the socle of R/Jt . Then s(R, aξ ) =
e
e
[pe ]
limt→∞ lime→∞ p1ed ℓR (R/(Jt : a⌈ξ(p −1)⌉ δtp )).
Proof. It is straightforward to check for each e ∈ N that
ξ
Iea
e
ξ
= (x ∈ R | φ(x1/p ) ∈ m for all φ ∈ Cea )
e
e
e
= (x ∈ R | (xy)1/p R ⊆ R1/p is not split for any y ∈ a⌈ξ(p −1)⌉ )
P
e
ξ
e
[pe ]
aξ
aξ
and it follows that Iea = t∈N It,e
where It,e
= (Jt : δtp a⌈(p −1)ξ⌉ ) for t, e ∈ N. Choosing
e+1
e
an element 0 6= c ∈ R such that ca⌈ξ(p −1)⌉ ⊆ (a⌈ξ(p −1)⌉ )[p] for all e ∈ N as in the proof
e
aξ
aξ [p]
aξ
aξ
aξ
, c((It,e
) ) ⊆ It,e+1
, and It,e
⊆ It+1,e
for all
of Theorem 4.11, we have that m[p ] ⊆ It,e
t, e ∈ N. The result now follows immediately from Theorem 6.3 as we know s(R, aξ ) =
ξ
lime→∞ p1ed ℓR (R/Iea ).
Corollary 6.10. Let (R, m, k) be a complete F -finite local normal domain of dimension d.
Suppose {Jt }t∈N is a descending chain of irreducible ideals cofinal with the powers of m,
and δt ∈ R generates the socle of R/Jt . If ∆ is an effective Q-divisor on Spec (R), then
e
[pe ]
s(R, ∆) = limt→∞ lime→∞ p1ed ℓR (R/(Jt R(⌈(pe − 1)∆⌉) : δtp )).
Proof. Again it is straightforward to check for each e ∈ N that
e
Ie(R,∆) = (x ∈ R | φ(x1/p ) ∈ m for all φ ∈ Ce(R,∆) )
e
e
= (x ∈ R | (x)1/p R ⊆ (R(⌈(pe − 1)∆⌉))1/p is not split)
P
e
(R,∆)
(R,∆)
(R,∆)
[pe ]
and it follows that Ie
= t∈N It,e where It,e = (Jt R(⌈(pe −1)∆⌉) : δtp ) for t, e ∈ N.
Choosing an element 0 6= c ∈ R so that divR (c) ≥ p⌈∆⌉ as in the proof of Theorem 4.12,
e
(R,∆)
(R,∆)
(R,∆)
(R,∆)
aξ
we have that we have that m[p ] ⊆ It,e , c((It,e )[p]) ⊆ It,e+1 , and It,e
⊆ It+1,e
for all
t, e ∈ N. The result now follows immediately from Theorem 6.3 as we know s(R, ∆) =
(R,∆)
).
lime→∞ p1ed ℓR (R/Ie
Lastly, note that Theorem 6.3 can also be applied outside the context of descending chains
of irreducible ideals, and gives yet another perspective on the original proof of the existence
of the F -signature.
Corollary 6.11. [Tuc12, Proof of Theorem 4.9] Let (R, m, k) be a complete F -finite local
e
e
domain of dimension d and IeF−sig = (r ∈ R | φ(r 1/p ) ∈ m for all φ ∈ Hom R (R1/p , R)).
Then s(R) = limt→∞ p1td eHK (ItF−sig ).
29
e
(ItF−sig )[p ]
t<e
[p]
. It is easily checked that m[p ] ⊆ It,e , It,e ⊆ It,e+1 ,
Proof. Let It,e =
F−sig
Ie
t≥e
e
F−sig
and It,e ⊆ It+1,e for all t, e ∈ N using that m[p ] ⊆ IeF−sig and (IeF−sig )[p] ⊆ Ie+1
for all e ∈ N.
The result now follows immediately from Theorem 6.3, as we have
e−t
1
1
1
ℓR (IeF−sig ) = lim lim ed ℓR (R/It,e ) = lim lim ed ℓR (R/It,e )
ed
e→∞ t→∞ p
t→∞ e→∞ p
e→∞ p
1
1
e−t
= lim lim ed ℓR (R/(ItF−sig )[p ] ) = lim td eHK (ItF−sig )
t→∞ e→∞ p
t→∞ p
s(R) =
lim
as desired.
7. Open Questions
In this section, we collect together some information on the important questions left unanswered in this article. We have seen that Hilbert-Kunz multiplicity and the F-signature enjoy
semi-continuity properties for F -finite rings; it is not difficult to extend these results to rings
which are essentially of finite type over an excellent local ring. More generally, however, this
raises the following question.
Question 7.1. If R is an excellent domain that is not F -finite or essentially of finite type
over an excellent local ring, do the Hilbert-Kunz multiplicity and F -signature determine
semicontinuous R-valued functions on Spec (R)?
In the case of F -signature, a positive answer would imply the openness of the strongly F regular locus for such a ring. Note that this question would seem closely related to the
existence of (locally and completely stable) test elements for excellent domains, which also
remains unanswered.
Perhaps the most important question left open in this article regarding the relationship
between Hilbert-Kunz multiplicity and F -signature is the following, which (in light of the
results of the previous section) is attributed to Watanabe and Yoshida in [WY04].
Question 7.2. Let (R, m, k) be a complete local F -finite normal domain. Do there exist
m-primary ideals I ( J so that eHK (I) − eHK (J) = s(R)?
For any ring such that Question 7.2 has a positive answer, it follows from weak F -regularity
is equivalent to strong F -regularity by the length criterion for tight closure [HH94, Theorem 8.17] (Corollary 5.6 above).
We see from Corollary 6.8 that Question 7.2 is true provided R is Q-Gorenstein on the
punctured spectrum. When in addition R is weakly F -regular, the stronger condition below
is satisfied as well.
Question 7.3. Let (R, m, k) be a complete local F -finite normal domain. Do there exist
m-primary ideals I ( J so that
e
e
e
frkR (R1/p )
ℓR (R/I [p ] ) − ℓR (R/J [p ] )
=
[k 1/pe : k]
ℓR (J/I)
for all e ∈ N?
30
The importance of Question 7.3 stems from the observation that it allows one to apply the
results of Huneke, McDermott, and Monsky [HMM04]; for an m-primary ideal I ⊆ R, they
e
show the existence of a constant α(I) ∈ R so that ℓ(R/I [p ] ) = eHK (I)ped + α(I)pe(d−1) +
O(pe(d−2) ). In other words, Hilbert-Kunz functions in normal local F -finite domains are
polynomial in pe to an extra degree. In particular, for any ring such that Question 7.3 has
a positive answer, so also does the following question.
Question 7.4. If (R, m, k) is a complete local F -finite normal domain, when does there exist
e
frkR (R1/p )
a positive constant α(R) ∈ R so that
= s(R)ped + α(R)pe(d−1) + O(pe(d−2) )?
[k 1/pe : k]
Finally, we have tried to emphasize the applicability of our techniques to the settings of
divisor and ideal pairs throughout, and the questions above are readily generalized to those
settings. Moreover, it may well be the case that answers to the questions above require the
use of such pairs (particularly to remove Gorenstein or Q-Gorensetein hypotheses). In this
direction, and in view of [Sch11], one could imagine a positive and constructive answer to
the question below could prove quite useful.
Question 7.5. Suppose (R, m, k) is an F -finite local normal domain and X = Spec (R). For
each ǫ > 0, does there exist an effective Q-divisor ∆ on X such that KX + ∆ is Q-Cartier
with index prime to p and so that s(R) − s(R, ∆) < ǫ? In case R is not local, can ∆ be
chosen globally to have this property locally over Spec (R)?
References
[Abe02]
[AE05]
[AE06]
[AE08]
[AL03]
[BB11]
[BE04]
[BH93]
[Bre13]
[BS13]
[BST12]
[BST13]
[Che43]
[EY11]
Ian M. Aberbach, Some conditions for the equivalence of weak and strong F -regularity, Comm.
Algebra 30 (2002), no. 4, 1635–1651. 26
Ian M. Aberbach and Florian Enescu, The structure of F-pure rings, Math. Z. 250 (2005), no. 4,
791–806. 22
, When does the F-signature exists?, Ann. Fac. Sci. Toulouse Math. (6) 15 (2006), no. 2,
195–201. MR 2244213 26
, Lower bounds for Hilbert-Kunz multiplicities in local rings of fixed dimension, Michigan
Math. J. 57 (2008), 1–16, Special volume in honor of Melvin Hochster. MR 2492437 1
Ian M. Aberbach and Graham J. Leuschke, The F -signature and strong F -regularity, Math. Res.
Lett. 10 (2003), no. 1, 51–56. MR 1960123 1, 2, 17, 20
Manuel Blickle and Gebhard Böckle, Cartier modules: finiteness results, J. Reine Angew. Math.
661 (2011), 85–123. MR 2863904 19
Manuel Blickle and Florian Enescu, On rings with small Hilbert-Kunz multiplicity, Proc. Amer.
Math. Soc. 132 (2004), no. 9, 2505–2509 (electronic). MR 2054773 1
Winfried Bruns and Jürgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39, Cambridge University Press, Cambridge, 1993. MR 1251956 18, 26
H. Brenner, Irrational Hilbert-Kunz multiplicities, ArXiv e-prints (2013). 1
Manuel Blickle and Karl Schwede, p−1 -linear maps in algebra and geometry, Commutative algebra,
Springer, New York, 2013, pp. 123–205. MR 3051373 3, 10, 11
Manuel Blickle, Karl Schwede, and Kevin Tucker, F -signature of pairs and the asymptotic behavior
of Frobenius splittings, Adv. Math. 231 (2012), no. 6, 3232–3258. MR 2980498 2, 3, 11, 22
, F -signature of pairs: continuity, p-fractals and minimal log discrepancies, J. Lond. Math.
Soc. (2) 87 (2013), no. 3, 802–818. MR 3073677 3, 11
Claude Chevalley, On the theory of local rings, Ann. of Math. (2) 44 (1943), 690–708. MR 0009603
18, 20
Florian Enescu and Yongwei Yao, The lower semicontinuity of the Frobenius splitting numbers,
Math. Proc. Cambridge Philos. Soc. 150 (2011), no. 1, 35–46. MR 2739072 9
31
[Gab04]
Ofer Gabber, Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, Walter
de Gruyter GmbH & Co. KG, Berlin, 2004, pp. 711–734. MR 2099084 19
[Gri01]
P. A. Grillet, Commutative semigroups, Advances in Mathematics (Dordrecht), vol. 2, Kluwer
Academic Publishers, Dordrecht, 2001. MR 2017849 13
[Har98] Nobuo Hara, A characterization of rational singularities in terms of injectivity of Frobenius maps,
Amer. J. Math. 120 (1998), no. 5, 981–996. MR 1646049 3
[HH90]
Melvin Hochster and Craig Huneke, Tight closure, invariant theory, and the Briançon-Skoda
theorem, J. Amer. Math. Soc. 3 (1990), no. 1, 31–116. MR 1017784 2, 6, 20, 24
[HH91]
, Tight closure and elements of small order in integral extensions, J. Pure Appl. Algebra
71 (1991), no. 2-3, 233–247. 18, 19
[HH94]
, F -regularity, test elements, and smooth base change, Trans. Amer. Math. Soc. 346 (1994),
no. 1, 1–62. MR 1273534 1, 2, 30
[HH00]
, Localization and test exponents for tight closure, Michigan Math. J. 48 (2000), 305–329,
Dedicated to William Fulton on the occasion of his 60th birthday. 6
[HL02]
Craig Huneke and Graham J. Leuschke, Two theorems about maximal Cohen-Macaulay modules,
Math. Ann. 324 (2002), no. 2, 391–404. MR 1933863 1, 2, 28
[HM93] C. Han and P. Monsky, Some surprising Hilbert-Kunz functions, Math. Z. 214 (1993), no. 1,
119–135. MR 1234602 1
[HMM04] Craig Huneke, Moira A. McDermott, and Paul Monsky, Hilbert-Kunz functions for normal rings,
Math. Res. Lett. 11 (2004), no. 4, 539–546. MR 2092906 31
[Hoc]
Melvin Hochster, Foundations of tight closure theory, http://www.math.lsa.umich.edu/~hochster/711F07/fndt
Lecture notes from a course taught at the University of Michigan, Fall 2007. 20, 24
[Hoc77]
, Cyclic purity versus purity in excellent Noetherian rings, Trans. Amer. Math. Soc. 231
(1977), no. 2, 463–488. MR 0463152 23, 26
[HS77]
Robin Hartshorne and Robert Speiser, Local cohomological dimension in characteristic p, Ann. of
Math. (2) 105 (1977), no. 1, 45–79. MR 0441962 19
[Hun96] Craig Huneke, Tight closure and its applications, CBMS Regional Conference Series in Mathematics, vol. 88, Published for the Conference Board of the Mathematical Sciences, Washington,
DC; by the American Mathematical Society, Providence, RI, 1996, With an appendix by Melvin
Hochster. 20
[HW02] Nobuo Hara and Kei-Ichi Watanabe, F-regular and F-pure rings vs. log terminal and log canonical
singularities, J. Algebraic Geom. 11 (2002), no. 2, 363–392. MR 1874118 3, 11, 21
[HY03]
Nobuo Hara and Ken-Ichi Yoshida, A generalization of tight closure and multiplier ideals, Trans.
Amer. Math. Soc. 355 (2003), no. 8, 3143–3174 (electronic). MR 1974679 3, 11, 21
[Kap74] Irving Kaplansky, Commutative rings, revised ed., The University of Chicago Press, Chicago,
Ill.-London, 1974. 28
[KS15]
K. Kurano and K. Shimomoto, An elementary proof of Cohen-Gabber theorem in the equal characteristic p > 0 case, ArXiv e-prints (2015). 5
[Kun69] Ernst Kunz, Characterizations of regular local rings for characteristic p, Amer. J. Math. 91 (1969),
772–784. MR 0252389 1
[Kun76]
, On Noetherian rings of characteristic p, Amer. J. Math. 98 (1976), no. 4, 999–1013.
MR 0432625 5, 6, 9
[Lyu97] Gennady Lyubeznik, F -modules: applications to local cohomology and D-modules in characteristic
p > 0, J. Reine Angew. Math. 491 (1997), 65–130. MR 1476089 19
[Mac96] Brian Cameron Maccrimmon, Strong F-regularity and boundedness questions in tight closure, ProQuest LLC, Ann Arbor, MI, 1996, Thesis (Ph.D.)–University of Michigan. 26
[Mon83] P. Monsky, The Hilbert-Kunz function, Math. Ann. 263 (1983), no. 1, 43–49. MR 697329 1, 6, 13
[MS97]
V. B. Mehta and V. Srinivas, A characterization of rational singularities, Asian J. Math. 1 (1997),
no. 2, 249–271. MR 1491985 3
[Pol15]
T. Polstra, Uniform Bounds in F-Finite Rings and Lower Semi-Continuity of the F-Signature,
ArXiv e-prints (2015). 1, 2, 7, 9, 14
[Sch11]
Karl Schwede, Test ideals in non-Q-Gorenstein rings, Trans. Amer. Math. Soc. 363 (2011), no. 11,
5925–5941. 31
32
[Smi97]
Karen E. Smith, F -rational rings have rational singularities, Amer. J. Math. 119 (1997), no. 1,
159–180. MR 1428062 3
[Smi16] Ilya Smirnov, Upper semi-continuity of the Hilbert-Kunz multiplicity, Compos. Math. 152 (2016),
no. 3, 477–488. MR 3477638 1, 2, 7
[ST12]
Karl Schwede and Kevin Tucker, A survey of test ideals, Progress in commutative algebra 2,
Walter de Gruyter, Berlin, 2012, pp. 39–99. MR 2932591 10
[ST14]
, On the behavior of test ideals under finite morphisms, J. Algebraic Geom. 23 (2014),
no. 3, 399–443. MR 3205587 2, 10
[ST15]
, Explicitly extending Frobenius splittings over finite maps, Comm. Algebra 43 (2015),
no. 10, 4070–4079. MR 3366560 2
[SVdB97] Karen E. Smith and Michel Van den Bergh, Simplicity of rings of differential operators in prime
characteristic, Proc. London Math. Soc. (3) 75 (1997), no. 1, 32–62. MR 1444312 1, 2
[Tak04] Shunsuke Takagi, An interpretation of multiplier ideals via tight closure, J. Algebraic Geom. 13
(2004), no. 2, 393–415. MR 2047704 3, 11, 21
[Tuc12] Kevin Tucker, F -signature exists, Invent. Math. 190 (2012), no. 3, 743–765. MR 2995185 1, 2, 6,
12, 22, 29
[WY00] Kei-ichi Watanabe and Ken-ichi Yoshida, Hilbert-Kunz multiplicity and an inequality between
multiplicity and colength, J. Algebra 230 (2000), no. 1, 295–317. MR 1774769 1
[WY04]
, Minimal relative Hilbert-Kunz multiplicity, Illinois J. Math. 48 (2004), no. 1, 273–294.
MR 2048225 1, 2, 24, 26, 30
[Yao06] Yongwei Yao, Observations on the F -signature of local rings of characteristic p, J. Algebra 299
(2006), no. 1, 198–218. MR 2225772 2, 26
E-mail address: [email protected]
Department of Mathematics, University of Missouri-Columbia, Columbia, MO 65211
E-mail address: [email protected]
Department of Mathematics, University of Illinois at Chicago, Chicago, IL 60607
33
| 0math.AC
|
Learn-Memorize-Recall-Reduce: A Robotic Cloud
Computing Paradigm
Shaoshan Liu, Bolin Ding, Jie Tang, Dawei Sun, Zhe Zhang, Grace Tsai, and Jean-Luc Gaudiot
ABSTRACT
The rise of robotic applications has led to the generation of
a huge volume of unstructured data, whereas the current
cloud infrastructure was designed to process limited
amounts of structured data. To address this problem, we
propose a learn-memorize-recall-reduce paradigm for
robotic cloud computing. The learning stage converts
incoming unstructured data into structured data; the
memorization stage provides effective storage for the
massive amount of data; the recall stage provides efficient
means to retrieve the raw data; while the reduction stage
provides means to make sense of this massive amount of
unstructured data with limited computing resources.
Keywords
Cloud Architecture; Distributed Computing; Storage;
Robotics
1. INTRODUCTION
Robots are mobile devices, and the rise of robotic
applications has imposed tremendous pressure on our
existing cloud infrastructure. For instance, every day, a
mobile phone sends out at least 30 MB of structured data to
the clouds of the service provider, the application operators
etc. (structured data means data that can be easily understood,
stored, and retrieved by machines). In contrast, even a very
simple robot can easily generate over 1 GB of unstructured
multimedia data per day. An extreme form of robot,
driverless cars, can generate as much as 2 GB of unstructured
data per second [1] (unstructured data means data that cannot
be easily understood, stored, and retrieved by machines).
Therefore, we are facing the urgent challenge of
designing and implementing a cloud architecture to process
this massive robotic unstructured data. For instance, in a
real-world scenario, in-home service robots may act in a
surveillance role, patrolling the users’ house and recording
captured videos. Then, users demand the capability of video
playback based on intelligent queries using time, location, as
well as objects in scene as inputs.
In sections 2, 3, and 4, we first present our
implementation of a robotic cloud infrastructure and discuss
how this infrastructure can meet this challenge. Specifically,
we generalize the requirements into a robotic cloud
computing paradigm - learn-memorize-recall, where
learning is about how to automatically understand the
unstructured data and convert it into structured data;
memorization is about how to effectively store the massive
amount of data, while recall is about how to efficiently
retrieve the data when needed.
Nonetheless, as the volumes of data collected from
robotic devices exponentially grow over time, having a great
robotic cloud infrastructure is not sufficient. Then in section
5, we delve into the final part of the proposed paradigm,
reduction, the utilization of data reduction techniques to
making sense of massive amounts of unstructured data with
a limited budget of computing resources.
2. ROBOTIC CLOUD ARCHITECTURE
To provide the learn-memorize-recall features, we have
designed and implemented a cloud architecture, which is
illustrated in Figure 1 below. The architecture consists of
the following components:
•
Robotic Client Devices: these devices capture
multimedia feeds and send the feeds to the cloud along
with their meta data.
•
Streaming Server: this server handles incoming
multimedia and streams on-demand live multimedia
feeds to users as requested.
•
Object Recognition: this is a deep-learning evaluation
engine for automatic extraction of semantic information
from incoming videos.
•
Key-Value Store: this key-value store organizes the
video feeds along with the learned/extracted semantic
information.
•
Query Engine: this query engine supports retrieval of
video feeds. One can search using any combination of
time, location, as well as extracted labels.
•
Business Analytics Engine: this engine generates highlevel statistics of all multimedia data. For example, one
can be interested in knowing which are the most
common objects appear in living rooms, etc.
•
Storage Layer: the storage layer needs to provide high
throughput for data persistence and low latency for fast
retrieval of video feeds. Also, it must manage
heterogeneous storage systems including S3, GCS,
Swift, HDFS, OSS, GlusterFS, and NFS.
of labels>>, and the value is a file path of the raw data in the
storage layer.
This poses two challenges: first, in real-world scenarios,
users vary in their choice of persistent storage for their data.
Some prefer Amazon S3, some their own deployment of
HDFS, others Ceph, etc. One way to get around this problem
is to create a set of APIs for each persistent storage, but this
would become impossible to manage as the number of
persistent storage options grows. A second, and probably
better way, to handle this is to create a unified storage layer
to abstract all underlying storages.
Figure 1: Robotic Cloud Architecture Overview
2.1 Learning
As multimedia data comes in to the cloud system, the first
task is to learn from the raw data and to extract semantic
information out of the multimedia data. For video streams,
we can extract object labels from frames and associate these
labels with the video stream. For audio streams, we can
extract sentences of the spoken language. Then this
semantic information can be used as keys, and the raw data
streams can be used as values, and together they are stored
in the key-value store.
To this end, Alluxio enables effective data management
across different storage systems through its use of
transparent naming and mounting API [6]. Transparent
naming maintains an identity between the Alluxio
namespace and the underlying storage system namespace.
When users create objects in the Alluxio namespace, they
can decide whether these objects should be persisted in the
underlying storage system. For objects that are persisted,
Alluxio preserve the object paths, relative to the underlying
storage system directory in which Alluxio objects are stored.
With this feature, we can now manage multiple
persistent storages using a single set of storage layer API,
which greatly simplifies the management of the
memorization part of the robotic cloud computing paradigm.
The second challenge is that the write throughput
directly impacts the performance of the whole system. If the
write speed of the memorization is slower than the detection
speed in learning stage, then it becomes the bottleneck and
leads to “memory loss.” We will discuss in the next section
how we can use Alluxio’s tiered storage feature, along with
write optimization to improve write throughput.
2.3 Recall
Figure 2: Video Stream Processing
As presented in Figure 2, we need a learning engine that
extracts object labels from a video stream. This engine needs
to be accurate in terms of recognition rate, and this engine
needs to be fast in order for us to capture as many objects as
possible in the video. Therefore, in this implementation, we
utilize faster r-cnn [3] network running on Caffe [4]. Faster
r-cnn introduces a Region Proposal Network (RPN), a fullyconvolutional network that simultaneously predicts object
bounds and object scores at each position, hence achieving
very high detection speed without sacrificing detection
accuracy.
2.2 Memorization
After extracting the semantic information from the raw
multimedia data, the extracted labels, along with the raw
data get stored in the key-value store for easy retrieval
(Figure 3). We implemented the key-value store using
MongoDB [5]. In this case, the key is the meta data
including <sessionID, timestamp, duration, location, <list
The last stage is Recall, in which the users can use any
combination of meta data, such as time, location, or detected
labels to accurately retrieve the target multimedia data. In
our implementation, we use MongoDB to perform the
intelligent search using the metadata as keys to first retrieve
a file path of the storage layer. Then we issue another
request to retrieve the raw data using the file path.
Therefore, the performance of the recall stage greatly
depends on the read performance of the storage layer. It is
helpful to exploit locality of the storage layer to improve
read performance. Ideally, the performance is highest if the
requested data is located in the memory of the local machine.
However, it is impossible to store all data in the memory of
the local machines. We need a mechanism to provide a
cache-like structure such that we have different levels of the
storage at different speed, including local memory, local
Solid State Drives (SSD), local Hard Disk Drives (HDD),
and remote storage.
This requirement can be fulfilled by Alluxio’s tiered
storage feature. With tiered storage, Alluxio can manage
multiple storage layers including Memory, SSD, and HDD.
Using tiered storage Alluxio can store more data in the
system at the same time, since memory capacity may be
limited in some deployments. With tiered storage, Alluxio
automatically manages blocks between all the configured
tiers, so users and administrators do not have to manually
manage the locations of the data. Users may specify their
own data management strategies by implementing allocators
and evictors.
In a way, the Memory layer of the tiered storage serves
as the top level cache, SSD serves as the second level cache,
HDD serves as the third level cache, while it is the persistent
storage is the last level storage. We will discuss in the next
section how we can use Alluxio’s tiered storage feature,
along with prefetching optimization to improve read latency.
Figure 3: Storage Layer Architecture
3. OPTIMIZATIONS
As mentioned in section 2, the storage layer write throughput
is critical to the performance of the memorization stage and
the storage layer read latency is critical to the performance
of the recall stage. In this section, we delve into the
optimizations we implement in the storage layer to improve
both write throughput and read latency.
3.1 Write Optimization
With the default Alluxio tiered storage implementation,
when a user writes a new block, it is written to the top tier
by default. If there is not enough space for the block in the
top tier, the system checks for free space in the next layer.
This gets repeated for each layer until free space is found.
For instance, if both the memory layer and the SSD layer are
full, free space may be found in the HDD layer. Then, the
evictor moves a block in the SSD layer to the HDD layer,
and then moves a block in the memory layer to the SSD
layer. Finally the block can be written back to the memory
layer.
This approach is inefficient when the tiered storage is
fully occupied as each time a new block is written, the
evictor needs to move blocks across all layers before the new
block can be written. This greatly reduces the write
throughput as it takes a long time to write each block in the
storage layer.
To allay this problem, we implement a Direct-Write
allocator, such that it writes a block to the first layer that has
free space. When we run a stress test to continually write
blocks into Alluxio, compared to the default allocator, our
Direct-Write approach reduces the write latency by half, and
consequently doubling the write throughput.
One could argue that the Write-Direct approach impacts
read performance since the new block is now not in the top
tier. However, in our experiences, it is very rarely that a
block gets read right after being written to storage. Also, if
the block gets read, say ten minutes later, by that time, it is
highly likely that the block has already been evicted to lower
level of the storage layer.
3.2 Read Optimization
To optimize read performance, we strive to keep as much
data in the frontend servers as possible, thus avoiding the
latency of fetching data blocks from remote backend storage
servers. We have collected over six months of text-based
query data from over 1,000 users. We discovered that,
without any optimization, we reached a hit rate of about 50%,
meaning that 50% of the queries are satisfied the the frontend
servers without hitting the remote backend storage servers,
while the other 50% end up hitting the remote servers, thus
leading to high latency.
An effective approach to improve this situation is to use
prefetching [2]. A simple improvement, when the frontend
servers are less busy, we prefetch data from the most
requested table into the frontend servers. This simple
improvement immediately boosts the hit rate to 70%. For
the next step, we break the day into 6 time periods, each is
four-hour long. For each time period, we prefetch the most
requested table from the same period in the previous day. By
doing this, we boost the hit rate to 80%.
This verifies that prefetching is indeed a very effective
technique in improving read latency. As we get more queries
for multimedia data in our storage, we plan to design more
fine-grained prefetching techniques to further improve the
performance of read operations. Some of the prefetching
strategies we plan to use include:
•
•
Label: since we now have label information associated
with each video stream, and we observe an initial trend
that people tend to search for certain labels, such as dog.
Prefetching the videos with the “hottest” label will be an
effective way to improve read performance.
Location: also, now that each video stream is associated
with location information, such as bedroom, living room,
etc., it will also be effective to prefetch the most
searched locations into the frontend storage servers. We
will explore these options in the next step.
4. PERFORMANCE AND SCALABILITY
In this section we delve into the details of the performance
and scalability of the implementation of the learn-memorizerecall robotic cloud computing paradigm. The machine
configuration used in this implementation consists of an Intel
Core-i7 CPU running at 3 GHz, and a Titan X GPU with 12
GB of graphics memory, and 32 GB of system memory.
4.1 Object Recognition Performance
First we look at the performance of the object recognition
engine. We stress the engine by feeding it video streams and
measure its performance. As shown in Figure 4, when under
stress, on a single server, it takes an average of 0.16 sec to
process an image. Also, this workload is GPU-bound, as it
uses 95% of GPU resources, but only 2~3% of CPU
resources and 3% of system memory.
Figure 4: Faster RCNN Detection Stress Test
In our real-world use case, the in-home service robots
move at a speed of about 30 centimeters per second.
Therefore, in most cases, we only need to extract labels from
an image every two seconds without missing an object in the
scene. This implies that with each server, we could support
10 simultaneous incoming video streams. This can be
further tuned based on the robot speed and user requirements.
4.2 Query Engine Performance
We now examine the performance of the query engine. We
stress the engine by launching 100 clients repeated send
queries to the query engine and we measure the response
time. As shown in Figure 5, when under stress, on a single
server, it takes on average of 4 ms to process a query. This
workload is CPU-bound, in that it uses 98% of CPU
resources, none of the GPU resources, and 2% of system
memory. This confirms that with one server, we could easily
handle over 100 users simultaneously.
we can write a video feed to storage. If the throughput is too
low, then the storage layer may become the bottleneck of the
whole multimedia data pipeline. As shown in Figure 6, with
Alluxio’s in-memory storage engine, we could easily
achieve greater than 650 MB/s throughput whereas with
Native File System, we could only achieve 120 MB/s. Using
remote HDFS, the performance is the worst, only sustaining
less than 20 MB/s throughput. This result indicates that for
the storage engine not to become the bottleneck of the
memorization part, it is important to keep most of the write
operations to hit the in-memory storage layer.
Figure 6: Alluxio Stress Test
Then we evaluate the video retrieval latency, using Alluxio’s
in-memory storage engine, we are able to retrieve a video
within 500 ms. However, when the video is stored in remote
machines, the latency can be as high as 20 seconds.
Therefore, using Alluxio to buffer “hot” video data could
reduce retrieval latencies by as much as 40 folds, which is
quite critical to the user experience, especially in the recall
stage.
4.4 Deployment and Scalability
After understanding the performance of the individual
components, we take a look at the deployment setup as well
as the scalability of this architecture.
Table 1: Resource Utilization
First, Table 1 summarizes the resource utilization of the
three components under stress. It is interesting that each
component stresses one type of system resource: the faster
R-CNN recognition engine stresses the GPU, MongoDB
stresses the CPU, while Alluxio stresses the system memory.
This indicates that it is best to co-locate these three services
on the same servers. This approach provides the following
benefits:
•
Figure 5: MongoDB Query Stress Test
4.3 Storage Performance
Then we examine the performance of the storage engine. We
first compare the throughput of a copy operation with
Alluxio, with the Native File System, and with Remote
HDFS. This parameter is critical as it determines how fast
•
Cost Efficiency: as we co-locate the services, we reduce
costs.
Performance: the storage layer, Alluxio, is now colocated with the same servers as the learning engine and
the query engine, which ensures high write throughput
and low read latency.
5.1 Data Reduction Techniques
Figure 7: Deployment Architecture
The deployment architecture is shown in Figure 7. It is
divided into two parts, the frontend servers and the backend
storage servers. The frontend servers process the incoming
data and write the raw data in Alluxio. Alluxio then
asynchronously persists the raw data in the backend storage
servers. Also, when the users generate new queries, the
frontend servers process the queries and retrieve the raw data
from Alluxio. If the requested data is buffered in Alluxio’s
tiered storage, it is immediately returned to the user.
Otherwise it is a cache miss and Alluxio requests the data
from the backend storage servers.
For Alluxio, we use a two-level deployment, allocating
20 GB of memory in the top tier and 200 GB of HDD in the
second tiered. With a 10-machine deployment, we provide
2.2 TB of cache space. Assuming 10 MB average video file
size, the system can buffer around 200,000 video files. Also,
with ten machines, we can simultaneously support 100 video
streams as well as 1,000 simultaneous queries. With this
setup, we are able to handle 10,000 users.
Next, we seek to answer the question as to whether we
could support one million users. The key to this question is
whether we could have a 1,000-machine Alluxio deployment
running stably. We conducted a stress test, with a 1,000instance Alluxio deployment and a background script to
repeatedly write to and read from the system. After running
the script for two weeks, we were able to confirm that
Alluxio could reliably scale to 1,000 instances. This
confirms that we can use Alluxio to support one million
users as our implementation scales.
Data reduction techniques, e.g., sampling, can be a feasible
solution to handle the above issue on robotic clouds,
especially when small errors are tolerable. For example,
when the data collected from robots is used to estimate
traffic volume/speed or to learn trajectory patterns, small
errors in the results are usually inevitable anyway due to the
noise in the initial data collection and in the learning stages.
For such tasks, the goal of our robotic cloud is to provide
data-analytics services with low storage cost and interactive
query response time at the cost of small errors in the
analytical results and answers.
To achieve this, we utilize techniques from approximate
query processing (AQP), which has been studied for decades
mostly in the context of relational databases. The goal of
AQP is to provide approximate answers to a subclass of SQL
queries with interactive response time and estimated error
bounds. This line of work has become very active recently
because of the explosion of data. The recent efforts result in
both innovative techniques and more mature sampling-based
systems, e.g., BlinkDB [13], QuickR [11], and Sample+Seek
[10].
Interested readers can refer to [12] for a brief survey of
the progresses. Sampling-based AQP techniques can be
characterized by where the sampler (for the data-reduction
purpose) is in the whole pipeline. In one way, the sampler
can be placed right before the data is streamed into the query
execution engine, or even pushed to the nodes of a query
plan tree with the execution engine (e.g., QuickR). In the
other way, samplers are used pre-computed samples of the
data tables, and SQL queries are rewritten to be executed on
these sample tables for fast response (e.g., BlinkDB and
Sample+Seek).
5.2 Usages in Robotic Clouds
In a robotic cloud, we have more options about where to
place the samplers and more considerations to be addressed
for each option. Following are our preliminary proposals.
•
5. NEXT STEP: DATA REDUCTION
As mentioned in the introduction, the volumes of data
collected from robotic devices can easily grow exponentially
over time, which on one hand, enables our robotic cloud to
provide data-analytics and decision-support services to users,
but on the other hand, imposes performance challenges on
all the three stages, learn-memorize-recall, in terms of both
storage and query response time. Although the architecture
presented in this paper is optimized for robotic workloads,
with a fixed and limited budget of computing resources, the
users eventually have to discard more and more data and
tolerate longer response time.
•
•
Sampling before learning. We can reduce the amount of
raw data collected from users by placing samplers in the
client devices or the streaming server. A tuple (e.g., an
image) in the raw data is selected into the object
recognition component only with some probability,
which is determined by some light-weight calculation
based on the tuple’s meta data, e.g., time and location.
Sampling before memorization. After the semantics
information is extracted from the raw data tuples, we
can pre-compute and store only a subsample into the
key-value store. For example, a row <sessionID,
timestamp, duration, location, <list of labels>> is
selected into the store with a probability determined by
the “list of labels”—we can select more sample rows for
some hot and more important labels and less sample
rows for the others.
Online sampling during recall. When a query is issued
in the recall stage, we can invoke a sampler to select
only a subset of rows from the key-value store into the
query execution engine (and of course, rewrite the query
to rescale the answer). For example, we may want to
count the number of times a label appears between a
time interval. A number of sampler can be borrowed
from AQP systems [12] to provide estimates of answers
and error bounds for such queries.
Note that we can either adopt all the three proposals at
the same time or only some of them, based on the different
types of underlying tasks and the amount of resource budget.
Finally, it is important to note that we always want to
guarantee that the errors are no more than some threshold in
our analytical results or answers, or at least, we want to
provide estimates of the errors. Such guarantees or
estimations are relatively easier for SQL-like aggregation
queries (e.g., counting the number of cars on a street), but
become more challenging for complex tasks (e.g., predicting
the trajectory of a car). We will start with adopting the
sampling-based data reduction techniques in our robotic
cloud to provide cheap and fast data-analytics services for
users with relatively simpler analytical workloads, and
extend them for more types of tasks in future.
6. CONCLUSIONS
We are facing the data explosion problem from robotic
applications.
Unlike existing mobile devices, robots
generate massive amount of unstructured data, which can not
be easily understood, stored, and retrieved by machines.
To address this problem, in this paper, we proposed a
learn-memorize-recall-reduce paradigm for robotic clouds:
the learning stage extracts semantic information from
incoming unstructured data and convert it into structured
data; the memorization stage provides effective storage for
the massive amount of data; the recall stage provides
efficient means to retrieve the raw data; and the reduction
stage provides means to making sense of massive amount of
unstructured data with limited computing resources.
7. REFERENCES
[1] S. Liu, J. Tang, C. Wang, Q. Wang, J-L. Gaudiot,
Implementing a Cloud Platform for Autonomous Driving.
arXiv preprint arXiv:1704.02696, 2017
[2] V. N. Padmanabhan and J. C. Mogul, Using predictive
prefetching to improve world wide web latency, Computer
Communications Review, vol. 26, pp. 22-36, July 1996.
[3] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards
real-time object detection with region proposal networks.
arXiv preprint arXiv:1506.01497, 2015.
[4] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R.
Girshick, S. Guadarrama, and T. Darrell. Caffe:
Convolutional architecture for fast feature embedding. arXiv
preprint arXiv:1408.5093, 2014
[5] K. Chodorow, MongoDB: the definitive guide. O’Reilly
Media Inc, 2013
[6] H. Li, A. Ghodsi, M. Zaharia, S. Shenker, and I. Stoica.
Reliable, Memory Speed Storage for Cluster Computing
Frameworks. In Proc. SoCC, 2014.
[7] Amazon S3, 2017. Amazon Simple Storage Service.
Available from http://aws.amazon.com/s3/
[8] D. Borthakur: HDFS architecture guide. HADOOP
APACHE PROJECT. 2008.
[9] S. Weil, S. Brandt, E. Miller, D. Long, and C. Maltzahn,
Ceph: A Scalable, High-Performance Distributed File
System, Proceedings of the 7th Conference on Operating
Systems Design and Implementation, November 2006.
[10] B. Ding, S. Huang, S. Chaudhuri, K. Chakrabarti, and C.
Wang, Sample + Seek: Approximating Aggregates with
Distribution Precision Guarantee, In Proc. SIGMOD, 2016.
[11] S. Kandula, A. Shanbhag, A. Vitorovic, M. Olma, R. Grandl,
S. Chaudhuri, and B. Ding, Quickr: Lazily Approximating
Complex Ad-Hoc Queries in Big Data Clusters, In Proc.
SIGMOD, 2016.
[12] S. Chaudhuri, B. Ding, and S. Kandula, Approximate Query
Processing: No Silver Bullet, In Proc. SIGMOD, 2017.
[13] S, Agarwal, B. Mozafari, A. Panda, H. Milner, S. Madden,
and I. Stoica, BlinkDB: Queries with Bounded Errors and
Bounded Response Times on Very Large Data, In Proc.
EuroSys, 2013.
Dr. Shaoshan Liu is currently the Co-Founder of PerceptIn,
working on developing the next-generation robotics platform.
Contact him at: [email protected]
Dr. Bolin Ding is currently a Researcher in the Data Management,
Exploration and Mining group at Microsoft Research, Redmond.
Contact him at: [email protected]
Dr. Jie Tang is the corresponding author and she is currently an
associate professor in the School of Computer Science and
Engineering of South China University of Technology,
Guangzhou, China. Contact her at: [email protected]
Dawei Sun is currently with Tsinghua University and PerceptIn,
working on Deep Learning and cloud infrastructures and
autonomous robots. Contact him at: [email protected]
Dr. Zhe Zhang is the co-founder of PerceptIn, working on
developing the next-generation robotics platform. Contact him at:
[email protected]
Dr. Grace Tsai is a founding engineer of PerceptIn, working on
developing the next-generation robotics platform. Contact her at:
[email protected]
Dr. Jean-Luc Gaudiot is professor in the Electrical Engineering and
Computer Science Department at the University of California,
Irvine and is currently serving as the 2017 President of the IEEE
Computer Society. Contact him at [email protected]
| 2cs.AI
|
UNIFORM BOUNDS IN F-FINITE RINGS AND LOWER
SEMI-CONTINUITY OF THE F-SIGNATURE
arXiv:1506.01073v2 [math.AC] 11 Dec 2015
THOMAS POLSTRA
Abstract. This paper establishes uniform bounds in characteristic p rings which are either
F-finite or essentially of finite type over an excellent local ring. These uniform bounds are
then used to show that the Hilbert-Kunz length functions and the normalized Frobenius
splitting numbers defined on the Spectrum of a ring converge uniformly to their limits,
namely the Hilbert-Kunz multiplicity function and the F-signature function. From this
we establish that the F-signature function is lower semi-continuous. Lower semi-continuity
of the F-signature of a pair is also established. We also give a new proof of the upper
semi-continuity of Hilbert-Kunz multiplicity, which was originally proven by Ilya Smirnov.
1. Introduction
Throughout this paper all rings are assumed to be commutative, Noetherian, with identity,
and of prime characteristic p. We shall reserve q to denote a power of p, i.e., q = pe for some
nonnegative integer e, and λ(−) denotes the standard length function.
If (R, m) is a local ring of dimension d, M a finitely generated R-module, and I an m1
[q]
M).
primary ideal of R, then the qth Hilbert-Kunz length of M at I is given by qdim(M
) λ(M/I
The Hilbert-Kunz multiplicity of M at I is defined by
1
M
eHK (I, M) := lim dim(M ) λ
.
q→∞ q
I [q] M
Paul Monsky showed this limit always exists in [13].
We say that a ring R is F-finite if the Frobenius endomorphism F : R → R which maps
r 7→ r p makes R a finite R-module. This is equivalent to R being module finite over F e (R)
for all e ≥ 1, and when R is reduced this is equivalent to R1/q being module finite over R
for all q. If R is an F-finite ring, then so is any finitely generated algebra and localization
over R. If (R, m, k) is local then we let α(R) = logp [k 1/p : k]. If R is not necessarily local
and P ∈ Spec (R), then we let α(P ) = α(RP ).
In this paper we will be interested in uniform properties of Hilbert-Kunz functions over
F-finite rings and rings essentially of finite type over an excellent local ring. For the sake
of simplicity, assume that I = m and M = R. Then not only does q1d λ(R/m[q] ) converge to
eHK (R) := eHK (m, R), it is the case that λ(R/m[q] R) = eHK (R)q d + O(q d−1 ), hence there
exists a C > 0 such that for all q, λ(R/m[q] R) ≤ Cq d . Now suppose that R is a not necessarily
local characteristic p ring. Then for each P ∈ Spec (R) there exits a constant C > 0 such
that for each q, λ(RP /P [q]RP ) ≤ Cq ht (P ) . This particular result is not very interesting
since the constant C depends upon P and is easily obtained from well known results about
Hilbert-Kunz length functions. Recently, I. Smirnov showed, if R is an excellent ring of
2010 Mathematics Subject Classification. 13A35, 13D40, 13F40, 14B05
1
characteristic p, that for each P ∈ Spec (R) there exists a constant C and element s ∈ R − P
such that for all Q ∈ D(s) ∩ V (P ), λRQ (RQ /Q[q] RQ ) ≤ Cq ht (Q) (see Lemma 14 in [14]). We
significantly improve this result in Proposition 3.3 and Proposition 4.3 for rings which are
either F-finite or are essentially of finite type over an excellent local ring, both of which are
large classes of excellent rings. A consequence of Proposition 3.3 and Proposition 4.3 is that,
if R is F-finite or essentially of finite type over an excellent local ring, then there exists a
constant C such that for all P ∈ Spec (R), λ(RP /P [q] RP ) ≤ Cq ht (P ) .
A map of primary ideals is a map I(−) : Spec (R) → {Ideals of R} such that for each
P ∈ Spec (R), I(P )RP is a P RP -primary. If M is a finitely generated R-module then we
shall denote by λM
q1 (I(−)), or simply λq1 (I(−)), if M is understood, to be the function from
the support of M, which we denote Supp (M), to the real numbers R, which maps a prime
1
P 7→ dim(M
λ(MP /I(P )[q1] MP ). We denote by eHK (I(−), M− ) the function which sends a
P)
q1
prime P ∈ Supp (M) to eHK (I(P ), MP ).
Let R, M, and I(−) be as above. Then it is easy to see that λq1 (I(−)) converges pointwise
to eHK (I(−), M− ) as q1 → ∞. Theorem 5.1 states that if R is an F-finite ring or essentially of
finite type over an excellent local ring, M is a finitely generated R-module, and I(−) a map of
M
M
primary ideals, then λM
q1 (I(−))/λ1 (I(−)) converges uniformly to eHK (I(−), M− )/λ1 (I(−))
as q1 → ∞. In particular, if M = R and I(P ) = P , then the qth Hilbert-Kunz function,
which sends a prime P 7→ λ(RP /P [q]RP ), converges uniformly to the Hilbert-Kunz multiplicity function, which sends a prime P 7→ eHK (RP ). In order to prove this we will need to
establish some uniform bounds in F-finite rings and in rings essentially of finite type over
an excellent local ring. Some of the uniform bounds established in Section 3 and Section 4
of this paper are related to, but are often improvements of, uniform bounds in Section 3 of
[14], which establishes the upper semi-continuity of Hilbert-Kunz multiplicity, and Section 3
of [15], which shows that the F-signature of a local ring exists.
I. Smirnov has recently shown in [14] that eHK (R− ) is upper semi-continuous on locally
equidimensional rings which are F-finite or essentially of finite type over an excellent local ring. In showing that the Hilbert-Kunz multiplicity function is the uniform limit of
upper semi-continuous functions on such rings, we easily recover Smirnov’s result. It is
still unknown if Hilbert-Kunz multiplicity is upper semi-continuous on an excellent locally
equidimensional ring.
Another interesting invariant defined on a local ring (R, m) of characteristic p is the Fsignature of R, defined originally in [10], by Huneke and Leuschke. For any module M and
any q = pe , we can view M as an R-module via restriction of scalars under F e which we
denote by F∗q M. In particular if R is F-finite and M = R, then F∗q R is a module finite Rmodule and we can write F∗q R ≃ Raq ⊕ Mq where Mq has no free R-summand. The number
aq
aq is called the qth Frobenius splitting number of R. We denote by bq = qα(R)
. The number
aq
sq := qα(R)+dim(R) is called the qth normalized Frobenius splitting number of R. Yao showed
there is a way to measure bq , in way that is well defined even for rings which are not F-finite,
hence one can define the qth normalized Frobenius splitting number for a ring which is not
F-finite ([16], Lemma 2.1). Huneke and Leuschke defined the F-signature of a local ring of
dimension d to be the limit limq→ qbqd , provided the limit exists. Kevin Tucker showed in [15]
that the F-signature of a local ring always exists.
2
If R is an F-finite ring, which is not necessarily local, then define the qth Frobenius splitting
number function aq : Spec (R) → R by letting aq (P ) be the qth Frobenius splitting number of
the local ring RP . If R is any characteristic p ring, then define the qth normalized Frobenius
splitting number function sq : Spec (R) → R by letting sq (P ) be the qth normalized Frobenius
splitting number of the local ring RP . We let bq (P ) = q ht (P ) sq (P ) and we let s : Spec (R) → R
be the F-signature function which sends a prime P 7→ s(RP ), the F-signature of the local
ring RP .
The problem of whether the F-signature function is a lower semi-continuous function with
respect to the Zariski topology has been of interest for quite some time. Recall that a
function f : X → R, where X is a topological space, is lower semi-continuous at x ∈ X
if for all ǫ > 0, there is an open neighborhood U of x such that f (x) − f (y) < ǫ for all
y ∈ U. In other words, a function f is lower semi-continuous at x if in a small enough open
neighborhood of x the numbers f (y) as y varies in the open neighborhood of x can only be
slightly smaller than f (x). We would like to briefly explain why it has been suspected that
the F-signature function should satisfy this property.
The F-signature detects subtle information about the severity of the singularity of a local
ring. Given a local ring (R, m), it always the case that 0 ≤ s(R) ≤ 1. Huneke and Leuschke
showed in [10] that s(R) = 1 if and only if R is a regular local ring. Aberbach and Leuschke
showed in [1] that s(R) > 0 if and only if R is strongly F-regular. Heuristically, the closer to
1 the F-signature of R is the "nicer" the singularity is, and the closer to 0 the "worse" the
singularity is. One expects that given a ring or a scheme with decent geometric properties,
that the severity of a singularity of a point is controlled in an open neighborhood of that point.
This is exactly what we should expect the F-signature function defined on the spectrum of
a "decent" ring, e.g. and excellent domain, to do. Given a prime P in the spectrum of
such a ring, we expect that in a small enough open neighborhood of the prime, that the
singularities found in that open neighborhood are not too much worse then the singularity
associated with P . Thus we should expect than in a small enough open neighborhood U of
P that s(Q) is at most ǫ closer to 0, which is precisely what lower semi-continuity of the
F-signature would say.
Another reason to expect the F-signature function to be lower semi-continuous is that
Enescu and Yao showed in [5] that under mild conditions, the qth normalized Frobenius
splitting number function is a lower semicontinuous functions. For example, they showed
that if R is a domain which is either F-finite or essentially of finite type over an excellent
local ring, then the qth normalized Frobenius number function is lower semi-continuous.
So after Kevin Tucker showed the F-signature to always exist, it has been known that the
F-signature function naturally arrises as the limit of lower semi-continuous functions.
Some light has been previously shed on the lower semi-continuity of the F-signature problem. Blickle, Schwede, and Tucker showed that if R is a regular and not necessarily local
F-finite ring with 0 6= f ∈ R and t ≥ 0, then the function Spec (R) → R defined by
P 7→ s(RP , f t ) is lower semi-continuous. See [3] for more details. The results in section 6 in
this paper recapture Blickle, Schewede, and Tucker’s result.
Theorem 5.6 proves that if R is either F-finite or is essentially of finite type over an
excellent local ring, then the qth normalized Frobenius splitting number functions converge
uniformly to the F-signature function as q → ∞. It will then follow by Enescu’s and Yao’s
3
work, [5], that the F-signature function will be lower semi-continuous on all such rings. Kevin
Tucker has independently found, and discussed with the author, an alternative proof of the
lower semi-continuity of the F-signature.
The paper is organized as follows. In section 2 we establish some preliminary results. In
particular, section 2 contains a generalized version of a Lemma of Sankar Dutta which is
crucial to the bounds given in section 3. Section 3 establishes uniform bounds of Hilbert-Kunz
length functions in F-finite rings. Section 3 is the most difficult section to work through,
but the bounds that are established lead to a proof that the F-signature function is lower
semi-continuous on F-finite rings and rings essentially of finite type over an excellent local
ring. Section 4 establishes the bounds in 3 for rings which are essentially of finite type over
an excellent local ring. In Section 5 we apply the results of section 3 and 4 to establish the
uniform convergence of Hilbert-Kunz length functions and normalized Frobenius splitting
numbers to their limits. In Section 6 lower semi-continuity of the F-signature of a pair,
(R, D), is established for all Cartier subalgebras D on an F-finite ring R.
2. Preliminary Results
If R is an F-finite ring which is locally equidimensional, then it was originally shown
by E. Kunz in [11] that the function Spec (R) → R which sends P 7→ α(P ) + ht (P ) is
constant on connected components of Spec (R). In particular, if R is an F-finite domain
then α(P ) + ht (P ) is constant on Spec (R). If R is an F-finite domain then we let γ(R) be
the constant α(P ) + ht (P ).
We will need a global version of a Lemma, first proved by Sankar Dutta, in order to
establish the uniform bounds found in Section 3 and 4. In [4], Sankar Dutta showed that
if (R, m) is an F-finite local domain of dimension d then there exists a finite set of nonzero
primes S(R) and a constant C such that for all q = pe there is a containment of R-modules
γ(R)
Rq
⊆ R1/q which has a prime filtration whose prime factors are isomorphic to R/P , where
P ∈ S(R), and such a prime factor appears no more than Cq γ(R) times in the filtration. In
γ(R)
particular, the length of the prime filtration of Rq
⊆ R1/q has length no more than
C|S(R)|q γ(R) . This result, for local domains whose residue field is perfect, is exercise 10.4 in
[9], whose proof is given in the second appendix by Karen Smith, and this result is explicitly
stated and proved in [8] as Lemma 4.
Remark 2.1. If R is an F-finite domain and P ∈ Spec (R) a nonzero prime, then γ(R/P ) =
1/p
logp [ PRRPP
: PRRPP ] = α(P ) < α(P ) + ht (P ) = γ(R).
Lemma 2.2. Let R be an F-finite domain. Then there exists a finite set of nonzero primes
S(R), and a constant C, such that for every q = pe ,
γ(R)
(1) there is a containment of R-modules Rq
⊆ R1/q ,
(2) which has a prime filtration whose prime factors are isomorphic to R/P , where P ∈
S(R),
(3) and for each P ∈ S(R), the prime factor R/P appears no more than Cq γ(R) times in
γ(R)
the prime filtration of the containment Rq
⊆ R1/q .
Proof. We shall prove the statement by induction on the Krull dimension of R. If the
dimension R is 0, then R is a field and the Lemma is trivial.
4
Now suppose that dim(R) > 0. Then R1/p is a torsion-free R-module of rank pγ(R) .
γ(R)
Hence there is an injection of R-modules Rp
⊆ R1/p so that the support of the cokernel
γ(R)
γ(R)
R1/p /Rp
consists of nonzero primes. Therefore Rp
⊆ R1/p has a prime filtration of the
following form with the quotients Mi /Mi−1 = R/Pi where Pi is a nonminimal prime of R,
Rp
γ(R)
= M0 ⊆ M1 ⊆ · · · ⊆ Mh = R1/p .
The quotients R/Pi are F-finite domains of smaller Krull dimension than R, and so we may
assume by induction that the resultPholds for each R/P
S i , with finite collection of primes
S(R/Pi ) and constant Ci . Let C ′ = Ci and S(R) = (S(R/Pi ) ∪ {Pi }). Observe that the
γ(R)
above filtration shows that Rp
⊆ R1/p has a prime filtration consisting of no more than
C ′ quotients isomorphic to R/P for each P ∈ S(R) and all prime factors are isomorphic
γ(R)
to R/P for some P ∈ S(R). We shall show by induction that Rq
⊆ R1/q has a prime
filtration whose prime factors are isomorphic to R/P with P ∈ S(R) with no more than
C ′ q γ(R) (1 + p1 + p12 + · · · + 1q ) quotients isomorphic to R/P for each P ∈ S(R).
γ(R)
Now suppose that Rq
= N0 ⊆ N1 ⊆ · · · ⊆ Nm = R1/q is a prime filtration of
γ(R)
Rq
⊆ R1/q whose prime factors are isomorphic to R/P with P ∈ S(R) with no more
than C ′ q γ(R) (1 + 1p + · · · + 1q ) quotients isomorphic to R/P for each P ∈ S(R). Take qth
roots of the modules in the filtration Rp
(R1/q )p
γ(R)
1/q
1/q
= M0
γ(R)
⊆ R1/p to get the following new filtration,
1/q
⊆ M1
1/q
⊆ · · · ⊆ Mh
= R1/pq .
1/q
Each of the quotients Mi /Mi+1 = (R/Pi )1/q . By induction there exists a prime filtration
1/q
1/q
of Mi−1 ⊆ Mi with precisely q γ(R) prime factors isomorphic to R/Pi and each other prime
factor is isomorphic to R/P for some P ∈ S(R/Pi ) and such a prime factor appears no more
γ(R)
than Ci q γ(R/Pi ) times in the filtration. Furthermore, the prime filtration Rq
= N0 ⊆ N1 ⊆
1/q
q γ(R) pγ(R)
(pq)γ(R)
1/q pγ(R)
· · · ⊆ Nm = R gives the following filtration of (R
)
=R
⊆ (R )
,
γ(R)
R(pq)
γ(R)
= N0p
γ(R)
⊆ N1p
γ(R)
γ(R)
p
⊆ · · · ⊆ Nm
= (R1/q )p
γ(R)
.
γ(R)
Hence R(pq)
⊆ (R1/q )p
has a prime filtration with prime factors isomorphic to R/P
with P ∈ S(R) and such a prime factor appears no more than C ′ (pq)γ(R) (1 + p1 + · · · + 1q )
times in the filtration.
γ(R)
Putting the above information together we get that there is an embedding of R(pq)
⊆
1/pq
R
with prime filtration whose prime factors are isomorphic to R/P with P ∈ S(R)
and there are no more than the following number of quotients isomorphic to R/P for each
P ∈ S(R):
′
C (pq)
γ(R)
h
X
1
1
+
Ci q γ(R/Pi ) .
1+ +···+
p
q
i=1
5
By Remark 2.1 we know that each γ(R/Pi ) ≤ γ(R) − 1, and so we have the following
estimates,
′
C (pq)
γ(R)
1
1
1+ +···+
p
q
+
=
=
≤
1
q
Ci q γ(R/Pi )
i=1
≤
Each of the sums 1 + 1p + · · · +
h
X
X
h
1
1
+
Ci q γ(R)−1
1+ +···+
C (pq)
p
q
i=1
1
1
C ′ (pq)γ(R) 1 + + · · · +
+ C ′ q γ(R)−1
p
q
1
1
1
′
γ(R)
1 + + · · · + γ(R)
C (pq)
p
q p
q
1
1
1
′
γ(R)
1+ +···+ +
C (pq)
.
p
q pq
γ(R)
′
≤ 1 + 21 + · · · +
q γ(R)
1/q
1
2e
≤ 2. It now follows by induction that
for every q, that the containment R
⊆ R will have a prime filtration whose factors are
isomorphic to R/P for some P ∈ S(R) with no more than C ′ (1 + p1 + · · · + 1q )q γ(R) ≤ 2C ′ q γ(R)
quotients isomorphic to R/P for each P ∈ S(R).
We will find it useful in Section 3 to have a version of Dutta’s Lemma with the inclusion
γ(R)
of Rq
⊆ R1/q reversed.
Corollary 2.3. Let R be an F-finite domain. Then there exists a finite set of nonzero primes
S(R), and a constant C, such that for every q = pe ,
γ(R)
(1) there is a containment of R-modules R1/q ⊆ Rq ,
(2) which has a prime filtration whose prime factors are isomorphic to R/P , where P ∈
S(R),
(3) and for each P ∈ S(R), the prime factor R/P appears no more than Cq γ(R) times in
γ(R)
the prime filtration of the containment R1/q ⊆ Rq .
γ(R)
Proof. Since R1/p is torsion-free of rank pγ(R) , there is an injection of R-modules R1/p ⊆ Rp
γ(R)
so that the support of the cokernel Rp /R1/p consists of nonzero primes. Therefore there
γ(R)
is prime filtration R1/p = M0 ⊆ M1 ⊆ · · · ⊆ Mh = Rp
with Mi /Mi−1 ≃ R/Pi with Pi
a nonzero prime ideal. By Remark 2.1 γ(R/Pi ) < γ(R). We let S(R/Pi ) and constant Ci
be the collection of primes and constant as described
the F-finite domain
P in Lemma 2.2 for S
′
R/Pi . As in the proof of Lemma 2.2 we let C =
Ci and S(M) = (S(R/Pi ) ∪ {Pi }).
γ(R)
Furthermore, once again as in Lemma 2.2, we can show by induction that R1/q ⊆ Rq
has
a prime filtration whose prime factors are isomorphic to R/P with P ∈ S(M) with no more
than C ′ q γ(R) (1 + p1 + p12 + · · · + 1q ) quotients isomorphic to R/P for each P ∈ S(R). The
above filtration of R1/p ⊆ Rp
γ(R)
shows the induction step when q = p.
6
γ(R)
Now suppose that R1/q = N0 ⊆ N1 ⊆ · · · ⊆ Nm = Rq
is a prime filtration of R1/q ⊆
γ(R)
Rq
such that each Nj /Ni−j ≃ R/Pj for some Pj ∈ S(R) and such a prime factor appears
γ(R)
no more than C ′ q γ(R) (1 + 1p + p12 + · · · + 1q ) times in the filtration. Therefore (Rp )1/q =
γ(R)
γ(R)
γ(R)
γ(R)
(R1/q )p
⊆ (Rq )p
= R(qp)
has a prime filtration with prime factors R/Pj with
Pj ∈ S(R) and such a prime factor appears no mare than C ′ (qp)γ(R) (1+ p1 + p12 +· · ·+ 1q ) times
in the filtration. Furthermore, the prime filtration R1/p = M0 ⊆ M1 ⊆ · · · ⊆ Mn = Rp
γ(R)
gives the following filtration of R1/pq = (R1/p )1/q ⊆ (Rp )1/q ,
1/q
(R1/p )1/q = M0
1/q
1/q
⊆ M1
⊆ · · · ⊆ Mn1/q = (Rp
γ(R)
γ(R)
)1/q .
1/q
Since Mi /Mi−1 ≃ (R/Pi )1/q , we apply Lemma 2.2 to know there is a prime filtration of
1/q
1/q
each Mi−1 ⊆ Mi whose prime factors come from S(R/Pi ) and such a prime factor appears
no more than Ci q γ(R/Pi ) ≤ C ′ q γ(R)−1 times in the filtration. Putting all of this information
γ(R)
together we get an embedding R1/pq ⊆ R(pq)
with a prime filtration whose prime factors
come from S(R) and such a prime factor appears no more than the following number in the
filtration,
X
h
1
1
1
1
γ(R) ′ γ(R)
1+ +···+
p
Cq
+
Ci q γ(R/Pi ) ≤ C ′ (qp)γ(R) (1 + + · · ·+ ) ≤ 2C ′ (qp)γ(R) .
p
q
p
pq
i=1
We combine of Lemma 2.2 and Corollary 2.3 into a single statement for convenience.
Corollary 2.4. Let R be an F-finite domain. There exists a finite set of nonzero primes
S(R) and a constant C such that for every q = pe , there is a containment of R-modules
γ(R)
γ(R)
R1/q ⊆ Rq
and Rq
⊆ R1/q which each has a prime filtration whose prime factors are
isomorphic to R/P , where P ∈ S(R), and such a prime factor appears no more than Cq γ(R)
times in the filtration.
We shall need the following two well know lemmas, whose proofs are given for the sake of
completion.
Lemma 2.5. Let (R, m, k) be an F-finite reduced local ring and let I be an m-primary ideal.
Then λ(R1/q /IR1/q ) = q α(R) λ(R/I [q] R).
Proof. Consider a prime filtration of I [q] R ⊆ R, say it is given by I [q] R = M0 ⊆ M1 ⊆
· · · ⊆ Mn = R with each Mi /Mi−1 ≃ k. Then by taking qth roots we get a filtration
1/q
1/q
1/q
1/q
1/q
IR1/q = M0 ⊆ M1 ⊆ · · · ⊆ Mn = R1/q with each quotient Mi /Mi−1 ≃ k 1/q . It follows
that λ(R1/q /IR1/q ) = q α(R) λ(R/I [q] R).
Lemma 2.6. Let (R, m, k) be a local characteristic p ring, I be an m-primary ideal, and M
a finitely generated R-module. Then
lim
1
q2 →∞ q dim(M )
2
dim(M )
λ(M/I [q1q2 ] M) = q1
7
eHK (I, M).
Proof. We only have to observe that
lim
q2 →∞
dim(M )
1
dim(M )
q2
λ(M/I
[q1 q2 ]
q1
M) = lim
λ(M/I [q1 q2 ] M)
q2 →∞ (q1 q2 )dim(M )
dim(M )
= q1
dim(M )
= q1
lim
q→∞
1
q dim(M )
λ(M/I [q] M)
eHK (I, M).
3. Uniform Bounds in F-Finite Rings
The goal of this section is to establish uniform bounds in not necessarily local F-finite
rings. The purpose of establishing these uniform bounds is to better understand the global
behavior of relative Hilbert-Kunz length functions, which can then be used to establish the
lower semi-continuity of the F-signature.
Remark 3.1. If (R, m) is a local ring, an m-primary pair of ideals will be a containment of
ideals of R, I ⊆ J, such that I is m-primary. Observe that either J is also m-primary or is
R itself. If I ⊆ J is an m-primary pair of ideals, then there is an ascending chain of ideals
I ⊆ (I, u1 ) ⊆ (I, u1 , u2) ⊆ · · · ⊆ (I, u1 , u2, ..., uλ(J/I) ) = J where ui+1 ∈ (I, u1 , ..., ui ) : m. We
shall let I0 = I and Ii = (I, u1 , ..., ui ) for 1 ≤ i ≤ λ(J/I).
Lemma 3.2. Let (R, m) be a local ring of characteristic p and M a finitely generated Rmodule. If I ⊆ J is an m-primary pair of ideals, then λ(J [q] M/I [q] M) ≤ λ(M/m[q] M)λ(J/I).
Pλ(J/I)
[q]
[q]
Proof. Observe that λ(J [q] M/I [q] M) = i=1
λ(Ii M/Ii−1 M), hence it is enough to show
that if I is m-primary and u ∈ (I : m), then λ((I, u)[q]M/I [q] M) ≤ λ(M/m[q] M). Well,
(I, u)[q]M/I [q] M ≃ M/(I [q] M :M uq ). Since u ∈ I : m we have that m[q] M ⊆ (I [q] M :M uq ),
hence λ(M/(I [q] M :M uq )) ≤ λ(M/m[q] M).
Proposition 3.3. Let R be an F-finite ring and M a finitely generated R-module. There
exists a constant C > 0 such that for all P ∈ Spec (R) and q = pe , if IRP ⊆ JRP is a
P RP -primary pair of ideals, then
[q]
JRP
J MP
dim(MP )
≤ Cq
λ
.
λ
I [q] MP
IRP
Proof. By Lemma 3.2 we only need to find a constant C such that for all P ∈ Spec (R) and
all q, λ(MP /P [q]MP ) ≤ Cq dim(MP ) . If M0 ⊆ M1P
⊆ · · · ⊆ Mn = M is a prime filtration of M
with Mi /Mi−1 ≃ R/Pi , then λ(MP /P [q]MP ) ≤ ni=1 λ(RP /(Pi + P [q])RP ). This reduces the
Proposition to showing that if R is an F-finite domain, then there is a constant C such that
for all P ∈ Spec (R), λ(RP /P [q] RP ) ≤ Cq ht (P ) .
Suppose that R is an F-finite domain. Let S(R) be the finite set of primes given by
Proposition 2.2 for the R-module R, and suppose that P ∈ Spec (R). By Lemma 2.5,
8
1/q
1/q
1/q
1/q
1
λ(RP /P RP ), so it is equivalent to show that λ(RP /P RP ) ≤
λ(RP /P [q]RP ) = qα(R
P)
γ(R)
Cq
for some C that does not depend on P . From the short exact sequence,
0 → Rq
γ(R)
→ R1/q → R1/q /Rq
γ(R)
→ 0,
we have that,
1/q
γ(R)
1/q
λ(RP /P RP ) ≤ λ(RPq
γ(R)
/P RPq
γ(R)
) + λ(R1/q /(RPq
γ(R)
= q γ(R) + λ(R1/q /(RPq
1/q
+ P RP ))
1/q
+ P RP )).
Therefore we only need to find a constant C, independent of P , such that
γ(R)
1/q
λ(R1/q /(RPq
+ P RP )) ≤ Cq γ(R) .
Before localizing at P we can apply Proposition 2.2 to know that there exists a filtration
γ(R)
γ(R)
of Rq
⊆ R1/q , say Rq
= N0 ⊆ N1 ⊆ · · · ⊆ Nn = R1/q , such that n ≤ C ′ |S(R)|q γ(R) ,
where C ′ is completely independent of P , and each Ni /Ni−1 ≃ R/Pi for some Pi ∈ S(R).
1/q
For convenience let Mi = (Ni )P . Localizing at P and adding P RP to each module Mi gives
γ(R)
1/q
1/q
1/q
1/q
+ P RP ⊆ RP whose factors are (Mi + P RP )/(Mi−1 + P RP ) ≃
a filtration of RPq
1/q
1/q
Mi /(Mi−1 + (P RP ∩ Mi )). Noticing that P RP ∩ Mi ⊇ P Mi we get that
1/q
1/q
1/q
λ (Mi + P RP )/(Mi−1 + P RP ) = λ Mi /(Mi−1 + (P RP ∩ Mi ))
≤ λ (Mi /(Mi−1 + P Mi ))
= λ(RP /(Pi RP + P RP ))
1/q
≤ λ(RP /P RP ) = 1.
1/q
It now follows that λ(RP /P RP ) ≤ (1 + C ′ |S(R)|)q γ(R) .
Corollary 3.4. Let R be an F-finite ring, N, M two finitely generated R-modules which are
isomorphic at minimal primes of R. Then there is a constant C such that for all P ∈ Spec (R)
and q = pe , if IRP ⊆ JRP is a P RP -primary pair of ideals, then
[q]
[q]
J MP
J NP
JRP
ht (P )−1
λ
−λ
.
≤ Cq
λ
I [q] MP
I [q] NP
IRP
Proof. Using the notation in Remark 3.1 and applying the triangle inequality
!
!
[q]
[q]
λ(J/I)
[q]
[q]
X
J NP
Ii MP
J MP
Ii NP
λ
−λ
≤
λ
−λ
.
[q]
[q]
I [q] MP
I [q] NP
Ii−1 MP
Ii−1 NP
i=1
Thus we may reduce ourselves to the scenario that J = (I, u) where u ∈ (I : P ). There
ϕ
ψ
are exact sequences M −
→ N → T1 → 0 and N −
→ M → T2 → 0, for which T1 , T2 are 0
when localized at minimal primes of R. Observe that ϕ(I [q] M :M uq ) ⊆ (I [q] N :N uq ) so that
9
there is induced map
MP
(I [q] MP :MP uq )
→
NP
(I [q] NP :NP uq )
whose cokernel, say (T1′ )P is naturally the
homomorphic image of (T1 )P . Thus we have the following commutative diagram.
ϕ
MP
NP
π1
NP
(I [q] NP :NP uq )
MP
(I [q] MP :MP uq )
NP
q
P :NP u )
that π1 (P [q]NP ) =
(T1 )P
of P [q]
. Thus
(T1 )P
Therefore λ
(I [q] N
−λ
MP
q
P : MP u )
(I [q] M
(T1 )P
0
π2
(T1′ )P
0
≤ λ((T1′ )P ). Observe that P [q] NP ⊆ (I [q] NP :NP
uq ) so
0 and therefore π2 (P [q] (T1 )P ) = 0. Hence (T1′ )P is the homomorphic
image
(I, u)[q]NP
(I, u)[q]MP
− λ
λ
I [q] MP
I [q] NP
NP
MP
(T1 )P
= λ
−λ
≤λ
.
(I [q] NP :NP uq )
(I [q] MP :MP uq )
P [q] (T1 )P
A similar argument applied to the exact sequence N → M → T2 → 0 implies that
[q]
[q]
(Ti )P
J MP
J NP
≤ max λ
λ
−λ
.
i=1,2
I [q] MP
I [q] NP
P [q] (Ti )P
The Corollary now follows by Proposition 3.3.
Corollary 3.5. Let R be an F-finite ring and 0 → M ′ → M → M ′′ → 0 a short exact
sequence of finitely generated R-modules. There exists a constant C such that for all P ∈
Spec (R) and q = pe , if IRP ⊆ JRP is a P RP -primary pair of ideals, then
[q] ′
[q] ′′
[q]
J MP
J MP
JRP
J MP
dim(MP )−1
−λ
.
−λ
≤ Cq
λ
λ
I [q] MP
I [q] MP′
I [q] MP′′
IRP
JRP
R M )RP
Proof. Observe that λ (J+Ann
≤
λ
. Therefore we can begin by replacing R
(I+Ann R M )RP
IRP
with R/Ann R M so
Lthat ht (P ) = dim MP for all P ∈ Spec (R). If R is reduced then M is
isomorphic to M ′ M ′′ at minimal primes of R and we can apply Corollary 3.4. Suppose
R is not reduced. Using a standard argument, we can reduce to the scenario that R is
reduced. See for example the proofs of Lemma 1.5 in [13] and Proposition 3.11 in [8]. Let
√ [q0 ]
= 0. Let F : R → R be the
e0 be a large enough integer so that for q0 = pe0 , 0
√
Frobenius endomorphism. Then F e0 (R) is abstractly isomorphic to the reduced ring R/ 0
and R is module finite over F e0 (R). Then for all P ∈ Spec (R) and IRP ⊆ P RP which is
P RP -primary,
MP
MP
MP
1
λ e
= λRP
= λRP
.
α(P ) F 0 (RP )
(I [q0] ∩ F e0 (R))[q] MP
(I [q0] ∩ F e0 (R))[q] MP
I [qq0] MP
q0
10
Theorem 3.6. Let R be an F-finite ring and M a finitely generated R-module. There exists
a constant C such that for all P ∈ Spec (R), for all q1 , q2 , if IRP ⊆ JRP is a P RP -primary
pair of ideals, then
[q1 ]
[q1 q2 ]
J MP
J
MP
JRP
ht (P )
dim(MP ) dim(MP )−1
λ
q2
−λ
.
≤ Cq2
q1
λ
I [q1 ] MP
I [q1 q2 ] MP
IRP
Proof. As in the proof of Corollary 3.5, we may replace R by R/Ann R M so that dim(MP ) =
ht P for all P ∈ Spec (R). If there is a short exact sequence 0 → M ′ → M → M ′′ → 0, then
λ
J [q1 ] MP
I [q1] MP
ht (P )
q2
−λ
J [q1 q2 ] MP
I [q1q2 ] MP
≤ A1 + A2 + A3 + A4 .
Where
A1 =
A2 =
A3 =
A4 =
L ′′
J [q1 ] (MP′
M )
ht (P )
L P′′
λ
−λ
q2
′
[q
]
1
I (MP
MP )
[q1 q2 ]
[q1 q2 ] ′ L ′′
J
MP
J
(MP
M )
L P′′
λ
−λ
′
[q
q
]
[q
q
]
1
2
1
2
I
MP
I
(MP
MP )
[q1 ] ′
[q1 q2 ] ′
J MP
J
MP
ht (P )
λ
q2
−λ
′
[q
]
[q
q
]
1
1
2
I MP
I
MP′
[q1 q2 ] ′′
[q1 ] ′′
J
MP
J MP
ht (P )
q2
−λ
.
λ
′′
[q
]
[q
q
]
1
1
2
I MP
I
MP′′
ht (P )−1
ht (P )
JRP
a constant C such that A1 ≤ Cq1
λ IRP q2
and A2 ≤
J [q1 ] MP
I [q1 ] MP
By Corollary 3.5 there is
P
. Therefore by considering a prime filtration of the module M, we can
C(q2 q1 )ht (P )−1 λ JR
IRP
reduce proving the theorem to the scenario that M = R/P for some prime P ∈ Spec (R),
i.e., we may assume that
M1/q= R is an F-finite domain. Observe that by Lemma 2.5,
[q q ]
J [q1 ] RP 2
1 2
1
λ JI [q1 q2 ] RRP = α(P
. Therefore the theorem is now reduced to showing that
)λ
[q1 ] 1/q2
P
q2
I
RP
there is a constant C independent of P, I, J, q1 , q2 such that
!
[q1 ]
[q1 ] 1/q2
J
R
J RP
JRP
γ(R)
γ(R) ht (P )−1
P
q2 − λ
.
λ
≤ Cq2 q1
λ
1/q
I [q1 ] RP
IRP
I [q1 ] RP 2
As in the proof of Corollary 3.4 we can further reduce to the scenario that J = (I, u) where
u ∈ (I : P ).
γ(R)
Let C, S(R) be as in Lemma 2.4 with corresponding inclusions of R-modules R1/q → Rq
γ(R)
and Rq
→ R1/q whose cokernels are T1 (q) and T2 (q) respectively. So there are exact
γ(R)
γ(R)
sequences 0 → R1/q → Rq
→ T1 (q) → 0 and 0 → Rq
→ R1/q → T2 (q) → 0 so
that both T1 (q) and T2 (q) have a prime filtration whose prime factors are isomorphic to
R/Q where Q ∈ S(R) and such a prime factor appears no more than Cq γ(R) times in the
filtration. As in the proof of Corollary 3.4 there will be the following commutative diagrams
11
with all vertical maps being surjective.
1/q2
RP
1/q2
RP
1/q
(I [q] RP 2 : 1/q2 uq )
R
P
q
γ(R)
RP2
γ(R)
RqP
(I [q] :
qγ(R)
q
R u )RP
q
γ(R)
RP2
T1 (q2 )P
0
T1′ (q2 )P
0
T2 (q2 )P
0
T2′ (q2 )P
0
γ(R)
RqP
γ(R)
q
(I [q] :R uq )RP
1/q2
RP
1/q
RP 2
1/q
(I [q] RP 2 : 1/q2 uq )
R
P
Ti (q2 )P
for i = 1, 2. It follows
Furthermore, Ti′ (q2 )P will be the homomorphic image of P [q]
Ti (q2 )P
that
!
[q1 ]
1/q
J [q1 ] RP 2
J RP
Ti (q2 )P
γ(R)
q2 − λ
.
λ
≤ max λ
1/q
i=1,2
I [q1 ] RP
P [q] Ti (q2 )P
I [q1 ] RP 2
Ti (q2 )P
RP
γ(R)
For each i = 1, 2, λ P [q]
≤
Cq
max
{
}. We can now apply
Q∈S(R)
[q]
Ti (q2 )P
(Q+P )RP
Proposition 3.3 to know that the desired bound exists.
4. Uniform Bounds in Rings Essentially of Finite Type Over an Excellent
Local Ring
The purpose of this section is to establish Proposition 3.3 and Theorem 3.6 for rings which
are essentially of finite type over an excellent local ring. The following well known Lemma
shall allow us to reduce our considerations to rings which are essentially of finite type over
a complete local ring.
Lemma 4.1. Let R → S be a faithfully flat homomorphism of characteristic p Noetherian
rings with regular fibers. Let M be a finitely generated R-module, P ∈ Spec (R) and IRP an
P RP -primary ideal. Then
1
1
MP
(S ⊗R M)Q
= ht (Q) λSQ
,
λR
q ht (P ) P I [q] MP
q
(I, x)[q] (S ⊗R M)Q
where Q is a prime of S lying over P and x is a regular system of parameters for SQ /P SQ .
12
S
(S⊗ M )
R
Q
P
≃ x[q]QSQ ⊗RP I [q]MM
. Since RP → SQ
Proof. The first thing to observe is that (I,x)[q] (S⊗
R M )Q
P
is flat and SQ /P SQ is regular, we have that
SQ
(S ⊗R M)Q
MP
=λSQ
λSQ
⊗RP [q]
(I, x)[q] (S ⊗R M)Q
x[q] SQ
I MP
=λSQ
=q
ht (Q)−ht (P )
λRP
λRP
MP
[q]
I MP
SQ
(P + x[q] )SQ
MP
[q]
I MP
.
Dividing both sides of the equation by q ht (Q) gives the desired result.
Suppose that R is essentially of finite type over the excellent local ring A. Let  denote
the completion of A with respect to its maximal ideal. Then R → Â ⊗A R is a faithfully
flat homomorphism with regular fibers ([12], Section 33, Lemma 4). This observation and
Lemma 4.1 allow us to reduce proving statements about rings essentially of finite type over
an excellent local ring to rings which are essentially of finite type over a complete local ring.
If R is essentially of finite type over a complete local ring A, then let Λ be a p-base of
the residue field of A. We shall let Γ be a cofinite subset of Λ. For each such Γ there is an
associated R-algebra, RΓ , which satisfies the following.
Theorem 4.2 ([7], Section 6). Let R be a characteristic p ring essentially of finite type
over a complete local ring. Then for each Γ ≤ Λ, RΓ is a faithfully flat, purely inseparable,
F-finite R-algebra.
To say that R → RΓ is purely inseparable is to say that for each s ∈ RΓ , there exists an
Γ
n ∈ N such that sn ∈ R. From this it follows that the induced map Spec
√ (R ) → Spec (R) is
a homeomorphism.√ The inverse map sends a prime P ∈ Spec (R) to P RΓ . If P ∈ Spec (R)
we shall let PΓ = P RΓ .
If R is essentially of finite type over a complete local ring, then for each Γ we have that
P RPΓΓ is PΓ RPΓΓ -primary. If R is essentially of finite type over an excellent local ring A, then
Γ shall represent a cofinite subset of a p-base for a coefficient field of Â. If R is essentially
of finite type over a complete local ring and M a finitely generated R-module, then we let
M Γ = RΓ ⊗R M.
Proposition 4.3. Let R be essentially of finite type over an excellent local ring and let M be
a finitely generated R-module. There exists a constant C > 0 such that for all P ∈ Spec (R)
and q = pe , if IRP ⊆ JRP is a P RP -primary pair of ideals, then
[q]
JRP
J MP
dim(MP )
≤ Cq
λ
.
λ
I [q] MP
IRP
13
Proof. By Lemma 4.1 and the remarks that follow, we may reduce to the scenario that
R is essentially of finite type over a complete local ring. Choose any Γ. Then for each
[q] M
P
P ∈ Spec (R) one sees that by tensoring a prime filtration of JI [q] M
with RPΓΓ that
P
λRΓP
Γ
!
J [q] MPΓΓ
I [q] MPΓΓ
= λRP
J [q] MP
I [q] MP
λRΓP (RPΓΓ /P RPΓΓ ).
Γ
We can now apply Proposition 3.3 to the F-finite ring RΓ so that we know there exists a
constant C such that for all P ∈ Spec (R) and for all q,
[q]
[q]
Γ
[q]
Γ
J
M
/I
M
λ
Γ
R
PΓ
PΓ
J MP
PΓ
λRP
=
Γ
Γ
[q]
I MP
λRΓP (RPΓ /P RPΓ )
Γ
ht (PΓ )
Cq
λRΓP JRPΓΓ /IRPΓΓ
Γ
≤
λRΓP (RPΓΓ /P RPΓΓ )
Γ
JRP
ht (P )
= Cq
λRP
.
IRP
Theorem 4.4. Let R be essentially of finite type over an excellent local ring and let M be
a finitely generated R-module. There exists a constant C such that for all P ∈ Spec (R), for
all q1 , q2 , if IRP ⊆ JRP is a P RP -primary pair of ideals, then
[q1 q2 ]
[q1 ]
J
MP
JRP
J MP
ht (P )
ht (P ) ht (P )−1
q2
−λ
.
≤ Cq2
q1
λ
λ
I [q1 ] MP
I [q1 q2 ] MP
IRP
Proof. The proof of this Theorem is identical to the proof of Proposition 4.3. Lemma 4.1
allows us to reduce to the scenario that R is essentially of finite type over an excellent local
ring. Pick a Γ and let C be as in Theorem 3.6 for the F-finite ring RΓ , then
[q1 ]
[q1 q2 ]
J MP
J
MP
ht (P )
λRP
q2
− λRP
I [q1] MP
I [q1 q2 ] MP
= λRΓP
≤
Γ
J [q1 ] MPΓΓ
I [q1 ] (MPΓΓ )
!
ht (P )
q2
ht (PΓ ) ht (PΓ )−1
q1
λRPΓ
Cq2
λRΓP
Γ
RΓ
P
Γ
P RΓ
P
Γ
− λRΓP
JRPΓ
IRPΓ
=
Γ
J [q1 q2 ] MPΓΓ
I [q1 q2 ] MPΓΓ
!,
ht (P ) ht (P )−1
Cq2
q1
λRP
λRΓP
Γ
JRP
IRP
RPΓΓ
P RPΓΓ
.
14
5. Uniform Convergence and Continuity Results
Theorem 5.1. Let R be either F-finite or essentially of finite type over an excellent local
ring and let M be a finitely generated R-module. Let I(−) be a map of primary ideals.
λ (I(−))
The sequence of functions λq11(I(−)) : Supp (M) → R, which sends a prime P ∈ Supp (M)
λ(MP /I(P )[q1 ] MP )
dim(MP
q1
)λ(RP /I(P )RP )
eHK (I(−),M− )
to
λ1 (I(−))
, converges uniformly to the scaled Hilbert-Kunz multiplicity function
, which sends a prime P ∈ Supp (M) to
eHK (I(P ),MP )
λ(RP /I(P )RP )
as q1 → ∞.
Proof. Given ǫ > 0, our goal is to show that there exists a q ′ such that for all P ∈ Supp (M)
1
1
and for all q1 ≥ q ′ , | λ(RP /I(P
λ (I(P )) − λ(RP /I(P
e (I(P ), MP )| < ǫ. After modding
)RP ) q1
)RP ) HK
out R by Ann R M, it follows by Theorems 3.6 and 4.4 that there exists a constant C > 0
such that, for all P ∈ Supp (M) and for all q1 , q2 ,
MP
MP
dim(MP )
q2
−λ
λ
I(P )[q1] MP
I(P )[q1q2 ] MP
≤
dim(MP ) dim(MP )−1
Cq2
q1
λ
RP
(I(P ) + Ann R (M))RP
≤
dim(MP ) dim(MP )−1
Cq2
q1
λ
RP
I(P )RP
.
dim(M )
P
Dividing both sides of the inequality by q2
, letting q2 → ∞, and applying Lemma
2.6, gives that for all P ∈ Supp (M) and for all q1 ,
RP
MP
dim(MP )
dim(MP )−1
− q1
eHK (I(P ), MP ) ≤ Cq1
λ
.
λ
I(P )[q1] MP
I(P )RP
Choose q ′ large enough that
dim(MP )
q1
C
q′
< ǫ and let q1 ≥ q ′ . Dividing the above inequality by
λ(RP /I(P )RP ) gives that for all P ∈ Supp (M) and all q1 ,
λM
C
eHK (I(P ), MP )
q1 (I(P ))
≤
−
< ǫ.
λ(RP /I(P )RP ) λ(RP /I(P )RP )
q1
Let n ∈ N and set fn (P ) =
1
dim(MP )
q1
λ(MP /I(P )[q1] MP ) where q1 = pn and let f be the limit
function f (P ) = eHK (I(P ), MP ). What Theorem 5.1 is saying is that there exists a strictly
1
positive function g : Spec (R) → R, namely g(P ) = λ(RP /I(P
, which does not depend on
)RP )
n, such that gfn converges uniformly to the function gf . If there exists a δ > 0 such that
for all P ∈ Spec (R) g(P ) ≥ δ, then fn converges uniformly to f . To see this we can choose
n so large so that for all |gfn − gf | < ǫδ. Then |fn − f | < ǫδ/g ≤ ǫδ/δ = ǫ. Using this
observation we obtain the following Corollary to Theorem 5.1.
Corollary 5.2. Let R be an F-finite ring of prime characteristic p > 0 and let M be a
finitely generated R-module. Let I(−) be a map of primary ideals. Suppose that there exists
15
a q such that P [q] ⊆ I(P ) for all P ∈ Supp (M), or more generally there exists a constant
D such that λ(RP /I(P )RP ) ≤ D for all P ∈ Supp (M). Then the sequence of functions
1
λ(MP /I(P )[q1] MP ), converges
λq1 (I(−)) : Supp (M) → R, which sends a prime P to dim(M
P)
q1
uniformly to the Hilbert-Kunz multiplicity function eHK (I(−), M− ), which sends a prime P
to eHK (I(P ), MP ).
Proof. By the above remarks we only need to find δ > 0 such that for all P ∈ Supp (M),
1
≥ δ, or equivalently that there exists a D such that for all P ∈ Supp (M),
λ(RP /I(P )RP )
λ(RP /I(P )RP ) ≤ D. We are assuming that for each P ∈ Spec (R) that P [q] ⊆ I(P ). Hence
by Lemma 3.3 there exists a constant C such that for all P ∈ Supp (M), λ(RP /I(P )RP ) ≤
λ(RP /P [q]RP ) ≤ Cq ht (P ) ≤ Cq dim(R) . Therefore D = Cq dim(R) works.
Corollary 5.2 gives an alternative proof of Smirnov’s result that if R is F-finite or essentially
of finite type over an excellent local ring, then eHK (−) is upper semi-continuous at primes
P such that RP is equidimensional.
Corollary 5.3. Let R be either F-finite or essentially of finite type over an excellent local
ring, then the Hilbert-Kunz function eHK (−) : Spec (R) → R≥1 which sends a prime P 7→
eHK (RP ) is upper semi-continuous at all P ∈ Spec (R) such that RP is equidimensional.
Proof. Consider the map of primary ideals I(−) which sends a prime P to P . Then P [1] =
P ⊆ I(P ) for each P ∈ Spec (R). Corollary 5.2 says that λq1 (−) converges uniformly to
eHK (−). E. Kunz originally showed in [11] that for each q1 the function λq1 (−) which sends
a prime P 7→ qht1(P ) λ(RP /P [q] RP ), is upper semi-continuous on all rings which are locally
equidimensional. If RP is equidimensional, then R being catenary implies that there is an
s ∈ R − P such that Rs is locally equidimensional. The s which works is 1 if min(RP ) =
min(R). If min(RP ) ( min(R), then just choose s ∈ ∩Q∈M in(R)−min(RP ) Q \ P . Therefore, if
RP is equidimensional, then in an open neighborhood of P , eHK (−) is the uniform limit of
upper semi-continuous functions, hence eHK (−) is upper semi-continuous as well.
Lemma 5.4. Let (R, m, k) be an excellent reduced local ring of dimension d. Let q1 , q2
equal pe1 and pe2 respectively and bq1 = q1d sq1 , bq1 q2 = (q1 q2 )d sq1 q2 , where sq1 and sq1 q2 are
the q1 th and q1 q2 th normalized Frobenius splitting numbers of R respectively. Then there
is an irreducible m-primary ideal I and u ∈ (I : m) such that bq1 = λ((I, u)[q1] /I [q1 ] ) and
bq1 q2 = λ((I, u)[q1q2 ] /I [q1 q2 ] ).
Proof. Let Ie = {r ∈ R | F∗e r⊗u = 0 in F∗e R⊗R ER (k)} where u generates the socle of ER (k).
Then bq = λ(R/Ie ), ([16], Remark 2.3). Since R is reduced and excellent, R is approximately
Gorenstein ([6], Theorem 1.7). So there exists a descending chain of irreducible m-primary
ideals {It }t∈N which is cofinal with {mt }t∈N . Let ut generate the socle mod It . Then Ie =
[q]
[q]
∪∞
: uqt ), therefore for each q there is a t0 such that for all t ≥ t0 , bq = λ(R/(It :
t=1 (It
[q]
uqt )) = λ((It , ut )[q] /It ).
16
Theorem 5.5. Let R be either F-finite or essentially of finite type over an excellent local
ring. There exists a constant C such that, for all P ∈ Spec (R), and for all q1 , q2 ,
ht (P )
ht (P ) ht (P )−1
q1
.
|bq1 (P )q2
− bq1 q2 (P )| ≤ Cq2
Proof. It is well known that if bq (P ) > 0 for some, equivalently for all, q, then RP is a
reduced ring. Therefore C = 0 is a constant which works for all P ∈ Spec (R) such that RP
is not reduced. If RP is reduced there exists an s ∈ R − P such that Rs is reduced. Therefore
by quasi-compactness of Spec (R), we may reduce our considerations to when R is a reduced
ring. The Theorem now follows by Lemma 5.4, Theorem 3.6, and Theorem 4.4.
Theorem 5.6. Let R be either F-finite or essentially of finite type over an excellent local ring.
The qth normalized Frobenius splitting number function, which maps a prime P 7→ sq (P ),
converges uniformly to the F-signature function, which maps a prime P 7→ s(RP ) as q → ∞.
Proof. Let ǫ > 0, let C be as in Theorem 5.5, and choose q so large that
P ∈ Spec (R) we have that
ht (P )
< ǫ. Then for all
ht (P ) ht (P )−1
q1
.
|bq1 (P )q2
Therefore
C
q
− bq1 q2 (P )| ≤ Cq2
bq1 q2 (P )
ht (P )−1
≤ Cq1
.
(q1 q2 )ht (P )
Letting q2 → ∞ we have that for all P ∈ Spec (R) that
ht (P )
bq1 (P ) − q1
ht (P )
|bq1 (P ) − q1
Hence for all q1 ≥ q and all P ∈ Spec (R),
bq1 (P )
ht (P )
q1
This verifies that
bq (P )
q ht (P )
ht (P )−1
s(RP )| ≤ Cq1
− s(RP ) ≤
.
C
C
≤
< ǫ.
q1
q
= sq (P ) converges uniformly to s(RP ) as q → ∞.
Theorem 5.7. Let R be either F-finite or essentially of finite type over an excellent local
ring. The F-signature function on Spec (R) is lower semi-continuous.
Proof. Let ǫ > 0 and let P ∈ Spec (R). If s(RP ) = 0 then it is the case that for all
Q ∈ Spec (R), that s(RP ) − s(RQ ) ≤ 0 < ǫ. Now suppose that s(RP ) > 0. Aberbach
and Leuschke showed in [1], along with Tucker’s proof of the existence of s(RP ) in [15], that
s(RP ) > 0 if and only if RP is strongly F-regular. In particular we have that RP is a domain.
There then exists an s ∈ R − P such that Rs is a domain. Enescu and Yao showed in [5]
that if S is a locally equidimensional ring which is either F-finite or essentially of finite type
over an excellent local ring, then the qth normalized Frobenius splitting number function is
lower semi-continuous on Spec (S). By Theorem 5.4, we have that in a neighborhood of P ,
the F-signature function is the uniform limit of lower semi-continuous functions, hence itself
is lower semi-continuous at P .
17
Observe that Theorem 5.6 applied to the maximal ideal of an excellent local ring (R, m, k)
bq
directly shows the sequence qdim(R)
is a Cauchy sequence.
6. Lower Semi-Continuity of F-signature of Pairs
In this section all rings under consideration will be F-finite. We want to establish the
lower semi-continuity of the F-signature of a pair (R, D) where D is a Cartier algebra, see
section 2 of [2] for a more in-depth look at the basic notions of a Cartier subalgebra. Our
main tool will be Proposition 3.3 in order to establish a uniform convergence result and the
desired lower semi-continuity.
L
Let Cq := HomR (F∗q R, R) and C R = q=pe ,e≥0 Cq . If ϕ ∈ Cq and ψ ∈ Cq′ then ϕ · ψ :=
′
′
ϕ ◦ F∗q ψ ∈ Cqq′ where F∗q ψ(F∗qq r) := F∗q ψ(F∗q r). We call C R the (total) Cartier algebra
of R. Note that with the multiplication defined on homogenous elements of C R makes C R
a noncommutative Fp -algebra. Even though the 0th graded piece of C R is Cp0 = C1 =
HomR (R, R) ≃ R, R is not central in C R , hence C R is not an R-algebra. We say that
D ⊆ C R is a Cartier subalgebra of R if D is a Fp -subalgebra of C R and D1 = C1 ≃ R.
Suppose
L that (R, m, k) is a local ring and D a Cartier subalgebra of R. Suppose that
F∗q R ≃
Mi as an R-module. The summand Mi is called a D-summand if Mi ≃ R and the
projection F∗q R → Mj ≃ R is an element of Dq . The qth F-splitting number of (R, D) is the
maximal number aD
q of D-summands appearing in the various direct sum decompositions
q
of F∗ R. Observe that aCq = aq for all q, the usual qth F-splitting number of R. For each
q = pe let IqD = {r ∈ R | ϕ(F∗q r) ∈ m for all ϕ ∈ Dq }. The following Lemma is a list of basic
properties about the sets IqD which can all be found in Section 3 of [2].
Lemma 6.1. Let (R, m, k) be a local F-finite ring and D a Cartier subalgebra and let q, q1 , q2
be various powers of p and ϕ ∈ Dq1 . Then
(1) IqD ⊆ R is an ideal,
(2) m[q] ⊆ IqD ,
(3) ϕ(F∗q1 IqD1 q2 ) ⊆ IqD2 ,
(4) λ(R/IqD ) =
aD
q
α(R)
q
.
Let (R, m, k) be local and D a Cartier subalgebra. Set ΓD to be the semigroup {q | aD
q 6= 0}.
The main result of Blickle, Schwede, and Tucker in [2] is that if (R, m, k) is local and D a
aD
q
1
Cartier subalgebra, then the limit limq∈ΓD →∞ qα(R)+dim(R)
= limq∈ΓD →∞ qdim(R)
λ(R/IqD ) exists,
it is called the F-signature of the pair (R, D), and is denoted s(R, D).
Suppose that R is F-finite and not necessarily local. Let D be a Cartier algebra of R.
Suppose that S ⊆ R is a multiplicatively closed set. Since R is F-finite, S −1 HomR (F∗q R, R) ≃
HomS −1 R (F∗q S −1 R, S −1 R). Therefore there is a naturally induced Cartier subalgebra S −1 D
of S −1 R such that (S −1 D)q = S −1 (Dq ). If P ∈ Spec (R) and s ∈ R we write DP and Ds
for the induced Cartier subalgebra of RP and Rs respectively. For each P ∈ Spec (R) let
aq (P, D) be the qth F-splitting number of (RP , DP ), sq (P, D) = aq (P, D)/q α(P )+ht (P ) , and let
s(P, D) = s(RP , DP ). Then sq (P, D) : Spec (R) → R converges to s(P, D) : Spec (R) → R
18
as q ∈ ΓD → ∞ as a limit of functions. For each q and P ∈ Spec (R), let
IqD (P ) = {r ∈ RP | ∀ϕ ∈ (DP )q , ϕ(F∗q r) ∈ P RP }.
Remark 6.2. If R is an F-finite ring, D a Cartier subalgebra, N, M ∈ N, f ∈ DqN1 , g ∈ DqM2 ,
then the natural map (g, g, ..., g) ◦F∗q2 f : F∗q1 q2 R → RN M is an element of DqN1 qM2 .
| {z }
N times
The remark follows from the assumption that if ϕ ∈ Dq1 and ψ ∈ Dq2 , then ψ ◦ F∗q2 ϕ ∈ Dq1 q2 .
Our first goal will to be establish a version of Corollary 2.3 for a pair (R, D) when R is an
F-finite domain. If R is a not necessarily local F-finite domain and D a Cartier subalgebra,
γ(R)
we let ΓD = ΓD0 . We will now only be interested in containments of R-modules R1/q ⊆ Rq
when q ∈ ΓD . Not only will we need to know that for each q ∈ ΓD that a prime filtration of
γ(R)
R1/q ⊆ Rq
has only finitely many prime factors up to isomorphism and such prime factors
only appears a controlled number of times, but we will need that each of the following q γ(R)
γ(R) πi
maps is an element of Dq , R1/q ⊆ Rq
−
→ R, where πi is the projection onto the ith factor.
Lemma 6.3. Let R be an F-finite domain and D a Cartier subalgebra. There exists a finite
set of nonzero primes S(R, D) and a constant C such that for every q ∈ ΓD ,
γ(R)
γ(R)
which is an element of Dqq ,
(1) there is a containment of R-modules R1/q ⊆ Rq
(2) which has a prime filtration whose prime factors are isomorphic to R/P , where P ∈
S(R, D),
(3) and for each P ∈ S(R, D), the prime factor R/P appears no more than Cq γ(R) times
γ(R)
in the prime filtration of the containment R1/q ⊆ Rq .
Proof. Let ΓD be generated by ΛD = {q1 , ..., qm } as a semigroup. For each qi ∈ ΛD we can
γ(R)
q
γ(R)
fix an embedding R1/qi ⊆ Rqi which is an element of Dqii which is an isomorphism when
1/q
q γ(R)
. As RW is
localized at 0. To see this, let W = R − 0 and q ∈ ΛD so that RW ≃ RW
1/q
1/q
1/q
a field, HomRW (RW , RW ) ≃ RW as an RW -module. Suppose that 0 6= ϕ ∈ Dq , then ϕW
1/q
1/q
1/q
generates HomRW (RW , RW ) ≃ RW as an RW -module. As D0 = HomR (R, R), we have that
1/q
1/q
q γ(R)
a RW -multiple of ϕW is still an element of (DW )q . Therefore the isomorphism RW ≃ RW
γ(R)
1/q
q γ(R)
is an element of (DW )qq . As R is an F-finite domain, the isomorphism RW ≃ RW
is the
1/q
q γ(R)
q γ(R)
localization of an embedding R ⊆ R
which is an element of Dq
.
γ(R)
For each q ∈ ΛD we consider a prime filtration of R1/q ⊆ Rq , say R1/q = N0 ⊆
γ(R)
N1 ⊆ · · · ⊆ Nn = Rq S, say Ni /Ni−1 ≃ R/Pi . Let S(R/Pq,i ) and
S Cq,i be as in Lemma
n
2.2 and let Sq (R, D) = i=1 S(R/Pq,i ) ∪ {Pq,i } and S(R, D) = q∈ΛD Sq (R, D). We can
P
Q
now set C ′ =
Cq,i . Every q ∈ ΓD can be expressed as qi ∈ΛD qiei where ei ∈ N. We
P
show by induction on
ei that for each q ∈ ΓD there is a containment of R-modules
γ(R)
1/q
q γ(R)
which is an element of Dqq , which has a prime filtration whose prime factors
R ⊆R
are isomorphic
to R/P , where P ∈ S(R, D), and such a prime factor appears no more than
P
1
′ γ(R)
ei = 1.
Cq
1 + p + · · · + 1q times in the filtration. This trivially holds for
Q
P
Now suppose that q = qi ∈ΛD qiei with
ei > 1. Without loss of generality we may
′
q
′
suppose that e1 ≥ 1 so that q = q1 ∈ ΓD . By induction, we can find R1/q = N0 ⊆ N1 ⊆
19
′γ(R)
· · · ⊆ Nm = Rq
is a prime filtration of an embedding R1/q ⊆ Rq
in Dqq′ , each
Nj /Ni−j ≃ R/Pj for some Pj ∈ S(R, D), and such a prime factor appears no more than
γ(R)
γ(R)
′
′
C ′ q ′γ(R) (1 + p1 + p12 + · · · + q1′ ) times in the filtration. Therefore (Rq1 )1/q = (R1/q )q1 ⊆
′γ(R)
′γ(R)
γ(R)
′
′γ(R)
γ(R)
(Rq
)q 1 = R q
has a prime filtration with prime factors R/Pj with Pj ∈ S(R, D) and
such a prime factor appears no mare than C ′ q γ(R) (1 + p1 + p12 + · · · + 1q ) times in the filtration.
γ(R)
Furthermore, the prime filtration R1/q1 = Nq1 ,0 ⊆ Nq1 ,1 ⊆ · · · ⊆ Nq1 ,n = Rq1
γ(R)
′
′
following filtration of R1/q = (R1/q1 )1/q ⊆ (Rq1 )1/q ,
′
1/q ′
1/q ′
′
γ(R)
(R1/q1 )1/q = Nq1 ,0 ⊆ Nq1 ,1 ⊆ · · · ⊆ Rq1/q
= (Rq1
1 ,n
1/q ′
1/q ′
gives the
′
)1/q .
′
Since Nq1 ,i /Nq1 ,i−1 ≃ (R/Pq1 ,i )1/q , we apply Lemma 2.2 to know there is a prime filtration
1/q ′
1/q ′
of each Nq1 ,i−1 ⊆ Nq1 ,i whose prime factors come from S(R/Pq1 ,i ) and such a prime factor
appears no more than Ci q γ(R/Pq1 ,i ) ≤ C ′ q γ(R)−1 times in the filtration. Putting all of this
γ(R)
γ(R)
information together we get an embedding R1/q ⊆ Rq , which is an element of Dqq
by
Remark 6.2, with a prime filtration whose prime factors come from S(R, D), and such a
prime factor appears no more than the following number in the filtration,
X
h
1
1
1 1
1
γ(R) ′ ′γ(R)
Ci q γ(R/Pq1 ,i ) ≤ C ′ q γ(R) (1 + + · · ·+ ′ + ) ≤ 2C ′ q γ(R) .
1+ +···+ ′ +
q1 C q
p
q
p
q q
i=1
Enescu and Yao showed that aq (P, C ) : Spec (R) → R, hence sq (P, C ), is lower semicontinuous on an F-finite ring which is locally equidimensional, (Corollary 2.5, [5]). We
provide a very similar proof that shows aq (P, D), hence sq (P, D), is a lower semicontinuous
function for any Cartier subalgebra D whenever R is locally equidimensional. It is well
known that a function f : X → R, X a topological space, is lower semi-continuous if and
only if for each r ∈ R the sets f −1 ((r, ∞)) = {x ∈ X | f (x) > r} is open in X.
Lower semi-continuity is a local condition. We may assume RP , hence R is reduced, else
aq (P, D) = 0 and aq (−, D) is trivially lower semi-continuous at P . Suppose that q ∈ ΓD ,
a (P,D) L
1/q
r ∈ R, and let P ∈ {Q ∈ Spec (R) | aq (Q, D) > r}. Then RP ≃ RPq
MP is such
1/q
that each of the aq (P, D) projections RP → RP is an element of (DP )q . It follows that
a (P,D) L
1/q
there is an s ∈ R − P such that Rs ≃ Rs q
Ms and each of the aD
q (P )-projections
1/q
′
′
Rs → Rs is an element of (Ds )q . Hence for all P ∈ D(s), aq (P , D) ≥ aq (P, D)) > r
D
and {Q ∈ Spec (R) | aD
q (Q) > r} is indeed an open set. This shows that aq (P ) is a lower
semi-continuous and so is sq (P, D) since aq (P, D) and sq (P, D) differ only by a constant on
connected components of Spec (R).
Consider the following condition we could impose on a Cartier subalgebra D.
(1)
D
(IqD (P ))[p] ⊆ Iqp
(P )
Suppose R is an F-finite domain and D a Cartier subalgebra. Then r ∈ IqD (P ) if and only
if ϕ(r 1/q ) ∈ P RP for all ϕ ∈ Dq ⊆ HomR (R1/q , R). Thus to impose condition (1) is to
impose that for each r ∈ IqD (P ) that ψ(r 1/q ) ∈ P RP for all ψ ∈ Dqp ⊆ HomR (R1/qp , R).
20
This condition is seen to be satisfied if for each ψ ∈ Dqp we require ϕ ◦ i ∈ Dq where i is the
natural inclusion R1/q ⊆ R1/qp .
Theorem 6.4. Let R be an F-finite domain and D a Cartier subalgebra of R. Then the Fsignature function which sends P ∈ Spec (R) to s(P, D) is lower semi-continuous. Moreover,
if the the Cartier subalgebra satisfies (1), then the function sq (P, D) converges uniformly to
the F-signature function s(P, D) as q ∈ ΓD → ∞.
Proof. Let C, S(R) be as in Lemma 6.3. Let q1 ∈ ΓD so that Dq1 6= 0. Let S(q1 ) be the
γ(R)
cokernel of R1/q1 → Rq1 . Therefore we have the following short exact sequences
γ(R)
0 → R1/q1 → Rq1
→ S(q1 ) → 0.
By Remark 6.2 we have exact sequences
q
1/q
γ(R)
RP 1
RP1
→ S̃(q1 ) → 0,
→
γ(R)
q1
IqD1 q2 (P )1/q1
D
Iq2 (P )RP
where S̃(q1 ) is the homomorphic image of S(q1 )P /IqD1 (P )S(q1 )P . Therefore by parts (2) and
(4) of Lemma 6.1,
aq2 (P, D) γ(R) aq1 q2 (P, D) α(P )
S(q1 )P
S(q1 )P
q1 −
≤λ
.
q
≤ λ(S̃(q1 )P ) ≤ λ
α(P )
(q1 q2 )α(P ) 1
IqD2 (P )S(q1 )P
P [q2] S(q1 )P
q2
By Proposition 3.3 there is a constant C1 , independent of P, q2 , such that
ht (P )−1
max λ RP /(Q + P [q2 ] )RP ≤ C1 q2
.
Q∈S(R)
It follows that
aq2 (P, D)
α(P )
q2
γ(R)
q1
−
aq1 q2 (P, D)
α(P )
q2
≤λ
S(q1 )P
[q
P 2] S(q1 )P
γ(R) ht (P )−1
q2
Dividing both sides of the inequality by q1
sq2 (P, D) − sq1 q2 (P, D) ≤
γ(R) ht (P )−1
q2
.
≤ CC1 |S(R)|q1
shows that
CC1 |S(R)|
.
q2
Letting q1 → ∞, and relabeling constants, shows that there is a constant C, independent of
P, q such that
C
sq (P, D) − s(P, D) < .
q
To see that s(−, D) is lower semi-continuous at P ∈ Spec (R) we may assume s(P, D) > 0,
else s(−, D) is trivially lower semi-continuous. Thus we may assume that RP is a strongly
F-regular domain. In particular, s(−, D) is the limit of lower semicontinuous functions in
an open neighborhood of P . Thus the lower semicontinuity of the sq (−, D) will now imply
the lower semicontinuity of s(−, D) since there is a constant C independent of Q ∈ Spec (R)
such that sq (Q, D) − s(Q, D) < Cq . Observe that up to this point in the proof we have not
used the assumption that the Cartier subalgebra D satisfies condition (1).
21
Now assume that the Cartier subalgebra D satisfies condition (1). Let C and S(R) be as
in Lemma 2.2 applied to the F-finite domain R. Let S(q1 ) be the cokernels of the inclusions
γ(R)
R q1
fq
1
−−→
R1/q1 . Then there are short exact sequences
fq
γ(R)
0 → R q1
1
−−→
R1/q1 → S(q1 ) → 0.
γ(R)
γ(R)
We claim that fq1 (IqD2 (P )Rq1 ) ⊆ IqD1 q2 (P )1/q1 . Let x ∈ Rq1
and r ∈ IqD2 (P ). Then
fq1 (rx)q1 = r q1 fq1 (x)q1 ∈ IqD2 (P )[q1] ⊆ IqD1 q2 (P ) by part (3) of Lemma 6.1 and the assumption
(1). Therefore there are induced exact sequences
q
γ(R)
RP1
fq
q
γ(R)
IqD2 (P )RP1
1
−−→
1/q
RP 1
→ S̃(q1 ) → 0.
IqD1 q2 (P )1/q1
1/q
Observe that by part (2) of Lemma 6.1 that P [q2] RP kills RP 2 /IqD1 q2 (P )1/q1 , hence P [q1] RP
kills S̃(q1 ). Therefore S̃(q1 ) is the homomorphic image of S(q1 )/P [q2] S(q1 )P . We can now
proceed as before to get a constant C independent of P and q such that
C
s(P, D) − sq (P, D) < .
q
Hence there is a constant C independent of P ∈ Spec (R) such that |s(P, D)−sq (P, D)| <
which implies sq (−, D) converges uniformly to s(−, D).
C
,
q
Acknowledgements
The author would like to Ian Aberbach for his support during the duration of this project
and suggesting the problem, Kevin Tucker for his suggestions which made the proofs much
simpler, Yongwei Yao for carefully reading an earlier version and providing feedback, and
Dale Cutkosky for his comments.
References
[1] Ian Aberbach and Graham Leuschke. The F-signature and strong F-regularity. Mathematical Research
Letters, 10(1):51–56, 2003.
[2] Manuel Blickle, Karl Schwede, and Kevin Tucker. F-signature of pairs and the asymptotic behavior of
frobenius splittings. Advances in Mathematics, 231(6):3232–3258, 2012.
[3] Manuel Blickle, Karl Schwede, and Kevin Tucker. F-signature of pairs: Continuity, p-fractals and minimal log discrepancies. Journal of the London Mathematical Society, 87(3):802–818, 2013.
[4] Sankar Dutta. Frobenius and multiplicities. Journal of Algebra, 85(2):424–448, 1983.
[5] Florian Enescu and Yongwei Yao. The lower semicontinuity of the Frobenius splitting numbers. Mathematical Proceedings of the Cambridge Philosophical Society, 150(1):35–46, 2011.
[6] Melvin Hochster. Cyclic Purity Versus Purity in Excellent Noetherian Rings. Transactions of the American Mathematical Society, 231(2):463–488, 1977.
[7] Melvin Hochster and Craig Huneke. F-Regularity, Test Elements, and Smooth Base Change. Transactions of the American Mathematical Society, 346(1):1–62, 1994.
[8] Craig Huneke. Hilbert-Kunz Multiplicity and the F-Signature. Springer, New York.
[9] Craig Huneke. Tight Closure and Its Applications, volume 88. CBMS Regional Conference Series in
Mathematics.
[10] Craig Huneke and Graham Leuschke. Two theorems about maximal Cohen-Macaulay modules. Mathematische Annalen, 324(2):391–404, 2002.
22
[11] Ernst Kunz. On Noetherian Rings of Characteristic p. American Journal of Mathematics, 98(4):999–
1013, 1976.
[12] Hideyuki Matsumura. Commutative Algebra, volume 56. Benjamin Cummings Publishing Co., Inc.
[13] Paul Monsky. The Hilbert-Kunz function. Mathematische Annalen, 263(1):43–49, 1983.
[14] Ilya Smirnov. Upper semi-continuity of the hilbert-kunz function. arXiv:1407.6476, 2014.
[15] Kevin Tucker. F −Signature Exists. Inventiones mathematicae, 190(3):743–765, 2012.
[16] Yongwei Yao. Observations on the F-signature of local rings of characteristic p. Journal of Algebra,
299(1):198–218, 2006.
Thomas Polstra, Department of Mathematics, University of Missouri-Columbia, Columbia,
MO 65211
E-mail address: [email protected]
23
| 0math.AC
|
arXiv:1305.0062v2 [q-bio.GN] 22 May 2013
Distilled Single Cell Genome Sequencing and De Novo
Assembly for Sparse Microbial Communities
Zeinab Taghavi∗†1 , Narjes S. Movahedi‡1 , Sorin Draghici§1,2 , and Hamidreza Chitsaz¶1
1
2
Department of Computer Science, Wayne State University, Detroit, MI 48202
Department of Obstetrics and Gynecology, Wayne State University, Detroit, MI 48202
May 24, 2013
Abstract
Identification of every single genome present in a microbial sample is an important and
challenging task with crucial applications. It is challenging because there are typically millions
of cells in a microbial sample, the vast majority of which elude cultivation. The most accurate
method to date is exhaustive single cell sequencing using multiple displacement amplification,
which is simply intractable for a large number of cells. However, there is hope for breaking this
barrier as the number of different cell types with distinct genome sequences is usually much
smaller than the number of cells.
Here, we present a novel divide and conquer method to sequence and de novo assemble
all distinct genomes present in a microbial sample with a sequencing cost and computational
complexity proportional to the number of genome types, rather than the number of cells. The
method is implemented in a tool called Squeezambler. We evaluated Squeezambler on simulated
data. The proposed divide and conquer method successfully reduces the cost of sequencing in
comparison with the naı̈ve exhaustive approach.
Availability: Squeezambler and datasets are available under
http://compbio.cs.wayne.edu/software/squeezambler/
1
Introduction
Critical applications, including the Human Microbiome Project (Methe et al., 2012), biothreat
detection, and combating antibiotic resistant pathogens, necessitate identification of all distinct
genome sequences in a bacterial sample. When prior knowledge is available about which organisms may be present in the sample, flow cytometry and 16S rRNA gene sequencing may be used.
However, metagenomics is usually used for analyzing the genomics of relatively abundant microbes
when no prior knowledge is given. Metagenomics consists in the study of the variation of species
in a complex microbial sample. Since the vast majority of environmental bacteria elude cultivation, metagenomics investigates microbial communities by sequencing sampled genome fragments
∗
To whom correspondence should be addressed.
[email protected]
‡
[email protected]
§
[email protected]
¶
[email protected]
†
1
I1,1
I1,2
A1,1 A1,2
A1,2 A1,1
I2,1
I3,1
I2,2
I3,2
A3,1 A3,4
I3,3
I3,4
A3,3 A3,4
I4,1
I4,2
I4,3
Figure 1: The divide and conquer algorithm for an example with 10 cells and 3 distinct genomes
shown in different colors. Each row corresponds to one sequencing round. The number of barcodes
in each round is the number of blue boxes in the corresponding row.
without the need for culturing. Such a heterogeneous pool of sequencing reads can also be assembled to yield a superposition of highly abundant genomes in the sample (Treangen et al., 2013).
There are two problems with metagenomics: (i) the resulting assembly contains multiple genomes
superimposed, and (ii) only highly abundant genomes survive the sampling process.
Advances in DNA amplification technology have enabled whole genome sequencing directly
from individual cells without requiring growth in culture. Single cell sequencing methods have enabled investigation of novel uncultured microbes (Kvist et al., 2007; Mussmann et al., 2007). These
culture-independent single cell studies are a powerful alternative to metagenomics studies. Genomic
sequencing from single bacterial genomes was first demonstrated with cells isolated by flow cytometry (Raghunathan et al., 2005), using multiple displacement amplification (MDA) (Dean et al.,
2002, 2001; Hosono et al., 2003). MDA is now the preferred method for whole genome amplification from single cells (Lasken, 2007; Ishoey et al., 2008). The first attempt to assemble a complete
bacterial genome from one cell (Zhang et al., 2006) further explored the challenges of assembly from
amplified DNA, including amplification bias and chimeric DNA rearrangements. Amplification bias
results in orders of magnitude difference in coverage (Raghunathan et al., 2005), and absence of
coverage in some regions. Chimera formation occurs during the DNA branching process by which
the φ29 DNA polymerase generates DNA amplification in MDA (Lasken and Stockwell, 2007). Subsequent studies continued to improve single cell assemblies (Marcy et al., 2007; Podar et al., 2007;
Hongoh et al., 2008; Rodrigue et al., 2009; Woyke et al., 2009). A nearly full potential of single cell
genome assembly has recently been realized by the work of Chitsaz et al., 2011 followed by those
of Bankevich et al., 2012 and Peng et al., 2012.
Due to recent progress in single cell DNA amplification techniques and de novo assembly algorithms, the genomes of all bacterial species in a sample can be captured one cell at a time.
However, there are often millions of cells per sample, in which case the naı̈ve deep sequencing of
every cell becomes prohibitively costly. Moreover, it is expected that deep sequencing of every cell
2
is often not necessary as the majority of biologically important samples are sparse in the sense that
many cells are biological replicates. Compressed (Candès and Tao, 2005, 2006; Donoho, 2006) and
distilled (adaptive sampling-and-refinement) sensing methods (Haupt et al., 2011; Wei and Hero,
2012) have been proposed in the last decade to exploit sparsity for reducing the cost of sensing and reconstructing signals in various spaces, ranging from Banach spaces to Boolean algebras
(Erlich et al., 2010; Stobbe and Krause, 2012). Inspired by those advances, we give an algorithm
in this paper that exploits sample sparsity to reduce the cost of sequencing without compromising
the accuracy of identification of all distinct genomes, even the ones that are minimally represented
in the sample.
2
Approach
A naı̈ve approach to solve the problem, which we call single cell co-assembly strategy, is to amplify
the genome of each cell, barcode them individually, pool them, sequence in one sequencing run,
and demultiplex based on the barcode. In this approach, each cell should be isolated and its DNA
extracted and amplified using multiple displacement amplification. Although there is currently
no high throughput device to perform these processes, one could envision automated microfluidic
devices that will be capable of high throughput separation, DNA extraction, amplification, and
barcoding of single cells in the near future. The output sequencing reads could then be co-assembled
using our tool HyDA (Movahedi et al., 2012). In HyDA, the read dataset of each cell is assigned a
unique color. All the colors are co-assembled in one colored de Bruijn graph. This approach requires
enough unique barcodes to tag the fragments of each cell. Also, barcodes attached to each fragment
are sequenced, which imposes additional sequencing cost. Fabrication of so many unique barcodes
becomes prohibitively expensive for a large number of cells and therefore, this naı̈ve approach will
not work.
The number of distinct genomes in a microbial community is often much less than the number
of cells. For example, Qin et al., 2010 estimated the number of detected distinct species in the
human gut to be in the order of 1,000, while the number of microbial cells in a human body,
most of which reside in the gut, is in the order of 100 trillion. We call this effect the sparsity
of distinct genomes in a sizable microbial population. The naı̈ve approach does not exploit the
sparsity to reduce the cost of sequencing. Here, we proposed to exploit the spareness by adopting
a divide and conquer strategy to reduce the amount of required barcodes and sequencing. After
isolation of each cell and extraction of the DNA, every DNA is amplified and stored separately,
e.g. in a microfluidic droplet. The main idea is to sample the amplified DNAs adaptively, which is
essentially allocating sequencing and barcoding resources dynamically over the course of multiple
sensing iterations. Initially, the algorithm has one group of cells, which is the entire sample. In each
iteration, each group is divided into two equally-sized subgroups. A small amount of DNA from
each cell in a subgroup is sampled, pooled, and sequenced. In practice, one barcode per subgroup
is used to multiplex and demultiplex the sequencing in one run. The amount of sampling from each
cell is computed based on the results of previous iterations. This is called resource allocation and
is similar to the one proposed by Haupt et al., 2011 and Wei and Hero, 2012. The resulting read
datasets, one per subgroup, are then co-assembled and compared using HyDA to decide pairwise
subsumption of subgroups. The cells in those subgroups that are subsumed by other subgroups,
even in previous iterations, are eliminated from further analysis. This procedure is continued until
each remaining group contains only one cell. The resulting non-redundant single element groups
capture all of the distinct genome sequences.
3
3
Methods
Although distinct genomes are often identified as different species, there are numerous cases where
distinct genomes are categorized as varied instances of the same species or even the same strain.
Instead of identification of strains and species which are currently phenotypic notions, the goal
of our approach is to find all distinct genomes in a sample. We define two genome sequences to
be distinct if the ratio of their differences over the whole genome size is above a threshold. That
threshold is input by the user and controls a trade-off between sensitivity and specificity.
Co-assembly and comparison of multiple input read datasets lie at the core of both approaches
we take in this paper. While there are assembly tools for single cell genomic data, such as SPAdes
(Bankevich et al., 2012) and IDBA-UD (Peng et al., 2012), and also co-assembly tools for normal
multicell genomic data such as Cortex (Iqbal et al., 2012), we use HyDA which is the only tool to
date that has both functionalities (Movahedi et al., 2012). However, the novel ideas proposed here
can also be implemented using other assemblers. For the sake of completeness, HyDA algorithm is
summarized in the following.
3.1
3.1.1
Co-assembly Algorithm
Construction and Condensation of the Colored De Bruijn Graph
The colored de Bruijn graph, a variation of the standard de Bruijn graph, is a combinatorial
structure that can be used to assemble a number of input read datasets, each represented by a
color, superimposed on a single de Bruin graph (Iqbal et al., 2012). The output of such assembly
methods is a number of assembled sequences (contigs) and the corresponding average multiplicity
in each color. Our de Bruijn graph of the input reads is stored in a hashed collection of splay
trees whose vertices are k-mers with an array of multiplicity counts (one entry per color), in- and
out-edges, and internal flags. A 1-in 1-out chain of k-mers is condensed into an equivalent long
sequence which is called a unitig. A maximal unitig, which cannot be extended further due to
a branch in the graph, is a contig. Note that in HyDA, our condensation is solely based on the
topology of the graph without any attention to the colored multiplicities. Ignoring multiplicities
for condensation is purposefully done, and constitutes the feature that allows the assembler to deal
with black out regions in single cell multiple displacement amplification (Chitsaz et al., 2011).
Contigs with low coverage are often caused by sequencing error. The low coverage contig
removal process is iterated with an increased cutoff in each round. In each iteration, those contigs
whose maximum coverage over all colors is less than the cutoff are eliminated, and the remaining
graph is recondensed. This causes some contigs to merge into larger ones with recomputed average
coverages. This process is similar to Velvet-SC’s low coverage contig removal, but instead of
considering one average coverage per contig, HyDA considers the maximum of average coverages
for all the colors of each contig (Chitsaz et al., 2011). In this case, only those contigs that have
low coverage in all colors are considered erroneous and removed. Another possible approach is to
eliminate those contigs for which the mean of average coverages for all colors is less than the cutoff.
However, if we were to follow this approach, a contig that is well covered in one color but is poorly
covered or absent in hundreds of colors would be lost since the mean would dilute the effect of
coverage in one color among hundreds of colors. This approach would not work for us here because
our goal is to be able to preserve rare contigs.
4
3.1.2
Redundancy Removal
To remove redundant genomes, we define a relation that is reminiscent of subset relation on the set of
contigs for each color. Note that our goal here is to remove redundant genomes, which are collections
of contigs, rather than to remove individually redundant contigs. Let A = {a1 , a2 , . . . , ar } be the
set of remaining contigs after iterative error removal. Let Mj (ai ) denote the average coverage of
contig ai in color j, for 1 ≤ i ≤ r and 1 ≤ j ≤ s, where s is the number of colors. Pick ǫ ≥ 0 and let
Aj = {ai ∈ A | Mj (ai ) > ǫ} ⊂ A be the set of contigs for each color j. The parameter ǫ determines
the trade-off between specificity and sensitivity. We chose ǫ = 0 in this study, but a non-zero ǫ
might be needed if there are erroneous or contaminant k-mers in one color that also occur in the
true genomic sequence of another color.
We define Dτ (Ai , Aj ) ∈ R on the set F = {A1 , A2 , . . . , As } as:
Dτ (Ai , Aj ) = τ −
kAi \Aj k
,
kAi k
(1)
in which Ai \Aj = {a ∈ Ai |a ∈
/ Aj }, and k · k denotes the total assembly size. We define:
Ai τ Aj iff 0 ≤ Dτ (Ai , Aj ),
(2)
in which τ ≥ 0. Particularly, 0 becomes the subset relation and can be used to detect and remove
redundant collections of contigs, i.e. those that are subsumed by a larger collection. However
in reality, the mathematical subset relation is not adequate as there are various types of noise
including sequencing errors, intrastrain variations such as SNPs and indels, contaminations added
in the amplification and sequencing process, and lack of coverage in some regions caused by the
MDA. Hence, the definition of subset should be loosened by choosing a small but nonzero value
for τ . Beside ǫ, the value of τ gives a trade-off between specificity and sensitivity of recognizing
distinct genomes. If τ is small, the algorithm detects two equivalent genomes as distinct, and if
τ is large, distinct genomes are declared equivalent. To see how τ is chosen, refer to Sections 3.2
and 4.2. The results are shown in Table 3. Finally, we compute a non-redundant set of assemblies
E = {Ai1 , Ai2 , . . . , Ait } ⊆ F such that for every distinct pair 1 ≤ a, b ≤ t, Aia 6τ Aib and
Aib 6τ Aia .
3.2
Divide and Conquer Strategy Exploiting Sparsity
Let n be the number of cells in the sample, and denote the cells by S i , i = 1, . . . , n. Our algorithm
aims to assemble all of the distinct genomes and identify at least one cell per distinct genome. To
reach this goal, our algorithm iteratively pools samples of amplicons from different cells, tags each
pool with a unique barcode, mixes the barcoded pools, and has the result sequenced. The objective
is to minimize the total number of bases required to be sequenced as well as the number of different
barcodes needed.
In the first iteration, we divide the n cells S 1 , . . . , S n into two sets
I1,1 = {S 1 , . . . , S ⌊n/2⌋ },
I1,2 = {S ⌊n/2⌋+1 , . . . , S n }.
Our algorithm samples equal amount of amplicons from each cell in I1,1 and I1,2 . The amplicons in
each set are pooled and tagged by two distinct barcodes. The barcoded amplicons are sequenced
to reach a desired number of base pairs. This number is an input parameter of our algorithm.
We define the total number of base pairs sequenced from I1,i to be b1,i , for i = 1, 2. The two
5
Table 1: The 9 chosen species for our simulation.
NCBI ID
NC 004663.1
NC 009614.1
NC 009615.1
NC 008532.1
NC 016776.1
FP929046.1
FP929051.1
FP929053.1
FP929055.1
Name
Bacteroides thetaiotaomicron VPI-5482 chromosome
Bacteroides vulgatus ATCC 8482 chromosome
Parabacteroides distasonis ATCC 8503 chromosome
Streptococcus thermophilus LMD-9
Bacteroides fragilis 638R
Faecalibacterium prausnitzii SL3/3
Ruminococcus bromii L2-63
Ruminococcus sp. SR1/5
Ruminococcus torques L2-14
Ref. Status
complete
complete
complete
complete
complete
draft
draft
draft
draft
Size
6.29
5.16
4.81
1.86
5.37
3.21
2.25
3.55
3.34
(bps)
M
M
M
M
M
M
M
M
M
No. of Cells
23
7
8
2
1
12
35
12
15
read datasets are co-assembled by HyDA using two colors. The result is two sets of contigs for each
color, A1,1 and A1,2 . We calculate Dτ1 (A1,1 , A1,2 ) and Dτ1 (A1,2 , A1,1 ) as defined in (1), in which
τ1 = τ / maxj |I1,j |, τ is an input parameter, and | · | is the set cardinality. We choose τ to be the
maximum allowable difference between the assembly of two single cells from the same strain. Based
on these values, we decide if the relations A1,1 τ1 A1,2 and A1,2 τ1 A1,1 hold. If A1,1 is a subset
of A1,2 , then all of the distinct genomes in I1,1 are present in I1,2 , therefore, the cells in I1,1 do not
need further sampling. This applies to I1,2 too, if A1,2 is a subset of A1,1 . If both relations hold,
one of I1,1 or I1,2 is eliminated arbitrarily from further analysis. Each remaining set I1,· is divided
into two subsets for analysis in the second iteration. Fig. 1 depicts an example of 10 cells with 3
distinct genomes shown in different colors.
The same splitting process occurs in the subsequent iterations. Assume Ii,1 , . . . , Ii,mi are the
remaining sets in iteration i. Each set Ii,j is sampled to produce bi,j base pairs, barcoded uniquely,
pooled, and sequenced. All of the new sequence datasets and those obtained in all previous iterations
are co-assembled. In the co-assembly, the previous datasets help to improve the assembly of the new
ones. The resulting contig set of Ii,j is denoted by Ai,j . For all j, k = 1 . . . mi , the relations Ai,j τi
Ai,k are evaluated, where τi is a threshold whose calculation will be explained below. The cells in
those Ii,j whose assemblies are subsumed will be removed from further analysis. All the remaining
ones are partitioned into two disjoint subsets. Denote the new subsets by Ii+1,1 , . . . , Ii+1,mi+1 . Note
that in iteration i, mi unique barcodes are needed. Therefore,
m = max mi
i
(3)
is the maximum number of barcodes required for the entire algorithm.
Parameters bi,j play a key role in the algorithm. We propose an adaptive calculation of bi,j to
minimize, without losing accuracy, the total base pairs sequenced:
X
bi,j .
(4)
b=
i,j
Assume Ii,j is a set that is created by dividing the set Ii−1,k in iteration i − 1 into two. We are
motivated to choose the total number of sequenced base pairs from cells in Ii,j to be proportional
to the total assembly size kAi−1,k k, i.e.
bi,j = c × kAi−1,k k,
(5)
where c is an input parameter indicating the estimated average coverage of each iteration. We say
Ai,j are accurate enough if the partial order relation τi can be assessed accurately. If c is large
6
and the assembly of iteration i − 1 is accurate enough, then in iteration i, adequate base pairs are
sequenced to allow an accurate enough assembly of set Ii,j .
Another factor that affects accuracy of the assessment of these relations is the choice of τi .
The threshold τi in the ith iteration is used to detect cells with similar genomes in spite of small
differences in their assemblies. We propose to use the following threshold:
τi =
τ
max1≤j≤mi |Ii,j |
.
(6)
Recall that τ is the maximum allowable difference between the assembly of two single cells with
similar genome sequences. To account for the worst possible case, it is assumed that there are |Ii,j |
distinct genomes in each group Ii,j . Therefore, max1≤j≤mi |Ii,j | captures the maximum number
of distinct genomes in Ii,j from any Ii,k . With these assumptions, τi is a conservative threshold.
This threshold will guarantee that two distinct genomes are detected, but it has the possibility
of detecting similar genomes as distinct. In the last iteration of the algorithm, when every group
consists of one cell, the threshold is τ . Note that the number of iterations, which is the number of
sequencing rounds, is always ⌈log2 n⌉.
3.3
Implementation
We implemented our algorithm in a tool called Squeezambler 1.0 in C++. Our tool and datasets
are available under http://compbio.cs.wayne.edu/software/squeezambler/
4
Results
Since we did not have access to real data, we tested our algorithm using simulated data. We
used our tool MDAsim 1.0 (Taghavi and Draghici, 2012) to simulate 100 multiple displacement
amplification processes (one process per cell) from 9 distinct genomes. The output of MDAsim 1.0
was fed into an Illumina read generator, ART (Huang et al., 2012), to generate Illumina reads, with
realistic errors, from the simulated amplicons. The set of generated reads for each cell were treated
like a microfluidic droplet from which samples without replacement are extracted in each iteration
of Squeezambler 1.0. We assume that MDA products are contamination free, which requires a
contaminant free automated microfluidic cell sorting, amplification, and sampling device.
4.1
Datasets
Totally 115 MDAs were simulated from 9 distinct genomes chosen from the list of species found
in a gut metagenomics study (Qin et al., 2010), that have a complete or draft reference genome.
Recall that we are simulating MDA from a reference genome, therefore, we needed a reference
genome for the chosen species. Table 1 summarizes the NCBI ID, name, size, reference status
(complete or draft), and the number of cells we have simulated. The number of cells is approximately
proportional to the abundance mean of the corresponding species in (Qin et al., 2010). We ran
MDAsim 1.0 with a diverse set of parameters, one for each cell, to capture the diverse nature of
MDA coverage bias. ART, an Illumina read generator, was then deployed to generate 100 bps
Illumina reads from the simulated amplicons. The amplification gain of MDAsim 1.0 was 300×
and that of ART was 8× from which 1/8 were selected randomly to obtain a total gain of 300×.
We assembled the dataset of each cell individually with HyDA 1.0, and the resulting assemblies
have between 0.1% and 4.2% missing reference bases measured by Gage (Salzberg et al., 2012),
which is similar to the real world situation with a successful MDA reaction (Chitsaz et al., 2011).
7
We made an error profile that matches the error statistics of Illumina HiSeq 2000 for ART. Using
our profile, ART injects on average 1% error into the reads, because of which we need to eliminate
erroneous contigs in the assembler. HyDA 1.0, and also its predecessor Velvet-SC, have an iterative
algorithm to remove low coverage contigs, which is explained in Section 3.1.1. In each iteration,
Squeezambler 1.0 provides HyDA 1.0 with a coverage cutoff as a percentage of the mean coverage
of each color. That percentage is constant in all iterations.
We designed three collections of cells, the statistics of which are summarized in Table 2. In the
first collection, there are 62 cells with 6 distinct genomes. In this collection, we put 22 different MDA
results from NC 004663.1 and 22 from FP929051.1 to play the role of highly abundant genomes in a
sample as well as 1 from NC 016776.1 and 2 from NC 008532.1 to represent low abundance genomes
in the same sample. In the first collection, the number of distinct genomes is approximately one
tenth of the number of cells but with a wide range of abundances. The second collection is the
sparsest collection among the three, where the number of distinct genomes is approximately one
twentieth of the number of cells. The third collection is the most diverse of the three, where the
number of distinct genomes is approximately one sixteenth of the number of cells.
Table 2: Our simulation setups: (i) 62 cells; 6 species, (ii) 97 cells; 5 species, and (iii) 112 cells; 7
species.
NCBI ID
NC 004663.1
NC 009614.1
NC 009615.1
NC 008532.1
NC 016776.1
FP929046.1
FP929051.1
FP929053.1
FP929055.1
4.2
62 cells; 6 species
22
36%
7
11%
8
13%
2
3%
1
1%
0
0%
22
36%
0
0%
0
0%
Abundance (%)
97 cells; 5 species
23
24%
0
0%
0
0%
0
0%
0
0%
12
12%
35
36%
12
12%
15
16%
112 cells; 7 species
23
21%
7
6%
8
7%
0
0%
0
0%
12
11%
35
31%
12
11%
15
13%
Simulation Results
We ran Squeezambler 1.0 for the three collections described above, the results of which are summarized in Table 3. Most of the Squeezambler 1.0 parameters were the same for all three collections. The assembly inclusion threshold constant per cell was chosen τ = 0.2 which means at most
20% of the assembly can vary among multiple instances of the same genome. Taking into account
the genomic sequence loss caused by the MDA, sampling of the amplicons, and sequencing errors,
20% is a reasonable choice. This is not an optimized value and is chosen based on the authors’
intuition. We chose τ conservatively in this work so that distinct genomes are not lost but some
equivalent genomes are detected as distinct. Finding the optimal value for τ needs a thorough
study which is beyond the scope of this paper.
The k-mer size for HyDA 1.0 was chosen to be k = 55 which is a common choice for the chosen
Illumina error profile (Chitsaz et al., 2011). The coverage cutoff, as a percentage of the coverage
mean, was chosen to be 100%, and the minimum contig length was 100 bps for HyDA 1.0. The
coverage mean is estimated based on the assembly size in the first iteration, which is often larger
than the actual genome size due to a myriad of erroneous low coverage k-mers. This causes the
initially estimated coverage mean to be a small fraction of the final coverage mean after error
removal.
8
Table 3: Squeezambler 1.0 results for the three setups summarized in Table 2. For some methods
we report the results for two different values of initial sequencing coverage per cell.
Setup
Method
62 cells;
6 distinct
genomes
97 cells;
5 distinct
genomes
single cell co-assembly
divide and conquer
Sequencing
per Cell∗ in the
1st Iteration
(bps)
63 M
7M
single cell co-assembly
single cell co-assembly
divide and conquer
divide and conquer
single cell co-assembly
single cell co-assembly
divide and conquer
63 M
31 M
1M
7M
63 M
31 M
1M
112 cells;
7 distinct
genomes
∗
b1,j /|I1,j |.
∗∗
b in (4).
∗∗∗
m in (3).
Total
Sequencing∗∗
(bps)
Max
Barcodes∗∗∗
3.9 G
3.0 G
6.1
3.0
3.2
2.9
7.1
3.5
7.1
G
G
G
G
G
G
G
Iterations
62
10
No. of
Predicted
Distinct
Genomes
6
8
97
97
17
10
112
112
33
5
11
5
6
7
14
11
1
1
7
7
1
1
7
1
6
Squeezambler 1.0 has an option to choose the number of initial groups in the first iteration,
g. If g is chosen to be equal to the number of cells, then Squeezambler 1.0 simulates the single
cell co-assembly of all the cells. If g = 2, then Squeezambler 1.0 simulates the divide and conquer
algorithm described in Section 3.2. Although experimenting with different g values may improve
the results, we do not have data for it.
Before any sequencing is done, the algorithm has no idea about the genome sizes, various distinct
genomes, and abundance of each genome. Therefore, an unbiased algorithm has to sequence from
each cell exactly the same amount right in the 1st iteration. That amount in our algorithm, denoted
by b1,j /|I1,j |, is an input parameter to Squeezambler 1.0 which was chosen to be between 1 Mbps
and 63 Mbps as reported in the third column of Table 3. In our setup, the size of the 9 distinct
genomes varies between 1.8 Mbps and 6.3 Mbps; see Table 1. Therefore, 1 Mbps sequencing
provides between 1/6× and 1/2× coverage, and 63 Mbps sequencing provides between 10× and
35× coverage.
The input parameter c, which controls the amount of sequencing in subsequent rounds, was
chosen to be c = 10, which means 10× expected coverage from each genome in each collection. We
observed that in practice 10× coverage of each distinct genome provides sufficient information for
reliable evaluation of . This is consistent with the Lander and Waterman, 1988 analysis, in which
the statistics of gaps and contigs in terms of coverage is characterized. Based on that analysis, 10×
coverage yields the entire genome without gap with high probability.
Our divide and conquer algorithm exhibits significant improvement in maximum barcodes, and
in most cases the total number of base pairs sequenced, over the single cell co-assembly method.
For the 97 cells, 5 distinct genomes collection, our divide and conquer algorithm requires only 2.9
Gbps sequencing and 10 barcodes in comparison to 3.0 Gbps sequencing and 97 barcodes consumed
by the single cell co-assembly method. Similarly for the 62 cells, 6 distinct genomes collection, our
divide and conquer algorithm requires only 3.0 Gbps sequencing and 10 barcodes in comparison to
3.9 Gbps sequencing and 62 barcodes required by the single cell co-assembly method. Even though
for the 112 cells, 7 distinct genomes collection, our divide and conquer algorithm outperforms single
cell co-assembly in terms of the number of barcodes, by 33 vs. 112, it requires 7.1 Gbps sequencing
which is more than that used by single cell co-assembly (3.5 Gbps).
In all of our experiments, all distinct genomes were correctly detected. Therefore, our results
9
exhibit ultimate sensitivity. However, in some experiments multiple cells with similar genomes
were identified as distinct, which is not an issue for our problem because based on the results of
Squeezambler 1.0, those cells that are identified as with distinct genomes can be deeply sequenced
and assembled afterwards. For the 62 cells, 6 distinct genomes collection, the number of detected
distinct genomes was between 6 and 8. For the 97 cells, 5 distinct genomes collection, that number
was between 5 and 11, and for the 112 cells, 7 distinct genomes collection, that was between 7 and
14. This specificity is reported in the sixth column of Table 3. Note that the number of sequencing
rounds (iterations) for single cell co-assembly is always 1, and for divide and conquer with g = 2,
it is ⌈log2 n⌉.
Due to the computational intensity of MDAsim 1.0, HyDA 1.0, and Squeezambler 1.0, we
report our results for only small examples in order to provide a proof of concept. We also chose
our parameters conservatively, and without optimization, so that we do not compromise accuracy.
Moreover, our examples have in the order of 100 cells and 6 distinct genomes, whereas real world
samples are much sparser as the number of cells may be in the order of billions and the number of
distinct genomes at most in the order of thousands. Therefore, we expect the reduction in the total
sequencing and maximum barcodes to be higher for real world applications than what we report
in this paper.
5
Conclusion
We presented an adaptive divide and conquer algorithm for distilled sequencing and de novo assembly of distinct genomes in a bacterial community, e.g. human gut microbiome. Samples derived
from such communities are often sparse in the sense that the number of distinct genomes is much
less than the number of cells. Our algorithm exploits sparsity to decrease the amount of sequencing
and the number of multiplexing barcodes needed for single cell sequencing and de novo assembly.
We implemented our algorithm in a tool which we call Squeezambler and performed simulation
experiments to demonstrate its power. Our results show that: i) the number of required barcodes
with our divide and conquer algorithm is less than that required by the naı̈ve approach, and that
ii) the amount of sequencing needed remains the same or decreases. Due to the computational
intensity of the problem, only small examples with low sparsity were studied in this work. Real
world samples are much sparser (∼ 1000 species in ∼ 1014 cells) than the examples here (∼ 5 species
in ∼ 100 cells). Also, the parameters used to run our tool were chosen conservatively and without
optimization. Therefore, we expect the improvement of our algorithm to be higher than what we
reported in this paper in real world situations. Squeezambler 1.0 identifies all distinct genomes
in the sample which are candidates for different strains/species. Those cells that are identified
as having distinct genomes need to be subsequently deeply sequenced and assembled in order to
obtain a more detailed assembly.
Acknowledgement
This work has been partially supported by the following grants: NIH RO1 DK089167, NIH STTR
R42GM087013, and NSF DBI-0965741 (to S.D.). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily represent
the views of any of the funding agencies.
10
References
Bankevich, A., et al. (2012). SPAdes: A New Genome Assembly Algorithm and Its Applications
to Single-Cell Sequencing. J. Comput. Biol., 19(5), 455–477.
Candès, E. and Tao, T. (2005). Decoding by linear programming. IEEE Transactions on Information Theory, 51(12), 4203–4215.
Candès, E. and Tao, T. (2006). Near-optimal signal recovery from random projections: Universal
encoding strategies? IEEE Transactions on Information Theory, 52(12), 5406–5425.
Chitsaz, H., et al. (2011). Efficient de novo assembly of single-cell bacterial genomes from short-read
data sets. Nature Biotech, 29(10), 915–921.
Dean, F. B., et al. (2001). Rapid amplification of plasmid and phage DNA using Phi 29 DNA
polymerase and multiply-primed rolling circle amplification. Genome Res., 11, 1095–1099.
Dean, F. B., et al. (2002). Comprehensive human genome amplification using multiple displacement
amplification. Proc. Natl. Acad. Sci. U.S.A., 99, 5261–5266.
Donoho, D. (2006). Compressed sensing. Information Theory, IEEE Transactions on, 52(4), 1289
–1306.
Erlich, Y., et al. (2010). Compressed genotyping. Information Theory, IEEE Transactions on,
56(2), 706 –723.
Haupt, J., et al. (2011). Distilled sensing: Adaptive sampling for sparse detection and estimation.
IEEE Transactions on Information Theory, 57(9), 6222–6235.
Hongoh, Y., et al. (2008). Complete genome of the uncultured Termite Group 1 bacteria in a single
host protist cell. Proc. Natl. Acad. Sci. U.S.A., 105, 5555–5560.
Hosono, S., et al. (2003). Unbiased whole-genome amplification directly from clinical samples.
Genome Res., 13, 954–964.
Huang, W., et al. (2012). ART: a next-generation sequencing read simulator. Bioinformatics,
28(4), 593–594.
Iqbal, Z., et al. (2012). De novo assembly and genotyping of variants using colored de bruijn graphs.
Nat Genetics, 44, 226 – 232.
Ishoey, T., et al. (2008). Genomic sequencing of single microbial cells from environmental samples.
Curr. Opin. Microbiol., 11, 198–204.
Kvist, T., et al. (2007). Specific single-cell isolation and genomic amplification of uncultured
microorganisms. Appl. Microbiol. Biotechnol., 74, 926–935.
Lander, E. S. and Waterman, M. S. (1988). Genomic mapping by fingerprinting random clones: a
mathematical analysis. Genomics, 2, 231–239.
Lasken, R. S. (2007). Single-cell genomic sequencing using Multiple Displacement Amplification.
Curr. Opin. Microbiol., 10, 510–516.
11
Lasken, R. S. and Stockwell, T. B. (2007). Mechanism of chimera formation during the Multiple
Displacement Amplification reaction. BMC Biotechnol., 7, 19.
Marcy, Y., et al. (2007). Dissecting biological ”dark matter” with single-cell genetic analysis of rare
and uncultivated TM7 microbes from the human mouth. Proc. Natl. Acad. Sci. U.S.A., 104,
11889–11894.
Methe, B. A., et al. (2012). A framework for human microbiome research. Nature, 486(7402),
215–221.
Movahedi, N. S., et al. (2012). De novo co-assembly of bacterial genomes from multiple single cells.
In IEEE Conference on Bioinformatics and Biomedicine, pages 561–565.
Mussmann, M., et al. (2007). Insights into the genome of large sulfur bacteria revealed by analysis
of single filaments. PLoS Biol., 5, e230.
Peng, Y., et al. (2012). IDBA-UD: a de novo assembler for single-cell and metagenomic sequencing
data with highly uneven depth. Bioinformatics, 28(11), 1420–1428.
Podar, M., et al. (2007). Targeted access to the genomes of low-abundance organisms in complex
microbial communities. Appl. Environ. Microbiol., 73, 3205–3214.
Qin, J., et al. (2010). A human gut microbial gene catalogue established by metagenomic sequencing. Nature, 464(7285), 59–65.
Raghunathan, A., et al. (2005). Genomic DNA amplification from a single bacterium. Appl.
Environ. Microbiol., 71, 3342–3347.
Rodrigue, S., et al. (2009). Whole genome amplification and de novo assembly of single bacterial
cells. PLoS ONE , 4, e6864.
Salzberg, S. L., et al. (2012). GAGE: A critical evaluation of genome assemblies and assembly
algorithms. Genome Res., 22(3), 557–567.
Stobbe, P. and Krause, A. (2012). Learning fourier sparse set functions. Journal of Machine
Learning Research - Proceedings Track , 22, 1125–1133.
Taghavi, Z. and Draghici, S. (2012). Mdasim: A multiple displacement amplification simulator. In
IEEE Conference on Bioinformatics and Biomedicine, pages 575–578.
Treangen, T. J., et al. (2013). MetAMOS: a modular and open source metagenomic assembly and
analysis pipeline. Genome Biol., 14(1), R2.
Wei, D. and Hero, A. (2012). Multistage adaptive estimation of sparse signals. In IEEE Statistical
Signal Processing Workshop (SSP), pages 153–156.
Woyke, T., et al. (2009). Assembling the marine metagenome, one cell at a time. PLoS ONE , 4,
e5299.
Zhang, K., et al. (2006). Sequencing genomes from single cells by polymerase cloning. Nat. Biotechnol., 24, 680–686.
12
| 5cs.CE
|
Virtual-to-Real: Learning to Control in Visual Semantic Segmentation
Zhang-Wei Hong1 , Yu-Ming Chen1 , Hsuan-Kung Yang1 , Shih-Yang Su1 , Tzu-Yun Shann1 ,
Yi-Hsiang Chang1 , Brian Hsi-Lin Ho1 , Chih-Chieh Tu1 , Tsu-Ching Hsiao1 , Hsin-Wei Hsiao1 ,
Sih-Pin Lai1 , Yueh-Chuan Chang2 , and Chun-Yi Lee1
arXiv:1802.00285v2 [cs.CV] 18 Mar 2018
1
Elsa Lab, Department of Computer Science, National Tsing Hua University, Hsinchu, Taiwan
2
National Chiao Tong University, Hsinchu, Taiwan
Abstract
Both modular and end-to-end learning-based approaches
have been proven effective in a variety of vision-based robotic
control tasks [Sadeghi and Levine, 2016; Finn and Levine,
2017; Gupta et al., 2017; Smolyanskiy et al., 2017; Zhu et
al., 2017]. A modular cognitive mapping and planning approach has been demonstrated successful in first-person visual navigation [Gupta et al., 2017]. Vision based reinforcement learning (RL) has been attempted to train an end-to-end
control policy for searching specific targets [Zhu et al., 2017].
Applying end-to-end supervised learning to navigate a drone
along a trail with human-labeled image-action pairs is presented in [Smolyanskiy et al., 2017]. A methodology of endto-end training a robot for object manipulation tasks using
unlabeled video data is described in [Finn and Levine, 2017].
While these learning-based approaches seem attractive, they
typically require a huge amount of training data. Collecting training data for learning a control policy in the physical world is usually costly and poses a number of challenges.
First, preparing large amounts of labeled data for supervised
learning takes considerable time and human efforts. Second,
RL relies on trial-and-error experiences, which restrict fragile
robots from dangerous tasks. Online training and fine-tuning
robots in the physical world also tend to be time-consuming,
limiting the learning efficiency of various RL algorithms.
An alternative approach to accelerate the learning efficiency and reduce the cost is training robots in virtual worlds.
Most of the recent works on robot learning collect training data from simulators [James and Johns, 2016; Rusu et
al., 2016; Sadeghi and Levine, 2016; Peng et al., 2017;
Tobin et al., 2017; Zhu et al., 2017]. However, the discrepancies between virtual and real worlds prohibit an agent
trained in a virtual world from transferring to the physical
world directly [James and Johns, 2016]. Images rendered
by low-fidelity simulators are unlikely to contain as much
rich information as real ones. Therefore, bridging the reality gap [Tobin et al., 2017] has been a challenging problem
in both computer vision and robotics. Many research efforts
have been devoted to tackling this problem by either domain
adaption (DA) [Rusu et al., 2016; Ghadirzadeh et al., 2017;
Zhang et al., 2017] or domain randomization (DR) [Sadeghi
and Levine, 2016; Peng et al., 2017; Tobin et al., 2017;
Zhang et al., 2017]. Both of these methods train agents
by simulators. DA fine-tunes simulator-trained models with
real world-data. DR, on the other hand, trains agents with
Collecting training data from the physical world
is usually time-consuming and even dangerous for
fragile robots, and thus, recent advances in robot
learning advocate the use of simulators as the
training platform. Unfortunately, the reality gap
between synthetic and real visual data prohibits
direct migration of the models trained in virtual
worlds to the real world. This paper proposes
a modular architecture for tackling the virtual-toreal problem. The proposed architecture separates
the learning model into a perception module and
a control policy module, and uses semantic image segmentation as the meta representation for
relating these two modules. The perception module translates the perceived RGB image to semantic image segmentation. The control policy module is implemented as a deep reinforcement learning agent, which performs actions based on the
translated image segmentation. Our architecture
is evaluated in an obstacle avoidance task and a
target following task. Experimental results show
that our architecture significantly outperforms all
of the baseline methods in both virtual and real
environments, and demonstrates a faster learning
curve than them. We also present a detailed analysis for a variety of variant configurations, and validate the transferability of our modular architecture.
1
Introduction
Visual perception based control has been attracting attention
in recent years for controlling robotic systems, as visual inputs contain rich information of the unstructured physical
world. It is usually necessary for an autonomous robot to
understand visual scene semantics to navigate to a specified
destination. Interpreting and representing visual inputs to
perform actions and interact with objects, however, are challenging for robots in unstructured environments as colored
images are typically complex and noisy [Maier et al., 2012;
Biswas and Veloso, 2012]. It is especially difficult to design
a rule-based robot satisfying such requirements.
Demo video: https://goo.gl/d2R3PN
1
guidance module adjusts the meta-state representations, such
that the behavior of the robot can be altered online without
replacing either the perception or control policy modules.
To demonstrate the effectiveness of the proposed modular architecture, we evaluate it against a number of baseline
methods on two benchmark tasks: obstacle avoidance and
target following. These tasks are essential for autonomous
robots, as the former requires collision-free navigation, while
the latter requires understanding of high-level semantics in
the environments. To validate the transferability of our architecture, we train our models and the baseline methods in
our simulators, and compare their performance on the benchmark tasks in both virtual and real worlds. Please note that
we focus on RL-based robotic control policy, though the control policy module can be implemented by various options
(e.g., imitation learning). Our results show that the proposed
architecture outperforms all the baseline methods in all the
benchmark tasks. Moreover, it enables an RL agent to learn
dramatically faster than all the baseline methods. By simply
replacing the perception module, we also show that no additional fine-tuning is necessary for the control policy module when migrating the operating environment. We further
demonstrate the versatility of the visual-guidance module in
Section 5. The contributions of this work are as follows:
randomized object textures, colors, light conditions, or even
physical dynamics in virtual worlds. The diversified simulated data enable the agents to generalize their models to
the real world. Unfortunately, collecting the real-world data
required by DA for control policy learning is exceptionally
time-consuming. Although DR does not require real-world
data, the technique lacks a systematic way to determine which
parameters are to be randomized. Furthermore, DR requires
considerably large number of training samples in the training
phase to achieve acceptable performance.
Instead of training a vision-based control policy in an endto-end fashion, we propose a new modular architecture to
address the reality gap in vision domain. We focus on the
problem of transferring deep neural network (DNN) models trained in simulated virtual worlds to the real world for
vision-based robotic control. We propose to separate the
learning model into a perception module and a control policy
module, and use semantic image segmentation as the metastate for relating these two modules. These two modules are
trained separately and independently. Each of them assumes
no prior knowledge of the other module. The perception module translates RGB images to semantic image segmentations.
This module can be any semantic image segmentation models pre-trained on either commonly available datasets [Cordts
et al., 2016; Zhou et al., 2017] or synthetic images generated
by simulators [Richter et al., 2017]. As annotated datasets
for semantic segmentation are widely available nowadays and
cross dataset domain adaptation has been addressed [Chen et
al., 2017], training and fine-tuning a perception module for
segmenting outdoor [Cordts et al., 2016] or indoor [Zhou et
al., 2017] scenes have become straightforward. The control
policy module employs deep RL methods and takes semantic
image segmentations as its inputs. In the training phase, the
control policy module only receives the image segmentations
rendered by low-fidelity simulators. The RL agent interacts
with the simulated environments and collects training data for
the control policy module. While in the execution phase, the
control policy module receives image segmentations from the
perception module, enabling the RL agent to interact with the
real world. As the image semantics rendered by the simulators and those generated by the perception module are invariant [You et al., 2017], the control policy learned from the simulators can be transferred to the real world directly. The proposed architecture provides better efficiency than the conventional learning-based methods, as the simulators are able to
render image segmentations at a higher frame rate, enabling
the RL agent to collect training or trial-and-error experiences
faster. It also allows the RL agent to learn faster, as semantic
image segmentation contains less noise than raw images.
Another advantage of the proposed architectures is its
modularity. Both the perception and control policy modules
can be flexibly replaced in a plug-and-play fashion, as long
as their meta-state representation (i.e., image segmentation)
formats are aligned with each other. Replacing the perception module allows a pre-trained control policy module to be
applied to different scenes. Replacing the control policy module, on the other hand, changes the policy of the robot in the
same environment. Furthermore, our architecture allows another visual-guidance module to be incorporated. The visual-
1. A new modular learning-based architecture which separates the vision-based robotic learning model into a perception module and a control policy module.
2. A novel concept for bridging the reality gap via the use
of semantic image segmentation, which serves as the
meta-state for relating the two modules.
3. A simple methodology for directly transferring the control policy learned in virtual environments to the real
world. The methodology does not require any domain
randomization or task-specific real-world data.
4. A way to migrate the operational environment of a robot
without further fine-tuning its control policy module.
5. A visual-guidance module for altering the behavior of
the robot via adjusting the meta-state representations.
The remainder of this paper is organized as follows. Section 2 introduces background material. Section 3 walks
through the proposed modular architecture. Section 4 describes the evaluation tasks and environments. Section 5
presents the experimental results. Section 6 concludes.
2
Background
This section briefly reviews background material on semantic
image segmentation, RL, and policy gradient methods.
2.1
Semantic Image Segmentation
The goal of semantic segmentation is to perform dense predictions at pixel level. It has received great attention due
to its applicability to a number of research domains (e.g.,
visual understanding, autonomous driving, and robotic control). Fully convolutional network (FCN) [Long et al., 2015]
pioneered to replace fully-connected (FC) layers by convolutional layers. A number of successive works have further
enhanced the accuracy and efficiency [Chen et al., 2016;
Zhao et al., 2017a; Zhao et al., 2017b; Paszke et al., 2016;
2
function is expressed as st = φpercept (xreal
; θpercept ), where
t
φpercept represents the semantic segmentation model, xreal
t
is the RGB image from the real world, and θpercept denotes
the parameters of the perception module. We directly employ
existing models, such as DeepLab [Chen et al., 2016] and
ICNet [Zhao et al., 2017a], and pre-train them on datasets
ADE20K [Zhou et al., 2017] and Cityscape [Cordts et al.,
2016] for indoor and outdoor scenes, respectively. The format
of meta-state representation st is the same for both phases.
The effect of using different semantic image segmentation
models as the perception module is analyzed in Section 5.
The benefit of using semantic segmentation as the metastate is threefold. First, although the visual appearances of
xreal
and xsyn
differ, their semantic segmentations are alt
t
most identical [You et al., 2017] (Fig. 1 (a)). This allows us to
use semantic segmentation to bridge the reality gap. Second,
semantic segmentation preserves only the most important information of objects in an RGB image, resulting in a more
structured representation of a scene. It also provides much
richer information than other sensory modalities, as they typically fail to cover crucial high-level scene semantics. Third,
the perception module in the execution phase only requires a
monocular camera, which is a lot cheaper than depth cameras
or laser range finders. As a result, semantic segmentation is
inherently an ideal option for meta-state representation.
Figure 1: Model architecture
He et al., 2015], making them more promising for robotic
control tasks. Unfortunately, relatively less attempts have
been made to incorporate them into robotic learning models.
2.2
RL and Policy Gradient Methods
RL trains an agent to interact with an environment E. An
RL agent observes a state s from E, takes an action a according to its policy π(a|s), and receives a reward r(s, a).
E then transits to a new state s0 . The agent’s objective is to
maximize its accumulated rewards
t with a discounted facPT G
τ −t
tor γ, expressed as Gt =
γ
r(sτ , aτ ), where t is
τ =t
the current timestep, and T is the total number of timesteps.
Policy gradient methods are RL approaches that directly opPT
timize π in the direction: ∇π t=0 log π(at |st )(Gt − b(st )),
where b(st ) is a baseline function. A common choice for
b(st ) is the value function V π (s) = E Gt |st = s, π]. This
approach is known as the actor-critic algorithm [Williams,
1992]. An asynchronous variant of the actor-critic algorithm,
namely asynchronous advantage actor-critic (A3C) [Mnih et
al., 2016], has been proven data-efficient and suitable for vision based control tasks [Wu and Tian, 2016; Parisotto and
Salakhutdinov, 2017; You et al., 2017; Zhu et al., 2017].
3
3.2
The control policy module is responsible for finding the best
control policy π for a specific task. In this work, it employs
the A3C algorithm, and takes st as its inputs (Fig. 1 (a)). In
the training phase, the model φcontrol is trained completely
in simulated environments, and learns π directly from st
rendered by low-fidelity simulators. While in the execution
phase, it receives st from the perception module, allowing
the RL agent to interact with the real world. As semantic segmentation is used as the meta state representation, it enables
virtual-to-real transferring of π without fine-tuning or DR.
Proposed Methodology
In this section, we present the modular architecture of our
model. We first provide an overview of the model architecture, followed by the implementation details of the modules.
Fig. 1 illustrates the proposed modular architecture. It consists of a perception module, a control policy module, and
a visual guidance module. Semantic image segmentation st
serves as the meta-state for related the former two modules,
as shown in Fig. 1 (a). The perception module generates st
from an RGB input image xt , which comes from different
sources in the training (xsyn
) and execution (xreal
) phases.
t
t
The control policy module takes st as its input, and reacts
with at according to π. The visual guidance module enables
high-level planning by modifying st , as shown in Fig. 1 (b).
3.1
Control Policy Module
3.3
Visual Guidance Module
The use of scene semantics as the meta-state gives the proposed architecture extra flexibility. Our modular architecture
allows a visual-guidance module to be augmented to guide
the control policy module to perform even more complex
tasks by manipulating the meta-state st . The manipulation
can be carried out by either human or another RL agent. One
can easily modify a robot’s task through manipulating the
class labels in st . This implies that a roadway navigation
robot can be transformed to a sidewalk navigation robot by
swapping the labels of the roadway and the sidewalk in st .
The visual guidance module can also alter a target following
robot’s objective online by modifying the target label to a new
one, as shown in Fig. 1 (b) (yellow chair). Note that visual
guidance does not require any retraining, fine-tuning, or extra
data in the above scenarios, while the previous works [Gupta
et al., 2017; Zhu et al., 2017] demand additional training samples to achieve the same level of flexibility.
Perception Module
The main function of the perception module is to generate
st , and passes it to the control policy module. In the training
stage, st is rendered by a segmentation shader from a synthetic image xsyn
. While in the execution phase, the percept
tion module could be any semantic segmentation model. Its
3
Action
(Linear (m/s),
Angular (rad/s))
Forward
(0.2, 0.0)
TurnLeft
(0.2, −0.2)
TurnRight
(0.2, 0.2)
Table 1: Mapping actions & robot velocities.
4
Model
Seg (Ours)
Seg-S (Ours)
DR-A3C
DR-A3C-S
Depth-A3C
Depth-A3C-S
ResNetA3C
Dimension
84×84×3
84×84×3×4
84×84×3
84×84×3×4
84×84×1
84×84×4
224×224×3
Format
RGB
Frame
RGB
Frame
Depth
Map
RGB
Frame
Table 2: Settings of model inputs.
Figure 2: Samples of evaluation scenes. From left to right: simple
corridor, cluttered hallway, and outdoor. Top to bottom rows are
RGB image, segmentation, and depth.
Virtual-to-Real Evaluation Setup
In this section, we present the evaluation setup in detail. We
evaluate the proposed architecture both in virtual and real environments. We begin with explaining the simulator used for
training and evaluation, followed by an overview of the baseline models and their settings. Finally, we discuss the tasks
and scenarios for performance evaluation.
Simulator. We use the Unity3D 1 engine to develop the virtual environments for training our RL agent. As Unity3D
supports simulation of rigid body dynamics (e.g., collision,
fraction, etc.), it allows virtual objects to behave in a realistic way. In the simulated environment, our RL agent receives
its observations in the form of scene semantics in first-person
perspective. It then determines the control actions (linear and
angular velocities) to take accordingly. The virtual environment responds to the actions by transiting to the next state and
giving reward signals. Please note that the geometric shape of
our simulated agent is set to be similar to that of the real robot.
Tasks. The proposed model is evaluated against the baselines in virtual and real environments on two benchmark
tasks: obstacle avoidance and target following. In both tasks,
an RL agent starts at a random position in the environment
at the beginning of each episode. At each timestep, it selects
an action from the three possible options: moving forward,
turning left, and turning right. Each action corresponds to the
linear and angular velocities of the robot specified in Table 1.
(1) Obstacle Avoidance: The agent’s goal is to navigate in a
diverse set of scenes, and avoid colliding with obstacles.
(2) Target Following: The agent’s objective is to follow a
moving target (e.g., human) while avoiding collisions.
Scenarios. We evaluate the models with the following three
scenarios. Fig. 2 illustrates a few sample scenes of them.
(1) Simple corridor: This scenario features indoor straight
passages, sharp turns, static obstacles (e.g., chairs, boxes, tables, walls, etc.), and moving obstacles (e.g., human).
(2) Cluttered hallway: This scenario features a hallway
crammed with static and moving obstacles for evaluating an
agent’s capability of avoiding collision in narrow space.
(3) Outdoor: This scenario features an outdoor roadway with
sidewalks, buildings, terrain, as well as moving cars and
pedestrians. This is used to evaluate how well the control
policy trained for the tasks can be transfered from an indoor
environment to an outdoor environment. Note that the agent
is not allowed to move on the sidewalks in this scenario.
1
(a) Obstacle avoidance task
(b) Target following task
Figure 3: Learning curve comparison.
Models. For all the experiments, our model is trained with
scene semantics generated from the simulator. We adopt DR,
depth perception, and ResNet [He et al., 2016] as the baseline
models. For the DR baseline, the texture of each mesh in a
scene is randomly chosen from 100 textures. For the depth
perception baseline, we use the depth maps generated from
the simulator. We set the minimum and maximum sensible
depths of the agent in the simulator to be 0.3m and 25m, respectively, which are close to those of the depth cameras (e.g.,
ZED) mounted on robots. We summarize the detailed settings in Table. 2. All the models in Table. 2 adopt the vanilla
A3C architecture [Mnih et al., 2016] except ResNet-A3C, in
which the convolutional layers are replaced by ResNet-101
pre-trained on ImageNet [Krizhevsky et al., 2012]. ResNetA3C is mainly used to justify that directly applies features
extracted by ResNet as the meta-state does not lead to the
same performance as ours. In Table. 2, we denote our model
as Seg, and the other baselines as DR-A3C, Depth-A3C, and
ResNet-A3C, respectively. The model named with ”-S” (e.g.,
DR-A3C-S) indicate that these models concatenate the latest
four frames as their inputs. Please note that we do not include
DA in our experiments, as we only focus on control policies
solely trained in simulated environments without further finetuning. We use Adam optimizer [Kingma and Ba, 2015], and
set both the learning rate and epsilon to 0.001. Each model is
trained for 5M frames, and the training data are collected by
16 worker threads for all experimental settings. For the evaluation tasks in the real world, we train DeepLabv2 [Chen et
al., 2016] on ADE20K [Zhou et al., 2017] and ICNet [Zhao
et al., 2017a] on Cityscapes [Cordts et al., 2016] as the indoor and outdoor perception modules, respectively. Please
note that the class labels in ADE20K are reduced as Table 5
for better accuracy and training efficiency.
Unity3D: https://unity3d.com/
4
Figure 4: Evaluation results in the Figure 5: Collision rate in the ob- Figure 6: Comparison of the tar- Figure 7: Comparison of the tarobstacle avoidance tasks.
stacle avoidance tasks.
get following tasks in simulation. get following tasks in real world.
5
Experimental Results and Analysis
baseline models tend to drive on the sidewalk, or rotate in
place. In contrast, our models focus on driving on the roadway. As a result, we conclude that using semantic segmentations as the meta state enables a control policy to be transfered
from an indoor environment to an outdoor environment.
Evaluation in the Real World. To test the virtual-to-real
transferability of the models, we replicate the settings of the
scenarios mentioned above in the real world. We measure
the performance by recording the number of collisions per
minute. We manually position the robot back to a random position when it collides with an obstacle. We report the results
in Fig. 5. In Simple Corridor, we observe that Depth-A3C
experience a huge performance drop from simulation to the
real world, since the depth map estimated in the real world
is too noisy for the model to adapt. The performances of
the other baselines are similarly decreased, while our methods still maintain the same level of performance. We exclude
Depth-A3C from the rest of the real world scenarios due to
the poor quality of depth estimation of the depth cameras in
complex environments. In Cluttered Hallway, while all of
the models degrade in performance, our methods still significantly outperform the baselines. In Outdoor, we observe that
both DR-A3C and ResNet-A3C fail to transfer their knowledge from simulation to the real world, and tend to bump into
cars and walls. We also notice that frame stacking leads to
lower performance, because it amplifies the errors resulting
from the changes in environmental dynamics. Additionally,
to demonstrate the robustness of the proposed architecture to
the quality of segmentation, we compare the performance of
our methods with different segmentation models. They include PSPNet [Zhao et al., 2017b] (high) and DeepLab [Chen
et al., 2016] (low) for indoor environments, and ICNet [Zhao
et al., 2017a] (high) and ENet [Paszke et al., 2016] (low) for
outdoor environments. Fig. 5 shows that our agents are not
quite sensitive to the quality of image segmentations. Note
that in Outdoor scenario, ICNet performs worse than ENet
due to its inevitable computational delay. From the results in
Figs. 3-5, we conclude that our methods are robust, generalizable, and transferable in the obstacle avoidance tasks.
In this section, we present the experimental results and their
implications. We comprehensively analyze the results and
perform ablative study of our methodology. In Section 5.1
and 5.2, we compare the performance of our architecture and
the baseline models in the tasks mentioned in Section 4. We
demonstrate the concept of visual guidance in Section 5.3.
Finally, we present a series of ablative studies in Section 5.4
5.1
Comparison in the Obstacle Avoidance Tasks
To compare the performance of our models against those of
the baselines in the obstacle avoidance tasks, we designed a
total of 16 indoor scenes for training the agents. Each scene
features a pre-defined set of attributes (e.g., hallway, static
obstacles), along with randomly located obstacles and moving objects. A scene is randomly selected for each training
episode. Each episode terminates after 1,000 timesteps, or
when a collision occurs. The agent receives a reward of 0.001
at each timesteps. Note that all the models are trained in
simulation. Fig. 3 plots the learning curves of the models.
Figs. 3 (a) and 3 (b) show the curves for the obstacle avoidance tasks and the target following tasks, respectively. While
most of the models achieve nearly optimal performance at
the end of the training phase, our methods learn significantly
faster than the baseline models. This is due to the fact that
scene semantics have simplified the original image representations to structured forms. On the contrary, the poor performance of ResNet-A3C indicates that the deep representations
extracted by ResNet does not necessarily improve the performance. In addition, we observe that frame stacking has
little impact on performance in the training phase. DepthA3C shows a slightly better performance than DR-A3C and
ResNet-A3C, implying the usefulness of depth information.
Evaluation in Simulation. Fig. 4 compares the mean rewards of the agents in three scenarios for seven types of models. The mean rewards are evaluated over 100 episodes, and
the scenes are different from those used at the training phase.
It can be observed that our methods outperform the other
baseline models in all of the three scenarios. In Simple Corridor, noticeable performance can be observed from the other
models. Most of the agents navigate well along the corridor, where there are relatively fewer obstacles compared to
the other scenarios. In a more challenging scenario Cluttered
Hallway, it can be seen that both Depth-A3C and DepthA3C-S experience a considerable drop in performance. This
is because depth map becomes too noisy in resolution to navigate in narrow space. In Outdoor, all models receive negative
mean rewards, except for our methods. We observe that the
5.2
Comparison in the Target Following Tasks
We further compare the performance of our methods against
those of the baselines in the target following tasks. For each
episode, a simulated scene featuring static and moving obstacles is similarly selected at random. An episode ends immediately after 1,000 timesteps, or when a collision with the target/obstacles occurs. The agent receives a reward of 0.001 at
each timestep if the target is within its sight, (1.0-”cumulative
5
reward”) if it touches the target, and 0.0 otherwise. At the beginning of each episode, the target is located in front of the
agent. The target then navigates to a randomly chosen destination along the path computed by the A* algorithm. The
ultimate goal of the agent is to avoid any collision on the way,
and keep up with the target. Note that for a fair comparison,
the geometry of the target is always fixed. Fig. 3 (b) plots the
learning curves of our models and the baselines in the training
phase. It can be observed that our models are much superior
to the other baseline models. We also notice that our agents
learn to chase the target person in the early stages of the training phase, while the baseline models never learn to follow
the target. We conclude that the superiority of our models
over the baselines results from semantic segmentation, which
makes the target easily recognizable by the agent.
(a) Switching target to person
Figure 8: Visual guidance for switching the following target.
Seg
Seg-S
VW (mean reward)
0.824
0.925
RW (success rate)
80%
90%
Table 3: Performance of visual guidance in the virtual world (VW)
and real world (RW) in the Switching-Target Following task.
experiments in both virtual and real environments. Both environments are the same as the ones in the Target Following
task. To demonstrate visual guidance, we randomly select an
object in the environment, and re-label it as the target at the
beginning of each episode. The initial location of the agent is
chosen such that the selected target falls in the field of view
(FOV) of the agent. The selected object is rendered to yellow,
as shown in Fig. 8. Fig. 8 illustrates the procedure of switching the target from a human to a chair. We measure the mean
rewards over 1,000 episodes in the virtual environments, and
the success rate over 10 episodes in the real world. Table 3
summarizes the results. The results reveal that our agents can
successfully catch the randomly specified targets in both environments, no matter what target size and shape are.
Evaluation in Simulation. To validate the generalizability,
we evaluate the agent in a virtual scene, which is not included
in the training sets. In each episode, the agent start from a
random position in the scene. We evaluate the mean rewards
over 100 episodes, and present the results in Fig. 6. Our
models achieve higher mean rewards than all the baselines,
which suggests the effectiveness of our models in capturing
and tracking a fixed moving object. We observe that DR-A3C
and ResNet-A3C always fail to follow the target at the fork of
a corridor, because they are easily distracted by other objects.
On the other hand, although Depth-A3C has higher mean rewards than the other baselines, it often fails when the target is
far away from it. From the results in Fig. 6, we conclude that
our models are generalizable and transferable to new scenes.
5.4
Robustness to Redundant Labels
We further show that our control policy module can retain its
performance, even if its input contains redundant labels. We
conduct obstacle avoidance experiments in Cluttered Hallway, as it is the most complex scenario for our agents. When
training in simulators, we add redundant class labels in semantic segmentations rendered by simulators. Given the
same training time, the agent is able to obtain similar performance as the results shown in Fig. 3 (a). For evaluation
in the real world, we compare the agent trained with redundant labels and the one used in Section 5.1 (i.e. trained with
reduced labels). For perception in the real world, we adopt
PSPNet as the perception module for both agents, and modify the output labels to be matched with the labels rendered by
simulators. The experimental results in Table 4 show that the
agents with redundant labels demonstrate better performance
than their counterparts shown in Section 5.1. We attribute
this improvement to the richer context of the scenes in image
segmentations with more labels.
Evaluation in the Real World To evaluate if a learned policy can be successfully transfered to the real world, we further conduct experiments in real indoor environments, which
consist of corridors, forks, sharp turns, and randomly placed
static obstacles. The target is a human who moves within the
environment. It is considered failed if the agent loses the target or collides with an obstacle, and successful if the agent
follows the target to the destination. We report the success
rate for each model in Fig. 7. The success rate is evaluated
over ten episodes. The results show that all the baseline models fail in real world scenarios. Specifically, as the models
trained with DR never learn a useful policy even in the simulated training environments, there is no doubt that they fail
in the real world tests. As for the cases of Depth-A3C and
Depth-A3C-S, despite the high mean rewards they attained in
the training phase, the learned policies do not transfer successfully to the real world. During the evaluation, we observe
that DR-A3C and ResNet-A3C agents treat the target person
as an obstacle, and avoid the target as it comes close. We also
notice that our models do not lose track of the target even at
sharp turns. We deduce that this is due to the temporal correlations preserved by frame stacking. Fig. 7 also provides
evidence that our architecture is robust to the quality of image segmentations even in the target following tasks.
5.3
(b) Switching target to chair
6
Conclusion
In this paper, we presented a new modular architecture for
transferring policies learned in simulators to the real world
for vision-based robotic control. We proposed to separate the
model into a perception module and a control policy module,
and introduce the concept of using semantic image segmentation as the meta state for relating these two modules. We
trained our model with a standard RL algorithm, and did not
apply any domain randomization technique. We performed
experiments in two benchmark tasks: an obstacle avoidance
task and a target following task, and demonstrated that our
proposed method outperforms the baseline models in both
Demonstration of Visual Guidance
In this section, we demonstrate the effectiveness of the visual
guidance module of our modular architecture by switching
the target in a Switching-Target Following task. We perform
6
Model
ReducedSeg
ReducedSeg-S
RedundantSeg
RedundantSeg-S
Collision
rate
0.7
0.9
Reduced
class labels
Wall
Floor
Ceil (Background)
0.1
Obstacle
0.2
Target
[James and Johns, 2016] S. James and E. Johns. 3D simulation for
robot arm control with deep Q-learning. arXiv:1609.03759, Sep.
2016.
[Kingma and Ba, 2015] D. P. Kingma and J. Ba. Adam: A method
for stochastic optimization. In Proc. Int. Conf. Learning Representations (ICLR), May 2015.
[Krizhevsky et al., 2012] A. Krizhevsky, I. Sutskever, and G. E.
Hinton. ImageNet classification with deep convolutional neural
networks. In Advances in Neural Information Processing Systems
(NIPS), pp. 1097-1105. Dec. 2012.
[Long et al., 2015] J. Long, E. Shelhamer, and T. Darrell. Fully
convolutional networks for semantic segmentation. In Proc.
IEEE Conf. Computer Vision and Pattern Recognition (CVPR),
pp. 3431-3440, Jul. 2015.
[Maier et al., 2012] D. Maier, A. Hornung, and M. Bennewitz.
Real-time navigation in 3D environments based on depth camera data. In Proc. Int. Conf. Humanoid Robots (Humanoids), pp.
692-697, Nov. 2012.
[Mnih et al., 2016] V. Mnih, A. P. Badia, M. Mirza, A. Graves,
T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proc. Int.
Conf. Machine Learning (ICML), pp. 1928-1937, Jul. 2016.
[Parisotto and Salakhutdinov, 2017] E. Parisotto and R. Salakhutdinov. Neural map: Structured memory for deep reinforcement
learning. arXiv:1702.08360, Feb. 2017.
[Paszke et al., 2016] A. Paszke, A. Chaurasia, S. Kim, and E. Culurciello. ENet: A deep neural network architecture for real-time
semantic segmentation. arXiv:1606.02147, 2016.
[Peng et al., 2017] X. B. Peng, M. Andrychowicz, W. Zaremba, and
P. Abbeel. Sim-to-real transfer of robotic control with dynamics
randomization. arXiv:1710.06537, Jun. 2017.
[Richter et al., 2017] S. R. Richter, Z. Hayder, and V. Koltun. Playing for benchmarks. In Proc. Int. Conf. Computer Vision (ICCV),
Oct 2017.
[Rusu et al., 2016] A. A. Rusu, M. Vecerik, T. Rothörl, N. Heess,
R. Pascanu, and R. Hadsell. Sim-to-real robot learning from pixels with progressive nets. In Proc. Conf. Robot Learning (CoRL),
pp. 262-270, Nov. 2016.
[Sadeghi and Levine, 2016] F. Sadeghi and S. Levine. CAD2 RL:
Real single-image flight without a single real image.
arXiv:1611.04201, Nov. 2016.
[Smolyanskiy et al., 2017] N. Smolyanskiy, A. Kamenev, J. Smith,
and S. Birchfield. Toward low-flying autonomous MAV trail navigation using deep neural networks for environmental awareness.
arXiv:1705.02550, May 2017.
[Tobin et al., 2017] J. Tobin, R. Fong, A. Ray, J. Schneider,
W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world.
In Proc. Int. Conf. Intelligent Robots and Systems (IROS), pp. 2330, Sep. 2017.
[Williams, 1992] R. J. Williams.
Simple statistical gradientfollowing algorithms for connectionist reinforcement learning. J.
Machine learning, vol. 8, no. 3-4, pp. 229-256, 1992.
[Wu and Tian, 2016] Y. Wu and Y. Tian. Training agent for firstperson shooter game with actor-critic curriculum learning. Mar.
2016.
[You et al., 2017] Y. You, X. Pan, Z. Wang, and C. Lu. Virtual to real reinforcement learning for autonomous driving.
arXiv:1704.03952, Sep. 2017.
Original class labels in the dataset
Window, Door, Fence, Pillar, Sign
board, Bulletin board
Road, Ground, Field, Path, Runway
Ceil
Bed, Cabinet, Sofa, Table, Curtain,
Chair, Shelf, Desk, Plant
(Depends on scenario)
Table 4: The results Table 5: Class reduction from ADE20K
in Cluttered Hallway dataset.
with redundant labels
virtual and real environments. We further validated the generalizability of our model in handling unfamiliar indoor and
outdoor scenes, and the transferability of our model from simulation to the real world without fine-tuning. Finally, in the
switching-target following task, we proved that our model is
flexible such that the target can be easily switched by visual
guidance.
References
[Biswas and Veloso, 2012] J. Biswas and M. Veloso. Depth camera
based indoor mobile robot localization and navigation. In Proc.
Int. Conf. Robotics and Automation (ICRA), pp. 1697-1702, May
2012.
[Chen et al., 2016] L.-C. Chen, G. Papandreou, I. Kokkinos,
K. Murphy, and A. L. Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and
fully connected CRFs. IEEE Trans. Pattern Analysis and Machine Intelligence, Jun. 2016.
[Chen et al., 2017] Y.-H. Chen, W.-Y. Chen, Y.-T. Chen, B.-C. Tsai,
Y.-C. F. Wang, and M. Sun. No more discrimination: Cross city
adaptation of road scene segmenters. In Proc. Int. Conf. Computer Vision (ICCV), pp. 2011-2020, Oct. 2017.
[Cordts et al., 2016] M. Cordts, M. Omran, S. Ramos, T. Rehfeld,
M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele.
The cityscapes dataset for semantic urban scene understanding.
In Proc. IEEE Conf. Computer Vision and Pattern Recognition
(CVPR), pp. 3213-3223, Jun. 2016.
[Finn and Levine, 2017] C. Finn and S. Levine. Deep visual foresight for planning robot motion. In Proc. Int. Conf. Robotics and
Automation (ICRA), pp. 2786-2793, May 2017.
[Ghadirzadeh et al., 2017] A. Ghadirzadeh, A. Maki, D. Kragic,
and M. Björkman. Deep predictive policy training using reinforcement learning. In Proc. Int. Conf. Intelligent Robots and
Systems (IROS), pp. 2351-2358, Sep. 2017.
[Gupta et al., 2017] S. Gupta, J. Davidson, S. Levine, R. Sukthankar, and J. Malik. Cognitive mapping and planning for visual
navigation. In Proc. IEEE Conf. Computer Vision and Pattern
Recognition (CVPR), Jul. 2017.
[He et al., 2015] K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. Trans. Pattern Analysis and Machine Intelligence (TPAMI),
vol. 37, no.9, pp. 1904-1916, Sep. 2015.
[He et al., 2016] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), pp. 770-778, Jun.
2016.
7
[Zhang et al., 2017] F. Zhang, J. Leitner, M. Milford, and P. Corke.
Sim-to-real transfer of visuo-motor policies for reaching in clutter: Domain randomization and adaptation with modular networks. arXiv:1709.05746, Sep. 2017.
[Zhao et al., 2017a] H. Zhao, X. Qi, X. Shen, J. Shi, and J. Jia. ICNet for real-time semantic segmentation on high-resolution images. arXiv:1704.08545, Apr. 2017.
[Zhao et al., 2017b] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia.
Pyramid scene parsing network. In Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), pp. 2881-2890,
Jul. 2017.
[Zhou et al., 2017] B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ADE20K dataset.
In Proc. IEEE Conf. Computer Vision and Pattern Recognition
(CVPR), pp. 5122-5130, Jul. 2017.
[Zhu et al., 2017] Y. Zhu, R. Mottaghi, E. Kolve, J. J. Lim,
A. Gupta, L. Fei-Fei, and A. Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In Proc.
Int. Conf. Robotics and Automation (ICRA), pp. 3357-3364, May
2017.
8
| 3cs.SY
|
Detecting and Correcting for Label Shift
with Black Box Predictors
arXiv:1802.03916v2 [cs.LG] 28 Feb 2018
Zachary C. Lipton1,2∗, Yu-Xiang Wang2∗, Alex Smola2
Carnegie Mellon University
Amazon AI
[email protected], [email protected], [email protected]
March 1, 2018
Abstract
Faced with distribution shift between training and test set, we wish to detect and quantify
the shift, and to correct our classifiers without test set labels. Motivated by medical diagnosis,
where diseases (targets), cause symptoms (observations), we focus on label shift, where the
label marginal p(y) changes but the conditional p(x|y) does not. We propose Black Box Shift
Estimation (BBSE) to estimate the test distribution p(y). BBSE exploits arbitrary black
box predictors to reduce dimensionality prior to shift correction. While better predictors give
tighter estimates, BBSE works even when predictors are biased, inaccurate, or uncalibrated, so
long as their confusion matrices are invertible. We prove BBSE’s consistency, bound its error,
and introduce a statistical test that uses BBSE to detect shift. We also leverage BBSE to
correct classifiers. Experiments demonstrate accurate estimates and improved prediction, even
on high-dimensional datasets of natural images.
1
Introduction
Assume that in August we train a pneumonia predictor. Our covariates consist of chest X-rays
administered in the previous year (distribution P ) and the labels binary indicators of whether a
physician diagnoses the patient with pneumonia. We train a convnet f to predict pneumonia given
an X-ray image. Assume that in the training set .2% of patients have pneumonia. We deploy f in
the clinic and for several months, it reliably predicts roughly .2% positive.
Fast-forward to January (distribution Q): Running f on the last week’s data, we find that 5% of
patients are predicted to have pneumonia! Because f remains fixed, the shift must owe to a change
in the marginal p(x), violating the familiar iid assumption. Absent familiar guarantees, we wonder:
Is f still accurate? What’s the real current rate of pneumonia? Shouldn’t our classifier, trained
under an obsolete prior, underestimate pneumonia when uncertain? Thus, we might suspect that
the real prevalence is greater than 5%.
∗ Equal
contribution
1
Given only labeled training data, and unlabeled test data, we desire to: (i) detect distribution
shift, (ii) quantify it, and (iii) correct our model to perform well on the new data. Absent assumptions
on how p(y, x) changes, the task is impossible. The covariate shift assumption posits
q(y, x) = q(x)p(y|x).
In our case, covariate shift suggests that the distribution of radiologic findings P(x) changes, but
that the conditional probability of pneumonia P(y|x) remains constant. It’s easy to see how this can
go wrong: What if our only feature were a cough? Normally, cough might not (strongly) indicate
pneumonia. But during an epidemic, P(pneumonia|cough) might go up substantially. Instead, we
posit that the labels P(y) shift, i.e. the manifestations of a disease do not change, the prevalence of
the disease does:
q(y, x) = q(y)p(x|y)
This distinction is well-known: Schölkopf et al. [16] observe that covariate shift corresponds to
causal learning (predicting effects), while label shift corresponds to anticausal learning (predicting
causes). We focus on label shift, motivated by diagnosis (diseases cause symptoms) and recognition
tasks (objects cause sensory observations). During a pneumonia outbreak, P(y|x) might rise, but
the manifestations of the disease p(x|y) might not change. Moreover, sometimes the very purpose of
learning is to detect changes in p(y). Examples include monitoring public health [7] and wildlife
populations [1]. Despite its importance, label shift is comparatively under-investigated: given
covariates x from both P and Q quantifying q(x)/p(x) may seem more intuitive.
We introduce Black Box Shift Estimation (BBSE) to estimate label shift using a black box predictor
f . BBSE estimates the ratios wl = q(yl )/p(yl ) for each label l, requiring only that the expected
confusion matrix, on P is invertible1 . We estimate ŵ by solving a linear system Ax = b where A is
the confusion matrix of f on the training data and b is the average output of f calculated on the
test data. We make the following contributions:
1. Consistency and error bounds for BBSE.
2. We apply BBSE to statistical tests for detecting distribution shift (under the label shift
assumption)
3. Model correction through importance-weighted Empirical Risk Minimization.
4. A comprehensive empirical validation of BBSE.
BBSE requires only labeled training data, unlabeled test data, and a black box predictor whose
confusion matrices are invertible under the training distribution. In comparison to past approaches
based on Kernel Mean Matching (KMM) [22], EM [4], and Bayesian inference [18], BBSE offers the
following advantages:
1. Accuracy does not depend on the data dimensionality.
2. Works with arbitrary black box predictors, even biased, uncalibrated, or inaccurate models.
3. Exploits deep learning without losing theoretical guarantees: better predictors provably lower
sample complexity.
1 For degenerate confusion matrices, a variant using continuous predictor responses (e.g. softmax outputs) may
still work.
2
4. Owing to generality, BBSE could be a standard diagnostic and corrective tool for arbitrary
ML models.
2
Prior Work
Despite its wide applicability, learning under label shift with unknown q(y) remains curiously
under-explored. Noting the difficulty of the problem, Storkey [18] proposes placing a (meta-)prior
over p(y) and inferring the posterior distribution from the test covariates. This requires explicitly
estimating p(x|y), which may not be feasible in high-dimensional datasets. Chan & Ng [4] infer
q(y) using EM but their method also requires estimating p(x|y). Schölkopf et al. [16] articulates
connections between label shift and anti-causal learning and Zhang et al. [22] extend the kernel
mean matching approach due to [9] to the label shift problem. When q(y) is known, label shift
simplifies to the problem of changing base rates [3, 6].
Covariate shift, also called sample selection bias, is well-studied [20, 11, 19, 9]. Rosenbaum & Rubin
[15] introduce Propensity Scoring to design unbiased experiments. Shimodaira [17] is one of the
earliest to propose correcting models via weighting examples in ERM by q(x)/p(x). Later works
estimate importance weights from the available data, e.g., Gretton et al. [9] propose kernel mean
matching to re-weight training points.
The earliest relevant work to ours comes from econometrics and addresses the use of non-random
samples to estimate behavior. Heckman [10] addresses sample selection bias, while [13] investigates
estimating parameters under, choice-based or endogenous stratified sampling, cases analogous to a
shift in the label distribution. Crucially, unlike our setting, they assume that the test set marginal
distributions are both known. Also related to our work is the foundational research on propensity
scoring [15] Finally, we note a connection to cognitive science work demonstrating that humans
classify items differently depending on the set of unlabeled items in which they appear [23].
Previous work requires estimating q(x), q(x)/p(x), or p(x|y), often relying on kernel methods, which
scale poorly with dataset size and underperform on high-dimensional data. In contrast, BBSE uses
black-box predictors to reduce dimensionality. It exploits advances in discriminative modeling and
does not degrade, either in accuracy or tractability, as data or feature dimensions grow.
3
Problem setup
We use x ∈ X = Rd and y ∈ Y to denote the feature and label variables. For simplicity, we assume
that Y is a discrete domain equivalent to {1, 2, ..., k}. Let P, Q be the source and target distributions
defined on X × Y. We use p, q to denote the probability density function (pdf) or probability mass
function (pmf) associated with P and Q respectively. The random variable of interest is clear from
context. For example, p(y) is the p.m.f. of y ∼ P and q(x) is the p.d.f. of x ∼ Q. Moreover, p(y = i)
and q(y = i) are short for PP (y = i) and PQ (y = i) respectively, where P(S) := E[1(S)] denotes the
probability of an event S and E[·] denotes the expectation. Subscripts P and Q on these operators
make the referenced distribution clear.
3
In standard supervised learning, the learner observes training data (x1 , y1 ), (x2 , y2 ), ..., (xn , yn )
drawn iid from a training (or source) distribution P . We denote the collection of feature vectors by
X ∈ Rn×d and the label by y. Under Domain Adaptation (DA), the learner additionally observes
a collection of samples X 0 = [x01 ; ...; x0m ] drawn iid from a test (or target) distribution Q. Our
objective in DA is to predict well for samples drawn from Q.
In general this task is impossible, e.g., P and Q might not share support. This paper considers DA
under 3 extra assumptions:
A.1 The label shift (also known as target shift) assumption
p(x|y) = q(x|y) ∀ x ∈ X , y ∈ Y.
A.2 For every y ∈ Y with q(y) > 0 we require p(y) > 0.2
A.3 Access to a black box predictor f : X → Y where the expected confusion matrix Cp (f ) is
invertible.
CP (f ) := p(f (x), y) ∈ R|Y|×|Y|
We now comment on the assumptions. A.1 corresponds to anti-causal learning. This assumption is
strong but reasonable in many practical situations, including medical diagnosis, where diseases cause
symptoms. It also applies when classifiers are trained on non-representative class distributions: Note
that while visions systems are commonly trained with balanced classes [5], the true class distribution
for real tasks is rarely uniform.
Assumption A.2 addresses identifiability, requiring that the target label distribution’s support be a
subset of training distribution’s. For discrete Y, this simply means that the training data should
contain examples from every class.
Assumption A.3 requires that the expected predictor outputs for each class be linearly independent.
This assumption holds in the typical case where the classifier predicts class yi more often given
images actually belong to yi than given images from any other class yj . In practice, f could be a
neural network, a boosted decision-tree or any other classifier trained on a holdout training data
set. We can verify at training time that the empirical estimated normalized confusion matrix is
invertible. Assumption A.3 generalizes naturally to soft-classifiers, where f outputs a probability
distribution supported on Y. Thus BBSE can be applied even when the confusion matrix
is degenerate.
Objective We wish to estimate w(y) := q(y)/p(y) for every y ∈ Y with training data, unlabeled
test data and a predictor f . This estimate
enables DA techniques under the importance-weighted
Pn
ERM framework, which solves min i=1 wi `(yi , xi ), using wi = q(xi , yi )/p(xi , yi ). Under the label
shift assumption, the importance weight wi = q(yi )/p(yi ). This task isn’t straightforward because
we don’t observe samples from q(y).
2 Assumes the absolute continuity of the (hidden) target label’s distribution with respect to the source’s, i.e.
dq(y)/dp(y) exists.
4
4
Main results
We now derive the main results for estimating w(y) and q(y). Assumption A.1 has the following
implication:
Lemma 1. Denote by ŷ = f (x) the output of a fixed function f : X → Y. If Assumption A.1 holds,
then
q(ŷ|y) = p(ŷ|y).
The proof is simple: recall that ŷ depends on y only via x. By A.1, p(x|y) = q(x|y) and thus
q(ŷ|y) = p(ŷ|y).
Next, combine the law of total probability and Lemma 1 and we arrive at
X
q(ŷ) =
q(ŷ|y)q(y)
y∈Y
=
X
p(ŷ|y)q(y) =
y∈Y
X
p(ŷ, y)
y∈Y
q(y)
.
p(y)
(1)
We estimate p(ŷ|y) and p(ŷ, y) using f and data from source distribution P , and q(ŷ) with unlabeled
test data drawn from target distribution Q. This leads to a novel method-of-moments approach for
consistent estimation of the shifted label distribution q(y) and the weights w(y).
Without loss of generality, we assume Y = {1, 2, ..., k}. Denote by µy , µŷ , µ̂ŷ , ν y , ν ŷ , ν̂ ŷ , w ∈ Rk
moments of p and q and their plug-in estimates, defined via
[µy ]i = q(y = i)
[ν y ]i = p(y = i)
[µŷ ]i = q(f (x) = i)
P
0
j {f (xj ) = i}
[µ̂ŷ ]i =
m
[ν ŷ ]i = p(f (x) = i)
P
j {f (xj ) = i}
[ν̂ ŷ ]i =
n
and [w]i = q(y = i)/p(y = i). Lastly define the covariance matrices Cŷ,y , Cŷ|y and Ĉŷ,y in Rk×k
via
[Cŷ,y ]ij = p(f (x) = i, y = j)
[Cŷ|y ]ij = p(f (x) = i|y = j)
1X
[Ĉŷ,y ]ij =
I{f (xl ) = i and yl = j}
n
l
We can now rewrite Equation (1) in matrix form:
µŷ = Cŷ|y µy = Cŷ,y w
Using plug-in maximum likelihood estimates of the above quantities yields the estimators
ŵ = Ĉ−1
ŷ,y µ̂ŷ and µ̂y = diag(ν̂ y )ŵ,
where ν̂ y is the plug-in estimator of ν y .
Next, we establish that the estimators are consistent.
5
a.s.
Proposition 2 (Consistency). If Assumption A.1, A.2, A.3 are true, then as n, m → ∞, ŵ −→ w
a.s.
and µ̂y −→ µy .
The proof (see Appendix A) uses the First Borel-Cantelli Lemma to show that the probability that
the entire sequence of empirical confusion matrices with data size n + 1, ..., ∞ are simultaneously
invertible converges to 1, thereby enabling us to use the continuous mapping theorem after applying
the strong law of large numbers to each component.
We now address our estimators’ convergence rates.
Theorem 3 (Error bounds). Assume that A.3 holds robustly. Let σmin be the smallest eigenvalue
−2
of Cŷ,y . There exists a constant C > 0 such that for all n > 80 log(n)σmin
, with probability at least
−10
−10
1 − 3kn
− 2km
we have
C
kwk2 log n k log m
2
kŵ − wk2 ≤ 2
+
(2)
σmin
n
m
kµ̂y − µy k2 ≤
Ckwk2 log n
+ kν y k2∞ kŵ − wk22
n
(3)
The bounds give practical insights (explored more in Section 7). In (2), the square error depends
on the sample size and is proportional to 1/n (or 1/m). There is also a kwk2 term that reflects
how different the source and target distributions are. In addition, σmin reflects the quality of the
given classifier f . For example, if f is a perfect classifier, then σmin = miny∈Y p(y). If f cannot
distinguish between certain classes at all, then Cŷ,y will be low-rank, σmin = 0, and the technique is
invalid, as expected.
We now parse the error bound of µ̂y in (3). The first term kwk2 /n is required even if we observe
the importance weight w exactly. The second term captures the additional error due to the fact
that we estimate w with predictor f . Note that kν y k2∞ ≤ 1 and can be as small as 1/k 2 when p(y)
is uniform. Note that when f correctly classifies each class with the same probability, e.g. 0.5, then
2
kν y k2 /σmin
is a constant and the bound cannot be improved.
Proof of Theorem 3. Assumption A.2 ensures that w < ∞.
−1
ŵ = Ĉ−1
(µŷ + E2 )
ŷ,y µ̂ŷ = (Cŷ,y + E1 )
−1
= w + [(Cŷ,y + E1 )−1 − C−1
E2
ŷ,y ]µŷ + (Cŷ,y + E1 )
By completing the square and Cauchy-Schwartz inequality,
T
−1
kŵ − wk2 ≤ 2µTŷ [(Cŷ,y + E1 )−1 − C−1
− C−1
ŷ,y ] [(Cŷ,y + E1 )
ŷ,y ]µŷ
+ 2E2 [(Cŷ,y + E1 )−1 ]T (Cŷ,y + E1 )−1 E2 .
By the Woodbury matrix identity, we get that
−1
−1
−1
−1 −1 −1
Ĉ−1
Cŷ,y .
ŷ,y = Cŷ,y + Cŷ,y [E1 + Cŷ,y ]
Substitute into the above inequality and use (1) to get
n
oT
−1 T
−1
−1 −1
−1
kŵ − wk2 ≤2w [E1−1 + C−1
]
[Cŷ,y
] × C−1
w
ŷ,y
ŷ,y [E1 + Cŷ,y ]
+ 2E2T [(Cŷ,y + E1 )−1 ]T (Cŷ,y + E1 )−1 E2
6
(4)
We now provide a high probability bound on the Euclidean norm of E2 , the operator norm of
−1
E1 , which will give us an operator norm bound of [E1−1 + C−1
and (Cŷ,y + E1 )−1 under our
ŷ,y ]
assumption on n, and these will yield a high probability bound on the square estimation error.
Pn
Operator norm of E1 . Note that Ĉŷ,y = n1 i=1 ef (xi ) eTyi , where ey is the standard basis with
1 at the index of y ∈ Y and 0 elsewhere. Clearly, Eef (xi ) eTyi = Cŷ,y . Denote Zi := ef (xi ) eTyi − Cŷ,y .
Check that kZi k2 ≤ kZi kF ≤ kZi k1,1 ≤ 2, max{kE[Zi ZTi ]k, kE[ZTi Zi ]k} ≤ 1, by matrix Bernstein
inequality (Lemma 7) we have for all t ≥ 0
t2
P(kE1 k ≥ t/n) ≤ 2ke− n+2t/3 .
√
Take t = 20n log n and use the assumption that n ≥ 4 log n/9 (which holds under our assumption
on n since σmin < 1). Then with probability at least 1 − 2kn−10
r
20 log n
kE1 k ≤
.
n
Using the assumption on n, we have kE1 k ≤ σmin /2
−1
k[E1−1 + C−1
k ≤ 2kE1 k ≤
ŷ,y ]
Also, we have k(Cŷ,y + E1 )−1 k ≤
√
2 20 log n
√
.
n
2
σmin .
Pm
1
0
0
Euclidean norm of E2 . Note that [E2 ]l = m
i=1 1(f (xi ) = l) − q(f (xi ) = l). By the standard
Hoeffding’s inequality and union bound argument, we have that with probability larger than
1 − 2km−10
√
10k log m
√
kE2 k = kµŷ − µ̂ŷ k2 ≤
m
Substitute into Equation (4), we get
kŵ − wk2 ≤
80k log m
80 log n
kwk2 +
2
2 m ,
σmin n
σmin
(5)
which holds with probability 1 − 2kn−10 − 2km−10 . We now turn to µ̂y . Recall that µ̂y = diag(ν̂ y )ŵ.
Let the estimation error of ν̂ y be E0 .
µ̂y = µy + diag(E0 )w + diag(ν y )(ŵ − w) + diag(E0 )(ŵ − w).
q
n
By Hoeffding’s inequlaity kE0 k∞ ≤ 20 log
with probability larger than 1 − kn−10 . Combining
n
with (5) yields
1
20kwk2 log n
+ kν y k2∞ kŵ − wk2 + O( 2 )
kµ̂y − µy k2 ≤
n
n
which holds with probability 1 − 3kn−10 − 2km−10 .
7
0.8
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
Uniform
KS-test with MLP
oracle KS-test
MMD B-test on p(x)
0.0
0.0
0.2
0.4
0.6
0.8
(a) CDF of p-value at δ = 0
1.0
0.0
0.2
0.4
0.6
0.8
1.0
(b) CDF of p-value at δ = 0.6
Power of the test (1- Type II error)
Uniform
KS-test with MLP
oracle KS-test
MMD B-test on p(x)
1.0
KS-test with MLP (1 epoch)
KS-test with MLP (5 epoch)
MMD B-test on p(x)
Oracle KS-test
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
(c) Power at α = 0.05
Figure 1: Label-shift detection on MNIST. Pane 1a illustrates that Type I error is correctly controlled
absent label shift. Pane 1b illustrates high power under mild label-shift. Pane 1c shows increased
power for better classifiers. We compare to kernel two-sample tests [21] and an (infeasible) oracle
two sample test that directly tests p(y) = q(y) with samples from each. The proposed test beats
directly testing in high-dimensions and nearly matches the oracle.
5
5.1
Application of the results
Black Box Shift Detection (BBSD)
Formally, detection can be cast as a hypothesis testing problem where the null hypothesis is
H0 : q(y) = p(y) and the alternative hypothesis is that H1 : q(y) 6= p(y). Recall that we observe
neither q(y) nor any samples from it. However, we do observe unlabeled data from the target
distribution and our predictor f .
Proposition 4 (Detecting label-shift). Under Assumption A.1, A.2 and for each classifier f
satisfying A.3 we have that q(y) = p(y) if and only if p(ŷ) = q(ŷ).
Proof. Plug P and Q into (1) and apply Lemma 1 with assumption A.1. The result follows directly
from our analysis in the proof of Proposition 2 that shows p(ŷ, y) is invertible under the assumptions
A.2 and A.3.
Thus, under weak assumptions, we can test H0 by running two-sample tests on readily available
samples from p(ŷ) and q(ŷ). Examples include the Kolmogorov-Smirnoff test, Anderson-Darling
or the Maximum Mean Discrepancy. In all tests, asymptotic distributions are known and we can
almost perfectly control the Type I error. The power of the test (1-Type II error) depends on the
classifier’s performance on distribution P , thereby allowing us to leverage recent progress in deep
learning to attack the classic problem of detecting non-stationarity in the data distribution.
One could also test whether p(x) = q(x). Under the label-shift assumption this is implied by
q(y) = p(y). The advantage of testing the distribution of f (x) instead of x is that we only need to
deal with a one-dimensional distribution. Per theory and experiments in [14] two-sample tests in
high dimensions are exponentially harder.
8
Estimating error of w at = 0.1
Estimating error of w at = 1.0
Estimating error of w at = 10.0
100
101
MSE
100
MSE
MSE
100
10
10
1
BBSE-hard
BBSE-soft
KMM
KMM timeout
10
500 1000 2000 4000 8000 16000 32000
n
Target accuracy at = 0.1
2
0.98
BBSE-hard
BBSE-soft
KMM
KMM timeout
10
2
BBSE-hard
BBSE-soft
KMM
KMM timeout
500 1000 2000 4000 8000 16000 32000
n
Target accuracy at = 1.0
0.98
n
Target accuracy at = 10.0
0.96
0.96
0.92
Unweighted
BBSE-hard
BBSE-soft
KMM
KMM timeout
0.90
0.88
0.86
500 1000 2000 4000 8000 16000 32000
n
0.94
0.92
Unweighted
BBSE-hard
BBSE-soft
KMM
KMM timeout
0.90
0.88
0.86
Target accuracy
0.94
0.94
Target accuracy
Target accuracy
1
500 1000 2000 4000 8000 16000 32000
0.98
0.96
10
1
0.92
0.90
Unweighted
BBSE-hard
BBSE-soft
KMM
KMM timeout
0.88
0.86
0.84
500 1000 2000 4000 8000 16000 32000
n
500 1000 2000 4000 8000 16000 32000
n
Figure 2: Estimation error (top row) and correction accuracy (bottom row) vs dataset size on MNIST
data compared to KMM [22] under Dirichlet shift (left to right) with α = {.1, 1.0, 10.0} (smaller α
means larger shift). BBSE confidence interval on 20 runs, KMM on 5 runs due to computation;
n = 8000 is largest feasible KMM experiment.
One surprising byproduct is that we can sometimes use this approach to detect covariate-shift,
concept-shift, and more general forms of nonstationarity.
Proposition 5 (Detecting general nonstationarity). For any fixed measurable f : X → Y
P = Q =⇒ p(x) = q(x) =⇒ p(ŷ) = q(ŷ).
This follows directly from the measurability of f .
While the converse is not true in general, p(ŷ) = q(ŷ) does imply that for every measurable
S ⊂ Y,
q(x ∈ f −1 (S)) = p(x ∈ f −1 (S)).
This suggests that testing Ĥ0 : p(ŷ) = q(ŷ) may help us to determine if there’s sufficient statistical
evidence that domain adaptation techniques are required.
9
Estimating error of w at = 0.7
101
Estimating error of w at = 0.5
Estimating error of w at = 0.3
101
100
MSE
MSE
MSE
100
100
10
10
1
BBSE-hard
BBSE-soft
KMM
KMM timeout
10
1
500 1000 2000 4000 8000 16000 32000
BBSE-hard
BBSE-soft
KMM
KMM timeout
BBSE-hard
BBSE-soft
KMM
KMM timeout
10
500 1000 2000 4000 8000 16000 32000
n
Target accuracy at = 0.7
0.98
0.975
n
Target accuracy at = 0.5
Unweighted
BBSE-hard
BBSE-soft
KMM
KMM timeout
0.850
0.825
0.800
500 1000 2000 4000 8000 16000 32000
n
n
Target accuracy at = 0.3
0.94
0.92
0.90
0.88
Unweighted
BBSE-hard
BBSE-soft
KMM
KMM timeout
0.86
0.84
0.82
500 1000 2000 4000 8000 16000 32000
n
Target accuracy
0.875
Target accuracy
Target accuracy
0.900
500 1000 2000 4000 8000 16000 32000
0.96
0.94
0.925
2
0.98
0.96
0.950
1
0.92
0.90
Unweighted
BBSE-hard
BBSE-soft
KMM
KMM timeout
0.88
0.86
0.84
500 1000 2000 4000 8000 16000 32000
n
Figure 3: Label-shift estimation and correction on MNIST data with simulated tweak-one shift with
parameter ρ.
5.2
Black Box Shift Correction (BBSC)
Our estimator also points to a systematic method of correcting for label-shift via importance-weighted
ERM. Specifically, we propose the following algorithm:
Algorithm 1 Domain adaptation via Label Shift
input Samples from source distribution X, y. Unlabeled data from target distribution X 0 . A class
of classifiers F. Hyperparameter 0 < δ < 1/k.
1. Randomly split the training data into two X1 , X2 ∈ Rn/2×d and y 1 , y 2 Rn/2 .
2. Use X1 , y 1 to train the classifier and obtain f ∈ F.
3. On the
data set X2 , y 2 , calculate the confusion matrix Ĉŷ,y . If ,
hold-out
if σmin Ĉŷ,y ≤ δ then
Set ŵ = 1.
else
Estimate ŵ = Ĉ−1
ŷ,y µ̂ŷ .
end if
4. Solve the importance weighted ERM on the X1 , y 1 with max(ŵ, 0) and obtain f˜.
output f˜
10
Note that for classes that occur rarely in the test set, BBSE may produce negative importance
weights. During ERM, a flipped sign would cause us to maximize loss, which is unbounded above.
Thus, we clip negative weights to 0.
Owing to its efficacy and generality, our approach can serve as a default tool to deal with domain
adaptation. It is one of the first things to try even when the label-shift assumption doesn’t hold. By
contrast, the heuristic method of using logistic-regression to construct importance weights [2] lacks
theoretical justification that the estimated weights are correct.
Even in the simpler problem of average treatment effect (ATE) estimation, it’s known that using
estimated propensity can lead to estimators with large variance [12]. The same issue applies in
supervised learning. We may prefer to live with the biased solution from the unweighted ERM
rather than suffer high variance from an unbiased weighted ERM. Our proposed approach offers a
consistent low-variance estimator under label shift.
6
Experiments
We experimentally demonstrate the power of BBSE with real data and simulated label shift. We
organize results into three categories — shift detection with BBSD, weight estimation with
BBSE, and classifier correction with BBSC. BBSE-hard denotes our method where f yields
classifications. In BBSE-soft, f outputs probabilities.
Label Shift Simulation To simulate distribution shift in our experiments, we adopt the following
protocols: First, we split the original data into train, validation, and test sets. Then, given
distributions p(y) and q(y), we generate each set by sampling with replacement from the appropriate
split. In knock-out shift, we knock out a fraction δ of data points from a given class from training
and validation sets. In tweak-one shift, we assign a probability ρ to one of the classes, the rest
of the mass is spread evenly among the other classes. In Dirichlet shift, we draw p(y) from a
Dirichlet distribution with concentration parameter α. With uniform p(y), Dirichlet shift is bigger
for smaller α.
Label-shift detection We conduct nonparametric two-sample tests as described in Section 5.1
using the MNIST handwritten digits data set. To simulate the label-shift, we randomly split the
training data into a training set, a validating set and a test set, each with 20,000 data points,
and apply knock-out shift on class y = 5 3 . Note that p(y) and q(y) differ increasingly as δ
grows large, making shift detection easier. We obtain f by training a two-layer ReLU-activated
Multilayer Perceptron (MLP) with 256 neurons on the training set for five epochs. We conduct a
two-sample test of whether the distribution of f (Validation Set) and f (Test Set) are the same using
the Kolmogorov-Smirnov test. The results, summarized in Figure 1, demonstrate that BBSD (1)
produces a p-value that distributes uniformly when δ = 0 4 (2) provides more power (less Type II
error) than the state-of-the-art kernel two-sample test that discriminates p(x) and q(x) at δ = 0.5,
and (3) gets better as we train the black-box predictor even more.
Weight estimation and label-shift correction We evaluate BBSE on MNIST by simulating
label shift and datasets of various sizes. Specifically, we split the training data set randomly in two,
3 Random
4 Thus
choice for illustration, method works on all classes.
we can control Type I error at any significance level.
11
Accuracy
0.9
0.8
Black Box Shift Correction on CIFAR10
BBSE-hard
Unweighted
1.0
0.9
0.8
Accuracy
1.0
0.7
0.6
0.5
Black Box Shift Correction on CIFAR10
BBSE-hard
Unweighted
0.7
0.6
0.5
0.4
0.4
0.3
0.3
0.0
0.2
0.4
0.6
0.8
1.0
1000
100
10
1
0.1
0.01
0.001
Figure 4: Accuracy of BBSC on CIFAR 10 with (left) tweak-one shift and (right) Dirichlet shift.
using first half to train f and the second half to estimate w. We use then use the full training set
for weighted ERM. As before, f is a two-layer MLP. For fair comparisons with baselines, the full
training data set is used throughout (since they do not need f without data splitting). We evaluate
our estimator ŵ against the ground truth w and by the prediction accuracy of BBSC on the test
set. To cover a variety of different types of label-shift, we take p(y) as a uniform distribution and
generate q(y) with Dirichlet shift for α = 0.1, 1.0, 10.0 (Figure 2).
Label-shift correction for CIFAR10 Next, we extend our experiments to the CIFAR dataset,
using the same MLP and this time allowing it to train for 10 epochs. We consider both tweak-one
and Dirichlet shift, and compare BBSE to the unweighted classifier under varying degrees of shift
(Figure 4). For the tweak-one experiment, we try ρ ∈ {0.0, 0.1, ..., 1.0}, averaging results over
all 10 choices of the tweaked label, and plotting the variance. For the Dirichlet experiments, we
sample 20 q(y) for every choice of α in the range {1000, 100, ..., .001}. Because the kernel-based
baselines cannot handle datasets of this size and dimensionality, here we compare only to unweighted
ERM.
Kernel mean matching (KMM) baselines We compare BBSE to the state-of-the-art kernel
mean matching (KMM) methods. For the detection experiments (Figure 1), our baseline is the
kernel B-test [21], an extension of the kernel max mean discrepancy (MMD) test due to Gretton
et al. [8] that boasts nearly linear-time computation and little loss in power. We compare BBSE to
a KMM approach Zhang et al. [22], that solves
min kCx|y (ν y ◦ w) − µx k2H ,
w
where we use operator Cx|y := E[φ(x)|ψ(y)] and function µx := EQ [φ(x)] to denote the kernel
embedding of p(x|y) and pQ (x) respectively. Note that under the label-shift assumption, Cx|y is
the same for P and Q. Also note that since Y is discrete, ψ(y) is simply the one-hot representation
of y, so ν y is the same as our definition before and Cx|y , ν y and µx must be estimated from finite
data. The proposal involves a constrained optimization by solving a Gaussian process regression
with automatic hyperparameter choices through marginal likelihood.
12
For fair comparison, we used the original authors’ implementations as baselines 5 and also used
the median trick to adaptively tune the RBF kernel’s hyperparameter. A key difference is that
BBSE matches the distribution of ŷ rather than distribution of x like [22] and we learn f through
supervised learning rather than by specifying a feature map φ by choosing a kernel up front.
Note that KMM, like many kernel methods, requires the construction and inversion of an n × n
Gram matrix, which has complexity of O(n3 ). This hinders its application to real-life machine
learning problems where n will often be 100s of thousands. In our experiments, we find that the
largest n for which we can feasibly run the KMM code is roughly 8, 000 and that is where we
unfortunately have to stop for the MNIST experiment. For the same reason, we cannot run KMM
for the CIFAR10 experiments. The MSE curves in Figure 2 for estimating w suggest that the
convergence rate of KMM is slower than BBSE by a polynomial factor and that BBSE better
handles large datasets.
7
Discussion
Constructing the training Set The error bounds on our estimates depend on the norm of the
true vector w(y) := q(y)/p(y). This confirms the common sense that absent any assumption on q(y),
and given the ability to select class-conditioned examples for annotations one should build a dataset
with uniform p(y). Then it’s always possible to apply BBSE successfully at test time to correct
f.
Sporadic Shift In some settings, p(y) might change only sporadically. In these cases, when no
label shift occurs, applying BBSC might damage the classifier. For these cases, we prose to combine
detection and estimation, correcting the classifier only when a shift has likely occurred.
Using known predictor In our experiments, f has been trained using a random split of the
data set, which makes BBSEto perform worse than baseline when the data set is extremely small.
In practice, especially in the context of web services, there could be a natural predictor f that
is currently being deployed whose training data were legacy and have little to do with the two
distributions that we are trying to distinguish. In this case, we do not lose that factor of 2 and we
do not suffer from the variance in training f with a small amount of data. This could allow us to
detect mild shift in distributions in very short period of time. Making it suitable for applications
such as financial market prediction.
BBSE with degenerate confusion matrices In practice, sometime confusion matrices will
be degenerate. For instance, when a class i is rare under P , and the features are only partially
predictive, we might find that p(f (x) = i) = 0. In these cases, two straightforward variations on the
black box method may still work: First, while our analysis focuses on confusion matrices, it easily
extends to any operator f , such as soft probabilities. If each class i, even if i is never the argmax
for any example, so long as p(ŷ = i|y = i) > p(ŷ = i|y = j) for any j 6= i, the soft confusion matrix
will be invertible. Even when we produce and operator with an invertible confusion matrix, two
options remain: We can merge c classes together, yielding a (k − c) × (k − c) invertible confusion
matrix. While we might not be able to estimate the frequencies of those c classes, we can estimate
the others accurately. Another possibility is to compute the pseudo-inverse.
5 https://github.com/wojzaremba/btest,
http://people.tuebingen.mpg.de/kzhang/Code-TarS.zip
13
Future Work As a next step, we plan to extend our methodology to the streaming setting. In
practice, label distributions tend to shift progressively, presenting a new challenge: if we apply BBSE
on trailing windows, then we face a trade-off. Looking far back increases m, lowering estimation
error, but the estimate will be less fresh. The use of propensity weights w on y makes BBSE
amenable to doubly-robust estimates, the typical bias-variance tradeoff, and related techniques,
common in covariate shift correction.
References
[1] Beijbom, O., Edmunds, P. J., Kline, D. I., Mitchell, B. G., & Kriegman, D. (2012). Automated
annotation of coral reef survey images. In Computer Vision and Pattern Recognition (CVPR).
[2] Bickel, S., Brückner, M., & Scheffer, T. (2009). Discriminative learning under covariate shift.
Journal of Machine Learning Research, 10 (Sep), 2137–2155.
[3] Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford university press.
[4] Chan, Y. S., & Ng, H. T. (2005). Word sense disambiguation with distribution estimation. In
Proceedings of the 19th international joint conference on Artificial intelligence, (pp. 1010–1015).
Morgan Kaufmann Publishers Inc.
[5] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). Imagenet: A large-scale
hierarchical image database. In CVPR.
[6] Elkan, C. (2001). The foundations of cost-sensitive learning. In IJCAI .
[7] Ginsberg, J., Mohebbi, M. H., Patel, R. S., Brammer, L., Smolinski, M. S., & Brilliant, L.
(2009). Detecting influenza epidemics using search engine query data. Nature.
[8] Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). A kernel
two-sample test. Journal of Machine Learning Research, 13 (Mar), 723–773.
[9] Gretton, A., Smola, A. J., Huang, J., Schmittfull, M., Borgwardt, K. M., & Schölkopf, B.
(2009). Covariate shift by kernel mean matching. Journal of Machine Learning Research.
[10] Heckman, J. J. (1977). Sample selection bias as a specification error (with an application to
the estimation of labor supply functions).
[11] Huang, J., Gretton, A., Borgwardt, K. M., Schölkopf, B., & Smola, A. J. (2007). Correcting
sample selection bias by unlabeled data. In Advances in neural information processing systems.
[12] Kang, J. D., & Schafer, J. L. (2007). Demystifying double robustness: A comparison of
alternative strategies for estimating a population mean from incomplete data. Statistical
science, 22 (4), 523–539.
[13] Manski, C. F., & Lerman, S. R. (1977). The estimation of choice probabilities from choice
based samples. Econometrica: Journal of the Econometric Society.
[14] Ramdas, A., Reddi, S. J., Póczos, B., Singh, A., & Wasserman, L. A. (2015). On the decreasing
power of kernel and distance based nonparametric hypothesis tests in high dimensions. In
AAAI .
14
[15] Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in
observational studies for causal effects. Biometrika, 70 (1), 41–55.
[16] Schölkopf, B., Janzing, D., Peters, J., Sgouritsa, E., Zhang, K., & Mooij, J. (2012). On causal
and anticausal learning. In International Coference on International Conference on Machine
Learning (ICML-12), (pp. 459–466). Omnipress.
[17] Shimodaira, H. (2000). Improving predictive inference under covariate shift by weighting the
log-likelihood function. Journal of statistical planning and inference.
[18] Storkey, A. (2009). When training and test sets are different: characterizing learning transfer.
Dataset shift in machine learning.
[19] Sugiyama, M., Nakajima, S., Kashima, H., Buenau, P. V., & Kawanabe, M. (2008). Direct
importance estimation with model selection and its application to covariate shift adaptation.
In Advances in neural information processing systems.
[20] Zadrozny, B. (2004). Learning and evaluating classifiers under sample selection bias. In
Proceedings of the twenty-first international conference on Machine learning, (p. 114). ACM.
[21] Zaremba, W., Gretton, A., & Blaschko, M. (2013). B-test: A non-parametric, low variance
kernel two-sample test. In Advances in neural information processing systems, (pp. 755–763).
[22] Zhang, K., Schölkopf, B., Muandet, K., & Wang, Z. (2013). Domain adaptation under target
and conditional shift. In International Conference on Machine Learning, (pp. 819–827).
[23] Zhu, X., Gibson, B. R., Jun, K.-S., Rogers, T. T., Harrison, J., & Kalish, C. (2010). Cognitive
models of test-item effects in human category learning. In ICML.
15
A
Proofs
We present the proofs of Lemma 1 and Proposition 2 in this Appendix.
Proof of Lemma 1. By the law of total probability
X
X
q(ŷ|y) =
q(ŷ|x, y)q(x|y) =
q(ŷ|x, y)p(x|y)
y∈Y
=
X
y∈Y
pf (ŷ|x)p(x|y) =
y∈Y
X
p(ŷ|x, y)p(x|y) = p(ŷ|y).
y∈Y
We applied A.1 to the second equality, and used the conditional independence ŷ ⊥
⊥ y|x under P and
Q together with p(ŷ|x) being determined by f , which is fixed.
Proof of Proposition 2. A.2 ensures that w < ∞. By Assumption A.3, Cŷ,y is invertible. Let δ > 0
be its smallest singular value. We bound the probability that Ĉŷ,y is not invertible:
P(Ĉŷ,y is not invertible) ≤ P(σmin (Ĉŷ,y ) < δ/2)
δ
≤ P(kĈŷ,y − Cŷ,y k2 ≥ δ/2) ≤ P(kĈŷ,y − Cŷ,y kF ≥ √ )
↑
2 k
pigeon hole
≤ P(∃(i, j) ∈ [k]2 , s.t.|[Ĉŷ,y ]i,j − [Cŷ,y ]i,j | ≥
↑
nδ 2
δ
) ≤ 2e− 4k3 .
1.5
2k
↑
pigeon hole
Hoeffding
P
By the convergence of geometric series n P(Ĉŷ,y is not invertible) < +∞. This allows us to invoke
the First Borel-Cantelli Lemma, which shows
P(Ĉŷ,y is not invertible i.o.) = 0.
(6)
This ensures that as n → ∞, Ĉŷ,y is invertible almost surely. By the strong law of large numbers
a.s.
a.s.
a.s.
(SLLN), as n → ∞ Ĉŷ,y −→ Cŷ,y and ν̂ y −→ ν y . Similarly, as m → ∞, µ̂ŷ −→ µŷ . Combining
these with (6) and applying the continuous mapping theorem with the fact that the inverse of an
invertible matrix is a continuous mapping we get that
a.s.
a.s.
ŵ = [Ĉŷ,y ]−1 µ̂ŷ −→ w, and µ̂y = diag(ν̂ y )ŵ −→ µy .
B
Concentration inequalities
Lemma 6 (Hoeffding’s
Pn inequality). Let x1 , ..., xn be independent random variables bounded by
[ai , bi ]. Then x̄ = n1 i=1 xi obeys for any t > 0
2n2 t2
P
P(|x̄ − E[x̄]| ≥ t) ≤ 2 exp − n
.
2
i=1 (bi − ai )
16
Lemma 7 (Matrix Bernstein Inequality (rectangular case)). Let Z1 , ..., Zn be independent random
matrices with dimension d1 × d2 and each satisfy
EZi = 0 and kZi k ≤ R
almost surely. Define the variance parameter
X
X
σ 2 − max{k
E[Zi ZTi ]k, k
E[ZTi Zi ]k}.
i
i
Then for all t ≥ 0,
!
P k
X
Zi k ≥ t
−t2
≤ (d1 + d2 ) · e σ2 +Rt/3 .
i
17
| 9cs.NE
|
Approximation Algorithms for Multi-Multiway Cut and
Multicut Problems on Directed Graphs
arXiv:1610.02336v3 [cs.DS] 7 Feb 2018
Ramin Yarinezhad
a Department
✩a
, Seyed Naser Hashemia,∗
of Mathematics and Computer Science, Amirkabir University of Technology,
Tehran, Iran
Abstract
In this paper, we present two approximation algorithms for the directed
multi-multiway cut and directed multicut problems. The so called region growing paradigm [1] is modified and used for these two cut problems on directed
graphs. By using this paradigm, we give for each problem an approximation algorithm such that both algorithms have the approximate factor O(k) the same
as the previous works done on these problems. However, the previous works
need to solve k linear programming, whereas our algorithms require only one
linear programming. Therefore, our algorithms improve the running time of the
previous algorithms.
Keywords: Approximation algorithm, Computational complexity, NP-hard
problems, Directed multi-multiway cut, Directed multicut cut
1. Introduction
In the following, we first review some of the important cut problems which
serve as a background for the problems focused in this paper. The undirected
multiway cut problem is defined on an undirected graph G = (V, E) with a
given set S = {s1 , ...s2 } ⊆ V of vertices called terminals and a weight function
ce , e ∈ E. Here, the goal is to find the minimum weight subset of edges so
✩
Email address: [email protected] (Ramin Yarinezhad)
author
Email address: [email protected] (Seyed Naser Hashemi )
∗ Corresponding
Preprint submitted to ..
February 9, 2018
that by deleting them, all terminals in S are disconnected. In other words,
there is not any path between any two considered terminals. It is proved that
this problem, for k ≥ 3, is NP-hard and MAX SNP-hard, for which a 2 − 2/k
factor approximation algorithm is given [5]. In [6], using a geometric relaxation,
an algorithm with an approximate factor of 1.5 − 1/k is introduced and it is
improved to 1.3438 − ǫk in [7].
For directed graphs, the version of the directed multiway cut problem is
defined. Likewise, given a set of terminals S = {s1 , ...s2 } ⊆ V and a weight
function ce , e ∈ E, we look for a minimum weight subset of edges which their
deletions disconnect all directed paths between each pair of terminals. Vazirani
and Yannakakis [3] showed that a directed multiway cut problem is also NP-hard
and MAX SNP-hard. They introduced an algorithm with a 2 log k approximate
factor. The best known approximation algorithm, presented by Noar and Zosin
[4], used a novel relaxation multiway flow to have an approximation algorithm
within a factor of 2.
The problem of undirected multicut is another well-known problem defined
on undirected graphs with a non-negative cost function ce , e ∈ E, and a set
of ordered pairs of vertices, namely; (s1 , t1 ), ..., (sk , tk ), which are called sourceterminal vertices. In this case, the attempt is to achieve a minimum cost subset
of edges so that removing them all sources become inaccessible from their corresponding terminals. For k ≥ 3 , it is shown that the problem is NP-hard
and MAX SNP-hard [5]. Garg, Vazirani, and Yannakakis [1] give, by the region
growing technique, an approximation algorithm with the O(log k) approximate
factor. In [18] for this problem with more constraints, an approximation algorithm has been proposed with approximation factor O(rlog 3/2 k), where r is a
part of the input instance.
The directed multicut problem is defined as follows: given a directed graph
G = (V, E), |V | = n with a non-negative function ce > 0, e ∈ E, and a set
of ordered pairs of vertices (s1 , t1 ), ..., (sk , tk ), we find a subset F ⊆ E with
minimum cost function so that their removal from the graph makes each pair
disconnected. That is, for any i, 1 ≤ i ≤ k, there is not any directed path from
2
si to ti in the graph G(V, E − F ).
Furthermore, if the desire is also to disconnect the paths from ti to si , for
any i, 1 ≤ i ≤ k, we have an alternative version of the multicut problem called
directed symmetric multicut problem.
As shown in [3], for k ≥ 2, the directed multicut problem is NP-hard and
MAX SNP-hard. In some papers have been shown that another version of this
problem is NP-hard [13]. In literature, most of the works on directed multicut
have been focused on the directed symmetric multicut problem [9, 10, 11, 12].
Even, Noar, Schieber and Sudan [11] presented an approximation algorithm
with a factor O((log)loglogk). In general, for a non-symmetric version, using the
√
technique region growing, an algorithm with the approximate factor O( nlogk)
is given [14]. For the general case, Gupta [15] introduced a simpler algorithm
√
and improved the approximate factor to O( n). Both problems above, studied
by [14, 15], use a linear programming relaxation to approximate the solution.
In the work of Saks, Samorodnitsky, Zosin [17], it is shown that the integrality
gap for the linear programming relaxation is O(k).
A more general problem on undirected graphs is the multi-multiway cut
problem in which the weight function w : E → ℜ+ and k sets S1 , S2 , ..., Sk
are given. Here, our aim is to obtain a minimum weight subset of edges whose
removal from the graph will disconnect all connections between the vertices in
each set Si , for 1 ≤ i ≤ k. For K = 1, this problem is an undirected multiway
cut problem, and if |Si | = 2, (1 ≤ i ≤ k), an undirected multicut problem is
obtained. Avidor and Langberg [2] showed that the undirected multi-multiway
cut problem is NP-hard and MAX SNP-hard, and by using the region growing
technique they could present an approximation algorithm within the factor of
O(logk). When the input graph is a tree, in [16] has been shown that this
problem is solvable in polynomial time, if the number of terminal sets is fixed
√
and in [8] has been presented an approximation algorithm with a factor O( k).
A directed version of the above problem is also defined namely as a directed multi-multiway cut problem. Similarly, for this problem a weight function w : E → ℜ+ on edges and k sets S1 , S2 , ..., Sk are given. We attempt to find
3
a minimum weight subset of edges whose removal from the graph will disconnect
all paths between the vertices in each set Si , for 1 ≤ i ≤ k. This problem generalizes the problems of directed multiway cut and directed symmetric multicut
(when k = 1 and |Si | = 2, respectively).
Since every instance of the directed multiway cut is defined as an instance of
the directed multi-multiway cut problem when k = 1, so the hardness proof for
the multiway cut problem implies that the directed multi-multiway cut problem
is also NP-hard and MAX SNP-hard.
Note that the problem of directed multi-multiway cut cannot be viewed as a
generalization of the undirected multi-multiway cut problem only by replacing
each undirected edge by two unparalleled directed edges. For example, consider
a tree with a root r, containing three leaves a, b, c, and assuming the weight of
each edge is equal to one. In this case, we get the optimal value, OP T = 2,
whereas substituting each edge by two directed edges gives OP T = 3, and this
proves that two problems above are not equivalent.
As described above, it is clear that the problems of directed multicut and
directed multi-multiway cut can be approximated by a factor O(k). But for
each of these problems, we require k linear programming to be solved in order
to obtain the desired approximation solution. In this paper, we show that we can
achieve the same result, i.e. an approximation with the factor O(k), by solving
only one linear programming. To achieve this goal, the so called paradigm of
region growing, introduced in [1] for undirected cut problems, is modified so
that it can be useful to produce an approximate solution of the multicut and
multi-multiway cut problems on directed graphs.
1.1. Organization
The rest of this paper is organized as follows: In section 2 we present a linear
programming relaxation for the Directed Multi-Multiway Cut problem which is
used in [2] and [3]. Section 3 contains necessary definitions and lemmas for the
algorithm Directed Multi-Multiway cut which proposed in section 4. Directed
Multicut Algorithm presented in section 5 and Conclusion is brought in section
4
6.
2. Linear Programming Relaxation for the Directed Multi-Multiway
Cut
We define a decision variable x(e) for each edge e which is as following: if
e belongs to directed multi-multiway cut, x(e) = 1, otherwise x(e) = 0. The
purpose is to find a directed multi-multiway cut with the minimum cost which
disconnects every directed path between two vertices in a group. We call that
the set of all directed paths between any two vertices belongs to a group, with
P . An integer program for the problem is given by:
minimize
X
w(e)x(e)
e∈E
subject to
X
e∈p
x(e) ≥ 1,
∀p ∈ P
x(e) ∈ {0, 1}, ∀e ∈ E
By relaxing this IP we obtain the following linear programming relaxation:
minimize
X
w(e)x(e)
e∈E
subject to
X
e∈p
x(e) ≥ 1, ∀p ∈ P
x(e) ≥ 0, ∀e ∈ E
In this LP, there is a constraint for each path. On the other hand, we may
have an exponential number of paths with respect to the input size and as a
result, exponential number of constraints. Nevertheless, we can solve this LP
in polynomial time, using the ellipsoid algorithm. For this LP, the separation
oracle operates as following: we get a solution x and assume that the length of
each edge e is equal to x(e). Then, we find the shortest directed path between
two vertices which are needed to be disconnected from each other. For example
(u, v), if the shortest path between u and v (either v → u or u → v) is more
P
than 1, then this constraint e∈p x(e) ≥ 1 is true for all paths between these
two vertices. Therefore, this LP can be solved in polynomial time.
5
To express the approximation algorithm for directed multi-multiway cut, we
need several definitions and lemmas which are presented in the next section.
3. Definitions and Lemmas
To round the solution of the mentioned LP and obtain a directed multimultiway cut, we use the region growing technique [1, 2]. Note that definitions
in [1, 2] are related to undirected graphs while definitions presented here, are
related to directed graphs.
We define a distance on edges and assume x is an optimal solution for the
LP. Let x(e) be the length of edge e. The distance between two vertices u and
v (either v → u or u → v), which is defined based on x(e), is the length of the
shortest path between them. We represent this shortest path with dist(u, v). If
there is no directed path between two vertices u and v, the value of dist(u, v) is
equal to the length of the shortest path in the graph, regardless of the direction
of edges. We define:
Bx (sij , r) = {v ∈ V : dist(sij , v) ≤ r},
where 1 ≤ i ≤ k, 1 ≤ j ≤ |Si | and r ∈ R. Bx (sij , r) is an area like a ball
with center sij and radius r. Assume that δ(s) is the set of all edges which
only one of their heads is in the set s. For a given radius r, let wt(δ(Bx (sij , r)))
be the sum of weights of all edges which one of their heads is in Bx (sij , r).
wt(δ(Bx (sij , r))) is defined as follows:
wt(δ(Bx (sij , r))) =
X
w(e),
e∈δ(Bx (sij ,r))
where w(e) is the weight of edge e. The same as [2], let ci (r) be the sum
of weights of directed edges whose one head only is inside these balls, where
1 ≤ i ≤ k. ci (r) is defined as follows:
6
|Si |
X
ci (r) =
wt(δ(Bx (sij , r))).
j=1
Assume that each edge e in the graph as being a pipe with cross-sectional
area w(e) and length x(e). Then, the product w(e)x(e) is equal to the volume
of edge e. Thus, the solution of LP is the minimum volume of edges such that
dist(u, v) ≥ 1, where u and v are in the same group, and there is a path between
them (either v → u or u → v). Let x be an optimal solution for the LP, and
P
V ∗ = e∈E w(e)x(e) be the volume of all edges. We know that V ∗ ≤ OP T
such that OP T is the optimal value for the IP. vi (r) is defined as follows:
vi (r) = βV ∗ +
|Si |
X
j=1
(
X
w(e)x(e) +
X
e=(u,v)∈E
u∈Bx (sij ,r)
v ∈B
/ x (sij ,r)
e=(u,v)∈E
u,v∈Bx (sij ,r)
w(e)(r − dist(sij , u))),
where β > 0 and is independent from r. We have to notice that an edge
may appear in ci (r) more than once. That means we may have δ(Bx (sij , r)) ∩
′
δ(Bx (sij ′ , r)) 6= ∅, for 1 ≤ j 6= j ≤ |Si |. Thus, ci (r) is an upper bound on the
cut. According to these definitions, we can express the Lemma 1, which is used
in the proof of Lemma 2.
Lemma 1: The function vi (r) is differentiable in (0, ∞) except some finite
number of points. The derivative of this function is ci (r).
Proof: The function vi (r) is not differentiable in points which the value of
function Bx (sij , r) changes. The function Bx (sij , r), changes for the values of
r in which there is a vertex v such that dist(sij , r) = r. Thus the number of
points in which the function vi (r) is not differentiable, is finite. Beside this,
according to the definition done for the function vi (r), the derivative of this
function is ci (r).
Lemma 2 says in directed graphs, we can always find a radius r < 12 , such
that the cost vi (r) is an upper bound for ci (r).
7
Lemma 2: Let x be a feasible solution for the LP, for every sij there is a
r<
1
2
and at least an α (α > 0) such that the following inequality is true:
ci (r) ≤ αvi (r).
The proof of this lemma is brought in the Section 4.1. We first present the
algorithm using this lemma.
4. Approximation Algorithm for Directed Multi-Multiway Cut
Our polynomial time approximation algorithm for directed multi-multiway
cut is described in Algorithm 1. The algorithm solves the LP first and finds
the optimal solution x. Then, the algorithm enters to a repetition loop and
till there exists a path between two vertices in a group, the algorithm works as
follows: assume that the set Si is chosen in this iteration. In the beginning, it
finds an r which satisfies the inequality of Lemma 2, and then it finds the set
of balls with the center of vertices inside the Si with the radius of r. Then it
puts the edges, which have been cut by these balls, in the answer set F .
Now the algorithm checks whether there are two vertices from another group
which are connected by a path and are in the same ball. If there are such vertices,
only the vertex in the center of each ball and its incident edges will be deleted
from the graph. Otherwise, all of the vertices in balls and incident edges with
them will be deleted from the graph.
8
Algorithm 1 Approximation Algorithm for Directed Multi-Multiway Cut
Result: A Directed Multi-Multiway Cut
F ←∅
Solve the LP and get the optimal solution x
while there is a path between sij ∈ Si and sij ′ ∈ Si , where 1 ≤ i ≤ k and
′
1 ≤ j 6= j ≤ |Si | do
Find ri such that ci (ri ) ≤ αvi (ri )
S|Si |
δ(Bx (sij , ri )) to F
Add j=1
S|Si |
if j=1 Bx (sij , ri ) contains two vertices u, v such that u, v ∈ Bx (sij , ri ),
where 1 ≤ j ≤ |Si | AND u, v ∈ Sm , where 1 ≤ m ≤ k and m 6= i AND
there is a path between u and v then
Remove sij , where 1 ≤ j ≤ |Si |, and incident edges with it form the graph
else
Remove
S|Si |
j=1
Bx (sij , ri ) and incident edges with it form the graph
end
∀l ∈ {1, ..., k}, Sl ← Sl ∩ V
end
Return F
Lemma 3: Algorithm 1 returns a Directed Multi-Multiway Cut.
Proof: For each ball like Bx (sij , r), where 1 ≤ i ≤ k, 1 ≤ j ≤ |Si |, there is
no vertex with the same group with sij in Bx (sij , r) because the radius of each
ball is smaller than
1
2.
The only case may lead to problems is that there are
two vertices u and v in one ball, which are the members of another group and
there is a path between them. In this case, only the central vertices and their
incident edges will be deleted from the graph. Therefore, the path between the
two vertices u and v will not be deleted from the graph. In the next iterations,
at least one of the edges of the path between u and v will be put in the answer
set.
9
Theorem 1: Algorithm 1 is a (α(1 + β)k)-approximation algorithm for
Directed Multi-Multiway Cut.
Proof: According to Lemma 2, we have ci (r) ≤ αvi (r), for every 1 ≤ i ≤ k.
Pk
Pk
Thus, i=1 ci (r) ≤ α i=1 vi (r). Besides, according to the definition of vi (r)
P
and algorithm, we have ki=1 vi (r) ≤ (kV ∗ + kβV ∗ ) .Thus:
F ≤
X
w(e) =
k
X
i=1
e∈F
ci (r) ≤ α
k
X
i=1
vi (r) ≤ α(1 + β)kV ∗ ≤ α(1 + β)kOP T.
4.1. The proof of Lemma 2 and finding the best values for α and β
The proof of Lemma 2: We use the contradiction method. Assume that
for every value of r <
1
2
and every α (α > 0) we have ci (r) > αvi (r). Thus we
have:
Z
0
1
2
ci (r) > αvi (r)
Z 12
ci (r)
dr > α
dr
vi (r)
0
According to Lemma 1, the function vi (r) is not differentiable at only a finite
number of point. We call these points r0 = 0 ≤ r1 ≤ ... ≤ rl ≤ rl+1 = 12 . Thus
we have:
Z
0
1
2
X
1 dvi (r)
(
)dr =
vi (r) dr
j=0
1
Z
≤ lnvi (
10
rj+1
rj
1 dvi (r)
(
)dr
vi (r) dr
1−
) − lnvi (0).
2
Since vi (r) is an increasing function, we have
1
≤ lnvi ( ) − lnvi (0),
2
as well vi (0) = βV ∗ as vi ( 12 ) ≤ βV ∗ + V ∗ . Thus, we have:
ln(
vi ( 12 )
βV ∗ + V ∗
α
)
≥
ln(
)>
βV ∗
vi (0)
2
α
β+1
)> .
ln(
β
2
(*)
In order to reach a contradiction, we have to choose values for α and β such
that the inequality (*) will not be true. On the other hand, the approximation
factor of the algorithm is dependent on these two parameters directly. So we
have to choose the appropriate value for α and β. Indeed, to find the best value
for α and β, we should solve the following nonlinear program:
minimize α(1 + β)
subject to ln(
β+1
α
)≤
β
2
α, β > 0.
We have solved this nonlinear program using Matlab software and found the
optimal value of α and β. These values are as follows α = 0.1 and β = 20.504. If
we put these values in the inequality (*), the contradiction is reached and Lemma
2 is proved. Using these values for α and β, the algorithm is an approximation
algorithm with factor (2.1504)k for the Directed Multi-Multiway cut problem.
5. Approximation Algorithm for Directed Multicut
Similar to the LP presented in the previous section can be provided an LP for
the Directed Multicut problem. We define a decision variable x(e) for each edge
11
e which is as follows. If e belongs to the directed multicut, x(e) = 1, Otherwise
x(e) = 0. The purpose is to find a directed multicut with the minimum weight
which cuts each directed path from si to ti for 1 ≤ i ≤ k. We represent the set
of all directed paths from si to ti for 1 ≤ i ≤ k with P . An linear programming
for the problem is as follows:
minimize
X
w(e)x(e)
e∈E
subject to
X
e∈p
x(e) ≥ 1, ∀p ∈ P
x(e) ≥ 0, ∀e ∈ E.
In this LP, there is a constraint for each path. On the other hand, we may
have an exponential number of paths with respect to the input size and as a
result, an exponential number of constraints. Nevertheless, we can solve this LP
in polynomial time, using the ellipsoid algorithm. For this LP, the separation
oracle operates as following: we get a solution x and assume that the length of
each edge e is equal to x(e). Then, we find the shortest directed path from si
to ti , for 1 ≤ i ≤ k. If the shortest path from si to ti , is more than 1, then the
P
constraint e∈p x(e) ≥ 1 is true for all paths from si to ti . So this LP can be
solved in polynomial time.
We provide a direct version of definitions like those used in region growth
technique in [1]. We define a distance on edges, assume that x is an optimal
solution for LP, let x(e) be the length of edge e. We show the shortest path
from u to v, which is based on x(e), with dist(u, v). If there is not any directed
path from u to v, the value of dist(u, v) is the shortest path between u and v
in the graph, without noticing the direction of edges. Now we define:
Bx (si , r) = {v ∈ V : dist(si , v) ≤ r}.
Bx (si , r) is an area like a ball with center si , where 1 ≤ i ≤ k, and radius
r ∈ R. Assume that the product of w(e)x(e) is equal to the volume of edge
e. Thus, the solution of the LP is the minimum volume of edges such that
12
dist(si , ti ) ≥ 1, for 1 ≤ i ≤ k. Assume that x is an optimal solution for the LP.
P
Let V ∗ = e∈E w(e)x(e) be the volume of all edges, indeed V ∗ is the optimal
value of LP. We know that V ∗ ≤ OP T such that OP T is the optimal value for
the IP. vx (si , r) is defined as follows:
X
vx (si , r) = βV ∗ +
X
w(e)x(e) +
e=(u,v)∈E
u∈Bx (si ,r)
v ∈B
/ x (si ,r)
e=(u,v)∈E
u,v∈Bx (si ,r)
w(e)(r − dist(si , u)).
Let δ(s) be the set of all edges which only one of their heads is in the set s.
For a given radius r, we define:
wt(δ(Bx (si , r))) =
X
w(e).
e∈δ(Bx (si ,r))
According to these definitions, we can express Lemma 4, which is used in
the proof of Lemma 5.
Lemma 4: The function vx (si , r) is differentiable in (0, ∞) except some
finite numbers of points. The derivative of this function is wt(δ(Bx (si , r))).
Lemma 5 demonstrates that in directed graphs, we can always find a radius
r < 21 , such that the cost vx (si , r) is an upper bound for wt(δ(Bx (si , r))).
Lemma 5: Assume that x is a feasible solution for LP. For every si there
is a r <
1
2
and at least an α (α > 0) such that the following inequality is true:
wt(δ(Bx (si , r))) ≤ αvx (si , r).
The proof of Lemma 4 and Lemma 5 is similar to the proof of Lemma 1 and
Lemma 2, respectively. In the rest of the paper, for simplicity we assume that
wt(r) = wt(δ(Bx (si , r))) and v(r) = vx (si , r).
Our polynomial time approximation algorithm for directed multicut is described in Algorithm 2. The algorithm first solves the LP and finds the optimal
solution x. In every iteration, the algorithm finds a pair which there is a path
13
between and finds an area with a radius that satisfies the condition in Lemma
5. Then, the algorithm puts the edges, which have been cut by the area, in set
F . If the area includes another pair (sj , tj ), and there is a path from sj to tj ,
then the algorithm removes only the central vertex of the area. Therefore, we
can cut the path between sj and tj in the next iterations.
Algorithm 2 Approximation Algorithm for Directed Multicut
Result: A Directed Multicut
F ←∅
Solve the LP and get an optimal solution x
while there is a path between si to ti , where 1 ≤ i ≤ k do
Grow a region S = Bx (si , r) until wt(r) ≤ αv(r)
Add δ(S) to F
if S contains a pair (sj , tj ), where 1 ≤ j ≤ k and i 6= j AND there is a
path form sj to tj then
Remove si and incident edges with si form the graph
else
Remove S and δ(S) form the graph
end
end
Return F
Lemma 6: Algorithm 2 returns a directed multicut.
Proof: Consider the ball Bx (si , r). The center of this ball is si . Vertex
ti cannot be in this ball because the radius of the ball is less than
1
2
(r < 21 ).
Beside this, according to the LP constraint, we know that the distance between
every pair (si , ti ), where 1 ≤ i ≤ k, should be more than 1. The only case may
lead to problems is that there is a pair (sj , tj ), where 1 ≤ j ≤ k and i 6= j, in
Bx (si , r) such that there is a path from sj to tj . In this case, the algorithm only
removes the central vertex si from the graph, thus the pair (sj , tj ) and also the
14
path between these vertices are still in the graph. In the rest of iterations, the
algorithm will make them disconnected.
Theorem 2: Algorithm 2 is an O(k)-approximation algorithm for the Directed Multicut problem.
Proof: We demonstrate the set of vertices in the ball Bx (si , r) with Bi . We
assume that Bi = ∅ when no ball is selected for vertex si . We also demonstrate
the set of cut edges for Bi with Fi . It means Fi is equal to δ(Bi ). Thus, we have
Sk
F = i=1 Fi . Assume that Vi is equal to the volume of all edges which are in
the ball Bi and also the volume of edges which have one head in Bi . According
to this definition, we have Vi ≥ vx (si , r) − βV ∗ because Vi includes the volume
of all edges in Fi . But vx (si , r) is contained only some part of these edges and
an addition value βV ∗ . According to Lemma 5 and the value chosen for r in
the algorithm, we have wt(Fi ) ≤ αvx (si , r) ≤ α(Vi + βV ∗ ). We know that the
algorithm may not remove the edges which are incident with vertices in Bi in
this iteration. Thus, an edge may belong to more than one area, on the other
Pk
hand, there are at most k areas. Therefore, i=1 Vi ≤ kV ∗ . So we have the
following inequalities:
X
w(e) =
e∈F
k
X
i=1
wt(Fi ) ≤ α
k
X
i=1
(Vi + βV ∗ ) ≤ α(1 + β)kV ∗ ≤ α(1 + β)kOP T.
Similar to section 4, the optimal value for α is 0.1 and β is 20.504. So
Algorithm 2 is an (2.1504)k-approximation algorithm for the Directed Multicut
problem.
6. Conclusions
In this paper, we design approximation algorithms for the directed multimultiway cut and directed multicut problems using the region growing technique
15
[1, 2]. By this paradigm, we give for each problem an O(k)-approximation
algorithm. The works previously done on these problems need to solve k linear
programming, whereas our algorithms require only one linear programming.
Both algorithms use the same linear programming relaxation. A question of
interest is to find the integrality gap of the linear programming relaxation for
these problems.
References
[1] Garg, N., Vazirani, V. V., Yannakakis, M. (1996). Approximate max-flow
min-(multi) cut theorems and their applications. SIAM Journal on Computing, 25(2), 235-251.
[2] Avidor, A., Langberg, M. (2007). The multi-multiway cut problem. Theoretical Computer Science, 377(1-3), 35-42.
[3] Garg, N., Vazirani, V. V., Yannakakis, M. (1994, July). Multiway cuts in directed and node weighted graphs. In International Colloquium on Automata,
Languages, and Programming (pp. 487-498). Springer, Berlin, Heidelberg.
[4] Naor, J., Zosin, L. (2001). A 2-approximation algorithm for the directed
multiway cut problem. SIAM Journal on Computing, 31(2), 477-482.
[5] Dahlhaus, E., Johnson, D. S., Papadimitriou, C. H., Seymour, P. D., Yannakakis, M. (1992, July). The complexity of multiway cuts. In Proceedings
of the twenty-fourth annual ACM symposium on Theory of computing (pp.
241-251). ACM.
[6] Clinescu, G., Karloff, H., Rabani, Y. (1998, May). An improved approximation algorithm for multiway cut. In Proceedings of the thirtieth annual
ACM symposium on Theory of computing (pp. 48-52). ACM.
[7] Karger, D. R., Klein, P., Stein, C., Thorup, M., Young, N. E. (2004). Rounding algorithms for a geometric embedding of minimum multiway cut. Mathematics of Operations Research, 29(3), 436-461.
16
[8] Zhang, P., Zhu, D., Luan, J. (2012). An approximation algorithm for the
Generalized k-Multicut problem. Discrete Applied Mathematics, 160(7-8),
1240-1247.
[9] Even, G., Naor, J. S., Rao, S., Schieber, B. (2000). Divide-and-conquer approximation algorithms via spreading metrics. Journal of the ACM (JACM),
47(4), 585-616.
[10] Leighton, T., Rao, S. (1988, October). An approximate max-flow min-cut
theorem for uniform multicommodity flow problems with applications to
approximation algorithms. In Foundations of Computer Science, 1988., 29th
Annual Symposium on (pp. 422-431). IEEE.
[11] Even, G., Naor, J. S., Schieber, B., Sudan, M. (1998). Approximating minimum feedback sets and multicuts in directed graphs. Algorithmica, 20(2),
151-174.
[12] Klein, P. N., Plotkin, S. A., Rao, S., Tardos, E. (1997). Approximation
algorithms for Steiner and directed multicuts. Journal of Algorithms, 22(2),
241-269.
[13] Bang-Jensen, J., Yeo, A. (2014). The complexity of multicut and mixed
multicut problems in (di) graphs. Theoretical Computer Science, 520, 8796.
[14] Cheriyan, J., Karloff, H., Rabani, Y. (2005). Approximating directed multicuts. Combinatorica, 25(3), 251-269.
[15] Gupta, A. (2003). Improved approximation algorithm for directed multicut.
In Proceedings of the Fourteenth Annual ACMSIAM Symposium on Discrete
Algorithms (SODA) (pp. 12-14).
[16] Kanj, I., Lin, G., Liu, T., Tong, W., Xia, G., Xu, J., ... Zhu, B. (2015).
Improved parameterized and exact algorithms for cut problems on trees.
Theoretical Computer Science, 607, 455-470.
17
[17] Saks, M., Samorodnitsky, A., Zosin, L. (2004). A lower bound on the integrality gap for minimum multicut in directed networks. Combinatorica,
24(3), 525-530.
[18] Chuzhoy, J., Makarychev, Y., Vijayaraghavan, A., Zhou, Y. (2016). Approximation algorithms and hardness of the k-route cut problem. ACM
Transactions on Algorithms (TALG), 12(1), 2.
18
| 8cs.DS
|
arXiv:1009.4683v1 [cs.CE] 23 Sep 2010
Efficient Computation of Optimal Trading Strategies
Malik Magdon-Ismail
Dept Computer Science
Rensselaer Polytechnic Institute
Room 207, Lally Bldg
110 8th Street
Troy, NY 12180
[email protected]
Victor Boyarshinov
Dept Computer Science
Rensselaer Polytechnic Institute
Room 207, Lally Bldg
110 8th Street
Troy, NY 12180
[email protected]
Abstract
Given the return series for a set of instruments, a trading strategy is a switching function
that transfers wealth from one instrument to another at specified times. We present efficient
algorithms for constructing (ex-post) trading strategies that are optimal with respect to the
total return, the Sterling ratio and the Sharpe ratio. Such ex-post optimal strategies are useful
analysis tools. They can be used to analyze the “profitability of a market” in terms of optimal
trading; to develop benchmarks against which real trading can be compared; and, within an
inductive framework, the optimal trades can be used to to teach learning systems (predictors)
which are then used to identify future trading opportunities.
1
Introduction
A trader has in mind the task of developing a trading system that optimizes some profit criterion,
the simplest being the total return. A more conservative approach is to optimize a risk adjusted
return. Widely followed measures of risk adjusted returns are the Sterling Ratio and Sharpe Ratio.
In an enviroment where markets exhibit frequent crashes and portfolios encounter sustained periods
of losses, the Sterling ratio and the Sharpe ratio have emerged as the leading performance measures
used in the industry. Given a set of instruments, a trading strategy is a switching function that
transfers the wealth from one instrument to another. In this paper, we consider the problem of
finding optimal trading strategies, i.e., trading strategies that maximize a given optimality criterion,
on historical data. In particular, we consider optimal strategies with respect to the total cumulative
return, as well as with respect to various risk adjusted measures of return (the Sterling ratio and
variants of the Sharpe ratio). Finding the optimal trading strategy for non-zero transactions cost
is a path dependent optimization problem even when the price time series is known. A brute force
1
approach to solving this problem would search through the space of all possible trading strategies,
keeping only the one satisfying the optimality criterion. Since the number of possible trading
strategies grows exponentially with time, the brute force approach leads to an exponential time
algorithm1 , which for all practical purposes is infeasible – even given the pace at which computing
power grows. The contribution in this work is to give efficient (polynomial time) algorithms to
compute the optimal trading strategy for various profit objectives, with or without constraints on
the number of trades that can ber made. Our motivations for constructing such optimal strategies
are:
(i) Knowing what the optimal trades are, one can take an inductive approach to real trading: on
historical data, one can construct the optimal trades; one can then correlate various market
and/or technical indicators with the optimal trades. These indicators can then be used to
identify future trading opportunities. In a sense, one can try to learn to predict good trading
opportunities based on indicators by emulating the optimal trading strategy. A host of such
activity within the inductive framework, goes under the name of financial engineering.
(ii) The optimal trading performance under certain trading constraints can be used as a benchmark for real trading systems. For example, how good is a trading system that makes ten
trades with a Sterling ratio of 4 over a given time period? One natural comparison is to
benchmark this trading strategy against a Sterling-optimal trading strategy that makes at
most ten trades over the same time period.
(iii) Optimal trading strategies (with or without constraints) can be used to quantitatively rank
various markets (and time scales) with respect to their profitability according to a given
criterion. So for example, one could determine the optimal time scale on which to trade a
particular market, or given a set of markets, which is the most (risk adjusted) profit-friendly.
(iv) Given a stochastic model for the behavior of a pair of instruments, one can use the efficient
algorithms presented here to construct ex-ante optimal strategies using simulation. To be
more specific, note that the optimal strategy constructed by our algorithms requires full
knowledge of the future price paths. The stochastic model can be used to generate sample
paths for the instruments. These sample paths can be used to compute the optimal trading
strategy given the current history and information set. One then has a sample set of future
paths and corresponding optimal trading strategies on which to base the current action. Note
that such a stochastic model for future prices would have to take into account correlations
(including auto-correlations) among the instruments.
1
The asymptotic running time of an algorithm is measured in terms of the input size n. If the input is a time
sequence of n price data points, then polynomial time algorithms have run time that is bounded by some polynomial
in n. Exponential time algorithms have running time greater than some exponentially growing function in n [6].
2
It is beyond the scope of the current discussion to develop these applications. Our main goal here
is to present the algorithms for obtaining optimal trading strategies, given a price time series.
1.1
Trading Model
We now make the preceeding discussion more precise. We consider optimal trading strategies on
two instruments, for concreteness, a stock S and a bond B with price histories {S0 , . . . , Sn } and
{B0 , . . . , Bn } over n consecutive time periods, ti , i ∈ {1, . . . , n}. Thus, for example, over time period
ti , the price of stock moved from Si−1 to Si . We denote the return sequence for the two instruments
by {s1 , . . . , sn } and {b1 , . . . , bn } respectively: si = log
Si
Si−1 ,
i
and correspondingly, bi = log BBi−1
. We
assume that one of the instruments is the benchmark instrument, and that all the equity is held
in the benchmark instrument at the begining and end of trading. The bond is usually considered
the benchmark instrument, and for illustration, we will follow this convention. The trivial trading
strategy is to simply hold onto bond for the entire duration of the trading period. It is useful to
define the excess return sequence for the stock, ŝi = si − bi . When the benchmark instrument is the
bond, the excess return as we defined it is the conventionally used one. However, one may want to
measure performances of a trading strategy with respect to the S&P 500 as benchmark instrument,
in which case the excess return would be determined relative to the S&P 500 return sequence.
The excess return sequence for the bond is just the sequence of zeros, b̂i = 0. Conventionally,
the performance of a strategy is measured relative to some trivial strategy, so the excess return
sequence will be the basis of most of our performance measures.
Definition 1.1 (Trading Strategy) A trading strategy T is a boolean n-dimensional vector indicating where the money is at the end of time period ti :
T [i] =
1
0
if money is in stock at the end of ti ,
if money is in bond at the end of ti .
We assume that T [0] = T [n] = 0, i.e., all the money begins and ends in bond. A trade is entered
at time ti if T [i] = 0 and T [i + 1] = 1. A trade is exited at time ti if T [i] = 1 and T [i + 1] = 0.
The number of trades made by a trading strategy is equal to the number of trades that are entered.
We make the following assumptions regarding the trading:
A1 [All or Nothing] : The position at all times is either entirely bond or entirely stock.
A2 [No Market Impact] : Trades can be placed without affecting the quoted price.
A3 [Fractional Market] : Arbitrary amounts of stock or bond can be bougnt or sold at any time.
A4 [Long Strategies] : We assume that we can only hold long positions in stock or bond.
3
Assumption A1 is in fact not the case in many trading funds, for it does not allow legging into a
trade, or holding positions in both instruments simultaneously. While this is technicaly a restriction,
for many optimality criteria (for example return optimal strategies), one can show that there always
exists an all-or-nothing optimal strategy. Thus, we maintain this simplifying assumption for our
discussion. Further, such assumptions are typically made in the literature on optimal trading (see
for example [14]). Assumptions A2–A4 are rather mild and quite accurate in most liquid markets,
for example foreign exchange. Assumption A3 is needed for A1, since if all the money should be
transfered to a stock position, this may necessitate the purchase of a fractional number of shares.
Note that if T [i − 1] 6= T [i], then at the begining of time period ti , the position was transfered from
one instrument to another. Such a transfer will incur an instantaneous per unit transaction cost
equal to the bid-ask spread of the instrument being transfered into. We assume that the bid-ask
spread is some fraction (fB for bond and fS for stock) of the bid price.
We denote the equity curve for a trading strategy T by the vector ET , i.e., ET [i] is the equity
at the end of time period ti , with ET [0] = 1. Corresponding to the equity curve is the excess return
sequence rT for the trading strategy T , i.e., for i ≥ 1
rT [i] = log
ET [i]
− bi .
ET [i − 1]
(1)
If we ignore the bid-ask spread, then the excess return in time period ti is given by
rT [i] = ŝi T [i] = (si − bi )T [i].
(2)
The bid-ask spread affects the return, reducing it by an amount depending on T [i − 1]. Denoting
this transactions cost attributable to T [i] by ∆[i], we have that
∆[i] = −T [i − 1](1 − T [i])fˆB − (1 − T [i − 1])T [i]fˆS ,
(3)
where fˆS = log(1+fS ) and fˆB = log(1+fB ). Thus, the bid-ask spread can be viewed as introducing
an instantaneous return of −fˆB or −fˆS whenever the position is switched. To exactly which time
period this transactions cost is applied may depend on the nature of the market, i.e., it may be
applied to rT [i], rT [i − 1] or rT [i + 1]. The nature of the results will not change significantly
for either of these options, so in our algorithms, we will generally make the choice that offers the
greatest technical simplicity. For a trading strategy T , we define the total return µ(T ), the sum of
the squared returns s2 (T ), the sum of squared deviations of the returns σ 2 (T ) and the maximum
4
drawdown M DD(T ) as follows,
µ(T ) =
n
X
rT [i],
(4)
s2 (T ) =
n
X
rT [i]2 ,
(5)
σ 2 (T ) =
n
X
i=1
i=1
rT [i] −
i=1
M DD(T ) =
max −
1≤k≤l≤n
2
1
1
µ(T ) = s2 (T ) − µ2 (T ),
n
n
l
X
(6)
rT [i].
(7)
i=k
When it is clear from the context what trading strategy we are talking about, we will generally
suppress the explicit dependence on T . The performance measures that we consider in this paper
are derived from these statistics. In particular, we are interested in the total return µ, the Sterling
ratio Strl, and variants of the Sharpe ratio, Shrp1 and Shrp2 :
Strl(T ) =
µ(T )
,
M DD(T )
Shrp1 (T ) =
µ(T )
,
σ(T )
Shrp2 (T ) =
µ(T )
.
σ 2 (T )
(8)
Shrp1 is the conventionally used Sharpe ratio. Shrp2 is a more risk averse performance measure, as
it is more sensitive to the variance in the returns. Often, Strl as we have defined it is refered to as
the Calmar ratio in the literature [8], and the Sterling ratio adds a constant (for example 10%) to
the M DD in the denominator [1]. Such a constant can easily be accomodated by our algorithms,
and so we will maintain this simpler definition for the Sterling ratio.
The contribution of this paper is efficient algorithms for computing optimal trading strategies.
We will use standard O() notation in stating our results: let n be the length of the returns sequences;
we say that the run time of an algorithm is O(f (n)) if, for some constant C, the runtime is ≤ Cf (n)
for any possible return sequences. If f (n) is linear (quadratic), we say that the runtime is linear
(quadratic). We will establish the following results.
Theorem 1.2 (Return Optimal Trading Strategies) A total return optimal trading strategy
can be computed in linear time. Specifically,
i. Unconstrained Trading. A trading strategy Tµ can be computed in O(n) such that for any
other strategy T , µ(Tµ ) ≥ µ(T ).
ii. Constrained Trading. A trading strategy TµK making at most K trades can be computed in
O(K · n) such that for any other strategy T K making at most K trades, µ(TµK ) ≥ µ(T K ).
Proof: See section 2.
5
Theorem 1.3 (Sterling Optimal Trading Strategies) A Sterling optimal trading strategy can
be computed in near linear time. Specifically,
i. Unconstrained Trading. A trading strategy TStrl can be computed in O(n log n) such that
for any other strategy T , Strl(TStrl ) ≥ Strl(T ).
K making at most K trades can be computed in
ii. Constrained Trading. A trading strategy TStrl
K ) ≥ Strl(T K ).
O(n log n) such that for any other strategy T K making at most K trades, Strl(TStrl
Proof: See section 3.
Theorem 1.4 (Sharpe Optimal Trading Strategies) A Sharpe optimal trading strategy can
be computed in near quadratic time. Specifically, trading strategies TShrp1 and TShrp2 can be found
in O(n2 log n) such that for any other strategy T , Shrp1 (TShrp1 ) ≥ Shrp1 (T ) and Shrp2 (TShrp2 ) ≥
Shrp2 (T )
Proof: See section 4.
In all cases, our proofs are constructive, and so immediately give the algorithms for performing the
desired computations. Next, we discuss the existing related work, followed by a detailed discussion
of the algorithms, along with all necessary proofs.
Related Work
The body of literature on optimal trading is so enormous that we only highlight here some representative papers. The reasearch on optimal trading falls into two broad categories. The first group
is on the more theoretical side where researchers assume that instrument prices satisfy some particular model, for example the prices are driven by a stochastic process of known form; the goal is to
derive closed-form solutions for the optimal trading strategy, or a set of equations that the optimal
strategy must follow. The main drawbacks of such theoretical approaches is that their prescriptions
can only be useful to the extent that the assumed models are correct. Our work does not make any
assumptions about the price dynamics to construct ex-post optimal trading strategies.
The second group of research which is more on the practical side is focused on exploring data
driven / learning methods for the prediction of future stock prices moves and trading opportunities.
Intelligent agents are designed by training on past data and their performance is compared with
some benchmark strategies. Our results furnish (i) optimal strategies on which to train intelligent
agents and (ii) benchmarks with which to compare their performance.
6
Theoretical Approaches
Boyd et al. in [11] consider the problem of single-period portfolio
optimization. They consider the maximization of the expected return subject to different types of
constraints on the portfolio (margin, diversification, budget constraints and limits on variance or
shortfall risk). Under certain assumptions on the returns distribution, they reduce the problem to
numerical convex optimization. Similarily, Thompson in [14] considered the problem of maximizing
the (expected) total cumulative return of a trading strategy under the assumption that the asset
price satisfies a stochastic differential equation of the form dSt = dBt + h(Xt )dt, where Bt is a
Brownian motion, h is a known function and Xi is a Markov Chain independent of the Brownian
motion. In this work, he assumes fixed transaction costs and imposes assumptions A1, A2, A4
on the trading. He also imposes a stricter version of our assumption A3: at any time, the trader
can have only 0 or 1 unit of stock. He proves that the optimal trading strategy is the solution of a
free-boundary problem, gives explicit solutions for several functions h and provides bounds on the
transaction cost above which it is optimal never to buy the asset at all.
Pliska et al. in [4] and Bielecki et al. in [3] considered the problems of optimal investment
with stochastic interest rates in simple economies of bonds and a single stock. They characterize
the optimal trading strategy in terms of a nonlinear quasi-variational inequality and develop a
numerical approaches to solving these equations.
Some work has been done within risk-return frameworks. Berkelaar and Kouwenberg in [2]
considered asset allocation in a return versus downside-risk framework, with closed-form solutions
for asset prices following geometric Brownian motions and constant interest rates. Liu in [10]
consider the optimal investment policy of a constant absolute risk aversion (CARA) investor who
faces fixed and proportional transaction costs when trading multiple uncorrelated risky assets.
Zakamouline in [16] studies the optimal portfolio selection problem using Markov chain approximation for a constant relative risk averse investor who faces fixed and proportional transaction
costs and maximizes expected utility of the investor’s end-of-period wealth. He identifies three
disjoint regions (Buy, Sell and No-Transaction) to describe the optimal strategy.
Choi and Liu in [13] considered trading tasks faced by an autonomous trading agent. An autonomous trading agent works as follows. First, it observes the state of the environment. According
to the environment state, the agent responds with an action, which in turn influences the current
environment state. In the next time step, the agent receives a feedback (reward or penalty) from
the environment and then perceives the next environment state. The optimal trading strategy for
the agent was constructed in terms of the agent’s expected utility (expected accumulated reward).
Cuoco et al. in [7] considered Value at Risk as a tool to measure and control the risk of the
trading portfolio. The problem of a dynamically consistent optimal porfolio choice subject to the
Value at Risk limits was formulated and they proved that the risk exposure of a trader subject to
a Value at Risk limit is always lower than that of an unconstrained trader and that the probability
7
of extreme losses is also decreased.
Mihatsch and Neuneier in [12] considered problem of optimization of a risk-sensitive expected
return of a Markov Decision Problem. Based on an extended set of optimality equations, risksensitive versions of various well-known reinforcement learning algorithms were formulated and
they showed that these algorithms converge with probability one under reasonable conditions.
Data Driven Approaches
Moody and Saffell in [9] presented methods for optimizing portfolios,
asset allocations and trading systems based on a direct reinforcement approach, which views optimal trading as a stochastic control problem. They developed reccurent reinforcement learning to
optimize risk-adjusted investment returns like the Sterling Ratio or Sharpe Ratio, while accounting
for the effects of transaction costs.
Liu et al. in [15] proposed a learning-based trading strategy for portfolio management, which
aims at maximizing the Sharpe Ratio by actively reallocating wealth among assets. The trading
decision is formulated as a non-linear function of the latest realized asset returns, and the function
cam be approximated by a neural network. In order to train the neural network, one requires a
Sharpe-Optimal trading strategy to provide the supervised learning method with target values. In
this work they used heuristic methods to obtain a locally Sharp-optimal trading strategy. The
transaction cost was not taken into consideration. Our methods can be considerably useful in the
determination of target trading strategies for such approaches.
2
Return-Optimal Trading Strategies
We use the notation [ti , tj ] to denote the set of time periods {ti , ti+1 , . . . , tj }. In order to compute
the return optimal strategies, we will use a dynamic programming approach to solve a more general
problem. Specifically, we will construct the return optimal strategies for every prefix of the returns
sequence. First we consider the case when there is no restriction on the number of trades, and
then the case when the number of trades is constrained to be at most K. Although we maintain
assumptions A1-A4 for simplicity, A1, A3 and A4 can be relaxed without much additional effort.
2.1
Overview of the Algorithm
The basic idea of the algorithm is to consider the optimal strategy to time period ti . This strategy
must end in either stock or bond. Suppose that it ends in stock, then it must arrive at the final
position in stock at ti by either passing through stock or bond at time ti−1 . Thus, the optimal
strategy which ends in stock at time ti must be either the optimal strategy which passes through
stock at time ti−1 followed by holding the stock for one more time period, or the optimal strategy
which passes through bond at time ti−1 and then makes a trade into the stock for the next time
8
period. Whichever is better among these two options yields the optimal strategy to time period
ti that ends in stock. A similar argument applies to the optimal strategy to time ti that ends in
bond. Thus, having computed the optimal strategies which end in stock and bond to time ti−1 , we
can compute the optimal strategies which end in stock and bond to time ti . This induction can be
propagated to obtain the final result.
When there are constraints on the number of trades, we only need to slightly modify the above
argument. We would like to compute the optimal strategy which ends (say) in stock and makes at
most K trades. Any such strategy has to be one of two possibilities: it makes at most K trades
ending in stock at time ti−1 , or it makes at most K − 1 trades, ending in bond at time ti−1 . If it
ended in bond, it can only make at most K − 1 trades because one additional trade will be required
to convert from bond at ti−1 to stock at ti . Thus the inductive construction will start with K = 0
which is to hold bond. Assuming we have computed all the optimal strategies for K = k to all
times {ti }, we can then compute all the optimal strategies for K = k + 1 to all times.
We now present the details and proofs of correctness for these two algorithms.
2.2
Unconstrained Return-Optimal Trading Strategies
First we give the main definitions that we will need in the dynamic programming algorithm to
compute the optimal strategy. Consider a return-optimal strategy for the first m time periods,
[t1 , tm ]. Define S[m, 0] (S[m, 1]) to be a return-optimal strategy over the first m periods ending
in bond (stock) at time tm . For ℓ ∈ {0, 1}, let µ[m, ℓ] denote the return of S[m, ℓ] over [t1 , tm ],
i.e., µ[m, ℓ] = µ(S[m, ℓ]). Let Prev[m, ℓ] denote the penultimate position of the optimal strategy
S[m, ℓ] just before the final time period tm .
The optimal strategy S[m, ℓ] must pass through either bond or stock at time period m − 1.
Thus, S[m, ℓ] must be the extension of one of the optimal strategies {S[m − 1, 0], S[m − 1, 1]} by
adding the position ℓ at time period tm . More specifically,
S[m, ℓ] =
{S[m − 1, 1], ℓ},
{S[m − 1, 0], ℓ}.
In particular, S[m, ℓ] will be the extension that yields the greatest total return. Using e̊q:ret and
e̊q:tx, we have that
µ({S[m − 1, 0], ℓ}) = µ[m − 1, 0] + ŝm ℓ − fˆS ℓ,
µ({S[m − 1, 1], ℓ}) = µ[m − 1, 1] + ŝm ℓ − fˆB (1 − ℓ).
9
Since µ[m, ℓ] is the maximum of these two values, we have the following recursion,
n
o
µ[m, ℓ] = max µ[m − 1, 0] + ŝm ℓ − fˆS ℓ, µ[m − 1, 1] + ŝm ℓ − fˆB (1 − ℓ) .
The position of the optimal strategy S[m, ℓ] just before time period m is given by the ending
position of the strategy that was extended. Thus,
0 if µ[m − 1, 0] + ŝm ℓ − fˆS ℓ ≥ µ[m − 1, 1] + ŝm ℓ − fˆB (1 − ℓ),
Prev[m, ℓ] =
1 otherwise.
If we already know µ[m − 1, 0] and µ[m − 1, 1], then we can compute µ[m, ℓ] and Prev[m, ℓ] for
ℓ ∈ {0, 1} in constant time. Further, we have that µ[1, 0] = 0 and µ[1, 1] = ŝ1 − fˆS , and so, by a
straight forward induction, we can prove the following lemma.
Lemma 2.1 Prev[m, ℓ] for all ℓ ∈ {0, 1} and m ≤ n can be computed in O(n).
The optimal strategy Tµ is exactly S[n, 0]. Prev[n, 0] gives the position at tn−1 , and the optimal
way to reach Prev[n, 0] at tn−1 is given by optimal strategy S[n − 1, Prev[n, 0]]. Continuing
backward in this fashion, it is easy to verify that we can reconstruct the full strategy Tµ using the
following backward recursion:
Tµ [n] = 0,
Tµ [m] = Prev[m + 1, Tµ [m + 1]], for 1 ≤ m < n.
Thus, a single backward scan is all that is required to compute Tµ [i] for all i ∈ {1, . . . , n}, which
is linear time, and so we have proved the first part of Theorem 1.2. Further, it is clear that the
algorithm requires memory that is linear in n to store Prev[m, ℓ]. While we have assumed that
the algorithm works with excess returns, the optimal strategy does not depend on this assumption,
thus the algorithm works correctly even with the actual return sequences. The generalization of
this algorithm to N > 2 instruments is straightforward by suitably generalizing a trading strategy.
S[m, ℓ] retains its definition, except now ℓ ∈ {0, . . . , N − 1}. To compute µ[m, ℓ] will need to take a
maximum over N terms depending on µ[m − 1, ℓ′ ], and so the algirithm will have runtime O(N n).
One concern with the unconstrained optimal strategy is that it may make too many trades. It
is thus useful to compute the optimal strategy that makes at most a given number of trades. We
discuss this next.
10
2.3
Constrained Return-Optimal Strategies
We suppose that the number of trades is constrained to be at most K. It is more convenient to
consider the number of jumps k, which we define as the sum of the number of trades entered and
the number exited. For a valid trading strategy, the number of trades entered equals the number
of trades exited, so k = 2K. Analogous to S[m, ℓ] in the previous section, we define S[m, k, ℓ] to be
the optimal trading strategy to time period tm that makes at most k jumps ending in instrument
ℓ. Let µ[m, k, ℓ] be the return of strategy S[m, k, ℓ], and let Prev[m, k, l] store the pair (k′ , ℓ′ ),
where ℓ′ is the penultimate position of S[m, k, ℓ] at tm−1 that leads to the end position ℓ, and k′
is the number of jumps made by the optimal strategy to time period tm−1 that was extended to
S[m, k, ℓ].
The algorithm once again follows from the observation that the the optimal strategy S[m, k, ℓ]
must pass through either bond or stock at tm−1 . A complication is that if the penultimate position
is bond and ℓ = 0, then at most k jumps can be used to get to thhe penultimate position, however,
if ℓ = 1, then only at most k − 1 jumps may be used. Similarily if the penultimate position is stock.
We thus get the following recursion,
o
n
µ[m, k, 0] = max µ[m − 1, k, 0], µ[m − 1, k − 1, 1] − fˆB ,
o
n
µ[m, k, 1] = max µ[m − 1, k − 1, 0] + ŝm − fˆS , µ[m − 1, k, 1] + ŝm .
This recursion is initialized with µ[m, 0, 0] = 0 and µ[m, 0, 1] = NULL for 1 ≤ m ≤ n. Once
µ[m, k, ℓ] is computed for all m, ℓ, then the above recursion allows us to compute µ[m, k + 1, ℓ] for
all m, ℓ. Thus, the computation of µ[m, k, ℓ] for 1 ≤ m ≤ n, 0 ≤ k ≤ 2K and ℓ ∈ {0, 1} can be
accomplished in O(nK). Once again, the strategy that was extended gives Prev[m, k, ℓ],
Prev[m, k, 0] =
(k, 0)
(k − 1, 1)
(k − 1, 0)
Prev[m, k, 1] =
(k, 1)
if µ[m − 1, k, 0] > µ[m − 1, k − 1, 1] − fˆB ,
otherwise.
if µ[m − 1, k − 1, 0] + ŝm − fˆS > µ[m − 1, k, 1] + ŝm ,
otherwise.
Since computing µ[m, k, ℓ] immediately gives Prev[m, k, ℓ], we have the following lemma,
Lemma 2.2 Prev[m, k, ℓ] for all ℓ ∈ {0, 1}, m ≤ n and k ≤ 2K can be computed in O(nK).
TµK is given by S[n, 2K, 0], and the full strategy can be reconstructed in a single backward scan
11
using the following backward recursion (we introduce an auxilliary vector κ),
TµK [n] = 0,
(κ[n − 1], TµK [n − 1]) = Prev[n, 2K, TµK [n]),
(κ[m], TµK [m]) = Prev[m + 1, κ[m + 1], TµK [m + 1]), for 1 ≤ m < n − 1.
Since the algorithm needs to store Prev[m, k, ℓ] for all m, k, the memory requirement is O(nK).
Once again, it is not hard to generalize this algorithm to work with N instruments, and the resulting
run time will be O(nN K).
3
Sterling-Optimal Trading Strategies
It will first be useful to discuss some of the M DD properties of the return-optimal strategy Tµ , as
these properties will have implications on our algorithm to determine Sterling-optimal strategies.
For a strategy T , it is useful to define the cumulative return series, CT [i] as the sum of the returns,
P
CT [i] = ij=1 rT [j]. Note that µ(Tµ ) = CTµ [n] ≥ CT [n] = µ(T ) for any strategy T . The equity
P
curve is given by ET [i] = exp CT [i] + ij=1 bj .
3.1
Upper bound on the Maximum drawdown of a Sterling Optimal Strategy
First, we will upper bound M DD(Tµ ). Intuitively, the M DD of the return optimal trading strategy
can not be larger than the bid-ask spread. This bound will be useful because the M DD of the Tµ
serves as an upper bound for the M DD of the Sterling-optimal strategy,
Lemma 3.1 M DD(TStrl ) ≤ M DD(Tµ ).
Proof: By definition,
µ(TStrl )
M DD(TStrl )
≥
µ(T )
M DD(T )
for any T . Thus, M DD(TStrl ) ≤
µ(TStrl )
µ(T ) M DD(T
) for
any T . Choosing T = Tµ and noting that µ(TStrl ) ≤ µ(Tµ ), we obtain the desired result.
Since the cost (in terms of the cumulative return) of entering and exiting a trade is −(fˆS + fˆB ), no
segment of the optimal trading strategy Tµ should lose more than this in return.
Lemma 3.2 For any i < j, CTµ [j] − CTµ [i] ≥ −(fˆS + fˆB ).
Proof: Suppose, for contradiction, that for some i < j, CTµ [j] − CTµ [i] < −(fˆS + fˆB ). By setting
Tµ [i + 1], . . . , Tµ [j] to be all equal to 0, it is easy to verify that the cumulative return of the strategy
must increase, which contradicts the optimality of Tµ .
12
For technical convenience, we will assume that the transactions cost when entering a trade is
assigned to the time period prior to the entry, and the transactions cost when exiting a trade is
assigned to the time period after the trade. Note that just prior to entering, the position is 0 and
so it will now have a return of −fˆS , and just after exiting, the position is 0, and will now have a
return of −fˆB .
Let fsp = fˆS + fˆB . From Lemma 3.2, no segment of the optimal strategy can lose more than fsp ,
and so this immediately gives an upper bound on M DD(Tµ ). For the trivial strategy that makes
no trades, the M DD is 0. If a strategy makes exactly one trade, then there is a drawdown of at
least fˆS at the begining, and of at least fˆB at the end. If at least two trades are made, then there
is a drawdown of at least fsp between the exit of one trade and the entry of another, and since
the drawdown cannot exceed fsp , the M DD must therefore equal fsp. We thus have the following
lemma.
Lemma 3.3 (M DD of T µ ) If Tµ makes no trades, M DD(Tµ ) = 0. If Tµ makes one trade,
max{fˆS , fˆB } ≤ M DD(Tµ ) ≤ fsp . If Tµ makes at least two trades, M DD(Tµ ) = fsp .
note that if we relax assumption A1, then by legging into a trade, it may be possible to decrease
the drawdown, in which case Lemma 3.3 would no longer be valid. We are now ready to discuss
the O(n log n) algorithms to obtain Sterling-optimal trading strategies. First we will consider
unconstrained Sterling optimal strategies, and then we will require number of trades ≤ K.
3.2
Overview of the Algorithm
Based on the results in the previous section, the main observation is that if the Sterling optimal
strategy makes two or more trades, then its M DD will be the bid-ask spread, and hence it must
be the return optimal strategy. Thus we only need to consider the case when the Sterling optimal
strategy makes exactly one trade. A useful observatioin is that no optimal strategy will exit a
trade in the middle of a sequence of positive returns or enter a trade in the middle of a sequence
of negative returns. Thus the return sequence can be contracted by combining any sequence of
positive returns into a single return, ans similarily any sequence of negative returns. Possible entry
points ai correspond to the beginning of a positive return and exit points the end of a sequence of
positive returns.
The main task is to find the Sterling optimal strategy making exactly one trade. We show
that this problem can be reduced to convex hull operations in 2-dimensions, and hence obtain an
O(n log n) algorithm for constructing the optimal solution.
In order to construct the Sterling optimal strategy which makes at most K trades, the basic
idea is to start with the return optimal strategy whose trade intervals cannot be enlarged (maximal
return optimal strategies). If this strategy makes at most K trades, then we are back to the
13
unconstrained case. Otherwise, we show that one can successively merge neighboring trades in the
return optimal strategy to obtain a sterling optimal strategy. The main difficulty in the algorithm
is to determine which trades to merge, and we will show that a greedy merging strategy is in fact
optimal.
3.3
Unconstrained Sterling-Optimal Strategies
For a degenerate trading system with all returns equal to zero, we define its Sterling ratio as 1. The
only trading system with a M DD of 0 is a degenerate trading system, so with this definition, the
Sterling ratio is defined for all possible trading systems. The computation of the Sterling-optimal
trading system breaks down into three cases, according to the number of trades its makes:
Sterling-optimal that makes zero trades. Sterling Ratio is 1.
Sterling-optimal that makes one trade. The trading strategy contains a single interval of 1’s.
Sterling-optimal that makes at least two trades. Any trading system that makes at least
two trades has an M DD ≥ fsp . Since M DD(Tµ ) ≤ fsp (Lemma 3.3), Tµ has the smallest
M DD among all such systems. Since it also has the highest total return, we conclude that if
the Sterling-optimal system makes at least two trades, then TStrl = Tµ .
The first case is trivially computed. The third case, i.e., the Sterling optimal strategy that makes
at least two trades can be computed in linear time using the dynamic programming algorithm to
compute Tµ . If we also compute the Sterling-optimal system that makes exactly one trade, then, we
solve our problem by taking the case with the maximum Sterling ratio. We now focus on finding
the trading strategy that makes only one trade and has greatest Sterling Ratio among all such
strategies.
Let T be a strategy that makes exactly one trade. The trade interval is the interval of time
periods, [ti , tj ] on which T = 1, i.e., the trade interval is an interval of 1’s in the trading strategy.
An elementary algorithm that considers all the O(n2 ) possible trade intervals, picking the best is
a quadratic time algorithm. The remainder of this section is devoted to providing an algorithm
which computes such a strategy in O(n log n) time, which will complete the proof of the first part
of Theorem 1.3. In fact the algorithm that we present is a much more general algorithm that
computes the single interval that optimizes a general class of optimality criteria. This algorithm
will be useful when we discuss the Sharpe-optimal strategy.
Consider a consecutive sequence of time periods ti , ti+1 , . . . , ti+k where k ≥ 1, with all the excess
returns non-negative and the last one positive, i.e., ŝi ≥ 0, ŝi+1 ≥ 0, . . . , ŝi+k > 0.
Lemma 3.4 Either the optimal single trade interval does not intersect these time periods, or an
optimal single interval can be chosen to contain this interval.
14
Proof:
Suppose that T [i + j] = 1 and T [i + j + 1] = 0 for some 0 ≤ j < k. Extend the trading
interval by setting T [i + j + 1] = 1, . . . , T [i + k] = 1, which adds positive return, without increasing
the M DD, contradicting the optimality of the original interval. On the other hand, suppose that
T [i + j] = 1 and T [i + j − 1] = 0 for some 0 < j ≤ k. Once again, by extending the trading interval,
setting T [i] = 1, . . . , T [i + j] = 1, we add non-negative returns, without increasing the M DD hence
this new interval is at least as good as the previous interval.
A similar result holds for a sequence of consecutive negative time periods, ti , ti+1 , . . . , ti+k where
k ≥ 1, with ŝi ≤ 0, ŝi+1 ≤ 0, . . . , ŝi+k < 0. If an optimal trading interval only intersects part of
these time periods, this intersection can be removed without decreasing the Sterling ratio. Thus,
by Lemma 3.4, any sequence of time periods with all returns non-negative (non-positive) can be
condensed into a single time period, t′i = ti + · · · + ti+k , with ŝ′i = ŝi + · · · + ŝi+k . Further, this
operation can be performed in linear time on the entire excess return sequence, so from now on
we assume without loss of generality that the excess return sequence consists of alternating time
periods of strictly positive and negative excess returns. If ŝi < 0, then ti cannot be the first 1 of
a trade, since by entering one time period later, we exclude only this negative return and do not
increase the M DD. Similarily, it cannot be the last 1 of a trade.
Lemma 3.5 The first 1 and the last 1 of the optimal trade interval must occus at time periods tf
and tl for which ŝf > 0 and ŝl > 0.
The pictorial illustration of this lemma is given below where we show the cumulative return curve.
The time instants ai are the possible entry points, and the time instants bi are the possible exit
points.
b6
Cumulative Return
b4
b3
b1
b5
b2
a4
a2
a1
a5
a6
a3
Time
Let the alternating sequence of entry and exit points be {a1 , b1 , a2 , b2 , . . . , ak , bk } (ai are the entry
points, and bi are the exit points). Note that after the preprocessing into alternating intervals,
k ≤ ⌈ n/2 ⌉. Notice that without loss of generality, we can throw away the first interval if it has a
negative return, as it can never be an entry point, and the last interval if it has a negative return,
for a similar reason. The optimal trade interval will be of the form (at , bt+r ), r ≥ 0.
Our algorithm for finding the Sterling-optimal interval will be to consider every possible starting
point ai , and find the Sterling-optimal interval with this point as starting point (i.e. we have to
15
find the end point of this interval). As the algorithm proceeds, we keep track of the best entry
point (and its corresponding exit point). The entry points at are processed from right to left.
After processing a new entry point at , we will modify the alternating sequence to facilitate faster
processing of the remaining points. More specifically, we will delete the processed entry point and
add a weight to the edge between bt−1 and bt to preserve all the necessary information – we cannot
simply delete the entry point at , since we have to keep track of maximum MDD that occurs in our
activity interval. Since between bt−1 and at we have a drawdown of bt−1 − at , we need to keep this
information in an edge weight connecting bt−1 to bt . At any stage of the algorithm, the edge weight
connecting bt−1 to bt will be equal to the MDD of the interval [bt−1 , bt ], and this MDD is realized on
a prefix of [bt−1 , bt ], i.e. M DD([bt−1 , bt ]) = C(bt−1 ) − C(x), for some x ∈ [bt−1 , bt ]. This property
is an invariant of the algorithm, which we will denote by (∗), and it will be maintained throughout
the algorithm. We will explicitly show this weight attached to bt in parentheses as (wt )bt , where
the value of wt appearing in parentheses indicates the weight.
We start our backward scan at the last entry point, ak , for which there is only one possible
interval (ak , bk ). We set the weight
wk = bk−1 − ak ,
which is the drawdown of the interval from bk−1 to bk . We also store the current best interval
(ak , bk , Strlk ),
where Strlk =
bk −ak −fsp
;
f
fsp = fˆS + fˆB ; and, f = max{fˆS , fˆB }. We now delete the possible entry
point ak from the sequence to give the processed sequence {a1 , b1 , ..., ak−1 , bk−1 , (wk )bk }. Note that
(ak , bk , Strlk ) is a one-step trade, but we keep it here for simplicity. We now proceed to processing
ak−1 and so the process will continue until all aj have been processed.
In the general case, suppose we have processed (backwards) all the entry points up to (including)
the entry point at+1 , and are currently processing entry point at . The weighted exit sequence is
{a1 , b1 , ..., at , bt , (wt+1 )bt+1 , . . . , (wt+m )bt+m },
where bt , . . . , bt+m are the possible exit points – note that t + m may not equal k due to possible
deletion of points which we discuss below. Assume that {bt+1 < . . . < bt+m }: this is true after
we have processed the first start point (since the sequence consists only of one point), and we
will maintain this condition by induction. If bt < bt+1 , then the entire sequence of exit points is
monotonically increasing. On the other hand, suppose that bt ≥ bt+1 . Then,
return(at , bt ) ≥ return(at , bt+1 ), and M DD(at , bt ) ≤ M DD(at , bt+1 ),
16
and so bt+1 need not be considered as an exit point for any optimal interval with entry point at or
earlier, because by stopping earlier at bt , we do not decrease the cumulative return, nor increase the
M DD. Hence, we can delete the possible exit point bt+1 . However, we must now update the weight
in (wt+2 )bt+2 to store the new drawdown between bt and bt+2 , wt+2 ← max{wt+1 , wt+2 + bt − bt+1 }.
Lemma 3.6 If bt ≥ bt+1 , the weighted exit sequence is updated as follows:
at , bt , (wt+1 )bt+1 , . . . , (wt+m )bt+m → at , bt , (max{wt+1 , wt+2 + bt − bt+1 })bt+2 , . . . , (wt+m )bt+m
The correctness of the weight updating rule above follows from the invariant (*) which is preserved
by the transformation.
This process is continued until the next exit point after bt is either above bt or there are no
remaining exit points after bt . In either event, the new sequence of exit points available for at is
strictly monotonically increasing (by induction). Observe that any deletion of a possible exit point
is a constant time operation. Further, since each deletion drops a point from the set {b1 , . . . , bk },
there can be at most k − 1 such deletions during the course of the entire algorithm. We thus have
the following lemma.
Lemma 3.7 When at is processed by the algorithm, the exit points bt < bt+1 < · · · are monotonically increasing. The cost of maintaining this condition for the entire algorithm is O(k) operations.
When processing at , the weighted exit sequence is {a1 , b1 , ..., at , bt , (wt+1 )bt+1 , . . . , (wt+m )bt+m }.
Suppose that wt+2 < wt+3 < . . . < wt+m . Initially this sequence is the empty sequence and so this
condition holds, and once again, by induction we will ensure that this condition will always hold.
Suppose that wt+1 ≥ wt+2 . By construction, bt+2 > bt+1 (Lemma 3.4), and so
return(at , bt+1 ) < return(at , bt+2 ), and M DD(at , bt+1 ) = M DD(at , bt+2 ).
Thus, no optimal interval can have entry point at (or earlier), and exit at bt+1 , because by exiting
at bt+2 , the M DD is not increased, however the total return is increased. Thus if wt+1 ≥ wt+2 , we
can remove the possible exit point bt+1 from the weighted exit sequence and update wt+2 ← wt+1 .
Note that this transformation also preserves the invariant (*).
Lemma 3.8 If wt+1 ≥ wt+2 , the weighted exit sequence is updated as follows:
at , bt , (wt+1 )bt+1 , . . . , (wt+m )bt+m → at , bt , (wt+1 )bt+2 , . . . , (wt+m )bt+m
We continue removing exit points in this way until either there is only one weight left in the weighted
exit sequence, or all the weights are strictly monotonically increasing (by induction).
17
Suppose that wt+1 ≤ f . In this case, we observe that bt cannot be the exit of an optimal interval
with entry at−r , where r ≥ 0. To see this note that if bt − at−r − fˆS < 0, then the return of this
interval is negative and this interval cannot be an optimal interval. If bt − at−r − fˆS ≥ 0 then since
the interval already has M DD of at least f , so by continuing to bt+1 , we do not increase the M DD
but strictly increase the return, hence it cannot be optimal to exit at bt .
Lemma 3.9 Let
at , bt , (wt+1 )bt+1 , . . . , (wt+m )bt+m
be the weighted exit sequence and let T ∈ [t, . . . , t + m] be such an index that wT < f and wT1 ≥ f .
Then no Sterling-optimal interval that starts at at or earlier can exit at {bt , . . . , bT −1 }.
Lemma 3.10 When at is processed by the algorithm, the weights wt+1 are monotonically increasing. The cost of mainitaining this condition for the entire algorithm is O(k) operations.
We thus assume from now on that when processing entry point at , the weighted exit sequence
{a1 , b1 , ..., at , bt , (wt+1 )bt+1 , . . . , (wt+m )bt+m } with m ≥ 0 satisfies the conditions of Lemmas 3.7 and
3.10. The first available exit gives the trade interval (at , bt ). If bt − at − fsp ≤ 0, i.e., if the return
is not positive, then this cannot possibly be an optimal interval. Otherwise, the Sterling Ratio is
Strlt =
bt − at − fsp
,
f
where fsp = fˆS + fˆB and f = max{fˆS , fˆB }. Now consider the exit points bT , where T > t and
suppose that bT − wT +1 < at . In this case, for r > 0,
return(at, bT +r ) < return(aT +1 , bT +r ),
and so no trade with entry at at , exiting at bT +r can possibly be optimal, since we could always
enter later at bT − wT +1 , exit at bT +r , and strictly increase the return without increasing the
drawdown. We are thus done with processing the entry point at , and we can proceed to at−1 after
updating weight wt and comparing
bT −at −fsp
f
with the current champion. Similarly, if bt̄ −wt̄+1 < at
for some t̄ ∈ [t, . . . , T − 1], we are done with the starting point at , and we can proceed to at−1
after updating weight wt . We assume that at any stage of the algorithm we keep the value of
mint̄∈[t,...,T −1] bt̄ − wt̄+1 and thus this check can be done in constant time for any given point at .
Thus, without loss of generality, we can assume that bT − wT +1 ≥ at and bt̄ − wt̄+1 ≥ at for all
t̄ ∈ [t, . . . , T − 1]. Since wT +1 ≥ f , we conclude that bT − at ≥ f . A trade, entering at at and
exiting at bT +r , r > 0 has total return bT +r − at − fsp. The next lemma gives the M DD.
18
Lemma 3.11 Assume that bT − wT +1 ≥ at and bt̄ − wt̄+1 ≥ at for all t̄ ∈ [t, . . . , T − 1]. The trade
(at , bT +r ), r > 0, has M DD = wT +r .
Proof:
The local maxima of the cumulative return sequence for this trade are {0, bt − at −
ˆ
fS , , . . . , bT − at − fˆS , . . . , bT +r − at − fˆS }. Since bt̄ − wt̄+1 − at ≥ 0 ∀t̄ ∈ [t, . . . , T − 1], M DD of the
interval [at , bT ] is equal to fˆS .
Since bT − at − fˆS ≥ 0 and since the sequence of exit points is strictly increasing,
M DD([at , bT +r ]) = max(M DD([at , bT ]), M DD([bT , bT +r ])) = max(fˆS , M DD([bT , bT +r )), fˆB )
where fˆB is the draw down after the last point bT +r .
Since the drawdown at any time in a trade is given by the the diffenence between the previous
maximum and the current cumulative return, M DD([bT , bT +r )) is at most maxi∈[1,r] wt+i . Since
the weights are monotonically increasing, we see that this drawdown ≤ wt+r , which is achieved in
the interval (bt+r−1 , bt+r ). Since wt+r ≥ f = max(fˆS , fˆB ) ∀r > 0, we conclude M DD([at , bT +r ]) =
wt+r .
Summary: For entry point at , the sequence of exit points bT +r , r ∈ [0, m] have cumulative
returns cr = bT +r − at − fsp and M DD’s dr = wT +r for r > 0 and d0 = f . The task is to maximize
cr /dr with respect to r. The sequences {cr } and {dr } are both strictly increasing. We now describe
a general algorithm for performing such a maximization.
3.3.1
Maximizing
cr
dr
Fix t and consider the set of trades (at , bt+r ), r ≥ 0. Corresponding to each trade is the pair
(dr , cr ). As discussed in the previous section, by construction, the sequences {dr } and {cr } are
strictly increasing. On the two dimensional plane, consider the set of points Pm = {(dr , cr )}m
r=0
defined by this set of pairs. Let p = (0, 0). Then the slope of the line joining p to (dr , cr ) is exactly
the Sterling ratio of the trade (at , bt+r ). Thus, finding the optimal trade is equivalent to finding the
upper-touching point from p to the convex hull of the set of points in Pm (see illustration below).
We call p the source point.
cr
p
dr
19
We get the same result if we define Pm = {(dr , bt+r )}m
r=0 , p = (0, at + fsp ). It is known that, given
the convex hull, this touching line can found in time O(log A) where A is number of the points
on the convex hull, [5]. It is easy to see that A ≤ m + 1 ≤ ⌈ n/2 ⌉. In order to compute this
touching point in O(log A) time, one must be able to efficiently search through the convex hull. we
now describe the mechanism for facilitating this search. The details are not important. What is
important is that this data structure can be constructed and efficiently maintained as we update
the set {dr , cr }. We will need to update this set as we process the points at backwards.
We accomplish the fast construction of the upper tangent point of the convex hull by maintaining
the convex hull as a doubly linked list, where each point in the convex hull maintains O(log A)
pointers to other points in the convex hull. More specifically, for point i we consider the set of
points {(dr , cr )}m
r=i (point i, (di , ci ) is the first point in this set). Corresponding to this set is some
convex hull Ci . Point i is also the first point in this convex hull (by the strict monotonicity of the
dr , cr ). Point i maintains a forward pointer to points at position 2j for j ≥ 1 in the convex hull
Ci of the points {(dr , cr )}m
r=i . Clearly point i maintains O(log m) forward pointers. Each point
also maintains backward pointers to any point that points forward to it. At point j, the backward
pointers specific to the convex hull starting at point i < j are maintained separatly for each i so
that constant time access to this set of pointers is possible. An array of an array of pointers suffices
to accomplish this. It is clear that the worst case memory requirement is O(m log m) pointers.
We now discuss the main operations we would like to be able to do on our set of points {(dr , cr )}
and the point p and still be able to update our convex hull data structure efficiently, and hence
efficiently compute the upper tangent line. First we recall that the dr are monotonically increasing.
Assume that each point (dr , cr ) stores all the necessary forward and backward pointers. Note that
the point (dr , cr ) stores a forward pointer to the point nxt(r), which is the next point in the convex
hull if all the points d0 , . . . , dr−1 were removed – remember that in this case, dr becomes leftmost,
and so must be in the convex hull. We will see that these assumptions are maintained inductively.
We state the next observation as a lemma.
Lemma 3.12 The initial convex hull with all the points is given by (d0 , c0 ) followed by the points
pointed to by nxt(0), nxt(nxt(0)), . . ..
As we move from at to process at−1 , the set of points {dr , cr } will change. However, they do not
change arbitrarily. As we will show, the changes can be reduced to combinations of the following
four operations on the set of points Pm .
(i) Translate p by some given vector v.
(ii) Translate all the points in {(dr , cr )} by some given vector v.
(iii) Remove the leftmost point (d0 , c0 ).
20
(iv) Add a new leftmost point (d−1 , c−1 ).
If we can efficiently update our data structure after any of these four operations on the points in
Pm , then we will be able to compute the upper tangent point efficiently for the new set. This is the
content of the next lemma. Since it is a rather technical lemma, we give the proof in the appendix.
Lemma 3.13 Assuming that p, {(dr , cr , nxtr )}m
r=1 are given, all the operations in (1)-(4) above
can be accomplished in time O(log m). Further, in the event that a point is added or deleted, all
necessary pointers are maintained.
We will now show that these four operations are all that is needed for maximizing the Sterling ratio.
Suppose that point at has been processed in the algorithm – i.e., the upper tangent point (optimal
trade with entry at at ) from pt = (0, at + fsp) to Pm = {(dr , bt+r )}m
r=0 has been computed. Now
consider the addition at−1 to construct the new weighted exit sequence. First, delete the leftmost
point (d0 , bT ) from the convex hull (operation (iii)). Lets consider all posible operations that may
take place when processinig at−1 . There are several possibilities, corresponding to the cases which
occur in the algorithm to ensure that the sequence {dr , cr } is strictly increasing.
1. bt−1 ≥ bt . We remove (leftmost) points bt+i (operation (iii)), i ≥ 0, until bt−1 < bt+i+1 , and the
′
new weight wt+i+1
may have increased (Lemma 3.6). Deletion of points bt+i from the weighted
exit sequence doesn’t change the convex hull until bt+i is a possible exit point. After this point,
deleting one point from the weighted exit sequence also deletes the corresponding leftmost point
of the convex hull. At the very end of the sequence of deletions, we have to update the MDD
′
of point bt+i+1 from wt+i+1 to wt+i+1
, this can be done by deletion of the (leftmost) point
′
(wt+i+1 , bt+i+1 ) and addition of new (leftmost) point (wt+i+1
, bt+i+1 ). The total number of such
removals during the entire algorithm is at most n − 1. When condition bt−1 < bt is satisfied,
proceed to the next stage.
2. bt−1 < bt .
(i) wt+1 ≤ wt . We remove (leftmost) points bt+i , i ≥ 0, until wt < wt+i+1 . Deletion of points
bt+i from the weighted exit sequence doesn’t change the convex hull until bt+i is a possible
exit point. After this point, deleting one point from the weighted exit sequence also deletes
the corresponding leftmost point of the convex hull. By Lemma 3.10, the total number of
such removals cannot exceed n − 1 over the course of the entire algorithm. When condition
wt < wt+1 is satisfied, proceed to the next stage.
(ii) bt−1 < bt and wt < wt+1 .
(a) f > wt . Add to the convex hull the point (f, bT ).
21
(b) f < wt . Add to the convex hull points (wt , bt ) and (f, bt−1 ).
The new source point is pt−1 = (0, at−1 + fsp ), which just corresponds to a shift of pt , and so
once the new convex hull for the new weighted exit sequence is computed, an additional O(log n)
operations are needed to find the new upper tangent point.
The total number of removals in the entire algorithm is O(n). For each new entry point, we
have at most a constant number of additions, and since the number of entry points is O(n), we see
that the total number of additions is O(n). By Lemma 3.13, we have that each operation takes
O(log n) in the worst case, thus the total run time is O(n log n). Collecting all the results together,
we can find the Sterling-optimal strategies making zero, one or more than one trade in O(n log n)
time, completing the proof of the first part of Theorem 1.3.
Note that by only maintaining exit points with weight at most some given constant, M DD0 , i.e.,
by truncating some region of the points to the right, this algorithm is easily extended to computing
the Sterling-optimal strategy that uses exactly one trade and has an M DD ≤ M DD0 .
Proposition 3.14 Given M DD0 , a Sterling-optimal strategy that uses exactly one trade and has
M DD ≤ M DD0 can be computed in O(n log n) time.
This result will be useful when we consider constrained Sterling-optimal strategies.
3.4
Constrained Sterling-Optimal Strategies
As with the return-optimal strategies, the unconstrained sterling-optimal strategies may make too
many trades. Here, we consider the the Sterling-optimal strategy that makes at most K trades,
K . We refer to such strategies as K-Sterling-optimal. First, we present some properties of this
TStrl
strategy, before giving an efficient algorithm to compute it.
A maximal return-optimal strategy Tµ∗ is a return-optimal strategy whose trade intervals cannot
be enlarged. Given any return-optimal strategy Tµ , in one (linear time) scan from left to right,
we can enlarge any trade intervals maximally to the right as long as they keep the same return.
Similarily, in one backward scan, we can extend all trade intervals maximally to the left. Since Tµ
can be computed in linear time, we conclude that
Lemma 3.15 A maximal return-optimal strategy Tµ∗ can be computed in linear time.
If any trade interval of a maximal return-optimal strategy is extended in either direction, then the
total return must strictly decrease. In the previous section, we gave an O(n log n) algorithm for
computing the Sterling-optimal strategy with exactly 1 trade. We also saw that if the unconstrained
Sterling-optimal strategy contains more than 1 trade, then it is Tµ∗ . Fix K, and let the number of
K = T ∗ , and we are done. Thus we only need
trades that Tµ∗ makes be K0 ≤ K. In this case TStrl
µ
22
to consider the case that 1 < K < K0 . Some important properties of Tµ∗ are summarized below.
Pi
0
When it is clear, We also use Tµ∗ to refer to the set of trading intervals {Ir }K
r=1 . Let Ci =
j=1 ŝj
denote the cumulative return sequence of the excess returns.
Lemma 3.16 Let Tµ∗ be maximal return-optimal. Let I be an interval [ti , tj ].
i. If I ∈ Tµ∗ , then,
Pj
k=i ŝk
− fsp ≥ 0 and M DD(I) ≤ fsp .
P
ii. Suppose I does not intersect with any interval in Tµ∗ and let the return of I ( jk=i ŝk ) be µ(I).
Then, µ(I) ≤ fsp. If I is adjacent to some interval of Tµ∗ , then µ(I) < 0. If I is adjacent to
two intervals in Tµ∗ , then µ(I) < −fsp.
iii. Let [tl , tr ] and [tl′ , tr′ ] be two consecutive trade intervals in Tµ∗ , l ≤ r < l′ ≤ r ′ . Then,
Cr − Cl′ > fsp, and for all r < q < l′ , Cl′ < Cq < Cr .
Let {(ai , bi )} denote the local minima and maxima of {Ci }, as in the previous section. Any trade
K must enter (exit) at a local minimum (maximum). Further, the entry (exit) point
of Tµ∗ or TStrl
must be a minimum (maximum) in the trade, otherwise we can shrink the trade, strictly increasing
the return without increasing the M DD.
K . Then C is a local minimum, C is
Lemma 3.17 Let I = [tl , tr ] be a trade interval of Tµ∗ or TStrl
r
l
a local maximum, and for any k, with l ≤ k ≤ r, Cl ≤ Ck ≤ Cr
K . Essentially, it states that the
We now give an important inclusion property of the strategy TStrl
Sterling optimal strategy can be constructed by merging some trades of the return optimal strategy.
We give the technical proof in the appendix (which will be the case with most of the technical results
in this section).
Proposition 3.18 Let Tµ∗ be a maximal return-optimal trading strategy. There exists a K-SterlingK , K > 1, with the following property: if I = [t , t ] is any trading interval in
optimal strategy TStrl
l r
K , then a prefix of I and a suffix of I are trades in the maximal return-optimal strategy T ∗ .
TStrl
µ
As a result of Proposition 3.18, we assume from now on that every interval of the sterling optimal
K is prefixed and suffixed by (not necessarily distinct) intervals from a maximal returnstrategy TStrl
optimal strategy that makes K0 trades.
K can be chosen to make exactly K trades.
Lemma 3.19 If 1 < K ≤ K0 then TStrl
K can be constructed: start with all the intervals of a maxiLemmas 3.18 and 3.19 indicate how TStrl
mal return-optimal strategy Tµ∗ and then merge some neighbouring intervals, keeping the merging
sequence that gave the best strategy. The number of possible merging sequences is exponential,
however, we will now show that an efficient greedy merging algorithm gives the correct result.
23
Given two consecutive non-adjacent intervals I1 = [tl1 , tr1 ], I2 = [tl2 , tr2 ], where I1 preceeds
I2 , define the bridge B(I1 , I2 ) = [tr1 , tl2 ] to be interval connecting I1 with I2 . If I1 and I2 are
intervals in a maximal return optimal strategy, then by Lemma 3.16, the M DD of the bridge
is Cr1 − Cl2 . Since Cr1 is a maximum over the interval [tl1 , tl2 ], and Cl2 is a minimum over the
interval [tr1 , tr2 ], we have that the M DD of the union of these three intervals, [tl1 , tr2 ] is given by
max{M DD(I1 ), Cr1 − Cl2 , M DD(I2 )}.
For every bridge B(I1 , I2 ), define the closure Cl(B(I1 , I2 )) to be the smallest interval J = [tl , tr ],
in the return sequence, satisfying the following three properties.
Cl1 . Cl ≤ Cm ≤ Cr for l ≤ m ≤ r, i.e., Cl is a minimum and Cr is a maximum in [tl , tr ].
Cl2 . I1 , I2 ⊂ J, i.e., J contains both I1 and I2 .
Cl3 . J is prefixed and suffixed by intervals from a maximal return-optimal strategy Tµ∗ .
Note that a bridge may not have closure. For example, if the two last intervals Il−1 , Il in Tµ∗ are
such that such that the end point Il is below the end point of Il−1 , then B(Il−1 , Il ) doesn’t have a
closure. The next lemma shows that if the closure J for a bridge exists, then not only is it unique,
but any other interval satisfying Cl1 - Cl3 contains J.
Lemma 3.20 For two intervals I1 , I2 , if Cl(B(I1 , I2 )) exists, then it is unique. Moreover, for any
other interval I satisfying Cl1 - Cl3 , Cl(B(I1 , I2 )) ⊆ I.
Proof:
Let J1 = [tl1 , tr1 ] and J2 = [tl2 , tr2 ] satisfy Cl1 - Cl3 . Without loss of generality, assume
that tl1 ≤ tl2 < tr1 ≤ tr2 . By construction, J1 ∩J2 = [tl2 , tr1 ] satisfies Cl1 - Cl3 . Now let Cl(B(I1 , I2 ))
be the intersection of all intervals that satisfy Cl 1 - Cl3 , concluding the proof.
Suppose that bridge B and B ′ are bridges in Tµ∗ and that Cl(B) contains B ′ . Then Cl(B) satisfies
Cl1 - Cl 3 with respect to B ′ and hence Cl(B) also contains Cl(B ′ ).
Lemma 3.21 Let B and B ′ be bridges in Tµ∗ . If B ′ ⊂ Cl(B), then Cl(B ′ ) ⊂ Cl(B).
K containing bridge B satisfies properties Cl - Cl (Lemma 3.17 & Proposition
Any interval in TStrl
1
3
3.18), immediately yielding the following proposition.
K and let B be a bridge in T ∗ .
Proposition 3.22 Let I ∈ TStrl
µ
i. If B ⊂ I, then Cl(B) ⊂ I.
ii. If B does not have a closure, then no K-Sterling-optimal strategy can contain B.
iii. A K-Sterling-optimal strategy with more than one trading interval and no bridges of Tµ∗ has
M DD = fsp . If it contains one or more bridges Bi of Tµ∗ , then M DD = maxi M DD(Cl(Bi )).
24
iv. The M DD of a K-Sterling-optimal strategy with more than one trading interval can be one of
at most T + 1 possible values where T is the number of bridges between the intervals of Tµ∗ .
Proof:
I ∈
K
TStrl
(i) and (ii) are immediate. (iv) follows from (iii), thus we only need to prove (iii). Let
contain the consecutive bridges B1 , . . . , BK , and hence their closures. From (i), it is
clear that M DD(I) ≥ maxi M DD(Cl(Bi )). It is also clear that I = ∪K
i Cl(Bi ). We show, by
strong induction on K, a more general statement than we need: suppose that I = ∪K
i Cl(Bi ), then
M DD(I) ≤ maxi M DD(Cl(Bi )). If K = 1 then I = Cl(B1 ) and the result is trivial; suppose
it is true for up to K − 1 consecutive bridges, K > 1, and suppose that I is the union of K
closures of consecutive bridges. Consider the first closure Cl(B1 ). Let I = [tl , tr ] and Cl(B1 ) =
[tl , tr′ ], tr′ ≤ tr . By definition of Cl(B1 ), Cr′ is a maximum over [tl , tr′ ]. Thus, M DD(I) =
max{M DD(Cl(B1 )), M DD([tr′ , tr ])}. If r = r ′ , then I = Cl(B1 ) and we are done. If r < r ′ ,
′
then tr′ +1 is the begining of some bridge Bκ . Let I ′ = ∪K
i=κ Cl(Bi ). Then, [tr ′ , tr ] ⊆ I and so
M DD([tr′ , tr ]) ≤ M DD(I ′ ). But I ′ is the union of at most K − 1 closures, so by the induction
hypothesis, M DD(I ′ ) ≤ maxi≥κ M DD(Cl(Bi )), concluding the proof.
We will distinguish between four types of bridges. Let I1 = [tl1 , tr1 ], I2 = [tl2 , tr2 ] be consecutive
intervals in Tµ∗ . The bridge B = B(I1 , I2 ) can be one of four types:
regular.
Cl1 ≤ Cl2 and Cr1 ≤ Cr2 , i.e., Cl(B) = [l1 , r2 ].
right irregular.
Cl1 ≤ Cl2 and Cr1 > Cr2 , i.e., Cl(B) contains the next bridge.
left irregular.
Cl1 > Cl2 and Cr1 ≤ Cr2 , i.e., Cl(B) contains the previous bridge.
irregular.
Cr1 > Cr2 and Cl1 > Cl2 , i.e., Cl(B) contains both the next and previous bridges.
We define the weight of the bridge W (B(I1 , I2 )) as follows:
Cr1 − Cl2
Cr − C
l2
1
W (B(I1 , I2 )) =
Cr1 − Cl2
+∞
if B(I1 , I2 ) is regular,
if B(I1 , I2 ) is left irregular and the previous bridge is right irregular.
if B(I1 , I2 ) is left irregular and the previous bridge is irregular.
otherwise.
The general idea behind our algorithm is to start with a maximal return-optimal strategy and
greedily merge pairs of intervals or pair of bridges according to the bridge weight, keeping track
of the best K intervals each time. When no more merging can occur, because we are down to K
intervals or all the bridge weights are ∞, we return the best K intervals we have seen so far. More
precisely, let Tµ∗ = {I1 , . . . , IK0 } be a maximal return-optimal trading strategy making K0 trades.
We denote this pool of trade intervals by P0 , the base pool. From pool Pi , we obtain pool Pi+1 by a
single merge according to the following rule. Let B = B(I1 , I2 ) be the bridge with smallest weight.
If B = ∞, stop (pool Pi+1 does not exist). Otherwise, there are two cases.
25
i. Regular merge: if B is regular, merge B with I1 and I2 to get a larger interval Inew = [tl1 , tr2 ].
We now update the status (type and weight) of any neighboring bridges as follows:
• Previous bridge changes from right-irregular to regular (update type and weight).
• Previous bridge B ′ changes irregular to left-irregular (update type). If the bridge previous
to B ′ is right-irregular or irregular then update weight.
• Next bridge changes from irregular to right-irregular (update type).
• Next bridge changes from left-irregular to to regular (update type and weight).
ii. Irregular merge: if B is left irregular, denoting the previous bridge B ∗ , merge the two bridges
and the interval between them into one bigger bridge Bnew = B ∗ ∪ I1 ∪ B. The status of bridges
other than B have not changed. The status and weight of B may need updating.
Intervals are formed only by regular merges, so it is easy to see that intervals resulting from this
merging procedure begin at a minimum and end at a maximum. New bridges are formed by
irregular merges, and the resulting bridge must begin at a maximum and end at a minimum. The
only bridge weights that could change are those that had weights of ∞. In such an event the weight
will drop, but not below the weight of the original bridge used in the merge that led to the update.
Lemma 3.23 Let bridge B with weight w be involved in a merge, and suppose that the weight of
bridge B ′ is updated in the process from ∞ to u. Then, w < u.
Proof: There are two cases:
i. The bridge involved in the merge was regular, i.e., two consecutive intervals I1 = [tl1 , tr1 ]
and I2 = [tl2 , tr2 ] are merged with their bridge B12 = B(I1 , I2 ) with W (B12 ) = w. Let the
preceeding interval be I0 = [tl0 , tr0 ] and the following interval be I3 = [tl3 , tr3 ] , and let the
preceeding and following bridges be B01 and B23 respectively. If B01 obtained finite weight, it
cannot be left irregular, as it would remain left irregular after the merge, and hence its weight
would still be ∞. Thus, we need only consider B01 right irregular or irregular (i.e., Cr0 > Cr1 ).
Its weight becomes u = Cr0 − Cl1 > Cr1 − Cl1 . Since B12 is regular, w = Cr1 − Cl2 < Cr1 − Cl1
and so w < u. If B23 obtained finite weight, then it could not be right regular or irregular as
it could not become regular or left irregular after the merge. Thus, we need only consider B23
left irregular (Cl2 > Cl3 ). Its weight becomes u = Cr2 − Cl3 > Cr2 − Cl2 . Since B12 is regular,
Cr1 ≤ Cr2 , and so u > Cr1 − Cl2 = w.
ii. The bridge involved in the merge was left-irregular, i.e., B12 = [tr1 , tl2 ] is left irregular, and
B01 = [tr0 , tl1 ] is either right-irregular or irregular (in both cases, Cr0 > Cr1 ). Let w = Cr1 −Cl2
be the weight of B12 . The merged bridge is B = B01 I1 B12 . If B has finite weight (i.e. it is
26
either regular or left-irregular), then its new weight is u = Cr0 − Cl2 > Cr1 − Cl2 = w. If
B is left-irregular or irregular, then it does not change the weight of any other bridge. If,
on the other hand, B became right-irregular or irregular, then it could affect the weight of
the following bridge B23 , if B23 was left-irregular (Cl2 > Cl3 ). In this case, the weight of
B23 = [tr2 , tl3 ] becomes v = Cr2 − Cl3 > Cr2 − Cl2 . But since B12 was left-irregular, Cr2 ≥ Cr1 ,
and so v > Cr1 − Cl2 = w.
The next lemma shows that if all the bridge weights become ∞, any further merged pairs of intervals
can never be part of a K-Sterling-optimal strategy.
Lemma 3.24 If all the bridge weights in a pool of intervals are ∞, then any further merged pairs
of intervals from this pool can never be part of a K-Sterling-optimal strategy.
Proof: (Lemma 3.24). Let Pr be pool obtained from P0 by some sequence of merging intervals with
bridges of finite weight, and suppose that all the bridges in Pr have infinite weight. In particular,
this means that none of the bridges in Pr are regular. Denote the bridges by B1 , . . . , Bm , and
consider bridge Bk . If Bk is right irregular or irregular, then all following bridges are either right
irregular or irregular since all bridges have finite weight. If a trading interval contains Bk , it must
contain Bk+1 (since Bk is right irregular or irregular), and so by induction, it must contain all the
following bridges (and their closures). But, the last bridge does not have a closure (as it is right
irregular or irregular), a contradiction. If on the other hand, Bk is left irregular, then all preceeding
bridges are left irregular as all bridges have infinite weight. If a trading interval contains Bk , it must
contain Bk−1 (since Bk is left irregular), and so by induction, it must contain all the preceeding
bridges (and their closures). But, the first bridge does not have a closure (as it is left irregular), a
contradiction. We conclude that Bk cannot be in any trading interval.
Each merge decreases the number of intervals and number of bridges by one. If we merge down
K can be chosen to be
to pool PK0 −K , we are left with exactly K intervals. We will show that TStrl
the best K trades (with respect to total return) in one of these pools. Specifically, define TjK to
be the K intervals in Pj with the highest total return. We say that a strategy is coarser than pool
Pi if the strategy can be obtained by a sequence of merges of some (or all) of the intervals in Pi .
Clearly, ∀i, Pi+1 is coarser than Pi , as Pi+1 is obtained from Pi after a single merge. Note that for
a strategy to be coarser than Pi , it need not contain every trade in Pi , however if it contains part
of any trade in Pi , then it contains the entire trade. Next, we show that after a merge, the M DD
of the remaining intervals is equal to the weight of the bridge involved in the merging.
27
Lemma 3.25 If pool Pi , i ≥ 1, was obtained from Pi−1 by a merge involving a bridge of weight w,
then the M DD of any interval in Pi is at most w. If the merge created a new interval (i.e., the
bridge was regular), then the M DD of the new interval is equal to w.
Proof:
In pool P0 , since any bridge is adjacent to two intervals of Tµ∗ , its weight is at least fsp
(Lemma 3.16). Consider sequence of pools P0 , P1 , . . . , Pr , where bridge Bi with weight W (Bi ) was
the minimum weight bridge involved in the merge that resulted in pool Pi from from Pi−1 . By
Lemma 3.23 bridges weights are non-decreasing, i.e., W (Bi ) ≤ W (Bi+1 ).
We now use induction on the index i. For i = 1, from Lemma 3.16, every interval in P0 has
M DD at most fsp . If P1 was obtained from P0 by an irregular merge, then all intervals of P1 are
intervals of P0 , with M DD at most fsp . Since W (B1 ) ≥ fsp , the claim holds. If the merge was
regular, then the M DD is W (B1 ) ≥ fsp and the M DD of all other intervals is at most fsp . Thus,
the claim holds for P1 .
Suppose the claim holds for all j < i and consider pool Pi which was obtained from Pi−1 using
a merge involving Bi . By the induction hypethesis, the M DD of any interval from Pi−1 is at
most W (Bi−1 ) ≤ W (Bi ). If Pi that was obtained by an irregular merge, every interval of Pi is an
interval of Pi−1 and thus has M DD at most W (Bi−1 ) ≤ W (Bi ). Suppose that Pi was obtained by
a regular merge – all intervals except the merged interval are intervals of Pi−1 . Consider the M DD
of the new interval, which is obtained by the regular merge I1 ∪ Bi ∪ I2 . Since new intervals are
created only through regular merges, it is easy to see by induction that property Cl1 holds for all
the intervals in Pi−1 , in particular it holds for I1 and I2 . Since Bi was regular, the M DD of the new
interval is max(M DD(I1 ), W (Bi ), M DD(I2 )). By the induction hypothesis, M DD(I1 ) ≤ W (Bi−1 )
and M DD(I2 ) ≤ W (Bi−1 ), thus, max(M DD(I1 ), W (Bi ), M DD(I2 )) = W (Bi ).
First, we show that if a K-Sterling-optimal strategy makes K trades, all of which are contained in
intervals of one of the pools Pi , then a K-Sterling-optimal strategy exists which is composed of the
K intervals with highest return in some pool Pj with j ≤ i.
Lemma 3.26 If K subintervals of the intervals of pool Pi are a K-Sterling-optimal strategy, then
for some j ≤ i, the K intervals with highest return of pool Pj are a K-Sterling-optimal strategy.
Proof: If Pi = P0 , then the claim is trivial. Suppose that i > 0, and let T = {I1 , . . . , IK } be the
K-Sterling-optimal strategy whose trades are all subintervals of intervals in Pi . Consider the set
B of all bridges in Tµ∗ that are contained in T , B = {Bi }ri=1 . We can assume that B is not empty
because if it were, then T is composed of intervals in Tµ∗ , in which case the top K intervals (with
respect to return) in Tµ∗ are clearly optimal. Since Pi contains all the intervals in T , Pi contains
all the bridges in B. Thus, there must exist j ≤ i such that Pj contains all the bridges in B and
no pool Pk , with k < j has this property, i.e., Pj was obtained from the previous pool by a regular
28
merge involving a bridge B ∗ which must contain some bridge Bl ∈ B. Let I be the interval in T
that contains Bl . Then, I must contain the whole bridge B ∗ , since if B ∗ is the result of irregular
merges, one of which involved bridge Bl , then B ∗ ⊂ Cl(Bl ), and Cl(Bl ) ⊆ I (Proposition 3.22).
Since B ⊂ I, M DD(T ) ≥ M DD(I) ≥ W (B ∗ ). By Lemma 3.25, since B ∗ was the last bridge
involved in a merge, the M DD of every interval in Pj is at most W (B ∗ ). Since every interval of
T is a subinterval of some interval in Pj , we conclude that T is contained in at most K intervals
of Pj . Let TjK be the top K intervals in Pj . Then, the return of T is at most the return of TjK .
Further, M DD(TjK ) ≤ W (B ∗ ) ≤ M DD(T ), and so Strl(TjK ) ≥ Strl(T ), concluding the proof.
We are now ready to prove the main result, which will lead to the greedy algorithm for constructing
a K-Sterling optimal strategy.
K
K
Theorem 3.27 Let j ∗ be such that Strl(TjK
∗ ) ≥ Strl(Tj ), ∀j. Then Tj ∗ is K-Sterling optimal.
Proof:
Let S0K be a K-Sterling-optimal strategy that makes K trades – by Lemma 3.19, such a
strategy must exist. If S0K has the same Sterling ratio as the trading strategy composed of the K
most profitable trades in P0 , then we are done. If not, then we know from Proposition 3.18 that
S0K is coarser than P0 . We prove the following statement for all k ≥ 1
K
that makes K trades
Q(k): Suppose there exists a K-Sterling-optimal strategy Sk−1
K is composed of K intervals of P , or there
and is coarser than Pk−1 . Then either Sk−1
k
exists a K-Sterling-optimal strategy SkK that makes K trades and is coarser than Pk .
We know that Q(1) is true. Suppose that Q(k) is true for all k ≥ 1, we then prove the proposition
K are composed of K intervals in
as follows. By an easy induction, we have that if none of the Sj−1
K making exactly K trades that is
Pj for all j ≤ m, then there is a K-Sterling-optimal strategy Sm
coarser than Pm . Suppose that we can construct a total of κ + 1 pools, Pi for 0 ≤ i ≤ κ ≤ K0 − K.
If κ < K0 − K then all the bridge weights in Pκ are infinite. If κ = K0 − K, then any further
merging leads to fewer than K intervals. In both cases, there cannot exist a K-Sterling-optimal
strategy that is coarser than Pκ . Therefore, for some j ∗ ≤ κ, the K-Sterling-optimal strategy SjK∗ −1
K that
is composed of K intervals of Pj ∗ . By Lemma 3.26, there is a K-Sterling-optimal strategy TStrl
is composed of the top K intervals of some pool Pl , where l ≤ j ∗ .
K
is coarser than
What remains is to show that Q(k) is true for all k ≥ 1. Suppose that Sk−1
Pk−1 and is not composed of K intervals in Pk . We show that there exists SkK that is coarser than
K
Pk . Since Sk−1
is coarser than Pk−1 , it contains at least one bridge B in Pk−1 with finite weight
(because if it contains an infinite weight bridge, then it either contains the preceeding or following
bridge; this argument continues analogously to the proof of Lemma 3.24 until we include a bridge
K
that contains B, and let Il and Ir be intervals in
of finite weight). Let I be the interval of Sk−1
Pk−1 (which are subintervals of I) connected by B. Let B ∗ be the bridge in Pk−1 with minimum
29
weight that was involved in the merge to get Pk from Pk−1 , and let Il∗ and Ir∗ be the intervals in
K is also coarser than P and we are done, so suppose
Pk−1 connected by B ∗ . If B ∗ = B then Sk−1
k
B ∗ 6= B. There are two possibilities:
K does not contain I ∗ or I ∗ , then S K ∩ (I ∗ ∪ B ∗ ∪ I ∗ ) = ∅ and
(i) B ∗ is a regular bridge. If Sk−1
r
r
l
k−1
l
K
K
thus Sk−1
itself is coarser than Pk , and can be chosen as SkK . Suppose that Sk−1
contains
K
Il∗ and not Ir∗ (similar argument if it contains Ir∗ and not Il∗ ). Thus some interval I ′ ∈ Sk−1
has as a suffix Il∗ . Suppose we construct SkK by replacing interval I ′ by interval I ′ ∪ B ∗ ∪ Ir∗ .
SkK is then coarser than Pk . Since B ∗ is regular, the return of I ′ ∪ B ∗ ∪ Ir∗ is at least
as big as the return of I ′ . Ir∗ is either an interval of P0 or was obtained by merging some
intervals of P0 through bridges with weight at most W (B ∗ ) (Lemma 3.25), and so M DD(Ir∗ ) ≤
W (B ∗ ). Since the maximum cumulative return for I ′ is attained at its right endpoint (Lemma
3.17) and the left endpoint of Ir∗ is a minimum in Ir∗ , we have that M DD(I ′ ∪ B ∗ ∪ Ir∗ ) =
max{M DD(I ′ ), W (B ∗ ), M DD(Ir∗ )} = max{M DD(I ′ ), W (B ∗ )}. Since W (B ∗ ) ≤ W (B), we
K ), and thus Strl(S K ) ≥ Strl(S K ), which means that
conclude that M DD(SkK ) ≤ M DD(Sk−1
k−1
k
K
contains both Il∗ and Ir∗ , and
SkK is also K-Sterling-Optimal. Finally, suppose that Sk−1
K
by removing bridge B and adding bridge B ∗ .
consider the strategy SkK obtained from Sk−1
K ) + W (B) − W (B ∗ ) ≥ µ(S K ). Since W (B) ≥ W (B ∗ ), the M DD cannot
µ(SkK ) = µ(Sk−1
k−1
have increased, and so SkK is K-Sterling-Optimal and coarser than Pk .
(ii) B ∗ is an irregular bridge. Since B ∗ = B(Il∗ , Ir∗ ) has finite weight, we can conclude that B ∗ is
∗ , I ∗ ) is right-irregular or irregular. Since
left-irregular and the previous bridge B− = B(Il−1
l
K does not contain B ∗ , by Lemma 3.17, there are two possibilities: S K does not contain
Sk−1
k−1
Il∗ , in which case it also does not contain bridge B− and so B− and B ∗ can be merged into one
∗
K , i.e., S K is also more coarse than P ; or, S K
bridge without influencing Sk−1
k
k−1 contains Il
k−1
as one of its intervals. In this case, since B ∗ is left-irregular, µ(Il∗ ) < W (B ∗ ) ≤ W (B), and so
K and breaking I into two subintervals by removing B from I results
by dropping Il∗ from Sk−1
in a profit increase of W (B) − µ(Il∗ ) > 0. Further, the M DD cannot increase, so the new
K , which contradicts
strategy makes K trades and has strictly greater Sterling ratio than Sk−1
K . Thus, S K cannot contain I ∗ as one of its intervals.
optimality of Sk−1
k−1
l
Thus Q(k) holds for all k ≥ 1, concluding the proof.
We are now ready to give the O(n log n) algorithm that establishes Theorem 1.3. First, we can
compute the optimal strategy that makes only one trade in O(n log n) (Section 3.3), and compare
this with the trivial strategy that makes no trades. It remains to compute the K-Sterling-optimal
strategy and pick the best. We show how to do this in O(n log n) time.
30
First we obtain Tµ∗ in linear time. Suppose Tµ∗ makes K0 > K trades (as otherwise Tµ∗ is our
solution). By Theorem 3.27, we only need to construct the pools P0 , P1 , . . ., maintaining the pool
with the optimal Sterling ratio for its top K trades, as explained in the following algorithm.
1:
Set i = 0; Sort (in decreasing order) the intervals from P0 according to profit; Sort all the
bridges with finite weight in increasing order. Let Bi be the minimum weight bridge in Pi ; Let
strategy Si consist of the top K intervals, and let Strlopt = Strl(Si );
2:
3:
4:
while Pi contains at least K intervals and at least one finite weight bridge do
if Bi = B(Il , Ir ) is regular then
Regular merge to obtain Pi+1 : remove Il , Ir , Bi from the interval and bridge orderings, and
add back I = Il ∪ Bi ∪ Ir into the interval ordering; compute µ(I) and M DD(I);
5:
6:
7:
Update neighboring bridge weigths and re-insert them back into the bridge ordering.
else if Bi = B(Il , Ir ) is left-regular then
Irregular merge to obtain Pi+1 : Let B− be the bridge left of Bi ; remove Il , B− , Bi from the
interval and bridge orderings. Create the new bridge B = B− ∪ Il ∪ Bi , compute W (B)
and insert B into the bridge ordering (note that W (B) may be ∞).
8:
end if
9:
i ← i + 1; update Strli ; if Strlopt < Strli , then Strlopt ← Strli .
10:
end while
The correctness of the algorithm follows from Theorem 3.27. We now analyse the run time of
an efficient implementation of the algorithm. P0 contains at most n intervals and bridges. Each
execution of the while loop reduces loop number of bridges and intervals by 1 each, so the while
loop is executed at most n/2 times. Merging two intervals is a constant time operation. The profit
of a new interval is the profit of the merged intervals minus the weight of the merging bridge (also
computable in constant time). The M DD of a new interval is the maximum of the M DD of the
merged intervals and the weight of the merging bridge (also computable in constant time). The
weight of a new bridge takes constant time to compute, and updating the weights of the neighbour
bridges is a constant time operation provided that pointers are maintained to them. These pointers
can be updated in constant time as well. Thus the run time within the while loop is dominated by
inserting into the bridge or interval orderings. At most a constant number of such such inserts into
a sorted list need to be done, and each is an O(log n) operation [6]. To efficiently implement step
9, we maintain two sorted lists of the top K intervals in the algorithm, sorted according to return
and M DD. These can be maintained in O(log K) operations. The first allows us to update the
total return of the top K intervals in constant time, and the second allows us to update the M DD
of the top K intervals (by taking the interval with largest M DD) in constant time. Thus the total
running time of the while loop is O(n log n + n log K) = O(n log n) The preprocessing (step 1) is
O(n), and so does not contribute to the asymptotic run time.
31
4
Sharpe Optimal Trading Strategies
Another popular measure of the portfolio’s risk-adjusted return is the Sharp Ratio. For trading
strategy T , we consider two versions of the Sharpe ratio, Shrp1 and Shrp2 .
Shrp1 (T ) =
µ(T )
,
σ(T )
Shrp2 (T ) =
µ(T )
.
σ 2 (T )
(9)
Note that Shrp2 is more conservative in that it penalizes large variances more heavily. We introduce
a simplified Sharpe ratio (SSR) S that will be instrumental to finding the optimal strategies,
S=
µ
.
s2
It is easy to check that maximizing Shrp1 is equivalent to maximizing
r̄
,
1 2
s
−r̄ 2
n
by
µ2
,
s2
and that Shrp2 is given
where r̄ is the mean return. We will relate the maximization of Shrp1 and Shrp2 to the
maximization of S.
Let T be a trading strategy that makes K trades, with trading intervals I1 , . . . , IK . Each trade
contributes a transactions cost of −fsp to the return sequence. In general, a trade contributes
fˆ2 + fˆ2 to s2 . However, we will assume that fsp ≪ 1 and so we will ignore the contribution of the
S
B
transactions cost to s2 . Alternatively, we can justify this by assuming that the transactions cost is
spread finely over many time intervals. The sum over these small time intervals is finite, equal to
−fsp , however, the sum of squares over these small time intervals can be made arbitrarily small.
Define the total return and sum of squared returns for each trading interval,
µi = µ(Ii ) =
X
s2i = s2 (Ii ) =
r[j],
j∈Ii
X
r[j]2 .
j∈Ii
We define Ai as the contribution of trade i to the mean return, and Bi as the contribution of trade
i to the mean squared return (ignoring the effect of the transactions cost), i.e., Ai = n1 (µi − fsp )
P
PK
and Bi = n1 s2i . We define A(T ) = K
k=1 Ai (note that r̄ = A(T )) and B(T ) =
k=1 Bi (note that
1 2
ns
4.1
= B(T )).
Maximizing the Simplified Sharpe Ratio S
We will need the following technical lemma, which can be proved by an easy induction.
Lemma 4.1 Let F = { ab11 , ab22 , . . . , abkk } be any set of fractions satisfying bi > 0 and
c
d
a1 +a2 +...+ak
b1 +b2 +...+bk
c
d
≤
ai
bi
≤
a
b,
≤ ab . The upper (resp. lower) bound is strict if at
least one of the fractions in F is strictly upper (resp. lower) bounded by ab resp dc .
for all i, where b, d > 0. Then,
≤
32
Let T ∗ be an SSR-optimal strategy making K > 1 trades with trading intervals I1 , . . . , IK .
∗
PK
S(T ) = Pi=1
K
Ai
i=1 Bi
Lemma 4.2
Proof:
Ai
Bi
=
A(T ∗ )
,
B(T ∗ )
is a constant for every interval i, i.e., every trade is equivalent.
Suppose that mini
Ai
Bi
<
Aj
Bj
for some j (strict inequality), and without loss of generality,
assume that the minimum is attained for interval I1 . By Lemma 4.1, if we remove I1 , we get that
PK
PK
i=2 Ai
i=1 Ai
< PK
= S(I2 ∪ · · · ∪ IK ),
S(I1 ∪ · · · ∪ IK ) = PK
B
B
i
i
i=1
i=2
which contradicts the optimality of T ∗ implying that mini
Ai
Bi
=
Aj
Bj
for all j.
Corollary 4.3 An SSR-optimal trading strategy making one trade exists.
Proposition 4.4 An SSR-optimal strategy making one trade can be found in O(n log n) time.
Proof:
By Corollary 4.3, we are guaranteed the existence of such a strategy. It suffices to find
P
P
the single interval I maximizing i∈I r[i]/ i∈I r[i]2 . Consider all intervals starting at position i
P
P
and define ck = kj=i r[j] and dk = kj=i r[j]2 . We wish to find k to maximize ck /dk . If we have
done this for position i, we now consider position i − 1. We show that the algorithm in Section
3.3.1 can be used. Trade intervals starting at i − 1 correspond to shifting all the ck by r[i − 1],
and all the dk by r[i − 1]2 . Both these operations simply correspond to shifting the origin point p
to p′ = p − (r[i − 1], r[i − 1]2 ). We then add a new leftmost point at p. Since each update takes
O(log n), and the optimal interval for the new points can be found in O(log n), the entire algorithm
runs in O(n log n).
4.2
Maximizing Shrp2
Ignoring the fsp 2 term in the denominator changes the denominator slightly, so we introduce the
slightly different quantity Shrp2 . Specifically,
Shrp2 (T ) =
2
d
n fsp
A(T )
,
+ B(T ) − A2 (T )
Shrp2 (T ) =
A(T )
,
B(T ) − A2 (T )
where d is the number of trades in T . By the Cauchy-Schwarz inequality, for any trading strategy,
P
P
r[i]2 ≥ n1 ( i r[i])2 . Since we are only interested in strategies for which A(T ) ≥ 0, we have
33
Lemma 4.5 For strategy T , if A(T ) > 0 then B(T ) − A2 (T ) > 0
We will show that maximizing Shrp2 is closely related to a constrained optimization of the SSR,
and that maximizing Shrp2 is not too far from maximizing Shrp2 .
Let Tµ∗ be return optimal, with return µ(Tµ∗ ) = µ∗ . For any 0 ≤ α ≤ µ∗ , we define the
constrained SSR-optimal strategy Tα as the strategy with maximum SSR among all strategies with
return at least α, i.e., A(Tα ) ≥ α and for all strategies T with A(T ≥ α), S(Tα ) ≥ S(T ). Note that
while an SSR-optimal strategy can be chosen with one trading interval, a constrained SSR-optimal
strategy may require more than one trading interval. We show that for some appropriate threshold
α, the constrained SSR-optimal strategy is a Shrp2 -optimal strategy.
Proposition 4.6 ∃0 ≤ α ≤ µ∗ such that the constrained SSR-optimal strategy Tα is Shrp2 -optimal.
Proof:
Let T be any Shrp2 -optimal strategy, and let α∗ = A(T ). Let Tα∗ be any constrained
SSR-optimal strategy. Then A(Tα∗ ) ≥ A(T ) and since S(Tα∗ ) ≥ S(T ), we have that
0 ≤ A(Tα∗ )B(T ) − A(T )B(Tα∗ ).
Suppose that Shrp2 (Tα∗ ) < Shrp2 (T ), then
0 ≤ A(Tα∗ )B(T ) − A(T )B(Tα∗ ) < A(Tα∗ )A(T ) · (A(T ) − A(Tα∗ )).
Both A(Tα∗ ) and A(T ) are > 0, otherwise both strategies are inferior to Tµ∗ ; thus A(T ) > A(Tα∗ ),
which is a contradiction. Therefore Shrp2 (Tα∗ ) ≥ Shrp2 (T ) and so Tα∗ is Shrp2 -optimal.
We will need the following technical property of any SSR-optimal interval (see appendix for proof).
Proposition 4.7 Let J be a subinterval of an SSR-optimal interval I. Then, µ(J) ≥ −fsp.
Further, if J is a prefix or suffix of I, then µ(J) > 0.
The next result establishes the intuitive result that adding an SSR-optimal interval to any trading
strategy can only improve the strategy.
Proposition 4.8 Let I0 be any SSR-optimal interval, and let T be any trading strategy. Let
T ′ = I0 ∪ T . Then A(T ′ ) ≥ A(T ) and S(T ′ ) ≥ S(T )
We can now give the intuition behind our algorithm. The starting point is Proposition 4.6, which
says that it suffices to look for constrained SSR-optimal strategies. So the natural first choice is
an unconstrained SSR-optimal interval T0 . Either this will be Shrp2 optimal or not. If not, it is
34
because it has too small a return. So our next step is to add to this interval a new interval (possibly
adjacent) with the property that the interval increases the return with smallest possible decrease in
SSR, resulting in strategy T1 . We repeat this process, constructing a sequence of trading strategies
T0 , T1 , . . . with the property that A(Ti ) > A(Ti−1 ), and among all other strategies T such that
A(T ) > A(Ti−1 ), S(Ti ) ≥ S(T ). We then pick the strategy Ti∗ with maximum Shrp2 ratio among
these strategies, which will be globally sharpe optimal.
Suppose that we have a current strategy, Ti . We need to determine the next piece to add to
this so that we increase the return, with smallest possible decrease in SSR. Let Ti be composed of
the intervals I0 , I1 , . . . , Id . We replace each of these intervals by a special symbol, $, to signify that
these regions are already included in the strategy. We thus obtain a generalized returns sequence,
one in which some intervals are replaced by the $ symbol. A generalized trading strategy on the
generalized return sequence must be composed of trades that do not contain the $ symbol. However
trades may be adjacent to the $ symbol. A trade interval I in a generalized trading strategy can be
isolated (not adjacent to any $ symbol), extending (adjacent to one $ symbol), or bridging (adjacent
to two $ symbols). In order to correctly account for the transactions cost, we need to change how
we compute A(I), so we introduce the new function Ā(I):
A(I)
Ā(I) = A(I) +
A(I) +
I is isolated
fsp
n
2fsp
n
I is extending
I is bridging
The generalized simplified Sharp ratio (GSSR) for generalized strategy T = {I1 , . . . , Id } is
P
Ā(Ii )
S̄(T ) = P i=1...d
i=1...d B(Ii )
Similar to the notion of a maximal return optimal strategy, we introduce the notion of a maximal SSR-optimal (or GSSR-optimal) interval as one which cannot be extended in either direction
without decreasing the SSR (or GSSR).
We now define generalized return sequences {R0 , R1 , . . .} as follows. R0 is just the original
returns sequence. Let Ii be a maximal GSSR-optimal interval for Ri . We obtain the generalized
sequence Ri+1 by replacing Ii ⊂ Ri with the symbol $. We define any set of generalized sequences
obtained in this way as monotone. We also refer to a member of a monotone set as monotone.
Let R0 , R1 , . . . , Rk be a monotone sequence of gerenalized returns sequences, and let I0 , I1 , . . . , Ik
be the maximal GSSR-optimal intervals corresponding to each sequence. By construction, Ii is a
maximal GSSR-optimal interval for Ri . We have defined Ā so that the SSR of the union of these
35
intervals in R0 is given by
Pd
SR0 (I0 ∪ I1 ∪ · · · ∪ Ik ) = Pi=1
d
ĀRi (Ii )
i=1 B(Ii )
,
where the subscript Ri indicates on which generalized return sequence the quantity is computed.
Lemma 4.9 S̄Ri (Ii ) ≥ S̄Ri+1 (Ii+1 )
Proof:
Suppose that S̄Ri (Ii ) < S̄Ri+1 (Ii+1 ), and let $i be the symbol that replaced Ii in Ri to
obtain Ri+1 . If Ii+1 is not adjacent with $i , then Ii is not GSSR-optimal in Ri , a contradiction. If
Ii+1 is adjacent with $i , then Ii ∪ Ii+1 has higher GSSR (by Lemma 4.1), so once again Ii is not
GSSR-optimal in Ri .
Now an easy induction, using Lemmas 4.1 and 4.9 gives,
Corollary 4.10 SR0 (I0 ∪ I1 ∪ · · · ∪ Ik ) ≥ S̄Rk (Ik ) for any k.
Analogous to Propositions 4.4, 4.7, 4.8, we have the following three propositions. Their proofs are
almost identical, so we omit them.
Proposition 4.11 A GSSR-optimal strategy making one trade exists, and all maximal GSSRoptimal trades can be found in O(N logN ) time.
Proposition 4.12 Let J be a subinterval of any GSSR-optimal interval I. Then µ(J) ≥ −fsp . If
J is a prefix or suffix of I that is not adjacent with the symbol ”$”, then µ(J) > 0.
Proposition 4.13 Let I0 be any GSSR-optimal interval, and let T be any generalized trading
strategy. Let T ′ = I0 ∪ T . Then, Ā(T ′ ) ≥ Ā(T ) and S̄(T ′ ) ≥ S̄(T ).
We now give the main result that will lead to the final algorithm to obtain the Shrp2 -optimal strategy. Its essential content is that given a monotone set of generalized returns sequences, R0 , R1 , . . .,
with corresponding GSSR-optimal intervals I0 , I1 , . . ., for some k, T = I0 ∪ I1 ∪ · · · ∪ Ik is Shrp2
optimal. We will need some preliminary results.
Proposition 4.14 For some k, T ∗ = I0 ∪ I1 ∪ · · · ∪ Ik is Shrp2 -optimal, where Ii are the GSSRoptimal intervals corresponding to a monotone set of generalized returns sequences.
Proof:
First we show that there exists a Shrp2 -optimal strategy T0 that contains I0 . Indeed,
let T be any Shrp2 -optimal strategy, and consider T0 = I0 ∪ T . By the Proposition 4.8, we have
S(T0 ) ≥ S(T ) and A(T0 ) ≥ A(T ) ≥ 0. Then,
Shrp2 (T0 ) − Shrp2 (T ) =
A(T0 )A(T )(A(T0 ) − A(T )) + B(T0 )B(T )(S(T0 ) − S(T ))
≥ 0,
(B(T0 ) − A2 (T0 ))(B(T ) − A2 (T ))
36
thus, T0 is Shrp2 -optimal.
Let Tk be a Shrp2 -optimal strategy that contains I0 ∪ · · · ∪ Ik . We know that T0 exists. If
Tk = I0 ∪ · · · ∪ Ik , then we are done. If , Tk = I0 ∪ · · · ∪ Ik ∪ T ′ , with T ′ 6= ∅, then we show that
there must exist a Shrp2 -optimal strategy Tk+1 which contains I0 ∪ · · · ∪ Ik+1 , i.e., there is some
other T ′′ ⊇ Ik+1 such that Tk+1 = I0 ∪ · · · ∪ Ik ∪ T ′′ is Shrp2 -optimal. The proposition then follows
by an easy induction.
Let T ′′ = T ′ ∪ Ik+1 . Then, ĀRk+1 (T ′′ ) ≥ ĀRk+1 (T ′ ) and S̄Rk+1 (T ′′ ) ≥ S̄Rk+1 (T ′ ) (Proposition
4.13). By Corollary 4.10 and the GSSR-optimality of Ik+1 , we have that
S(I0 ∪ . . . ∪ Ik ) ≥ S̄Rk+1 (Ik+1 ) ≥ S̄Rk+1 (T ′′ ) ≥ S̄Rk+1 (T ′ )
From now on, we will drop the Rk+1 subscript. Let A = A(I0 ∪ . . . ∪ Ik ), B = B(I0 ∪ . . . ∪ Ik ),
A′ = Ā(T ′ ), B ′ = B(T ′ ), A′′ = Ā(T ′′ ) and B ′′ = B(T ′′ ). Let Shrp2 = Shrp2 (I0 ∪ . . . ∪ Ik ),
′
′′
Shrp2 = Shrp2 (I0 ∪ . . . ∪ Ik ∪ T ′ ) and Shrp2 = Shrp2 (I0 ∪ . . . ∪ Ik ∪ T ′′ ). Thus,
Shrp2 =
A
,
B − A2
′
Shrp2 =
A + A′
, and
B + B ′ − (A + A′ )2
′′
Shrp2 =
A + A′′
.
B + B ′′ − (A + A′′ )2
Let A′′ = α′′ A and A′ = α′ A, where α′′ ≥ α′ > 0. Then, by direct computation, one obtains
A
′
Shrp2 =
B+
A2
+
α′
B′
1+α′ ( α′
−B−
A2 )
′
Since Shrp2 > Shrp2 , we conclude that
B′
α′
≥
B ′′
α′′ ,
and since α′′ ≥ α′ > 0,
α′′
1 + α′′
′′
α′′
1+α′′
≥
, and
B′
2
α′ − B − A < 0.
α′
1+α′ > 0, therefore
B ′′
− B − A2
α′′
α′
≤
1 + α′
A
′′
Shrp2 =
B+
A2
+
α′′
B ′′
1+α′′ ( α′′
− B − A2 )
,
Since S̄(T ′′ ) ≥ S̄(T ′ ), we have that
B′
− B − A2
α′
< 0,
′
and so Shrp2 ≥ Shrp2 , concluding the proof.
By Proposition 4.14, a Shrp2 -optimal trading strategy can be obtained by constructing the strategies
Tk , and then picking the one with the maximum value for Shrp2 . The next proposition shows that
this can be done in O(n2 logn) time.
Proposition 4.15 A Shrp2 -optimal trading strategy can be found in time O(n2 logn).
Proof:
Ii can be obtained in O(n log n) time (Proposition 4.11). Since there are at most n such
intervals (since each must be non-empty), obtaining all the intervals is in O(n2 log n).
Given the intervals, a single scan can be used to obtain the k for which Tk is Shrp2 -optimal.
37
One can improve the runtime to O(n2 ) if O(n2 ) memory is available, however, we do not discuss
the details.
4.2.1
Approximation Ratio
We have given an algorithm that obtains a Shrp2 -optimal strategy. A modification to the algorithm
constructs the hierarchy Ti and pick the one with the highest value of Shrp2 . Suppose we have
a Shrp2 -optimal strategy T and let T ∗ be a Shrp2 -optimal strategy. Then by Proposition 4.6, it
must be that A(T ∗ ) > A(T ) and that S(T ∗ ) ≤ S(T ). Since Shrp2 (T ) ≥ Shrp2 (T ∗ ), we have that
A∗ (B − A2 ) − A(B ∗ − A∗ 2 ) ≤ 0, where A = A(T ), A∗ = A(T ∗ ), B = B(T ), B ∗ = B(T ∗ ). We can
evaluate Shrp2 (T ∗ ) − Shrp2 (T ) to obtain
0 ≤ Shrp2 (T ∗ ) − Shrp2 (T ) ≤ Shrp∗2 ·
d
n fsp
2
d
n fsp
2
+ B − A2
When B − A2 = O(1), which is usually the case, we see that this is a very accurate approximation
(since fsp ≪ 1).
4.3
Maximizing Shrp1
Once again, we introduce the slightly different quantity Shrp1 ,
Shrp1 (T ) = q
A(T )
2
d
n fsp
Shrp1 (T ) = p
,
+ B(T ) −
A2 (T
)
A(T )
B(T ) − A2 (T )
.
2
We will optimize Shrp1 (T ). Since maximizing Shrp1 (T ) is equivalent to minimizing 1/Shrp1 (T ) the
problem reduces to maximizing
Q(T ) =
A2 (T )
B(T )
The entire algorithm is analogous to that for maximizing Shrp2 in the previous section, we only
need to prove the analogs of Propositions 4.6 and 4.14.
Proposition 4.16 ∃0 ≤ α ≤ µ∗ such that the constrained SSR-optimal strategy Tα is Q-optimal.
Proof:
Let T be Shrp1 -optimal, and let α∗ = A(T ). Let Tα∗ be a corresponding constrained
SSR-optimal strategy. A(Tα∗ ) ≥ A(T ) and
that
A2 (Tα∗ )
B(Tα∗ )
≥
A2 (T )
B(T ) ,
A(Tα∗ )
B(Tα∗ )
≥
A(T )
B(T ) .
Multiplying these two inequalities gives
i.e. Tα∗ is also Q-optimal.
Proposition 4.17 For some k, T ∗ = I0 ∪I1 ∪· · ·∪Ik is Q-optimal., where Ii are the GSSR-optimal
intervals corresponding to a monotone set of generalized returns sequences.
38
The proof is very similar to tho proof of Proposition 4.14. Let T be Q-optimal, and let
Proof:
T0 = I0 ∪ T . Then A(T0 ) ≥ A(T ) and S(T0 ) ≥ S(T ). Multiplying these two inequalities give that
Q(T0 ) ≥ Q(T ), or that T0 is also Q-optimal.
Let Tk be a Q-optimal strategy that contains I0 ∪ · · · ∪ Ik . Introduce T ′ , T ′′ = Ik+1 ∪ T ′
as in the proof of Proposition 4.14. Let Q = Q(I0 ∪ . . . ∪ Ik ), Q′ = Q(I0 ∪ . . . ∪ Ik ∪ T ′ ) and
Q′′ = Q(I0 ∪ . . . ∪ Ik ∪ T ′′ ),
Q=
A2
,
B
Q′ =
(A + A′ )2
, and
B + B′
(A + A′′ )2
.
B + B ′′
Q′′ =
Following exactly the same logic as in the proof to Proposition 4.14, we only need to show that
Q′′ ≥ Q′ . Let A′′ = α′′ A and A′ = α′ A, where α′′ ≥ α′ > 0.
B ′′
α′′ (α′′ +2)
≤
Q′ =
B′
α′ (α′ +1) .
≥
A′
B′
implies that
B ′′
α′′
≤
B′
α′ ,
and so
By direct computation, one obtains
B+ 1−
A2
1
(1+α′ )2
Since Q′ > Q, it must be that
have that
A′′
B ′′
1
1−
(1 + α′′ )2
Q′′ =
,
B+ 1−
A2
1
(1+α′′ )2
B′
α′ (α′ +2)
−B
B′
α′ (α′ +2)
− B < 0. Since α′′ ≥ α′ , 1 −
B ′′
−B
α′ (α′′ + 2)
≤
1
1−
(1 + α′ )2
1
(1+α′′ )2
B ′′
α′′ (α′′ +2)
≥ 1−
B′
−B
α′ (α′ + 2)
−B
.
1
,
(1+α′ )2
so we
< 0,
which implies that Q′′ ≥ Q′ .
4.3.1
Approximation Ratio
Once again, a modification to the algorithm constructs the hierarchy Ti and picks the one with
the highest one with value of Shrp1 . Suppose we have a Shrp1 -optimal strategy T and let T ∗ be a
Shrp1 -optimal strategy. By direct computation, and using the fact that Shrp1 (T ) ≥ Shrp1 (T ∗ ) =⇒
A∗ 2 B − A2 B ∗ ≤ 0, we get
0≤
Shrp21 (T ∗ )
which gives an approximation ratio of
−
q
Shrp21 (T
)≤
dfsp 2
2
∗
n
Shrp1 (T )
dfsp 2
n +
1 − O(fsp 2 ) when B = O(1).
39
B
5
Discusion
Our main goal was to provide the theoretical basis for the computation of a posteriori optimal
trading strategies, with respect to various criteria. In particular, we have presented the algorithms,
together with the proofs of their correctness. The highlights of our contributions are that return and
sterling optimal strategies can be computed very efficiently, even with constraints on the number
of trades. Sharpe optimal strategies prove to be much tougher to compute. However, for slightly
modified notions of the Sharpe ratio, where one ignores the impact of bid-ask spread squared we
can compute the optimal strategy efficiently. This is a reasonable approach since in most cases, the
bid-ask spread is ∼ 10−4 . We also show that this modified optimal strategy is not far from optimal
with respect to the unmodified Sharpe ratio.
Interesting topics of future research are to use these optimal strategies to learn how to trade
optimally, which was the original motivation of this work. Further, one could use them to benchmark trading strategies as well as markets. A natural open problem is whether Sharpe optimal
strategies can be computed under constraints on the number of trades. We suspect that the monotone hierarchy we created with respect to the SSR has an important role to play, but the result has
been elusive.
The algorithms presented here can be useful if one could relax Assumption A1. We have used
such an assumption because it considerably simplifies the analysis, and in many cases, the optimal
strategy is in fact an all-or-nothing strategy.
At the core of some of our algorithms is a new technique for optimizing quotients over intervals
of a sequence. This technique is based on relating the problem to convex set operations, and for our
purposes has direct application to optimizing the M DD, the simplified Sharpe ratio (SSR), which
is an integral component in optimizing the Sharpe ratio, and the Downside Deviation Ratio (DDR).
This technique may be of more general use in optimizing other financially important criteria. As a
result, such an algorithm may be of independent interest.
References
[1] Online encyclopedia of financial terms. http://www.investopedia.com, 2004.
[2] Arjan Berkelaar and Roy Kouwenberg. Dynamic asset allocation and downside-risk aversion.
citeseer.ist.psu.edu/berkelaar00dynamic.html.
[3] Tomasz R. Bielecki, Jean-Philippe Chancelier, Stanley Pliska, and Agnes Sulem. Risk sensitive
portfolio optimization with transaction costs. To appear, 2004.
40
[4] Tomasz R. Bielecki, Stanley Pliska, and Jiongmin Yong. Optimal investment decisions for a
portfolio with a rolling horizon bond and a discount bond. To appear, 2004.
[5] Timothy M. Chan. Output-sensitive results on convex hulls, extreme points, and related
problems. In Proc. 11th Symposium on Computational Geometry, pages 10 – 19, 1995.
[6] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction
to Algorithms. Mcgraw-Hill, 2001.
[7] Sergei Issaenko. Domenico Cuoco, Hua He. Optimal dynamic trading strategies with risk
limits. SSRN Electronic Paper Collection, 2001.
[8] Bhaswar Gupta and Manolis Chatiras. The interdependence of managed futures risk measures.
pages 203–220. Center for International Securities and Derivatives Markets, October 2003.
[9] Matthew Saffell. John Moody. Learning to trade via direct reinforcement. IEEE Transactions
on Neural Networks, 12(4):875–889, 2001.
[10] Hong Liu. Optimal consumption and investment with transaction costs and multiple risky
assets. The Journal of Finance, 59(1):289, 2004.
[11] Stephen Boyd Miguel Sousa Lobo, Maryam Fazel. Portfolio optimization with linear and fixed
transaction costs. Annals of Operations Research, special issue on financial optimization, to
appear, 2006.
[12] Oliver Mihatsch and Ralph Neuneier. Risk-sensitive reinforcement learning. Mach. Learn.,
49(2-3):267–290, 2002.
[13] Jiming Liu. Samuel P. M. Choi. Optimal time-constrained trading strategies for autonomous
agents. In Proceedings of International ICSC Symposium on Multi-agents and Mobile Agents
in Virtual Organizations and E-Commerce (MAMA 2000), 2000.
[14] G. W. P. Thompson. Optimal trading of an asset driven by a hidden markov process in
the presence of fixed transaction cost. In Collection of research papers. Judge Institute of
Management Working Papers, 2002.
[15] Jiqing Han. Yang Liu, Xiaohui Yu. Sharp ratio-oriented active trading: A learning approach.
SSRN Electronic Paper Collection, 2001.
[16] Valeri I. Zakamouline. Optimal portfolio selection with both fixed and proportional transaction
costs for a CRRA investor with finite horizon. Norwegian School of Economics and Business
Administration, 2002.
41
A
Technical Proofs
Lemma 3.13.
Proof: Let A = O(m) be the size of the current convex hull. For (1), we do not change the points at
all, we simply compute the new tangent point for p′ = p+v, which can be accomplished in O(log A).
(2) is equivalent to shifting p by −v. To prove (3), notice that if we remove (d0 , c0 ), then the new
leftmost point becomes (d1 , c1 ) and we immediately have the new convex hull nxt(1), nxt(nxt(1)), . . ..
Thus we can find the new upper tangent point in O(log A′ ) = O(log m), where A′ is the size of
the new convex hull. Further, deleting (d0 , c0 ) requires first deleting the backward pointers of the
points that it points to O(log A), and then deleting the point itself, and its forward pointers (it
has no backward pointers), O(log A). To prove (4), note that when we add (d−1 , c−1 ), nxt(−1)
is exactly the upper tangent point from p′ = (d−1 , c−1 ) to the current convex hull. This can be
computed in O(log A). We now need to add all the necessary pointers into the data structure. For
each forward pointer we add, we will add the coresponding backward pointer as well. We need a
pointer at position 2j in the convex hull of (d−1 , c−1 ). But this is exactly the point at position 2j −1
in the convex hull of point nxt(−1). Since nxt(−1) maintains a pointer to point 2j in its convex
hull, and this point will have a backward pointer by one step of this same convex hull, we can
construct the forward an backward pointer for point 2j in the convex hull of (d−1 , c−1 ) in constant
time, requiring total time O(log A′ ) = O(log m) to construct all the new forward and backward
pointers, where A′ is the size of the new convex hull. We now construct the new upper tangent
point from p to the new convex hull of (d−1 , c−1 ) in O(log A′ ) time. The entire process is therefore
O(log m).
The algorithm that we have just described is a general purpose algorithm for efficiently maintaining
the upper tangent point to any set of points, as long as only a limited set of operations is allowed
on the set of points and the source point.
Proposition 3.18.
Proof:
First we show that for every trading interval I ∗ = [ta , tb ] in Tµ∗ with I ∩ I ∗ 6= ∅, one can
K such that I ∗ ⊆ I. Suppose to the contrary, that for some I ∗ , either t < t and t ≥ t or
pick TStrl
a
l
b
l
K so that I ∗ ⊆ I.
ta ≤ tr and tb > tr . We will extend I without decreasing the Sterling ratio of TStrl
Suppose ta < tl and tb ≥ tl (a similar argument holds for ta ≤ tr and tb > tr ). There are two cases:
K : Applying Lemma 3.17 to I ∗ , we have: C ≤ C .
i. I ∗ does not intersects any other interval of TStrl
a
l
Thus by extending I to [ta , tr ], the return of the interval cannot decrease. Since M DD(I ∗ ) ≤
K ), since we already have that M DD(T K ) ≥
fsp , this extension cannot increase the M DD(TStrl
Strl
fsp .
42
K : ∃I ′ = [t ′ , t ′ ] ∈ T K such that
ii. I ∗ intersects with the previous trading interval of strategy TStrl
r
l
Strl
Pl−1
∗
ta ≤ tr′ < tl . Since [tr′ +1 , tl−1 ] is a subinterval of I , j=r′ +1 ŝj ≥ −fsp (Lemma 3.16). If we
K , we save on the transaction cost
merge I and I ′ by adding the interval [tr′ +1 , tl−1 ] into TStrl
of fsp, and so the total return will not decrease. We show that the M DD has not increased.
Since Cr′ is a maximum in [tl′ , tr′ ], the drawdown for all points in [tr′ +1 , tl ] is at most fsp . Since
Cl is a minimum in [tl , tr ], we conclude that the drawdown for any point in [tl , tr ] is at most
K ) ≥ f , we conclude that this merger does not increase
max{fsp , M DD(I)}. Since M DD(TStrl
sp
the M DD.
K by removing I, without increasing the
Note that µ(I) ≥ 0 otherwise we improve the return of TStrl
K cannot possibly be optimal. Thus, without loss of generality, we assume that
M DD, ans so TStrl
the return of I is positive. Suppose that I ∩ Tµ∗ = ∅. Then, by adding I to Tµ∗ , we strictly increase
K contains at least
the return, a contradiction on the optimality of Tµ∗ . Thus, every interval of TStrl
one interval of Tµ∗ . Now consider the maximal prefix Pmax of I that does not overlap with any
interval of Tµ∗ . Since we know that I contains some interval of Tµ∗ , we conclude that this maximal
prefix must be adjacent to some interval of Tµ∗ . By Lemma 3.16, this interval has strictly negative
K , without increasing its M DD.
return, so removing it from I strictly increase the return of TStrl
K , thus, P
This contradicts the optimality of TStrl
max must be empty. Similarily, the maximal suffix
of I that is non-intersecting with Tµ∗ must be empty, concluding the proof.
Lemma 3.19.
Proof:
If K = K0 , then Tµ∗ itself is K-Sterling-optimal. If K < K0 , we show that if the number
of trades made is less than K, we can always add one more interval without decreasing the Sterling
K cannot contain all the intervals of T ∗ , as otherwise (by
ratio of the strategy. First, note that TStrl
µ
K contains two consecutive intervals
the pigeonhole principle) at least one interval I = [tl , tr ] of TStrl
I1 = [tl1 , tr1 ] and I1 = [tl2 , tr2 ] of Tµ∗ . The region between these two intervals has return less than
−fsp (Lemma 3.16), so breaking up I into the two intervals [tl , tr1 ] and [tl2 , tr ] will strictly increase
K . If T K does not
the return, without increasing the M DD, contradicting the optimality of TStrl
Strl
contain some interval of Tµ∗ , then by adding this interval, we do not decrease the return or the
M DD (Lemma 3.16), since the M DD is already ≥ fsp .
Proposition 4.7.
Proof: If J is a prefix or suffix of I and µ(J) ≤ 0, then deleting J from I gives at least as much
return, with smaller sum of squared returns, contradicting the SSR-optimality of I. Suppose that
I = L ∪ J ∪ R where L and R are nonempty subintervals of I. If
43
µ(J)+µ(R)
s2 (J)+s2 (R)
<
−fsp +µ(L)
,
s2 (L)
then by
Lemma 4.1,
S(I) =
−fsp + µ(L) + µ(J) + µ(R)
−fsp + µ(L)
<
= S(L)
2
2
2
s (L) + s (J) + s (R)
s2 (L)
(*)
This contradicts the optimality of I, so we have
−fsp + µ(L)
µ(J) + µ(R)
≥
.
2
2
s (J) + s (R)
s2 (L)
Now, suppose that µ(J) < −fsp . Using e̊q:* and Lemma 4.1, we find that
S(I) ≤
µ(J) + µ(R)
−fsp + µ(R)
−fsp + µ(R)
< 2
<
= S(R),
s2 (J) + s2 (R)
s (J) + s2 (R)
s2 (R)
because s2 (J) > 0. This contradicts the SSR-optimality of I, so µ(J) ≥ −fsp.
Proposition 4.8.
Proof:
If I0 is contained in some interval of T , then there is nothing to prove, so suppose that
T ∪ I0 6= T , and that T contains d ≥ 1 trades I1 , . . . , Id . Note that S(T ) = A(T )/B(T ). If I0 and
T do not intersect, then T ′ = I0 ∪ T = {I0 , I1 , . . . , Id }. A(T ′ ) = A(T ) + A(I0 ) ≥ A(T ), because
A(I0 ) ≥ 0. Since I0 is SSR-optimal, S(I0 ) =
S(T ′ ) =
A(I0 )
B(I0 )
≥
A(T )
B(T )
= S(T ), so by lemma 4.1,
A(T ) + A(I0 )
A(T )
A(T ′ )
=
≥
= S(T ).
B(T ′ )
B(T ) + B(I0 )
B(T )
Suppose that T ∩ I0 6= ∅. We can decompose T into four parts (each part could be empty):
T = S1 ∪ S2 ∪ Il ∪ Ir , where S1 contains intervals that do not intersect with I0 , S2 contains intervals
that are contained in I0 , Il is not contained in I0 but overlaps I0 on the left, and Ir is not contained
in I0 but overlaps I0 on the right. T ′ = I0 ∪ T = S1 ∪ Il ∪ I0 ∪ Ir , i.e., adding I0 combines all
the trades in {S2 , Il , Ir } into one trade. Since the internal regions of I0 have return at least −fsp
and any prefix and suffix of I0 has positive return (Proposition 4.7), we see that merging any two
consecutive trades overlapping I0 decreases the number of trades by one, hence increases the return
by fsp and the added interval loses at most fsp , hence this merge can only increase A(T ). If either
Il or Ir are empty, then we are addionally adding a prefix or suffix of I0 without changing the
number of trades, which also increases A(T ), thus we see that A(T ′ ) ≥ A(T ).
44
Let’s introduce the following definitions,
1
(µ(Il ∩ I0 ) + µ(Ir ∩ I0 ))
n
1
= A(S2 ) + (µ(Il ∩ I0 ) − fsp + µ(Ir ∩ I0 ) − fsp )
n
1
= B(S1 ) + (s2 (Il ∩ I0 ) + s2 (Ir ∩ I0 ))
n
1
= B(S2 ) + (s2 (Il ∩ I0 ) + s2 (Ir ∩ I0 )),
n
A1 = A(S1 ) +
A2
B1
B2
where I0 is the complement of I0 . Letting A0 = A(I0 ) and B0 = B(I0 ), we then have
S(T ) =
A1 + A2
A1 + A0
, and S(T ′ ) =
.
B1 + B2
B1 + B0
Note that S(S2 ∪ (Il ∩ I0 ) ∪ (Ir ∩ I0 )) =
1
n µ(Il ∩ I0 )
1 2
n s (Il ∩ I0 )
If not, then suppose (for example) that
S(I0 ∪ Il ) =
A2
B2 ,
≤
so by the optimality of I0 ,
A0
, and
B0
1
µ(Il ∩I0 )
n
1 2
s
(Il ∩I0 )
n
>
1
n µ(Ir ∩ I0 )
1 2
n s (Ir ∩ I0 )
A0
B0 .
A0 + n1 µ(Il ∩ I0 )
B0 +
1 2
n s (Il
∩ I0 )
≤
A2
B2
≤
A0
B0 .
We show that
A0
.
B0
(**)
Then,
>
A0
= S(I0 )
B0
contradicting the SSR-optimality of I0 . Again, by the optimality of I0 , S(S1 ) =
A(S1 )
B(S2 )
≤
A0
B0
=
S(I0 ). it also has to be sharper than strategy S1 . Thus, using (**) and Lemma 4.1 we have that
A1
B1
≤
A0
B0 .
Because A2 is obtained from the returns of a collection of subintervals of I0 , it follows
from Proposition 4.7 that A2 ≤ A0 . Now suppose that S(T ) > S(T ′ ), i.e.,
(A1 + A2 )(B1 + B0 ) − (A1 + A0 )(B1 + B2 ) > 0.
Since A2 ≤ A0 , it follows that B2 ≤ B0 . Rearranging terms in the equation above, we have that
A2 B1 − A1 B2
B0 (B1 + B2 )
>
≥
A0
A1 + A2
−
,
B0 B1 + B2
A1 + A2
A2 B1 − A1 B2
A2
−
=
.
B2 B1 + B2
B2 (B1 + B2 )
Since S(I0 ) ≥ S(T ), the first inequality shows that A2 B1 − A1 B2 > 0. The second inequality then
implies that B2 > B0 , a contradiction.
45
| 5cs.CE
|
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
arXiv:1802.08113v4 [math.OC] 31 Mar 2018
Please note that where the full-text provided is the Author Accepted Manuscript or Post-Print version this may differ from the
final Published version.
To cite this publication, please use the final published version.
Personal use of this material is permitted. Permission from the author(s) and/or copyright holder(s), must be obtained for all other
uses, in any current or future media, including reprinting or republishing this material for advertising or promotional purposes.
Publication information:
Date of Submission: January 2016.
Date of Acceptance: September 2016.
Available Online: September 2016.
Please contact us and provide details if you believe this document breaches copyrights. We will remove access to the work immediately and investigate your claim.
Adaptive synchronization of unknown nonlinear networked systems
with prescribed performance
Hashim A. Hashima , Sami El-Ferikb and Frank L. Lewisc
a
Department of Electrical and Computer Engineering, University of Western Ontario, London, ON, Canada, N6A-5B9.
Email: [email protected]
b
King Fahd University of Petroleum and Minerals, Dhahran, 31261, Saudi Arabia.
c
UTA Research Institute, The University of Texas at Arlington 7300 Jack Newell Blvd. S, Ft. Worth, Texas 76118.
This paper proposes an adaptive tracking control with prescribed performance function for
distributive cooperative control of highly nonlinear multi-agent systems. The use of such
approach confines the tracking error within a large predefined set to a predefined smaller
set. The key idea is to transform the constrained system into unconstrained one through
the transformation of the output error. Agents’ dynamics are assumed unknown, and the
controller is developed for a strongly connected structured network. The proposed controller
allows all agents to follow the trajectory of the leader node, while satisfying the necessary
dynamic requirements. The proposed approach guarantees uniform ultimate boundedness for
the transformed error as well as a bounded adaptive estimate of the unknown parameters and
dynamics. Simulations include two examples to validate the robustness and smoothness of the
proposed controller against highly nonlinear heterogeneous multi-agent system with uncertain
time-variant parameters and external disturbances.
Keywords: Prescribed performance, Transformed error, Multi-agents, Distributed adaptive
control, Consensus, Transient, Steady-state error, Networked Systems, Distributed Adaptive
Control, Robustness.
1
Introduction
In recent years, distributive cooperative control has gained
popularity and attention among control researchers owing to its
capabilities to mimic the social behavior of animals such as bees
swarming, birds flocking, ants foraging, fish schooling and so forth.
Indeed, the control scheme enables a group of agents to perform
a task that can be daunting for an individual agent in a simpler
and faster manner. Besides, the cooperation between agents allows
information exchange, improving performance and increasing
productivity through collaboration, which emulates the standard
behavior in social groups. Cooperative control contributes to the
betterments of many applications such as the control of autonomous
mobile robot vehicles in energy and mineral explorations, space
studies, surveillance and other areas. Agents are connected by
a communication network and exchange useful information. In
such case, they are considered as nodes. The group of agents may
follow one or more real or virtual leaders. The network formed
by all nodes creates a graph, which can be directed or undirected.
Undirected graphs refer to no difference between nodes while
directed graphs define the direction of the flow of information
between the node and its neighbors.
In the literature, Fax & Murray (2004) and Ren et al. (2005)
can be considered among the first pioneering studies addressing
the consensus in multi-agent systems. Several other scholars
contributed to the development in this field. For instance, the
consensus of passive nonlinear systems has been addressed in
Chopra & Spong (2006). Many research work such as Olfati-Saber
et al. (2007)-Zhou et al. (2015) investigated node consensus of
cooperative tracking problem. Distributed tracking control for
linear heterogeneous agents of MIMO systems with parameter
uncertainties was established in Y. Zhao et al. (2015). Cooperative
tracking control for a single node has been studied in Das & Lewis
(2010); Cao & Ren (2012) and in the case of high-order dynamics
in Zhang & Lewis (2012). The work in Das & Lewis (2010); Zhang
& Lewis (2012) developed a neuro-adaptive distributed control for
heterogeneous agents with unknown nonlinear dynamics connected
through a digraph. In Das & Lewis (2010), the authors considered
nodes with first-order dynamic. Later on, in Zhang & Lewis (2012),
high order systems have been addressed. In all previous studies,
the input function in the node dynamics was assumed known.
On the other hand, cooperative tracking control problems
of systems with unknown input function have been studied
Theodoridis et al. (2012) and El-Ferik et al. (2014). In Theodoridis
et al. (2012), neuro-adaptive fuzzy was proposed to approximate
unknown nonlinear dynamics and input functions. The centers of
the output membership functions are determined based on off-line
trials. A very fundamental assumption in all these studies is the one
that considers the unknown nonlinear dynamics as well as the input
function as linear in parameters (LIP) (see for instance F. W. Lewis
et al. (1998) or El-Ferik et al. (2014)). The goal is to guarantee the
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
ultimate stability of the tracking error. Recently, there are several
studies that were published addressing different issues related to
adaptive control of multi-agent systems. These challenges include
actuator fault (see for instance Na (2013), Na et al. (2014),Tong et
al. (2014), Tong et al. (2015), Y. Li & Tong (2015) and L. Zhao
& Jia (2016), switching network topology Yang et al. (2016),
Predictor-based adaptive control W. Wang et al. (2016), etc. All
these challenging practical issues could benefit from prescribed
performance framework to guarantee performance. In particular, a
practical implementation of neuro-adaptive prescribed performance
control to compensate for friction using a turn table servo system
has been reported in Na et al. (2014). More implementation of such
control approaches is really needed.
The distributed control of multi-agents attempts to tackle
unknown nonlinearities, unmodeled dynamics, uncertainties, and
disturbances. Estimation of the closed loop characteristics such
as transient and steady state error is almost impossible to be
represented analytically Bechlioulis et al. (2014). Alternatively,
prescribed performance has been proposed as a means to seclude
the error to an arbitrarily small set, where the convergence is
constrained to a given range. The key idea in the approach is to
transform the error from the restricted space to the unconstrained
one. The following section gives necessary details about the
method. At this stage, it is worth to mention that prescribed
performance approach aims to satisfy the following objectives.
The convergence error has to be less that the predefined value; the
transformed error is bounded; the maximum overshoot is less than
the prescribed constant; the system’s controlled output is smooth;
and the control signal is both bounded and smooth.
Developing a cooperative control approach for multi-agent
systems with prescribed performance has many benefits. In this
context, the specified performance ensures that the consensus
output error starts within a large predefined set and then converges systematically into a predefined narrow set Bechlioulis &
Rovithakis (2008). During transient and steady-state, the tracking
error satisfies a known time-varying performance. Adaptive
cooperative control with prescribed performance has then the
ability to increase the robustness of the system’s behavior and to
reduce the control effort. The proper selection of the upper and
lower bounds of the prescribed performance functions guarantees
the convergence of error within predefined limits smoothly
and systematically. In the literature, robust adaptive control
with prescribed performance function for feedback linearizable
systems has been designed in Bechlioulis & Rovithakis (2008).
The design of neuro-adaptive controllers to handle unknown
nonlinearities and disturbances has been considered in Bechlioulis
& Rovithakis (2009)-Yang et al. (2015) for different applications.
The application of prescribed performance scheme with neural
approximation included strict-feedback systems Bechlioulis &
Rovithakis (2009), affine systems J. Wang et al. (2010), high
order nonlinear systems Bechlioulis & Rovithakis (2014). Each of
these studies considered different assumptions on the input matrix
continuity. Further refinement of these results improved the neural
network weights previously tuned using trial and errors to avoid
neural nets in the controller design and redevelop adaptive control
with prescribed performance based on fuzzy adaptive tuning in Sun
& Liu (2014) and model reference adaptive control in Mohamed
(2014). All previous studies considered a single autonomous
system. Recently, Bechlioulis et al. (2014) designed a control of a
platoon with unknown nonlinear dynamics. Thus, the agents are
set in a straight line, and each node sees only the one in front. This
represents a particular network structure and a special formation
for the nodes.
In this work, we propose a robust adaptive distributive control
with prescribed performance for a group of nodes connected
through a directed communication graph with known topology.
The control law is fully distributed based on the fact that the control
law of each agent respects the graph’s topology and uses only the
allowed local neighborhood information. Thus, the leader does not
communicate with all the nodes. In our work, we consider a general
network form characterized by its L and B matrices. The formation
of the multi-agent systems can be anything including platooning.
The synchronization error between nodes follows a prescribed
performance to satisfy predefined characteristics imposed by the
designer. Each node contains unknown nonlinear dynamics and
time-varying uncertainties. The controller is developed to meet
a predefined transient response and specific characteristics of the
steady-state synchronization error for each node. The original form
of the prescribed performance as originally proposed in Bechlioulis
& Rovithakis (2008) is modified in this work to overcome the
chattering in the control signal due to the interaction between
nodes caused by the consensus algorithm the exchanging state
information between nodes. The new approach guarantees stable
dynamics with non-oscillatory, limited and smooth control signal.
The rest of the paper is organized as follows. Section 2 presents
preliminaries of graph theory. Problem formulation and the
derivation of the local error synchronization equation follow in
Section 3. Section 4 contains the control law development as well
as the stability proof of the connected graph. Simulations results are
presented in Section 5. Conclusion and future work are in Section 6.
Notations:The following symbols are used throughout the
paper.
|·|
: absolute value of a real number;
k·k
: Euclidean norm of a vector;
k · kF
: Frobenius norm of a matrix;
tr{·}
: trace of a matrix;
σ(·)
: set of singular values of a matrix, with the maximum
singular value λ̄ and the minimum singular value λ;
P > 0 : indicates that the matrix P is positive definite;
(P ≥ 0) (positive semi-definite);
N
: set {1, ..., N};
Im
: identity matrix of order m.
1
: unity vector [1, . . . , 1]> ∈ Rn where n is the required
appropriate dimension.
2
Basic graph theory
A graph is denoted by G = (V, E) with a nonempty finite set
of nodes (or vertices) V = {V1 , V2 , . . . , Vn }, and a set of edges
(or arcs) E ⊆ V × V. (Vi , V j ) ∈ E if there is an edge from
node i to node j. The topology of a weighted graph is often described by the adjacency matrix A = [ai j ] ∈ RN×N with weights
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
ai j > 0 if (V j , Vi ) ∈ E: otherwise ai j = 0. Throughout the paper, the topology of the communication network is fixed, i.e. A is
time-invariant, and the self-connectivity element aii = 0. A graph
can be directed or undirected. A directed graph is called digraph.
The weight in-degree of a node i is defined as the sum of i-th
P
row of A, i.e., di = Nj=1 ai j . Define the diagonal in-degree matrix D = diag(d1 , . . . , dN ) ∈ RN×N and the graph Laplacian matrix
L = D−A. The set of neighbors of a node i is Ni = { j|(V j ×Vi ) ∈ E}.
If node j is a neighbor of node i, then node i can get the information from node j , but not necessarily vice-versa. For an undirected
graph, the neighborhood requires a mutual relation. A direct path
from node i to node j is a sequence of successive edges in the form
{(Vi , Vk ), (Vk , Vl ), . . . , (Vm , V j }. A digraph has a spanning tree if
there is a node (called the root) having a possible direct path to every
other node in the graph. A digraph is strongly connected if for any
ordered pair of nodes [Vi , V j ] with i , j, there is a directed path
from node i to node j (for more details, see Ren & Beard (2008) or
F. L. Lewis et al. (2013)).
B = diag{bi } and 1 = [1, . . . , 1]> ∈ RN . Note that x̃ = x − 1 · x0 , and
f (x0 , t) = 1 · f (x0 , t). For more details, the proof of equation (5) is
stated in F. L. Lewis et al. (2013).
The derivative error dynamics of (5) is
ė = (L + B) (Am x̃ + f (x) + Bm u + w − f (x0 , t))
(6)
Remark 1 The communication graph is considered strongly connected. Thus, if bi , 0 for at least one i, i = 1, . . . , N then (L + B)
is an irreducible diagonally dominant M-matrix and hence nonsingular Qu (2009).
Remark 2 (see F. L. Lewis et al. (2013)) If agent state is xi ∈ Rn
and the leader state x0 ∈ Rn where n > 0, then e, x ∈ RnN and
equation (5) will be
e = ((L + B) ⊗ IN ) (x − 1x0 )
(7)
where ⊗ is the Kronecker product.
3
Problem Formulation
Consider the following nonlinear dynamics for the ith node
ẋi = Ami xi + Bmi ui + fi (xi ) + wi
(1)
where the state node is xi ∈ R, the control signal node is ui ∈ R
and the unknown disturbance for each node is wi ∈ R. Ami ∈ R and
Bmi ∈ R are known constants; fi (xi ) ∈ R is the unknown nonlinear
part of the dynamics and assumed to be Lipschitz. From (1), the
global dynamic can be written as
ẋ = Am x + Bm u + f (x) + w
(2)
where x = [x1 , . . . , xN ]> ∈ RN , u = [u1 , . . . , uN ]> ∈ RN ,
f (x) = [ f1 (x1 ), . . . , fN (xN )]> ∈ RN , w = [w1 , . . . , wN ]> ∈ RN ,
Am = diag{Am1 , . . . , AmN }, and Bm = diag{Bm1 , . . . , BmN }. x0 is the
leader’s state and it represents the desired synchronization trajectory
according to the following equation
Also, one should note that B , 0 for a strongly connected graph
with
||β|| ≤ ||e||/λ (L + B)
(8)
where λ (L + B) is the minimum singular value of (L + B)
F. L. Lewis et al. (2013).
A performance function ρ (t) is associated with the error component e (t) and is defined as a smooth function such as ρ (t) : R+ →
R+ is a positive decreasing function lim ρ (t) = ρ∞ > 0. The pret→∞
scribed performance function can be written as
ρi (t) = (ρi0 − ρi∞ ) exp−`i t +ρi∞
(9)
where ρi0 , ρi∞ and `i are appropriately defined positive constants.
In order to overcome the difficulty caused through the synchronization algorithm and achieve the desired prescribed performance, the
following time varying constraints are proposed:
(3)
− δi ρi (t) < ei (t) < ρi (t) ,
i f ei (t) > 0
(10)
where x0 ∈ R is the leader state node, f (x0 , t) ∈ R is the nonlinear part of the leader node’s dynamic. The model presented in
(2)-(3) is very similar to the one treated in Das & Lewis (2010),
with the exception that the dynamic of the agent as described in (2)
is more representative of real systems. The local synchronization
error function for agent i can be described as in X. Li et al. (2004)Khoo et al. (2009).
X
ei =
ai j (xi − x j ) + bi (xi − x0 )
(4)
− ρi (t) < ei (t) < δi ρi (t) ,
i f ei (t) < 0
(11)
ẋ0 = Am x0 + f (x0 , t)
j∈Ni
for all t ≥ 0 and 0 < δi ≤ 1, and i = 1, ..., N.
Remark 3 The dynamic constraints (10) and (11) represent a modification of the ones in Bechlioulis & Rovithakis (2008), and Mohamed (2014). In these papers, the constraints are conditioned on
e(0) as follows
− δρ (t) < e (t) < ρ (t) ,
i f e(0) > 0
(12)
− ρ (t) < e (t) < δρ (t) ,
i f e(0) < 0
(13)
where ai j ≥ 0 and aii = 0. ai j > 0 if agent i is directed to agent
j, bi ≥ 0. The network is such that bi > 0 for at least one agent i.
Hence, equation (4) can be written in the global error form as
Figure (1) shows the tracking error of controller with prescribed
performance as it transits from a large to a smaller set in accordance with equations (10) and (11).
e = − (L + B) (1x0 − x) = (L + B) (x − 1x0 )
Due to the interaction between agents’ dynamics, such constraints
will lead to instability. Upon crossing this reference, the system becomes unstable under the original formulation (12) and (13). However, the switching based on ei (t) provides the necessary control to
keep the system stable.
= (L + B) ( x̃)
(5)
where the global error is e = [e1 , . . . , eN ]> ∈ RN , global state vector
is 1x0 = x0 ∈ RN , the Laplacian matrix is L ∈ RN×N , B ∈ RN×N with
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
Figure 1. Graphical representation of tracking error with prescribed performance (a) Prescribed performance of (12); (b)
Prescribed performance of (13).
Remark 4 In the control with prescribed performance framework
as presented as in Bechlioulis & Rovithakis (2008); Mohamed
(2014), the knowledge of the sign of ei (0) is sufficient to maintain
the same robust controller for all t > 0 and satisfy the performance
constraints (no switching occurs after t = 0). However, in the case
of multi-agent systems, the synchronization error (7) creates a coupling between the different states of each agent. The interactions
created may force the synchronization error to violate the desired
performance constraints and exit the compact set if keep the same
controller based solely on the sign ei (0), i = 1, . . . , N. Switching
is rather needed at any time t to keep the error within the compact
sets.
One should notice from Figure (1) that the tracking error in the
case of multi-agent systems may exceed the lower (or upper) bound
(in green color). Upon crossing this constraint, the system becomes
unstable under the control based on ei (0). However, the switching
based on ei (t) provides the necessary control effort to keep the system stable.
In order to transform the constrained error of the nonlinear system (10) and (11) to an unconstrained one, a transformed error i is
defined as
!
ei (t)
i = ψ
(14)
ρi (t)
or equivalently,
ei (t) = ρi (t) S (i )
(15)
where i , S i (.) and ψ−1
i (.) are all smooth functions, i = 1, 2, . . . , N.
S (.) = ψ−1 (.) and S (.) satisfy the following properties:
Property 1
1. S i (i ) is smooth and strictly increasing.
2. −δi < S (i ) < δ̄i ,
−δ̄i < S (i ) < δi ,
3.
i f ei (t) ≥ 0
i f ei (t) < 0
limi →−∞ S (i ) = −δi
limi →+∞ S (i ) = δ̄i ,
limi →−∞ S (i ) = −δ̄i
lim
S ( ) = δ ,
i →+∞
i
i
if
ei (t) ≥ 0
if
ei (t) < 0
where
δ̄i expi −δi exp−i
, δ̄i > δi if ei (t) ≥ 0
expi + exp−i
S (i ) =
δi expi −δ̄i exp−i
, δi > δ̄i if ei (t) < 0
expi + exp−i
Now, consider the general form of the smooth function
S (i ) =
δ̄i expi −δi exp−i
expi + exp−i
and the transformed error
!
ei (t)
i =S −1
ρi (t)
δ + ei (t) /ρi (t)
ln i
,
1
δ̄i − ei (t) /ρi (t)
=
δi + ei (t) /ρi (t)
2
,
ln
δ̄i − ei (t) /ρi (t)
(16)
(17)
with δ̄i > δi if ei (t) ≥ 0 (18)
with
δi > δ̄i if ei (t) < 0
Remark 5 In the previous set of equations, δi and δ̄i exchange
values depending on the sign of ei (t). One should note that the
highest value of both involves subtracting the absolute value of
ei (t) /ρi (t) and the lowest includes the addition of the absolute
value of ei (t) /ρi (t).
Using this remark and the fact that ρi (t) > 0, equation (18) can be
rewritten as
δ + |ei (t) |/ρi (t)
ln i
,
if ei (t) ≥ 0
1 δ̄i − |ei (t) |/ρi (t)
with δ̄i > δi
i =
δi + |ei (t) |/ρi (t)
2
(t)
,
if
e
<
0
−
ln
i
δ̄i − |ei (t) |/ρi (t)
(19)
Thus, the transformed error can be expressed in more compact
form as follows:
!
δ + |ei (t) |/ρi (t)
1
i = sign (ei (t) /ρi (t)) · ln i
,with
δ̄i > δi
2
δ̄i − |ei (t) |/ρi (t)
(20)
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
And to attenuate the effect of chattering, the following form of the
transformed error will be considered
!
!
δ + |ei (t) |/ρi (t)
1
ξei (t)
i = √ er f
· ln i
,with
δ̄i > δi
ρi (t)
δ̄i − |ei (t) |/ρi (t)
2 π
(21)
Rξe 2
2
−a
where er f (ξe/ρ) = √π e da. ξ > 0 is a design parameter.
0
Remark 6 The primary role of ξ is to make the er f (ξe) as close as
possible to sign (e). Ideally, ξ is selected to be as big as possible.
For instance, |er f (ξe)| u 1 when |e| > ∆ = 2ξ . Therefore, if ξ = 200
then |er f (e) | u 1 when |e| > 0.01. However, while the derivative
is smooth the more one selects a big ξ the more there is a risk of
chattering.
For simplification, let x (t) = x, e (t) = e, (t) = and ρ (t) = ρ.
After algebraic manipulations, the derivative of transformed error
when |e|/ρ ≥ ∆/ξ can be approximated by:
!
!
1
1
ei ρ̇i
1
˙i =
ėi −
(22)
+
2ρi δi + |ei |/ρi δ̄i − |ei |/ρi
ρi
Remark 7 As mentioned earlier, the selection of the high gain ξ
can make the absolute value of the error function converge to 1 for
a small value of the ratio |e (t) |/ρ = ∆/ρ. In our analysis, we will
use (22) to show that the control will generate a UUB error dynamic
that will converge to a ball around zero with a radius that can be
made as small as desired depending on the selection of ξ. Thus, the
error may not converge to zero.
Let
ri =
1
1
1
+
2ρi δi + |ei |/ρi δ̄i − |ei |/ρi
!
(23)
From (6) and (22), the global synchronization of the transformed
error can be obtained as
˙ = R (L + B) (Am x + f (x) + Bm u + w − f (x0 , t)) − Υ̇ Υ−1 e (t)
(24)
where the control at the level of each node is of the form ui =
−ci + ν; the value of ν represents the part of the control action necessary to tackle the uncertainties and takes into account the estimation errors in the adaptation rule (see 35); = [1 , . . . , N ]> ∈ RN ,
Υ = diag{ρi (t)} and Υ̇ = diag{ρ̇i (t)}, i = 1, . . . , N; R is such that
R = diag[r1 (t) , . . . , rN (t)] with R > 0 and Ṙ < 0; Υ̇ Υ−1 < 0
with limt→∞ Υ̇ Υ−1 = 0. Before proceeding further, the following
definitions are needed (see Das & Lewis (2010)).
Definition 1 The global error e (t) ∈ RN is uniformly ultimately
bounded (UUB) if there exists a compact set Ω ⊂ RN so that
∀e(t0 ) ∈ Ω there exists a bound B and a time t f (B, e(t0 )), both independent of t0 ≥ 0, such that ||e (t) || ≤ B so that ∀t > t0 + t f .
4
Adaptive Projection Approximation
Using linear parametrization of nonlinear systems (for more details see A.8 in Hovakimyan & Cao (2010) )The agent i’s nonlinear
dynamics in (1) can be written as
ẋi = Ami xi + Bmi ui + θi ||xi ||∞ + σi (xi , t)
(25)
with θi ∈ R is an unknown but bounded time varying parameter
and σi ∈ R is the part that includes all unknown nonlinearities and
external disturbances, w. Θi and ∆i are known compact sets where
θi ∈ Θi and σi ∈ ∆i . In the remaining of the paper, the following
assumptions will be considered.
Assumption 1 Hovakimyan & Cao (2010)
1. Leader’s states are bounded by ||x0 || ≤ X0 .
2. Leader’s nonlinear dynamic is unknown and bounded such
as || f (x0 , t)|| ≤ F M .
0
3. Uniform boundedness of the unknown parameters: ||θ (t) || ≤
θ M and ||σ (t) || ≤ σ M for all t > 0
4. Uniform boundedness of the rate of variation of parameters:
θ (t) and σ (t) are continuously differentiable with uniformly
bounded derivatives. ||θ̇ (t) || ≤ dθ < ∞ and ||σ̇ (t) || ≤ dσ <
∞ for all t ≥ 0.
One should note that the values of the estimation bounds are not
necessary known.
Assumption 2 Matrix Ami , Bmi are known and B−1
mi exists.
Let θ̂ and σ̂ be the approximation of θ and σ respectively. Then,
θ̃i = θi − θ̂i
(26)
σ̃i = σi − σ̂i
(27)
Remark 8 The communication graph is considered strongly connected. Thus, if bi , 0 for at least one i, i = 1, . . . , N then (L + B)
is an irreducible diagonally dominant M-matrix and hence nonsingular Das & Lewis (2010). The control signal of local agent i can
be given by
Lemma 1 (see F. L. Lewis et al. (2013) for more details.) Let L
be an irreducible matrix and B , 0 such as (L + B) is nonsingular,
then we can define
q = [q1 , . . . , qN ]> = (L + B)−1 · 1
(28)
P = diag{pi } = diag{1/qi }
(29)
Then, P > 0 and the matrix Q defined as
Q = P (L + B) + (L + B)> P = P S (L + B) + (L + B)> S P (30)
is also positive definite with S = P−1
Definition 2 The control node trajectory x0 (t) given by (1) is cooperative UUB with respect to solutions of node dynamics (3) if there
exists a compact set Ω ⊂ RN so that ∀(xi (t0) − x0 (t0)) ∈ Ω, there
exist a bound B and a time t f (B, (x(t0 )x0 (t0 ))), both independent of
t0 ≥ 0, such that ||x(t0 )x0 (t0 )|| ≤ B, ∀i, ∀t > t0 + t f .
The gist of the idea is that Q = S (L + B) + (L + B)> S is diagonally
strictly dominant, and since it is a symmetric M-matrix, then it is
positive definite. Based on this lemma, the following Preposition
holds
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
Proposition 1 Let R a positive definite diagonal matrix, and L, B,
P and S as defined in Lemma 1, then the matrix Q defined as
Q = P R (L + B) + (L + B)> R P
(31)
is positive definite.
Proof:
Since (L + B) is is a nonsingular M-matrix and R > 0 is diagonal,
then R (L + B) is a non-singular M-Matrix.
(L + B) q = 1 > 0
Let P1 = P R = R P and Q = P1 (L + B) + (L + B)> P1 . Using
˙ respectively, one
equations (36)-(39) and (41) to replace ˙ , θ̃˙ and σ̃
can write
1
V̇ = − c > Q − > P1 (L + B) f (x0 , t)) − kθ̃> θ̃ − kσ̃> σ̃
2
+ kσ̃> σ + kθ̃> θ + > P1 Aθ̃ kxk∞ + > P1 Aσ̃
(44)
− > P1 Υ̇ Υ−1 e (t) + θ̃> Γ−1 θ̇ + σ̃> Γ−1 σ̇
On the other hand,
(32)
Let S = diag{qi } then
e (t) = Υ S ()
Therefore
R (L + B) S 1 = R (L + B) q = R1 > 0
(33)
which means strict diagonal dominance of R (L + B) S .
Q = P R (L + B)+(L + B)> R P = P R (L + B) S + S (L + B)> R P
(34)
R (L + B) S + S (L + B)> R is symmetric and strictly diagonally
dominant. Therefore, Q is positive definite.
The control signal of local nodes is given by
ui = B−1
mi −ci − Ami x̃i − θ̂i ||xi ||∞ − σ̂i
where the control gain c > 0 and the overall control signal
u = B−1
m −c − Am x̃ − θ̂ kxk∞ − σ̂
(35)
(45)
− P Υ̇ S () + θ̃ Γ θ̇ + σ̃ Γ σ̇
>
>
−1
>
−1
one should note that Λ (t) = −P Υ̇ is a positive definite diagonal
lim
matrix for ∀ t and t −−→ ∞, Λ (t) = 0
1
V̇ = − c > Q − > P1 (L + B) f (x0 , t)) − kθ̃> θ̃ − kσ̃> σ̃
2
+ kσ̃> σ + kθ̃> θ + > P1 Aθ̃ kxk∞ + > P1 Aσ̃ + > Λ δ̄
(46)
+ θ̃ Γ θ̇ + σ̃ Γ σ̇
>
(36)
with kxk∞ = [||x1 ||∞ , . . . , ||xN ||∞ ] . Let the adaptive estimates of θ̂
and σ̂ updated according to
>
θ̂˙i = Γi xi iT pi ri (di + bi ) − kΓi θ̂i
(37)
>
>
˙ i = Γi iT pi ri (di + bi ) − kΓi σ̂i
σ̂
1
V̇ = − c > Q − > P1 (L + B) f (x0 , t)) − kθ̃> θ̃ − kσ̃> σ̃
2
+ kσ̃> σ + kθ̃> θ + > P1 Aθ̃ kxk∞ + > P1 Aσ̃
(38)
−1
>
−1
δ̄ = max{δ̄1 , . . . , δ̄N }.
1
2
V̇ ≤ − cλ (Q) kk2 + λ̄ (P1 ) λ̄ (L + B) F M kk − k θ̃ − k kσ̃k2
2
+ λ̄ (P1 ) λ̄(A)x M θ̃ kk + λ̄ (P1 ) λ̄(A) kk kσ̃k + k kσ̃k σ M
+ k θ̃ θ M + δ̄λ̄ (Λ) kk + λ̄ Γ−1 dθ θ̃ + λ̄ Γ−1 dσ kσ̃k
(47)
+
with Γi ∈ R and k > 0. c and k are scalar design parameters.
then
Theorem 1 Consider the strong connected digraph of the network
in (1) with adaptive estimates in (37) and (38) satisfying Assumptions (1 and under the control law , then the distributed multi-agent
system is UUB stable if the tuning gain k and c satisfy the following
conditions
cλ (Q)
(39)
k=
2
and
1
cλ (Q) > (xm + 1)λ̄(P)λ̄(A)
(40)
2
with P and Q are defined in Lemma 1.
Proof: using (36), equation (6) becomes
ė = (L + B) (−c + θ̃ kxk∞ + σ̃ − f (x0 , t))
1 >
1
1
P + θ̃> Γ−1 θ̃ + σ̃> Γ−1 σ̃
2
2
2
(41)
(42)
with P > 0 as defined in (29), γ ∈ R+ was mentioned in (37) and
Γ = diag{γi }. The derivative of (42) is
˙
V̇ = > P ˙ + θ̃> Γ−1 θ̃˙ + σ̃> Γ−1 σ̃
θ̃
1
i
2 cλ (Q)
kσ̃k − 12 λ̄ (P1 ) λ̄(A)x M
1
− 2 λ̄ (P1 ) λ̄(A)
h
+ λ̄ (P1 ) λ̄ (L + B) F M + δ̄λ̄ (Λ)
h
Define z = kk
θ̃
− 12 λ̄ (P1 ) λ̄(A)x M
k
0
kθ M + λ̄ Γ−1 dθ
− 21 λ̄ (P1 ) λ̄(A) kk
0
θ̃
kσ̃k
k
i kk
kσ M + λ̄ Γ−1 dσ θ̃
kσ̃k
(48)
i>
kσ̃k , then equation (48) can be written as
V̇ ≤ − z> Hz + h> z
(49)
and V̇ ≤ 0 if and only if H is positive definite and
Consider the following Lyapunov candidate function
V=
h
V̇ ≤ − kk
(43)
kzk >
khk
λ (H)
(50)
The Lyapunov candidate in (42) can be described by
−1
λ
Γ
λ
Γ−1
1
1
2
λ(P) kk2 +
θ̃ +
kσ̃k2 ≤ V ≤ λ̄(P) kk2
2
2
2
2
(51)
−1
λ̄ Γ−1
λ̄
Γ
2
2
+
θ̃ +
kσ̃k
2
2
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
or
1h
kk
2
1h
kk
2
λ(P)
i
kσ̃k 0
0
λ̄(P)
i
kσ̃k 0
0
θ̃
θ̃
Let
0 kk
0
θ̃ ≤ V ≤
λ Γ−1 kσ̃k
(52)
0 kk
0
θ̃
λ̄ Γ−1 kσ̃k
0
λ Γ−1
0
0
λ̄ Γ−1
0
λ(P)
Πmin = 0
0
λ̄(P)
Πmax = 0
0
0
λ Γ−1
0
0
λ̄ Γ−1
0
Also from (53), we have
s
kzk ≤
s
2V
,
λ(S )
kzk ≥
2V
λ̄(S̄ )
Then, equation (49) can be written as
√
V̇ ≤ −Hα V + hα V
with Hα =
2λ(H)
λ̄(Π̄)
and hα = √2khk which equivalent to
λ(Π)
√
V ≤ exp−Hα t/2
p
V(0) −
!
hα
hα
+
Hα
Hα
That can be written as
p
p
√
hα
V ≤ V(0) ≤ V(0) +
Hα
(52) is equivalent to
1
1
λ(Πmin ) kzk2 ≤ V ≤ λ̄(Πmax ) kzk2
2
2
(53)
(61)
√
√
√
2d V
≤ −Hα V + hα
dt
0
0
λ Γ−1
0
0
λ̄ Γ−1
(60)
(62)
(63)
(64)
Finally, the algorithm of nonlinear single node dynamics such as
equation (1) can be summarized briefly as
1. Define the system known parameters Ami , Bmi .
then
V>
1
khk2
λ(Πmin )
2
λ (H)2
(54)
2. Define the control design parameters such as Γi , pi , di , bi , k and
c.
3. Evaluate local error synchronization from equation (4).
Defining c =
2
,
λ(Q)
γ1 = 12 λ̄ (P1 ) λ̄(A)x M , γ2 = 21 λ̄ (P1 ) λ̄(A)
6. Evaluate transformed error from equation (21).
kθ M + λ̄ Γ−1 dθ
Substitute (48) to (55), we have
−γ1
c
k
H = −γ1
−γ2
0
cλ (Q) −
i
kσ M + λ̄ Γ−1 dσ kzk
(55)
−γ2
0
k
1
cλ (Q)
2
and
+ 1)λ̄(P)λ̄(A)
2
kzk >
λ̄ (P1 ) λ̄ (L + B) F M + δ̄λ̄ (Λ) λ̄(ρ̇) + kθ M + λ̄ Γ
−1
dθ + kσ M + λ̄ Γ
−1
dσ
λ (H)
which implies
θ̃ >
λ̄ (P1 ) λ̄ (L + B) F M + δ̄λ̄ (Λ) + kθ M + λ̄ Γ−1 dθ + kσ M + λ̄ Γ−1 dσ
λ (H)
λ̄ (P1 ) λ̄ (L + B) F M + δ̄λ̄ (Λ) + kθ M + λ̄ Γ−1 dθ + kσ M + λ̄ Γ−1 dσ
λ (H)
λ̄ (P1 ) λ̄ (L + B) F M + δ̄λ̄ (Λ) + kθ M + λ̄ Γ−1 dθ + kσ M + λ̄ Γ−1 dσ
λ (H)
(66)
Examples and Simulations
Example 1: Consider the digraph composed of five nodes
strongly connected and having a single leader connected to node
3. The pining gains between connected nodes are assumed equal to
1 as in Figure 2. The nonlinear dynamics of the different agents are
as follows
ẋ1 = x13 + u1 + a1 cos (t)
ẋ2 = x22 + u2 + a2 cos (t)
(57)
ẋ3 = x34 + u3 + a3 cos (t)
(58)
ẋ5 = x55 + u5 + a5 cos (t)
(59)
ai , i = 1, . . . , 5 are bounded randomly generated constant amplitudes. The leader dynamics was selected ẋ0 = f0 (x0 , t) = 0 with
ẋ4 = x4 + u4 + a4 cos (t)
kσ̃k >
Remark 9 If we have xi ∈ Rn , n > 1, then ui ∈ Rn , fi (xi ) ∈ Rn ,
θi ∈ Rn , σi ∈ Rn , ri = diag{ri1 , . . . , rin } ∈ Rn×n , then the problem
can be extended easily and the estimated weight will be written as
>
(65)
θ̂˙i = Γi ||xi ||∞ iT (pi In )ri ((di + bi )In ) − kΓi θ̂i
5
(56)
kk >
8. Evaluate adaptive estimates from equations (37) and (38).
>
˙ i = Γi iT (pi In )ri ((di + bi )In ) − kΓi σ̂i
σ̂
1
(x
2 m
And from 50, taking || · ||1 of h
7. Evaluate control signal from equation (35).
9. Go to step 2.
where H is positive definite matrix. If we select k =
cλ (Q) > 12 (x M + 1)λ̄(P)λ̄(A), then, we will have
λ (H) =
4. Evaluate the prescribed performance function from equation (9).
5. Evaluate ri from equation (23).
−γ1 −γ2
c
k
0 kzk
V̇ ≤ − kzk> −γ1
−γ2
0
k
h
+ λ̄ (P1 ) λ̄ (L + B) F M + δ̄λ̄ (Λ)
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
1
Agents
5
4
2
3
Leader
0
Figure 2. Strongly connected graph of one leader and five
agents.
desired consensus value equal to 2. Nonlinearities and disturbances
are assumed to be unknown in all nodes. The control parameters
of the system are ρ∞ = 0.05 × 11×5 , ρ0 = 7 × 11×5 , l = 7 × 11×5 ,
Γ = 150I5×5 , δ̄ = 7, δ = 1, c = 100, k = 0.8, α = 20 and x0 = 2,
x(0) = [0.8230, −0.9001, −2.5351, −1.4567, −0.7553]> .
Figures 3 and 4 show the output performance, control signal and
transformed error respectively for the proposed control algorithm
using (18) for the transformed error. Figure 3 shows the severe
chattering in the control effort. Although the oscillation of synchronization error values satisfy the prescribed performance conditions,
the switching in error signs caused switching in transformed errors
as shown in Figure 4 which consequently causes chattering in the
control signal as clearly revealed in Figure 3.
The proposed control with the new prescribed performance
function as in (21) Figure 5 and 6 show the output performance,
control signal and transformed error respectively. A significant
improvement in the control effort and transformed error can be
clearly observed.
Example 2: (MIMO case) Consider the same problem as in Figure 2 with 3 inputs and 3 outputs nonlinear systems. The nonlinear
dynamics of the graph are now
ẋ j = Ax j + Bu j + θ j x j + f j (x j ) + D j (t)
h
i
θ j = θ1j θ2j θ3j ,
2c1, j sin(0.4c1, j t)cos(0.3t)
3c1, j sin(0.5t)
2
1
θ j = 0.9sin(0.2c2, j t) , θ j = 2.5sin(0.3c2, j t) + 0.3cos (t) ,
0.6c3, j cos(0.15t)
0.5sin(0.13c3, j t)
0.7sin(0.2c1, j t)
3
1.0sin(0.1c2, j t)
θ j =
,
1.5cos(0.7c3, j t) + 1.6c3, j sin(0.3t)
a, b, c are matrices that were selected with different input elements
to introduce heterogeneity into the system and therefore different
control efforts have to be implemented.
1.5 0.5 0.7 1.3 0.7
a = 0.5 1.4 0.1 1.3 2.4 ,
2.8 1.4 0.6 0.7 0.6
0.5 1.5 1.1 1.6 0.3
b = 0.7 1.2 1.3 0.5 0.3 ,
1.1 1.4 1.6 0.6 1.0
1.5 2.5 0.5 1.7 0.7
c = 0.5 1.7 1.1 0.3 0.4 ,
0.8 0.4 2.2 0.9 1.4
The leader’s dynamics is selected such that x0
=
[3cos(0.7t), 2cos(0.8t), 1.5cos (t)]> .
The other parameters of the problem are defined as ρ∞ = 0.05 × 13×5 ,
ρ0 = 7 × 13×5 , l = 7 × 13×5 , Γ = 150I5×5 , δ̄ = 7, δ = 1,
c = 100, k = 0.8, α = 50. Initial conditions of x(0) =
[1.6399, 1.6639, −2.1864, 0.1160, −2.7805, −2.2175, −0.1489, 2.2989,
−1.3038, 0.5571, −0.5959, 1.6760, −2.4743, 0.0488, 0.8288].
The robustness of the proposed controller against time variant
uncertainties in parameters, time-variant disturbances and high
nonlinearities are tested in this example considering the formula
in (21). Figure 7 shows the output performance of the proposed
controller for the MIMO case. The control input in the connected
graph is shown in Figure 8. Errors and transformed errors for the
three outputs are depicted in Figure 9, 10 and 11. Figure 12 shows
the phase plane plot starting from different initial conditions and the
synchronization to the desired trajectory. The results demonstrate
the performance of the proposed robust controller.
y j = Cx j
where x j ∈ R3×1 is the state vector, u j ∈ R3×1 is the input vector,
y j ∈ R3×1 is the output vector. A, BandC ∈ R3×3 are known constant
matrices.
−20 22 0
15 0 , B = I3×3 , C = I3×3 ,
A = 0
0
0 −3
f j (x j ) ∈ R3×1 is the system nonlinear vector, D j (t) ∈ R3×1 is the
system disturbance vector, θ j ∈ R3×1 . Each of f j (x j ), D j (t) , θ j are
assumed to be completely unknown.
a1, j x3, j x1, j + 0.2sin(x1, j a1, j )
f j (x j ) = −a2, j x1, j x3, j − 0.2a2, j cos(a2, j x3, j t)x1, j ,
a3, j x1, j x2, j
1 + b1, j sin(b1, j t)
1.2cos(b2, j t)
D j (t) =
,
sin(0.5b3, j t) + cos(b3, j t) − 1
6
Conclusion
In this paper, a distributed adaptive tracking control of nonlinear
uncertain multi-agent systems with prescribed performance is proposed. Under such controller, the tracking error is confined from
within a predefined large set to a smaller set according to a given
performance. Agents’ dynamics were assumed unknown. The control law is fully distributed based on the fact that the control of each
agent respects the strongly connected graph’s topology and includes
only the allowed local neighborhood information. The proposed
approach guarantees uniform ultimate boundedness for the transformed error. Simulations include two examples to validate the robustness and smoothness of the proposed controller against highly
nonlinear heterogeneous multi-agent system with time-variant uncertain parameters and external disturbances. In future work, control of multi-agents with networks that are weakly connected or
have variable topology will be studied. Under such controllers L
and B could time varying and an additional but practical challenges.
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
2
yi
1
0
Agent1
Leader
−1
0
0.1
0.2
0.3
0.1
0.2
0.3
Agent2
0.4
Agent3
Agent4
0.5
Agent5
0.6
0.7
200
100
ui
0
−100
−200
Agent1
0
Agent2
0.4
Agent3
Agent4
0.5
Agent5
0.6
0.7
Time(sec)
Figure 3. Output performance and control signal using (18) for the transformed error.
ε1
ρ1
e1
5
0
−5
0
5
δρ1
ε2 & e2
ε1 & e1
10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0
−5
−10
0
0
−5
−10
0
0.1
0.2
0.3
10
ε4 & e4
ε3 & e3
5
ε2
ε3
0.1
0.2
0.3
ρ3
e3
0.4
0.5
δρ3
0.6
0.7
ρ2
e2
0.4
ε4
0.5
δρ2
0.6
ρ4
e4
0.7
δρ4
5
0
−5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Time(sec)
ε5 & e5
5
0
−5
−10
0
ε5
0.1
0.2
0.3
ρ5
e5
0.4
0.5
δρ5
0.6
0.7
Time(sec)
Figure 4. Error and Transformed Error using (18) for the transformed error.
Systems subjects to actuator failure, saturation or hysteresis will
benefit from prescribed performance control and represent interesting areas of further development. Implementation of such control
approach on a real system is an area needing research effort.
Acknowledgment
The authors would like to acknowledge the support of King
Fahd University of Petroleum and Minerals under project Number
IN141048.
7
References
Bechlioulis, C. P., Dimarogonas, D. V., & Kyriakopoulos,
K. J. (2014). Robust control of large vehicular platoons
with prescribed transient and steady state performance. In
Decision and control (cdc), 2014 ieee 53rd annual conference on (pp. 3689–3694).
1
Bechlioulis, C. P., & Rovithakis, G. A. (2008). Robust
adaptive control of feedback linearizable MIMO nonlinear
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
2
y1
1
0
−1
−2
−3
0
Agent1
Leader
0.1
0.2
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
Time(sec)
Agent2
0.6
Agent3
0.7
Agent4
0.8
Agent5
0.9
1
100
50
u1
0
−50
−100
−150
0
Agent1
Agent2
0.6
Agent3
0.7
Agent4
0.8
Agent5
0.9
1
0.2
0.4
2
0
−2
−4
−6
0
ε5 & e5
ε1
6
4
2
0
−2
0
e1
ρ1(t)
δρ1
ε2 & e2
6
4
2
0
−2
0
ε3
0.6
0.8
e3
ρ3(t)
0.2
0.4
0.6
0.8
ε5
0.2
e5
0.4
0.6
Time(sec)
ρ(t)
0.8
2
0
−2
−4
−6
1
0
1
6
4
2
0
−2
0
δρ3
ε4 & e4
ε3 & e3
ε1 & e1
Figure 5. Output performance and control signal with the new prescribed performance function as in (21).
ε2
0.2
e2
ρ2(t)
0.6
0.8
e4
ρ4(t)
0.4
0.6
Time(sec)
0.8
0.4
ε4
0.2
δρ2
1
δρ4
1
δρ5
1
Figure 6. Error and transformed error with the new prescribed performance function as in (21).
systems with prescribed performance. Automatic Control,
IEEE Transactions on, 53(9), 2090–2099.
1, 3, 4
low-complexity global approximation-free control scheme
with prescribed performance for unknown pure feedback
systems. Automatica, 50(4), 1217–1226.
1
Bechlioulis, C. P., & Rovithakis, G. A. (2009). Adaptive
control with guaranteed transient and steady state tracking error bounds for strict feedback systems. Automatica,
45(2), 532–538.
1
Cao, Y., & Ren, W. (2012). Distributed coordinated tracking
with reduced interaction via a variable structure approach.
Automatic Control, IEEE Transactions on, 57(1), 33–48.
1
Bechlioulis, C. P., & Rovithakis, G. A.
Chopra, N., & Spong, M. W. (2006). Passivity-based control
(2014).
A
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
yi1
2
0
−2
−4
0
Leader
1
2
3
4
Agent1
5
6
Leader
Agent1
5
6
Leader
Agent1
5
6
Agent2
7
Agent3
Agent4
8
9
Agent3
Agent4
8
9
Agent3
Agent4
8
9
Agent3
Agent4
8
9
Agent3
Agent4
8
9
Agent3
Agent4
8
9
Agent5
10
yi2
2
0
−2
0
1
2
3
4
Agent2
7
Agent5
10
yi3
1
0
−1
−2
0
1
2
3
4
Agent2
7
Agent5
10
Time(sec)
Figure 7. Output performance in the MIMO case using (21).
ui1
100
0
−100
−200
0
Agent1
1
2
3
4
5
1
2
3
4
5
1
2
3
4
5
6
Agent2
7
Agent5
10
ui2
100
0
−100
0
Agent1
6
Agent2
7
Agent5
10
100
ui3
0
−100
−200
−300
0
Agent1
6
Agent2
7
Agent5
10
Time(sec)
Figure 8. Control Signal in the MIMO case using (21).
of multi-agent systems. In Advances in robot control (pp.
107–134). Springer.
1
Das, A., & Lewis, F. L. (2010). Distributed adaptive control for synchronization of unknown nonlinear networked
systems. Automatica, 46(12), 2014–2021.
1, 3, 3, 8
El-Ferik, S., Qureshi, A., & Lewis, F. L. (2014). Neuroadaptive cooperative tracking control of unknown higher-
order affine nonlinear systems. Automatica, 50(3), 798–
808.
1
Fax, J. A., & Murray, R. M. (2004). Information flow and
cooperative control of vehicle formations. Automatic Control, IEEE Transactions on, 49(9), 1465–1476.
1
Hovakimyan, N., & Cao, C. (2010). L1 adaptive control the-
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
ε11 & e11
e11
ρ11(t)
δρ11
4
2
0
−2
0
2
ε21 & e21
ε11
6
0
−2
−4
ε21
−6
2
4
6
8
10
0
2
−2
−4
ε31
−6
0
ε41
6
0
ε41 & e41
ε31 & e31
2
2
e31
4
6
ρ31(t)
8
δρ31
ρ21(t)
6
8
e41
ρ41(t)
6
8
e22
ρ22(t)
6
8
e42
ρ42(t)
6
8
δρ21
10
δρ41
4
2
0
−2
0
10
e21
4
2
4
10
Time(sec)
ε51 & e51
2
0
−2
−4
ε51
−6
0
2
e51
ρ51(t)
6
8
4
δρ51
10
Time(sec)
Figure 9. Error and Transformed Error in the MIMO case using (21) for x1, j where j = 1, . . . , 5.
δρ12
2
0
0
−2
−4
ε22
−6
2
4
6
ε32
6
e32
8
ρ32(t)
10
δρ32
4
2
0
−2
0
2
0
2
4
4
6
8
10
2
4
10
0
−2
−4
ε42
−6
2
δρ22
2
ε42 & e42
ε12 & e12
ρ12(t)
4
−2
0
ε32 & e32
e12
ε22 & e22
ε12
6
0
δρ42
10
Time(sec)
ε52
ε52 & e52
6
e52
ρ52(t)
6
8
δρ52
4
2
0
−2
0
2
4
10
Time(sec)
Figure 10. Error and Transformed Error in the MIMO case using (21) for x2, j where j = 1, . . . , 5.
ory: Guaranteed robustness with fast adaptation (Vol. 21).
Siam.
4, 1
mal and adaptive design approaches. Springer Science &
Business Media.
2, 3, 2, 3, 1
Khoo, S., Xie, L., & Man, Z. (2009). Robust finite-time consensus tracking algorithm for multirobot systems. Mechatronics, IEEE/ASME Transactions on, 14(2), 219–228.
3
Lewis, F. W., Jagannathan, S., & Yesildirak, A. (1998). Neural network control of robot manipulators and non-linear
systems. CRC Press.
1
Lewis, F. L., Zhang, H., Hengster-Movric, K., & Das, A.
(2013). Cooperative control of multi-agent systems: opti-
Li, X., Wang, X., & Chen, G. (2004). Pinning a complex dynamical network to its equilibrium. Circuits and Systems
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
2
0
ε23 & e23
ε13 & e13
2
−2
−4
ε13
−6
0
2
e13
ρ13(t)
6
8
4
δρ13
0
−2
−4
ε23
−6
10
0
2
ε43
6
ε43 & e43
ε33 & e33
2
0
−2
−4
ε32
−6
0
2
e32
ρ32(t)
6
8
4
δρ32
ρ23(t)
6
8
e43
ρ43(t)
6
8
δρ23
10
δρ43
4
2
0
−2
0
10
e23
4
2
4
10
Time(sec)
ε53 & e53
2
0
−2
−4
ε53
−6
0
2
e53
ρ53(t)
6
8
4
δρ53
10
Time(sec)
Figure 11. Error and Transformed Error in the MIMO case using (21) for x3, j where j = 1, . . . , 5
2
1
Leader
Agent1
0
x3
Agent2
Agent3
−1
Agent4
Agent5
−2
o START
−3
3
∆ END
2
1
0
x2
−1
−2
−3
−3
−2
−1
1
0
2
3
x1
Figure 12. Phase Plan for motion synchronization in the MIMO system case under different initial conditions.
I: Regular Papers, IEEE Transactions on, 51(10), 2074–
2087.
3
Li, Y., & Tong, S. (2015). Prescribed performance adaptive
fuzzy output-feedback dynamic surface control for nonlinear large-scale systems with time delays. Information
Sciences, 292, 125–142.
1
Mohamed, H. A. H. (2014). Improved robust adaptive con-
trol of high-order nonlinear systems with guaranteed performance (M.Sc). King Fahd University Of Petroleum &
Minerals.
1, 3, 4
Na, J. (2013). Adaptive prescribed performance control of
nonlinear systems with unknown dead zone. International
Journal of Adaptive Control and Signal Processing, 27(5),
426–446.
1
To cite this article: Hashim A. Hashim, Sami El-Ferik, and Frank L. Lewis. "Adaptive synchronisation of unknown
nonlinear networked systems with prescribed performance." International Journal of Systems Science 48, no. 4
(2017): 885-898.
Na, J., Chen, Q., Ren, X., & Guo, Y. (2014). Adaptive prescribed performance motion control of servo mechanisms
with friction compensation. Industrial Electronics, IEEE
Transactions on, 61(1), 486–494.
1
Olfati-Saber, R., Fax, J. A., & Murray, R. M. (2007). Consensus and cooperation in networked multi-agent systems.
Proceedings of the IEEE, 95(1), 215–233.
1
Qu, Z. (2009). Cooperative control of dynamical systems:
applications to autonomous vehicles. Springer Science &
Business Media.
1
Ren, W., & Beard, R. W. (2008). Distributed consensus in
multi-vehicle cooperative control. Springer.
2
Ren, W., Beard, R. W., & others. (2005). Consensus seeking
in multiagent systems under dynamically changing interaction topologies. IEEE Transactions on automatic control, 50(5), 655–661.
1
Sun, Y., & Liu, H. (2014). Fuzzy Adaptive Prescribed Performance Control for MIMO Uncertain Chaotic Systems
in Nonstrict Feedback Form. Discrete Dynamics in Nature
and Society, 2014.
1
Theodoridis, D. C., Boutalis, Y. S., & Christodoulou, M. A.
(2012). Direct adaptive neuro-fuzzy trajectory tracking
of uncertain nonlinear systems. International Journal of
Adaptive Control and Signal Processing, 26(7), 660–688.
1
Wang, J., Hovakimyan, N., & Cao, C. (2010). Verifiable
adaptive flight control: unmanned combat aerial vehicle
and aerial refueling. Journal of guidance, control, and
dynamics, 33(1), 75–87.
1
Wang, W., Wang, D., & Peng, Z. (2016). Predictor-based
adaptive dynamic surface control for consensus of uncertain nonlinear systems in strict-feedback form. International Journal of Adaptive Control and Signal Processing.
1
Yang, Y., Ge, C., Wang, H., Li, X., & Hua, C. (2015). Adaptive neural network based prescribed performance control
for teleoperation system under input saturation. Journal of
the Franklin Institute, 352(5), 1850–1866.
1
Yang, Y., Yue, D., & Dou, C. (2016). Distributed adaptive output consensus control of a class of heterogeneous
multi-agent systems under switching directed topologies.
Information Sciences, 345, 294–312.
1
Zhang, H., & Lewis, F. L. (2012). Adaptive cooperative
tracking control of higher-order nonlinear systems with
unknown dynamics. Automatica, 48(7), 1432–1439.
1
Zhao, L., & Jia, Y. (2016). Neural network-based adaptive
consensus tracking control for multi-agent systems under
actuator faults. International Journal of Systems Science,
47(8), 1931–1942.
1
Tong, S., Sui, S., & Li, Y. (2015). Fuzzy adaptive output
feedback control of mimo nonlinear systems with partial
tracking errors constrained. Fuzzy Systems, IEEE Transactions on, 23(4), 729–742.
1
Zhao, Y., Duan, Z. S., Wen, G. H., Li, Z. K., & Chen, G. R.
(2015). Fully distributed tracking control for non-identical
multi-agent systems with matching uncertainty. International Journal of Adaptive Control and Signal Processing, 29(8), 1024–1037. Retrieved from http://dx.doi
.org/10.1002/acs.2520 doi: 10.1002/acs.2520
1
Tong, S., Wang, T., & Li, Y. (2014). Fuzzy adaptive actuator
failure compensation control of uncertain stochastic nonlinear systems with unmodeled dynamics. Fuzzy Systems,
IEEE Transactions on, 22(3), 563–574.
1
Zhou, N., Xia, Y., Fu, M., & Li, Y. (2015). Distributed cooperative control design for finite-time attitude synchronisation of rigid spacecraft. IET Control Theory & Applications.
1
| 3cs.SY
|
Logical Methods in Computer Science
Vol. 6 (4:8) 2010, pp. 1–43
www.lmcs-online.org
Submitted
Published
Nov. 16, 2009
Dec. 18, 2010
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
TAKESHI TSUKADA a AND ATSUSHI IGARASHI b
a
Graduate School of Information Science, Tohoku University
e-mail address: [email protected]
b
Graduate School of Informatics, Kyoto University
e-mail address: [email protected]
Abstract. Taha and Nielsen have developed a multi-stage calculus λα with a sound type
system using the notion of environment classifiers. They are special identifiers, with which
code fragments and variable declarations are annotated, and their scoping mechanism is
used to ensure statically that certain code fragments are closed and safely runnable.
In this paper, we investigate the Curry-Howard isomorphism for environment classifiers by developing a typed λ-calculus λ⊲ . It corresponds to multi-modal logic that allows
quantification by transition variables—a counterpart of classifiers—which range over (possibly empty) sequences of labeled transitions between possible worlds. This interpretation
will reduce the “run” construct—which has a special typing rule in λα —and embedding
of closed code into other code fragments of different stages—which would be only realized by the cross-stage persistence operator in λα —to merely a special case of classifier
application. λ⊲ enjoys not only basic properties including subject reduction, confluence,
and strong normalization but also an important property as a multi-stage calculus: timeordered normalization of full reduction.
Then, we develop a big-step evaluation semantics for an ML-like language based on λ⊲
with its type system and prove that the evaluation of a well-typed λ⊲ program is properly
staged. We also identify a fragment of the language, where erasure evaluation is possible.
Finally, we show that the proof system augmented with a classical axiom is sound and
complete with respect to a Kripke semantics of the logic.
1. Introduction
A number of programming languages and systems that support manipulation of programs as data [1, 2, 3, 4, 5] have been developed in the last two decades. A popular language
abstraction in these languages consists of the Lisp-like quasiquotation mechanism to create and compose code fragments and a function to run them like eval in Lisp. For those
languages and systems, a number of type systems for so-called “multi-stage” calculi have
been studied [5, 6, 7, 8, 9, 10, 11] to guarantee safety of generated programs even before the
generating program runs.
1998 ACM Subject Classification: D.3.3, F.3.3, F.4.1.
Key words and phrases: Curry-Howard correspondence, Environment classifiers, Modal logic, Multi-stage
calculus.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-6 (4:8) 2010
CC
T. Tsukada and A. Igarashi
Creative Commons
2
T. TSUKADA AND A. IGARASHI
Among them, some seminal work on the principled design of type systems for multistage calculi is due to Davies [7] and Davies and Pfenning [12, 8]. They discovered the CurryHoward isomorphism between modal/temporal logics and multi-stage calculi by identifying
(1) modal operators in modal logic with type constructors for code fragments treated as
data and, in the case of temporal logic, (2) the notion of time with computation stages.
For example, the calculus λ [7], which can be thought as a reformulation of Glück and
Jørgensen’s calculus for multi-level generating extensions [6] by using explicit quasiquote
and unquote in the language, corresponds to a fragment of linear-time temporal logic (LTL)
with the temporal operator “next” (written ) [13]. Here, linearly ordered time corresponds
to the level of nesting of quasiquotations, and a modal formula A to the type of code of
type A. It, however, does not treat eval; in fact, the code type in λ represents code
values, whose bodies are open, that is, may have free variables, so simply adding eval to
the calculus does not work—execution may fail by referencing free variables in the code.
The calculus developed by Davies and Pfenning [12, 8], on the other hand, corresponds to
(intuitionistic) modal logic S4 (only with the necessity operator ), in which a formula A
is considered the type of closed code values of type A. It supports safe eval since every code
is closed, but inability to deal with open code hampers generation of efficient code. The
following work by Taha and others [5, 14, 15, 9, 16] sought various forms of combinations
of the two systems above to develop expressive type systems for multi-stage calculi.
Finally, Taha and Nielsen [9] developed a multi-stage calculus λα , which was later modified to make type inference possible [16] and implemented as a basis of MetaOCaml. The
calculus λα has a strong type system while supporting open code, run, (which corresponds
to eval), and a mechanism called cross-stage persistence (CSP), which allows a value to be
embedded in a code fragment evaluated later. They introduced the notion of environment
classifiers (or, simply, classifiers), which are special identifiers with which code fragments
and variable declarations are annotated, to the type system. A key idea is to reduce the
closedness checking of a code fragment (which is useful to guarantee the safety of run)
to the freshness checking of a classifier. Unfortunately, however, correspondence to a logic
is not clear for λα any longer, resulting in somewhat ad-hoc typing rules and complicated
operational semantics, which would be difficult to adapt to different settings.
In this paper, we investigate a Curry-Howard isomorphism for environment classifiers
by developing a typed λ-calculus λ⊲ . As a computational calculus, λ⊲ is equipped with
quasiquotation (annotated with environment classifiers) and abstraction over environment
classifiers just like λα , with application of a classifier abstraction to a possibly empty sequence
of environment classifiers, which makes λ⊲ different from λα . Intuitively, (the type system
of) λ⊲ can be considered a proof system of a multi-modal logic to reason about deterministic
labeled transition systems. Here, modal operators are indexed with transition labels, and
so the logic is multi-modal. One notable feature of the logic is that it has quantification
that allows one to express “for any state transitions,” where a state transition is a possibly
empty sequence of labels. This quantifier corresponds to types for classifier abstractions,
used to ensure freshness of classifiers, which correspond to transition labels (and variables
ranging over their sequences).
A pleasant effect of this logical interpretation—in particular, interpreting environment
classifiers as variables ranging over sequences of transition labels—is that it will reduce the
run construct—which has a peculiar typing rule in λα —and embedding of closed code into
other code fragments of different stages—which would be only realized by the CSP operator
in λα —to merely a special case of classifier application.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
3
Our technical contributions can be summarized as follows:
• Identification of a modal logic that corresponds to (a computational calculus with) environment classifiers;
• Development of a new typed λ-calculus λ⊲ , naturally emerged from the correspondence,
with its syntax, operational (small-step reduction and big-step evaluation) semantics, and
type system;
• Proofs of basic properties, which a multi-stage calculus is expected to enjoy; and
• Proofs of soundness and completeness of the proof system (augmented with a classical
axiom) with respect to a Kripke semantics of the logic.
Our calculus λ⊲ not only enjoys the basic properties such as subject reduction, confluence,
and strong normalization but also time-ordered normalization [7, 10], which says (full)
reduction to a normal form can always be performed according to the order of stages. We
extend λ⊲ with base types and recursion, define a big-step evaluation semantics as a basis
of a multi-stage programming language such as MetaOCaml, and prove the evaluation of a
well-typed program is safe and staged, i.e., if a program of a code type evaluates to a result,
it is a code value whose body is a well-typed program, again. We also develop erasure
semantics, where information on classifiers is (mostly) discarded, and identify a subset of
the language, where the original and erasure semantics agree, by an alternative type system.
It turns out that the subset is rather similar to λi [16], whose type system is used in the
current implementation of MetaOCaml.
One missing feature in λ⊲ is CSP for all types of values but we do not think it is a
big problem. First, CSP for primitive types such as integers is easy to add as a primitive;
CSP for function types is also possible as long as they do not deal with open code, which,
we believe, is usually the case. Second, as mentioned above, embedding closed code into
code fragments of later stages is supported by a different means. It does not seem very
easy to add CSP for open code to λ⊲ , but we think it is rarely needed. For more detail, see
Section 6.3.
We can obtain a natural deduction proof system of a new logic that corresponds to the
calculus λ⊲ just by removing terms from typing rules, as usual. It is also easy to see that
terms and reduction in the calculus correspond to proofs and proof normalization in the
logic, respectively.
Of course, we should answer an important question: “What does this logic really mean?”
We will elaborate the intuitive meaning of formulae in Section 2 and proof rules can be
understood according to this informal interpretation but, to answer this question more precisely, one has to give a semantics and prove the proof system is sound and complete with
respect to the semantics. However, the logic is intuitionistic and it is not straightforward
to give (Kripke) semantics [17]. So, instead of Kripke semantics of the logic directly corresponding to λ⊲ , we give that of a classical version of the proof system, which has a proof
rule for double negation elimination and prove that the proof system is sound and complete
with it in Section 5. Even though the semantics does not really correspond to λ⊲ , it justifies
our informal interpretation.
Organization of the Paper. In Section 2, we review λα and informally describe how
the features of its type system correspond to those of a logic. In Section 3, we define the
multi-stage calculus λ⊲ and prove basic properties including subject reduction, strong normalization, confluence, and time-ordered normalization. In Section 4, we define MiniML⊲ ,
4
T. TSUKADA AND A. IGARASHI
an extension of λ⊲ with base types and recursion, with its big-step semantics and prove that
the big-step semantics implements staged execution. We also investigate erasure semantics
of a subset of MiniML⊲ here. In Section 5, we formally define (a classical version of) the
logic and prove soundness and completeness of the proof system (augmented with a classical
rule) with respect to a Kripke semantics. Lastly, we discuss related work and conclude.
2. Interpreting Environment Classifiers in a Modal Logic
In this section, we informally describe how environment classifiers can be interpreted
in a modal logic. We start with reviewing Davies’ λ [7] to get an intuition of how notions
in a modal logic correspond to those in a multi-stage calculus. Then, along with reviewing
main ideas of environment classifiers, we describe our logic informally and how our calculus
λ⊲ is different from λα by Taha and Nielsen [9].
2.1. λ : Multi-Stage Calculus Based on LTL. Davies has developed the typed multistage calculus λ , which corresponds to a fragment of intuitionistic LTL by the CurryHoward isomorphism. It can be considered the λ-calculus with a Lisp-like quasiquotation
mechanism. We first review linear-time temporal logic and the correspondence between the
logic and the calculus.
In LTL, the truth of propositions may depend on discrete and linearly ordered time, i.e.,
a given time has a unique time that follows it. Some of the standard temporal operators
are
(to mean “next”), (to mean “always”), and U (to mean “until”). The Kripke
semantics of (classical) LTL can be given by taking the set of natural numbers as possible
worlds;1 then, for example, the semantics of
is given by: n
τ if and only if n + 1 τ ,
where n τ is the satisfaction relation, which means “τ is true in world—or, at time—n.”
In addition to the usual Curry-Howard correspondence between propositions and types
and between proofs and terms, Davies has pointed out additional correspondences between
time and computation stages (i.e., levels of nested quotations) and between the temporal
operator
and the type constructor meaning “the type of code of”. So, for example,
τ1 → τ2 , which means “if τ1 holds at next time, then τ2 holds at next time,” is considered
the type of functions that take a piece of code of type τ1 and return another piece of code of
type τ2 . According to this intuition, he has developed λ , corresponding to the fragment
of LTL only with .
λ has two new term constructors next M and prev M , which correspond to the
introduction and elimination rules of , respectively. The type judgment of λ is of the
form Γ ⊢n M : τ , where Γ is a context, M is a term, τ is a type (a proposition of LTL,
only with ) and n is a natural number indicating a stage. A context, which corresponds
to assumptions, is a mapping from variables to pairs of a type and a natural number, since
the truth of a proposition depends on time. The key typing rules are those for next and
prev:
Γ ⊢n+1 M : τ
Γ ⊢n next M : τ
Γ ⊢n M : τ
.
Γ ⊢n+1 prev M : τ
1Note that this is equivalent to another, perhaps more standard presentation as a sublogic of CTL∗ [13].
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
5
The former means that, if M is of type τ at stage n + 1, then, at stage n, next M is code
of type τ ; the latter is its converse. Computationally, next and prev can be considered
quasiquote and unquote, respectively. So, in addition to the standard β-reduction, λ has
the reduction rule prev (next M ) −→ M , which cancels next by prev.
The code types in λ are often called open code types, since the quoted code may
contain free variables, so naively adding the construct to “run” quoted code does not work,
since it may cause unbound variable errors.
Although the logic is considered intuitionistic, Davies has only shown that the proof
system augmented with double negation elimination is equivalent to a standard axiomatic
formulation [13], which is known to be sound and complete with the Kripke semantics described above. Kojima and Igarashi [18, 19] have studied the semantics of intuitionistic LTL
and shown that the proof system obtained from λ is sound and complete with the given
semantics. Even though the Kripke semantics discussed here does not really correspond to
the logic obtained from the calculus, it certainly helps understand intuition behind the logic
and we will continue to use Kripke semantics in what follows for an explanatory purpose.
2.2. Multi-Modal Logic for Environment Classifiers. Taha and Nielsen [9] have introduced environment classifiers to develop λα , which has quasiquotation, run, and CSP
with a strong type system. We explain how λα can be derived from λ .2 Environment
classifiers are a special kind of identifiers with which code types and quoting are annotated:
for each classifier α, there are a type constructor hτ iα for code and a term constructor
hM iα to quote M . Then, a stage is naturally expressed by a sequence of classifiers, and a
type judgment is of the form Γ ⊢A M : τ , where natural numbers in a λ type judgment
are replaced with sequences A of classifiers. So, the typing rules of quoting and unquoting
(written ˜M ) in λα are given as follows:
Γ ⊢Aα M : τ
Γ ⊢A hM iα : hτ iα
Γ ⊢A M : hτ iα
.
Γ ⊢Aα ˜M : τ
Obviously, this is a generalization of λ : if only one classifier is allowed, then the calculus
is essentially λ .
The corresponding logic would also be a generalization of LTL, in which there are
several “dimensions” of linearly ordered time. A Kripke frame for the logic is given by a
transition system in which each transition relation is a map. More formally, a frame is a
α
triple (S, L, {−→| α ∈ L}) where S is the (non-empty) set of states, L is the set of labels,
α
and −→ ∈ S → S for each α ∈ L. Then, the semantics of hτ iα is given by: s hτ iα if and
α
only if s′ τ for any s′ such that s −→ s′ , where s and s′ are states.
α
The calculus λ has also a scoping mechanism for classifiers and it plays a central role
to guarantee safety of run. The term (α)M , which binds α in M , declares that α is used
locally in M and such a local classifier can be instantiated with another classifier by term
M [β]. We show typing rules for them with one for run below:
Γ ⊢A M : τ
α∈
/ FV(Γ, A)
A
Γ ⊢ (α)M : (α)τ
Γ ⊢A M : (α)τ
Γ ⊢A M [β] : τ [α := β]
Γ ⊢A M : (α)hτ iα
.
Γ ⊢A run M : (α)τ
2Unlike the original presentation, classifiers do not appear explicitly in contexts here. The typing rules
shown are accordingly adapted.
6
T. TSUKADA AND A. IGARASHI
The rule for (α)M requires that α does not occur in the context—the term M has no free
variable3 labeled α—and gives a type of the form (α)τ , which Taha and Nielsen called
α-closed type, which characterizes a relaxed notion of closedness. For example, the term
hλx : b.xiα is a closed term, so this term is α-closed and the judgment ∅ ⊢ε (α)hλx : b.xiα :
(α)hb → biα is valid. The term hxiα , however, is not α-closed because this term has free
variable x in the stage α, but β-closed (if β 6= α) because there is no free variable in the
stage containing the classifier β. The rule for run M says that an α-closed code fragment
annotated with α can be run. Note that h·iα (but not (α)·) is removed in the type of
run M . Taha and Nielsen have shown that α-closedness is sufficient to guarantee safety of
run.
When this system is to be interpreted as logic, it is fairly clear that (α)τ is a kind of
universal quantifier, as Taha and Nielsen have also pointed out [9]. Then, the question is
“What does a classifier range over?”, which has not really been answered so far. Another
interesting question is “How can the typing rule for run be read logically?”
One plausible answer to the first question is that “classifiers range over the set of
transition labels”. This interpretation matches the rule for M [β] and it seems that the
typing rules without run (with a classical axiom) are sound and complete with the Kripke
semantics that defines s (α)τ by s τ [α := β] for all β ∈ L. However, it is then difficult
to explain the rule for run.
The key idea to solve this problem is to have classifiers range over the set of finite (and
possibly empty) sequences of transition labels and to allow a classifier abstraction (α)M to
be applied to also sequences of classifiers. Then, run will be unified to a special case of
application of a classifier abstraction to the empty sequence. More concretely, we change
the term M [β] to M [B], where B is a possibly empty sequence of classifiers (the left rule
below). When B is empty and τ is hτ0 iα (assuming τ0 do not include α), the rule (as shown
as the right rule below) can be thought as the typing rule of (another version of) run, since
α-closed code of τ0 becomes simply τ0 (without (α)· as in the original λα ).
Γ ⊢A M : (α)τ
Γ ⊢A M [B] : τ [α := B]
Γ ⊢A M : (α)hτ0 iα
Γ ⊢A M [ε] : τ0
Another benefit of this change is that CSP for closed code (or embedding of persistent
code [10]) can be easily expressed. For example, if x is of the type (α)hintiα , then it can be
used as code computing an integer at different stages as in, say, h· · · (˜x[α]) + 3 · · · h· · · 4 +
( ˜˜ x[αβ]) · · · iβ · · · iα . So, once a programmer obtains closed code, she can use it at any
later stage. While our calculus λ⊲ does not have a primitive of CSP for all types, we can
express CSP in many cases. In Section 6.3, we discuss this subject in more detail.
Correspondingly, the semantics is now given by v, ρ; s τ where v is a valuation for
propositional variables and ρ is a mapping from classifiers to sequences of transition labels.
Then, v, ρ; s
hτ iα is defined by v, ρ; s′
τ where s′ is reachable from s through the
sequence ρ(α) of transitions and v, ρ; s (α)τ by: v, ρ[A/α]; s τ for any sequence A of
labels (ρ[A/α] updates the value of α to be A). In Section 5, we give a formal definition
of the Kripke semantics and show that the proof system, based in the ideas above, with
double negation elimination is sound and complete with respect to it.
3It is important to distinguish labels of free variables from free occurrences of classifiers. For example,
the term hλx : b.xiα has free occurrence of the classifier α, but no free variable labeled by α because there
is no free variable at all.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
7
3. The Calculus λ⊲
In this section, we define the calculus λ⊲ , based on the ideas described in the previous
section: we first define its syntax, type system, and small-step full reduction semantics
and states some basic properties; then we prove the time-ordered normalization property.
Finally, we give an example of programming in λ⊲ . We intentionally make notations for
type and term constructors different from λα because their precise meanings are different;
it is also to avoid confusion when we compare the two calculi.
3.1. Syntax. Let Σ be a countably infinite set of transition variables, ranged over by α
and β. A transition, denoted by A and B, is a finite sequence of transition variables; we
write ε for the empty sequence and AB for the concatenation of the two transitions. We
write Σ∗ for the set of transitions. A transition is often called a stage. We write FTV(A)
for the set of transition variables in A, defined by FTV(α1 α2 . . . αn ) = {αi | 1 ≤ i ≤ n}.
Let PV be the set of base types (corresponding to propositional variables), ranged over
by b. The set Φ of types, ranged over by τ and σ, is defined by the following grammar:
Types
τ ::= b | τ → τ | ⊲α τ | ∀α.τ .
A type is a base type, a function type, a code type, which corresponds to h·iα of λα , or
an α-closed type, which corresponds to (α)τ . The transition variable α of ∀α.τ is bound
in τ . In what follows, we assume tacit renaming of bound variables in types. The type
constructor ⊲α connects tighter than → and → tighter than ∀: for example, ⊲α τ → σ means
(⊲α τ ) → σ and ∀α.τ → σ means ∀α.(τ → σ). We write FTV(τ ) for the set of free transition
variables, which is defined in a straightforward manner.
Let Υ be a countably infinite set of variables, ranged over by x and y. The set of terms,
ranged over by M and N , is defined by the following grammar:
Terms
M ::= x | M M | λx : τ.M | ◮α M | ◭α M | Λα.M | M A .
In addition to the standard λ-terms, there are four more terms, which correspond to hM iα ,
˜M , (α)M , and M [β] of λα (respectively, in the order presented). Note that, unlike ˜M
in λα , the term ◭α M for unquote is also annotated. This annotation is needed because
a single transition variable can be instantiated with a sequence, in which case a single
unquote has to be duplicated accordingly. The variable x in λx : τ.M and the transition
variable α in Λα.M are bound in M . Bound variables are tacitly renamed to avoid variable
capture in substitution. We write FTV(M ) for the set of free transition variables, which is
defined in a straightforward manner, e.g., FTV(◮α M ) = FTV(◭α M ) = FTV(M ) ∪ {α},
FTV(Λα.M ) = FTV(M ) − {α} and FTV(M A) = FTV(M ) ∪ FTV(A).
3.2. Type System. As mentioned above, a type judgment and variable declarations in a
context are annotated with stages. A context Γ is a finite set {x1 : τ1 @A1 , . . . , xn : τn @An },
where xi are distinct variables. We often omit braces {}. We write FTV(Γ)
S for the set of
free transition variables in Γ, defined by: FTV({xi : τi @Ai | 1 ≤ i ≤ n}) = ni=1 (FTV(τi ) ∪
FTV(Ai )).
A type judgment is of the form Γ ⊢A M : τ , read “term M is given type τ under context
Γ at stage A.” Figure 1 presents the typing rules to derive type judgments. The notation
τ [α := B], used in the rule (Ins), is capture-avoiding substitution of transition B for α in
τ . When α in ⊲α is replaced by a transition, we identify ⊲ε τ with τ and ⊲AB τ with ⊲A ⊲B τ .
For example, (⊲α ∀α. ⊲α b)[α := ε] = ∀α. ⊲α b and (∀α. ⊲β b)[β := αα] = ∀α′ . ⊲α ⊲α b.
8
T. TSUKADA AND A. IGARASHI
Γ, x : τ @A ⊢A x : τ
(Var)
Γ, x : τ @A ⊢A M : σ
(Abs)
Γ ⊢A λx : τ.M : τ → σ
Γ ⊢A M : τ → σ
Γ ⊢A N : τ
(App)
Γ ⊢A M N : σ
Γ ⊢Aα M : τ
(◮)
Γ ⊢A ◮α M : ⊲α τ
Γ ⊢A M : ⊲α τ
(◭)
Γ ⊢Aα ◭α M : τ
Γ ⊢A M : τ
α∈
/ FTV(Γ) ∪ FTV(A)
(Gen)
A
Γ ⊢ Λα.M : ∀α.τ
Γ ⊢A M : ∀α.τ
(Ins)
Γ ⊢A M B : τ [α := B]
Figure 1: Typing rules.
The first three rules on the left are mostly standard except for stage annotations.
The conditions on stage annotations are similar to those in most multi-stage calculi: The
rule (Var) means that variables can appear only at the stage in which those variables
are declared. and the rule (Abs) requires the stage of the parameter to be the same as
that of the body and, correpondingly, the rule (App) requires M and N are typed at the
same stage. The next two rules (◮) and (◭) are for quoting and unquoting and already
explained in the previous section. The last two rules (Gen) and (Ins) are for generalization
and instantiation of a transition variable, respectively. They resemble the introduction and
elimination rules of ∀x.A(x) in first-order predicate logic: the side condition of the rule
(Gen) ensures that the choice of α is independent of the context. Computationally, this
side condition expresses α-closedness of M , that means M has no free variable which has
annotation α in its type or its stage. This is a weaker form of closedness, which means M
has no free variable at all.
3.3. Reduction. We will introduce full reduction M −→ N , read “M reduces to N in
one step,” and prove basic properties including subject reduction, confluence and strong
normalization.
Before giving the definition of reduction, we define substitution. Since the calculus
has binders for term variables and transition variables, we need two kinds of substitutions
for both kinds of variables. Substitution M [x := N ] for a term variable is the standard
capture-avoiding one, and its definition is omitted here. Substitution M [α := A] of A for α
is defined similarly to τ [α := A]. We show representative cases below:
(λx : τ.M )[α := A] = λx : (τ [α := A]).(M [α := A])
(M B)[α := A] = (M [α := A]) (B[α := A])
(◮β M )[α := A] = ◮β[α:=A] (M [α := A])
(◭β M )[α := A] = ◭β[α:=A] (M [α := A]),
where ◮α1 ...αn M is an abbreviation for ◮α1 · · · ◮αn M , and ◭α1 ...αn M for ◭αn · · · ◭α1 M .
In particular, (◮α M )[α := ε] = (◭α M )[α := ε] = M [α := ε]. Note that, when a transition variable in ◭ is replaced, the order of transition variables is reversed, because this
is the inverse operation of ◮. This is similar to the inversion operation in group theory:
−1
−1
(a1 a2 . . . an )−1 = a−1
n an−1 . . . a1 .
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
9
The reduction relation M −→ N is the least relation closed under the following three
computation rules
(λx : τ.M ) N −→ M [x := N ]
◭α (◮α M ) −→ M
(Λα.M ) A −→ M [α := A]
and congruence rules, which are omitted here. In addition to the standard β-reduction,
there are two rules: the second one, which is already explained previously, cancels quote by
unquote and the last one, instantiation of a transition variable, is similar to polymorphic
function application in System F. Note that the reduction is full—reduction occurs under
any context in any stage. This reduction relation can be thought as (non-deterministic)
proof normalization, which should preserve types, be confluent and strongly normalizing.
T
Then, we will define another reduction relation as a triple M −→ N , with T standing for
the stage of reduction in Section 3.5, as done in λ [7] and λ
[10], to prove time-ordered
normalization.
3.4. Basic Properties. We will prove three basic properties, namely, subject reduction,
strong normalization and confluence.
The key lemma is, as usual, Substitution Lemma, which says substitution preserves
typing. We will prove such a property for each kind of substitution. We define substitution
Γ[α := A] for contexts as follows:
{xi : τi @Ai }[α := A] = {xi : τi [α := A]@Ai [α := A]}
Lemma 3.1 (Substitution Lemma).
(1) If Γ, x : σ@B ⊢A M : τ and Γ ⊢B N : σ, then Γ ⊢A M [x := N ] : τ .
(2) If Γ ⊢A M : τ , then Γ[α := B] ⊢A[α:=B] M [α := B] : τ [α := B].
Proof. Easy induction on the typing rules. We only show main cases.
The proof of (1):
• Case M = x: It is the case that τ = σ and A = B. So, what we have to show is
B
Γ ⊢ N : τ , which is already assumed.
• Case M = M1 M2 : By the typing rules, we know that Γ, x : σ@B ⊢A M1 : τ0 → τ
and Γ, x : σ@B ⊢A M2 : τ0 for some τ0 . By the induction hypothesis, Γ ⊢A (M1 [x := N ]) :
τ0 → τ and Γ ⊢A (M2 [x := N ]) : τ0 . By the rule (App), we have Γ ⊢A (M1 M2 )[x := N ] : τ .
• Case M = ◮α M0 : By the typing rules, we know that τ = ⊲α τ0 and Γ, x : σ@B ⊢Aα
M0 : τ0 . By the induction hypothesis, Γ ⊢Aα M0 [x := N ] : τ0 . So, we obtain Γ ⊢A
(◮α M0 )[x := N ] : ⊲α τ0 .
The proof of (2):
• Case M = x: By the typing rules, x : τ @A ∈ Γ. So, x : (τ [α := B]) : (A[α := B]) ∈
Γ[α := B]. Therefore, Γ[α := B] ⊢A[α:=B] x : τ [α := B].
• Case M = ◮β M0 : By the typing rules, we have τ = ⊲β τ0 and Γ ⊢Aβ M0 : τ0 . By the
induction hypothesis, Γ[α := B] ⊢(A[α:=B])(β[α:=B]) M0 [α := B] : τ0 [α := B]. By applying
the rule (◮) as many times as required, we obtain Γ[α := B] ⊢A[α:=B] ◮β[α:=B] (M0 [α := B]) :
⊲β[α:=B] (τ0 [α := B]). By the definition of substitution, this is equal to what we have to show.
10
T. TSUKADA AND A. IGARASHI
• Case M = Λβ.M0 : By the typing rules, we have τ = ∀β.τ0 and Γ ⊢A M0 : τ0 . Moreover we can assume without loss of generality that β ∈
/ FTV(Γ) ∪ FTV(A) ∪ FTV(B) ∪ {α}.
A[α:=B]
By the induction hypothesis, we have Γ[α := B] ⊢
M0 [α := B] : τ0 [α := B]. Because
β∈
/ FTV(Γ[α := B]) ∪ FTV(A[α := B]), we obtain Γ[α := B] ⊢A[α:=B] (Λβ.(M0 [α := B])) :
∀β.(τ0 [α := B]). This judgement is equal to what we have to show.
Theorem 3.2 (Subject Reduction). If Γ ⊢A M : τ and M −→ M ′ , then Γ ⊢A M ′ : τ .
Proof. By straightforward induction on the derivation of M −→ M ′ , using Substitution
Lemma (Lemma 3.1). We only show three base cases and omit induction steps.
• Case M = (λx : τ0 .M0 ) M1 −→ M0 [x := M1 ] = M ′ : Because Γ ⊢A (λx : τ0 .M0 ) M1 :
τ , we have Γ, x : τ0 @A ⊢A M0 : τ and Γ ⊢A M1 : τ0 . By Substitution Lemma (1), we obtain
Γ ⊢A M0 [x := M1 ] : τ .
• Case M = (Λα.M0 ) B −→ M0 [α := B] = M ′ : Because Γ ⊢A (Λα.M0 ) B : τ , we
have α ∈
/ FTV(Γ) ∪ FTV(A) and Γ ⊢A M0 : τ0 such that τ0 [α := B] = τ . By the second
statement of Substitution Lemma (Lemma 3.1), we obtain Γ[α := B] ⊢A[α:=B] M0 [α := B] :
τ0 [α := B]. Because α ∈
/ FTV(Γ) ∪ FTV(A), Γ[α := B] = Γ and A[α := B] = A. Therefore,
A
Γ ⊢ M0 [α := B] : τ .
Γ
• Case M = ◭α ◮α M0 −→ M0 = M ′ : By replacing M in the assumption, we obtain
◭α ◮α M0 : τ . So we have Γ ⊢A M0 : τ as required.
⊢A
Theorem 3.3 (Strong Normalization). Let M be a typable term. There is no infinite
reduction sequence M −→ N1 −→ N2 −→ · · · .
Proof. We construct a term ♮(M ) of the simply typed λ-calculus (λ→ ) as follows:
♮(x)
♮(λx : τ.N )
♮(N1 N2 )
♮(◮α N )
♮(◭α N )
♮(Λα.N )
♮(N A)
=
=
=
=
=
=
=
x
λx.♮(N )
♮(N1 ) ♮(N2 )
♮(N )
♮(N )
♮(N )
♮(N )
We can easily prove the following propositions by induction of the structure of M :
(1) If M is typable in λ⊲ , then ♮(M ) is typable in λ→ .
(2) Suppose M −→ M ′ . If this is a β-reduction step, then ♮(M ) −→ ♮(M ′ ) in λ→ . Otherwise, ♮(M ) = ♮(M ′ ).
Now, assume there exists an infinite reduction sequence from a typable term M . It is
clear that there are infinitely many β-reduction. By (2), there exists an infinite reduction
sequence from ♮(M ), which is typable by (1). This contradicts the strong normalization
property of λ→ .
The last property we will show is confluence. We prove this by using parallel reduction
and complete development [20]. We define the parallel reduction relation M =⇒ N as in
Figure 2. Notice that, the rule (P-◭ ◮) allows more than one nested pairs of quoting and
unquoting to be cancelled in one step: For example, ◭α1 ◭α2 . . . ◭αn ◮αn . . . ◮α2 ◮α1 x =⇒
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
x =⇒ x
(P-Var)
11
M =⇒ N
(P-TApp)
M A =⇒ N A
M =⇒ N
(P-Lam)
λx : τ.M =⇒ λx : τ.N
M =⇒ N
(P-Ins)
(Λα.M ) A =⇒ N [α := A]
M1 =⇒ N1
M2 =⇒ N2
(P-App)
M1 M2 =⇒ N1 N2
M =⇒ N
(P-◮)
◮α M =⇒ ◮α N
M1 =⇒ N1
M2 =⇒ N2
(P-Beta)
(λx : τ.M2 ) M2 =⇒ N1 [x := N2 ]
M =⇒ N
(P-◭)
◭α M =⇒ ◭α N
M =⇒ N
(P-Gen)
Λα.M =⇒ Λα.N
M =⇒ N
(P-◭ ◮)
◭A ◮A M =⇒ N
Figure 2: Rules for Parallel Reduction.
x. It is not very standard in the sense that parallel reduction usually does not allow “hidden”
redices (that is, redices that appear only after some other reduction steps) to be contracted
in one step. We require this definition because a transition variable α can be replaced with
a sequence A of transition variables during reduction. If A in (P-◭ ◮) were α, Lemma 3.5
(2) below would not hold any longer.
The following lemma relates the reduction relation and the parallel reduction relation.
Lemma 3.4. (−→) ⊆ (=⇒) ⊆ (−→∗ )
Proof. (−→) ⊆ (=⇒) can be shown by induction on the derivation M −→ N . We can prove
(=⇒) ⊆ (−→∗ ) by induction on the structure of M =⇒ N .
Thanks to this lemma, we know that confluence of −→ is equivalent to confluence of
=⇒. We prove confluence of =⇒ by showing that =⇒ enjoys the diamond property. The
following properties of parallel reduction are useful.
Lemma 3.5.
(1) If M1 =⇒ N1 and M2 =⇒ N2 , then M1 [x := M2 ] =⇒ N1 [x := N2 ].
(2) If M =⇒ N , then M [α := A] =⇒ N [α := A].
Proof. Easy induction on the structure of the derivation M1 =⇒ N1 and M =⇒ N , respectively.
Now, we define the notion of complete development and show its key property. The
complete development M ⋆ of M is defined by induction as in Figure 3.
Lemma 3.6. If M =⇒ N , then N =⇒ M ⋆ .
Proof. By induction on the derivation of M =⇒ N with case analysis on M . We show only
interesting cases.
• Case (P-Lam): We have M = λx : τ.M0 and N = λx : τ.N0 with M0 =⇒ N0 . By
the induction hypothesis, we have N0 =⇒ M0 ⋆ . So, by applying (P-Lam) rule, we obtain
λx : τ.N0 =⇒ λx : τ.M0 ⋆ , as required.
12
T. TSUKADA AND A. IGARASHI
x⋆
(λx : τ.M )⋆
((λx : τ.M ) N )⋆
(M N )⋆
(◮α M )⋆
(◭A ◮A M )⋆
(◭α M )⋆
(Λα.M )⋆
((Λα.M ) A)⋆
(M A)⋆
=
=
=
=
=
=
=
=
=
=
x
λx : τ.M ⋆
M ⋆ [x := N ⋆ ]
M⋆ N⋆
◮α M ⋆
M⋆
◭α M ⋆
Λα.M ⋆
M ⋆ [α := A]
M⋆ A
(if M 6= λx : τ.M ′ )
(if M 6= ◭A ◮Aα M ′ )
(if M 6= Λα.M ′ )
Figure 3: Definition of Complete Development.
• Case (P-TApp): We have M = M0 A and N = N0 A with M0 =⇒ N0 . There are
two subcases.
Assume M0 6= Λα.M1 . By the induction hypothesis, we have N0 =⇒ M0 ⋆ . By applying
(P-TApp) rule, we obtain N0 A =⇒ M0 ⋆ A.
Assume M0 = Λα.M1 for some M1 . By the induction hypothesis, we have N0 =⇒ M0 ⋆ =
Λα.M1 ⋆ . By the definition of parallel reduction, we have that N0 = Λα.N1 for some N1 and
N1 =⇒ M1 ⋆ .
So, by applying (P-TIns) rule, we obtain (Λα.N1 ) A =⇒ M1 ⋆ [α := A] = M ⋆ .
• Case (P-Ins): We have M = (Λα.M0 ) A and N = N0 [α := A] with M0 =⇒ N0 .
By the induction hypothesis, we have N0 =⇒ M0 ⋆ . By applying Lemma 3.5, we obtain
N0 [α := A] =⇒ M0 ⋆ [α := A]
It is easy to show diamond property of =⇒ by using Lemma 3.6.
Lemma 3.7. If M =⇒ N1 and M =⇒ N2 , then there exists N which satisfies N1 =⇒ N
and N2 =⇒ N .
Proof. Choose M ⋆ as N and use the previous lemma.
Theorem 3.8 (Confluence). If M −→∗ N1 and M −→∗ N2 , then there exists N such that
N1 −→∗ N and N2 −→∗ N .
Proof. By Lemma 3.4, we have
(−→) ⊆ (=⇒) ⊆ (−→∗ ).
So we obtain
(−→∗ ) ⊆ (=⇒∗ ) ⊆ (−→∗ ),
which implies (−→∗ ) = (=⇒∗ ). Therefore what we should show is confluence of =⇒, which
is an easy consequence of Lemma 3.7.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
13
3.5. Annotated Reduction and Time-Ordered Normalization. We introduce the
notion of stages into reduction and prove the property called time-ordered normalization [7,
10]. Intuitively, it says that normalization can be done in the increasing order of stages
and does not need to ‘go back’ to earlier stages. In other words, once all redices at some
stage are contracted, subsequent reductions never yield a new redex at that stage. To state
time-ordered normalization formally, we first introduce the notion of paths from one stage
to another and a new reduction relation, annotated with paths to represent the stage at
which reduction occurs.
A path represents how the stage of a subterm is reached from the stage of a given term.
For example, if Γ ⊢α M and Γ′ ⊢αβ N for a subterm N of M , then we say the path from (the
stage of) M to (that of) N is β. The stage of a subterm may not be able to be expressed by
a transition (a sequence of transition variables), however: For example, consider the path
from ◭α M to M . We introduce formal inverses α−1 to deal with such cases: the path from
the stage of ◭α M to that of M is represented by α−1 . Similarly, the path from ◮α ◭β M
to M will be αβ −1 .
Formally, the set of paths, ranged over by T and U , is the free group generated by the
set of transition variables Σ. In other words, a path is a finite sequence ξ1 ξ2 . . . ξn , where
ξi = α or α−1 , such that it includes no subsequence of the form αα−1 or α−1 α for any α.
We define (α−1 )−1 = α and
−1
(ξ1 · · · ξn−1 ) · (ξn+2 . . . ξm )
(if ξn = ξn+1
)
(ξ1 · · · ξn ) · (ξn+1 . . . ξm ) =
−1
(ξ1 · · · ξn−1 ) · (ξn ξn+1 . . . ξm ) (if ξn 6= ξn+1 ).
The empty sequence ε is the unit element for the operation T · U . We simply write T U for
−1
. . . ξ1−1 .
T · U . We define (ξ1 ξ2 . . . ξn )−1 = ξn−1 ξn−1
We say a path T is positive if T does not contain formal inverses, in other words, the
canonical form of T is in Σ∗ . We can naturally identify the positive paths with transitions
and use metavariables A and B for positive paths. We write T ≤ U when there exists a
positive path A which satisfies T A = U . Clearly, ε ≤ T if and only if T is positive.
T
The annotated reduction relation is a triple of the form M −→ N , where M and N
are terms and T is a path from the stage of M to that of its redex—more precisely, that
of the constructor destructed by the reduction, since the stage of a redex and that of its
constructor may be different as in ◮α in redex ◭α ◮α M . The definition of the annotated
reduction, presented in Figure 4, is mostly straightforward. For example, α−1 is given to
◭α ◮α M −→ M (the rule (AR-Quote)), because the path to the constructor ◮α is α−1 .
As for the rule (AR-◮), the path from M to the constructor destructed by the reduction
is T and the path from ◮α M to M is α, hence the path from ◮α M to the constructor is
given by their concatenation αT . The rule (AR-◭) is similar. The rule (AR-Gen) is the
most interesting. First of all, α is bound here, so, we cannot propagate T in the premise to
the conclusion to prevent α from escaping its scope. We have found that replacing α with
ε, which is the earliest possible stage, is a reasonable choice, especially for time-ordered
normalization.
Annotated reduction is closely related to reduction defined in the previous section. It
T
is easy to see that M −→ N if and only if there exists T such that M −→ N . Furthermore,
such T is unique.
The next theorem shows that any reduction occurs indeed at a positive stage.
T
Theorem 3.9. If Γ ⊢A M : τ and M −→ N , then ε ≤ AT .
14
T. TSUKADA AND A. IGARASHI
ε
(λx : τ.M ) N −→ M [x := N ]
ε
(Λα.M ) A −→ M [α := A]
α−1
T
N −→ N ′
(AR-Beta)
T
M N −→ M N ′
T
(AR-Trans)
M −→ M ′
αT
◮α M −→ ◮α M ′
(AR-Quote)
α−1 T
◭α M −→ ◭α M ′
T
M −→ M ′
λx : τ.M −→ λx : τ.M ′
(AR-Abs)
T
M N −→ M ′ N
(AR-◭)
T
M −→ M ′
T [α:=ε]
T
M −→ M ′
(AR-◮)
T
M −→ M ′
◭α ◮α M −→ M
T
(AR-App2)
(AR-Gen)
Λα.M −→ Λα.M
(AR-App1)
T
M −→ M ′
T
M A −→ M ′ A
(AR-Ins)
Figure 4: Annotated Reduction.
T
Proof. Easy induction on the structure of M −→ N .
U
We say M is T -normal when there are no U ≤ T and N such that M −→ N . Then,
we can state time-ordered normalization as follows:
Theorem 3.10 (Time Ordered Normalization). Let M be a typable term. If M is T -normal
and M −→∗ N , then N is T -normal.
Proof. See Appendix A.
As its corollary, we know that for any reduction to a normal form from a typable term M
is “rearranged” according to an increasing order between stages. Moreover, this increasing
order can be any total order that respects ≤, i.e., includes ≤ as a subset.
Corollary 3.11. Let M be a typable term and be a total order that respects ≤. Then,
T1
T2
Tn
there is a reduction sequence M −→
N1 −→
· · · −→
Nn , which satisfies T1 T2 · · · Tn
and Nn is a normal form.
3.6. Programming in λ⊲ . We give an example of programming in λ⊲ . The example is the
power function, which is a classical example in multi-stage calculi and partial evaluation.
We augment λ⊲ with integers, Booleans, arithmetic and comparison operators, if-then-else,
a fixed point operator fix, and let. In the next section, we will formalize such a language
(without let) as MiniML⊲ and study its evaluation in more detail. For readability, we often
omit type annotations and put terms under quotation in shaded boxes.
We start with the ordinary power function without staging.
let power0 : int → int → int
= fix f. λn. λx. if n = 0 then 1 else x ∗ (f (n − 1) x)
Our purpose is to get a code generator power∀ that takes the exponent n and returns
(closed, hence runnable) code of λx.x ∗ x ∗ . . . x ∗ 1, which computes xn without recursion.
Here, we follow the construction of code generators in the previous work [15, 14].
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
15
First, we construct a code manipulator power1 : int → ⊲α int → ⊲α int, which takes an
integer n and a piece of integer code and then outputs a piece of code which concatenates the
input code by “∗” n times. It can be obtained by changing type annotation and introducing
quasiquotation.
let power1 : int → ⊲α int → ⊲α int
= fix f. λn. λx : ⊲α int.
if n = 0 then (◮α 1) else ◮α ((◭α x) ∗ (◭α f (n − 1) x))
Then, from power1 , we can construct a code generator powerα of type int → ⊲α (int → int),
which means it takes an integer and returns code of a function.
let powerα : int → ⊲α (int → int)
= λn. ◮α λx :int. ◭α (power1 n (◮α x))
It indeed behaves as a code generator: for example, powerα 3 would evaluate to the code
value ◮α λx : int .x ∗ (x ∗ (x ∗ 1)).
This construction is independent of the choice of the stage α. So, by abstracting α at
appropriate places in power1 and powerα , we can obtain the desired code generator, whose
return type is a closed code type ∀γ. ⊲γ (int → int).
let power2 : ∀β. int → ⊲β int → ⊲β int
= Λβ. fix f. λn. λx : ⊲β int.
if n = 0 then (◮β 1) else ◮β ((◭β x) ∗ (◭β f (n − 1) x))
let power∀ : int → ∀γ. ⊲γ (int → int)
= λn. Λγ. ◮γ λx :int. ◭γ (power2 γ n (◮γ x))
The output from power∀ is usable in any stage. For example, if we want code of a cube
function at the later stage, say A, then we write power∀ 3 A. In particular, when A is the
empty sequence ε, power∀ 3 ε : int → int evaluates to a function closure which computes
x ∗ x ∗ x ∗ 1 from the input x. The former corresponds to CSP (of closed code) and the
latter to run.
4. MiniML⊲
We extend λ⊲ and define an ML-like functional language MiniML⊲ , which has, in addition to the features of λ⊲ , integers, arithmetic and comparison operations, Booleans,
conditional expressions, and the (call-by-value) fixed-point combinator fix. We define the
type system and big-step evaluation semantics for MiniML⊲ and prove type soundness. In
this semantics, bindings of transition variables have to be maintained at run time. So, we
investigate a fragment of MiniML⊲ , in which programs can be executed by mostly forgetting
information on transition variables. We give another type system, which identifies such a
fragment, and erasure translation, which removes transitions from terms, and alternative
evaluation semantics for erased terms. Then, we prove the erasure property, which says
program executions before and after erasure agree.
16
T. TSUKADA AND A. IGARASHI
4.1. Syntax and Type System. The syntax of types and terms of MiniML⊲ is defined
as follows, where n and bv are metavariables ranging over integers and Boolean constants
true and false.
Types
Terms
τ ::= int | bool | τ → τ | ⊲α τ | ∀α.τ
M ::= x | n | bv | M = M | M + M | M − M | M ∗ M
| if M then M else M | fix f : τ → σ.M | λx : τ.M | M M
| ◮α M | ◭α M | Λα.M | M A
The type system is given as a straightforward extension of that of λ⊲ . We show typing
rules for the additional constructs.
bv ∈ {true, false}
n∈Z
(IntC)
(BoolC)
A
Γ ⊢ n : int
Γ ⊢A bv : bool
Γ ⊢A M : int
Γ ⊢A N : int
(Eq)
A
Γ ⊢ M = N : bool
Γ ⊢A M : int
Γ ⊢A N : int
♦ ∈ {+, −, ∗}
(IntOp)
A
Γ ⊢ M ♦ N : int
Γ ⊢A M : bool
Γ ⊢ A N1 : τ
Γ ⊢ A N2 : τ
(If)
Γ ⊢A if M then N1 else N2
Γ, f : τ → σ@A ⊢A M : τ → σ
(Fix)
Γ ⊢A fix f : τ → σ.M : τ → σ
We use the same notations for term substitution M [x := N ], and transition substitution
τ [α := A] and M [α := A] and other auxiliary notions, which can be similarly defined. It is
easy to prove that MiniML⊲ also enjoys Substitution Lemma.
Lemma 4.1 (Substitution Lemma).
(1) If Γ, x : σ@B ⊢A M : τ and Γ ⊢B N : σ, then Γ ⊢A M [x := N ] : τ .
(2) If Γ ⊢A M : τ , then, Γ[α := B] ⊢A[α:=B] M [α := B] : τ [α := B].
Proof. The proof is essentially the same as that of Substitution Lemma for λ⊲ (Lemma 3.1).
4.2. Evaluation and Type Soundness. Now, we give a big-step semantics and prove
that the execution of a well-typed program is properly divided into stages. The judgment
has the form ⊢A M ⇓ R, read “evaluating term M at stage A yields result R,” where R is
either err, which stands for a run-time error, or a value v, defined below. Values are given
via a family of sets V A indexed by transitions, that is, stages. The family V A is defined by
the following grammar:
vε ∈ V ε
::= n | true | false | λx : τ.M | ◮α V α | Λα.V ε
v A ∈ V A (A 6= ε) ::= x | n | true | false | λx : τ.V A | fix f : τ → σ.V A
| V A V A | ◮α V Aα | Λα.V A | V A B
′
| ◭α V A (if A′ α = A and A′ 6= ε)
The index A represents the current stage in which a value is typed. So, the index changes
under quoting and unquoting. Note that a value at a higher stage (that is, under quotation)
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
17
include free variables, applications and instantiation since computation is suspended. For
example, x y ∈ V α and so ◮α x y ∈ V ε .
Figure 5 shows the evaluation rules. Notice that metavariables M or N for terms
(not values) are used on the right side of ⇓, since it is not immediately clear that a result
is really a value of a proper form (or err)—we will prove such a property as a theorem.
The evaluation is left-to-right and call-by-value. The rules in Figure 5(1) are for ordinary
evaluation. The rule for ◭α M means that quote is canceled by unquote; since the resulting
term M ′ belongs to the stage α (inside quotation), α is attached to the conclusion. As
seen in the rule for Λα.M , Λ does not delay the evaluation of the body. The rule about
instantiation of a transition abstraction is straightforward. The rules for stages later than ε,
which are in Figure 5(2), are all similar: since the term to be evaluated is inside quotation,
each term constructor is left as it is and only subterms of stage ε will be evaluated. We also
need rules for handling erroneous terms, such as:
⊢ε M ⇓ M ′
M′ ∈
/Z
♦ ∈ {+, −, ∗, =}
ε
⊢ M ♦N ⇓ err
⊢A M ⇓ err
♦ ∈ {+, −, ∗, =}
A
⊢ M ♦N ⇓ err
They are shown in Appendix B.
We show a few properties of the big-step semantics. The first theorem says that evaluation is deterministic.
Theorem 4.2. If ⊢A M ⇓ R and ⊢A M ⇓ R′ , then R = R′ .
Proof. By straightforward induction on the derivation of ⊢A M ⇓ R.
The second theorem below says that, unless the result is err, the result must be a value
even though the rules do not say it is the case.
Theorem 4.3. Suppose ⊢A M ⇓ R. Then, either R = err or R ∈ V A .
Proof. By easy induction on the derivation ⊢A M ⇓ R.
The last property is type soundness and its corollary that if a well-typed program of
a code type yields a result, then the result is a quoted term, whose body is also typable
at stage ε. Unlike a usual setting where only closed terms are considered programs, free
variables at non-ε stages are considered symbols and do not cause unbound variable errors
in MiniML⊲ , so we relax the notion of programs to include terms that contain such symbolic
variables. We say that Γ is ε-free if it satisfies A 6= ε for any x : τ @A ∈ Γ; then, a program
is a term which is typed under an ε-free environment. In the statement of Type Soundness
Theorem, we also use the notation Γ−A , defined by: Γ−A = {x : τ @B | x : τ @AB ∈ Γ}.
Theorem 4.4 (Type Soundness). If Γ is ε-free and Γ ⊢ε M : τ and ⊢ε M ⇓ R, then R = v
and v ∈ V ε for some v and Γ ⊢ε v : τ . In particular, if τ = ⊲α τ0 , then v = ◮α N and
Γ−α ⊢ε N : τ0 .
Proof. See Section C.1.
18
T. TSUKADA AND A. IGARASHI
⊢ε M ⇓ m
⊢ε N ⇓ n
m=n
ε
⊢ M = N ⇓ true
⊢ε M ⇓ m
⊢ε M ⇓ m
⊢ε N ⇓ n
m 6= n
ε
⊢ M = N ⇓ false
⊢ε N ⇓ n
k = m♦n (where ♦ ∈ {+, −, ∗})
⊢ε M ♦N ⇓ k
⊢ε M ⇓ false
⊢ε N2 ⇓ N2′
⊢ε if M then N1 else N2 ⇓ N2′
⊢ε M ⇓ true
⊢ε N1 ⇓ N1′
⊢ε if M then N1 else N2 ⇓ N1′
⊢ε M ⇓ λx : τ.M ′
⊢ε λx : τ.M ⇓ λx : τ.M
⊢ε
⊢α M ⇓ M ′
◮α M ⇓ ◮α M ′
⊢ε N ⇓ N ′
⊢ε M ′ [x := N ′ ] ⇓ M ′′
⊢ε M N ⇓ M ′′
⊢ε M ⇓ ◮α M ′
⊢α ◭α M ⇓ M ′
⊢ε
⊢ε M ⇓ Λα.M ′
⊢ε M ′ [α := B] ⇓ M ′′
⊢ε M B ⇓ M ′′
⊢ε M ⇓ M ′
Λα.M ⇓ Λα.M ′
c ∈ Z ∪ {true, false}
⊢ε c ⇓ c
⊢ε M [f := fix f : τ → σ.M ] ⇓ M ′
⊢ε fix f : τ → σ.M ⇓ M ′
(1) Rules for ordinary evaluation.
⊢A M ⇓ M ′
⊢A N ⇓ N ′
♦ ∈ {+, −, ∗, =}
M ♦N ⇓ M ′ ♦N ′
⊢A
⊢A M ⇓ M ′
⊢A N1 ⇓ N1′
⊢A N2 ⇓ N2′
⊢A if M then N1 else N2 ⇓ if M ′ then N1′ else N2′
⊢A x ⇓ x
⊢A M ⇓ M ′
⊢A λx : τ.M ⇓ λx : τ.M ′
⊢Aα M ⇓ M ′
A
⊢ ◮α M ⇓ ◮α M ′
⊢A M ⇓ M ′
⊢Aα ◭α M ⇓ ◭α M ′
⊢A M ⇓ M ′
⊢A Λα.M ⇓ Λα.M ′
⊢A
⊢A M ⇓ M ′
⊢A N ⇓ N ′
⊢A M N ⇓ M ′ N ′
⊢A M ⇓ M ′
⊢A M B ⇓ M ′ B
c ∈ Z ∪ {true, false}
⊢A c ⇓ c
⊢A M ⇓ M ′
fix f : τ → σ.M ⇓ fix f : τ → σ.M ′
(2) Rules for evaluation inside quotation. Here, A 6= ε.
Figure 5: Big-Step Semantics of MiniML⊲ .
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
19
4.3. Staged Transition Variables for Erasure Property. The evaluation of MiniML⊲
introduced above relies on the annotation of transition variables. For example, consider two
terms
M1 = (Λα.◮α (λx.1)(◮β ((fix f.f )2))) ε
M2 = (Λα.◮α (λx.1)(◮α ((fix f.f )2))) ε.
The only difference is the annotation on ◮((fix f.f )2), but ⊢ε M1 ⇓ 1 whereas there is no
term N such that ⊢ε M2 ⇓ N . In other words, the evaluation of M1 terminates but that of
M2 diverges. Therefore, we must record how transition variables are bound to transitions
during evaluation.
From the implementation point of view, it is desirable that evaluation is insensitive to
the annotation as much as possible to avoid overhead. In λα [9], environment classifiers
can be regarded as completely static citizens so that the evaluation does not require them,
although the authors do not explicitly state it. The property that the evaluation goes well
even if we erase the annotations is called erasure property. The previous example shows
that the erasure property does not hold for MiniML⊲ . Since the argument B, especially,
its length, in an instantiation M B is significant at run time, we cannot erase transitions
completely. So, we consider a slightly weaker notion of erasure, which removes transition
variables only from ◮, ◭ and Λ and replaces the transition B in M B with its length. The
goal of this section is to find a practically meaningful subset of MiniML⊲ , which enjoys the
erasure property under the translation sketched above.
The reasons why the erasure property is broken are (1) Λ-bound transition variables
are used “too far” from the binder, as is the case in M2 and (2) the “depth” of quoting
◮α can be changed by using instantiation with a transition, whose length is not 1. In the
case of M2 , there is an occurrence of transition variable α far from the binder and α is
instantiated by ε, whose length is 0. So, to ensure the erasure property, it is enough to
prevent both (1) and (2) from holding at once, in other words, to guarantee that Λ-bound
transition variables occur near the binder or to restrict instantiations to only transitions of
length 1.
Based on this observation, we will introduce two instantiation rules. The first rule is
for instantiation of transition variables used only near the binder. We can change the depth
of quoting by using this rule, but this rule can be applied only in limited situations. The
second rule is for instantiation of transition variables by transitions, whose lengths are 1.
This rule can be applied to any ∀-types, but we cannot change the depth of quoting. We
introduce a new term constructor M [α] to distinguish from the former.
The first instantiation rule requires some control on the occurrences of transition variables. We enforce one additional restriction, which requires that transition variables be also
staged like term variables. This restriction rejects a type with nested occurrences of ⊲α ,
such as ∀α. ⊲α ⊲α τ , whose term would have a distant use of ◮α . This restriction is closely
related to the distinction between open and closed code types in λi [16].
We define a new type system with staged transition variables. We need two changes to
deal with the stages for transition variables. First, we introduce environments for transition
variables. A transition environment is a set of the form {α1 @A1 , . . . , αn @An }, where αi are
distinct transition variables. An intuitive meaning of α@A is that the valid occurrence of α
is always of the form Aα. The second change is the annotation for the universal quantifier.
The new syntax for universal quantification is ∀α@A.τ , where A is the (positive) path from
the current stage to the stage in which α is usable.
20
T. TSUKADA AND A. IGARASHI
∆ ⊢s A
b ∈ {int, bool}
(ST-Base)
∆ ⊢A
s b
∆ ⊢A
∆ ⊢A
s τ
s σ
(ST-Imp)
A
∆ ⊢s τ → σ
∆ ⊢Aα
s τ
(ST-Code)
A
∆ ⊢s ⊲α τ
∆, α@AB ⊢A
s τ
(ST-Univ)
∀α@B.τ
∆ ⊢A
s
Figure 6: The definition of well-formed types under the transition environment ∆.
Next, we define well-formed transitions, transition environments, types, and type environments to ensure every use of a transition variable is valid. We say a transition A =
α1 . . . αn is well formed under a transition environment ∆ if, for any i ≤ n, αi @α1 . . . αi−1 ∈
∆. We say ∆ is well formed if, for any α@A ∈ ∆, A is well formed under ∆, i.e., all stages
where transition variables are declared are well formed. This definition avoids the circular
definition of transition variables, e.g., α@β, β@α. We write ∆ ⊢s A if A is well formed
under ∆, and ⊢s ∆ if ∆ is a well-formed transition environment.
The judgment of the form ∆ ⊢A
s τ means “type τ is well formed at stage A under ∆”,
and defined by the rules in Figure 6. The base types int and bool are always well formed
at any well-formed stage. The rules for τ → σ and ⊲α τ resemble the typing rules (Abs) and
(◮), respectively. The type ∀α@B.τ , which binds a new transition variable α, is well formed
at A under ∆ if τ is well formed under the transition variables environment extended by
the new transition variable declaration α@AB. Finally, we define well-formedness of type
environment Γ under ∆, written ∆ ⊢s Γ, by: Γ is well formed under ∆ if and only if ∆ is
well formed and, for any x : τ @A ∈ Γ, τ is well formed at A under ∆ (i.e., ∆ ⊢A
s τ ).
Figure 7 shows the typing rules that differ from the previous type system (except the
addition of ∆). They have additional premises about well-formedness. The rule (S-Var)
requires the well-formedness of environment Γ, x : τ @A, which will require well-formedness
of the type τ at A and the transition environment ∆. The rules (S-Num) and (S-Bool)
require the well-formedness of the environment Γ and the stage A, which ensures the wellformedness of the base types. The typing rule (S-Gen) records the path from the current
stage to the stage in which α is usable. This information is used by the rules (S-Ins1) and
(S-Ins2). As mentioned above, there are two kinds of transition instantiation rules and
corresponding term constructors. The first one (S-Ins1) is computationally meaningful, in
other words it may change the depth of quoting, but can be used only in limited situations.
The second one (S-Ins2) does not change the depth of quoting, so this is computationally
meaningless and we can use anytime. Here, substitution for a transition variable α in
∀α@A.τ (among other types) is defined as follows:
(∀α@A.τ )[β := B] = ∀α@(A[β := B]).(τ [β := B])
A
It is easy to see that Γ; ∆ ⊢A
s M : τ implies Γ ⊢ M : τ .
Now, we define the syntax for erased terms, terms after erasure and the erasure translation from λ⊲ terms to erased terms, and the big-step semantics for erased terms. The
syntax of erased terms, ranged over by M ♭ , is as follows:
Erased Terms
M ♭ ::= x | n | b | M ♭ = M ♭ | M ♭ + M ♭ | M ♭ − M ♭ | M ♭ ∗ M ♭
| if M ♭ then M ♭ else M ♭ | fix f.M ♭ | λx.M ♭ | M ♭ M ♭
| ◮ M ♭ | ◭ M ♭ | ΛM ♭ | M ♭ [] | M ♭ n (n ≥ 0)
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
21
α∈
/ FTV(Γ) ∪ FTV(∆)
Γ; ∆, α@AB ⊢A
s M :τ
(S-Gen)
Γ; ∆ ⊢A
Λα.M
:
∀α@B.τ
s
∆ ⊢s Γ, x : τ @A
(S-Var)
Γ, x : τ @A; ∆ ⊢A
s x:τ
∆ ⊢s Γ
∆ ⊢s A
(S-Num)
A
Γ; ∆ ⊢s n : int
Γ; ∆ ⊢A
s M : ∀α@ε.τ
∆ ⊢s AB
τ = ⊲α σ
(S-Ins1)
A
Γ; ∆ ⊢s M B : τ [α := B]
∆ ⊢s Γ
∆ ⊢s A
bv ∈ {true, false}
(S-Bool)
Γ; ∆ ⊢A
s bv : bool
Γ; ∆ ⊢A
s M : ∀α@B.τ
β@AB ∈ ∆
(S-Ins2)
Γ; ∆ ⊢A
s M [β] : τ [α := β]
Figure 7: The typing rules which differ from the previous type system.
The erasing function ♭(·) from terms to erased erased terms is defined as follows:
♭(c)
♭(M ♦N )
♭(if M then N1 else N2 )
♭(fix f : τ → σ.M )
♭(λx : τ.M )
♭(M N )
♭(◮α M )
♭(◭α M )
♭(Λα.M )
♭(M [β])
♭(M A)
n
=
=
=
=
=
=
=
=
=
=
=
c
(c ∈ Z ∪ {true, false})
♭(M )♦♭(N )
(♦ ∈ {+, −, ∗, =})
if ♭(M ) then ♭(N1 ) else ♭(N2 )
fix f.♭(M )
λx.♭(M )
♭(M ) ♭(N )
◮ ♭(M )
◭ ♭(M )
Λ♭(M )
♭(M ) []
♭(M ) n (n is the length of A).
n
z }| {
z }| {
So, ♭(◮A M ) = ◮ . . . ◮ ♭(M ) and ♭(◭A M ) = ◭ . . . ◭ ♭(M ) where n is the length of A.
The erasure semantics is essentially the same as the ordinary evaluation semantics in
Section 4.2, except for the two differences: one is that the stage A of ⊢A M ⇓ N is replaced
by the natural number n, which is the length of A; and the other is that the rule for M n at
the stage 0. In this case, M must be evaluated to Λ ◮ M ′ and ◮ at its head is duplicated
by n times. We show only main rules below.
⊢n+1
M♭ ⇓ N♭
⊢0e M ♭ ⇓ ΛN ♭
e
⊢ne ◮ M ♭ ⇓ ◮ N ♭
⊢0e M ♭ [] ⇓ N ♭
n
z }| {
⊢0e M ♭ ⇓ Λ ◮ M0♭
⊢0e ◮ ◮ · · · ◮ M0♭ ⇓ N ♭
⊢0e M ♭ n ⇓ N ♭
⊢0e M ♭ ⇓ ◮ N ♭
⊢1e ◭ M ♭ ⇓ N ♭
Finally, we state the erasure property: the erasure semantics is equivalent to the semantics with transition variables for terms typed under the new type system.
Theorem 4.5 (Erasure Property). Suppose Γ is ε-free and Γ; ∆ ⊢εs M : τ . Then,
(1) if ⊢ε M ⇓ N , then ⊢0e ♭(M ) ⇓ ♭(N ); and
22
T. TSUKADA AND A. IGARASHI
(2) if ⊢0e ♭(M ) ⇓ N ♭ , then there is some N ′ such that ⊢ε M ⇓ N ′ and N ♭ = ♭(N ′ ).
Proof. See Appendix C.2.
We believe that the calculus with this new type system does not lose much expressiveness for practical use; in fact, the example of power functions in Section 3.6 can be typed
with the new type system.
5. Kripke Semantics for λ⊲ and Logical Completeness
In this section, we formalize the Kripke semantics discussed in Section 2 and prove completeness of a classical version of the proof system obtained from λ⊲ to justify the informal
interpretation of types in λ⊲ as formulae of a logic. We augment the set of propositions
(namely types) with falsity and the proof rules with double negation elimination. It is left
for future work to study the semantics of the intuitionistic version, of which recent work on
Kripke semantics for intuitionistic LTL [18, 19] can be a basis.
First, we (re)define the set of propositions and the natural deduction proof system.
Then, we proceed to the formal definition of the Kripke semantics and prove soundness and
completeness of the proof system. Finally, we examine another rule for the double negation
elimination.
5.1. Natural Deduction. The set Φ⊥ , ranged over by φ and ψ, of propositions is given
by the grammar for Φ extended with a new constant ⊥.
The natural deduction system can be obtained by forgetting variables and terms in
the typing rules. We add the following new rule, which is the ordinary double negation
elimination rule, adapted for this setting:
Γ, (φ → ⊥)@A ⊢B ⊥
(⊥-E) .
Γ ⊢A φ
5.2. Kripke Semantics and Completeness. As mentioned in Section 2, the Kripke
a
semantics for this logic is based on a functional transition system T = (S, L, {−→ | a ∈ L})
where S is the (non-empty) countable set of states, L is the countable set of labels, and
a1 ···an ′
a
s if there exist s1 , . . . , sn−1 such that
−→ ∈ S → S for each label a ∈ L. We write s −→
an−1
a1
a2
an ′
s −→ s1 −→ · · · −→ sn−1 −→ s . Actually, given s, a1 , . . . , an , s′ always exists in this setting
ai
because −→
is a total function for all 1 ≤ i ≤ n.
To interpret a proposition, we need two valuations, one for propositional variables and
the other for transition variables. The former is a total function v ∈ S × PV → {0, 1}; the
latter is a total function ρ ∈ Σ → L∗ , where L∗ is the set of all finite sequences of labels.
Then, we define the satisfaction relation T , v, ρ; s φ, where s ∈ S is a state, as follows:
T , v, ρ; s
T , v, ρ; s
T , v, ρ; s
T , v, ρ; s
T , v, ρ; s
p
⊥
φ→ψ
⊲α φ
∀α.φ
iff
iff
iff
iff
v(s, p) = 1
never occurs
T , v, ρ; s 6 φ or T , v, ρ; s
, v, ρ; s′
ρ(α)
ψ
s′
T
φ where s −→
for all A ∈ L∗ , T , v, ρ[A/α]; s
φ
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
23
Here, ρ[A/α] is defined by: ρ[A/α](α) = A and ρ[A/α](β) = ρ(β) (for β 6= α). The
satisfaction relation is extended pointwise to contexts Γ (possibly infinite sets of pairs of a
proposition and a transition4) by:
T , v, ρ; s
Γ iff T , v, ρ; s
The local consequence relation Γ
Γ
φ iff T , v, ρ; s
⊲A φ for all φ@A ∈ Γ .
φ is defined by:
Γ implies T , v, ρ; s
φ for any T , v, ρ, s .
Then, the natural deduction proof system is sound and complete with respect to the
local consequence relation. The proof is similar to the one for first-order predicate logic:
we use the standard techniques of Skolemization and Herbrand structure.
Theorem 5.1 (Soundness). If Γ ⊢ε φ, then Γ
φ.
Proof. By induction on the derivation of Γ ⊢ε φ.
Theorem 5.2 (Completeness). If Γ
φ, then Γ ⊢ε φ.
Proof. We give a proof sketch; more detailed proofs are found in Appendix D.
We assume Γ 0ε φ. We construct a transition system T , two valuations v and ρ and
a state s such that T , v, ρ; s
Γ and T , v, ρ; s 1 φ. The construction is similar to the
construction of counter models in first-order predicate logic.
First, we prove the following proposition.
Γ 0ε ⊥ implies there exists T , v, ρ and s such that T , v, ρ; s
Γ.
We can reduce this proposition to the completeness of quantifier-free logic (the logic without
the quantifier over transition variables) by using Skolemization and Herbrand Universes.
Then, we prove this theorem by contraposition. Because Γ 0ε φ, we have Γ, φ → ⊥@ε 0ε
⊥. Therefore, there exist T , v, ρ and s which satisfy T , v, ρ; s Γ and T , v, ρ; s φ → ⊥.
Then, T , v, ρ; s 1 φ because T , v, ρ; s φ → ⊥. Therefore, Γ 1 φ.
5.3. Alternative Semantics. We can give an alternative deduction rule for double negation elimination.
Γ, (φ → ⊥)@A ⊢A ⊥
(⊥-E)
Γ ⊢A φ
The difference is in the stage of the premise. This rule requires that the stage of ⊥ is equal to
the stage of φ → ⊥, but in the rule in Section 5.1 the stage of ⊥ is arbitrary. This restriction
makes the proof system weaker: for example, in this setting ⊲α is not self-dual, that means
¬ ⊲α ¬φ ↔ ⊲α φ is not provable (here ¬φ is an abbreviation of φ → ⊥), while under the
previous rules ⊲α is self-dual. The difference is equivalent to the axiom ⊲A ⊥ ↔ ⊲B ⊥ (or
a weaker form ∀α.(⊲α ⊥ → ⊥)) in the sense that this system with this axiom is equivalent
to the previous proof system. This axiom corresponds to the axiom ¬A ↔ ¬
A in
linear-time temporal logic, due to Stirling [13].
4We allow Γ to be infinite for technical convenience. For an infinite context Γ, we write Γ ⊢A φ if there
exists a finite context Γ′ ⊆ Γ such that Γ′ ⊢A φ. The following argument holds if we restrict Γ to be finite
because the logic is compact, i.e., Γ is unsatisfiable if and only if there exists a finite subset Γ′ ⊆ Γ such that
Γ′ is also unsatisfiable. For more detail, see Appendix D.
24
T. TSUKADA AND A. IGARASHI
[[b]] = b
[[τ → σ]] = [[τ ]] → [[σ]]
[[ τ ]] = ⊲α [[τ ]]
[[x]]
[[λx : τ.M ]]
[[M N]]
[[next M ]]
[[prev M ]]
=
=
=
=
=
x
λx : [[τ ]].[[M ]]
[[M ]] [[N ]]
◮α [[M ]]
◭α [[M ]]
[[·]] = ·
[[Γ, x : A@n]] = [[Γ]], x : [[A]]@αn
Figure 8: Embedding from λ
to λ⊲ . α is a fixed transition variable.
There is a corresponding semantics, with respect to which the new proof system is sound
and complete. In fact, this is achieved by a minor change: we allow transition functions
ρ(α)
a
−→ to be partial. As a result, there can be no s′ such that s −→ s′ . In this setting, the
a1 ...an
as there is no s′ which satisfies
semantics for ⊲α φ has to be modified. We define s −→
6
a1 ...an ′
s −→ s , and T , v, ρ; s ⊲α φ as follows:
T , v, ρ; s
⊲α φ
iff
ρ(α)
s −→
6
or T , v, ρ; s′
ρ(α)
φ where s −→ s′
Completeness and soundness are proved similarly.
6. Comparing with other multi-stage calculi
In this section, we will compare λ⊲ with closely related calculi λ [7], the Kripke-style
modal λ-calculus [8], λα [9] and λi [16]. The first two calculi are based on Curry-Howard
correspondence between multi-stage calculi and modal logics and our work can be considered
a generalization of them. In fact, there are embeddings from these two calculi to λ⊲ . The
calculi λα and λi are multi-stage calculi with environment classifiers. We discuss several
differences among these two calculi and λ⊲ . Although it does not seem possible to give
(straightforward) embeddings from them, due to the presence of CSP, we will show that an
embedding from the CSP-free fragment of λi to λ⊲ is possible.
6.1. Comparing with λ . λ [7] is a multi-stage calculus corresponding to linear-time
temporal logic (LTL). As already mentioned in Section 2, λ is obtained by using only one
transition variable, say α. Then the modal operator
to mean “next” in LTL corresponds
to the modal operator ⊲α and the stage n in LTL corresponds to αn and so on. We define
an embedding [[·]] from λ into λ⊲ in Figure 8. This embedding is essentially the same as
that from λ calculus into λα , given by Taha and Nielsen [9].
The following two theorems show the correctness of the embedding.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
25
[[b]] = b
[[τ → σ]] = [[τ ]] → [[σ]]
[[τ ]] = ∀α. ⊲α [[τ ]]
[[x]]A = x
[[λx : τ.M ]]A = λx : [[τ ]].[[M ]]A
[[M N ]]A = [[M ]]A [[N ]]A
[[box M ]]A = Λα. ◮α [[M ]]Aα
A
[[unboxn M ]]
(where α ∈
/ FTV(A))
A′
= ◭B [[M ]] B
(where A = A′ B and the length of B is n)
[[x1 : τ1 , . . . xn : τn ]]A = x1 : [[τ1 ]]@A, . . . , xn : [[τn ]]@A
[[Γ0 ; · · · ; Γn ]]α1 ···αn
= [[Γ0 ]]ε , . . . , [[Γi ]]α1 ···αi , . . . , [[Γn ]]α1 ···αn
Figure 9: Embedding from the Kripke-style modal λ-calculus to λ⊲ .
n
Theorem 6.1. If Γ ⊢n M : τ in λ , then [[Γ]] ⊢α [[M ]] : [[τ ]].
Proof. By induction on the type derivation.
Theorem 6.2. Let M , N be λ
terms. Then, M −→ N iff [[M ]] −→ [[N ]].
Proof. By induction on the structure of M .
Moreover, by giving a suitable definition of reduction annotated with time for λ , we
can easily show that the embedding preserves the stage of reduction.
We can construct a reverse mapping, a type- and reduction-preserving embedding from
the quantifier-free fragment of λ⊲ to λ , by simply forgetting annotations of transition
variables. Moreover, the quantifier-free fragment of λ⊲ with only one transition variable
is isomorphic to λ in the sense that there is a bijection that preserves typability and
reduction.
6.2. Comparing with Calculus based on S4. Davies and Pfenning [8, 12] develop calculi
that correspond to intuitionistic modal logic S4 (only with ). The type τ represents
closed code values, which thus can be run or embedded in code of any later stages, as is
possible in λ⊲ . We compare λ⊲ with one of their calculi (what they call the Kripke-style
modal λ-calculus in Section 4 of [8]), in which, there are box and unboxn for quoting and
unquoting, respectively (see Pfenning and Davies [8] for details). An embedding [[·]] from
this calculus into λ⊲ is given in Figure 9.
The following theorems state the correctness of the embedding.
Theorem 6.3. Consider a sequence A of distinct transition variables of length n. Then
Γ0 ; Γ1 ; . . . ; Γn ⊢ M : τ implies [[Γ0 ; . . . ; Γn ]]A ⊢A [[M ]]A : [[τ ]].
Proof. An easy induction on the type derivation Γ0 ; . . . ; Γn ⊢ M : τ .
26
T. TSUKADA AND A. IGARASHI
Theorem 6.4. Let A be a sequence of distinct transition variables of length n. Provided
that Γ0 ; . . . ; Γn ⊢ M : τ is derivable in the Kripke-style modal λ-calculus, then M −→β N
iff [[M ]]A −→ [[N ]]A .
Proof. By induction on the structure of M .
Taha and Nielsen have shown a similar embedding from the Kripke-style calculus into
λα . In their embedding, the translation of unbox, which corresponds to the elimination rule
for , is slightly more involved than ours, since they use run and the CSP operator. In our
embedding, on the other hand, unbox is expressed uniformly by M B, which corresponds
to the elimination rule for ∀.
6.3. Comparing with λα and λi . Comparing λ⊲ with λα [9], we can point out two differences: the meaning of run and the absence of CSP primitive.
In λα , run is a primitive, while, in λ⊲ , run M is defined as a syntax sugar for M ε.
The following rules are typing rules for run in λα and λ⊲ , respectively.
Γ ⊢A M : (α)hτ iα
Γ ⊢A run M : (α)τ
Γ ⊢A M : ∀α.τ
Γ ⊢A M ε : τ [α := ε]
Aside from the presence of a binder (α), which is not essential, an important difference is
how code type constructors are removed in the conclusion. In λα , run removes only the
outermost bracket annotated with α, while, in λ⊲ , M ε removes all code-type constructor
⊲α in τ . This difference in typing rules is also reflected in reductions.
run (α)hviα −→ (α)v
(Λα.v) ε −→ v[α := ε]
(Here we assume that v does not contain the CSP constructor to simplify the argument—
See [9] for the complete definition.) So, it does not seem very easy to give an embedding of
either direction.
From a practical point of view, we do not think this difference is very significant. It is
not clear when one wants to share one environment classifier or transition variable among
different stages. If the use of classifiers or transition variables is staged, as we discussed
in Section 4.3, then the difference is very little. In fact, the current implementation of
MetaOCaml is based on λi , which can be considered a subset of both λα and λ⊲ (if CSP is
ignored), and this fact shows that the difference is practically insignificant.
As another difference, λα allows the CSP constructor % to be applied to any terms to
embed the value of the term inside a quotation. It is easy to see that φ@ε ⊢α φ is not
provable in general and so such a universal CSP operator would not be expressible in λ⊲ .
However, we can support CSP for many types. CSP for values of base types such as integers
and Booleans is easy. CSP for function closures is also possible if they do not contain open
code in their bodies or environment. We can deal with CSP for closed code, i.e., the terms
which have types of the form ∀α. ⊲α τ with α ∈
/ FTV(τ ), as syntactic sugar in λ⊲ . The
α
following rules are for CSP in λ and for CSP as syntactics sugar in λ⊲ .
Γ ⊢A M : τ
Γ ⊢Aα % M : τ
Γ ⊢A M : ∀β. ⊲β τ
β∈
/ FTV(τ )
Γ ⊢Aα Λβ.(◭α (M αβ)) : ∀β. ⊲β τ
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
[[b]]
[[τ → σ]]
[[hτ iα ]]
[[hτ i]]
=
=
=
=
b
[[τ ]] → [[σ]]
⊲α [[τ ]]
∀α. ⊲α [[τ ]]
27
(where α ∈
/ FV(τ ))
[[x]]A = x
[[λx.M ]]A = λx : τ.[[M ]]A
A
[[M N ]]
[[hM i]]
[[˜M ]]
[[run M ]]
[[open M ]]
[[close M ]]
=
=
=
=
=
=
A
(τ should be chosen appropriately)
A
[[M ]] [[N ]]
◮α [[M ]] (α should be chosen appropriately)
◭α [[M ]] (α should be chosen appropriately)
[[M ]] ε
[[M ]] α (α should be chosen appropriately)
Λα.[[M ]] (α should be chosen appropriately)
[[x1 : τ1A1 , . . . xn : τnAn ]] = x1 : [[τ1 ]]@A1 , . . . , xn : [[τn ]]@An
Figure 10: Embedding from λi without CSP primitive to λ⊲ . Precisely speaking, in order
to recover τ and α, which appear only on the right hand side, this embedding
takes type derivations of λi rather than terms as an input.
The only problematic case is CSP for open code and, as mentioned above, functions containing open code, but we think it is rarely needed.
λi [16] is developed as a subset of λα where type inference is possible (although there
are slight differences in syntax and typing). The difference between λ⊲ and λi , is smaller
than that between λ⊲ and λα . In fact, we can construct an embedding that preserves typing
from λi (without the CSP operator) into λ⊲ , by observing that the executable code type
hτ i in λi corresponds to∀α. ⊲α τ (where α ∈
/ FTV(τ )) in λ⊲ . Figure 10 shows the complete
description of the embedding. Precisely speaking, it takes a type derivation rather than
terms: For example, the rule for open means
##
""
Γ ⊢A M : hτ i
= [[Γ ⊢A M : hτ i]] α,
Γ ⊢A open M : hτ iα
(note that α does not appear in the term open M and it comes from the derivation) and
the rule for close means
##
""
Γ ⊢A M : hτ iα
α∈
/ FV(Γ, A, τ )
= Λα.[[Γ ⊢A M : hτ iα ]].
Γ ⊢A close M : hτ i
Theorem 6.5. For any λi term M , which does not contain %, if Γ ⊢A M : τ in λi , then
[[Γ]] ⊢A [[M ]] : [[τ ]].
Proof. Easy induction on the type derivation.
28
T. TSUKADA AND A. IGARASHI
Preservation of the semantics is hard to discuss precisely. First of all, the semantics
of λi is not given in [16] in spite of the subtle syntactical differences between λα and λi .
However, as far as we guess, the semantics of λi seems very close to the erasure semantics
in Section 4.3, and then we expect to have preservation of the semantics.
7. Related Work
Multi-Stage Calculi Based on Modal Logics and Their Extensions. Our work can
be considered a generalization of the previous work on the Curry-Howard isomorphism
between multi-stage calculi and modal logics [7, 8, 10]. As we have seen in Section 6, there
are embeddings from λ and λ to λ⊲ .
The restriction of λ that all code be closed precludes the definition of a code generator
like power∀ , which generates both efficient and runnable code. Nanevski and Pfenning [21]
have extended λ with the notion of names, similar to the symbols in Lisp, and remedied
the defect of λ by allowing newly generated names (not variables) to appear in closed
code.
Taha and Sheard [5] added run and CSP to λ and developed MetaML, but its type
system was not strong enough—run may fail at run-time. Then, Moggi, Taha, Benaissa,
and Sheard [14] developed the calculus AIM (“An Idealized MetaML”), in which there are
types for both open and closed code; it was then simplified to λBN , which replaced closed
code types with closedness types for closed terms that are not necessarily code. Both calculi
are based on categorical models and have sound type systems. The notion of α-closedness
in λα can be considered a generalization of λBN ’s closed types. In fact, the typing rule for
run in λBN is similar to the one in λα . Although some of these calculi have sound type
systems, it is hard to regard them as logic, mainly due to the presence of CSP, which delays
the stage of the type judgment to any later stage, and the typing rule for run (as discussed
in Section 6.3).
More recently, Yuse and Igarashi have proposed the calculus λ [10] by combining
λ and λ , while maintaining the Curry-Howard isomorphism. The main idea was to
consider LTL with modalities “always” () and “next” ( ), which represent closed and
open code types, respectively. It is similar to AIM in this respect. Although λ is based
on logic, it cannot be embedded into λ⊲ simply by combining the two embeddings above.
In fact, in λ , both directions of τ ↔ τ are provable, whereas neither direction
of (∀α. ⊲α ⊲β τ ) ↔ ⊲β ∀α. ⊲α τ is provable in λ⊲ . However, in λ it seems impossible to
program a code generator like power∀ , which generates specialized code used at any stage;
the best possible one presented can generate specialized code used only at any later stage,
so running the specialized code is not possible.
It is considered not easy to develop a sound type system for staging constructs with
side effects. Calcagno, Moggi, and Sheard developed a sound type system for a multi-stage
calculus with references using closed types [22]. It is interesting to study whether their
closedness condition can be relaxed by using α-closedness.
Other Multi-Stage Calculi. Kim, Yi, and Calcagno’s λpoly
open [11] is a rather powerful
multi-stage calculus with open and closed code fragments, intentionally variable-capturing
substitution, lifting values into code, and even references and ML-style type inference. The
type structure of λpoly
open is rather different: Since variables in code values can escape from
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
29
its initial binder’s scope and get captured by other binders, a code type records the names
of free variables and their types, as well as the type of the whole code. It is not clear
how (a pure fragment of) the calculus can be related to other foundational calculi; possible
directions may be to use the calculus of contexts [23] by Sato, Sakurai, and Kameyama,
and the contextual modal type theory by Nanevski, Pfenning, and Pientka [24].
Viera and Pardo [25] have proposed a multi-stage language with intensional code analysis, that is, pattern matching on code. The language requires typechecking at run-time.
More recently, Kameyama, Kiselyov, and Shan [26] have developed an extension of (a
two-level version of) λ with control operators shift/reset [27], which enable an interesting
pattern of code generation such as let-insertion [28] in the direct style. It will be interesting
to investigate how this calculus extends to dynamic code execution (i.e., run).
Modal Logics. As we discussed above, the -fragment of modal logic, the -fragment
of LTL can be embedded into our logic, and the -fragment of LTL and our logic is
incomparable.
Our logic has three characteristic features: (1) it is multi-modal, (2) it has universal
quantification over modalities and (3) modal operators are “relative”, meaning their semantics depends on the possible world at which they are interpreted. Most other logics do not
have all of these features.
Dynamic logic [29] is a multi-modal logic for reasoning about programs. Its modal
operators are [α] for each program α, and [α]φ means “when α halts, φ must stand after
execution of α from the current state”. Dynamic logic is multi-modal and its modal operators are “relative”, but does not have quantification over programs. Therefore, there is
no formula in Dynamic logic which would correspond to ∀α. ⊲α ⊲α φ. There is, however, a
formula which is expressive in Dynamic logic but not in our logic: e.g., a Dynamic logic
formula [α∗ ]φ, which means intuitively φ ∧ [α]φ ∧ [α][α]φ ∧ . . . , cannot be expressed in our
logic.
Hybrid logic [30] is a modal logic with a new kind of atomic formula called nominals,
each of which must be true exactly one state in any model (therefore, a nominal names a
state). For each nominal i, @i is a modal operator and @i φ means “φ stands at the state
denoted by i”. Hybrid logic has a universal quantifier over nominals (and another binder
↓: ↓ x.φ means “let x stand for the nominal for the current world, then φ stands”). Hybrid
logic differs from our logic, in that modal operators @i indicate worlds directly, hence are
not “relative”. In Hybrid logic, @i @j φ ↔ @j φ, but ⊲α ⊲β φ and ⊲β φ are not equivalent in
our logic.
8. Conclusion and Future Work
We have studied a logical aspect of environment classifiers by developing a simply typed
multi-stage calculus λ⊲ with environment classifiers. This calculus corresponds to a multimodal logic with quantifier over transitions by the Curry-Howard isomorphism and satisfies
time-ordered normalization as well as basic properties such as subject reduction, confluence,
and strong normalization. The classical proof system is sound and complete with respect
to the Kripke semantics. Our calculus simplifies the previous calculus λα of environment
classifiers by reducing run and some uses of CSP to an extension of another construct. We
believe our work helps clarify the semantics of environment classifiers.
30
T. TSUKADA AND A. IGARASHI
We have also studied evaluation of (a slight extension of) λ⊲ and shown staged execution
of a program is possible. Also, it is shown that erasure execution is possible for a certain
subset, which is close to λi , the basis of MetaOCaml.
From a theoretical perspective, it is interesting to study the semantics of the intuitionistic version of the logic, as mentioned earlier, and also the calculus corresponding to the
classical version of the logic. It is known that the naive combination of staging constructs
and control operators is problematic since bound variables in quotation may escape from
its scope by a control operator. We expect that a logical analysis, like the one presented
here and Reed and Pfenning [31], will help analyze the problem.
From a practical perspective, one feature missing from λ⊲ is CSP for all types. As
argued in the introduction, we think typical use of CSP is rather limited and so easy to
support. Type inference for full λ⊲ would not be possible for the same reason as λα [16].
However, it would be easy to applying type inference for λilet [16] to a similar subset of λ⊲ .
Acknowledgments. This work was begun while the first author was at Kyoto University.
We would like to thank Lintaro Ina, Naoki Kobayashi, Ryosuke Sato, Naokata Shikuma,
and anonymous reviewers for useful comments. This work was supported in part by Grantin-Aid for Scientific Research No. 21300005.
References
[1] Jones, N.D., Gomard, C.K., Sestoft, P.: Partial Evaluation and Automatic Program Generation.
Prentice-Hall (1993)
[2] Consel, C., Lawall, J.L., Meur, A.F.L.: A tour of Tempo: A program specializer for the C language.
Science of Computer Programming 52(1-3) (2004) 341–370
[3] Wickline, P., Lee, P., Pfenning, F.: Run-time code generation and Modal-ML. In: Proceedings of
1998 ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI’98),
Montreal, Quebec, Canada (1998) 224–235
[4] Poletto, M., Hsieh, W.C., Engler, D.R., Kaashoek, M.F.: ‘C and tcc: A language and compiler for
dynamic code generation. ACM Transactions on Programming Languages and Systems 21(2) (1999)
324–369
[5] Taha, W., Sheard, T.: MetaML and multi-stage programming with explicit annotations. Theoretical
Computer Science 248 (2000) 211–242
[6] Glück, R., Jørgensen, J.: Efficient multi-level generating extensions for program specialization. In:
Proceedings of Programming Languages, Implementations, Logics and Programs (PLILP’95). Volume
982 of LNCS. (1995) 259–278
[7] Davies, R.: A temporal-logic approach to binding-time analysis. In: Proceedings of the Eleventh Annual
IEEE Symposium on Logic in Computer Science (LICS 1996), IEEE Computer Society Press (1996)
184–195
[8] Davies, R., Pfenning, F.: A modal analysis of staged computation. J. ACM 48(3) (2001) 555–604
[9] Taha, W., Nielsen, M.F.: Environment classifiers. In: Proceedings of the 30th ACM SIGPLAN-SIGACT
Symposium on Principles of Programming Languages (POPL ’03), New York, NY, USA, ACM (2003)
26–37
[10] Yuse, Y., Igarashi, A.: A modal type system for multi-level generating extensions with persistent code.
In: Proceedings of Eighth ACM-SIGPLAN International Symposium on Principles and Practice of
Declarative Programming (PPDP’06). (2006) 201–212
[11] Kim, I.S., Yi, K., Calcagno, C.: A polymorphic modal type system for lisp-like multi-staged languages.
In: Proceedings of the 33rd ACM SIGPLAN-SIGACT symposium on Principles of programming languages (POPL ’06), New York, NY, USA, ACM (2006) 257–268
[12] Davies, R., Pfenning, F.: A modal analysis of staged computation. In: Proceedings of the 23rd ACM
SIGPLAN-SIGACT symposium on Principles of programming languages (POPL ’96). (1996) 258–270
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
31
[13] Stirling, C.: Modal and temporal logics. In: Handbook of Logic in Computer Science. Volume 2. Oxford
University Press (1992) 477–563
[14] Moggi, E., Taha, W., Benaissa, Z.E.A., Sheard, T.: An idealized MetaML: Simpler, and more expressive.
In: Proceedings of European Symposium on Programming (ESOP’99). Volume 1576 of LNCS. (1999)
193–207
[15] Benaissa, Z.E.A., Moggi, E., Taha, W., Sheard, T.: Logical modalities and multi-stage programming.
In: Proceedings of Workshop on Intuitionistic Modal Logics and Applications (IMLA’99). (1999)
[16] Calcagno, C., Moggi, E., Taha, W.: ML-like inference for classifiers. In: Proceedings of European
Symposium on Programming (ESOP ’04), Springer-Verlag (2004) 79–93
[17] Plotkin, G., Stirling, C.: A framework for intuitionistic modal logics. In: Proceedings of Theoretical
Aspects of Reasoning about Knowledge, Monterey, CA, Morgan Kaufmann (1986)
[18] Kojima, K., Igarashi, A.: On constructive linear-time temporal logic. In: Proceedings of Intuitionistic
Modal Logics and Applications Workshop (IMLA ’08), Pittsburgh, PA (2008)
[19] Kojima, K., Igarashi, A.: Constructive linear-time temporal logic: Proof systems and Kripke semantics.
Information and Computation (2010) Accepted for publication.
[20] Takahashi, M.: Parallel reductions in λ-calculus. Information and Computation 118(1) (1995) 120–127
[21] Nanevski, A., Pfenning, F.: Staged computation with names and necessity. Journal of Functional
Programming 15(5) (2005) 893–939
[22] Calcagno, C., Moggi, E., Sheard, T.: Closed types for a safe imperative MetaML. Journal of Functional
Programming 13(3) (2003) 545–571
[23] Sato, M., Sakurai, T., Kameyama, Y.: A simply typed context calculus with first-class environments.
Journal of Functional and Logic Programming 2002(4) (2002) 1–41
[24] Nanevski, A., Pfenning, F., Pientka, B.: Contextual modal type theory. ACM Transactions on Computational Logic 9(3) (2008)
[25] Viera, M., Pardo, A.: A multi-stage language with intensional analysis. In: Proceedings of Fifth International Conference on Generative Programming and Component Engineering (GPCE’06), Portland,
OR (2006) 11–20
[26] Kameyama, Y., Kiselyov, O., chieh Shan, C.: Shifting the stage: Staging with delimited control. In:
Proceedings of ACM Symposium on Partial Evaluation and Semantics-Based Program Manipulation
(PEPM’09), Savannah, GA (2009)
[27] Danvy, O., Filinski, A.: Abstracting control. In: Proceedings of ACM Conference on Lisp and Functional
Programming. (1990) 151–160
[28] Lawall, J.L., Danvy, O.: Continuation-based partial evaluation. In: Proceedings of ACM Conference on
Lisp and Functional Programming. (1994) 227–238
[29] Harel, D., Kozen, D., Tiuryn, J.: Dynamic logic. In Gabbay, D., Guenther, F., eds.: Handbook of
Philosophical Logic. Volume 4. 2nd edn. Springer-Verlag (2002) 99–218
[30] Areces, C., ten Cate, B.: Hybrid logics. In Blackburn, P., Wolter, F., van Benthem, J., eds.: Handbook
of Modal Logics. Elsevier (2006) 821–868
[31] Reed, J., Pfenning, F.: Intuitionistic letcc via labelled deduction. In: Proceedings of Workshop on
Methods for Modalities (M4M). (2007)
32
T. TSUKADA AND A. IGARASHI
Appendix A. Proof of Time Ordered Normalization (Theorem 3.10)
First, we give an inductive characterization of the class of T -normal terms. A judgment
of the form ∆ ⊢⇓T M , where ∆ is a finite set of transition variables, means “M is T -normal
under bound transition variables ∆.” We distinguish transition variables bound outside of
M since they are interpreted as the empty transition in annotated reduction. We also need
an auxiliary judgment of the form ∆ ⊢ ▽T M , read “M is T -neutral under bound transition
variables ∆”. We write T [∆ := ε] for T [α1 := ε] . . . [αn := ε], where ∆ = {α1 , . . . , αn }. The
inference rules for these judgments are shown in Figure 11.
We first prove that this proof system characterize T -normal forms as in Lemma A.2,
which is obtained as a special case of the lemma below. In what follows, FTV(T ) denotes
the set of transition variables in the path T .
Lemma A.1. Let M be a typable term, T be a path, ∆ be a finite set of transition variables.
Then, the following two conditions are equivalent:
U
(1) For any term N and path U , if M −→ N , then U [∆ := ε] T [∆ := ε].
(2) ∆ ⊢⇓T M .
Proof. We show (1) =⇒ (2) by induction on M .
• Case M = x: ∆ ⊢⇓T M trivially holds.
U
• Case M = ◮α M ′ : We first show that M ′ −→ N ′ implies U [∆ := ε] (α−1 T )[∆ := ε]
αU
U
for any N ′ and U . Assume M ′ −→ N ′ . Then, we have ◮α M ′ −→ ◮α N ′ . By (1), we have
(αU )[∆ := ε] T [∆ := ε] and then U [∆ := ε] (α−1 T )[∆ := ε]. m By the induction
−1
hypothesis, we have ∆ ⊢⇓α T M ′ , so ∆ ⊢⇓T ◮α M ′ .
• Case M = ◭α M ′ : Similarly to the case above, we have ∆ ⊢⇓αT M ′ . What remains
to show is ∆ ⊢ ▽αT M ′ .
Since M is typable, a possible form of M ′ is a variable, an application, an instantiation,
or ◮α M ′′ for some M ′′ . In the first three cases, ∆ ⊢ ▽αT M ′ is trivial. If M = ◮α M ′′ ,
α−1
then we have M ′ −→ M ′′ . Then, by (1), we have α−1 [∆ := ε] 6≤ T [∆ := ε] and then
ε 6≤ (αT )[∆ := ε]. By the first rule for · ⊢ ▽(·) ·, we have ∆ ⊢ ▽αT M ′ .
• Case M = Λα.M ′ :
ε T [∆ := ε]
∆ ⊢ ▽T M
∆ ⊢⇓T x
∆ ⊢⇓T M
∆ ⊢⇓T λx : τ.M
∆ ⊢⇓T M
∆ ⊢⇓αT ◮α M
M = x or M0 M1 or M0 A
∆ ⊢ ▽T M
∆ ⊢⇓T M
∆ ⊢ ▽T M
∆ ⊢⇓T M N
∆ ⊢⇓T N
∆ ⊢⇓T M
∆ ⊢ ▽T M
−1
∆ ⊢⇓α T ◭α M
∆, α ⊢⇓T M
α∈
/ FTV(T ) ∪ ∆
T
∆ ⊢⇓ Λα.M
∆ ⊢⇓T M
∆ ⊢ ▽T M
∆ ⊢⇓T M A
Figure 11: T -neutral terms and T -normal terms.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
33
U [α:=ε]
U
/ FTV(T ) ∪ ∆. Then, we have Λα.M ′ −→ Λα.N ′ . By
Assume M ′ −→ N ′ and α ∈
(1), (U [α := ε])[∆ := ε] T [∆ := ε]. We have (U [α := ε])[∆ := ε] = U [∆, α := ε] and
T [∆, α := ε] = T [∆ := ε] because α ∈
/ FTV(T ). So, U [∆, α := ε] T [∆, α := ε]. Then, we
have ∆, α ⊢⇓T M ′ by the induction hypothesis. So, ∆ ⊢⇓T Λα.M ′ as required.
Other cases are similar.
The proof of (2) ⇒ (1) is by easy induction on the structure of the derivation ∆ ⊢⇓T M .
We show the case M = ◭α M ′ as a representative case.
Assume ∆ ⊢⇓T ◭α M ′ . By definition, we have ∆ ⊢⇓αT M ′ and ∆ ⊢ ▽αT M ′ . Assume
U
◭α M ′ −→ N . There are two cases.
U
αU
• Case ◭α M ′ −→ ◭α N ′ with M ′ −→ N ′ : Because ∆ ⊢⇓αT M ′ , by induction hypothesis we have (αU )[∆ := ε] (αT )[∆ := ε]. This implies that U [∆ := ε] T [∆ := ε].
α−1
• Case M ′ = ◮α M ′′ and ◭α ◮α M ′′ −→ M ′′ : Since ∆ ⊢ ▽αT ◮α M ′′ , we have ε
(αT )[∆ := ε]. This equation is equivalent to α−1 [∆ := ε] T [∆ := ε].
Lemma A.2. Suppose M is typable. Then, M is T -normal if and only if ∅ ⊢⇓T M .
Proof. Immediate from Lemma A.1.
Now, we prove that reduction preserves T -normality and T -neutrality, as in Lemma A.5,
from which Theorem 3.10 immediately follows. Before that, we prove that (term and
transition) substitution preserves T -normality and T -neutrality under a certain condition.
Lemma A.3. Suppose Γ, x : σ@B ⊢A M : τ and Γ ⊢B N : σ and B[∆ := ε] (AT )[∆ := ε].
−1
If ∆ ⊢⇓T M and ∆ ⊢⇓B AT N , then ∆ ⊢⇓T (M [x := N ]). Similarly, if ∆ ⊢ ▽T M , then
∆ ⊢ ▽T (M [x := N ]).
Proof. By induction on M . We show only the main cases.
• Case M = y: The subcase x 6= y is trivial. Assume x = y. Because Γ, x : σ@B ⊢A x :
τ , it is the case that A = B. So B −1 AT = T and ∆ ⊢⇓T N . Moreover, we have ∆ ⊢ ▽T N
because ε T [∆ := ε], which follows from B[∆ := ε] (AT )[∆ := ε] and B = A.
• Case M = ◮α M ′ : From the type derivation of M , we have τ = ⊲α τ0 and Γ, x :
−1
σ@B ⊢Aα M ′ : τ0 . From ∆ ⊢⇓T ◮α M ′ , we obtain ∆ ⊢⇓α T M ′ . Moreover, B[∆ := ε]
(AT )[∆ := ε] = ((Aα)(α−1 T ))[∆ := ε] holds. By the induction hypothesis, we now have
−1
∆ ⊢⇓α T (M ′ [x := N ]). Therefore ∆ ⊢⇓T (◮α M ′ )[x := N ].
If ∆ ⊢ ▽T ◮α M ′ , it is the case that ε T [∆ := ε]. Therefore ∆ ⊢ ▽T (◮α M ′ )[x := N ].
• Case M = Λα.M ′ : Without loss of generality, we can assume that α ∈
/ FTV(A) ∪
FTV(N ) ∪ FTV(Γ) ∪ FTV(T ) ∪ FTV(B) ∪ ∆. From the type derivation of M , we have τ =
∀α.τ0 and Γ, x : σ@B ⊢A M ′ : τ0 . From ∆ ⊢⇓T M , we also have ∆, α ⊢⇓T M ′ . Because α
is fresh and B[∆ := ε] (AT )[∆ := ε], we have B[∆, α := ε] (AT )[∆, α := ε]. By the induction hypothesis, we have ∆, α ⊢⇓T M ′ [x := N ], which implies ∆ ⊢⇓T (Λα.M ′ )[x := N].
If ∆ ⊢ ▽T Λα.M ′ , it is the case that ε T [∆ := ε]. Therefore ∆ ⊢ ▽T (Λα.M ′ )[x := N ].
34
T. TSUKADA AND A. IGARASHI
Lemma A.4. Suppose Γ ⊢A M : τ and α ∈
/ FTV(AT ).
T
T
[α:=B]
(1) If ∆, α ⊢⇓ M , then ∆ ⊢⇓
(M [α := B]); and
(2) if ∆, α ⊢ ▽T M , then ∆ ⊢ ▽T [α:=B] (M [α := B]).
Proof. We first show (2) by case analysis on the last rule used to derive ∆, α ⊢ ▽T M .
• Case ∆, α ⊢ ▽T M with ε T [∆, α := ε]: Because α ∈
/ FTV(AT ) and ε ≤ A,
we have A = A′ αn and T = α−n T ′ and α ∈
/ FTV(A′ ) ∪ FTV(T ′ ) for some n ≥ 0.
Then, ε T [∆, α := ε] = (α−n T ′ )[∆, α := ε] = T ′ [∆ := ε]. Therefore B −n [∆ := ε]
(B −n [∆ := ε])(T ′ [∆ := ε]) = T [α := B][∆ := ε]. It follows that ε T [α := B][∆ := ε] because B −n [∆ := ε] ≤ ε. So we have ∆ ⊢ ▽T [α:=B] (M [α := B]).
• The other case is easy.
Then, (1) is proved by induction on the derivation of ∆, α ⊢⇓T M . We show only the main
cases.
• Case ∆, α ⊢⇓T N0 N1 : Applying the induction hypothesis and (2) to, respectively,
∆ ⊢⇓T [α:=B] (N0 [α := B])
∆, α ⊢⇓T N0
∆, α ⊢ ▽T N0
∆ ⊢ ▽T [α:=B] (N0 [α := B])
results in
∆, α ⊢⇓T N1
∆ ⊢⇓T [α:=B] (N1 [α := B]).
So we obtain ∆ ⊢⇓T [α:=B] (N0 N1 )[α := B].
• Case ∆, α ⊢⇓T ◭β M ′ : It is the case that ∆, α ⊢⇓βT M ′ and ∆, α ⊢ ▽βT M ′ . Because
′
/ FTV(AT ) = FTV(A′ βT ),
Γ ⊢A ◭β M ′ : τ , we have A = A′ β and Γ ⊢A M ′ : ⊲β τ . Since α ∈
by applying the induction hypothesis and (2), we have ∆ ⊢⇓(βT )[α:=B] M ′ [α := B] and
∆ ⊢ ▽(βT )[α:=B] M ′ [α := B]. Therefore ∆ ⊢⇓T [α:=B] (◮β[α:=B] (M ′ [α := B])).
U
Lemma A.5. Suppose Γ ⊢A M : τ and M −→ N and U [∆ := ε] T [∆ := ε]. If ∆ ⊢⇓T M ,
then ∆ ⊢⇓T N . Similarly, if ∆ ⊢ ▽T M , then ∆ ⊢ ▽T N .
U
Proof. By induction on the derivation of M −→ N .
ε
• Case M = (λx : σ.P ) Q −→ P [x := Q] = N : It is the case that U = ε and so
ε T [∆ := ε]. Assume ∆ ⊢⇓T M . Then we have ∆ ⊢⇓T P and ∆ ⊢ ▽T P and ∆ ⊢⇓T Q.
From ε T [∆ := ε], it follows that A[∆ := ε] (AT )[∆ := ε]. By Lemma A.3, we have
∆ ⊢⇓T (P [x := Q]). Moreover, we have ∆ ⊢ ▽T N because U [∆ := ε] = ε T [∆ := ε].
ε
• Case M = (Λα.P ) B −→ P [α := B] = N : We can assume without loss of generality
that α ∈
/ FTV(Γ) ∪ FTV(A) ∪ FTV(T ) ∪ ∆. Assume ∆ ⊢⇓T M . Then we have ∆, α ⊢⇓T P .
Because α ∈
/ FTV(AT ), by applying Lemma A.4, we have ∆ ⊢⇓T [α:=B] P [α := B]. Because
α∈
/ FTV(T ), T [α := B] = T . So we have ∆ ⊢⇓T (P [α := B]).
The proof of ∆ ⊢ ▽T N is similar to the prior case since U = ε.
U
α−1 U
• Case M = ◮α P −→ ◮α P ′ = N with P −→ P ′ : Assume ∆ ⊢⇓T ◮α P . Then
−1
∆ ⊢⇓α T P . The typability of M implies the typability of P . We have (α−1 U )[∆ := ε]
(α−1 T )[∆ := ε] because U [∆ := ε] T [∆ := ε]. So, by the induction hypothesis, we obtain
−1
∆ ⊢⇓α T P ′ . Therefore ∆ ⊢⇓T N .
Assume ∆ ⊢ ▽T ◮α P . It must be the case that ε T [∆ := ε]. Therefore ∆ ⊢ ▽T ◮α P ′ .
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
35
U′
U
• Case M = Λα.P −→ Λα.P ′ = N with P −→ P ′ and U = U ′ [α := ε]: We can assume
without loss of generality that α ∈
/ FTV(Γ)∪FTV(A)∪FTV(T )∪∆. Moreover α ∈
/ FTV(U )
because U = U ′ [α := ε]. It follows that
U ′ [∆, α := ε] = (U ′ [α := ε])[∆ := ε]
(by the definition of U )
= U [∆ := ε]
(by assumption)
T [∆ := ε]
(because α ∈
/ FTV(T ))
= T [∆, α := ε]
Assume that ∆ ⊢⇓T M . Then ∆, α ⊢⇓T P . The typability of M implies the typability
of P . So, by the induction hypothesis, we have ∆ ⊢⇓T P ′ , which implies ∆ ⊢⇓T N .
The proof of the remaining part is similar to other cases.
Appendix B. List of Error-generating and Error-propagating Rules
The List of Error-generating Rules.
⊢ε M ⇓ M ′
⊢ε M ⇓ M ′
M′ ∈ Z
M′ ∈
/Z
♦ ∈ {+, −, ∗, =}
ε
⊢ M ♦N ⇓ err
⊢ε N ⇓ N ′
N′ ∈
/Z
♦ ∈ {+, −, ∗, =}
ε
⊢ M ♦N ⇓ err
⊢ε M ⇓ M ′
M′ ∈
/ {true, false}
ε
⊢ if M then N1 else N2 ⇓ err
⊢ε M ⇓ M ′
M ′ 6= λx : τ.M ′′
⊢ε M N ⇓ err
⊢ε x ⇓ err
⊢ε M ⇓ M ′
M ′ 6= Λα.M ′′
⊢ε M A ⇓ err
⊢ε M ⇓ M ′
M ′ 6= ◮α M ′′
α
⊢ ◭α M ⇓ err
⊢ε ◭α M ⇓ err
The List of Error-propagating Rules.
⊢A M ⇓ err
♦ ∈ {+, −, ∗, =}
A
⊢ M ♦N ⇓ err
⊢A M ⇓ M ′
⊢A M ⇓ err
⊢A if M then N1 else N2 ⇓ err
⊢ε M ⇓ false
⊢ε N2 ⇓ err
ε
⊢ if M then N1 else N2 ⇓ err
⊢A M ⇓ M ′
⊢A N ⇓ err
♦ ∈ {+, −, ∗, =}
A
⊢ M ♦N ⇓ err
⊢ε M ⇓ true
⊢ε N1 ⇓ err
⊢ε if M then N1 else N2 ⇓ err
⊢A M ⇓ M ′
⊢A N1 ⇓ err
A 6= ε
A
⊢ if M then N1 else N2 ⇓ err
⊢A N1′ ⇓ N1′
⊢A N2 ⇓ err
⊢A if M then N1 else N2 ⇓ err
A 6= ε
⊢A M ⇓ err
A 6= ε
A
⊢ λx : τ.M ⇓ err
36
T. TSUKADA AND A. IGARASHI
⊢A M ⇓ err
⊢A M N ⇓ err
⊢A M ⇓ M ′
⊢A N ⇓ err
⊢A M N ⇓ err
⊢A M ⇓ err
◭α M ⇓ err
⊢Aα
⊢A M ⇓ err
Λα.M ⇓ err
⊢A
⊢ε M [f := fix f : τ → σ.M ] ⇓ err
⊢ε fix f : τ → σ.M ⇓ err
⊢Aα M ⇓ err
⊢A ◮α M ⇓ err
⊢A M ⇓ err
⊢A M B ⇓ err
⊢A M ⇓ err
A 6= ε
A
⊢ fix f : τ → σ.M ⇓ err
Appendix C. Proofs of Properties about mini-ML
C.1. Proof of Theorem 4.4 (Type Soundness of MiniML⊲ ).
Lemma C.1. If Γ ⊢αA v : τ and v ∈ V αA , then Γ−α ⊢A v : τ .
Proof. By induction on the structure of v. We show only cases for v = x and v = ◭β v ′ .
• Case v = x: We have x : τ @αA ∈ Γ. Because x : τ @A ∈ Γ−α , Γ−α ⊢A x : τ .
• Case v = ◭β v ′ : We have A 6= ε because ◭β v ′ ∈ V αA . That means A = A′ β for some
′
′
A′ and Γ ⊢αA β ◭β v ′ : τ . So we have Γ ⊢αA v ′ : ⊲β τ . By the induction hypothesis,
′
′
Γ−α ⊢A v ′ : ⊲β τ . Therefore, by (◭), we have Γ−α ⊢A β ◭β v ′ : τ .
Proof of Theorem 4.4. The first part of the theorem
If Γ is ε-free and Γ ⊢A M : τ and ⊢A M ⇓ R, then R = v and v ∈ V A for
some v and Γ ⊢A v : τ .
is proved by induction on the derivation of ⊢A M ⇓ R with case analysis on the form of M
and the last rule used to derive ⊢A M ⇓ R.
We only show representative cases here.
• Case M = x
(A 6= ε)
− Subcase A
⊢ x⇓x
(1) Immediate.
− Subcase ε
⊢ x ⇓ err
(1) We have x : τ @ε ∈ Γ because Γ ⊢ε x : τ but this contradicts the assumption that
Γ is ε-free.
• Case M = ◭α N
⊢ε N ⇓ ◮α N ′
− Subcase α
⊢ ◭α N ⇓ N ′
(1) We have Γ ⊢ε N : ⊲α τ because Γ ⊢ε ◭α N .
(2) By the induction hypothesis, Γ ⊢ε ◮α N ′ : ⊲α τ and ◮α N ′ ∈ V ε .
(3) So we have Γ ⊢α N ′ : τ and N ′ ∈ V α .
⊢A N ⇓ N ′
(A 6= ε)
− Subcase Aα
⊢ ◭α N ⇓ ◭α N ′
(1) We have Γ ⊢A N : ⊲α τ because Γ ⊢A ◭α N : τ .
(2) By the induction hypothesis, Γ ⊢A N ′ : ⊲α τ and N ′ ∈ V A .
(3) Therefore, Γ ⊢Aα ◭α N ′ : τ by (◭) and ◭α N ′ ∈ V Aα .
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
37
⊢ε N ⇓ N ′
N ′ 6= ◮α N ′′
⊢α ◭α N ⇓ err
(1) We have Γ ⊢ε N : ⊲α τ because Γ ⊢ε ◭α N : τ .
(2) By the induction hypothesis, Γ ⊢ε N ′ : ⊲α τ and N ′ ∈ V ε , which is a contradiction.
• Case M = Λα.N
⊢A N ⇓ N ′
− Subcase A
⊢ Λα.N ⇓ Λα.N ′
(1) We have τ = ∀α.τ ′ and Γ ⊢A N : τ ′ for some τ ′ and α ∈
/ FTV(Γ) ∪ FTV(A)
A
because Γ ⊢ Λα.N : τ .
(2) By the induction hypothesis, Γ ⊢A N ′ : τ ′ and N ′ ∈ V A .
(3) Therefore, Γ ⊢A Λα.N ′ : τ by (Gen) and ΛαN ′ ∈ V A .
• Case M = N B
⊢ε N ⇓ Λα.N ′
⊢ε N ′ [α := B] ⇓ M
− Subcase
ε
⊢ NB⇓M
(1) We have τ = σ[α := B] and Γ ⊢ε N : ∀α.σ for some σ and α, because Γ ⊢ε N B : τ .
(2) By the induction hypothesis, Γ ⊢ε Λα.N ′ : ∀α.σ. Therefore Γ ⊢ε N ′ : σ and
α∈
/ FTV(Γ).
(3) By Substitution Lemma (Lemma 3.1) and α ∈
/ FTV(Γ), Γ ⊢ε N ′ [α := B] : τ .
(4) By the induction hypothesis, Γ ⊢ε M : τ and M ∈ V ε .
⊢A N ⇓ N ′
(A 6= ε)
− Subcase A
⊢ N B ⇓ N′ B
(1) We have τ = σ[α := B] and Γ ⊢A N : ∀α.σ for some σ and α, because Γ ⊢A N B : τ .
(2) By the induction hypothesis, Γ ⊢A N ′ : ∀α.σ and N ′ ∈ V A .
(3) Therefore, Γ ⊢A N ′ B : τ by (Ins) and N ′ B ∈ V A .
⊢ε N ⇓ N ′
N ′ 6= Λα.N ′′
− Subcase
⊢ε N B ⇓ err
(1) We have τ = σ[α := B] and Γ ⊢ε N : ∀α.σ for some σ and α, because Γ ⊢ε N B : τ .
(2) By the induction hypothesis, Γ ⊢ε N ′ : ∀α.σ and N ′ ∈ V A , which is a contradiction.
Now, we prove the second part. By the first part, Γ ⊢ε v : ⊲α τ and v ∈ V ε . Therefore
v = ◮α v ′ and v ′ ∈ V α . From the typing rules, we have Γ ⊢α v ′ : τ . Then, we have
Γ−α ⊢ε v ′ : τ by Lemma C.1.
− Subcase
C.2. Proof of Theorem 4.5. We first prove type soundness for the new type system
(Lemma C.8). Although the new type system identifies a subset of well-typed λ⊲ terms,
it has to be proved again, since we want to gurantee that the evaluation of a term typed
under the new type system results in (if converges) a value that is also typed under it. We
start with proving various substitution lemmas.
Lemma C.2 (Term Substitution Preserves Typing). If Γ, x : σ@B; ∆ ⊢A
s M : τ and
B
A
Γ; ∆ ⊢s M : σ, then Γ; ∆ ⊢s M [x := N ] : τ .
Proof. Similar to the proof of Lemma 3.1.
Next, we show that substitution for transition variables preserves various kinds of wellformedness and typing.
Lemma C.3. If ∆, α@B ⊢s A and ∆ ⊢s BC, then ∆[α := C] ⊢s A[α := C].
38
T. TSUKADA AND A. IGARASHI
Proof. By induction of the length of A. The base case is trivial because ∆ ⊢s ε for any ∆.
So, we show the induction step.
Let A = α1 . . . αn . Assume ∆, α@B ⊢s A. It is easy to see that ∆, α@B ⊢s α1 . . . αn
implies ∆, α@B ⊢s α1 . . . αn−1 . Then ∆[α := C] ⊢s (α1 . . . αn−1 )[α := C] by the induction
hypothesis. We have αn @α1 . . . αn−1 ∈ (∆, α@B) from the definition of ∆, α@B ⊢s A. The
proof is divided into two cases.
• Case αn 6= α: It is the case that αn @α1 . . . αn−1 ∈ ∆. The conclusion is trivial,
because αn @(α1 . . . αn−1 )[α := C] ∈ ∆[α := C].
• Case αn = α: It is the case that α1 . . . αn−1 = B. Let C = β1 . . . βm . Because
∆ ⊢s BC, we have
∆0 = {α1 @ε, α2 @α1 , . . . , αn−1 @α1 . . . αn−2 , β1 @B, β2 @Bβ1 , . . . , βm @Bβ1 . . . βm−1 } ⊆ ∆.
Let ∆1 = ∆ − ∆0 . Note that α ∈
/ {α1 , . . . , αn−1 , β1 , . . . , βm } = FTV(∆0 ) and ∆0 ⊢s BC.
Therefore we have ∆[α := C] = ∆0 [α := C] ∪ ∆1 [α := C] = ∆0 ∪ ∆1 [α := C]. Because
(Bα)[α := C] = BC and ∆0 ⊢s BC, we obtain ∆[α := C] ⊢s A[α := C].
Lemma C.4. If ⊢s ∆, α@B and ∆ ⊢s BC, then ⊢s ∆[α := C].
Proof. Let β@A ∈ ∆[α := C]. We show ∆[α := C] ⊢s A.
There exists some A′ such that β@A′ ∈ ∆ and A = A′ [α := C] because β@A ∈
∆[α := C]. Since ⊢s ∆, α@B, we have ∆, α@B ⊢s A′ . So by Lemma C.3, we have
∆[α := C] ⊢s A′ [α := C] as required.
A[α:=C]
Lemma C.5. If ∆, α@B ⊢A
s τ and ∆ ⊢s BC, then ∆[α := C] ⊢s
τ [α := C].
Proof. By induction on the structure of the type τ .
• Case τ = int: Because ∆, α@B ⊢A
s int, we have ∆, α@B ⊢s A. So by Lemma C.3, we
A[α:=C]
obtain ∆[α := C] ⊢s A[α := C]. Therefore we obtain ∆[α := C] ⊢s
int by applying
(ST-Base).
Aβ
• Case τ = ⊲β σ: Because ∆, α@B ⊢A
s ⊲β σ, we have ∆, α@B ⊢s σ. By the induction
(Aβ)[α:=C]
hypothesis, we have ∆[α := C] ⊢s
σ[α := C]. Therefore by applying (ST-Code)
A[α:=C]
as many times as needed, we obtain ∆[α := c] ⊢s
⊲β[α:=C] (σ[α := C]).
′
′
A
• Case τ = ∀β@A′ .σ: Because ∆, α@B ⊢A
s ∀β@A .σ, we have ∆, α@B, β@AA ⊢s
A[α:=C]
σ. By the induction hypothesis, we have ∆[α := C], β@(AA′ )[α := C] ⊢s
σ[α := C].
A[α:=C]
′
Hence by applying (ST-Univ), we obtain ∆[α := C] ⊢s
∀β@(A [α := C]).(σ[α := C]).
Lemma C.6. If ∆, α@B ⊢s Γ and ∆ ⊢s BC, then ∆[α := C] ⊢s Γ[α := C].
Proof. Take x : τ @A ∈ Γ. By the well-formedness of Γ, we have ∆, α@B ⊢A
s τ . By
A[α:=C]
Lemma C.5, we have ∆[α := C] ⊢s
τ [α := C], as required.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
39
Now we show the substitution lemma for transition variables. However, transition
substitution does not preserve typing in general, and, even worse, transition substitution is
undefined in some cases. This is because of the distinction between two kinds of applications
M [α] and M A. For example, if we substitute A for α into M [α], the result is M [A], which
is not a valid term.
So, the substitution lemma we will prove is a restricted one. Transition substitution
preserves typing in the following two cases:
(1) Substitution of a single transition variable, M [α := β] for any M .
(2) Substitution of an arbitrary transition A at the initial stage in a value, v[α := A], where
v is a value (an element of V ε ) and α@ε.
They are what we need to prove type soundness.
Lemma C.7.
A[α:=β]
(1) If Γ; ∆, α@B ⊢A
M [α := β] :
s M : τ and β@B ∈ ∆, then Γ[α := β]; ∆[α := β] ⊢s
τ [α := β].
A[α:=B]
(2) For any v ∈ V A , if Γ; ∆, α@ε ⊢A
s v : τ and ∆ ⊢s B, then Γ[α := B]; ∆[α := B] ⊢s
v[α := B] : τ [α := B].
Proof. We can prove (1) by induction on the type derivation.
• Case (S-Var): We have M = x and ∆, α@B ⊢s Γ, x : τ @A. By Lemma C.6, we have
∆[α := β] ⊢s Γ[α := β], x : τ [α := β]@A[α := β]. By applying the (S-Var) rule, we obtain
A[α:=β]
Γ[α := β]; ∆[α := β] ⊢s
x : τ [α := β].
• Case (S-Gen): We have M = Λγ.M ′ and τ = ∀γ@C.σ and Γ; ∆, α@B, γ@AC ⊢A
s
A[α:=β]
: σ. By the induction hypothesis, we have Γ[α := β]; ∆[α := β], γ@(AC)[α := β] ⊢s
M ′ [α := β] : σ[α := β] (note that α 6= γ). By applying the (S-Gen) rule, we obtain
A[α:=β]
Γ[α := β]; ∆[α := β] ⊢s
Λγ.(M ′ [α := β]) : ∀γ@(C[α := β]).σ[α := β].
M′
• Case (S-Ins2): We have M = M ′ [γ1 ], and there is some σ and γ0 and C such
′
that τ = σ[γ0 := γ1 ] and Γ; ∆, α@B ⊢A
s M : ∀γ0 @C.σ and γ1 @AC ∈ ∆. We can assume
without loss of generality γ0 ∈
/ FTV(C) ∪ {α, β}. By the induction hypothesis, we have
A[α:=β]
Γ[α := β]; ∆[α := β] ⊢s
M ′ [α := β] : (∀γ0 @C.σ)[α := β]. Since (∀γ0 @C.σ)[α := β] =
∀γ0 @(C[α := β]).(σ[α := β]) and γ1 [α := β]@(AC)[α := β] ∈ ∆[α := β], we can apply (SA[α:=β]
Ins2) to obtain Γ[α := β]; ∆[α := β] ⊢s
(M ′ [γ1 ])[α := β] : σ[α := β][γ0 := γ1 [α := β]].
Because γ0 ∈
/ {α, β}, we have
σ[α := β][γ0 := γ1 [α := β]] = σ[γ0 := γ1 ][α := β]
= τ [α := β].
The proof of (2) is the same as the proof of (1), except for the case (S-Ins2). We prove
this case. Let v = v ′ [γ1 ]. To prove this case, it is important to notice that γ1 6= α. This fact
can be shown by the following observation. Assume that γ1 = α. Then, the stage A of v
must be ε because the stage of α is ε, but v ∈ V ε contradicts the fact that v is a transition
application. Therefore γ1 6= α. The rest of the proof is the same as (1).
Then we show the preservation of ⊢s by evaluation.
A
A and
Lemma C.8 (Type Soundness (2)). If Γ; ∆ ⊢A
s M and ⊢ M ⇓ R, then R = v
A
A
A
A
Γ; ∆ ⊢s v for some v ∈ V .
40
T. TSUKADA AND A. IGARASHI
Proof. Similar to the proof of Theorem 4.4.
We also have the counterpart of Lemma C.1 for the new type system. We define
∆−α = {β@B | β@αB ∈ ∆}.
αA , then we have Γ−α ; ∆−α ⊢A M : τ .
Lemma C.9. If Γ; ∆ ⊢αA
s M : τ and M ∈ V
s
Proof. Similar to the proof of Lemma C.1.
Finally, we prove Theorem 4.5 via a more general property for which proof by induction
works.
Theorem C.10. If Γ is ε-free and Γ; ∆ ⊢A
s M : τ and n is the length of A, then
(1) if ⊢A M ⇓ N , then ⊢ne ♭(M ) ⇓ ♭(N ); and
(2) if ⊢ne ♭(M ) ⇓ N , there is some N ′ such that ⊢A M ⇓ N ′ and N = ♭(N ′ ).
Proof. By induction on the derivation of ⊢A M ⇓ N (for (1)) or ⊢ne ♭(M ) ⇓ N (for (2)) with
case analysis on the last rule used in the derivation. Actually, the only interesting case is
when M = M0 B and ⊢A M0 B ⇓ N .
• Proof of (1): Assume M = M0 B and ⊢A M0 B ⇓ N . The case A 6= ε is easy. Assume
A = ε. So, the last derivation step for ⊢ε M0 B ⇓ N is of the form
⊢ε M0 ⇓ Λα.M1
⊢ε M1 [α := B] ⇓ N
.
⊢ε M0 B ⇓ N
Since Γ; ∆ ⊢εs M0 B : τ , we have Γ; ∆ ⊢εs M0 : ∀α@ε. ⊲α σ and τ = σ[α := B] for some
σ. By Lemma C.8, we have Γ; ∆ ⊢εs Λα.M1 : ∀α@ε. ⊲α σ and Λα.M1 ∈ V ε . Therefore,
Γ; ∆, α@ε ⊢εs M1 : ⊲α σ and M1 ∈ V ε . Because M1 ∈ V ε , we have M1 = ◮α M2 with
M2 ∈ V α . So Γ; ∆, α@ε ⊢αs M2 : σ. By Lemma C.9, Γ−α ; (∆, α@ε)−α ⊢εs M2 : σ. Because
(∆, α@ε)−α = ∆−α , we have α ∈
/ FTV((∆, α@ε)−α ), and therefore α ∈
/ FTV(M2 ). So
(◮α M2 )[α := B] = ◮B M2 . Then, we know that the last derivation step must have a
more specific shape:
⊢ε M0 ⇓ Λα.◮α M2
⊢ε ◮B M2 ⇓ N
⊢ε M0 B ⇓ N
m
⊢0e
⊢0e
z }| {
◮ . . . ◮ ♭(M2 ) ⇓
♭(M0 ) ⇓ Λ ◮ ♭(M2 ) and
m
z }| {
♭(N ), where m is the length of B, because ♭(◮B M2 ) = ◮ . . . ◮ ♭(M2 ). Finally, we obtain
⊢0e ♭(M0 ) m ⇓ ♭(N ) from them.
By the induction hypothesis, we have
• Proof of (2): Assume M = M0 B and ⊢ne ♭(M0 B) ⇓ N ♭ . The case n 6= 0 is easy. We
assume n = 0. Then, the last derivation step is of the form:
⊢0e
where m is the length of
some σ.
m
z }| { ♭
0
♭
⊢e ◮ . . . ◮ M2
♭(M0 ) ⇓ Λ ◮ M2
0
⊢e ♭(M0 ) m ⇓ N ♭
B. Since Γ; ∆ ⊢εs M0 B : τ , we have
⇓ N♭
Γ; ∆ ⊢εs M0 : ∀α@ε. ⊲α σ for
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
41
By the induction hypothesis, there is some M1 such that ⊢ε M0 ⇓ M1 and Λ ◮ M2♭ =
♭(M1 ). So, M1 = Λα. ◮β M2 for some α and β and M2 such that ♭(M2 ) = M2♭ . Actually,
α = β since M0 has the type ∀α@ε.⊲α σ and evaluation preserves types. By the same argument as in the proof of (1), we have α ∈
/ FTV(M2 ). Therefore, (◮α M2 )[α := B] = ◮B M2 .
m
z }| {
Because ♭(◮B M2 ) = ◮ . . . ◮ M2♭ , by the induction hypothesis, we have ⊢ε ◮B M2 ⇓ N
for some N with ♭(N ) = N ♭ .
Proof of Theorem 4.5. Theorem 4.5 is a special case of Theorem C.10.
Appendix D. Proof of Completeness (Theorem 5.2)
D.1. Completeness in the Quantifier-Free Setting. We first prove completeness in the
quantifier-free setting: the set of propositions has no quantifiers and the deduction rules
and semantics has no rule for quantifiers.
Formally, the set Φ−∀
⊥ of quantifier-free propositions is defined by the following grammar:
Propositions
φ ∈ Φ−∀
⊥ ::= b | ⊥ | φ → φ | ⊲α φ .
The natural deduction system consists of (Var), (Abs), (App), (◮), (◭) and (⊥-E). The
Kripke semantics for this fragment is essentially the same as the full logic, but the proposition φ in the satisfaction relation T , v, ρ; s φ is restricted to be quantifier-free one.
∗
An assumption Γ is a (possibly infinite) set {φi @Ai } where φi ∈ Φ−∀
⊥ and Ai ∈ Σ . We
ε
A
say an assumption Γ is consistent if Γ 0 ⊥. (When Γ is infinite, Γ ⊢ φ means that there
exists a finite assumption Γ′ ⊂ Γ such that Γ′ ⊢A φ.) We say Γ is maximally consistent
when Γ is consistent and maximal (under the ordinary set inclusion ordering).
Lemma D.1. For any consistent assumption Γ, there is a maximally consistent assumption
Γ′ s.t. Γ′ ⊇ Γ.
Proof. We apply Zorn’s Lemma to the set Θ = {∆ | Γ ⊆ ∆ and ∆ is consistent}. To apply
Zorn’s lemma, S
we prove that every totally ordered
subsets Ξ ⊆ Θ has its upper bound in
S
Θ, by showing ΞS∈ Θ. It is clear that Γ ⊆ S Ξ.
S
We show that Ξ isSconsistent. Assume Ξ is inconsistent, i.e. Ξ ⊢ε ⊥. Then there
is a finite
subset ∆ ⊆ Ξ such that ∆ ⊢ε ⊥. Let {φi @Ai | 0 ≤ i ≤ n} = ∆. Because
S
∆ ⊆ Ξ, for every φi @Ai ∈ ∆, there is an assumption Ξi ∈ Ξ which satisfies φi @Ai ∈ Ξi .
Because the number n of elements in ∆ is finite and Ξ is totally ordered, there is a maximal
assumption Ξj ⊇ Ξi (0 ≤ i ≤ n). Therefore ∆ ⊆ Ξj and Ξj is inconsistent, and this causes
a contradiction.
42
T. TSUKADA AND A. IGARASHI
Let Γ be maximally consistent. We construct a transition system TΓ and valuations ρΓ
α
α
and vΓ from it. The transition system5 TΓ is (Σ∗ , Σ, {→ | α ∈ Σ}), where A → Aα, and
valuation ρΓ for transition variables is defined by ρΓ (α) = α. We define valuation vΓ for
propositional variables as vΓ (A, p) = 1 if and only if p@A ∈ Γ.
Lemma D.2. TΓ , vΓ , ρΓ ; ε
⊲A φ if and only if φ@A ∈ Γ.
Proof. Easy induction with respect to the proposition φ.
Theorem D.3 (Completeness for the quantifier-free fragment). If Γ
φ then Γ ⊢ε φ.
Proof. We prove by contraposition. Assume Γ 0ε φ. Then Γ, (φ → ⊥)@ε is consistent.
By Lemma D.1 there is a maximally consistent assumption Γ′ ⊇ Γ ∪ {(φ → ⊥)@ε}. By
φ → ⊥. So, TΓ′ , vΓ′ , ρΓ′ ; ε 1 φ and,
Γ and TΓ′ , vΓ′ , ρΓ′ ; ε
Lemma D.2, TΓ′ , vΓ′ , ρΓ′ ; ε
therefore, Γ 1 φ.
D.2. Completeness of the Full Logic. We prove completeness for the full logic. The
way to prove completeness is similar to that for first-order predicate logic.
We abbreviate φ → ⊥ to ¬φ, (¬φ) → ψ to φ ∨ ψ, ¬(φ → ¬ψ) to φ ∧ ψ and ¬∀α.¬φ to
∃α.φ. The deduction rules for these connectives can be given as in ordinary classical logic,
e.g. as follows:
Γ ⊢A φ
Γ ⊢A ψ
(∧-I)
Γ ⊢A φ ∧ ψ
Γ, φ@A ⊢A ψ
α∈
/ FTV(Γ, ψ@A)
(∃-E) .
A
Γ⊢ ψ
Γ ⊢A ∃α.φ
We abbreviate ∀α1 ∀α2 . . . ∀αn φ to ∀n ᾱφ. We also write φ as ∀0 ᾱφ. We say a proposition
is in prenex normal form when it has the form
∀n0 ᾱ0 ∃β1 ∀n1 ᾱ1 . . . ∃βm ∀nm+1 ᾱm+1 ψ
and ψ has no quantifier. It is easy to see that for any proposition φ there is an equivalent
proposition φ′ (i.e., ⊢ε φ ↔ φ′ ) which is in prenex normal form. Therefore we assume
without loss of generality that all propositions in an assumption are in prenex normal form.
We construct a quantifier-free assumption ∆(Γ) from an assumption Γ with quantifiers.
Assume that all binding transition variables are different from other binding variables and
free variables in Γ.
First we construct “Herbrand universe” from Γ. Herbrand universe is a term algebra,
which is freely generated algebra from a signature. Let Σ0 be the set of free transition
variables in Γ and Σ1 be the set of binding transition variables in Γ. The set SΓ is defined
as follows:
k−1
SΓ = {(βk , Σi=1
ni ) | ∀n0 ᾱ0 ∃β1 . . . ∃βk . . . ∀nm ᾱm ψ@A ∈ Γ}
We regard dom(SΓ ) as the set of pairs of function symbols and its arities. If there is no
constant (i.e., arity 0 function) in SΓ and Σ0 = Σ1 = ∅, then we add a constant (c, 0) to SΓ .
If there is no function symbol whose arity is not 0, then we add a function symbol (f, 1) to
SΓ . We define the set HΓ as the least set which satisfies the following condition:
• if (β, n) ∈ SΓ and αi ∈ HΓ ∪ Σ0 ∪ Σ1 (1 ≤ i ≤ n), then β(α1 , . . . , αn ) ∈ HΓ
5If we want to prove completeness of the logic in Sec. 5.3, we define T as ({A ∈ Σ∗ | ⊥@A ∈
/
α
α
Γ}, Σ, {→ | α ∈ Σ}), where A → Aα if ⊥@Aα ∈
/ Γ and otherwise undefined.
A LOGICAL FOUNDATION FOR ENVIRONMENT CLASSIFIERS
43
Second we regard Σ as the “Herbrand universe” by following observation. Because Σ
is countably infinite and Σ0 and Σ1 is countable, we assume without loss of generality that
Σ\(Σ0 ∪ Σ1 ) is countably infinite. It is easy to see that HΓ is countably infinite. Therefore
there is an bijection HΓ → (Σ\(Σ0 ∪ Σ1 )). By using this bijection, we regard any element
in HΓ as an element in Σ. So Σ0 ∪ Σ1 ∪ HΓ = Σ and Σ0 ∩ HΓ = Σ1 ∩ HΓ = ∅.
Then we define ∆(Γ) as follows:
∆(Γ) = {ψ [ᾱ0 , β1 , ᾱ1 , β2 , . . . ᾱm := γ̄0 , β1 (γ̄0 ), γ̄1 , β2 (γ̄0 , γ̄1 ), . . . , γ̄m ]@A
| ∀n0 ᾱ0 ∃β1 . . . ∀nm ᾱm ψ@A ∈ Γ, γij ∈ Σ}
Lemma D.4. If Γ 0ε ⊥, then ∆(Γ) 0ε ⊥ in quantifier-free logic.
Proof. We prove this lemma by contraposition.
Assume that ∆(Γ) ⊢ε ⊥. Then there is a finite subset {φi @Ai | 0 ≤ i ≤ n} ⊆ ∆
such that {φi @Ai | 0 ≤ i ≤ n} ⊢ε ⊥. From the construction of ∆(Γ), for every φi ,
there are θi and γij ∈ Σ such that θi @A ∈ Γ and θi = ∀n0 ᾱ0 ∃β1 . . . ∀nm ᾱm ψi and φi =
ψi [ᾱ0 , β1 , ᾱ1 , β2 , . . . ᾱm := γ̄0 , β1 (γ̄0 ), γ̄1 , β2 (γ̄0 , γ̄1 ), . . . , γ̄m ]. Then we can get {θi @Ai | 0 ≤
i ≤ n} ⊢ε ⊥ by using (Ins) and (∃-E) in an appropriate order. Therefore Γ ⊢ε ⊥.
Lemma D.5. If T , v, ρ; ε
∆(Γ), then T , v, ρ; ε
Γ.
Proof. Let ∀n0 ᾱ0 ∃β1 . . . ∀nm ᾱm ψ@A ∈ Γ. We can get the following proposition by easy
induction on the number of existential quantifiers.
for all γi,j ∈ ΣT , v, ρ; ε
∃βk ∀nk ᾱk . . . ∀nm ᾱn ψ[ᾱ0 , β1 , . . . , ᾱk := γ̄0 , β1 (γ̄0 ), . . . , γ̄k ]
This lemma is an easy consequence of this proposition.
Proof of Theorem 5.2. We prove by contraposition. Assume Γ 0ε φ. Then Γ, ¬φ@ε 0ε ⊥.
By Lemma D.4, ∆(Γ ∪ {¬φ@ε}) is a consistent assumption in the quantifier-free fragment.
Therefore, from completeness of the quantifier-free fragment, there is a model such that
T , v, ρ; ε ∆(Γ ∪ {¬φ@ε}). By Lemma D.5, T , v, ρ; ε Γ and T , v, ρ; ε ¬φ. Therefore
T , v, ρ; ε 1 φ. So Γ 1 φ.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
Submitted to the Annals of Statistics
arXiv:1611.09804v1 [stat.ME] 29 Nov 2016
BEST LINEAR UNBIASED ESTIMATORS IN
CONTINUOUS TIME REGRESSION MODELS
By Holger Dette‡ , Andrey Pepelyshev§ and Anatoly
Zhigljavsky§
Ruhr-Universität Bochum‡ and Cardiff University§
In this paper the problem of best linear unbiased estimation is
investigated for continuous-time regression models. We prove several
general statements concerning the explicit form of the best linear unbiased estimator (BLUE), in particular when the error process is a
smooth process with one or several derivatives of the response process available for construction of the estimators. We derive the explicit form of the BLUE for many specific models including the cases
of continuous autoregressive errors of order two and integrated error processes (such as integrated Brownian motion). The results are
illustrated by several examples.
1. Introduction. Consider a continuous-time linear regression model of
the form
(1.1)
y(t) = θ T f (t) + ǫ(t) ,
t ∈ [A, B],
where θ ∈ Rm is a vector of unknown parameters, f (t) = (f1 (t), . . . , fm (t))T
is a vector of linearly independent functions defined on some interval, say
[A, B], and ǫ = {ǫ(t)|t ∈ [A, B]} is a random error process with E[ǫ(t)] = 0
for all t ∈ [A, B] and covariances E[ǫ(t)ǫ(s)] = K(t, s). We will assume that ǫ
has continuous (in the mean-square sense) derivatives ǫ(i) (i = 0, 1, . . . , q)
up to order q, where q is a non-negative integer.
This paper is devoted to studying the best linear unbiased estimator (BLUE)
of the parameter θ in the general setting and in many specific instances.
Understanding of the explicit form of the BLUE has profound significance
on general estimation theory and on asymptotically optimal design for (at
least) three reasons. Firstly, the efficiency of the ordinary least squares estimator, the discrete BLUE and other unbiased estimators can be computed
exactly. Secondly, as pointed out in a series of papers Sacks and Ylvisaker
(1966, 1968, 1970), the explicit form of the BLUE is the key ingredient for
Primary 62J05; secondary 31A10
Keywords and phrases: linear regression, correlated observations, signed measures, optimal design, BLUE, AR processes, continuous autoregressive model
1
2
H. DETTE ET AL.
constructing the (asymptotically) optimal exact designs in the regression
model
(1.2) y(ti ) = θ T f (ti ) + ǫ(ti ) ,
A ≤ t1 < t2 . . . < tN −1 < tN ≤ B ,
with E[ǫ(ti )ǫ(tj )] = K(ti , tj ). Thirdly, simple and very efficient estimators
for the parameter θ in the regression model (6.7) can be derived from the
continuous BLUE, like the extended signed least squares estimator investigated in Dette et al. (2016a) and the estimators based on approximation of
stochastic integrals proposed in Dette et al. (2017).
There are many classical papers dealing with construction of the BLUE,
mainly in the case of a non-differentiable error process; that is, in model
(6.8) with q = 0. In this situation, it is well understood that solving specific
instances of an equation of Wiener-Hopf type
Z B
(1.3)
K(t, s)ζ(dt) = f (s),
A
for an m-dimensional vector ζ of signed measures implies an explicit construction of the BLUE in the continuous-time model (6.8). This equation
was first considered in a seminal paper of Grenander (1950) for the case of
the location-scale model y(t) = θ + ǫ(t), i.e. m = 1, f1 (t) = 1. For a general
regression model with m ≥ 1 regression functions (and q = 0), the BLUE
was extensively discussed in Grenander (1954) and Rosenblatt (1956) who
considered stationary processes in discrete time, where the spectral representation of the error process was heavily used for the construction of the
estimators. In this and many other papers including Pisarenko and Rozanov
(1963); Kholevo (1969); Hannan (1975) the subject of the study was concentrated around the spectral representation of the estimators and hence
the results in these references are only applicable to very specific models.
A more direct investigation of the BLUE in the location scale model (with
q = 0) can be found in Hajek (1956), where equation (1.3) for the BLUE was
solved for a few simple kernels. The most influential paper on properties of
continuous BLUE and its relation to the reproducing kernel Hilbert spaces
(RKHS) is Parzen (1961). A relation between discrete and continuous BLUE
has been further addressed in Anderson (1970). An excellent survey of classical results on the BLUE is given in the book of Näther (1985), Sect. 2.3 and
Chapter 4 (for the location scale model). Formally, Theorem 2.3 of Näther
(1985) includes the case when the derivatives of the process y(t) are available (q ≥ 0); this is made possible by the use of generalized functions which
may contain derivatives of the Dirac delta-function. This theorem, however,
provides only a sufficient condition for an estimator to be the BLUE.
BLUE IN CONTINUOUS TIME REGRESSION MODELS
3
The main reason why results on the BLUE in model (6.8) are so difficult
to obtain consists in the fact that - except in the location-scale model - the
functional to be minimized is not convex, so that the usual arguments are
not applicable. The main examples, where the explicit form of the BLUE
was known before the publication of the monograph by Näther (1985), are
listed in Sect. 2.3 of this book. In most of these examples either a Markovian
structure of the error process is assumed or one-dimensional location scale
model is studied. Section 2.6 of our paper updates this list and gives a short
outline of previously known cases where the explicit form of the BLUE was
known until now.
There was also an extensive study of the relation between solutions of
the Wiener-Hopf equations and the BLUE through the RKHS theory, see
Parzen (1961); Sacks and Ylvisaker (1966, 1968, 1970) for an early or Ritter
(2000) for a more recent reference. If q = 0 then the main RKHS assumption is usually formulated as the existence of a solution, say ζ0 , of
equation (1.3), where the measure ζ0 is continuous and has no atoms, see
Berlinet and Thomas-Agnan (2011) for the theory of the RKHS. As shown in
the present paper, this almost never happens for the commonly used covariance kernels and regression functions (a single general exception from this
observation is given in Proposition 2.4). Note also that the numerical construction of the continuous BLUE is difficult even for q = 0 and m = 1, see
e.g. Ramm and Charlot (1980) and a remark on p.80 in Sacks and Ylvisaker
(1966). For q > 0, the problem of numerical construction of the BLUE is
severely ill-posed and hence is extremely hard.
The main purpose of this paper is to provide further insights into the
structure of the BLUE (and its covariance matrix) from the observations
{Y (t)|t ∈ T } (and its q derivatives) in continuous-time regression models of
the form (6.8), where the set T ⊆ [A, B] defines the region where the process is observed. By generalizing the celebrated Gauss-Markov theorem, we
derive new characterizations for the BLUE which can be used to determine
its explicit form and the corresponding covariance matrix in numerous models. In particular, we do not have to restrict ourselves to one-dimensional
regression models and to Markovian error processes. Thus our results require minimal assumptions regarding the regression function and the error
process. Important new examples, where the BLUE can be determined explicitly, include the process with triangular covariance function (2.7), general
integrated processes (in particular, integrated Brownian motion) and continuous autoregressive processes including the Matérn kernels with parameters
3/2 and 5/2.
The remaining part of this paper is organized as follows. In Section 2 we
4
H. DETTE ET AL.
develop a consistent general theory of best linear unbiased estimation using
signed matrix measures. We are able to circumvent the convexity problem
and derive several important characterizations and properties of the BLUE.
In particular, in Theorem 2.1 we provide necessary and sufficient conditions
for an estimator to be BLUE when q ≥ 0; in Theorem 2.2 such conditions are
derived for q = 0, T ⊂ Rd with d ≥ 1 and very general assumptions about
the vector of regression functions f (·) and the covariance kernel K(·, ·).
Section 3 is devoted to models, where the error process has one derivative.
In particular, we derive an explicit form of the BLUE, see Theorems 3.1 and
3.2, and obtain the BLUE for specific types of smooth kernels. In Section 3.4
we consider regression models with a continuous-time autoregressive (AR)
error process of order 2 (i.e. CAR(2)) in more detail. Moreover, in an online
supplement [see Dette et al. (2016b)] we demonstrate that the covariance
matrix of the BLUE in this model can be obtained as a limit of the covariance
matrices of the BLUE in discrete regression models (6.7) with observations
at equidistant points and a discrete AR(2) error process. In Section 4 we
give some insight into the structure of the BLUE when the error process
is more than once differentiable. Some numerical illustrations are given in
Section 5, while technical proofs can be found in Section 6.
2. General linear estimators and the BLUE.
2.1. Linear estimators and their properties. Consider the regression model
(6.8) with covariance kernel K(t, s) = E[ǫ(t)ǫ(s)]. Suppose that we can observe the process {y(t)|t ∈ T } along with its q ≥ 0 mean square derivatives
{y (i) (t)|t ∈ T } for i = 1, . . . , q, where the design set T is a Borel subset of
some interval [A, B] with −∞ ≤ A < B ≤ ∞. This is possible when the kernel K(t, s) is q times continuously differentiable on the square [A, B]×[A, B]
and the vector-function f (t) = (f1 (t), . . . , fm (t))T is q times differentiable
on the interval [A, B] with derivatives f (1) , . . . f (q) (f (0) = f ). Throughout
this paper we will also assume that the functions f1 , . . . , fm are linearly
independent on T .
Let Y (t) = {(y (0) (t), . . . , y (q) (t))T } be the observation vector containing the
process y(t) = y (0) (t) and its q derivatives. Denote by YT = {Y (t) : t ∈ T }
the set of all available observations. The general linear estimator of the
parameter θ in the regression model (6.8) can be defined as
Z
q Z
X
(2.1)
y (i) (t)Gi (dt),
G(dt)Y (t) =
θ̂G =
T
i=0
T
where G(dt) = (G0 (dt), . . . , Gq (dt)) is a matrix of size m × (q + 1). The
columns of this matrix are signed vector-measures G0 (dt), . . . , Gq (dt) defined
BLUE IN CONTINUOUS TIME REGRESSION MODELS
5
on Borel subsets of T (all vector-measures in this paper are signed and have
length m).
The following lemma shows a simple way of constructing unbiased estimators; this lemma will also be used for deriving the BLUE in many examples.
The proof is given in Section 6.
Lemma 2.1. Let ζ0 , . . . , ζq be some signed vector-measures defined on T
such that the m × m matrix
(2.2)
C=
q Z
X
i=0
T
T
ζi (dt) f (i) (t)
is non-degenerate. Define G = (G0 , . . . , Gq ), where Gi are the signed vectormeasures and Gi (dt) = C −1 ζi (dt) for i = 0, . . . , q. Then the estimator θ̂G is
unbiased.
The covariance matrix of any unbiased estimator θ̂G of the form (2.1) is
given by
Z Z
(2.3)
G(dt)K(t, s)GT (ds)
Var(θ̂G ) =
T
T
q Z Z
q X
X
∂ i+j K(t, s)
=
Gi (dt)GTj (ds) ,
∂ti ∂sj
T T
i=0 j=0
where
K(t, s) =
∂ i+j K(t, s)
∂ti ∂sj
q
i,j=0
=
q
E[ǫ(i) (t)ǫ(j) (s)]
i,j=0
is the matrix consisting of the derivatives of K.
2.2. The BLUE. If there exists a set of signed
R vector-measures, say G =
(G0 , . . . , Gq ), such that the estimator θ̂G = T G(dt)Y (t) is unbiased and
R
Var(θ̂H ) ≥ Var(θ̂G ), where θ̂H = T H(dt)Y (t) is any other linear unbiased
estimator which uses the observations YT , then θ̂G is called the best linear
unbiased estimator (BLUE) for the regression model (6.8) using the set of
observations YT . The BLUE depends on the kernel K, the vector-function f ,
the set T where the observations are taken and on the number q of available
derivatives of the process {y(t)|t ∈ T }.
The following theorem is a generalization of the celebrated Gauss-Markov
theorem (which is usually formulated for the case when q = 0 and T is
6
H. DETTE ET AL.
finite) and gives a necessary and sufficient condition for an estimator to be
the BLUE. In this theorem and below we denote the partial derivatives of
the kernel K(t, s) with respect to the first component by
K (i) (t, s) =
∂ i K(t, s)
.
∂ti
The proof of the theorem can be found in Section 6.
Theorem 2.1. Consider the regression model (6.8), where the error process
{ǫ(t)|t ∈ [A, B]} has a covariance kernel K(·, ·) ∈ C q ([A, B] × [A, B]) and
f (·) ∈ C q ([A, B]) for some q ≥ 0. Suppose that the process {y(t)|t ∈ [A, B]}
along with its q derivatives can
R be observed at all t ∈ T ⊆ [A, B].
An unbiased estimator θ̂G = T G(dt)Y (t) is BLUE if and only if the equality
(2.4)
q Z
X
i=0
K (i) (t, s)Gi (dt) = Df (s),
T
is fulfilled for all s ∈ T , where D is some m × m matrix. In this case,
D = Var(θ̂G ) with Var(θ̂G ) defined in (2.3).
The next proposition is slightly weaker than Theorem 2.1 (here the covariance matrix of the BLUE is assumed to be non-degenerate) but will be very
useful in further considerations.
Proposition 2.1. Let the assumptions of Theorem 2.1 be satisfied and
let ζ0 , . . . , ζq be signed vector-measures defined on T such that the matrix
C defined in (2.2) is non-degenerate. Define G = (G0 , . . . , Gq ), Gi (dt) =
R
C −1 ζi (dt) for i = 0, . . . , q. The estimator θ̂G = T G(dt)Y (t) is the BLUE
if and only if
(2.5)
q Z
X
i=0
K (i) (t, s)ζi (dt) = f (s)
T
for all s ∈ T . In this case, the covariance matrix of θ̂G is Var(θ̂G ) = C −1 .
2.3. Grenander’s theorem and its generalizations.
When T = [A, B], q = 0, m = 1 and the regression model (6.8) is the
location-scale model y(t) = α + ε(t), Theorem 2.1 is known as Grenander’s
theorem [see Grenander (1950) and Section 4.3 in Näther (1985)]. In this
special case Grenander’s theorem has been generalised by Näther (1985) to
the case when T ⊂ Rd [see Theorem 4.3 in this reference]. The reason why
BLUE IN CONTINUOUS TIME REGRESSION MODELS
7
Grenander’s and Näther’s theorems only deal with the location-scale model
is caused by the fact that for this model the convexity of the functional to be
minimized is easy to establish. For general regression models the convexity
argument is not directly applicable and hence the problem is much harder.
For the case of one-dimensional processes, Theorem 2.1 generalizes Grenander’s theorem to arbitrary m-parameter regression models of the form (6.8)
and the case of arbitrary q ≥ 0. Another generalization of the Grenander’s
theorem is given below; it deals with a general m-parameter regression model
(6.8) with a continuous error process (i.e. q = 0) and a d-dimensional set
T ⊂ Rd ; that is, the case where y(t) is a random field. Note that the conditions on the vector of regression functions f (·) in Theorem 2.2 are weaker
(when d = 1) than the conditions on f (·) in Theorem 2.1 applied in the case
q = 0.
Theorem 2.2. Consider the regression model y(t) = θ T f (t) + ǫ(t), where
t ∈ T ⊂ Rd , the error process ǫ(t) has covariance kernel K(·, ·) and f : T → Rm
is a vector of bounded integrable and linearly independent functions. Suppose
that the process y(t) can be observed at all t ∈ TR and let G be a signed vectormeasure on T , such that the estimator θ̂G = T G(dt)Y (t) is unbiased. θ̂G
is a BLUE if and only if the equality
Z
K(t, s)G(dt) = Df (s)
T
holds for all s ∈ T for some m × m matrix D. In this case, D = Var(θ̂G ),
where Var(θ̂G ) is the covariance matrix of θ̂G defined by (2.3).
The proof of this theorem is a simple extension of the proof of Theorem 2.1
with q = 0 to general T ⊂ Rd and left to the reader.
2.4. Properties of the BLUE.
(P1) Let θ̂G1 and θ̂G2 be BLUEs for the same regression model (6.8) and the
same q but for two different design sets T1 and T2 such that T1 ⊆ T2 .
Then Var(θ̂G1 ) ≥ Var(θ̂G2 ).
(P2) Let θ̂G1 and θ̂G2 be BLUEs for the same regression model (6.8) and
the same design set T but for two different values of q, say, q1 and q2 ,
where 0 ≤ q1 ≤ q2 . Then Var(θ̂G1 ) ≥ Var(θ̂G2 ).
(P3) Let θ̂G with G = (G0 , . . . , Gq ) be a BLUE for the regression model
(6.8), design space T and given q ≥ 0. Define g(t) = Lf (t), where
L is a non-degenerate m × m matrix, and a signed vector-measure
H = (H0 , . . . , Hq ) with Hi (dt) = L−1 Gi (dt) for i = 0, . . . , q. Then θ̂H
8
H. DETTE ET AL.
is a BLUE for the regression model y(t) = β T g(t) + ε(t) with the same
T
y(t), ε(t), T and q. The covariance matrix of θ̂H is L−1 Var(θ̂G )L−1 .
(P4) If T = [A, B] and a BLUE θ̂G is defined by the matrix-measure G
that has smooth enough continuous parts, then we can choose another
representation θ̂H of the same BLUE, which is defined by the matrixmeasure H = (H0 , H1 , . . . , Hq ) with vector-measures H1 , . . . , Hq having no continuous parts.
(P5) Let ζ0 , . . . , ζq satisfy the equation (2.5) for all s ∈ T , for some vectorfunction f (·), design set T and given q ≥ 0. Define C = Cf by (2.2).
Let g(·) be some other q times differentiable vector-function on the
interval [A, B]. Assume that for all s ∈ T , signed vector-measures
η0 , . . . , ηq satisfy the equation
q Z
X
(2.6)
K (i) (t, s)ηi (dt) = g(s);
i=0
T
g(·), the same design
that is, the equation (2.5) for the vector-function
P R
set T and the same q. Define Cg = qi=0 T g (i) (t)ηiT (dt), which is the
matrix (2.2) with ηi substituted for ζi and g(·) substituted for f (·).
If the matrix C = Cf + Cg is non-degenerate, then we define the set
of signed vector-measures G = (G0 , . . . , Gq ) by Gi = C −1 (ζi + ηi ),
i = 0, . . . , q, yielding the estimator θ̂G . This estimator is a BLUE for
the regression model y(t) = θ T [f (t) + g(t)] + ε(t), t ∈ T .
Properties (P1)–(P3) are obvious. The property (P4) is a particular case of
the discussion of Section 2.5. To prove (P5) we simply add the equations
(2.5) and (2.6) and then use Proposition 2.1.
We believe that the properties (P4) and (P5) have never been noticed before
and both these properties are very important for understanding best linear
unbiased estimators in the continuous-time regression model (6.8) and especially for constructing a BLUE for new models from the cases when a BLUE
is known for simpler models. As an example, assume that all functions in
the vector f are not constant and set g(t) = c, where c is a constant vector.
Then, if we know the BLUE for f and another BLUE for the location-scale
model, we can use property (P5) to construct BLUE for θ T (f (t) + c). This
is an essential part of the proof of Theorem 3.2, which allows obtaining the
explicit form of the BLUE for the integrated error processes from the explicit form of the BLUE for the corresponding non-integrated errors (which
is a much easier problem).
2.5. Non-uniqueness. Let us show that if T = [A, B] then, under the additional smoothness conditions, for a given set of signed vector-measures
BLUE IN CONTINUOUS TIME REGRESSION MODELS
9
G = (G0 , G1 , . . . , Gq ) on T we can find another set of measures H =
(H0 , H1 , . . . , Hq ) such that the signed vector-measures H1 , . . . , Hq have no
continuous parts but the expectations and covariance matrices of the estimators θ̂G and θ̂H coincide.
For this purpose, let q > 0, G0 , . . . , Gq be some signed vector-measures and
for some i ∈ {1, . . . , m}, the signed measure Gi (dt) has the form
Gi (dt) = Qi (dt) + ϕ(t)dt,
where Qi (dt) is a signed vector-measure and ϕ ∈ C i ([A, B]) (that is, ϕ is
an i times differentiable vector-function on the interval [A, B]). Define the
matrix H = (H0 , . . . , Hq ), where the columns of H are the following signed
vector-measures:
h
i
H0 (dt) = G0 (dt) + (−1)i ϕ(i) (t)dt − ϕ(i−1) (A)δA (dt) + ϕ(i−1) (B)δB (dt) ,
Hi (dt) = Qi (dt), Hj (dt) = Gj (dt), for j = i + 1, . . . , q and
h
i
Hj (dt) = Gj (dt) + (−1)i−j−1 ϕ(i−j−1) (A)δA (dt) − ϕ(i−j−1) (B)δB (dt)
for j = 1, . . . , i − 1, where δA (dt) and δB (dt) are the Dirac delta-measures
concentrated at the points A and B, respectively. The proof of the following
result is given in Section 6.
Lemma 2.2. In the notation
R above, the expectations
R and covariance matrices of the estimators θ̂G = G(dt)Y (t) and θ̂H = H(dt)Y (t) coincide.
By Lemma 2.2 we can restrict the search of linear unbiased estimators to
estimators θ̂G of the form (2.1), where the components G1 , . . . , Gq of the
signed matrix-measure G = (G0 , . . . , Gq ) have no continuous parts.
2.6. Several examples of the BLUE for non-differentiable error processes.
For the sake of completeness we first consider the case when the errors in
model (6.8) follow a Markov process; this includes the case of continuous
autoregressive errors of order 1. In presenting these results we follow Näther
(1985) and Dette et al. (2016a).
Proposition 2.2. Consider the regression model (6.8) with covariance
kernel K(t, s) = u(t)v(s) for t ≤ s and K(t, s) = v(t)u(s) for t > s, where
u(·) and v(·) are positive functions such that q(t) = u(t)/v(t) is monotonically increasing. Define the signed vector-measure ζ(dt) = zA δA (dt) +
10
H. DETTE ET AL.
zB δB (dt) + z(t)dt with
h f (A)u′ (A)
i
1
′
−
f
(A)
,
v 2 (A)q ′ (A)
u(A)
h′ (B)
1 h h′ (t) i′
,
z
=
,
z(t) = −
B
v(t) q ′ (t)
v(B)q ′ (B)
zA =
where the vector-function
h(·) is defined by h(t) = f (t)/v(t). Assume that
R
the matrix C = T f (t)ζ T (dt) is non-degenerate. Then the estimate θ̂G with
G(dt) = C −1 ζ(dt) is a BLUE with covariance matrix C −1 .
In the following statement we provide an explicit expression for the BLUE in
a special case of the covariance kernel K(t, s) such that K(t, s) 6= u(t)v(s).
This statement provides the first example where an explicit form of the
BLUE and its covariance matrix can be obtained for a non-Markovian error
process. The proof is given in Section 6.
Proposition 2.3. Consider the regression model (6.8) on the interval T =
[A, B] with errors having the covariance function K(t, s) = 1 + λ1 t − λ2 s,
where t ≤ s, λ1 ≥ λ2 and λ2 (B − A) ≤ 1. Define the signed vector-measure
ζ(dt) = zA δA (dt) + zB δB (dt) + z(t)dt by
λ21 f (A) + λ1 λ2 f (B)
/(λ1 + λ2 ),
λ1 + λ2 + λ21 A − λ22 B)
λ1 λ2 f (A) + λ22 f (B)
zB = f (1) (B) +
/(λ1 + λ2 )
λ1 + λ2 + λ21 A − λ22 B)
R
and suppose that the matrix C = T f (t)ζ T (dt) is non-degenerate. Then the
estimator θ̂G with G(dt) = C −1 ζ(dt) is a BLUE with covariance matrix C −1 .
z(t) = −
f (2) (t)
,
λ1 + λ2
zA =
− f (1) (A) +
If λ1 = λ2 and [A, B] = [0, 1] in Proposition 2.3 then we obtain the following
case when the kernel is
(2.7)
K(t, s) = max(1 − λ|t − s|, 0) .
Optimal designs for this covariance kernel (with λ = 1) have been considered
in [Sect. 6.5 in Näther (1985)], Müller and Pázman (2003) and Fedorov and Müller
(2007).
Example 2.1. Consider the regression model (6.8) on the interval T =
[0, 1] with errors having the covariance kernel (2.7) with λ ≤ 1. Define the
signed vector-measure
ζ(dt) = [−f (1) (0)/(2λ)+fλ ]δ0 (dt)+[f (1) (1)/(2λ)+fλ ]δ1 (dt)−[f (2) (t)/(2λ)]dt ,
BLUE IN CONTINUOUS TIME REGRESSION MODELS
11
R
where fλ = (f (0)+f (1))/(4−2λ). Assume that the matrix C = T f (t)ζ T (dt)
is non-degenerate. Then the estimator θ̂G with G(dt) = C −1 ζ(dt) is a BLUE;
the covariance matrix of this estimator is given by C −1 .
Consider now the case when the regression functions are linear combinations
of eigenfunctions from the Mercer’s theorem. Note that a similar approach
was used in Dette et al. (2013) for the construction of optimal designs for
the signed least squares estimators. Let T = [A, B] and ν be a measure
on the corresponding Borel field with positive density. Consider the integral
operator
(2.8)
TK (h)(·) =
Z
B
A
K(t, ·)h(t)ν(dt)
on L2 (ν, [A, B]), which defines a symmetric, compact self-adjoint operator.
In this case Mercer’s Theorem [see e.g. Kanwal (1997)] shows that there exist
a countable number of orthonormal (with respect to ν(dt)) eigenfunctions
φ1 , φ2 , . . . with positive eigenvalues λ1 , λ2 , . . . of the integral operator TK .
The next statement follows directly from Proposition 2.1.
Proposition 2.4. Let
1 , φ2 , . . . be the eigenfunctions of the integral operPφ
∞
ator (2.8) and f (t) = ℓ=1 qℓ φℓ (t) with some sequence {qℓ }ℓ∈N in Rm such
that f1 (t), . . . P
, fm (x) are linearly independent.
Then the estimator θ̂G with
P∞
−1
−1
T
q
φ
(t)ν(dt)
and
C
=
G(dt) = C −1 ∞
λ
ℓ
ℓ
ℓ=1 λℓ qℓ qℓ is a BLUE with
ℓ=1 ℓ
−1
covariance matrix C .
Proposition 2.4 provides a way of constructing the covariance kernels for
which the measure defining the BLUE does not have any atoms. An example
of such kernels is the following.
Example 2.2. Consider the regression model (6.8) with m = 1, f (t) ≡ 1,
t ∈ T = [−1, 1], and the covariance kernel K(t, s) = 1 + κpα,β (t)pα,β (s),
α+β
where κ > 0, α, β > −1 are some constants and pα,β (t) = α−β
2 + (1 + 2 )t
is the Jacobi polynomial of degree 1. Then the estimator θ̂G with G(dt) =
const · (1 − t)α (1 + t)β dt is a BLUE.
3. BLUE for processes with trajectories in C 1 [A, B]. In this section, we assume that the error process is exactly once continuously differentiable (in the mean-square sense).
12
H. DETTE ET AL.
3.1. A general statement. Consider the regression model (6.8) and a linear
estimator in the form
Z
Z
(3.1)
y (1) (t)G1 (dt),
y(t)G0 (dt) +
θ̂G0 ,G1 =
T
T
where G0 (dt) and G1 (dt) are signed vector-measures. The following corollary
is a specialization of Proposition 2.1 when q = 1.
Corollary 3.1. Consider the regression model (6.8) with the covariance
kernel K(t, s) and such that y (1) (t) exists in the mean-square sense for all
t ∈ [A, B]. Suppose that y(t) and y (1) (t) can be observed at all t ∈ T . Assume
that there exist vector-measures ζ0 and ζ1 such that the equality
Z
Z
K (1) (t, s)ζ1 (dt) = f (s),
K(t, s)ζ0 (dt) +
T
T
is fulfilled for all s ∈ T , and such that the matrix
Z
Z
T
f (1) (t)ζ1T (dt)
f (t)ζ0 (dt) +
C=
T
T
is non-degenerate. Then the estimator θ̂G0 ,G1 defined in (3.1) with Gi =
C −1 ζi (i = 0, 1) is a BLUE with covariance matrix C −1 .
The next theorem provides sufficient conditions for vector-measures of some
particular form to define a BLUE by (3.1) for the case T = [A, B]. This
theorem, which is proved in Section 6, will be useful for several choices of
the covariance kernel below. Define the vector-function
z(t) = (τ0 f (t) − τ2 f (2) (t) + f (4) (t))/s3 ,
and vectors
zA =
zB =
z1,A =
z1,B =
f (3) (A) − γ1,A f (1) (A) + γ0,A f (A) /s3 ,
− f (3) (B) + γ1,B f (1) (B) + γ0,B f (B) /s3 ,
− f (2) (A) + β1,A f (1) (A) − β0,A f (A) /s3 ,
f (2) (B) + β1,B f (1) (B) + β0,B f (B) /s3 ,
where τ0 , τ2 , γ0,A , γ1,A , β0,A , β1,A , γ0,B , γ1,B , β0,B , β1,B , s3 are some constants
and s3 = K (3) (s−, s) − K (3) (s+, s). Define the functions
J1 (s) = − γ1,A K(A, s) + β1,A K (1) (A, s) + τ2 K(A, s) − K (2) (A, s),
(3.2)
J2 (s) = γ0,A K(A, s) − β0,A K (1) (A, s) − τ2 K (1) (A, s) + K (3) (A, s),
J3 (s) = − γ1,B K(B, s) + β1,B K (1) (B, s) − τ2 K(B, s) + K (2) (B, s),
J4 (s) = γ0,B K(B, s) − β0,B K (1) (B, s) + τ2 K (1) (B, s) − K (3) (B, s).
BLUE IN CONTINUOUS TIME REGRESSION MODELS
13
Theorem 3.1. Consider the regression model (6.8) on the interval T =
[A, B] with errors having the covariance kernel K(t, s). Suppose that the
vector of regression functions f is four times differentiable and K(t, s) is
also four times differentiable for t 6= s such that
K (i) (s−, s) − K (i) (s+, s) = 0, i = 0, 1, 2,
K (3) (s−, s) − K (3) (s+, s) 6= 0.
With the notation of the previous paragraph define the vector-measures
ζ0 (dt) = zA δA (dt) + zB δB (dt) + z(t)dt,
ζ1 (dt) = z1,A δA (dt) + z1,B δB (dt).
Assume that there exist constants τ0 , τ2 , γ0,A , γ1,A , β0,A , β1,A , γ0,B , γ1,B , β0,B , β1,B
such that (i) the identity
τ0 K(t, s) − τ2 K (2) (t, s) + K (4) (t, s) ≡ 0
(3.3)
holds for all t, s ∈ [A, B], (ii) the identity J1 (s) R+ J2 (s) + J3 (s)R + J4 (s) ≡ 0
holds for all s ∈ [A, B], and (iii) the matrix C = T f (t)ζ0T (dt)+ T f (1) (t)ζ1T (dt)
is non-degenerate. Then the estimator θ̂G0 ,G1 defined in (3.1) with Gi (dt) =
C −1 ζi (dt) (i = 0, 1) is a BLUE with covariance matrix C −1 .
3.2. Two examples for integrated error processes. In this section we illustrate the application of our results calculating the BLUE when errors follow
an integrated Brownian motion and an integrated process with triangularshape kernel. All results of this section can be verified by a direct application
of Theorem 3.1. We first consider the case of Brownian motion, where the
integrated covariance kernel is given by
Z tZ s
K(t, s) =
min(t′ , s′ )dt′ ds′
a
(3.4)
a
max(t, s)(min(t, s)2 − a2 ) a2 (min(t, s) − a) min(t, s)3 − a3
=
−
−
2
2
6
and 0 ≤ a ≤ A.
Proposition 3.1. Consider the regression model (6.8) with integrated covariance kernel given by (3.4) and suppose that f is four times differentiable
on the interval [A, B]. Define the signed vector-measures
ζ0 (dt) = zA δA (dt) + zB δB (dt) + z(t)dt,
ζ1 (dt) = z1,A δA (dt) + z1,B δB (dt),
14
H. DETTE ET AL.
where z(t) = f (4) (t),
12A
6(A + a)
f (1) (A) +
f (A),
2
(A + 3a)(A − a)
(A + 3a)(A − a)3
4(A + 2a)
6(A + a)
= −f (2) (A) +
f (1) (A) −
f (A),
(A + 3a)(A − a)
(A + 3a)(A − a)2
zA = f (3) (A) −
z1,A
zB = −f (3) (B),
z1,B = f (2) (B).
R
RB
Assume that the matrix C = A f (t)ζ0T (dt)+ T f (1) (t)ζ1T (dt) is non-degenerate.
Then the estimator θ̂G0 ,G1 defined in (3.1) with Gi (dt) = C −1 ζi (dt) is a
BLUE with covariance matrix C −1 .
The next example is a particular case of Proposition 3.1 when a = 0.
Example 3.1. Consider the regression model (6.8) on T = [A, B] with
integrated covariance kernel
(3.5)
K(t, s) = min(t, s)2 (3 max(t, s) − min(t, s))/6 .
Suppose that f is differentiable four times. Define the vector-measures ζ0 (dt) =
zA δA (dt) + zB δB (dt) + z(t)dt and ζ1 (dt) = z1,A δA (dt) + z1,B δB (dt), where
z(t) = f (4) (t),
12
6 (1)
f (A) + 3 f (A),
2
A
A
6
4 (1)
(2)
= −f (A) + f (A) − 2 f (A),
A
A
= −f (3) (B), z1,B = f (2) (B).
zA = f (3) (A) −
z1,A
zB
RB
RB
Assume that the matrix C = A f (t)ζ0T (dt)+ A f (1) (t)ζ1T (dt) is non-degenerate.
Then the estimator θ̂G0 ,G1 defined in (3.1) with Gi (dt) = C −1 ζi (dt) is a
BLUE with covariance matrix C −1 .
Consider now the integrated triangular-shape kernel
Z tZ s
K(t, s) =
max{0, 1 − λ|t′ − s′ |}dt′ ds′
0
0
(3.6)
= ts − λ min(t, s) 3 max(t, s)2 − 3ts + 2 min(t, s)2 /6.
Proposition 3.2. Consider the regression model (6.8) on T = [A, B] with
integrated covariance kernel (3.6), where λ(B − A) < 1. Suppose that f is
BLUE IN CONTINUOUS TIME REGRESSION MODELS
15
four times differentiable. Define the signed vector-measures
ζ0 (dt) = zA δA (dt) + zB δB (dt) + z(t)dt,
ζ1 (dt) = z1,A δA (dt) + z1,B δB (dt),
where z(t) = f (4) (t)/(2λ) and
i
6λ (1)
12κ1
6κ2 (1)
f
(A)
+
f
(B)
+
f
(A)
/(2λ),
A2 κ4
Aκ4
A3 κ4
i
h
2λ (1)
6κ2
4κ3 (1)
(2)
f (A) −
f (B) − 2 f (A) /(2λ),
= − f (A) +
Aκ4
κ4
A κ4
h
i
2λ (1)
4λ (1)
6λ
= f (2) (B) −
f (A) +
f (B) +
f (A) /(2λ),
κ4
κ4
Aκ4
(3)
= −f (B)/(2λ), κj = Aλ − jBλ + 2j.
zA =
z1,A
z1,B
zB
h
f (3) (A) −
RB
RB
Assume that the matrix C = A f (t)ζ0T (dt)+ A f (1) (t)ζ1T (dt) is non-degenerate.
Then the estimator θ̂G0 ,G1 defined in (3.1) with Gi (dt) = C −1 ζi (dt) is a
BLUE with covariance matrix C −1 .
3.3. Explicit form of the BLUE for the integrated processes. We conclude
this section establishing a direct link between the BLUE for models with nondifferentiable error processes and the BLUE for regression models with an
integrated kernel (3.9). Note that this extends the class of kernels considered
in Sacks and Ylvisaker (1970) in a nontrivial way.
Consider the regression model (6.8) with a non-differentiable error process
with covariance kernel K(t, s) and BLUE
Z
y(t)G0 (dt).
θ̂G0 =
T
From Proposition 2.1 we have for the vector-measure ζ0 (dt) satisfying (2.5)
and defining the BLUE
Z B
(3.7)
K(t, s)ζ0 (dt) = f (s)
A
−1
R
and Var(θ̂G0 ) = C −1 = T f (t)ζ0T (dt) . The unbiasedness condition for
the measure G0 (dt) = C −1 ζ0 (dt) is
Z
f (t)GT0 (dt) = Im .
T
16
H. DETTE ET AL.
Define the integrated process as follows:
Z t
Z t
ye(t) =
y(u)du, fe(t) =
f (u)du,
a
a
εe(t) =
Z
t
ε(u)du
a
with some a ≤ A (meaning that the regression vector-function and the error
process are defined on [a, B] but observed on [A, B]) so that
fe(1) (t) = f (t),
Consider the regression model
ye(1) (t) = y(t),
εe(1) (t) = ε(t) .
ỹ(t) = θ T f˜(t) + ε̃(t),
(3.8)
which has the integrated covariance kernel
Z tZ s
R(t, s) =
(3.9)
K(u, v)dudv.
a
a
The proof of the following result is given in Section 6.
Theorem 3.2. Let the vector-measure ζ0 satisfy the equality (3.7) and define the BLUE θ̂G0 with G0 (dt) = C −1 ζ0 (dt) in the regression model (6.8)
with covariance kernel K(·, ·). Let the measures η0 , η1 satisfy the equality
Z
Z
(3.10)
R(1) (t, s)η1 (dt) = 1
R(t, s)η0 (dt) +
T
T
for all s ∈ T . Define the vector-measures
R A R B ζ̃0 = −cη0 and ζ̃1 = −cη1 + ζ0 ,
where the vector c is given by c = a [ A K(t, s)ζ0 (dt) − f (s)]ds. Then the
estimator θ̂G̃0 ,G̃1 defined in (3.1) with G̃i (dt) = C̃ −1 ζ̃i (dt) (i = 1, 2), where
R
R
C̃ = f˜(t)ζ̃0T (dt) + f˜(1) (t)ζ̃1T (dt), is a BLUE in the regression model (3.8)
with integrated covariance kernel (3.9).
Repeated application of Theorem 3.2 extends the results to the case of several times integrated processes.
If a = A in (3.9) we have c = 0 in Theorem 3.2 and in this case, the
statement of Theorem 3.2 can be proved easily. Moreover, in this case the
class of kernels defined by (3.9) is exactly the class of kernels considered
in equation (1.5) and (1.6) of Sacks and Ylvisaker (1970) for once differentiable processes (k = 1 in their notation). We emphasize that the class of
kernels considered here is much richer than the class of kernels considered
in Sacks and Ylvisaker (1970).
BLUE IN CONTINUOUS TIME REGRESSION MODELS
17
3.4. BLUE for AR(2) errors. Consider the continuous-time regression model
(6.8), which can be observed at all t ∈ [A, B], where the error process is a
continuous autoregressive (CAR) process of order 2. Formally, a CAR(2)
process is defined as a solution of the linear stochastic differential equation
of the form
(3.11)
dε(1) (t) = ã1 ε(1) (t) + ã2 ε(t) + σ02 dW (t),
where Var(ε(t)) = σ 2 and W (t) is a standard Wiener process, [see Brockwell et al.
(2007)]. Note that the process {ε(t)|t ∈ [A, B]} defined by (3.11) has a continuous derivative and, consequently, the process {y(t) = θ T f (t) + ε(t)| t ∈
[A, B]}, is a continuously differentiable process with drift on the interval
[A, B]. In this section we derive the explicit form for the continuous BLUE
using Theorem 3.1. An alternative approach would be to use the coefficients
of the equation (3.11) as indicated in Parzen (1961).
There are in fact three different forms of the autocorrelation function ρ(t) =
K(0, t) of CAR(2) processes [see e.g. formulas (14)–(16) in He and Wang
(1989)], which are given by
(3.12)
ρ1 (t) =
λ1
λ2
e−λ1 |t| −
e−λ2 |t| ,
λ2 − λ1
λ2 − λ1
where λ1 6= λ2 , λ1 > 0, λ2 > 0, by
n
o
λ
ρ2 (t) = e−λ|t| cos(ω|t|) + sin(ω|t|) ,
(3.13)
ω
where λ > 0, ω > 0, and by
(3.14)
ρ3 (t) = e−λ|t| (1 + λ|t|) ,
where λ > 0. Note that the kernel (6.13) is widely known as Matérn kernel with parameter 3/2, which has numerous applications in spatial statistics [see Rasmussen and Williams (2006)] and computer experiments [see
Pronzato and Müller (2012)]. In the following results, which are proved in
Section 6.7, we specify the BLUE for the CAR(2) model.
Proposition 3.3. Consider the regression model (6.8) with CAR(2) errors, where the covariance kernel K(t, s) = ρ(t − s) has the form (6.11).
Suppose that f is a vector of linearly independent, four times differentiable
functions on the interval [A, B]. Then the conditions of Theorem 3.1 are
satisfied for s3 = 2λ1 λ2 (λ1 + λ2 ), τ0 = λ21 λ22 , τ2 = λ21 + λ22 , βj,A = βj,B = βj
and γj,A = γj,B = γj for j = 0, 1, where β1 = λ1 + λ2 , γ1 = λ21 + λ1 λ2 + λ22 ,
β0 = λ1 λ2 and γ0 = λ1 λ2 (λ1 + λ2 ).
18
H. DETTE ET AL.
Proposition 3.4. Consider the regression model (6.8) with CAR(2) errors, where the covariance kernel K(t, s) = ρ(t − s) has the form (6.12).
Suppose that f is a vector of linearly independent, four times differentiable
functions. Then the conditions of Theorem 3.1 hold for s3 = 4λ(λ2 + ω 2 ),
τ0 = (λ2 + ω 2 )2 , τ2 = 2(λ2 − ω 2 ), βj,A = βj,B = βj and γj,A = γj,B = γj
for j = 0, 1, where β1 = 2λ, γ1 = γ1 = 3λ2 − ω 2 , β0 = λ2 + ω 2 and
γ0 = 2λ(λ2 + ω 2 ).
The BLUE for the covariance kernel in the form (6.13) is obtained from
either Proposition 3.3 with λ1 = λ2 = λ or Proposition 3.4 with ω = 0.
Remark 3.1. In the online supplement Dette et al. (2016b) we consider
the regression model (6.7) with a discrete AR(2) error process. Although
the discretised CAR(2) process follows an ARMA(2, 1) model rather than an
AR(2) [see He and Wang (1989)] we will be able to establish the connection
between the BLUE in the discrete and continuous-time models and hence
derive the limiting form of the discrete BLUE and its covariance matrix.
4. Models with more than once differentiable error processes. If
T = [A, B] and q > 1 then solving the Wiener-Hopf type equation (2.5)
numerically is virtually impossible in view of the fact that the problem
is severely ill-posed. Derivation of explicit forms of the BLUE for smooth
kernels with q > 1 is hence extremely important. We did not find any general
results on the form of the BLUE in such cases. In particular, the well-known
paper Sacks and Ylvisaker (1970) dealing with these kernels does not contain
any specific examples. In Theorem 3.2 we have already established a general
result that can be used for deriving explicit forms for the BLUE for q > 1
times integrated kernels, which can be used repeatedly for this purpose. We
can also formulate a result similar to Theorem 3.1. However, already for
q = 2, even a formulation of such theorem would take a couple of pages and
hence its usefulness would be very doubtful.
In this section, we indicate how the general methodologies developed in the
previous sections can be extended to error processes with q > 1 by two
examples: twice integrated Brownian motion and CAR(p) error models with
p ≥ 3, but other cases can be treated very similarly.
4.1. Twice integrated Brownian motion.
Proposition 4.1. Consider the regression model (6.8) where the error
process is the twice integrated Brownian motion with the covariance kernel
K(t, s) = t5 /5! − st4 /4! + s2 t3 /12, t < s.
BLUE IN CONTINUOUS TIME REGRESSION MODELS
19
Suppose that f is 6 times differentiable and define the vector-measures
ζ0 (dt) = zA δA (dt) + zB δB (dt) + z(t)dt,
ζ1 (dt) = z1,A δA (dt) + z1,B δB (dt),
ζ2 (dt) = z2,A δA (dt) + z2,B δB (dt),
where z(t) = f (6) (t),
zA = (A5 f (5) (A) − 60A2 f (2) (A) + 360Af (1) (A) − 720f (A))/A5 ,
z1,A = −(A4 f (4) (A) − 36A2 f (2) (A) + 192Af (1) (A) − 360f (A))/A4 ,
z2,A = (A3 f (3) (A) − 9A2 f (2) (A) + 36Af (1) (A) − 60f (A))/A3 ,
zB = −f (5) (B), z1,B = f (4) (B), z2,B = −f (3) (B).
Then the estimator θ̂G0 ,G1 ,G2 defined by (2.1) (for q = 2) with Gi (dt) =
C −1 ζi (dt) (i = 0, 1, 2),
Z
Z
Z
(1)
T
T
f (2) (t)ζ2T (dt),
f (t)ζ1 (dt) +
f (t)ζ0 (dt) +
C=
T
T
T
is the BLUE with covariance matrix C −1 .
4.2. CAR(p) models with p≥ 3. Consider the regression model (6.8), which
can be observed at all t ∈ [A, B] and the error process has the continuous
autoregressive (CAR) structure of order p. Formally, a CAR(p) process is a
solution of the linear stochastic differential equation of the form
dε(p−1) (t) = ã1 ε(p−1) (t) + . . . + ãp ε(t) + σ02 dW (t),
where Var(ε(t)) = σ 2 and W is a standard Wiener process, [see Brockwell et al.
(2007)]. Note that the process ε has continuous derivatives ε(1) (t), . . . , ε(p−1) (t)
at the point t and, consequently, the process {y(t) = θ T f (t)+ε(t)| t ∈ [A, B]}
is continuously differentiable p − 1 times on the interval [A, B] with drift
θ T f (t). Define the vector-functions
z(t) = (τ0 f (t) + τ2 f (2) (t) + . . . + f (2p) (t))/s2p−1 ,
and vectors
zj,A =
zj,B =
X2p−j−1
l=0
X2p−j−1
l=0
γl,j,A f (j)(A)/s2p−1 ,
γl,j,B f (j) (B)/s2p−1
for j = 0, 1, . . . , p − 1, where s2p−1 = K (2p−1) (s−, s) − K (2p−1) (s+, s).
20
H. DETTE ET AL.
Proposition 4.2. Consider the regression model (6.8) with CAR(p) errors. Define the vector-measures
ζ0 (dt) = z0,A δA (dt) + z0,B δB (dt) + z(t)dt,
ζj (dt) = zj,A δA (dt) + zj,B δB (dt), j = 1, . . . , p − 1,
for j = 1, . . . , p−1. Then there exist constants τ0 , τ2 . . . , τ2(p−1) and γl,j,A , γl,j,B ,
such that the estimator θ̂G0 ,G1 ,...,Gp−1 defined by (2.1) (for q = p − 1) with
Gj (dt) = C −1 ζj (dt) (i = 0, 1, . . . , p − 1),
C=
Z
T
f (t)ζ0T (dt) +
p−1 Z
X
j=1
T
f (j) (t)ζjT (dt),
is a BLUE with covariance matrix C −1 .
Let us consider the construction of a BLUE for model (6.8) with a CAR(3)
error process in more detail. One of several possible forms for the covariance
function for the CAR(3) process is given by
(4.1)
ρ(t) = c1 e−λ1 |t| + c2 e−λ2 |t| + c3 e−λ3 |t| ,
where λ1 , λ2 , λ3 are the roots of the autoregressive polynomial ã(z) = z 3 +
ã1 z 2 + ã2 z + ã3 ,
cj =
1
kj
, kj = ′
,
k1 + k2 + k3
ã (λj )ã(−λj )
λi 6= λj , λi > 0, i, j = 1, . . . , 3, see Brockwell (2001). Specifically, we have
λ2 λ3 (λ2 + λ3 )
,
(λ1 − λ2 )(λ1 − λ3 )(λ1 + λ2 + λ3 )
λ1 λ3 (λ1 + λ3 )
,
c2 =
(λ2 − λ1 )(λ2 − λ3 )(λ1 + λ2 + λ3 )
λ1 λ2 (λ1 + λ2 )
c3 =
.
(λ3 − λ1 )(λ3 − λ2 )(λ1 + λ2 + λ3 )
c1 =
In this case, a BLUE is given in Proposition 4.2 with the following parameters:
τ0 = −λ21 λ22 λ23 , τ2 = λ21 λ22 + λ21 λ23 + λ22 λ23 , τ4 = −λ21 − λ22 − λ23 ,
2λ1 λ2 λ3 (λ1 + λ2 )(λ1 + λ3 )(λ2 + λ3 )
=2
s5 =
λ1 + λ2 + λ3
Q
i λi
Q
i6=j (λi + λj )
P
,
i λi
BLUE IN CONTINUOUS TIME REGRESSION MODELS
21
P
Q
z0,A = f (5) (A) − i λ2i f (3) (A) − i λi f (2) (A)
P
Q P
Q P
+[ i6=j λ2i λ2j + i λi i λi ]f (1) (A) − i λi i6=j λi λj f (A)
P
Q
Q P
z1,A = −f (4) (A) + i,j λi λj f (2) (A) − i6=j (λi + λj )f (1) (A) + i λi i λi f (A)
P
P
Q
z2,A = f (3) (A) − i λi f (2) (A) + i6=j λi λj f (1) (A) − i λi f (A)
Q
P
−z0,B = f (5) (B) − i λ2i f (3) (B) − i λi f (2) (B)
Q P
Q P
P
+[ i6=j λ2i λ2j + i λi i λi ]f (1) (B) − i λi i6=j λi λj f (B)
P
Q
Q P
−z1,B = −f (4) (B) + i,j λi λj f (2) (B) − i6=j (λi + λj )f (1) (B) + i λi i λi f (B)
P
P
Q
−z2,B = f (3) (B) − i λi f (2) (B) + i6=j λi λj f (1) (B) − i λi f (B)
If we set λ1 = λ2 = λ3 = λ then the above formulas give the explicit form
of the BLUE for the Matérn
kernel with
√ 5/2; that is, the kernel
√
parameter
defined by ρ(t) = 1 + 5tλ + 5t2 λ2 /3 exp − 5tλ .
5. Numerical study. In this section, we describe some numerical results
on comparison of the accuracy of various estimators for the parameters in
the regression models (6.8) with [A, B] = [1, 2] and the integrated Brownian
motion as error process. The kernel K(t, s) is given in (3.5) and the explicit
form of the covariance matrix of the continuous BLUE can be found in
Example 3.1. We denote this estimator by θ̂cont.BLU E . We are interested in
the efficiency of various estimators for this differentiable error process. For a
given N (in the tables, we use N = 3, 5, 10), we consider the following four
estimators that use 2N observations:
• θ̂BLU E (N, N ): discrete BLUE based on observations y(t1 ), . . . , y(tN ),
y ′ (t1 ), . . . , y ′ (tN ) with ti = 1 + (i − 1)/(N − 1), i = 1, . . . , N . This
estimator uses N observations of the original process and its derivative
(at equidistant points).
• θ̂BLU E (2N − 2, 2): discrete BLUE based on observations y(t1 ), . . . ,
y(t2N −2 ), y ′ (1), y ′ (2) with ti = 1 + (i − 1)/(2N − 3), i = 1, . . . , 2N − 3.
This estimator uses 2N − 2 observations of the original process (at
equidistant points) and observations of its derivative at the boundary
points of the design space.
• θ̂BLU E (2N, 0): discrete BLUE based on observations y(t1 ), . . . , y(t2N )
with ti = 1 + (i − 1)/(2N − 1), i = 1, . . . , 2N . This estimator uses
2N observations of the original process (at equidistant points) and no
observations from its derivative.
22
H. DETTE ET AL.
• θ̂OLSE (2N, 0): ordinary least square estimator (OLSE) based on observations y(t1 ), . . . , y(t2N ) with ti = 1 + (i − 1)/(2N − 1), i = 1, . . . , 2N .
This estimator uses 2N observations of the original process (at equidistant points) and no observations from its derivative.
In Table 1 – 3 we use the results derived in this paper to calculate the
efficiencies
(5.1)
Eff(θ̃) =
Var(θ̂cont.BLU E )
,
Var(θ̃)
where θ̃ is one of the four estimators under consideration. In particular we
consider three different scenarios for the drift in model (6.8) defined by
(5.2)
m = 1, f (t) = 1
(5.3)
m = 3, f (t) = (1, sin(3π), cos(3π))T
(5.4)
m = 5, f (t) = (1, t, t2 , 1/t, 1/t2 )T
Table 1
The efficiency defined by (5.1) for four different estimators based on 2N observations.
The drift function is given by (5.2)
θ̃
N = 3 N = 5 N = 10
θ̂BLU E (N, N )
1
1
1
θ̂BLU E (2N − 2, 2)
1
1
1
θ̂BLU E (2N, 0)
0.8593 0.9147 0.9570
θ̂OLSE (2N, 0)
0.0732 0.0733 0.0734
Table 2
The efficiency defined by (5.1) for four different estimators based on 2N observations.
The drift function is given by (5.3)
θ̃
θ̂BLU E (N, N )
θ̂BLU E (2N − 2, 2)
θ̂BLU E (2N, 0)
θ̂OLSE (2N, 0)
N=3
0.41246
0.45573
0.47796
0.00113
N =5
0.92907
0.98706
0.77195
0.00137
N = 10
0.99680
0.99972
0.89641
0.00218
The results of Table 1 – 3 are very typical for many regression models with
differentiable error processes (i.e. q = 1) and can be summarized as follows.
Any BLUE is far superior to the OLSE and any BLUE becomes very efficient
BLUE IN CONTINUOUS TIME REGRESSION MODELS
23
Table 3
The efficiency defined by (5.1) for four different estimators based on 2N observations.
The drift function is given by (5.4)
θ̃
θ̂BLU E (N, N )
θ̂BLU E (2N − 2, 2)
θ̂BLU E (2N, 0)
θ̂OLSE (2N, 0)
N=3
0.69608
0.86903
0.10040
0.08873
N =5
0.95988
0.99379
0.33338
0.14103
N = 10
0.99791
0.99981
0.62529
0.11890
when N is large. Moreover, the use of information from the derivatives in
constructing BLUEs makes them more efficient than the BLUE which only
uses values of {y(t)|t ∈ T }. We also emphasize that the BLUEs which use
more than two values of the derivative y ′ of the process have lower efficiency
than the BLUE that uses exactly two values of derivatives, y ′ (A) and y ′ (B).
Therefore the best way of constructing the BLUE for N observations in the
interval [A, B] is to emulate the asymptotic BLUE: that is, to use y ′ (A)
and y ′ (B) but for the other N − 2 observations use values of the process
{y(t)|t ∈ T }.
6. Appendix.
T is
6.1. Proof of Lemma 2.1. The mean of θ̂G
T
E[θ̂G
]
=θ
T
Z
T
F (t)G (dt) = θ
T
T
q Z
X
i=0
T
f (i) (t)GTi (dt) .
This implies that the estimator θ̂G is unbiased if and only if
q Z
X
(6.1)
f (i) (t)GTi (dt) = Im .
i=0
T
Since Gi = C −1 ζi , we have
q Z
q Z
X
X
T
T
(i)
T
f (i) (t)ζiT (dt)C −1 = C T C −1 = Im ,
f (t)Gi (dt) =
i=0
T
i=0
T
which completes the proof.
6.2. Proof of Theorem 2.1 .
I. We will call a signed matrix-measure G unbiased if the associated estimator θbG defined in (2.1) is unbiased. The set of all unbiased signed matrixmeasures will be denoted by S. This set is convex.
24
H. DETTE ET AL.
The covariance matrix of any estimator θbG is the matrix-valued function
φ(G) = Var(θ̂G ) defined in (2.3). The BLUE minimizes this matrix-valued
function on the set S.
Introduce the vector-function d : T × S → Rm by
d(s, G) =
q Z
X
T
j=0
K (j) (t, s)Gj (dt) − φ(G)f (s) .
The validity of (2.4) for all s ∈ T is equivalent to the validity of d(s, G) =
0m×1 for all s ∈ T . Hence we are going to prove that θbG is the BLUE if
and only if d(s, G) = 0m×1 for all s ∈ T . For this purpose we will need the
following auxiliary result.
Lemma 6.1.
For any G ∈ S we have
Z
d(s, G)GT (ds) = 0m×m ,
T
where d(s, G) = (d(s, G), d(1) (s, G), . . . , d(q) (s, G)) is a m × (q + 1) matrix.
Proof of Lemma 6.1 Using the unbiasedness condition (6.1) for G, we
have
Z
Z Z
Z
T
T
F (s)GT (ds)
G(dt)K(t, s)G (ds) − φ(G)
d(s, G)G (ds) =
T
T
T
T
= φ(G) − φ(G)Im = 0m×m
where F (s) = (f (s), f (1) (s), . . . , f (q) (s)).
For any two measures G and H in S, denote
Z Z
G(dt)K(t, s)H T (ds)
Φ(G, H) =
T
T
which is a matrix of size m × m. Note that φ(G) = Φ(G, G) for any G ∈ S.
For any two matrix-measures G and H in S and any real α, the matrixvalued function
φ((1 − α)G + αH) = (1 − α)2 φ(G) + α2 φ(H) + α(1 − α) [Φ(G, H) + Φ(H, G)]
is quadratic in α. Also we have ∂ 2 φ((1 − α)G + αH)/∂α2 = 2φ(G − H) ≥ 0
and hence φ(·) is a matrix-convex function on the space S (see e.g. Hansen and Tomiyama
(2007) for properties of matrix-convex functions).
BLUE IN CONTINUOUS TIME REGRESSION MODELS
25
Since the matrix-function φ((1 − α)G + αH) is quadratic and convex in
α ∈ R, the assertion that G is the optimal matrix measure minimizing φ(·)
on S, is equivalent to
∂φ((1 − α)G + αH)
∂α
(6.2)
α=0
= 0, ∀ H ∈ S .
The directional derivative of φ((1 − α)G + αH) as α → 0 is
(6.3)
∂
φ((1 − α)G + αH)
∂α
α=0
= Φ(G, H) + Φ(H, G) − 2φ(G).
R
To rewrite (6.3), we note that T d(s, G)H T (ds) can be written as
Z
Z
(6.4)
F (s)H T (ds)
d(s, G)H T (ds) = Φ(G, H) − φ(G)
T
T
= Φ(G, H) − φ(G),
where in the last equality we have used the unbiasedness condition (6.1) for
H. Using (6.3), (6.4) and the fact that the matrix Φ(H, G) − φ(G) is the
transpose of Φ(G, H) − φ(G) we obtain
T
Z
Z
∂
d(s, G)H T (ds) .
d(s, G)H T (ds) +
=
φ((1−α)G + αH)
α=0
∂α
T
T
(6.5)
Consequently, if d(s, G) = 0m×1 for all s ∈ T , then (6.2) holds and hence G
gives the BLUE.
II. Assume now that G gives the BLUE θ̂G . This implies, first, that (6.2)
holds and second, for all c ∈ Rm cT φ(G)c ≤ cT φ(H)c, for any H ∈ S. Let
us deduce that d(s, G) = 0m×1 for all s ∈ T (which is equivalent to validity
of (2.4)). We are going to prove this by contradiction.
Assume that there exists s0 ∈ T such that d(s0 , G) 6= 0. Define the signed
matrix-measure ζ = (ζ0 , ζ1 , . . . , ζq ) with ζ0 (ds) = G0 (ds) + κd(s0 , G)δs0 (ds),
κ 6= 0, and ζi (ds) = Gi (ds) for
R i = 1, . .T. , q.
Since G is unbiased, CG = T G(dt)F
(t) = Im . For any small positive or
R
small negative κ, the matrix Cζ = T ζ(dt)F T (t) = Im + κd(s0 , G)f T (s0 ) is
non-degenerate and its eigenvalues are close to 1.
In view of Lemma 2.1, H(ds) = Cζ−1 ζ(ds) is an unbiased matrix-measure.
Using the identity (6.5) and Lemma 6.1 we obtain for the measure Gα =
(1 − α)G + αH:
∂φ(Gα )
∂α
T
α=0
= κd(s0 , G)dT (s0 , G)Cζ−1 + κCζ−1 d(s0 , G)dT (s0 , G).
26
H. DETTE ET AL.
Write this as ∂φ(Gα )/∂α α=0 = κ(X0 AT + AX0 ), where A = Cζ−1 and
X0 = d(s0 , G)dT (s0 , G) is a symmetric matrix.
For any given A, the homogeneous Lyapunov matrix equation XAT +AX = 0
has only the trivial solution X = 0 if and only if A and −A have no common
eigenvalues, see [§3, Ch. 8 in Gantmacher (1959)]; this is the case when
A = Cζ−1 and κ is small enough.
This yields that for X = X0 , the matrix X0 AT +AX0 is a non-zero symmetric
matrix. Therefore, there exists a vector c ∈ Rm such that the directional
derivative of cT φ(Gα )c is non-zero. For any such c, cT φ(Gα )c < cT φ(G)c for
either small positive or small negative α and hence θ̂G is not the BLUE.
Thus, the assumption of the existence of an s0 ∈ T such that d(s0 , G) 6= 0
yields a contradiction to the fact that G gives the BLUE. This completes
the proof that the equality (2.4) is necessary and sufficient for the estimator
θ̂G to be the BLUE .
6.3. Proof of Lemma 2.2. We repeat i times the integration by parts formula
Z
Z
B
(i)
(i−1)
ψ (i−1) (t)ϕ(1) (t)dt
ψ (t)ϕ(t)dt = ψ
(t)ϕ(t) −
A
T
T
for any differentiable function ψ(t). This gives
Z
i
X
(−1)j−1 ψ (i−j) (t)ϕ(j−1) (t)
ψ (i) (t)ϕ(t)dt =
T
j=1
B
A
+ (−1)i
Z
ψ(t)ϕ(i) (t)dt.
T
Using the above equality with ψ(t) = y (i) (t) we obtain that the expectation of two estimators coincide. Also, using the above equality with ψ(t) =
K (i) (t, s) we obtain that the covariance matrices of the two estimators coincide.
6.4. Proof of Proposition 2.3. Straightforward calculus shows that
Z
Z
K(t, s)f (2) (t)dt/(λ1 + λ2 )
K(t, s)ζ(dt) = K(A, s)zA + K(B, s)zB −
T
T
= K(A, s)zA + K(B, s)zB
+ − K(t, s)f (1) (t)|sA + K (1) (t, s)f (t)|s−
A
(1)
− K(t, s)f (1) (t)|B
(t, s)f (t)|B
s +K
s+ /(λ1 + λ2 )
= (1 + λ1 A − λ2 s)zA + (1 + λ1 s − λ2 B)zB + f (s)
+ K(A, s)f (1) (A) − K (1) (A, s)f (A)
− K(B, s)f (1) (B) + K (1) (B, s)f (B) /(λ1 +λ2 ) = f (s).
Therefore, the conditions of Proposition 2.1 are fulfilled.
BLUE IN CONTINUOUS TIME REGRESSION MODELS
27
6.5. Proof of Theorem 3.1. It is easy to see that θ̂G0 ,G1 is unbiased. Further
we are going to use Corollary 3.1 which gives the sufficient condition for an
estimator to be the BLUE. We will show that the identity
Z B
Z B
(6.6)
K (1) (t, s)ζ1 (dt) = f (s)
LHS =
K(t, s)ζ0 (dt) +
A
A
holds for all s ∈ [A, B]. By the definition of the measure ζ it follows that
LHS = zA K(A, s) + zB K(B, s) + IA + IB + z1,A K (1) (A, s) + z1,B K (1) (B, s),
Rs
RB
where IA = A K(t, s)z(t)dt, IB = s K(t, s)z(t)dt. Indeed, for the vectorfunction z(t) = τ0 f (t) − τ2 f (2) (t) + f (4) (t), we have
Z s
Z s
Z s
(2)
s3 IA = τ0 K(t, s)f (t)dt − τ2 K(t, s)f (t)dt + K(t, s)f (4) (t)dt
A
A
ZAs
= τ0 K(t, s)f (t)dt − τ2 K(t, s)f (1) (t)|sA + τ2 K (1) (t, s)f (t)|sA
A
Z s
−τ2 K (2) (t, s)f (t)dt + K(t, s)f (3) (t)|sA − K (1) (t, s)f (2) (t)|s−
A
A
Z s
(3)
K (4) (t, s)f (t)dt.
(t, s)f (t)|s−
+K (2) (t, s)f (1) (t)|s−
A +
A −K
A
By construction, the coefficients τ0 , τ2 , are chosen such that the equality (3.3)
holds for all t ∈ [A, B] and any s, implying that integrals in the expression
for IA are cancelled. Thus, we obtain
s3 IA = +τ2 K(A, s)f (1) (A) − τ2 K (1) (A, s)f (A) − K(A, s)f (3) (A)
+K (1) (A, s)f (2) (A) − K (2) (A, s)f (1) (A) + K (3) (A, s)f (A)
−τ2 K(s−, s)f (1) (s) + τ2 K (1) (s−, s)f (s) + K(s−, s)f (3) (s)
−K (1) (s−, s)f (2) (s) + K (2) (s−, s)f (1) (s) − K (3) (s−, s)f (s).
Similarly we have
s3 IB = −τ2 K(B, s)f (1) (B) + τ2 K (1) (B, s)f (B) + K(B, s)f (3) (B)
−K (1) (B, s)f (2) (B) + K (2) (B, s)f (1) (B) − K (3) (B, s)f (B)
+τ2 K(s+, s)f (1) (s) − τ2 K (1) (s+, s)f (s) − K(s+, s)f (3) (s)
+K (1) (s+, s)f (2) (s) − K (2) (s+, s)f (1) (s) + K (3) (s+, s)f (s).
28
H. DETTE ET AL.
Using the assumption on the derivatives of the covariance kernel K(t, s), we
obtain
s3 (IA + IB ) = τ2 K(A, s)f (1) (A) − τ2 K (1) (A, s)f (A) − K(A, s)f (3) (A)
+K (1) (A, s)f (2) (A) − K (2) (A, s)f (1) (A) + K (3) (A, s)f (A)
−τ2 K(B, s)f (1) (B) + τ2 K (1) (B, s)f (B) + K(B, s)f (3) (B)
−K (1) (B, s)f (2) (B) + K (2) (B, s)f (1) (B) − K (3) (B, s)f (B) + s3 f (s).
Also we have
s3 (zA K(A, s) + z1,A K (1) (A, s)) =
= f (3) (A) − γ1,A f (1) (A) + γ0,A f (A) K(A, s)
+ − f (2) (A) + β1,A f (1) (A) − β0,A f (A) K (1) (A, s)
= f (3) (A)K(A, s) + (−γ1,A K(A, s) + β1,A K (1) (A, s))f (1) (A)
−K (1) (A, s)f (2) (A) + (γ0,A K(A, s) − β0,A K (1) (A, s))f (A),
and a similar result at the point t = B. Putting these expressions into
(6.6) and using the assumption that constants γ1,A , β1,A , γ0,A , β0,A and
γ1,B , β1,B , γ0,B , β0,B are chosen such that the sum of the functions defined
in (3.2) is identically equal to zero, we obtain
Z
B
K(t, s)ζ0 (dt) +
A
Z
B
K (1) (t, s)ζ1 (dt) = f (s);
A
this completes the proof.
6.6. Proof of Theorem 3.2. Observing (3.7) the vector c is can be written
as
Z A Z B
K(t,
s)ζ
(dt)
−
f
(s)
ds
c =
0
a
A
Z A Z B
Z s Z B
=
K(t, s′ )ζ0 (dt) − f (s′ ) ds′ +
K(t, s′ )ζ0 (dt) − f (s′ ) ds′
a
=
Z
B
A
Z
A
s
a
K(t, s′ )ds′ ζ0 (dt) −
Z
A
s
a
f (s′ )ds′ =
Z
A
B
A
R(1) (t, s)ζ0 (dt) − f˜(s).
We now show that equation (2.5) in Proposition 2.1 holds for K = R, q =
1, f = f˜ and ζi = ζ̃i . Observing (3.10) and the definition of ζ̃i in Theorem 3.2
BLUE IN CONTINUOUS TIME REGRESSION MODELS
29
we obtain
Z
Z
R(1) (t, s)ζ̃1 (dt)
R(t, s)ζ̃0 (dt) +
T
T
Z
Z
Z
(1)
R(1) (t, s)ζ0 (dt)
R (t, s)η1 (dt) +
R(t, s)η0 (dt) +
= −c
T
T
T
= −c · 1 + fe(s) + c = fe(s).
6.7. Proof of Propositions 3.3 and 3.4. For the sake of brevity we only give
a proof of Proposition 3.3, the other result follows by similar arguments. Direct calculus gives s3 = K (3) (s+, s)−K (3) (s−, s) = 2λ1 λ2 (λ1 +λ2 ). Then we
obtain that the identity (3.3) holds for τ0 = λ21 λ22 and τ2 = λ21 + λ22 . Straightforward calculations show that identities (3.2) hold with the specified values
of constants γ1 , γ0 , β1 , β0 .
Acknowledgements. This work has been supported in part by the Collaborative Research Center “Statistical modeling of nonlinear dynamic processes” (SFB 823, Project C2) of the German Research Foundation (DFG)
and the National Institute Of General Medical Sciences of the National
Institutes of Health under Award Number R01GM107639. The content is
solely the responsibility of the authors and does not necessarily represent
the official views of the National Institutes of Health. The work of Andrey
Pepelyshev was partly supported by the project “Actual problems of design
and analysis for regression models” (6.38.435.2015) of St. Petersburg State
University. The authors would like to thank Martina Stein, who typed parts
of this manuscript with considerable technical expertise.
References.
Anderson, T. W. (1970). Efficient estimation of regression coefficients in time series.
Technical report, DTIC Document.
Berlinet, A. and Thomas-Agnan, C. (2011). Reproducing Kernel Hilbert Spaces in Probability and Statistics. Springer.
Brockwell, P. (2001). Continuous-time ARMA processes. Handbook of Statistics, 19:249–
276.
Brockwell, P., Davis, R., and Yang, Y. (2007). Continuous-time Gaussian autoregression.
Statistica Sinica, 17(1):63.
Dette, H., Konstantinou, M., and Zhigljavsky, A. (2017). A simple approach to optimal
designs for correlated observations. The Annals of Statistics. To appear.
Dette, H., Pepelyshev, A., and Zhigljavsky, A. (2013). Optimal design for linear models
with correlated observations. The Annals of Statistics, 41(1):143–176.
Dette, H., Pepelyshev, A., and Zhigljavsky, A. (2016a). Optimal designs in regression with
correlated errors. The Annals of Statistics, 44(1):113–152.
Dette, H., Pepelyshev, A., and Zhigljavsky, A. (2016b). Supplement to “best linear unbiased estimators in continuous time regression models”. Annals of Statistics.
30
H. DETTE ET AL.
Fedorov, V. V. and Müller, W. G. (2007). Optimum design for correlated fields via covariance kernel expansions. In mODa 8-Advances in Model-Oriented Design and Analysis,
pages 57–66. Springer.
Gantmacher, F. R. (1959). The Theory of Matrices. Chelsea, New York.
Grenander, U. (1950). Stochastic processes and statistical inference. Ark. Mat., 1:195–277.
Grenander, U. (1954). On the estimation of regression coefficients in the case of an autocorrelated disturbance. The Annals of Mathematical Statistics, 25(2):252–272.
Hajek, J. (1956). Linear estimation of the mean value of a stationary random process with
convex correlation function. Czechoslovak Mathematical Journal, 6(81):94–117.
Hannan, E. (1975). Linear regression in continuous time. Journal of the Australian Mathematical Society (Series A), 19(02):146–159.
Hansen, F. and Tomiyama, J. (2007). Differential analysis of matrix convex functions.
Linear Algebra and its Applications, 420(1):102–116.
He, S. and Wang, J. (1989). On embedding a discrete-parameter ARMA model in a
continuous-parameter ARMA model. Journal of Time Series Analysis, 10(4):315–323.
Kanwal, R. (1997). Linear Integral Equations. Birkhauser, Boston.
Kholevo, A. (1969). On estimates of regression coefficients. Theory of Probability & Its
Applications, 14(1):79–104.
Müller, W. G. and Pázman, A. (2003). Measures for designs in experiments with correlated
errors. Biometrika, 90(2):423–434.
Näther, W. (1985). Effective Observation of Random Fields. Teubner Verlagsgesellschaft,
Leipzig.
Parzen, E. (1961). An approach to time series analysis. The Annals of Mathematical
Statistics, pages 951–989.
Pisarenko, V. F. and Rozanov, J. A. (1963). On certain problems for stationary processes
leading to integral equations related to Wiener-Hopf equations. Problemy Peredaci
Informacii, 14:113–135.
Pronzato, L. and Müller, W. G. (2012). Design of computer experiments: space filling and
beyond. Statistics and Computing, 22(3):681–701.
Ramm, A. G. and Charlot, J. (1980). Theory and Applications of Some New Classes of
Integral Equations. Springer.
Rasmussen, C. and Williams, C. (2006). Gaussian processes for machine learning. MIT
Press.
Ritter, K. (2000). Average-Case Analysis of Numerical Problems. Springer.
Rosenblatt, M. (1956). Some regression problems in time series analysis. In Proceedings
of the Third Berkeley Symposium on Mathematical Statistics and Probability, Volume
1: Contributions to the Theory of Statistics, pages 165–186. University of California.
Sacks, J. and Ylvisaker, N. D. (1966). Designs for regression problems with correlated
errors. Annals of Mathematical Statistics, 37:66–89.
Sacks, J. and Ylvisaker, N. D. (1968). Designs for regression problems with correlated
errors; many parameters. Annals of Mathematical Statistics, 39:49–69.
Sacks, J. and Ylvisaker, N. D. (1970). Statistical designs and integral approximation.
Proc. 12th Bienn. Semin. Can. Math. Congr.(R. Pyke, ed.), pages 115–136.
Submitted to the Annals of Statistics
SUPPLEMENT TO “BEST LINEAR UNBIASED
ESTIMATORS IN CONTINUOUS TIME REGRESSION
MODELS”
By Holger Dette‡ , Andrey Pepelyshev§ and Anatoly
Zhigljavsky§
Ruhr-Universität Bochum‡ and Cardiff University§
We demonstrate that the covariance matrix of the BLUE in the
continuous-time regression model model with a CAR(2) error process
can be obtained as limit of the covariance matrix of a BLUE in the
discrete regression model with observations at equidistant points and
a discrete AR(2) error process.
Here we investigate the approximation of the BLUE for continuous-time
regression time models with a CAR(2) error process (see Section 4 of the
paper) by the BLUE in the model
(6.7) y(ti ) = θ T f (ti ) + ǫ(ti ) ,
A ≤ t1 < t2 . . . < tN −1 < tN ≤ B ,
where the errors follow a discrete AR(2) process. This model will be abbreviated as DAR(2) throughout this section . The main difficulty to establish the
connection between the discrete and continuous AR(2) cased lies in the fact
that the discretised CAR(2) process follows an ARMA(2, 1) model rather
than the AR(2), see He and Wang (1989). To be precise, assume that the
observations in the continuous-time regression model
(6.8)
y(t) = θ T f (t) + ǫ(t) ,
t ∈ [A, B],
are taken at N equidistant points of the form
(6.9)
tj = A + (j − 1)∆ , (j = 1, . . . , N )
on the interval [A, B], where ∆ = (B − A)/(N − 1), and that the errors
ǫ1 , . . . , ǫN satisfy the discrete AR(2) equation
(6.10)
ǫj − a1 ǫj−1 − a2 ǫj−2 = ςj ,
Primary 62K05; secondary 31A10
Keywords and phrases: linear regression, correlated observations, signed measures, optimal design, BLUE, AR processes, continuous autoregressive model
31
32
H. DETTE ET AL.
where ςj are Gaussian independent identically distributed random variables
with mean 0 and variance σς2 = σ 2 (1 + a2 )((1 − a2 ) − a21 )/(1 − a2 ). Here
we make a usual assumption that the equation (6.10) defines the AR(2)
process for j ∈ {. . . , −2, −1, 0, 1, 2, . . .} but we only take the values such that
j ∈ {1, 2, . . . , N }. Let rk = E[ǫj ǫj+k ] denote the autocovariance function of
the AR(2) process {ǫ1 , . . . , ǫN } and assume without loss of generality that
σ 2 = 1.
There are in fact three different forms of the autocovariance function (note
that we assume throughout σ 2 = 1) of CAR(2) processes [see e.g. formulas
(14)–(16) in He and Wang (1989)], which are given by
(6.11)
ρ1 (t) =
λ1
λ2
e−λ1 |t| −
e−λ2 |t| ,
λ2 − λ1
λ2 − λ1
where λ1 6= λ2 , λ1 > 0, λ2 > 0, by
n
o
λ
ρ2 (t) = e−λ|t| cos(q|t|) + sin(q|t|) ,
q
(6.12)
where λ > 0, q > 0, and by
ρ3 (t) = e−λ|t| (1 + λ|t|) ,
(6.13)
where λ > 0.
Also, there are three forms of autocovariance functions of the discrete AR(2)
process of the form (6.10) [see formulas (11)–(13) in He and Wang (1989)],
which are given by
(6.14)
(1)
rk = Cpk1 + (1 − C)pk2 ,
C=
(1 − p22 )p1
,
(1 − p22 )p1 − (1 − p21 )p2
where j ≥ 0, p1 6= p2 , 0 < |p1 |, |p2 | < 1; by
(6.15)
(2)
rk = pk cos(bk) + C sin(bk) ,
C = cot(b)
1 − p2
,
1 + p2
where 0 < p < 1, 0 < b < 2π and b 6= π, and finally by
(6.16)
(3)
rk = pk (1 + kC),
C=
1 − p2
,
1 + p2
where 0 < |p| < 1. Each form of the autocovariance function should be
considered individually. However, we can formulate the following general
statement.
SUPPLEMENT TO “BLUE IN CONTINUOUS TIME REGRESSION MODELS” 33
Theorem 6.1. Consider the multi-parameter model (6.8) such that the
errors follow the AR(2) model. Assume that f (·) is four times continuously
differentiable. Define the following constants depending on the form of the
autocovariance function rk . If rk is of the form (6.14), set
ln(p2 )
ln(p1 )
, λ2 = −
,
∆
∆
= λ21 λ22 , τ2 = λ21 + λ22 , β1 = λ1 + λ2 , β0 = λ1 λ2 ,
λ1 = −
τ0
γ1 = λ21 + λ1 λ2 + λ22 , γ0 = λ1 λ2 (λ1 + λ2 ), s3 = 2λ1 λ2 (λ1 + λ2 ).
If rk is of the form (6.15), set
b
ln(p)
, q=− ,
∆
∆
2
2 2
= (λ + q ) , τ2 = 2(λ2 − q 2 ), β1 = 2λ, β0 = λ2 + q 2 ,
λ = −
τ0
γ1 = 3λ2 − q 2 , γ0 = 2λ(λ2 + q 2 ), s3 = 4λ(λ2 + q 2 ).
If rk is of the form (6.16), set
ln(p)
, τ0 = λ4 , τ2 = 2λ2 , β1 = 2λ, β0 = λ2 ,
∆
= 3λ2 , γ0 = 2λ3 , s3 = 4λ3 .
λ = −
γ1
For large N , the discrete BLUE θ̂BLUE,N based on N observations at the points
(6.9) can be approximated by the continuous estimator
∗
′
′
Z
θ̂ = D z1,B y (B)+z1,A y (A)+zA y(A)+zB y(B)+ z(t)y(t)dt
T
where
Z
−1
∗
(1)
T
(1)
T
T
T
.
D = f (B)z1,B +f (A)z1,A +f (A)zA +f (B)zB + f (t)z T (t)dt
T
Moreover, for this approximation, we have D ∗ = limN →∞ Var(θ̂BLUE,N ), i.e.
D ∗ is the limit of the variance of the discrete BLUE as N → ∞. Here the
quantities z(t), zA , zB , z1,A and z1,B in the continuous estimator are defined
by
z(t) = − τ2 f (2) (t) − τ0 f (t) − f (4) (t) /s3 ,
zA = f (3) (A) − γ1 f (1) (A) + γ0 f (A) /s3 ,
zB = − f (3) (B) + γ1 f (1) (B) + γ0 f (B) /s3 ,
z1,A = − f (2) (A) + β1 f (1) (A) − β0 f (A) /s3 ,
z1,B = f (2) (B) + β1 f (1) (B) + β0 f (B) /s3 .
34
H. DETTE ET AL.
Proof. It is well known that the inverse of the covariance matrix Σ =
(E[ǫj ǫk ])j,k of the discrete AR(2) process is a five-diagonal matrix, i.e.
(6.17)
Σ−1
k11
k21
k2
1
0
= .
S ..
k12
k22
k1
k2
..
.
k2
k1
k0
k1
..
.
0
0
0
0 0
k2 0
k1 k2
k0 k1
.. ..
.
.
k2 k1
0 k2
0 0
0
0
0
k2
..
.
k0
k1
k2
...
...
...
,
..
.
k1 k2
k22 k12
k21 k11
where the non-vanishing elements are given by k0 = 1 + a21 + a22 , k1 =
−a1 + a1 a2 , k2 = −a2 , k11 = 1, k12 = k21 = −a1 , k22 = 1 + a21 and
S = (1 + a1 − a2 )(1 − a1 − a2 )(1 + a2 )/(1 − a2 ). Using the explicit form (6.17)
for Σ−1 we immediately obtain the following result.
Corollary 6.1. Consider the linear regression model (6.8) with observations at N equidistant points (6.9) and errors that follow the discrete AR(2)
model (6.10). Let hi be the i-th column of matrix H = X T Σ and fi = f (ti ).
Then the vectors h1 , . . . , hN can be represented explicitly as follows:
h1 =
h2 =
hN =
hN −1 =
hi =
1
S
1
S
1
S
1
S
1
S
(k11 f1 + k12 f2 + k2 f3 ) ,
(k21 f1 + k22 f2 + k1 f3 + k2 f4 ) ,
(k11 fN + k21 fN −1 + k2 fN −2 ) ,
(k12 fN + k22 fN −1 + k1 fN −2 + k4 fN −3 ) ,
(k2 fi−2 + k1 fi−1 + k0 fi + k1 fi+1 + k2 fi+2 )
for i = 3, . . . , N − 2.
For the approximation of hi , we have to study the behavior of the coefficients
which depend on the autocovariance function rk of the AR(2) process (6.10).
In the following subsections we will consider the different types of autocovariance functions separately and prove Theorem 6.1 by deriving approximations
for the vectors hi .
SUPPLEMENT TO “BLUE IN CONTINUOUS TIME REGRESSION MODELS” 35
Autocovariances of the form (6.14). From Corollary 6.1 we obtain that
Shi = −a2 fi−2 + (a1 a2 − a1 )fi−1 + (1 + a21 + a22 )fi + (a1 a2 − a1 )fi+1 − a2 fi+2
= a2 (2fi − fi−2 − fi+2 ) − (a1 a2 − a1 )(2fi − fi−1 − fi+1 )
+(1 + a21 + a22 − 2a2 + 2a1 a2 − 2a1 )fi
= a2 (2fi − fi−2 − fi+2 ) − (a1 a2 − a1 )(2fi − fi−1 − fi+1 )
+(a1 + a2 − 1)2 fi
for i = 3, 4, . . . , N −2. Now consider the case when the autocovariance structure of the errors has the form (6.14) for fixed N . Suppose that the parameters of the autocovariance function (6.14) satisfy p1 6= p2 , 0 < p1 , p2 < 1.
We do not discuss the case with negative p1 or negative p2 because discrete
AR(2) processes with such parameters do not have continuous real-valued
analogues. From the Yule-Walker equations we obtain that the coefficients
a1 and a2 in (6.10) are given by
a1 = r1
(6.18)
(1)
1 − r2
,
1 − r12
(1)
a2 =
r2 − r12
,
1 − r12
where r1 = r1 and r2 = r2 are defined by (6.14). With the notation
λ1 = − log(p1 )/∆ and λ2 = − log(p2 )/∆ with ∆ = (B − A)/N we obtain
(6.19)
p1 = e−λ1 ∆ ,
p2 = e−λ2 ∆ .
We will assume that λ1 and λ2 are fixed but ∆ is small and consider the
properties of different quantities as ∆ → 0. By a straightforward Taylor
expansion we obtain the approximations
a1 = a1 (∆) = 2 − (λ1 + λ2 )∆ + (λ21 + λ22 )∆2 /2 + O(∆3 ),
a2 = a2 (∆) = −1 + (λ1 + λ2 )∆ − (λ1 + λ2 )2 ∆2 /2 + O(∆3 ),
S = S(∆) = 2λ1 λ2 (λ1 + λ2 )∆3 + O(∆4 ),
1
λ1 + λ2 2
λ2
(6.20) C = C(∆) =
+ λ1 λ2
∆ + O(∆4 ).
λ2 − λ1 6
λ1 − λ2
Consequently (observing (6.19) and (6.20)), for large N the continuous
AR(2) process with autocovariances (6.11) can be considered as an approximation to the discrete AR(2) process with autocovariances (6.14).
Since S = O(∆3 ), a1 = 2 + O(∆) and a2 = −1 + O(∆), it follows
S
1
1
1
hi
= f (4) (ti ) − 4a2 2 f (2) (ti ) + (a1 a2 − a1 ) 2 f (2) (ti ) + 4 (a1 + a2 − 1)2 fi + O(∆)
4
∆
∆
∆
∆
1
1
(4)
(2)
= f (ti ) + 2 (a1 a2 − a1 − 4a2 )f (ti ) + 4 (a1 + a2 − 1)2 fi + O(∆)
∆
∆
(4)
2
2 (2)
2 2
= f (ti ) − (λ1 + λ2 )f (ti ) + λ1 λ2 fi + O(∆).
36
H. DETTE ET AL.
Thus, the vectors hi , i = 3, . . . , N − 2, are approximated by the vectorfunction
1
(λ21 + λ22 )f (2) (t) − λ21 λ22 f (t) − f (4) (t) ,
z(t) = −
s3
where s3 = 2λ1 λ2 (λ1 + λ2 ). For the boundary points we obtain
Sh1 = f1 − a1 f2 − a2 f3
= (−2f2 + f3 + f1 ) + (λ1 + λ2 )(f2 − f3 )∆
+((−1/2f2 + 1/2f3 )λ21 + f3 λ1 λ2 + (−1/2f2 + 1/2f3 )λ22 )∆2
and
+((1/6f2 − 1/6f3 )λ31 − 1/2f3 λ21 λ2 − 1/2f3 λ1 λ22 + (1/6f2 − 1/6f3 )λ32 )∆3 + O(∆4 )
= f (2) (t2 ) − (λ1 + λ2 )f (1) (t2 ) + f3 λ1 λ2 ∆2 + O(∆3 )
Sh2 = −a1 f1 + (1 + a21 )f2 + (a1 a2 − a1 )f3 − a2 f4
= (−2f1 + f4 + 5f2 − 4f3 ) + (λ1 + λ2 )(f1 − 4f2 + 4f3 − f4 )∆
+((−1/2f1 + 1/2f4 − 3f3 + 3f2 )λ21 + (2f2 − 4f3 + f4 )λ2 λ1
+ (−1/2f1 + 1/2f4 − 3f3 + 3f2 )λ22 )∆2
+((1/6f1 − 5/3f2 + 5/3f3 − 1/6f4 )λ31 + (−f2 + 3f3 − 1/2f4 )λ2 λ21
=
=
+ (−f2 + 3f3 − 1/2f4 )λ22 λ1 + (1/6f1 − 5/3f2 + 5/3f3 − 1/6f4 )λ32 )∆3 + O(∆4 )
f (2) (t3 ) − 2f (2) (t2 ) + (λ1 + λ2 )(3f (1) (t2 ) − f (1) (t1 ) − f (1) (t3 )) − f3 λ1 λ2 ∆2 + O(∆3 )
− f (2) (t2 ) + (λ1 + λ2 )f (1) (t2 ) − f3 λ1 λ2 ∆2 + O(∆3 )
Thus, we can see that
h1 = −h2 + O(1) = z1,A
1
+ O(1),
∆
where
z1,A =
1
− f (2) (A) + (λ1 + λ2 )f (1) (A) − λ1 λ2 f (A) .
s3
This means that the vectors h1 and h2 at t1 and t2 are large in absolute
value and have different signs. Similarly, we have
hN = −hN −1 + O(1) = z1,B
1
+ O(1)
∆
where
z1,B =
1 (2)
f (B) + (λ1 + λ2 )f (1) (B) + λ1 λ2 f (B) .
s3
SUPPLEMENT TO “BLUE IN CONTINUOUS TIME REGRESSION MODELS” 37
To do a finer approximation, it is necessary to investigate the quantity
g := Sh1 + Sh2 ,
which is of order O(1). Indeed, we have
g = (3f2 − 3f3 − f1 + f4 ) + (λ2 + λ1 )(f1 − 3f2 + 3f3 − f4 )∆
+((−f1 + f4 − 5f3 + 5f2 )/2(λ21 + λ22 ) + (2f2 − 3f3 + f4 )λ2 λ1 )∆2
+((f1 − 9f2 + 9f3 − f4 )/6(λ31 + λ32 )
+(−2f2 + 5f3 − f4 )/2(λ21 λ2 + λ1 λ22 ))∆3 + O(∆4 )
= f (3) (t1 )∆3 + O(∆4 ) + (−f (1) (t1 )(λ21 + λ22 ) − f (1) (t1 )λ2 λ1 )∆3
+f (t1 )(λ21 λ2 + λ1 λ22 )∆3 + O(∆4 )
=
f (3) (t1 ) − (λ21 + λ1 λ2 + λ22 )f (1) (t1 ) + λ1 λ2 (λ1 + λ2 )f (t1 ) ∆3 + O(∆4 )
and, consequently,
h1 + h2 =
1 (3)
f (t1 ) − (λ21 + λ1 λ2 + λ22 )f (1) (t1 ) + λ1 λ2 (λ1 + λ2 )f (t1 ) + O(∆),
s3
where s3 = 2λ1 λ2 (λ1 +λ2 ). Therefore, if ∆ → 0, it follows that h1 +h2 ≈ zA ,
where
1 (3)
zA =
f (A) − (λ21 + λ1 λ2 + λ22 )f (1) (A) + λ1 λ2 (λ1 + λ2 )f (A) .
s3
Similarly, we obtain hN + hN −1 ≈ zB if ∆ → 0, where
zB =
1
− f (3) (B) + (λ21 + λ1 λ2 + λ22 )f (1) (B) + λ1 λ2 (λ1 + λ2 )f (B) .
s3
Autocovariances of the form (6.15). Consider the autocovariance function of the form (6.15), then the coefficients a1 and a2 are given by (6.18)
(2)
(2)
where r1 = r1 and r2 = r2 are defined by (6.15). With the notations
λ = − log p/∆ and q = b/∆ (or equivalently p = e−λ∆ and b = q∆) we
obtain by a Taylor expansion
a1 = 2 − 2λ∆ + (λ2 − q 2 )∆2 + O(∆3 ),
a2 = −1 + 2λ∆ − 2λ2 ∆2 + O(∆3 ),
S = 4λ(λ2 + q 2 )∆3 + O(∆4 )
and
C=
λ λ(λ2 + q 2 ) 2
−
∆ + O(∆4 )
q
3q
38
H. DETTE ET AL.
as ∆ → 0. Similarly, we have
S
1
1
hi
= f (4) (ti ) + 2 (a1 a2 − a1 − 4a2 )f (2) (ti ) + 4 (a1 + a2 − 1)2 fi + O(∆)
∆4
∆
∆
= −2(λ2 − q 2 )f (2) (ti ) + (λ2 + q 2 )2 fi + O(∆).
Thus, the optimal weights hi , i = 3, . . . , N − 2, are approximated by the
signed density
z(t) = −
1
2(λ2 − q 2 )f (2) (t) − (λ2 + q 2 )2 f (t) − f (4) (t) ,
s3
where s3 = 4λ(λ2 + q 2 ). Similarly, we obtain that
1
+ O(1),
∆
1
= −hN −1 + O(1) = z1,B + O(1),
∆
h1 = −h2 + O(1) = z1,A
hN
where
1
− f (2) (A) + 2λf (1) (A) − (λ2 + q 2 )f (A) ,
s3
1 (2)
f (B) + 2λf (1) (B) + (λ2 + q 2 )f (B) .
=
s3
z1,A =
z1,B
Calculating g := Sh1 + Sh2 we have
g = (3f2 − f1 − 3f3 + f4 ) + 2λ(f1 − 3f2 + 3f3 − f4 )∆
+((−f1 + 7f2 − 8f3 + 2f4 )λ2 + q 2 (f1 − 3f2 + 2f3 ))∆2
+((−f1 + 7f2 − 4f3 )λq 2 + (f1 − 15f2 + 24f3 − 4f4 )/3λ3 )∆3 + O(∆4 )
= f (3) (t1 )∆3 − (3λ2 − q 2 )f (1) (t1 )∆3 + 2λ(λ2 + q 2 )f (t1 )∆3 + O(∆4 ).
Therefore, it follows that h1 + h2 ≈ PA if ∆ → 0, where
zA =
1 (3)
f (A) − (3λ2 − q 2 )f (1) (A) + 2λ(λ2 + q 2 )f (A) ,
s3
and s3 = 4λ(λ2 + q 2 ). Similarly, we obtain the approximation hN + hN −1 ≈
PB if ∆ → 0, where
zB =
1
− f (3) (B) + (3λ2 − q 2 )f (1) (B) + 2λ(λ2 + q 2 )f (B) .
s3
Autocovariances of the form (6.16). For the autocovariance function
(6.16) the coefficients a1 and a2 in the AR(2) process are given by (6.18)
SUPPLEMENT TO “BLUE IN CONTINUOUS TIME REGRESSION MODELS” 39
(3)
(3)
where r1 = r1 and r2 = r2 are defined by (6.16). With the notation
λ = − log p/∆ (or equivalently p = e−λ∆ ) we obtain the Taylor expansions
a1 = 2 − 2λ∆ + λ2 ∆2 + O(∆3 ),
a2 = −1 + 2λ∆ − 2λ2 ∆2 + O(∆3 ),
S = 4λ3 ∆3 + O(∆4 ),
λ3
C = λ∆ − ∆3 + O(∆5 )
3
as ∆ → 0. Similar calculations as given in the previous paragraphs give
S
hi
1
1
= f (4) (ti ) + 2 (a1 a2 − a1 − 4a2 )f (2) (ti ) + 4 (a1 + a2 − 1)2 fi + O(∆)
4
∆
∆
∆
= f (4) (ti ) − 2λ2 f (2) (ti ) + λ4 fi + O(∆).
Thus, the vectors hi , i = 3, . . . , N − 2, are approximated by the signed
density
z(t) = −
1
2λ2 f (2) (t) − λ4 f (t) − f (4) (t) ,
s3
where s3 = 4λ3 . For the remaining vectors h1 , h2 , hN −1 and hN we obtain
1
+ O(1),
∆
1
= −hN −1 + O(1) = z1,B + O(1),
∆
h1 = −h2 + O(1) = z1,A
hN
with
1
− f (2) (A) + 2λf (1) (A) − λ2 f (A) ,
s3
1
f (2) (B) + 2λf (1) (B) + λ2 f (B) .
=
s3 f
z1,A =
z1,B
Calculating g := Sh1 + Sh2 we have
g = (3f2 − 3f3 − f1 + f4 ) + 2λ(f1 − 3f2 + 3f3 − f4 )∆
−λ2 (f1 − 7f2 + 8f3 − 2f4 )∆2
+1/3λ3 (f1 − 15f2 + 24f3 − 4f4 )∆3 + O(∆4 )
= f (3) (t1 )∆3 − 3λ2 f (1) (t1 )∆3 + 2λ3 f (t1 )∆3 + O(∆4 ).
Therefore, if ∆ → 0, it follows that h1 + h2 ≈ zA , where
zA =
1 (3)
f (A) − 3λ2 f (1) (A) + 2λ3 f (A) ,
s3
40
H. DETTE ET AL.
and s3 = 4λ3 . Similarly, we obtain the approximation hN + hN −1 ≈ zB if
∆ → 0, where
zB =
1
− f (3) (B) + 3λ2 f (1) (B) + 2λ3 f (B) .
s3
Ruhr-Universität Bochum
Fakultät für Mathematik
44780 Bochum
Germany
E-mail: [email protected]
School of Mathematics
Cardiff University
Cardiff, CF24 4AG
UK
E-mail: [email protected]
E-mail: [email protected]
| 10math.ST
|
arXiv:1803.08037v2 [cs.DS] 24 Mar 2018
Similar Elements and Metric Labeling on Complete Graphs
Pedro F. Felzenszwalb
Brown University
Providence, RI, USA
[email protected]
March 28, 2018
We consider a problem that involves finding similar elements in a collection of sets. The problem
is motivated by applications in machine learning and pattern recognition (see, e.g. [3]). Intuitively
we would like to discover something in common among a collection of sets, even when the sets have
empty intersection. A solution involves selecting an element from each set such that the selected
elements are close to each other under an appropriate metric. We formulate an optimization problem
that captures this notion and give an efficient approximation algorithm that finds a solution within
a factor of 2 of the optimal solution.
The similar elements problem is a special case of the metric labeling problem defined in [2] and
we also give an efficient 2-approximation algorithm for the metric labeling problem on complete
graphs. Metric labeling on complete graphs generalizes the similar elements problem to include
costs for selecting elements in each set.
The algorithms described here are similar to the “center star” method for multiple sequence
alignment described in [1].
Beyond producing solutions with good theoretical guarantees, the algorithms described here are
also practical. A version of the algorithm for the similar elements problem has been implemented
and used to find objects in a collection of photographs [4].
1
Similar Elements
Let X be a (possibly infinite) set and d be a metric on X. Let S1 , . . . , Sn be n finite subsets of
X. The goal of the similar elements problem is to select an element from each set Si such that
the selected elements are close to each other under the metric d. One motivation is for discovering
something in common among the sets S1 , . . . , Sn even when they have empty intersection.
We formalize the problem as the minimization of the sum of pairwise distances among selected
elements. Let x = (x1 , . . . , xn ) with xi ∈ Si . Define the similar elements objective as,
X
c(x) =
d(xi , xj ).
(1)
1≤i,j≤n
Let x∗ = argminx c(x) be an optimal solution for the similar elements problem.
Optimizing c(x) appears to be difficult, but we can define easier problems if we ignore some of
the pairwise distances in the objective. In particular we define n different “star-graph” objective
1
functions as follows. For each 1 ≤ r ≤ n define the objective cr (x) to account only for the terms in
c(x) involving xr ,
X
cr (x) =
d(xr , xj ).
(2)
j6=r
xr
cr (x)
Let
= argminx
be an optimal solution for the optimization problem defined by cr (x).
We can compute xr efficiently using a simple form of dynamic programming, by first computing xrr
and then computing xrj for j 6= r.
xrr = argmin
X
min d(xr , xj ),
(3)
xr ∈Sr j6=r xj ∈Sj
xrj = argmin d(xrr , xj ).
(4)
xj ∈Sj
Each of the n “star-graph” objective functions leads to a possible solution. We then select from
among the solutions x1 , . . . , xn as follows,
r̂ = argmin cr (xr ),
(5)
1≤r≤n
r
x̂ = x .
(6)
Theorem 1. The algorithm described above finds a 2-approximate solution for the similar elements
problem. That is,
c(x̂) ≤ 2c(x∗ ).
Proof. First note that,
c(x) =
n
X
cr (x).
r=1
Since the minimum of a set of values is at most the average, and xr minimizes cr (x),
min cr (xr ) ≤
1≤r≤n
n
n
r=1
r=1
1X r ∗
1
1X r r
c (x ) ≤
c (x ) = c(x∗ ).
n
n
n
By the triangle inequality we have
c(x) =
X
1≤i,j≤n
d(xi , xj ) ≤
X
(d(xi , xr ) + d(xr , xj )) = 2n
1≤i,j≤n
n
X
d(xr , xl ) = 2ncr (x).
l=1
Therefore
c(x̂) ≤ 2ncr̂ (x̂) = 2n min cr (xr ) ≤ 2c(x∗ ).
1≤r≤n
To analyze the running time of the algorithm we assume the distances d(p, q) between pairs of
elements in S = S1 ∪ · · · ∪ Sn are either pre-computed and given as part of the input, or they can
each be computed in O(1) time.
Let k = max1≤i≤n |Si |. The first stage of the algorithm involves n optimization problems that
can be solved in O(nk2 ) time each. The second stage of the algorithm involves selecting one of the
n solutions, and takes O(n2 ) time.
2
Remark 2. If each of the sets S1 , . . . , Sn has size at most k the running time of the approximation
algorithm for the similar elements problem is O(n2 k2 ).
The bottleneck of the algorithm is the evaluation of the minimizations over xj ∈ Sj in (3) and
(4). This computation is equivalent to a nearest-neighbor computation, where we want to find
a point from a set S ⊆ X that is closest to a query point q ∈ X. When the nearest-neighbor
computation can be done efficiently (with an appropriate data structure) the running time of the
similar elements approximation algorithm can be reduced.
2
Metric Labeling on Complete Graphs
Let G = (V, E) be an undirected simple graph on n nodes V = {1, . . . , n}. Let L be a finite set
of labels with |L| = k and d be a metric on L. For i ∈ V let mi be a non-negative function
mapping labels to real values. The unweighted metric labeling problem on G is to find a labeling
x = (x1 , . . . , xn ) ∈ Ln minimizing
X
X
c(x) =
mi (xi ) +
d(xi , xj ).
(7)
i∈V
{i,j}∈E
Let x∗ = argminx c(x). This optimization problem can be solved in polynomial time using
dynamic programming if G is a tree. Here we consider the case when G is the complete graph
and give an efficient 2-approximation algorithm based on the solution of several metric labeling
problems on star graphs.
For each r ∈ V define a different objective function, cr (x), corresponding to a metric labeling
problem on a star graph with vertex set V rooted at r,
cr (x) =
X mi (xi )
n
i∈V
+
X
j∈V \{r}
d(xr , xj )
.
2
(8)
Let xr = argminx cr (x). We can solve this optimization problem in O(nk 2 ) time using a simple
form of dynamic programming. First compute an optimal label for the root vertex using one step
of dynamic programming,
X
mj (xj ) d(xr , xj )
mr (xr )
+
min
+
.
(9)
xrr = argmin
xj ∈L
n
n
2
xr ∈L
j∈V \{r}
Then compute xrj for j ∈ V \ {r},
xrj
= argmin
xj ∈L
mj (xj ) d(xrr , xj )
+
n
2
.
(10)
Optimizing each cr (x) separately leads to n possible solutions x1 , . . . , xn , and we select one of
them as follows,
r̂ = argmin cr (xr ),
r∈V
r
x̂ = x .
3
(11)
(12)
Theorem 3. The algorithm described above finds a 2-approximate solution for the metric labeling
problem on a complete graph. That is,
c(x̂) ≤ 2c(x∗ ).
Proof. First note that,
c(x) =
n
X
cr (x).
r=1
Since the minimum of a set of values is at most the average, and xr minimizes cr (x),
n
n
1X r ∗
1
1X r r
c (x ) ≤
c (x ) = c(x∗ ).
min c (x ) ≤
1≤r≤n
n r=1
n r=1
n
r
r
Since d is a metric and mi is non-negative,
X
X
c(x) =
mi (xi ) +
d(xi , xj )
i∈V
{i,j}∈E
d(xi , xj )
2
i∈V
(i,j)∈V 2
X
X
d(xi , xr ) d(xr , xj )
≤
mi (xi ) +
+
2
2
2
=
X
=
X
X
mi (xi ) +
i∈V
(i,j)∈V
mi (xi ) + 2n
i∈V
≤ 2n
l∈V \{r}
X mi (xi )
i∈V
r
X
n
+ 2n
d(xr , xl )
2
X
l∈V \{r}
d(xr , xl )
2
= 2nc (x).
Therefore
c(x̂) ≤ 2ncr̂ (x̂) = 2n min cr (xr ) ≤ 2c(x∗ ).
1≤r≤n
The first stage of the algorithm involves n optimization problems that can be solved in O(nk2 )
time each. The second stage involves selecting one of the n solutions, and takes O(n2 ) time.
Remark 4. The running time of the approximation algorithm for the metric labeling problem on
complete graphs is O(n2 k2 ).
Acknowledgments
We thank Caroline Klivans, Sarah Sachs, Anna Grim, Robert Kleinberg and Yang Yuan for helpful
discussions about the contents of this report. This material is based upon work supported by the
National Science Foundation under Grant No. 1447413.
4
References
[1] Dan Gusfield. Efficient methods for multiple sequence alignment with guaranteed error bounds.
Bulletin of Mathematical Biology, 55(1):141–154, 1993.
[2] Jon Kleinberg and Eva Tardos. Approximation algorithms for classification problems with pairwise relationships: Metric labeling and markov random fields. Journal of the ACM, 49(5):616–
639, 2002.
[3] Oded Maron and Aparna Lakshmi Ratan. Multiple-instance learning for natural scene classification. In International Conference on Machine Learning, volume 98, pages 341–349, 1998.
[4] Sarah Sachs. Similar-part approximation using invariant feature descriptors. Undergraduate
Honors Thesis, Brown University, 2016.
5
| 8cs.DS
|
Resource Allocation for Solar Powered UAV
Communication Systems
(Invited Paper)
arXiv:1801.07188v1 [cs.IT] 22 Jan 2018
Yan Sun, Derrick Wing Kwan Ng, Dongfang Xu, Linglong Dai, and Robert Schober
Abstract—In this paper, we investigate the resource allocation
design for multicarrier (MC) systems employing a solar powered
unmanned aerial vehicle (UAV) for providing communication
services to multiple downlink users. We study the joint design of the
three-dimensional positioning of the UAV and the power and subcarrier allocation for maximization of the system sum throughput.
The algorithm design is formulated as a mixed-integer non-convex
optimization problem, which requires a prohibitive computational
complexity for obtaining the globally optimal solution. Therefore,
a low-complexity suboptimal iterative solution based on successive
convex approximation is proposed. Simulation results confirm that
the proposed suboptimal algorithm achieves a substantially higher
system sum throughput compared to several baseline schemes.
I. I NTRODUCTION
Future wireless communication systems are envisioned to
provide ubiquitous and sustainable high data-rate communication services [1], [2]. However, in some cases, deploying
conventional terrestrial infrastructure (e.g. base stations (BSs)) is
not cost-effective or not feasible. For example, it is not possible
to deploy fixed BSs in a timely and economical manner in temporary hotspots, disaster areas, and complex terrains. To handle
this issue, aerial communication systems based on unmanned
aerial vehicles (UAVs) have been proposed as a promising
new paradigm to facilitate fast and flexible deployment due
to their excellent maneuverability [3]–[7]. In particular, UAVs
equipped with on-board wireless transmitters can fly over the
target area and provide communication services. Moreover,
since UAVs enjoy high mobility, they can adjust their aerial
position according to the real-time locations of the users which
introduces additional degrees of freedom for improving system
performance. In [3], the authors investigated UAV trajectory
design for minimization of the mission completion time in
multicasting systems. The authors of [4] proposed a suboptimal
joint trajectory, power allocation, and user scheduling algorithm
for maximization of the minimum user throughput in multiUAV systems. In [5], a suboptimal joint trajectory and power
allocation algorithm was proposed for maximization of the
system secrecy rate in a UAV communication system. The threedimensional (3-D) positioning of UAVs for maximization of the
number of served users and the coverage area was studied in [6]
and [7], respectively. However, the UAV-based communication
systems considered in [3]–[7] are powered by on-board batteries,
leading to limited operation time. Specifically, the UAVs in [3]–
[7] are required to return to the ground frequently for recharging
their batteries. Hence, these designs cannot guarantee stable
and sustainable communication services which creates a system
performance bottleneck.
To overcome this shortcoming, solar powered UAVs have
received significant attention due to their potential to realize
perpetual flight [8], [9]. In particular, solar panels installed on
the UAVs harvest the received solar energy and convert it to
electrical energy for long endurance flights. The authors of [8]
Y. Sun, D. Xu, and R. Schober are with Friedrich-Alexander-University
Erlangen-Nürnberg (FAU), Germany. D. W. K. Ng is with University of New
South Wales, Australia. L. Dai is with Tsinghua University, Beijing, China.
and [9] have developed solar powered UAV prototypes and
demonstrated the possibility of continuous flight for 28 hours.
However, the amount of harvested solar energy is affected by
the flight altitude of the UAV. For example, the atmospheric
transmittance decreases for lower altitudes leading to a smaller
amount of harvested solar energy [10]. Besides, the intensity
of solar energy can significantly decrease if the light passes
through clouds, resulting in reduced solar energy flux at the
solar panels [10], [11]. Therefore, UAVs flying at higher altitude
can generally harvest more solar energy than those flying at
lower altitude. In [12], the authors studied the optimal trajectory of solar-powered UAVs for maximization of the harvest
solar power taking into account the atmospheric transmittance.
However, the influence of clouds on solar powered UAVs was
not considered in [12]. Moreover, [12] focused only on flight
control of solar powered UAVs, whereas communication design
was not considered. In fact, since higher flight altitudes lead
to a more severe path loss for air-to-ground communications,
there is a tradeoff between harvesting more solar energy and
improving communication performance. This tradeoff does not
exist in conventional UAV communication systems [3]–[7] and
the results derived in [3]–[7] cannot be applied for solar powered UAV communication systems. Moreover, multicarrier (MC)
techniques are expected to play an important role also in future
multiuser communication systems [13]–[20]. However, resource
allocation designs for BS-based MC communication systems
[13]–[20] cannot be directly applied to solar powered MC
UAV communication systems, where the power and subcarrier
allocation is coupled with the aerial positioning of the UAVs. In
fact, the joint design of the 3-D positioning and the power and
subcarrier allocation for solar powered MC UAV communication
systems is an open research problem.
In this paper, we address the above issues. To this end, the
resource allocation algorithm design for solar powered MC UAV
communication systems is formulated as a combinatorial nonconvex optimization problem for maximization of the system
sum throughput. The considered optimization problem is in
general intractable and obtaining the globally optimal solution
may result in prohibitive computational complexity. Therefore,
we develop an efficient suboptimal resource allocation algorithm
based on successive convex approximation to strike a balance
between computational complexity and optimality.
II. S YSTEM M ODEL
In this section, we first present the considered MC UAV
communication system model. Then, we discuss the solar energy
harvesting model adopted for resource allocation design.
A. Notation
We use boldface lower case letters to denote vectors. C
denotes the set of complex numbers; RN ×1 denotes the set of
all N × 1 vectors with real entries; R+ denotes the set of nonnegative real numbers; |·| and k·k denote the absolute value of
a complex scalar and the Euclidean vector norm, respectively;
LOWTRAN 7 [23]. The atmospheric transmittance at altitude z
can be empirically approximated as follows [12], [24]:
Lup
Cloud
φ(z) = α − βe−z/δ ,
z
Llow
User 1
User 2
Solar light
Wireless signal
Fig. 1. A solar powered MC UAV communication system with one UAV
transmitter and K = 2 downlink users.
E{·} denotes statistical expectation; the circularly symmetric
complex Gaussian distribution with mean µ and variance σ 2
is denoted by CN (µ, σ 2 ); and ∼ stands for “distributed as”;
∇x f (x) denotes the gradient vector of function f (x) whose
components are the partial derivatives of f (x).
B. MC UAV Communication System Model
The considered MC UAV wireless communication system
comprises one rotary-wing UAV-mounted transmitter [21] and K
downlink (DL) users. The UAV-mounted transmitter and the DL
users are single-antenna half-duplex devices, cf. Figure 1. The
UAV is equipped with solar panels which harvest solar energy
and convert it to electrical energy. The harvested energy is used
for providing communication services and powering the flight
operation of the UAV. The system bandwidth is divided into NF
orthogonal subcarriers. We assume that each subcarrier can be
allocated to at most one user.
In the considered system, the path loss of the communication
link between the UAV and user k ∈ {1, . . . , K} is modeled as
ρk = ̺kr − rk k−2 , where r = (x, y, z) and rk = (xk , yk , 0)
represent the 3-D Cartesian coordinates of the UAV and user k,
respectively. In particular, (x, y) and (xk , yk ) are the horizontal
coordinates of the UAV and user k, respectively, while z denotes
c
)2 , where c is the
the altitude of the UAV. Besides, ̺ = ( 4πf
0
speed of light and f0 is the center frequency of the carrier signal.
Therefore, in each scheduling time slot, the received signal
at downlink user k on subcarrier
i ∈ {1, . . . , NF } is given by
p
i hi
̺p
k k i
d + nik ,
(1)
uik =
kr − rk k k
where dik ∈ C denotes the data symbol transmitted from the
UAV to user k on subcarrier i and we assume E{|dik |2 } = 1
without loss of generality. pik ∈ R+ is the transmit power
for the signal transmitted to user k on subcarrier i. hik ∈ C
denotes the shadowing and small scale fading coefficient for
the link between the UAV and user k on subcarrier i [22].
nik ∼ CN (0, σk2 ) denotes the complex additive white Gaussian
noise (AWGN) on subcarrier i at user k. Besides, we assume
that the channel state hik , ∀k, i, is perfectly known at the UAV to
unveil a performance upper bound for MC UAV communication
systems.
C. Solar Energy Harvesting
The considered MC UAV communication system is powered
by the harvested solar energy. In general, the amount of harvested solar energy is affected by the atmospheric transmittance
and clouds in the air [10], [11]. In particular, the atmospheric
transmittance increases with the altitude, as higher altitudes lead
to higher solar intensity such that more solar energy can be
collected by the solar panels. The atmospheric transmittance for
a given altitude is deterministic and has a value between zero
and one which can be calculated by using the software tool
(2)
where α is the maximum value of the atmospheric transmittance,
β is the extinction coefficient of the atmosphere, and δ is the
scale height of the earth. Besides, the solar intensity is reduced
if there is a cloud between the sun and the solar panel. The
attenuation of solar light passing through a cloud can be modeled
as [11]:
cloud
ϕ(dcloud ) = e−βc d
,
(3)
where βc ≥ 0 denotes the absorption coefficient modeling
the optical characteristics of the cloud and dcloud denotes the
distance that the solar light propagates through the cloud.
Therefore, the electrical output power of the solar panels at
altitude z is
modeled as [10]–[12]:
z ≥ Lup ,
ηSGφ(z)ϕ(0),
P solar (z) = ηSGφ(z)ϕ(Lup − z),
Llow ≤ z < Lup , (4)
ηSGφ(z)ϕ(Lup − Llow ), z < Llow ,
where η and S denote the energy harvesting efficiency and the
area of the solar panels, respectively. Constant G denotes the
average solar radiation on earth. Lup and Llow are the altitudes
of the upper and lower boundaries of the cloud, respectively.
III. P ROBLEM F ORMULATION
AND
S OLUTION
In this section, after defining the adopted performance measure, we formulate the resource allocation problem. Then, we
propose an iterative algorithm to solve the proposed problem.
A. Achievable Data Rate
Assuming subcarrier i is allocated to user k, the achievable
data rate on subcarrier i is given by:
Hki pik
,
(5)
Uki (pik , sik , r) = sik log2 1 +
kr − rk k2
where Hki = ̺|hik |2 /σk2 . Variable sik ∈ {0, 1} is the binary
subcarrier allocation indicator. Specifically, sik = 1 if user k is
allocated to subcarrier i, and sik = 0, otherwise.
B. Optimization Problem Formulation
In this paper, we maximize the system sum throughput via
optimizing the 3-D position and the power and subcarrier allocation of the UAV in a given time slot. The problem formulation
is given as follows: X
NF X
K
Hki pik
i
(6)
maximize
s
log
1
+
k
2
kr − rk k2
sik ,pik ,r
i=1 k=1
NF X
K
X
s.t. C1:
sik pik +PUAV ≤ P solar (z), C2: pik ≥ 0, ∀i, k,
i=1 k=1
C3:
C5:
NF X
K
X
sik pik ≤ Pmax ,
i=1 k=1
sik ∈{0, 1}, ∀i, k,
C4: zmin ≤ z ≤ zmax ,
C6:
K
X
sik ≤1, ∀i.
k=1
Constraint C1 is the power constraint of the UAV where constant PUAV represents the power required for maintaining the
operation of the UAV. Constraint C2 is the non-negative transmit
power constraint. Pmax in constraint C3 denotes the maximum
transmit power of the UAV-mounted transmitter as imposed by
restrictions on the transmit spectrum mask to limit the amount
of out-of-cell interference in the DL. zmin and zmax in constraint
C4 denote the minimum and the maximum flight altitude of the
UAV imposed by regulation. Constraints C5 and C6 are imposed
to guarantee that each subcarrier is allocated to at most one user.
For facilitating the presentation, we rewrite the power available
from solar energy
in (4) as:
harvesting
z ≥ Lup ,
A − Be−z/δ ,
P solar (z) = M (z) − BC1 e(βc −1/δ)z , Llow ≤ z < Lup , (7)
AC2 − BC2 e−z/δ ,
z < Llow ,
where A = ηSGα, B = ηSGβ, C1 = e−βc Lup , C2 =
e−βc (Lup −Llow ) , and M (z) = AC1 eβc z . For the considered
communication system, we note that there is a fundamental
tradeoff between harvesting solar energy and improving communication performance. In particular, the UAV can harvest
more solar energy by climbing up to higher altitudes. However,
flying at higher altitude leads to a larger path loss for the
communication links between the UAV and the users which
results in a degradation of the system performance.
The problem in (6) is a mixed-integer non-convex problem
due to the non-convex constraint C1, the binary constraint C5,
and the non-convex objective function. In general, mixed-integer
non-convex optimization problems cannot be solved optimally in
a computationally efficient manner. Thus, in the next section, we
propose a successive convex approximation based suboptimal
scheme for the considered problem.
C. Joint 3-D Position, Power, and Subcarrier Optimization
In problem (6), the binary constraint C5 and constraint C6
are imposed to allocate at most one user to each subcarrier
which is an obstacle for the design of a computationally efficient
resource allocation algorithm. In this section, we transform
problem (6) into an equivalent form while relaxing constraints
C5 and C6. First, we temporarily assume that each subcarrier
can be allocated to multiple users and introduce the following
new utility function for user k on subcarrier
i:
!
Hki
i
p̃
2 k
kr−r
k
k
, (8)
Ũki (p̃ik , r) = log2 1 + H i P
K
k
i
m6=k p̃m + 1
kr−rk k2
where p̃ik ∈ R+ denotes the transmit power for the signal
transmitted to user k on subcarrier i. In fact, (8) represents the
achievable rate of user k on subcarrier i where
i is
Psubcarrier
Hi
K
allocated to K DL users and the term kr−rkk k2 m6=k p̃im in (8)
represents the multiuser interference at user k from the K − 1
co-channel users. Then, adopting the utility function in (8), we
formulate a modified optimization problem for maximizing the
system sum throughput:
!
Hki
i
NF X
K
X
kr−rk k2 p̃k
maximize
log2 1 + H i P
K
p̃,r
k
i
i=1 k=1
m6=k p̃m + 1
kr−rk k2
NF X
K
X
f
s.t. C1:
p̃ik + PUAV ≤ P solar (z),
i=1 k=1
f p̃ik ≥ 0, ∀i, k,
C2:
NF K×1
f
C3:
NF X
K
X
p̃ik ≤ Pmax ,
C4,
(9)
i=1 k=1
where p̃ ∈ R
is the collection of all p̃ik . We note that conf C3
f and C4 in (9) have the same physical meaning as
straints C1–
the constraints C1–C4 in problem (6), respectively. Constraints
C5 and C6 are not imposed in (9) due to the modified subcarrier
allocation strategy that allows the multiplexing of multiple users
on each subcarrier. We note that the problem formulations in (9)
and (6) are equivalent when in (9) on each subcarrier at most one
of the powers p̃ik is non-zero. Now, we introduce the following
theorem which reveals the equivalence between (9) and (6).
Theorem 1: The optimal subcarrier assignment strategy for
maximizing the system sum throughput in (9) assigns each subcarrier to the user with the best channel gain and no subcarrier
is shared by multiple users.
Algorithm 1 Successive Convex Approximation
1: Initialize the iteration index j = 1 and initial point p̃(1) and θ(1)
2: repeat
3:
For given p̃(j) and θ(j) , solve the convex problem in (14) and
store the resulting solution {p̃ and θ}
4:
Set j = j + 1 and p̃(j) = p̃ and θ(j) = θ
5: until convergence
6: Obtain final resource allocation policy p̃∗ = p̃(j) , θ∗ = θ(j)
Proof: The proof closely follows the proof in [25, Appendix]
and is omitted here due to the space limitation.
Intuitively, if multiple users are activated on a subcarrier, the
PK
Hi
interference term kr−rkk k2 m6=k p̃im will severely degrade the
system sum throughput. Therefore, problems (6) and (9) are
equivalent in the sense that they yield the same optimal solution.
Hence, we focus on the solution of problem (9). We note that
the fractional term p̃ik /kr − rk k2 in the objective function of
(9) is an obstacle to solving (9) efficiently. We overcome this
difficulty by rewriting (9) in the following equivalent form:
NF X
K
X
Hki p̃ik
maximize
log2 1 + PK
i i
p̃,r,θ
m6=k Hk p̃m + θk
i=1 k=1
f C3,
f C4, C7: kr − rk k2 ≤ θk ,
s.t. C1–
(10)
where θk is an auxiliary variable and θ ∈ RK×1 is the collection
of all θk . Now, the remaining non-convexity of problem (10) is
f and the objective function. We note that
due to constraint C1
(10) can be rewritten as a difference of convex programming
problem [26]:
NF X
K
K
X
X
minimize −
Hki p̃im + θk − G(p̃, θ)
log2
p̃,r,θ
m=1
i=1 k=1
f C3,
f C4, C7,
s.t. C2,
f
C1:
NF X
K
X
p̃ik +PUAV ≤P solar (z), (11)
i=1 k=1
P
K
i i
where G(p̃, θ) = − i=1 k=1 log2
m6=k Hk p̃m + θk .
We note that the problems in (10) and (11) are equivalent. We
can obtain a locally optimal solution of (11) by applying successive convex approximation [26]. In particular, for any feasible
point p̃(j) and θ (j) , we replace G(p̃, θ) and P solar (z) with
their global underestimations G(p̃, θ, p̃(j) , θ (j) ) and P solar (z),
respectively, where
PNF PK
G(p̃, θ, p̃(j) , θ (j) ) = G(p̃(j) , θ (j) ) + ∇p̃ G(p̃, θ)(p̃ − p̃(j) )
+ ∇θ G(p̃, θ)(θ − θ (j) ),
(12)
z≥Lup ,
A − Be−z/δ ,
solar
(z)= M (z,z (j) )−BC1 e(βc −1/δ)z , Llow ≤z<Lup, (13)
P
AC2 − BC2 e−z/δ ,
z<Llow .
(j)
(j)
Here, M (z, z (j) ) = AC1 eβc z + AC1 βc eβc z (z − z (j) ) is the
global underestimation of M (z) = AC1 eβc z in (7). Then, for
any given p̃(j) and θ (j) , we can obtain a lower bound for (11)
by solving the following optimization problem:
NF X
K
K
X
X
minimize −
log2
Hki p̃im + θk − G(p̃, θ, p̃(j) , θ (j) )
p̃,r,θ
i=1 k=1
f C3,
f C4, C7,
s.t. C2,
m=1
K
NF X
X
f
C1:
p̃ik +PUAV ≤P solar (z), (14)
i=1 k=1
Then, we successively tighten the obtained lower bound by
applying the iterative algorithm summarized in Algorithm 1.
The proposed iterative algorithm converges to a locally optimal solution of (11) and has polynomial time computational
complexity [26]. We note that standard convex program solvers
2 GHz and 5 MHz
64 and 78 kHz
0.8978, 0.2804 [10]
1367 W/m2 , 8000 m
700 m and 1400 m [11]
0.01 [11]
100 m and 1500 m
0.4
−110 dBm
200 W
S = 1 m2
14
Average system sum throughput (bits/s/Hz)
17
Proposed scheme
Baseline scheme 1
Baseline scheme 2
16
Average system sum throughput (bit/s/Hz)
TABLE I
S YSTEM PARAMETERS
Carrier center frequency and bandwidth
Number and bandwidth of subcarriers
Parameters for atmospheric transmittance, α, β
Average solar radiation and scale height, G, δ
Altitude of cloud, Llow and Lup
Absorption coefficient of cloud, βc
Altitude limitation for UAV, zmin and zmax
Efficiency of solar panels, η
Receiver noise power, σk2
Power requirement for operating UAV, PUAV
15
14
Average system
sum throughput
improvement
Proposed scheme
13
12
Baseline scheme 1
11
10
Baseline scheme 2
9
8
7
Proposed scheme
6
12
1
Baseline scheme 1
10
Baseline scheme 2
S = 0.7 m2
Proposed scheme, S = 1 m2
Baseline scheme 1, S = 1 m 2
Baseline scheme 2, S = 1 m 2
Proposed scheme, S = 0.7 m2
Baseline scheme 1, S = 0.7 m2
Baseline scheme 2, S = 0.7 m2
4
Average system sum
throughput improvement
25
30
4
5
6
7
8
Fig. 3. Average system sum throughput (bits/s/Hz) versus the number of users
for different resource allocation schemes with Pmax = 40 dBm and S = 1 m2 .
6
0
20
3
Number of users
8
2
2
35
40
45
50
Maximum transmit power (dBm)
Fig. 2. Average system sum throughput (bits/s/Hz) versus the maximum transmit
power of the UAV (dBm), Pmax , for different resource allocation schemes and
K = 3.
such as CVX [27] can be used for efficiently solving the convex
problem in (14).
Note that we can determine the subcarrier allocation policy
from the obtained p̃∗ in line 6 of Algorithm 1. In particular, we
note that p̃ik is larger than zero only if DL user k is allocated to
subcarrier i. Thus, the subcarrier allocation policy is obtained
as: sik = 1 if p̃ik > 0, and sik = 0, otherwise.
IV. S IMULATION R ESULTS
In this section, we evaluate the system performance of the
proposed scheme via simulations. The adopted simulation parameters are given in Table I. We consider a single cell where
the K DL users are randomly and uniformly distributed within
in the cell boundary of 1500 meters and the entire service area
is covered by clouds. In each slot, the fading coefficients of the
channels between the UAV and the DL users on each subcarrier
are independent and identically distributed random variables
following a Rician distribution with Rician factor 3 dB. We
obtained the results shown in this section by averaging over
5000 realizations of multipath fading. Besides, we also consider
the performance of two baseline schemes for comparison. For
baseline scheme 1, we set (x, y)=(0, 0), i.e., the origin of the
cell, and then jointly optimize the flight altitude z, power pik ,
and subcarrier allocation sik of the UAV communication system.
For baseline scheme 2, the user on each subcarrier is selected
randomly and we optimize the 3-D position of the UAV and the
power allocated to the users.
In Figure 2, we investigate the average system sum throughput
versus the maximum transmit power at the UAV, Pmax , for
K = 3 DL users and different solar panel sizes S. The
average system throughputs of the proposed scheme and all
baseline schemes increase monotonically with the maximum
transmit power Pmax . In fact, for the proposed scheme and the
baseline schemes, the UAV can fly to higher altitudes to harvest
more solar energy when the maximum transmit power Pmax
increases. Thus, the proposed scheme and the baseline schemes
can effectively exploit the increased transmit power budget
to improve the received signal-to-interference-plus-noise ratio
(SINR) at the users. Besides, for a smaller value of S, there is a
diminishing return in the average system sum throughput for the
proposed scheme and all baseline schemes when Pmax exceeds
35 dBm and the average system sum throughput approaches
a constant in the large Pmax regime. In fact, since the output
power of the solar panels is smaller for smaller solar panel
sizes, a UAV equipped with a smaller solar panel has to fly
at a higher altitude to harvest the same amount of solar energy
as a UAV equipped with a larger solar panel, which leads to a
severe path loss for air-to-ground communications and causes
a performance degradation. Furthermore, as can be observed,
the proposed scheme achieves a considerably higher average
system sum throughput than baseline schemes 1 and 2 due to
the joint optimization of the 3-D position and the power and
subcarrier allocation. In particular, for baseline scheme 1, the
horizontal coordinates of the UAV (x, y) are fixed leading to
fewer available degrees of freedom for improving the average
system sum throughput. For baseline scheme 2, although the
adopted random subcarrier allocation policy provides fairness,
it causes a poor utilization of the system resources.
In Figure 3, we study the average system sum throughput
versus the number of DL users for Pmax = 40 dBm and S = 1
m2 . As can be observed, the average system sum throughput
for the proposed scheme and baseline scheme 1 increase with
the number of users since these schemes are able to exploit
multiuser diversity. The performance of baseline scheme 2 is
independent of the number of users since it employs a random
subcarrier allocation policy. Moreover, it can be observed from
Figure 3 that the average system sum throughput of the proposed
scheme grows faster with the number of users than that of
baseline scheme 1. In fact, for baseline scheme 1, the UAV
cannot adjust its horizontal coordinates (x, y) according to the
locations of the users which limits its capability to exploit the
multiuser diversity introduced by the different locations of the
users.
V. C ONCLUSION
In this paper, we studied the joint optimization of the 3D position and the power and subcarrier allocation for solar
powered MC UAV communication systems. The objective of the
resulting mixed-integer non-convex optimization problem was
the maximization of the system sum throughput. A suboptimal
resource allocation algorithm design based on successive convex
approximation was proposed. Simulation results unveiled that
the proposed scheme for solar powered UAV systems achieves
a significant improvement in system performance compared to
two baseline schemes.
R EFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong,
and J. C. Zhang, “What Will 5G Be?” IEEE J. Select. Areas Commun.,
vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[2] V. W. S. Wong, R. Schober, D. W. K. Ng, and L.-C. Wang, Key
Technologies for 5G Wireless Systems, 1st ed. Cambridge University
Press, 2017.
[3] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput Maximization for UAVEnabled Mobile Relaying Systems,” IEEE Trans. Commun., vol. 64, no. 12,
pp. 4983–4996, Dec. 2016.
[4] Q. Wu, Y. Zeng, and R. Zhang, “Joint Trajectory and Communication Design for Multi-UAV Enabled Wireless Networks,” arXiv preprint
arXiv:1705.02723, 2017.
[5] G. Zhang, Q. Wu, M. Cui, and R. Zhang, “Securing UAV Communications
via Trajectory Optimization,” in Proc. IEEE Global Commun. Conf., Dec.
2017, pp. 1–5.
[6] R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D Placement of An Aerial Base Station in Next Generation Cellular Networks,”
in Proc. IEEE Intern. Commun. Conf., May 2016, pp. 1–5.
[7] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient Deployment
of Multiple Unmanned Aerial Vehicles for Optimal Wireless Coverage,”
IEEE Commun. Lett., vol. 20, no. 8, pp. 1647–1650, Aug. 2016.
[8] P. Oettershagen, A. Melzer, T. Mantel, K. Rudin, T. Stastny, B. Wawrzacz,
T. Hinzmann, K. Alexis, and R. Siegwart, “Perpetual Flight with A Small
Solar-Powered UAV: Flight Results, Performance Analysis and Model
Validation,” in Proc. IEEE Aerosp. Conf., Mar. 2016, pp. 1–8.
[9] S. Morton, R. D’Sa, and N. Papanikolopoulos, “Solar Powered UAV:
Design and Experiments,” in Proc. IEEE/RSJ Intern. Conf. on Intel. Robot.
and Systems, Sep. 2015, pp. 2460–2466.
[10] J. A. Duffie and W. A. Beckman, Solar Engineering of Thermal Processes,
3rd ed. Hoboken, NJ, USA: John Wiley & Sons, 2013.
[11] A. Kokhanovsky, “Optical Properties of Terrestrial Clouds,” Earth-Science
Reviews, vol. 64, no. 3, pp. 189–241, Feb. 2004.
[12] J. S. Lee and K. H. Yu, “Optimal Path Planning of Solar-Powered UAV
Using Gravitational Potential Energy,” IEEE Trans. Aerosp. and Electron.
Systems, vol. 53, no. 3, pp. 1442–1451, Jun. 2017.
[13] D. W. K. Ng, E. S. Lo, and R. Schober, “Energy-Efficient Resource
Allocation in OFDMA Systems with Hybrid Energy Harvesting Base
Station,” IEEE Trans. Wireless Commun., vol. 12, no. 7, pp. 3412–3427,
Jul. 2013.
[14] Y. Sun, D. W. K. Ng, Z. Ding, and R. Schober, “Optimal Joint Power
and Subcarrier Allocation for Full-Duplex Multicarrier Non-Orthogonal
Multiple Access Systems,” IEEE Trans. Commun., vol. 65, no. 3, pp. 1077–
1091, Mar. 2017.
[15] L. Xiang, D. W. K. Ng, T. Islam, R. Schober, V. W. S. Wong, and J. Wang,
“Cross-Layer Optimization of Fast Video Delivery in Cache- and BufferEnabled Relaying Networks,” IEEE Trans. Veh. Technol., vol. 66, no. 12,
pp. 11 366–11 382, Dec. 2017.
[16] J. Zhu, J. Wang, Y. Huang, S. He, X. You, and L. Yang, “On Optimal Power
Allocation for Downlink Non-Orthogonal Multiple Access Systems,” IEEE
J. Select. Areas Commun., vol. 35, no. 12, pp. 2744–2757, Dec. 2017.
[17] D. W. K. Ng, E. S. Lo, and R. Schober, “Wireless Information and Power
Transfer: Energy Efficiency Optimization in OFDMA Systems,” Tech.
Rep. 12, Dec. 2013.
[18] A. Bannour, C. Sacchi, and Y. Sun, “MIMO-OFDM Based Energy Harvesting Cooperative Communications Using Coalitional Game Algorithm,”
IEEE Trans. Veh. Technol., vol. 66, no. 12, pp. 11 166–11 179, Dec. 2017.
[19] Y. Sun, D. W. K. Ng, J. Zhu, and R. Schober, “Robust and Secure Resource
Allocation for Full-Duplex MISO Multicarrier NOMA Systems,” arXiv
preprint arXiv:1710.01391, 2017.
[20] X. Zhou, C. K. Ho, and R. Zhang, “Wireless Power Meets Energy Harvesting: A Joint Energy Allocation Approach in OFDM-Based System,”
IEEE Trans. Wireless Commun., vol. 15, no. 5, pp. 3481–3491, May 2016.
[21] “LTE
Unmanned
Aircraft
Systems,”
Qualcomm
Technologies,
Tech.
Rep.,
May
2017,
[Online]
https://www.qualcomm.com/media/documents/files/lte-unmanned-aircraft-systems-trial-report.pdf.
[22] A. Al-Hourani, S. Kandeepan, and A. Jamalipour, “Modeling Air-toGround Path Loss for Low Altitude Platforms in Urban Environments,” in
Proc. IEEE Global Commun. Conf., Dec. 2014, pp. 2898–2904.
[23] F. X. Kneizys, E. Shettle, L. Abreu, J. Chetwynd, and G. Anderson, “Users
Guide to LOWTRAN 7,” Air Force Geophysics Lab, Tech. Rep., 1988.
[24] A. J. LaRocca, “Methods of Calculating Atmospheric Transmittance and
Radiance in the Infrared,” Proceedings of the IEEE, vol. 63, no. 1, pp.
75–94, Jan. 1975.
[25] J. Jang and K. B. Lee, “Transmit Power Adaptation for Multiuser OFDM
Systems,” IEEE J. Select. Areas Commun., vol. 21, no. 2, pp. 171–178,
Feb. 2003.
[26] Q. T. Dinh and M. Diehl, “Local Convergence of Sequential Convex
Programming for Nonconvex Optimization,” in Recent Advances in Optimization and its Applications in Engineering. Springer, 2010, pp. 93–102.
[27] M. Grant and S. Boyd, “CVX: Matlab Software for Disciplined Convex
Programming, version 2.1,” [Online] http://cvxr.com/cvx, Mar. 2014.
| 7cs.IT
|
Bidirectional deep-readout echo state networks
Filippo M. Bianchi1 , Simone Scardapane2 , Sigurd Løkse1 and Robert Jenssen1
∗
arXiv:1711.06509v3 [cs.NE] 13 Feb 2018
1- Machine Learning Group – UiT the Arctic University of Norway
2- Sapienza University of Rome – Dept. of Information Engineering,
Electronics and Telecommunications (DIET)
Abstract. We propose a deep architecture for the classification of multivariate time series. By means of a recurrent and untrained reservoir we
generate a vectorial representation that embeds temporal relationships in
the data. To improve the memorization capability, we implement a bidirectional reservoir, whose last state captures also past dependencies in the
input. We apply dimensionality reduction to the final reservoir states to
obtain compressed fixed size representations of the time series. These are
subsequently fed into a deep feedforward network trained to perform the
final classification. We test our architecture on benchmark datasets and
on a real-world use-case of blood samples classification. Results show that
our method performs better than a standard echo state network and, at
the same time, achieves results comparable to a fully-trained recurrent
network, but with a faster training.
1
Introduction
Reservoir computing (RC) is an established paradigm for modeling nonlinear
temporal sequences [1]. In machine learning tasks, echo state networks (ESNs)
are the most common RC models, wherein the input sequence is projected to
a high-dimensional space through the use of a (fixed) nonlinear recurrent reservoir [1]. Learning is performed by applying simple linear techniques in the
high-dimensional reservoir space. The lack of flexibility in the recurrent part is
balanced by a range of advantages, including faster training compared to other
recurrent neural networks (RNNs). In tasks requiring a limited amount of temporal memory, ESNs achieve state-of-the-art results in many real-world scenarios
constrained by time budgets, low-power hardware and limited data [2]. On the
other hand, fully-trained RNNs trade architectural and training complexity with
more accurate representations and a larger memory capability [3].
We propose an architecture for time series classification called Bidirectional
Deep-readout ESN (BDESN), which combines the training speed of RC with the
accuracy of trainable RNNs. We equip our model with a bidirectional reservoir
to curtail the short-term memory limitation in ESNs. Bidirectional architectures
have been successfully applied in RNNs to extract temporal features from the
time series that accounts also for dependencies very far in time [4]. Differently
from the classic linear readout found in ESNs, we generate the desired output
∗ This work is partially funded by the Norwegian Research Council FRIPRO grant no.
239844 on the Next Generation Learning Machines and IKTPLUSS grant no. 270738 Deep
Learning for Health.
by processing the temporal features with a deep feedforward network trained
with gradient descent. Although we loose a closed form solution, the whole
architecture can still be trained quickly, since the recurrent part is fixed and the
time consuming unfolding procedure [5] is avoided. Additionally, we can leverage
recent advancements in deep learning to speed up the learning [6]. The temporal
features generated by the bidirecitonal reservoir are projected into a subspace
before being fed into to feedforward network, to reduce model and training
complexity [7]. Deep feedforward networks on top of RNNs demonstrated to
enhance the representative power of the state space of a recurrent model [8].
However, so far their application in reservoir computing has been limited, due
to difficulties in training [1]. We note that our model differs from the recently
proposed deep ESNs [9, 10], which are built by concatenating multiple (fixed)
reservoirs, with no modification of the adaptable part. The two ideas are in fact
complementary, and can in principle be used together.
We empirically validate our model against standard ESNs and fully-trainable
RNNs on multiple benchmark datasets. We show that BDESN vastly outperforms an ESN, achieving a competitive accuracy with respect to RNNs, while
being orders of magnitude faster to train.
2
Methodology
The BDESN assigns to a multivariate time series (MTS) x = {xt }Tt=0 a class label
c through a three-step procedure. First, a special recurrent reservoir generates
a representation of x that has fixed (large) size and embeds temporal features.
Then, a dimensionality reduction algorithm projects the reservoir outputs in a
lower dimensional space. Finally, a multilayer perceptron (MLP) classifies the
vectorial representation of x. In Fig. 1 we depict the whole architecture, whose
details are discussed in the following.
PCA
0
𝑇
𝑥
ℎത 𝑇
𝑥′
Input
𝑐
Reservoir
Last
state
Dim.
Reduct.
MLP
Fig. 1. Schematic illustration of the BDESN architecture.
2.1
Extracting features with a reservoir
The reservoir is governed by the following state-update equation
ht = f (Wh ht−1 + Wi xt ),
(1)
where ht is the internal time-dependent state, which combines the current input xt and the previous computations ht−1 . The function f is a nonlinear
activation (usually a tanh), Wh is a sparse matrix that defines the recurrent
self-connections in the reservoir, and Wi defines input connections. Both matrices are randomly generated and left untrained, and the reservoir behavior
is controlled mainly by three hyperparameters. These are the state size N , the
spectral radius ρ of Wh , and scaling of the inputs ω. Through an optimal tuning
of these hyperparameters the reservoir produces rich dynamics and its internal
states can be used to solve many prediction and classification tasks [1].
The last state hT generated by the reservoir, after the whole input x is processed, is a high-level representation of fixed size that embeds the temporal dependencies of x. Since the reservoir trades its internal stability with a vanishing
memory of the past inputs [11], at time T the state maintains scarce information about the first inputs. To alleviate this issue, we feed to the same reservoir
also the MTS in reverse order, x0 = {xT −t }Tt=0 and we generate a new representation h0 T that is more influenced by the first inputs. A final representation
T
is obtained by concatenating the two states, h̄T = [hT ; h0 T ] . A bidirectional
reservoir has been recently used for time series prediction [12]. However, the
architecture proposed in this work is significantly different, mainly because we
deal with classification tasks.
2.2
Dimensionality Reduction
A dimensionality reduction method maps data from high to a lower dimensional
space, usually relying on unsupervised criteria. In a classification setting supervised techniques can also be adopted to improve the dimensionality reduction
procedure, although we do not consider them in this paper.
Popular algorithms are principal component analysis (PCA) and kernel PCA,
which project data on the first d eigenvectors of a covariance matrix. PCA and
kPCA have been already successfully applied to ESN reservoir states, improving
its prediction and modeling capabilities [7]. In BDESN, dimensionality reduction provides a strong regularization that prevents overfitting in the classifier
operating on the reservoir states.
2.3
Multilayer Perceptron
In a standard ESN, the output layer is a linear readout that is quickly trained
by solving a convex optimization problem. However, we hypothesize that a simple linear model does not possess sufficient representational power for modeling
the high-level embeddings resulting from our reservoir space. For this reason,
in BDESN the classification is performed by a MLP with L layers on the representations generated at the previous steps. The number of layers in the MLP
thus determines a “feedforward” depth to the RNN [8]. Deep MLPs are known
for their capability of disentangling factors of variations from high-dimensional
spaces, and can nowadays be trained efficiently with the use of sophisticated
regularization techniques. In particular, we leverage two procedures, namely
dropout and the minimization of the L2 norm of the model parameters (i.e.,
weight decay). Thanks to the dimensionality reduction step, we can greatly reduce the number of parameters in the MLP, hence its variance, speed up the
computation and obtain a faster convergence during training.
3
Experiments
To evaluate the proposed architecture, we compare the classification accuracy
and the training time of BDESN with two different RNNs. The first is a standard
ESN, configured with the same reservoir (only unidirectional) of BDESN and
with linear readout trained with ridge regression. The second is a network
where the vectorial representations are generated by a recurrent layer of gated
recurrent units (GRUs), which are fed into a MLP equivalent to the one of
BDESN. Contrarily to the reservoir, the GRU layer is trained supervisedly.
Each RNN depends on different hyperparameters, which are optimized by
means of random search. The optimized RNNs are trained 10 times, using random and independent model initializations on a Intel Core i7-6850K CPU@4GHz.
The code is implemented in Tensorflow and is publicly available online, along
with a detailed description of the experimental setup and validation procedure1 .
3.1
Benchmark datasets
We perform classification of MTS from different datasets, whose details are reported in Tab. 1.
Dataset #V Train
Test
DistPhal 1
ECG
2
Libras 2
Ch.Traj. 3
Wafer
6
Jp.Vow. 12
139
100
180
2558
896
370
400
100
180
300
298
270
Table 1. Time series dataset
#C T min T max Source details. Column 2 to 5 report
the number of variables
3 80
80
UCR
(#V ), samples in training
2 39
152 UCR
and test set, and number of
15 45
45
UCI
classes (#C), respectively.
20 109 205 UCI
Tmin is the length of the
2 104 198 UCR
shortest MTS in the dataset
9 7
29
UCI
and Tmax the longest MTS.
Classification accuracy on the test sets and training time (in minutes) obtained by each RNN are reported in Tab. 2.
Results show that BDESN consistently performs better than ESN, by obtaining an increment of the accuracy up to 30 percentage points (Character
Trajectory dataset). BDESN outperforms GRU, sometimes slightly, in several
classification tasks, except in the Wafer dataset. Average GRU accuracy is
higher also in ECG, but the difference is not significant due to the high standard deviation in the results. While BDESN is slower than ESN, it can still be
trained up to 70 times faster than GRU, providing a huge impact in terms of
the required computational resources.
1 github.com/FilippoMB/Bidirectional-Deep-reservoir-Echo-State-Network
Table 2. Classification accuracy and training time (minutes) obtained by each RNN on
different datasets in 10 independent runs. Best average accuracies are in bold.
Dataset
ESN
GRU
Accuracy Time Accuracy Time
BDESN
Accuracy Time
DistPhal
ECG
Libras
Ch.Traj.
Wafer
Jp. Vow.
71.5±3.3
65.6±4.63
76.2±2.77
43.4±2.72
89.7±0.9
85.8±2.34
73.7±0.91
78.5±1.5
78.4±1.9
77.6±5.53
90.3±0.8
94.72±1.12
0.11
0.03
0.03
0.23
0.06
0.06
69.4±2.4
79.1±4.3
72.7±4.8
72.1±6.4
98.2±0.9
94.45±1.88
34.33
10.35
5.27
40.28
35.38
6.03
1.15
0.22
0.53
0.63
0.76
0.58
Training Loss
Objective function
100
10
2
10
4
10
6
Fig. 2. Cross entropy
loss assumed by GRU
and BDESN during
training. In BDESN,
the loss is more stable
and, in average,
assumes lower values.
GRU
BDESN
0
10000
20000 30000
Iteration
40000
50000
Even if GRU during the training phase may reach a lower cross entropy on
the training set, in BDESN the average loss is usually lower and the training is
more stable. For example, in Fig. 2 we report the cross entropy values reached
during training on the Japanese Vowels dataset. We observe that GRU is more
unstable and converges to less accurate values than BDESN.
3.2
Classification of blood samples time series
We test our method also on a real-world dataset of blood measurements collected from patients undergoing a gastrointestinal surgery at University Hospital of North Norway in 2004–2012. Each patient in the dataset is represented
by a MTS of 10 blood samples measurements extracted within 20 days after
surgery. We consider the problem of classifying patients with and without surgical site infections from their blood samples. The MTS contains missing data,
corresponding to measurements not collected for a given patient in one day of
the observation period, which are replaced with mean-imputation. The dataset
consists of 883 MTS, of which 232 are relative to patients with infections.
RNN
F1 score
ESN
0.734±0.041
GRU
0.779±0.018
BDESN 0.761±0.054
Time (mins)
0.01
10.54
1.87
Table 3. F1 score and time
of the different models on the
blood samples dataset. Best
average result is in bold.
In Tab. 3 we report the F1 score and time obtained by the different models
on the blood samples data. Since the dataset is imbalanced, F1 score in this
case is a more suitable measure of performance than accuracy. We observe that
the performance of GRU and BDESN are comparable and both are better than
ESN. GRU mean F1 score is somehow higher and, by looking at the standard
deviation, we notice its performance to be slightly more stable. However, GRU
trades this moderate improvement with a significant increment in training time,
which is 1 and 2 orders of magnitude higher than BDESN and ESN, respectively.
4
Conclusions
We proposed a novel recurrent architecture for time series classification, the bidirectional deep-readout ESN, which captures dependencies in the data forward
and backward in time, by means of a large untrained reservoir. The representation yielded by the reservoir is firstly regularized with PCA and then classified by
a MLP trained supervisedly. Our approach trains much faster than a RNN configured with GRU cells to generate the same kind of representations. We tested
our architecture on several benchmark datasets and one real-world use-case of
time series classification. Results indicate that BDESN performance is much
better than ESN and sometime even better than a fully trained GRU network.
References
[1] M. Lukoševičius and H. Jaeger. Reservoir computing approaches to recurrent neural
network training. Computer Science Review, 2009.
[2] S. Scardapane and D. Wang. Randomness in neural networks: an overview. Wiley
Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 2017.
[3] I. Sutskever, O. Vinyals, and Q. V Le. Sequence to sequence learning with neural networks.
In Advances in Neural Information Processing Systems 27. 2014.
[4] A. Graves and J. Schmidhuber. Framewise phoneme classification with bidirectional
LSTM and other neural network architectures. Neural Networks, 2005.
[5] F. M. Bianchi, E. Maiorino, M. C. Kampffmeyer, A. Rizzi, and R. Jenssen. Recurrent
Neural Networks for Short-Term Load Forecasting: An Overview and Comparative Analysis. Springer, 2017.
[6] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT press, 2016.
[7] S. Løkse, F. M. Bianchi, and R. Jenssen. Training echo state networks with regularization
through dimensionality reduction. Cognitive Computation, 2017.
[8] R. Pascanu, C. Gulcehre, K. Cho, and Y. Bengio. How to construct deep recurrent neural
networks. arXiv preprint arXiv:1312.6026, 2013.
[9] C. Gallicchio and A. Micheli. Deep reservoir computing: A critical analysis. In ESANN,
2016.
[10] C. Gallicchio, A. Micheli, and L. Pedrelli. Deep reservoir computing: a critical experimental analysis. Neurocomputing, 2017.
[11] F. M. Bianchi, L. Livi, and C. Alippi. Investigating echo-state networks dynamics by
means of recurrence analysis. IEEE Transactions on Neural Networks and Learning
Systems, 2016.
[12] A. Rodan, A. F. Sheta, and H. Faris. Bidirectional reservoir networks trained using SVM+
privileged information for manufacturing process modeling. Soft Computing, 2017.
| 9cs.NE
|
System and source identification from operational vehicle responses:
A novel modal model accounting for the track-vehicle interaction✩
Giovanni De Filippis, Davide Palmieri, Leonardo Soria∗, and Luigi Mangialardi
Dipartimento di Meccanica, Matematica e Management, Politecnico di Bari, Viale Japigia 182, 70126 Bari, Italy
arXiv:1702.08325v1 [cs.SY] 7 Jan 2017
Abstract
Operational Modal Analysis (OMA) is a powerful tool, widely used in the fields of structural identification and health
monitoring, and certainly eligible for identifying the real in-operation behaviour of vehicle systems. Several attempts
can be found in the literature, for which the usage of algorithms based on the classical OMA formulation has been
strained for the identification of passenger cars and industrial trucks. The interest is mainly focused on the assessment
of suspension behaviour and, thus, on the identification of the so-called vehicle rigid body modes. But issues arise
when the operational identification of a vehicle system is performed, basically related to the nature of the loads
induced by the roughness of rolling profiles. The forces exerted on the wheels, in fact, depending on their location,
are affected by time and/or spatial correlation, and, more over, do not fit the form of white noise sequences. Thus,
the nature of the excitation strongly violate the hypotheses on which the formulation of classical OMA modal model
relies, leading to pronounced modelling errors and, in turn, to poorly estimated modal parameters. In this paper, we
develop a specialised modal model, that we refer to as the Track-Vehicle Interaction Modal Model, able to incorporate
the character of road/rail inputs acting on vehicles during operation. Since in this novel modal model the relationship
between vehicle system outputs and modal parameters is given explicitly, the development of new specific curve fitting
techniques, in the time-lag or frequency domain, is now possible, making available simple and cost-effective tools for
vehicle operational identification. More over, a second, but not less important outcome of the proposed modal model
is the usage of the resulting techniques for the indirect characterisation of rolling surface roughness, that can be used
to improve comfort and safety.
Keywords: system identification, modal model, vehicle systems, surface roughness, correlated inputs
1. Introduction
The identification of the real in-operation behaviour is of utmost importance in all the stages of the vehicle
design process, from model updating, extensively used as a tool for improving the accuracy of dynamics simulations,
to optimisation strategies and techniques, product life-cycle management implementations, and the verification of
controllers’ performance [1, 2, 3]. In this context, a wheeled vehicle, cruising at constant speed, on a certain road
profile, is a common working occurrence for designers, in which the vehicle is mainly subjected to the external loads
exerted by the interaction with the rolling rough surface. From an experimental point of view, since unknown steadystate inputs are applied to the system, Operational Modal Analysis (OMA), even referred to as in-Operation or
Output-only Modal Analysis, is certainly eligible to perform the system structural identification, moving from output
data only [4, 5]. Specifically, in the case of linear systems subjected to operational loads, OMA allows for estimating
the so-called modal parameters, that is the resonance frequencies, the damping ratios, the mode shapes, and the
operational reference vectors, that can be utilised, in turn, to obtain a mathematical model of the system under test,
generally referred to as the modal model [6].
Ordinary vehicle systems generally exhibit a very complex dynamics, mainly owing to nonlinearities related to
shock absorbers’ operation and to the kinematics imposed by the suspension design [7, 8]. Thus, modal models
synthesised under the real operational loadings actually represent an equivalent linearisation of the tested nonlinear
systems around interesting and representative working points [9, 10]. Different speeds, manoeuvres, as cruising or
constant radius cornering, the presence of payloads, and of different adopted designs of suspension result in different
equivalent linearised models.
A novel, but even challenging application of OMA is being proposed in the framework of vehicle ride dynamics [11,
12]. Since modal parameter estimation allows for quantifying the overall damping associated to the vehicle rigidbody modes, OMA of vehicle systems can be employed as a tool for assessing the performance of different suspension
systems. More over, in the case of semi-active and active suspensions, the application of OMA is extremely interesting
for the validation and optimisation of controllers’ performance [12].
✩ Manuscript
submitted to the Journal of Sound and Vibration, in the present form, on April 1, 2016.
author. Tel.: +390805962813
Email address: [email protected] (Leonardo Soria)
∗ Corresponding
Preprint submitted to Elsevier
April 1, 2016
During the last decade, several robust methods have been developed for estimating reliable modal parameters from
output-only data [13]. Basically, these methods rely on the Natural Excitation Technique (NExT) assumptions [14,
15, 16, 17]: (i) The unknown loads acting on the system have to fit the form of white noise sequences, (ii) in case
of multi-point excitation, the external inputs are required to be strictly uncorrelated. The NExT assumptions are
commonly satisfied in civil engineering applications, as is the case of high rise buildings and suspension bridges, where
typically the structures are excited by environmental loadings due to traffic, wind and waves [18].
In this paper, we first focus on the issues that arise when an operational structural identification of a vehicle
system is performed by using classical OMA based techniques. These issues are related to several aspects: (i) The
presence of high modal density and closely-spaced modes; (ii) the high amount of damping owing to the presence
of shock-absorbers; (iii) the nature of the loads induced by the rolling surface. Specifically, we show that the latter
point implies a strong violation of the NExT assumptions. In fact, the trend of each excitation, transmitted to the
vehicle through the wheel stiffness and damping, is quite far from the flat frequency shaping of white noise [19, 20].
In addition, depending on their location, the inputs acting on the wheels result affected by time and/or spatial
correlation [21, 22, 23, 24, 25, 26]. As a consequence, the application of classical OMA methodologies for postprocessing the vehicle output responses may lead to pronounced modelling errors and, therefore, to poorly estimated
modal parameters.
Thus, by assuming that the rolling surface is an homogeneous Gaussian random field [27, 28], we formulate of a
novel OMA modal model, referred to as the Track-Vehicle Interaction Modal Model (TVIMM), specialised and suitable
for developing new identification procedures aimed at the estimation of the modal parameters of a vehicle system.
More over, we show that the coefficients of the adopted empiric surface model can be even estimated as a further
outcome of our OMA processing. Since direct measurements performed by using specific profilometers can result to
be expensive, this second application of the model has relevant interest in road and railway health monitoring. In this
field, in fact, the indirect characterisation of surface roughness can be employed as a tool for improving safety and
comfort [29].
The rest of the paper is organised as follows. In Section 2, we obtain the representation of surface-induced
forces exerted on a vehicle and discuss the issues arising when identification procedures based on the classical OMA
formulation are used to process the output responses of vehicle systems. In Section 3, we propose a novel OMA
formulation, specifically designed for the operational identification of vehicle systems, based on a specialised modal
model accounting for the track-vehicle interaction. In Section 4, we offer a numerical demonstration of the developed
formulation. Concluding remarks are summarised in Section 5. In Appendix A, we provide mathematical proof of
some fundamental equations used in Section 3.
2. System and source identification from operational vehicle responses
The nature of loadings exerted on a vehicle by the rolling surface plays a fundamental role in the process of
operational identification of the system. We here exploit the properties of homogeneous Gaussian random fields to
derive a model of the surface-induced forces. By comparing the resulting surface inputs with those permitted by
complying the NExT assumptions, we conclude that a specialised OMA formulation is needed, able to provide a
correct modelling of the system and of the main excitation source, that is the surface roughness.
2.1. Classical OMA approach
By processing the output responses of a N degrees of freedom (dofs) system, OMA leads to the estimation of the
T
∈ CN ×1 and the operational
modal parameters, that is the poles λn , the modal vectors ψ n = ψ1n · · · ψN n
T
∈ CN ×1 (with n = 1, . . . , N ), where the symbol (.)T indicates matrix
reference vectors φn = φ1n · · · φN n
transposition. The computation of system’s poles λn is of fundamental interest as they contain information about the
resonance frequencies fn = ωn /2π and the damping factors ζn
p
(1)
λn , λ∗n = −ζn ωn ± iωn 1 − ζn2 ,
where (.)∗ indicates complex-conjugation. The application of methods of system identification to OMA has given rise
to a large variety of estimation techniques [6, 30], comprising, generally, several steps. The most advanced procedures
are based on the usage of the modal model, commonly utilised to complete the identification step and for validating
the results of the estimation process [17]. Recently, a new approach has been proposed, allowing for the computation
of modal parameters directly from the modal model [31].
In the case of classical OMA, the modal model is formulated both in the time-lag τ and in the angular frequency
ω = 2πf domains, by referring to correlation functions and power spectral densities (PSDs), respectively. Thus, by
considering a system of external forces relying on the NExT assumptions, the input PSD matrix S f (ω) ∈ CN ×N can
be written as
S1
0
..
(2)
S f (ω) =
.
.
0
2
SN
In fact, since the inputs are required to be strictly uncorrelated white noise sequences, the matrix in Eq. (2) has only
constant entries along the main diagonal. With referring to a N dofs system subjected to the operational loadings
Eq. (2), the formulation of the modal model can be given in terms of output correlation matrix Rq (τ ) ∈ RN ×N and
output PSD matrix S q (ω) ∈ CN ×N , as
Rq (τ ) =
2N
X
φn ψ Tn e+λn τ h (τ ) + ψ n φTn e−λn τ h (−τ ) ,
(3)
n=1
S q (ω) =
where h(.) indicates the Heaviside step function.
2N
X
φn ψ Tn
ψ n φTn
+
,
iω − λn
−iω − λn
n=1
(4)
2.2. Random fields for surface profiles modelling
Generally, in case of OMA, input loads are uncontrollable and, in addition, remain unmeasured. For this reason,
the analyst has to verify that the operational loads are suitable to adequately and effectively excite the system in the
frequency range of interest.
A vehicle system in steady-state working conditions is subjected to both external and on-board excitations [1].
On-board forces are related to the operation of rotating parts and are specifically imputable to engine operation and to
non-uniformities in assemblies and components of the driveline. These forces introduce harmonics in the spectra, whose
frequencies are expected to depend on the engine rotating velocity and, in turn, on the vehicle speed. External loads
are mainly related to surface roughness, that represent the only present source able to generate an effective broad band
excitation. Specifically, the cut-off frequency of the input spectrum depends on roughness spatial frequency content,
vehicle velocity and wheel dynamic behaviour.
Random fields for surface modelling have extensively been studied and the usage of different approaches has
emerged [27, 28]. Basically, the following assumptions are commonly adopted: (i) the surface roughness is an homogeneous random field; (ii) the height of the asperities satisfies to a zero-mean Gaussian distribution; (iii) the pavement
unevenness is an ergodic random process. Homogeneity implies that the statistical properties of surface roughness
are independent on spatial observations. This assumption has implications similar to stationarity for one-dimensional
random processes. The second assumption ensures that the output responses of a linear system subjected to pavement excitations satisfy, as well, to a zero-mean Gaussian distribution. Ergodicity guarantees that average is equal to
expectation calculated over the whole ensemble.
In simple track-vehicle interaction models, the unilateral contact point hypothesis is often utilised to describe the
forces transmitted to the vehicle through the wheel stiffness and damping. This assumption is made without loss of
generality, by considering that the distributed contact in the wheel-pavement interface acts as a low-pass filter, whose
bandwidth is governed by the contact interface itself [32]. In this case, the surface roughness can be supposed to be
a single-track random process, describing the longitudinal profile along the wheel path in the travelling direction [19].
Generally, an empiric parametric model is adopted to fit the measured PSD, and the captured surface roughness
is classified or employed in the simulations. The non-smoothed PSD is often approximated by a simple function,
involving the usage of only few parameters. A literature survey on existing different approximations for longitudinal
road profiles is presented in Ref. [20].
When more than one profile is needed, the ordinary coherence function is introduced to express the relationship
between multiple tracks [33]. Considering two parallel surface profiles di (x) and dj (x), with x the longitudinal spatial
variable, separated by a distance Wp , the ordinary coherence function is defined in the angular spatial frequency
domain ν as
Sdi dj (ν)
Γp (ν) = p
,
(5)
Sdi (ν) Sdj (ν)
where Sdi (ν) and Sdj (ν) are the auto-PSDs of di (x) and dj (x), respectively, and Sdi dj (ν) is the cross-PSD between
the two profiles. By definition, the coherence function is a real even function ranging from 0 to 1. As a consequence,
two separated surface profiles perfectly overlap at wavelengths corresponding to amplitude values of the coherence
function equal to 1. The expression of cross-PSD is obtained from Eq. (5) as
q
(6)
Sdi dj (ν) = Γp (ν) Sdi (ν) Sdj (ν)e−iβp (ν) ,
where βp (ν) is the difference between the phases related to di (x) and dj (x).
Based on the homogeneity assumption, the following properties can be exploited for simplifying the description of
multiple tracks [27, 28]:
(i) The auto-PSDs related to parallel surface profiles are coinciding and, thus, equal to the same function
Sdi (ν) = Sdj (ν) = Sd (ν) ;
3
(7)
(ii) the cross-PSD between two parallel surface profiles is a real and even function depending on auto-PSD and
coherence function
(8)
Sdi dj (ν) = Γp (ν) Sd (ν) ,
in which it is specifically useful to notice the cancellation of the phase difference βp (ν) in Eq. (8).
2.3. Problem statement
We here consider the generic Nt -wheel vehicle system represented in Fig. 1, with Nt = 4, even if the following
considerations can be easily extended to systems equipped with Nt 6= 4 wheels. The considered geometry includes
Nw = 3 different trackwidths Wp (with p = 1, . . . , Nw ) and Nl = 3 different wheelbases Ll (with l = 1, . . . , Nl ). We
hypothesise that the vehicle travels at constant velocity V on an homogeneous Gaussian surface. By writing Eqs.
from (5) to (8) in the angular frequency domain ω = νV , we obtain the following PSD matrix of surface-induced
displacements S r (ω) ∈ CNt ×Nt
SdA dB (ω) · · · SdA dD (ω)
SdA (ω)
∗
..
Sd d (ω)
SdB (ω)
.
A B
S r (ω) =
..
..
..
.
.
.
∗
SdA dD (ω)
···
···
SdD (ω)
+iωτ1
Γ2 (ω)
Γ1 (ω) e−iωτ2
1
Γ1 (ω) e
Γ1 (ω) e−iωτ1
1
e−iωτ3
Γ3 (ω) e−iωτ1
(9)
=
Sd (ω) ,
..
.
1
Γ (ω) e+iωτ2
3
Γ2 (ω)
···
1
···
where τl = Ll /V indicate the time-delays between the inputs acting on different axles.
To describe the forces transmitted to the vehicle through the wheel stiffness and damping, we first introduce the
following static gain matrix Gf r ∈ RN ×2Nt
(N −N )×N
t
t
0 (N −Nt )×Nt
0
Gf r =
,
(10)
Kt
Ct
where K t and C t ∈ RNt ×Nt are diagonal matrices in which stiffness and damping terms are, respectively, collected.
We, second, account for the PSDs of the derived processes, in addition to Eq. (9), in case tyre damping is included in
the wheel model. We, finally, obtain the following representation for the input PSD matrix related to surface-induced
forces S f (ω) ∈ CN ×N
1
iω
S f (ω) = Gf r (D (ω) ⊗ S r (ω)) GTfr with D (ω) =
,
(11)
−iω ω 2
where the matrix D(ω) ∈ C4×4 allows for taking into account the contributions related to road-induced velocities
and the symbol ⊗ denotes the Kronecker product. We stress that the sign of the imaginary part of Eq. (11) depends
on the adopted definition of correlation function, where a change of definition in the time-lag domain leads to a
complex-conjugate expression in the frequency domain.
X
L3
L1
Y
L2
d1
A
W1
d2
V
W2
W3
C
B
D
d3
Figure 1: Geometry of a generic Nt -wheel vehicle system.
By comparing Eqs. (2) and (11), we conclude that for a vehicle system subjected to operational loadings, the modal
model resulting from the classical OMA approach (Eqs. (3) and (4)) is no more valid. The arisen issues are mainly
4
related to the nature of the loads induced by the rolling surface, that violates the NExT assumptions. First, we notice
that forces applied to each wheel are basically coloured excitations, with magnitude inversely proportional to the
frequency raised to a certain power, and shaping depending on the auto-PSD of the road profile (Sd (ω) Eq. (7)) and
on the wheel parameters (Gf r Eq. (10)). In addition, being systems equipped with more than one wheel, forces are
affected by time and/or spatial correlation: (i) wheels mounted on the same axle are subjected to spatially correlated
inputs (Γp (ω) in Eq. (9)); (ii) wheels travelling on the same path and located on different axles are subjected to
time correlated inputs (e±iωτl in Eq. (9)); (iii) wheels travelling on separated paths and located on different axles are
subjected to time and spatially correlated inputs (Γp (ω) e±iωτl in Eq. (9)).
Owing to the aforementioned effects, the application of classical OMA methodologies to vehicle responses may
lead to pronounced modelling errors and, therefore, to poorly estimated modal parameters. The formulation of a
specialised modal model for correctly describing the track-vehicle interaction is required and, in turn, a specific OMA
formulation is needed to understand how Eqs. (3) and (4) are modified. We stress that a modal model providing
the relationship between the generic output of the system and its modal parameters in an explicit form is needed for
the formulation of a whatever procedure for modal parameter estimation based on suitable curve fitting algorithms.
We more over comment that since operational vehicle responses incorporate information about the surface roughness,
the identification procedures based on this novel OMA formulation would allow for estimating, in addition to modal
parameters, even the coefficients of the adopted empiric surface models.
3. The Track-Vehicle Interaction Modal Model
We here utilise a 7 dofs system (N = 7), generally referred to as the full-car model, to introduce the theoretical
background of the proposed Track-Vehicle Interaction Modal Model. To obtain the analytical expression of the generic
system output, we solve a Duhamel integral in modal coordinates. Moving from Eq. (11), we compute the correlation
matrix of surface-induced forces and achieve the TVIMM formulation by using properties of convolution integrals and
Fourier transform.
3.1. Full-car model
The full-car model (Fig. 2) is the simplest mathematical description of a four-wheel vehicle (Nt = 4), whose
predicted output responses incorporate all the effects of the issues discussed in Section 2, which make no longer
possible the usage of classical OMA methodologies for post-processing. Since full-car model offers a good trade-off
between model complexity and accuracy, this linear lumped-parameters system is commonly utilised for simulating
the ride dynamics of passenger and race cars [2, 3]. The model preserves the multi-input nature of road excitation
and, different from half-car model, allows for evaluating the contributions to vehicle responses due to roll disturbance
produced by two parallel tracks. The geometry comprises one single trackwidth W1 and one single wheelbase L1 . In
Fig. 2, we indicate the dofs of the model by using the following notation: zs , ϕs , and θs represent the heave, the roll
and the pitch rigid body motions of the sprung mass (the body); zs1 , zs2 , zs3 , zs4 denote the vertical displacements
at the four corners of the sprung mass (strut mounts to body); zu1 , zu1 , zu3 , zu4 are the rattle displacements of the
unsprung masses (the wheels). More over, with regards to system parameters, ms , jsx , jsy represent the mass and the
moments of inertia associated with the body; mu1 , mu2 , mu3 , mu4 denote the unsprung masses; kf , kr , cf , cr are the
stiffness and damping coefficients of the absorbers; kf t , krt , cf t , crt indicate the stiffness and damping coefficients of
the tyres; L1f , L1r are the distances of the front and rear axle, respectively, from the center of gravity of the unsprung
mass, the sum of which equals the wheelbase.
3.2. Duhamel integral in modal coordinates
Since vehicles are generally non-proportional damping systems, we adopt a representation in state-space form.
Thus, we recast the motion equations of a N dofs system into an equivalent set of 2N first-order differential equations.
Specifically, by denoting with q(t) ∈ RN ×1 the vector of Lagrangian coordinates and with f (t) ∈ RN ×1 the vector of
external loads, we write the well-known set of system dynamics equations as
M q̈(t) + C q̇(t) + K q(t) = f (t),
(12)
where t is the time variable, and M , C, and K ∈ RN ×N are the mass, damping, and stiffness matrices, respectively.
By adding the following further set of differential equations
M q̇(t) − M q̇(t) = 0 N ×N ,
(13)
and by making the substitutions
x(t) =
q(t)
q̇(t)
ẋ(t) =
q̇(t)
q̈(t)
u(t) =
f (t)
0 N ×1
,
(14)
we obtain the set of system dynamics equations in state-space form
P ẋ(t) + Qx(t) = u(t),
5
(15)
L1
L1r
L1f
zs1
zs
zs3
ϕs
θs
V
xs
kf
zs2
cf
W1
ys
kr
cr
zu1
mu1
kf
kf t
mu3
cf
zu2
cf t
zs4
zu3
ms jsx jsy
kr
cr
zu4
crt
krt
mu2
mu4
d1
A
kf t
C
cf t
crt
krt
d2
B
D
Figure 2: Schematic representation of the full-car model.
where P and Q ∈ R2N ×2N are partitioned as
C
M
P =
M 0 N ×N
Q=
K
0 N ×N
0 N ×N
−M
.
(16)
Here, the product −P −1 Q leads to the so-called state matrix A. The eigenvalue decomposition of the state matrix
A is
A = V ΛV −1 ,
(17)
where Λ ∈ C 2N ×2N is the eigenvalue matrix, containing the system poles
Λ = diag (λ1 , . . . , λN , λ1 ∗ , . . . , λN ∗ ) ,
(18)
with diag(.) denoting a diagonal matrix, while V ∈ C 2N ×2N is the eigenvector matrix, having the following structure
ψ1
···
ψN
ψ ∗1
···
ψ ∗N
V =
.
(19)
λ1 ψ 1 · · · λN ψ N λ∗1 ψ ∗1 · · · λ∗N ψ ∗N
We decouple the equations of motion by using the following coordinate transformation from the physical to the modal
space
X
2N
ψn
q(t)
pn (t) ,
(20)
x(t) = V p(t) ⇒
=
λn ψ n
q̇(t)
n=1
T
where p(t) = p1 (t) · · · p2N (t)
∈ RN ×1 is the modal state vector. By substituting Eq. (20) in Eq. (15) and
T
pre-multiplying both sides by V , we obtain 2N independent differential equations, that we collect in the following
compact form
(21)
M a ṗ(t) + M b p(t) = V T u(t),
where M a and M b ∈ C2N ×2N are two diagonal matrices generally referred to as modal a and modal b. Specifically,
M a = V T P V = diag (ma1 , . . . , maN , m∗a1 , . . . , m∗aN )
(22)
M b = V T QV = diag (mb1 , . . . , mbN , m∗b1 , . . . , m∗bN ) ,
(23)
and
while
V T u (t) =
ψ1
···
ψN
ψ ∗1
···
ψ ∗N
T
f (t) .
(24)
By considering the generic system motion equation of set Eq. (21) referred to the n-th mode of vibration,
man ṗn (t) + mbn pn (t) = ψ n T f (t) with
6
mbn = −λn man ,
(25)
and by assuming zero initial conditions, we retrieve the solution in the form
ψ T
pn (t) = n
man
Zt
f (ǫ)eλn (t−ε) dε.
(26)
−∞
Eq. (26), usually written in physical coordinates, is the well-known Duhamel integral. By combining Eqs. (20) and (26),
we obtain the system response in terms of the physical Lagrangian coordinates
Zt
2N
X
ψnψn T
ψ n pn (t) =
q(t) =
f (ρ)eλn (t−ρ) dρ.
m
an
n=1
n=1
2N
X
(27)
−∞
We comment that Eq. (26) can be interpreted as a convolution integral encompassing the impulse response matrix
g(t) ∈ RN ×N of the system under study
2N
X
ψ n ψ n T λn t
g (t) =
e .
(28)
man
n=1
3.3. Input and output correlation functions
By considering the generic outputs qi (t) and qj (t) in Eq. (27), evaluated at the separated time instants t and t + τ ,
respectively,
Zt
2N
X
ψin ψ n T
f (ρ) eλn (t−ρ) dρ,
(29)
qi (t) =
m
an
n=1
−∞
and
t+τ
Z
2N
X
ψjm ψ m T
f (σ) eλm (t+τ −σ) dσ,
qj (t + τ ) =
m
am
m=1
(30)
−∞
where the following relations hold between the dummy integration variables
ρ→t
and σ → t + τ
σ =ρ+τ
⇔
and τ = σ − ρ,
(31)
we derive the resulting output cross-correlation function
Rqi qj (τ ) = E [qi (t) qj (t + τ )]
t+τ Z t
Z
2N X
2N
h
i
X
ψin ψjm
=
ψ n T E f (ρ) f T (ρ + τ ) ψ m eλn (t−ρ) eλm (t+τ −σ) dρdσ
man mam
n=1 m=1
−∞ −∞
2N
2N X
X
ψin ψjm
=
m
an mam
n=1 m=1
t+τ Z t
Z
ψ n T Rf (τ ) ψ m eλn (t−ρ) eλm (t+τ −σ) dρdσ,
(32)
−∞ −∞
where the symbol E[.] indicates the expectation computed over the ensemble and Rf (τ ) ∈ RN ×N is the input correlation matrix related to the external forces, which, by definition, is the inverse Fourier transform of the input PSD
matrix Eq. (11).
By particularising to the case of full-car model (Fig. 2) the stiffness and damping entries of the static gain matrix
Gf r Eq. (10)
K t = diag(kf t , kf t , krt , krt )
C t = diag(cf t , cf t , crt , crt ),
(33)
we decompose the PSD matrix of surface-induced displacements S r (ω) through the following summation
S r (ω) = Sd (ω)
Nw
X
Γp (ω)∆p (ω)
with Γ0 (ω) = 1,
(34)
p=0
where Nw = 1 in the case of full car model and the matrices ∆0 (ω) and ∆1 (ω) ∈ CNt ×Nt are defined as
0
1
0
e−iωτ1
0
1
0
e−iωτ1
,
∆0 (ω) =
e+iωτ1
0
1
0
+iωτ1
0
1
0
e
7
(35)
∆1 (ω) =
0
1
0
e
1
0
e
+iωτ1
+iωτ1
0
e−iωτ1
0
.
1
0
0
e−iωτ1
0
1
(36)
By substituting Eq. (34) in Eq. (11), we obtain the Fourier pair S f (ω) and Rf (τ ) in the case of interest
S f (ω) = Sd (ω)
Nw
X
Γp (ω) Gf r (D (ω) ⊗ ∆p (ω)) GTfr
with Γ0 (ω) = 1,
(37)
p=0
and
Rf (τ ) = Rd (τ ) ∗
Nw
X
p=0
Hp (τ ) ∗ Gf r (d (τ ) ⊗ δ p (τ )) GTfr
with
H0 (τ ) = 1,
(38)
where Rd (τ ) is the auto-correlation function associated to surface profiles and Hp (τ ) (with p 6= 0) represents the
inverse Fourier transform (IFT) of the ordinary coherence function Γp (ω) referred to two parallel tracks. The terms
d(τ ) ∈ R4×4 , δ 0 (τ ) and δ 1 (τ ) ∈ RNt ×Nt correspond to a representation in the time-lag domain of the matrices included
in Eq. (37), that is
"
#
d(.)
1
dτ
d(τ ) =
,
(39)
2
− d(.)
− ddτ(.)
2
dτ
δ (τ )
0
δ (τ − τ1 )
0
0
δ (τ )
0
δ (τ − τ1 )
,
δ 0 (τ ) =
(40)
δ (τ + τ1 )
0
δ (τ )
0
0
δ (τ + τ1 )
0
δ (τ )
0
δ (τ )
0
δ (τ − τ1 )
δ (τ )
0
δ (τ − τ1 )
0
,
δ 1 (τ ) =
(41)
0
δ (τ + τ1 )
0
δ (τ )
δ (τ + τ1 )
0
δ (τ )
0
2
d (.)
where δ(.) indicates the Dirac delta function, and the symbols d(.)
dτ and d2 τ denote the first and second order derivative
operators, respectively. By combining Eqs. (38) and (32), we achieve the final expression of the cross-correlation
function Rqi qj (τ )
t+τ Z t
Z
2N X
2N
X
ψin ψjm
Rqi qj (τ ) = Rd (τ ) ∗
Hp (τ ) ∗
ψ Tn Gf r (d (τ ) ⊗ δ p (τ )) GTfr ψ m eλn (t−ρ) eλm (t+τ −σ) dρdσ.
m
m
an
am
p=0
n=1 m=1
Nw
X
−∞ −∞
(42)
3.4. Modal model in the time-lag and frequency domains
Moving from Eq. (42), by means of calculations detailed in Appendix A, we obtain the relationship between the
pl
operational modal parameters (λn , ψin , αpin , βin
, χpl
in ) and the generic output cross-correlation function Rqi qj (τ ). We
here rephrase the expression of the output correlation matrix Rq (τ ), in the following compact matrix notation
Rq (τ ) =
2N
X
ϕ̄n (τ )ψ Tn e+λn τ + ψ n ϕ̄Tn (−τ )e−λn τ ,
(43)
n=1
where the terms ϕ̄n (τ ) are the lag-dependent operational reference vectors, defined as
ϕ̄n (τ ) = Rd (τ ) ∗
NW
X
Hp (τ ) ∗
αpn h (τ )
p=0
+
NL
X
+λn τl
β pl
h (τ
ne
+ τl ) +
−λn τl
h (τ
χpl
ne
l=1
!
− τl )
with H0 (τ ) = 1.
(44)
By Fourier transforming Eq. (43), we obtain the expression of the output PSD matrix S q (ω)
2N
X
ψ ϕT (−ω)
ϕn (ω)ψ Tn
+ n n
,
S q (ω) =
iω − λn
−iω − λn
n=1
where the terms ϕn (ω) are the frequency-dependent operational reference vectors, defined as
!
NL
NW
X
X
pl
−iωτl
βn e+iωτl + χpl
Γp (ω) αpn +
ϕn (ω) = Sd (ω)
with Γ0 (ω) = 1.
ne
p=0
l=1
8
(45)
(46)
The modal model that we introduce through Eqs. (43) and (45), namely the TVIMM, incorporates a combination
of several contributions: (i) Rd (τ ) and Sd (ω) describe the statistical properties of the single surface profiles; (ii)
e±λn τl h (τ ± τl ) and e±iωτl account for the time correlation between the inputs applied to wheels travelling on the
same path and mounted on different axles; (iii) Hp (τ ) and Γp (ω) encompass the spatial correlation due to loads acting
on wheels travelling on separated paths. Thus, the vehicle output responses contain information on the dynamics of
the system and on its interaction with the surface profiles. Not considering this fundamental last contribution from
the resulting modal model (Eqs. (43) and (45)) increase the modelling errors of the identification procedure. As a
consequence, the unknown parameters result poorly estimated.
We consider useful to stress that although Eqs. (43) and (45) have been obtained by only considering the rigid-body
modes of the full-car model, this novel OMA formulation is of general validity and can be utilised for real-world vehicle
systems. Specifically, it is able to include and describe, without loss of generality, even the presence of deformable-body
modes of vibration.
3.5. Comments on the usage of half spectra
Several robust OMA algorithms in the frequency domain rely on the estimation of the so called half spectra [18],
defined as
2N
X
φin ψjn
.
(47)
(ω)
=
Sq(+)
i qj
iω − λn
n=1
The main advantages of this representation consist in: (i) The usage of lower model orders without affecting the
quality of the fitting; (ii) the adoption of well-known windowing functions for reducing the effect of leakage and the
influence of samples at the higher time-lags, which are the most affected by noise.
(+)
In the case of the classical OMA formulation, half spectra Sqi qj (ω) are computed by Fourier transforming the
correlation functions corresponding to positive time-lags
(τ ) =
Rq(+)
i qj
2N
X
φin ψjn e+λn τ h (τ ) .
(48)
n=1
We here prove that in presence of time correlation this property does no longer hold. In fact, the lag-dependent
operational reference vectors ϕ̄Tn (−τ ) do not vanish at the positive time-lags
ϕ̄Tn (−τ ) = Rd (τ ) ∗
NW
X
p=0
Hp (τ ) ∗
NL
X
T
−λn τl
β pl
h (−τ + τl ) for
n e
τ > 0,
(49)
l=1
implying that the Fourier transform of Eq. 43 for τ > 0 does not lead to the first fraction summation in Eq. 45.
4. Numerical demonstration
In this section, we offer a numerical demonstration of the effectiveness of Eqs. (43) and (45). Specifically, we first
compute the responses of the full-car model to surface-induced excitations by using the following implicit analytical
expression holding in the frequency domain
S q (ω) = G∗qf (ω) S f (ω) GTqf (ω) ,
(50)
representing the well-known input-output formula [28]. Since the full-car model utilised in this study is a linear
mutliple-input mutliple-output system, the output PSD matrix S q (ω) can be calculated at each frequency through
Eq. (50). We derive the frequency response function matrix Gqf (ω) ∈ CN ×N , referred to the lumped-parameters
system of Fig. 2, by means of the following relation
Gqf (ω) = −ω 2 M + iωC + K
−1
.
(51)
Then, we compare the obtained curves with those resulting from the TVIMM. We stress that the comparison, here
provided for demonstration purposes, is made between quantities computed by following two different routes: The
former is based on Eq. (50), in which the relationship between the generic output of the system and its modal
parameters in not explicitly given; the latter is based on the proposed OMA for vehicles’ modal model Eqs. (43)
and (45), in which the relationship is instead provided in an explicit form. Only this second formulation allows
for developing specialised curve fitting techniques, needed, in turn, for modal parameter estimation. We more over
comment that, based on the different approximations proposed in the literature, semi-analytical expressions of Sd (ω)
and Γp (ω) have to be provided, in order to form the input PSD matrix of surface-induced forces S f (ω) (Eq. (11)).
9
4.1. Matrix formulation of the simulated vehicle model
For the full-car model, two alternative sets of degrees of freedom can be used to describe the dynamics of the
sprung mass: (i) The heave zs , the roll ϕs , and the pitch θs or (ii) the vertical translations at the right-front zs1 ,
left-front zs2 , and right-rear zs3 corners. In what follows, we analyse both the two cases by introducing two different
Lagrangian coordinate vectors, respectively
q 1 (t) =
zs
ϕs
θs
zu1
zu3
zu2
zu4
The vector of surface-induced forces f (t) is
T
q 2 (t) =
zs1
zs2
zs3
zu2
zu1
cf t d˙A (t) + kf t dA (t) cf t d˙B (t) + kf t dB (t) crt d˙C (t) + krt dC (t)
zu3
zu4
T
.
crt d˙D (t) + krt dD (t)
(52)
T
.
(53)
The mass, damping and stiffness matrices of the considered vehicle model, related to the first set of Lagrangian
coordinates q 1 (t), are given by
M 1 = diag (ms , jsx , jsy , mu1 , mu2 , mu3 , mu4 ) ,
(54)
f (t) =
0
0 0
2 (cf + cr )
0
2 (cr L1r − cf L1f )
C1 =
−cf
−cf
−cr
−cr
0
0.5 (cf + cr ) W12
0
0.5 cf W1
−0.5 cf W1
0.5 cr W1
−0.5 cr W1
2 (kf + kr )
0
0
0.5
(k
+
kr ) W12
f
0
2 (kr L1r − kf L1f )
K1 =
−kf
0.5 kf W1
−k
−0.5
kf W1
f
−kr
0.5 kr W1
−kr
−0.5 kr W1
2 (cr L1r − cf L1f )
0
2 cf L21f + cr L21r
cf L1f
cf L1f
−cr L1r
−cr L1r
−cf
0.5 cf W1
−cf
−0.5 cf W1
cf L1f
cf + cf t
0
0
0
cf L1f
0
cf + cf t
0
0
−kf
0.5 kf W1
−kf
−0.5 kf W1
kf L1f
kf + kf t
0
0
0
kf L1f
0
kf + kf t
0
0
2 (kr L1r − kf L1f )
0
2 kf L21f + kr L21r
kf L1f
kf L1f
−kr L1r
−kr L1r
−cr
−0.5 cr W1
−cr L1r
−cr L1r
,
0
0
0
0
cr + crt
0
0
cr + crt
(55)
−kr
−kr
0.5 kr W1 −0.5 kr W1
−kr L1r
−kr L1r
,
0
0
0
0
kr + krt
0
0
kr + krt
(56)
−cr
0.5 cr W1
while those related to the second set q 2 (t), can be expressed by
M 2 = M 1 T −1
where T ∈ RN ×N is the transformation matrix
1
1
T =
1
C 2 = C 1 T −1
K 2 = K 1 T −1 ,
−0.5 W1 −L1f
0.5 W1 −L1f 0 4×4
−0.5 W1 L1r
0 4×4
I 4×4
.
(58)
The parameters of the full-car model employed in the simulation [3] are listed in Tab. 1.
Quantity
W1
L1f
L1r
ms
jsx
jsy
kf
kr
cf
Unit
m
m
m
kg
kg m2
kg m2
kN/m
kN/m
kN s/m
Value
1.490
1.064
1.596
1150
530
1630
15.750
14
1.475
Quantity
cr
mu1
mu2
mu3
mu4
kf t
krt
cf t
crt
Unit
kN s/m
kg
kg
kg
kg
kN/m
kN/m
kN s/m
kN s/m
Table 1: Parameters of the full-car model [3].
10
(57)
Value
1.310
57.500
57.500
57.500
57.500
140
140
0.150
0.150
4.2. Single-profile and coherence-based models for parallel road paths
We obtain the input PSD matrix of surface-induced forces by imposing that the auto-PSD of single surface profiles
Sd (ω) be equal to the ISO 8608 class C approximation. According to ISO 8608 [33], in fact, roughness level is classified
from A to H, and the form of the fitted one-sided auto-PSD is given as follows
Sd (ν) = S0
ν
ν0
−e
for νa ≤ ν ≤ νb ,
(59)
where ν0 = 1 rad/m denotes the reference angular spatial frequency and S0 (m3 ) is the amplitude of the PSD for
ν = ν0 . A constant velocity auto-PSD can be obtained by imposing the undulation exponent e equal to 2. When
this occurrence is satisfied, the auto-PSD of surface-induced displacements is calculated through a simple integration
of a flat-spectrum velocity signal. Rearranging Eq. (59), the expression of the PSD in the angular frequency domain
becomes
−2
Sd (ν = ω/V )
S0 ω
Sd (ω) =
=
for ωa ≤ ω ≤ ωb ,
(60)
V
V ω0
where ω0 = ν0 V indicates the reference angular frequency.
Since in the classification of road profiles the angular spatial frequency usually ranges from νa = 0.063 rad/m to
νb = 17.77 rad/m [33] (corresponding to wavelengths from 100 m to 0.35 m, respectively), by considering the usual
range of speed values from 10 to 30 m/s, we have that the angular frequency varies in the interesting range from
ωa = 1.89 rad/s (0.30 Hz) to ωb = 177.7 rad/s (28.3 Hz). Thus, to account for these restrictions, we introduce lower
and upper limits in Eqs. (59) and (60).
The ISO 8608 class C auto-PSD and auto-correlation function are shown in Figs. 3a and 3b, respectively. In
general, an analytical or numerical expression of the auto-correlation function Rd (τ ) can be obtained through inverse
Fourier transformation of Eq. (60). Specifically, the curve depicted in Fig. 3b, representing Rd (τ ) for the ISO 8608
case, is build up by numerically computing the IFT of Sd (ω).
(b)
(a)
−5
3
−4
2
Amplitude (m )
2
2
Amplitude (m /Hz)
Log
10
x 10
10
10
−6
1
0
−8
−1
10
−1
10
0
10
1
10
2
−10
Log
Frequency (Hz)
−5
0
Time Lag (s)
5
10
Figure 3: ISO 8608 - class C approximation (ν0 = 1 rad/m, S0 = 16 · 10−6 m3 ) for V = 20 m/s: (a) auto-PSD and
(b) auto-correlation function.
With regards to the ordinary coherence function, we hypothesize that Γp (ω) in Eq. (11) have the trend of the
fitting model proposed by Bogsjö [24], and illustrated in Fig. 4a. Experimental data acquired on very different road
tracks have surprisingly shown a good agreement with this one-parameter exponentially decreasing function. At a
constant velocity, the original exponential model can be rearranged in the angular frequency domain as
Γp (ω) = e
−µWp ω
2πV
for
ω ≥ 0.
(61)
From practical considerations, it is basically possible to conclude that Γp (ω) has to approach 1 for long wavelengths
(small ω) and 0 for short wavelengths (large ω). More over, as the two tracks become closer, that is Wp → 0, then
Γp (ω) → 1. The curve in Fig. 4b, related to Hp (τ ), is even build up by numerically computing the IFT of Γp (ω).
4.3. Analysis of the results
By solving Eqs. from (16) to (19), we compute resonance frequencies (fn ), damping ratios (ζn ) and modal vectors
(ψ n ) of the full-car model. The obtained modes of vibration are reported in Fig. 5 for increasing values of the resonance
frequency. While the computed modal parameters depend on the physical parameters of the exploited vehicle model,
some considerations of general validity can be made, owing to well-known vehicle design procedures and to the fact
that suspensions and wheels mounted on different axles have usually similar stiffness and damping values. Specifically,
11
(a)
(b)
1
7
0.9
6
0.8
5
Amplitude
Amplitude
0.7
0.6
0.5
0.4
0.3
3
2
1
0.2
0
0.1
0
0.3
4
5
10
15
20
Frequency (Hz)
25
−1
−10
30
−5
0
Time Lag (s)
5
10
Figure 4: Bogsjö exponential model (µ = 3.8) for V = 20 m/s: (a) ordinary coherence function and (b) its IFT.
we notice the presence of high modal density and closely-spaced modes, that represent a first practical issue for the
identification of vehicle systems. In the frequency band ranging from 0 to 20 Hz two main sets of rigid-body modes do
exist. The three modes involving the chassis motion are all approximately located at 1 Hz, while the four modes related
to rattle motions of front and rear axles are found in the range from 8 to 15 Hz depending on the vehicle parameter
values (8 Hz in the case of values collected in Tab. 1). A second issue is caused by the presence of shock-absorbers
with high amount of damping injected in the system, which is also a typical occurrence in vehicle systems.
In Fig. 6, we show some output cross-PSDs of the full-car model. Owing to the high damping ratios, the resonance
peaks, related to modes of vibration, are not clearly evidenced in the magnitude plots. More over, time and spatial
correlations of surface-induced excitations produce relevant effects in form of distortions of cross-spectra. The time
delay between the loads acting on the front and rear axles introduces the most significant of these distortions. In
particular, we find (i) humps appearing in the magnitude of the cross-PSDs related to chassis dofs (Figs. 6a and 6b)
and (ii) the phase plots referred to the cross-PSDs of front and rear axles (Figs. 6c and 6d) revealing the typical
saw-tooth trend caused by time-delayed signals. These distortions, well-known in the field of vehicle ride dynamics,
are imputable to the so-called wheelbase filtering effect [21]. We comment that it should be expected that the presence
in the output spectra of these distortion effects could hamper the correct operation of curve fitting methodologies
based on the classical OMA modal model, that specifically relies on the NExT hypotheses.
By inspection of Fig. 6, we find that the curves computed by using the proposed TVIMM perfectly overlap those
obtained through the usage of the input-output formula. This proves the effectiveness of the proposed Track-Vehicle
Interaction Modal Model in the frequency domain (Eq. (45)). More over, some output cross-correlation functions of
the full-car model are depicted in Fig. 7 and even in the case of time-lag domain (Eq. (43)), we obtain the expected
overlapping of the curves.
5. Conclusions
Aim of the present paper is to introduce a novel OMA modal model, referred to as the Track-Vehicle Interaction
Modal Model, needed for developing operational identification techniques, specifically designed for the case of vehicle
systems and road/rail surface roughness. The effectiveness of the proposed formulation has been proved by simulating
the response of a typical four-wheel vehicle model. Moving from the obtained expressions, specific estimation algorithms
operating in the time-lag or in the frequency domain can be developed and implemented for practical purposes,
providing new powerful tools to vehicle designers. By means of the resulting techniques, in fact, a map of vehicle
system poles corresponding to the various operating conditions, as mainly at the different cruising speeds, can be
achieved. The excitations produced during durability tests, in controlled conditions, could be utilised for providing
the required operational loadings. Since these tests are usually already performed in scheduled test campaigns foreseen
for vehicle design purposes, the usage of estimation techniques based on the modal model proposed in the paper would
result simple and cost-effective. To this aim, both equipped laboratories and specific proving grounds could be
successfully employed. As a further outcome of the proposed methodology, the resulting techniques could be used for
the indirect characterisation of surface roughness, and, in turn, as a tool for improving safety and comfort.
Acknowledgments
The authors Giovanni De Filippis and Davide Palmieri have attended a full Doctoral Program at Politecnico di
Bari, with PhD scholarships funded, respectively, through the PON01 02380 and the PON02 00576 3333585, by the
Ministry of Education, University, and Research - MIUR, which is gratefully acknowledged.
12
(a) Heave: 1.064 Hz, 27.954 %
(b) Roll: 1.221 Hz, 32.189 %
(c) Pitch: 1.295 Hz, 34.295 %
(d) Rear axle hop: 8.044 Hz, 26.031 %
(e) Front axle roll: 8.129 Hz, 27.794 %
(f) Front axle hop: 8.134 Hz, 28.372 %
(g) Rear axle roll: 8.207 Hz, 25.354 %
Figure 5: Modes of vibration of the full-car model. Blue and red lines indicate the undeformed and deformed shapes,
respectively.
13
(a)
(b)
Input−output formula
TVIMM
−5
−5
10
Magnitude (m2/Hz)
Log
Magnitude (m2/Hz)
Log
10
−10
10
−10
10
−15
10
0.3
−15
5
10
15
20
Frequency (Hz)
25
10
30
Phase (rad)
Phase (rad)
3
2
1
0
0.3
5
10
15
20
Frequency (Hz)
25
0.3
10
15
20
Frequency (Hz)
25
30
5
10
15
20
Frequency (Hz)
25
30
5
10
15
20
Frequency (Hz)
25
30
5
10
15
20
Frequency (Hz)
25
30
2
0
−2
30
(c)
5
0.3
(d)
−4
10
10
10
Phase (rad)
10
−6
−8
−10
−12
0.3
5
10
15
20
Frequency (Hz)
25
10
30
2
0
−2
0.3
10
Phase (rad)
10
Magnitude (m2/Hz)
Log
Magnitude (m2/Hz)
Log
10
5
10
15
20
Frequency (Hz)
25
−5
−10
−15
0.3
2
0
−2
0.3
30
Figure 6: Output cross-PSDs of the full-car model (Fig. 2). Comparison between the input-output formula and the
TVIMM: (a) Szs zu1 (f ), (b) Szs2 zs3 (f ), (c) Szu1 zu3 (f ) and (d) Szu1 zu4 (f ).
14
(a)
(b)
−5
4
x 10
Input−output formula
TVIMM
1
0
−1
−10
−5
0
Time Lag (s)
5
10
1
0
−2
−15
15
(d)
−5
x 10
−10
−5
0
Time Lag (s)
5
10
15
−10
−5
0
Time Lag (s)
5
10
15
−5
4
x 10
3
2
2
Amplitude (m )
2
Amplitude (m )
2
−1
−2
−15
3
3
2
2
(c)
x 10
4
Amplitude (m )
3
Amplitude (m2)
−5
5
1
0
2
1
0
−1
−15
−1
−10
−5
0
Time Lag (s)
5
10
15
−15
Figure 7: Output cross-correlation functions of the full-car model (Fig. 2). Comparison between the input-output
formula and the TVIMM: (a) Rzs zu1 (τ ), (b) Rzs2 zs3 (τ ), (c) Rzu1 zu3 (τ ) and (d) Rzu1 zu4 (τ ).
15
References
[1] T. D. Gillespie, Fundamentals of vehicle dynamics, Vol. 2005, SAE International, 1992.
[2] W. F. Milliken, D. L. Milliken, Race car vehicle dynamics, Vol. 400, Society of Automotive Engineers Warrendale,
1995.
[3] M. Guiggiani, The Science of Vehicle Dynamics, Springer, 2014.
[4] L. Hermans, H. Van der Auweraer, Modal testing and analysis of structures under operational conditions: industrial applications, Mechanical Systems and Signal Processing 13 (2) (1999) 193–216.
doi:10.1006/mssp.1998.1211.
[5] R. Brincker, P. H. Kirkegaard, Special issue on operational modal analysis, Mechanical Systems
and Signal Processing 24 (5) (2010) 1209 – 1212, special Issue:
Operational Modal Analysis.
doi:http://dx.doi.org/10.1016/j.ymssp.2010.03.005.
[6] W. Heylen, S. Lammens, P. Sas, Modal Analysis Theory and Testing, Katholieke Universteit Leuven, Departement
Werktuigkunde, 2007.
[7] J. Wallaschek, Dynamics of non-linear automobile shock-absorbers, International Journal of Non-Linear Mechanics
25 (2) (1990) 299–308. doi:10.1016/0020-7462(90)90059-I.
[8] C. Surace, K. Worden, G. Tomlinson, An improved nonlinear model for an automotive shock absorber, Nonlinear
Dynamics 3 (6) (1992) 413–429. doi:10.1007/BF00045646.
URL http://link.springer.com/article/10.1007/BF00045646
[9] T. Caughey, Equivalent linearization techniques, Journal of the Acoustical Society of America 35 (11) (1963)
1706–1711. doi:10.1121/1.1918794.
[10] P. Hagedorn, J. Wallaschek, On equivalent harmonic and stochastic linearization for nonlinear shockabsorbers, in:
Nonlinear Stochastic Dynamic Engineering Systems, Springer, 1988, pp. 23–32.
doi:10.1007/978-3-642-83334-2_2.
[11] V. Cossalter, R. Lot, F. Maggio, The modal analysis of a motorcycle in straight running and on a curve, Meccanica
39 (1) (2004) 1–16.
[12] L. Soria, B. Peeters, J. Anthonis, H. Van der Auweraer, Operational modal analysis and the performance assessment of vehicle suspension systems, Shock and Vibration 19 (5) (2012) 1099–1113. doi:10.3233/SAV-2012-0715.
[13] E. Parloo, Application of frequency-domain system identification techniques in the field of operational modal
analysis, Ph.D. thesis, Dept. of Mechanical Engineering, Vrije Universiteit Brussel, Belgium (2003).
[14] G. H. James, T. G. Carne, J. P. Lauffer, A. R. Nord, Modal testing using natural excitation, in: Poceedings of
the 10th International Modal Analysis Conference (IMAC), San Diego, California, 3rd–7th February 1992.
[15] F. Shen, M. Zheng, D. Feng Shi, F. Xu, Using the cross-correlation technique to extract modal parameters on
response-only data, Journal of Sound and Vibration 259 (5) (2003) 1163–1179. doi:10.1006/jsvi.2002.5203.
[16] J. M. Caicedo, S. J. Dyke, E. A. Johnson, Natural excitation technique and eigensystem realization algorithm for
phase i of the iasc-asce benchmark problem: Simulated data, Journal of Engineering Mechanics 130 (1) (2004)
49–60.
[17] B. Peeters, H. Van der Auweraer, Polymax: a revolution in operational modal analysis, in: Proceedings of the 1st
International Operational Modal Analysis Conference (IOMAC), Copenhagen, Denmark, 26th–27th April 2005.
[18] B. Peeters, H. Van der Auweraer, F. Vanhollebeke, P. Guillaume, Operational modal analysis for estimating the dynamic pro
Shock and Vibration 14 (4) (2007) 283–303.
URL http://iospress.metapress.com/content/4124g26592755887/
[19] S. Türkay, H. Akçay, A study of random vibration characteristics of the quarter-car model, Journal of Sound and
Vibration 282 (1) (2005) 111–124. doi:10.1016/j.jsv.2004.02.049.
[20] P. Andren, Power spectral density approximations of longitudinal road profiles, International Journal of Vehicle
Design 40 (1) (2006) 2–14.
URL http://inderscience.metapress.com/content/fhtfujv9yadktayw/
[21] A. A. Butkunas, Power spectral density and ride evaluation, Tech. rep., SAE Technical Paper (1966).
16
[22] D. Styles, C. Dodds, Simulation of random environments for structural dynamics testing, Experimental Mechanics
16 (11) (1976) 416–424. doi:10.1007/BF02410349.
[23] D. Ammon, Problems in road surface modelling, Vehicle System Dynamics: International Journal of Vehicle
Mechanics and Mobility 20 (S1) (1992) 28–41. doi:10.1080/00423119208969386.
[24] K. Bogsjö, Coherence of road roughness in left and right wheel-path, Vehicle System Dynamics: International
Journal of Vehicle Mechanics and Mobility 46 (S1) (2008) 599–609. doi:10.1080/00423110802018289.
[25] K. Song, X. Chen, Y. Lin, Wheelbase filtering effect on vehicle ride dynamics, in: Proceedings of the FISITA
2012 World Automotive Congress, Springer, 2013, pp. 1183–1195. doi:10.1007/978-3-642-33738-3_22.
[26] G. De Filippis, D. Palmieri, L. Soria, L. Mangialardi, B. Peeters, H. Van der Auweraer, Operational modal
analysis of passenger cars: Effect of the correlation between front and rear inputs, in: Proceedings of the 11th
International Conference on Recent Advances in Structural Dynamics (RASD), Pisa, Italy, 1st–3rd July 2013.
[27] C. Dodds, J. Robson, The description of road surface roughness, Journal of Sound and Vibration 31 (2) (1973)
175–183. doi:10.1016/S0022-460X(73)80373-6.
[28] D. Newland, An introduction to random vibrations, spectral and wavelet analysis, Longman Scientific & Technical
(Harlow, Essex, England and New York), 1993.
[29] A. González, E. J. O’brien, Y.-Y. Li, K. Cashell, The use of vehicle acceleration measurements to estimate road
roughness, Vehicle System Dynamics 46 (6) (2008) 483–499.
[30] B. Peeters, G. De Roeck, Stochastic system identification for operational modal analysis: a review, Journal of
Dynamic Systems, Measurement, and Control 123 (4) (2001) 659–667.
[31] M. El-Kafafy, A. Elsawaf, B. Peeters, T. Vampola, P. Guillaume, Modal parameter identification using particle
swarm optimization, International Journal of Systems Applications, Engineering & Development 9 (2015) 103–
109.
[32] L. Sun, W. Kenis, W. Wang, Stochastic spatial excitation induced by a distributed contact
on homogenous gaussian random fields, Journal of Engineering Mechanics 132 (7) (2006) 714–722.
doi:10.1061/(ASCE)0733-9399(2006)132:7(714).
[33] ISO, Mechanical vibration - road surface profiles - reporting of measured data, iso 8608:1995 (e) (1995-09-01).
A. Mathematical derivation of Eqs. from (43) to (46)
Based on Eq. (10), we write the following equality
ψ Tn Gf r (d (τ ) ⊗ δ p (τ )) GTfr ψ m = ψ tT
K t δ p (τ ) − C t δ̇ p (τ ) K t + K t δ̇ p (τ ) − C t δ̈ p (τ ) C t ψ tm ,
n
(A1)
where ψ tn ∈ C1×Nt is a vector containing the modal components related to the wheels, that is, for the full-car model,
T
.
(A2)
ψ tn = ψ4n ψ5n ψ6n ψ7n
By substituting in Eq. (42), we get
Rqi qj (τ ) = Rd (τ ) ∗
2N X
2N
X
ψin ψjm
m
an mam
n=1 m=1
Nw
X
Hp (τ ) ∗
p=0
ψ tT
n
K t J ap
(τ ) −
C t J bp
!
t
b
c
(τ ) K t + K t J p (τ ) − C t J p (τ ) C t ψ m , (A3)
where the terms J ap (τ ), J bp (τ ) and J cp (τ ) ∈ RNt ×Nt are matrices that encompass convolution integrals involving Dirac
delta functions, located at different time-lag values (Eqs. (40) and (41)), and their derivatives. We rephrase the
convolution integrals referred to J ap (τ ) as
Nw
X
p=0
Hp (τ ) ∗J ap
(τ ) =
Nw
X
Hp (τ ) ∗
p=0
t+T
Z
Zt
δ p (τ ) eλn (t−ρ) eλm (t+τ −σ) dρdσ
−∞ −∞
E1 (τ )
0
= H0 (τ ) ∗
E2 (τ )
0
0
E1 (τ )
0
E2 (τ )
E3 (τ )
0
E1 (τ )
0
0
0
E1 (τ )
E3 (τ )
+ H1 (τ ) ∗
0
0
E1 (τ )
E2 (τ )
17
E1 (τ )
0
E2 (τ )
0
0
E3 (τ )
0
E1 (τ )
E3 (τ )
0
,
E1 (τ )
0
(A4)
where the functions E1 (τ ), E2 (τ ) and E3 (τ ) represent the explicit solutions of the convolution integrals, that is
eλm τ h (τ ) + e−λn τ h (τ )
,
λn + λm
(A5)
eλm (τ +τ1 ) h (τ + τ1 ) + e−λn (τ +τ1 ) h (−τ − τ1 )
,
λn + λm
(A6)
E1 (τ ) = −
E2 (τ ) = −
eλm (τ −τ1 ) h (τ − τ1 ) + e−λn (τ −τ1 ) h (−τ + τ1 )
.
λn + λm
For example, Eq. (A6) can be proved by considering the following resolution scheme
(A7)
E3 (τ ) = −
t+τ Zt
Z
E2 (τ ) =
δ (τ + τ1 ) eλn (t−ρ) eλm (t+τ −σ) dρdσ
−∞ −∞
=
t+τ
Z
e
t+τ
Z
eλm (t+τ −σ) eλn (t−σ+τ1 ) h (t − σ − τ1 ) dσ.
λm (t+τ −σ)
dσ
−∞
=
Zt
δ (σ − ρ + τ1 ) eλn (t−ρ) dρ
−∞
(A8)
−∞
Thus, by introducing the change of variable
σ̄ = σ + τ1 − t
(A9)
the last integral in Eq. (A8) can be solved as
E2 (τ ) = e
λm (τ +τ1 )
τZ+τ1
e−(λn +λm )σ̄ h (−σ̄) dσ̄
−∞
=
τZ+τ1
e
−(λn +λm )σ̄
−∞
=
τZ+τ1
τ +τ1
e−(λn +λm )σ̄
e−(λn +λm )σ̄
h (−σ̄)
δ (−σ̄) dσ̄
−
h (−σ̄) dσ̄ =
− (λn + λm )
− (λn + λm )
−∞
e−(λn +λm )(τ +τ1 ) h (−τ − τ1 )
+
− (λn + λm )
−∞
τZ+τ1
−∞
e−(λn +λm )σ̄
δ (−σ̄) (−dσ̄)
− (λn + λm )
e−(λn +λm )(τ +τ1 ) h (−τ − τ1 ) + h (τ + τ1 )
.
=
− (λn + λm )
(A10)
Moving from Eq. (A4), we derive the expressions of the convolution integrals referred to J bp (τ ) and J cp (τ ), that are
Nw
X
Hp (τ ) ∗J bp
(τ ) =
Hp (τ ) ∗
p=0
p=0
Nw
X
Nw
X
Hp (τ ) ∗J cp
(τ ) =
Nw
X
p=0
p=0
t+T
Z
Zt
δ̇ p (τ ) eλn (t−ρ) eλm (t+τ −σ) dρdσ = −λn
Hp (τ ) ∗
Zt
Hp (τ ) ∗J ap (τ ) ,
(A11)
p=0
−∞ −∞
t+T
Z
Nw
X
δ̈ p (τ ) eλn (t−ρ) eλm (t+τ −σ) dρdσ = λ2n
Nw
X
Hp (τ ) ∗J ap (τ ),
(A12)
p=0
−∞ −∞
where the solutions of these convolution integrals can be retrieved from that of Eq. (A4) by using the properties of
Dirac delta function derivatives.
Eqs. (A4), (A11) and (A12) lead to the following equality
Nw
X
p=0
K t J ap (τ ) − C t J bp (τ ) K t + K t J bp (τ ) − C t J cp (τ ) C t ψ tm =
Hp (τ ) ∗ ψ tT
n
= H0 (τ ) ∗
2
− c2rt λ2n (ψ6m ψ6n + ψ7m ψ7n ) E1 (τ ) +
kf2 t − c2f t λ2n (ψ4m ψ4n + ψ5m ψ5n ) + krt
+ ((kf t − cf t λn ) (krt + crt λn ) (ψ4m ψ6n + ψ5m ψ7n )) E2 (τ ) +
= + ((kf t + cf t λn ) (krt − crt λn ) (ψ6m ψ4n + ψ7m ψ5n )) E3 (τ )) +
2
− c2rt λ2n (ψ6m ψ7n + ψ7m ψ6n ) E1 (τ ) +
+ H1 (τ ) ∗
kf2 t − c2f t λ2n (ψ4m ψ5n + ψ5m ψ4n ) + krt
+ ((kf t − cf t λn ) (krt + crt λn ) (ψ4m ψ7n + ψ5m ψ6n )) E2 (τ ) +
+ ((kf t + cf t λn ) (krt − crt λn ) (ψ6m ψ5n + ψ7m ψ4n )) E3 (τ )) .
18
(A13)
By substituting this expression in Eq. (42), we observe that the generic cross-correlation function Rqi qj (τ ) can be
rewritten as
!
NL
Nw
2N
X
X
X
p
pl +λn τl
pl −λn τl
Rqi qj (τ ) = Rd (τ ) ∗
αin h (τ ) +
βin e
h (τ + τl ) + γin e
h (τ − τl ) e+λn τ ψjn +
Hp (τ ) ∗
p=0
+
2N
X
ᾱpjn h (−τ )
n=1
+
n=1
NL
X
l=1
pl +λn τl
β̄jn
e
h (−τ
+ τl ) +
pl −λn τl
γ̄jn
e
h (−τ
!
− τl ) e
l=1
−λn τ
ψin
!!
,
(A14)
pl
pl
pl
pl
pl
where the coefficients αpl
in , βin , γin , ᾱin , β̄in , γ̄in play the same role as the operational reference factors φin in Eq. (3).
The number of coefficients in Eq. (A14) can be significantly decreased by recalling that
Rqi qj (τ ) = Rqj qi (−τ )
pl
αpl
in = ᾱin
⇔
pl
pl
βin
= β̄in
pl
pl
= γ̄in
.
γin
(A15)
Thus, for the full-car model we have
2
2N ψ
c2f t λ2m − kf2 t (ψ4m ψ4n + ψ5m ψ5n ) + c2rt λ2m − krt
(ψ6m ψ6n + ψ7m ψ7n )
X
im
,
α0in =
mAn mAm (λn + λm )
m=1
01
βin
=
2N
X
ψim (cf t λm − kf t ) (crt λm + krt ) (ψ6m ψ4n + ψ7m ψ5n )
,
mAn mAm (λn + λm )
m=1
(A17)
2N
X
ψim (cf t λm + kf t ) (crt λm − krt ) (ψ4m ψ6n + ψ5m ψ7n )
,
mAn mAm (λn + λm )
m=1
2
2N ψ
c2f t λ2m − kf2 t (ψ4m ψ5n + ψ5m ψ4n ) + c2rt λ2m − krt
(ψ6m ψ7n + ψ7m ψ6n )
X
im
01
γin
=
α1in =
mAn mAm (λn + λm )
m=1
11
βin
=
11
γin
=
(A16)
2N
X
ψim (cf t λm − kf t ) (crt λm + krt ) (ψ6m ψ5n + ψ7m ψ4n )
,
mAn mAm (λn + λm )
m=1
(A18)
,
(A19)
(A20)
2N
X
ψim (cf t λm + kf t ) (crt λm − krt ) (ψ4m ψ7n + ψ5m ψ6n )
.
mAn mAm (λn + λm )
m=1
(A21)
Finally, rearranging Eq. (A14) in a more suitable form we obtain
Rqi qj (τ ) =
2N
X
ϕ̄in (τ ) ψjn e+λn τ + ψin ϕ̄jn (−τ ) e−λn τ ,
(A22)
n=1
where ϕ̄in (τ ) is a lag-dependent operational reference factor, defined as
ϕ̄in (τ ) = Rd (τ ) ∗
Nw
X
p=0
Hp (τ ) ∗
αpin h (τ )
+
NL
X
l=1
19
pl +λn τl
βin
e
h (τ
+ τl ) +
pl −λn τl
γin
e
h (τ
!
− τl ) .
(A23)
| 3cs.SY
|
arXiv:1307.7810v1 [q-bio.QM] 30 Jul 2013
Accurate Decoding of Pooled Sequenced Data
Using Compressed Sensing
Denisa Duma1 , Mary Wootters2 , Anna C. Gilbert2 ,
Hung Q. Ngo3 , Atri Rudra3 , Matthew Alpert1 ,
Timothy J. Close4 , Gianfranco Ciardo1 , and Stefano Lonardi1
1
Dept. of Computer Science and Eng., University of California, Riverside, CA 92521
2
Dept. of Mathematics, The University of Michigan, Ann Arbor, MI 48109
3
Dept. of Computer Science and Eng., Univ. at Buffalo, SUNY, Buffalo, NY 14260
4
Dept. of Botany & Plant Sciences, University of California, Riverside, CA 92521
Abstract. In order to overcome the limitations imposed by DNA barcoding when multiplexing a large number of samples in the current generation of high-throughput sequencing instruments, we have recently proposed a new protocol that leverages advances in combinatorial pooling
design (group testing) [9]. We have also demonstrated how this new protocol would enable de novo selective sequencing and assembly of large,
highly-repetitive genomes. Here we address the problem of decoding
pooled sequenced data obtained from such a protocol. Our algorithm
employs a synergistic combination of ideas from compressed sensing and
the decoding of error-correcting codes. Experimental results on synthetic
data for the rice genome and real data for the barley genome show that
our novel decoding algorithm enables significantly higher quality assemblies than the previous approach.
Keywords: second/next-generation sequencing, pooled sequencing, compressed sensing, error-correcting codes
1
Introduction
The second generation of DNA sequencing instruments offer unprecedented throughput and extremely low cost per base, but read lengths are much shorter compared to Sanger sequencing. An additional limitation is the small number of
distinct samples that these instruments can accommodate (e.g., two sets of eight
lanes on the Illumina HiSeq). When the sequencing task involves a large number
of individual samples, a common solution is to employ DNA barcoding to “multiplex” samples within a single lane. DNA barcoding, however, does not scale
readily to thousands of samples. As the number of samples reaches the hundreds, exhaustive DNA barcoding becomes time consuming, error-prone, and
expensive. Additionally, the resulting distribution of reads for each barcoded
sample can be severely skewed (see, e.g., [1]).
Combinatorial pooling design or group testing allows one to achieve multiplexing without exhaustive barcoding. In group testing, a design or scheme is
a set of tests (or pools) each of which is a subset of a large collection of items
that needs to be tested for the presence of (a few) ‘defective’ items. The result
of testing a pool is a boolean value indicating whether the pool contains at least
one defective. The goal of group testing is to decode the information obtained by
testing all the pools in order to determine the precise identity of the defectives,
despite the fact that the defectives and non-defectives are mixed together. The
challenge is to achieve this goal while, at the same time, minimizing the number
of pools needed. Recently, compressed sensing (CS) has emerged as a powerful
technique for solving the decoding problem when the results of testing the pools
are more than boolean outcomes, for instance, real or complex values.
Combinatorial pooling has been used previously in the context of genome
analysis (see, e.g., [5,6,7,2,12]), but not for de novo genome sequencing. Our proposed pooling method for genome sequencing and assembly was first described
in [9] and has generated considerable attention. It was used to produce one of
the critical datasets for the first draft sequence of the barley genome [14]. In our
sequencing protocol, thousands of BAC clones are pooled according to a combinatorial design so that, at the outset of sequencing, one can ‘decode’ each read to
its source BACs. The underlying idea is to encode the identity of a BAC within
the pooling pattern rather than by its association with a specific DNA barcode.
We should stress that combinatorial pooling is not necessarily an alternative to
DNA barcoding, and both methods have advantages and disadvantages. They
can be used together to increase the number of samples that can be handled and
benefit from the advantages of both.
In this paper we address the problem of decoding pooled sequenced data
obtained from a protocol such as the one in [9]. While the main objective is to
achieve the highest possible accuracy in assigning a read to the correct BAC,
given that one sequencing run can generate hundreds of millions of reads, the decoding procedure has also to be time- and space-efficient. Since in [9] we pooled
BAC clones according to the Shifted Transversal Design [15] which is a ReedSolomon based pooling design, our proposed decoding approach combines ideas
from the fields of compressive sensing and decoding of error-correcting codes.
Specifically, given the result of ‘testing’ (in this case, sequencing) pools of genomic BAC clones, we aggregate read frequency information across the pools
and cast the problem as a compressed sensing problem where the unknowns are
the BAC assignments of the reads. We solve (decode) for the unknown assignments using a list recovery strategy as used in the decoding of error-correcting
codes. Reed-Solomon codes are known to be good list-recoverable codes which
can also tolerate a large fraction of errors. We also show that using readily available information about the reads like overlap and mate pair information can
improve the accuracy of the decoding. Experimental results on synthetic reads
from the rice genome as well as real sequencing reads from the barley genome
show that the decoding accuracy of our new method is almost identical to that of
HashFilter [9]. However, when the assembly quality of individual BAC clones
is the metric of choice, the decoding accuracy of the method proposed here is
significantly better than HashFilter.
2
Related Work
The resemblance between our work and the closest related research efforts using combinatorial pooling and compressed sensing ideas stops at the pooling
of sequencing data. Our application domain, pooling scheme employed and algorithmic approach to decoding, are completely different. To the best of our
knowledge, all compressed sensing work in the domain of genomics deals with
the problem of genotyping large population samples, whereas our work deals
with de novo genome sequencing. For instance in [5], the authors employ a pooling scheme based on the Chinese Remainder Theorem (CRT) to identify carriers
of rare alleles in large cohorts of individuals. The pooling scheme allows the
detection of mutants within a pool, and by combining information across pools
one is able to determine the identity of carriers. In true group testing style, the
unknown carrier identities are encoded by a boolean vector of length equal to
the number of individuals, where a value of one indicates a carrier and zero a
normal individual. To decode their pooling scheme and find the unknown vector, the authors devise a greedy decoding method called Minimum Discrepancy
Decoder. In [6], loopy belief propagation decoding is used for the same pooling
scheme. A similar application domain is described in [12], where the authors
identify carriers of rare SNPs in a group of individuals pooled with a random
pooling scheme (Bernoulli matrix) and use the Gradient Projection for Sparse
Reconstruction (GPSR) algorithm to decode the pooling scheme and recover the
unknown carrier identities. The same problem is tackled in [11] with a pooling
design inspired from the theory of error correcting codes. However, this design is
only able to identify a single rare-allele carrier within a group. In [2], the authors
organize domain-specific (linear) constraints into a compressed sensing matrix
which they use together with GPSR decoding to determine the frequency of each
bacterial species present in a metagenomic mixture.
3
Preliminaries
As mentioned in the introduction, in [9] we pool DNA samples (BAC clones)
according to a combinatorial pooling scheme, then sequence the pools using
high-throughput sequencing instruments. In this paper we show how to efficiently
recover the sequence content of each BAC by combining ideas from the theory
of sparse signal recovery or compressed sensing (CS) as well as from the large
body of work developed for the decoding of error-correcting codes.
Formally, a combinatorial pooling design (or pooling scheme) can be represented by a binary matrix Φ with m rows (corresponding to pools) and n
columns (corresponding to items to be pooled), where entry (i, j) is 1 if item j
is present in pool i, 0 otherwise. The matrix Φ is called the design matrix, sensing matrix or measurement matrix by various authors in the literature. In this
paper we only use the first two names to designate Φ. An important property
of a combinatorial pooling design is its decodability d (also called disjunctness),
which is the maximum number of ‘defectives’ it guarantees to reliably identify.
Let w be a subset of the columns (pooled variables) of the design matrix Φ
and p(w) be the set of rows (pools) that contain at least one variable in w: the
matrix Φ is said to be d-decodable (d-disjunct) if for any choice of w1 and w2
with |w1 | = 1, |w2 | = d and w1 6⊂ w2 , we have that p(w1 ) 6⊆ p(w2 ).
In this paper, we pool BACs using the combinatorial pooling scheme called
Shifted Transversal Design (STD) [15]. STD is a layered design, i.e., the rows of
the design matrix are organized into multiple redundant layers such that each
pooled variable appears only once in each layer, that is, a layer is a partition
of the set of variables. STD is defined by parameters (q, L, Γ ) where L is the
number of layers, q is a prime number equal to the number of pools (rows) in
each layer and Γ is the compression level of the design. Thus, in order to pool
n variables, STD uses a total of m = q × L pools. The set of L pools defines a
unique pooling pattern for each variable which can be used to retrieve its identity.
This set of L integers is called the signature of the variable. The compression
level Γ is defined to be the smallest integer such that q Γ +1 ≥ n. STD has the
desirable property that any two variables co-occur in at most Γ pools, therefore
by choosing a small value for Γ one can make STD pooling extremely robust to
errors. The parameter Γ is also related to the decodability of the design through
the equation d = ⌊(L − 1)/Γ ⌋. Therefore, Γ can be seen as a trade-off parameter:
the larger it is, the more items can be tested (up to q Γ +1 ), but fewer defectives
can be reliably identified (up to ⌊(L − 1)/Γ ⌋). For more details on the pooling
scheme and its properties please refer to [15].
In order to decode measurements obtained through STD (i.e., reconstruct
the sequence content of pooled BACs) we borrow ideas from compressed sensing
(CS), an area of signal processing that describes conditions and efficient methods
for capturing sparse signals from a small number of aggregated measurements
[6]. Unlike combinatorial group testing, in compressed sensing measurements can
be more general than boolean values, allowing recovery of hidden variables which
are real or complex-valued. Specifically, in CS we look for an unknown vector
or signal x = (x1 , x2 , . . . , xn ) which is s-sparse, i.e., has at most s non-zero
entries. We are given a vector y = (y1 , y2 , . . . , ym ) of measurements (m ≪ n),
which is the product between the (known) design matrix Φ and the unknown
vector x, that is y = Φx. Under certain conditions on Φ, by using the measurements y, the assumption on the sparsity of x and information encoded by Φ,
it is possible to recover the original sparse vector x. The latter equation corresponds to the ideal case when the data is noise-free. In practice, if the signal
x is not as sparse as needed and if measurements are corrupted by noise, the
equation becomes y = Φx + ǫ. In CS theory there are two main approaches
for solving the latter equation, namely linear programming (LP) decoding and
greedy pursuit decoding. Greedy pursuit algorithms have faster decoding time
than LP-based approaches, frequently sub-linear in the length of x (although
for specially designed matrices). Their main disadvantages is that they usually
require a slightly larger number of measurements and do not offer the same uniformity and stability guarantees as LP decoding. Greedy pursuits are iterative
algorithms which proceed in a series of steps: (1) identify the locations of the
largest coefficients of x by greedy selection, (2) estimate their values, (3) update
y by subtracting the contribution of estimated values from it, and iterate (1-3)
until some convergence criterion is met. Usually O(s) iterations, where s is the
sparsity of x, suffice [17]. Updating y amounts to solving a least squares problem
in each iteration.
The most well known greedy decoding algorithm is Orthogonal Matching
Pursuit (OMP) [16], which has spawned many variations. In OMP, the greedy
rule selects in each iteration the largest coordinate of ΦT y, i.e., the column of Φ
which is the most correlated with y. In this paper, we are interested in a variant of
OMP called Simultaneous Orthogonal Matching Pursuit (S-OMP). S-OMP is different from OMP in that it approximates multiple sparse signals x1 , x2 , . . . , xK
simultaneously by using multiple linear combinations, y1 , y2 , . . . , yK , of the sensing matrix Φ [17]. The unknown signals {xk }k∈{1,··· ,K} as well as measurement vectors {yk }k∈{1,··· ,K} can be represented by matrices X ∈ Rn×K and
Y ∈ Rm×K . Intuitively, by jointly exploiting information provided by Y, SOMP is able to achieve better approximation error especially when the signals
to be approximated are corrupted by noise which is not statistically independent
[17].
The mapping from the CS setting into our problem follows naturally and
we give here a simplified and intuitive version of it. The detailed model will
be introduced in the next section. The variables to be pooled are BAC clones.
Each column of the design matrix corresponds to a BAC to be pooled and
each row corresponds to a pool. For each read r (to be decoded) there is an
unknown s-sparse vector x which represents at most s BACs which could have
generated r. The vector of measurements y (frequency vector ) of length m gives
for each read r, the number of times r appears in each of the m pools. The use
of numerical measurements (read counts) rather than boolean values indicating
the presence or the absence of r from a pool is in accordance with CS theory
and offers additional valuable information for decoding. To carry out the latter,
we use a S-OMP style algorithm but replace the greedy selection rule by a
list recovery criterion. Briefly, we obtain a list of candidate BACs for read r
as those columns of Φ whose non-zero coordinates consistently correspond to
the heaviest-magnitude measurements in each layer of y [10]. This allows for a
finer-grained usage of the values of y on a layer-by-layer basis rather than as a
whole. Additionally, by requiring that the condition holds for at least l layers
with l ≤ L, one can make the algorithm more robust to the noise in vector y.
4
Decoding Algorithms
In this section we present our decoding algorithms that assign reads back to the
BACs from which they were derived. Recall that we have n BACs pooled into m
pools according to STD and each BAC is pooled in exactly L pools. The input
data to the decoding algorithm consists of (1) m datasets containing the reads
obtained from sequencing the m pools, and (2) the parameters of the pooling
design, including the signatures of all n BACs. We will assume that each read
r may originate from up to s BACs with s ≪ n; ideally, we can make the same
assumption for each k-mer (a k-mer is a substring of r of length k) of r, provided
that k is ‘large enough’. In practice, this will not be true for all k-mers (e.g.,
some k-mers are highly repetitive), and we will address this issue later in this
document.
We start by preprocessing the reads to correct sequencing errors in order
to improve the accuracy of read decoding. For this task, we employ SGA [13],
which internally employs a k-mer based error correction strategy. An additional
benefit of error correction is that it reduces the total number of distinct k-mers
present in the set of reads. After the application of SGA, there still remains a
small proportion of erroneous k-mers, which we discard because they will likely
introduce noise in the decoding process. An advantage of pooled sequencing is
that erroneous k-mers are easy to identify because they appear in fewer than
L pools. To be conservative, we only discard k-mers appearing in fewer than γ
pools where γ ≤ L is a user-defined parameter (see Section 5.1 for details on the
choice of this parameter). The closer γ is to L the more likely it is that a k-mer
that appears in γ pools is correct, but missing from the remaining L − γ pools
due to sequencing errors. Henceforth, we will call a k-mer valid if it appears in
a number of pools in the range [γ, sL] where s is the sparsity parameter. Any
k-mer occurring in more than sL pools is considered highly repetitive, and will
likely not be useful in the decoding process. The decoding algorithm we employ
can safely ignore these repetitive k-mers.
To carry out the decoding, we first compute the frequencies of all the k-mers
in all the m pools. Specifically, we decompose all SGA-corrected reads into kmers by sliding a window of length k (there are |r| − k + 1 such windows for
each read r). For each distinct k-mer, we count the number of times it appears
in each of the m pools, and store the sequence of the k-mer along with its vector
of m counts into a hash table. We refer to the vector of counts of a k-mer as its
frequency vector.
We are now ready to apply our CS-style decoding algorithm. We are given
a large number of reads divided into m sets (pools). For each read r, we want
to determine which of the n BACs is the source. Since we decomposed r into its
constitutive k-mers, we can represent the pool counts of all its k-mers by a frequency matrix Yr . Matrix Yr is a non-negative integer matrix where the number
of columns is equal to the number Kr of k-mers in r, the number of rows is equal
to the numbers m of pools, and entry (i, j) reports the number of times the j th
k-mer of r appears in pool i. The input to the decoding algorithm for read r is
given by (1) the frequency matrix Yr , (2) the design matrix Φ ∈ {0, 1}m×n, and
(3) the maximum number s of BACs which could have generated r. To decode r
means to find a matrix Xr ∈ Zn×Kr such that Xr = argminX ||ΦX − Yr ||2 with
the constrain that Xr is row-sparse, i.e., it has at most s non-zero rows (one for
each source BAC).
Since finding the source BACs for a read is sufficient for our purposes, we can
reduce the problem of finding matrix X to the problem of finding its row support
S(X), which is the union of the supports of its columns. The support Supp(X:,j )
of a column j of X is the set of indices i such that Xi,j 6= 0. In our case, the
non-zero indices represent the set of BACs which generated the read (and by
transitivity its constitutive k-mers). Since this set has cardinality at most s, in
the ideal case, X is row-sparse with support size at most s. In practice, the same
k-mer can be shared by multiple reads and therefore the number of non-zero
indices can differ from s. By taking a conservative approach, we search for a
good s-sparse approximation of S(X), whose quality we evaluate according to
the following definition.
Definition: A non-empty set S is good for X if for any column j of X, we
have S ⊂ Supp(X:,j ).
Our decoding Algorithm 1 finds S in two steps, namely Filter and Estimate, which are explained next.
Filter (Algorithm 2) is a one-iteration S-OMP style algorithm in which
multiple candidate BACs are selected (we tried performing multiple iterations
without significant improvement in the results). Whereas S-OMP selects one
BAC per iteration as the column of Φ most correlated (inner product) with all
the columns of Y, our algorithm employs a list recovery criterion to obtain an
approximation X̃r of Xr . Specifically, for each column y of Yr and for each layer
l ∈ [1, . . . , L], we select a set Sl of candidate pools for that layer as follows. We
choose set Sl by considering the h highest-magnitude coordinates of y in layer
l and selecting the corresponding pools. BACs whose signature pools belong to
all L sets Sl are kept while the rest of them are removed, i.e., their X̃-entries are
set to zero. Finally, for the BACs that are not filtered out, the X̃-entry estimate
follows the min-count estimate. The value of h should be chosen to be Θ(s):
h = 3s is sufficient even for noisy data [10].
Next, the Estimate (Algorithm 3) algorithm determines Sr by computing
a score for each BAC. Based on the computed scores, we select and return the
top s BACs as the final support Sr of Xr . Read r is then assigned to all the
BACs in Sr . The scoring function we employ for each BAC b is the number of
k-mers “voting” for b, i.e., having a frequency of at least τ in each pool in the
signature of b. The value we used for τ is given in Section 5. If we consider the
rows of X̃r as vectors of length Kr , our scoring function is simply the l0 norm
of these vectors, after zeroing out all the entries smaller than τ . We also tried
Algorithm 1: FindSupport (Φ, Yr , h, s)
Input : Φ ∈ {0, 1}m×n , Yr ∈ Nm×Kr and sparsity s such that
Xr = argminX ||ΦX − Yr ||2 for a s-row-sparse matrix Xr ∈ Nn×Kr ;
h ≤ q the number of entries per layer considered by list recovery
Output: A non-empty set Sr with |Sr | ≤ s which is good for Xr
1
2
3
X̃r ← Filter(Φ, Yr , h)
Sr ← Estimate(X̃r , s)
return Sr
Algorithm 2: Filter(Φ, Yr , h)
Input : Φ ∈ {0, 1}m×n , Yr ∈ Nm×Kr , parameter h
Output: An approximation X̃r for Xr
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
// Recall that Φ has L layers with q pools each
// For a column y of Yr , denote by y[l]i the ith entry in layer l
X̃r ← 0
for k = 1, . . . , Kr do
Let y = Yr:,k be the kth column of Yr
for l = 1, . . . , L do
Sl ← set of h indices i ∈ {1, . . . , q} such that the corresponding counts
y[l]i are the h heaviest-magnitude counts in layer l of column y
for b = 1, . . . , n do
layersMatched ← 0
Let φ = Φ:,b be the bth column of Φ
for l = 1, . . . , L do
if the unique i such that φ[l]i = 1 belongs to Sl then
layersMatched ← layersMatched + 1
if layersMatched = L then
X̃b,k ← minφp =1 {yp }
l1 and l2 norms without observing significant improvements in the accuracy of
read assignments.
Observe that algorithms FindSupport, Filter and Estimate process one
read at a time. Since there is no dependency between the reads, processing
multiple reads in parallel is trivial. However, better total running time, improved
decoding accuracy as well as a smaller number of non-decodable reads can be
achieved by jointly decoding multiple reads at once. The idea is to use additional
sources of information about the reads, namely (1) read overlaps and (2) matepair information. For the former, if we can determine clusters of reads that are
mutually overlapping, we can then decode all the reads within a cluster as a
single unit. Not only this strategy increases the decoding speed, but it also has
Algorithm 3: Estimate(X̃r , s)
Input : X̃r , sparsity parameter s
Output: Support set Sr , with |Sr | ≤ s
1
2
3
4
for b = 1, . . . , n do
score(b) ← |{k : X̃b,k ≥ τ }|
Sr ← set of indices b with the highest s scores
return Sr
the potential to improve the accuracy of read assignments because while some of
the reads in the cluster might have sequencing errors, the others might be able
to ‘compensate’. Thus, we can have more confidence in the vote of high-quality
shared k-mers. There is, however, the possibility that overlaps are misleading.
For instance, overlaps between repetitive reads might lead one to assign them to
the same cluster while in reality these reads belong to different BACs. To reduce
the impact of this issue we allow any read that belongs to multiple clusters to be
decoded multiple times and take the intersection of the multiple assignments as
the final assignment for the read. If a read does not overlap any other read (which
could be explained due to the presence of several sequencing errors) we revert
to the single read decoding strategy. In order to build the clusters we compute
all pairwise read overlaps using SGA [13], whose parameters are discussed in
Section 5.
In order to apply FindSupport on a cluster c of reads, we need to gather
the frequency matrix Yc for c. Since the total number of k-mers within a cluster
can be quite large as the clusters themselves can be quite large, and each kmer can be shared by a subset of the reads in the cluster, we build Yc on the
most frequently shared valid k-mers in the cluster. Our experiments indicate
that retaining a number of k-mers equal to the numbers of k-mers used in the
decoding of individual reads is sufficient. When reads within a cluster do not
share a sufficient number of valid k-mers, we break the cluster into singletons
and decode its reads individually. We denote by µ the minimum number of valid
k-mers required to attempt decoding of both clusters and individual reads. The
choice of this parameter is also discussed in Section 5.
We can also use mate pair information to improve the decoding, if reads are
sequenced as paired-ends (PE). The mate resolution strategy (MRS) we employ
is straightforward. Given a PE read r, (1) if the assignment of one of the mates
of r is empty, we assign r to the BACs of the non-empty mate; (2) if both
mates of r have BAC assignments and the intersection of these assignments is
non-empty, we assign r to the BACs in the intersection; (3) if both mates of r
have BAC assignments and their intersection is empty, we discard both mates.
In what follows, we will use RBD to refer to the read based-decoding and CBD
to refer to the cluster-based decoding versions of our algorithm. CBD with MRS
is summarized in Algorithm 4.
5
Experimental Results
While our algorithms can be used to decode any set of DNA samples pooled
according to STD, in this paper, we evaluate their performance on sets of BAC
clones selected in such a way that they cover the genome (or a portion thereof)
with minimum redundancy. In other words, the BACs we use form a minimum
tiling path (MTP) of the genome. The construction of a MTP for a given genome
requires a physical map, but both are well-known procedures and we will not
discuss them here (see, e.g., [4] and references therein). Once the set of MTP
BAC clones has been identified, we (1) pool them according to STD, (2) sequence
Algorithm 4: ClusterFindSupport(Φ, C, {Yc }c∈C , h, s)
Input : Φ ∈ {0, 1}m×n , parameter h, sparsity parameter s, set C of all
clusters, frequency matrix Yc for each cluster c ∈ C
Output: A support set Sr with |Sr | ≤ s for each read r
1
2
3
4
5
6
7
8
9
10
11
12
13
14
for each cluster c ∈ C do
Sc ←FindSupport(Φ, Yc , h, s)
for each read r ∈ c do
if Sr = ∅ then Sr ← Sc
else Sr ← Sr ∩ Sc // Take intersection of all assignments to r
// MRS
for each PE read (r1 , r2 ) do
if Sr1 = ∅ then Sr1 ← Sr2
if Sr2 = ∅ then Sr2 ← Sr1
if Sr1 6= ∅ and Sr2 6= ∅ then
Sr1 ,r2 ← Sr1 ∩ Sr2
if Sr1 ,r2 6= ∅ then
Sr1 ← Sr1 ,r2
Sr2 ← Sr1 ,r2
the resulting pools, (3) apply our decoding algorithm to assign reads back to their
source BACs. Step (3) makes it possible to assemble reads BAC-by-BAC, thus
simplifying the genome assembly problem and increasing the accuracy of the
resulting BAC assemblies [9].
Recall that CS decoding requires the unknown assignment vector x to be ssparse. Since we use MTP BAC clones, if the MTP was truly a set of minimally
overlapping clones, setting s equal to 2 would be sufficient; we set it equal to
3 instead to account for imperfections in the construction of the MTP and to
obtain additional protection against errors. Figure 1 illustrates the three cases
(read belongs to one BAC, two BACs or three BACs) we will be dealing with
during decoding, and how it affects our STD parameter choice.
Next, we present experimental evaluations where we pool BAC clones using the following STD parameters. Taking into consideration the need for a
3-decodable pooling design for MTP BACs, we choose parameters q = 13, L = 7
and Γ = 2, so that m = qL = 91, n = q Γ +1 = 2197 and d = ⌊(L − 1)/Γ ⌋ = 3.
Fig. 1. The three cases we are dealing with during read decoding.
In words, we pool 2197 BACs in 91 pools distributed in 7 layers of 13 pools
each. Each BAC is pooled in exactly 7 pools and each pool contains q Γ = 169
BACs. Recall that we call the set of L pools to which a BAC is assigned the
BAC signature. In the case of STD, any two-BAC signatures can share at most
Γ = 2 pools and any three-BAC signatures can share at most 3Γ = 6 pools.
5.1
Simulation results on the rice genome
To simulate our combinatorial pooling protocol and subsequent decoding, we
used the genome of rice (Oryza sativa) which is about 390 Mb and fully sequenced. We started from an MTP of 3,827 BAC clones selected from a real
physical map library for rice of 22,474 clones. The average BAC length in the
MTP was ≈ 150kB. Overall the clones in the MTP spanned 91% of the rice
genome. We pooled a subset of 2,197 of these BACs into 91 pools according to
the pooling parameters defined above. The resulting pools were ‘sequenced’ in
silico using SimSeq, which is a high-quality short read simulator used to generate the synthetic data for Assemblathon [3]. SimSeq uses error profiles derived
from real Illumina data to inject “realistic” substitution errors. For each pool,
we generated 106 PE reads of 100 bases each with an average insert size of 300
bases. A total of 200M usable bases gave an expected ≈ 8× sequencing depth for
a BAC in a pool. As each BAC is present in 7 pools, this is an expected ≈ 56×
combined coverage before decoding. After decoding however, since a read can
be assigned to more than one BAC, the actual average BAC sequencing depth
became 91.68× for RBD, 93× for CBD and 97.91× for CBD with MRS.
To simulate our current workflow, we first performed error-correction on the
synthetic reads using SGA [13] with k-mer size parameter k = 26. Then, the
hash table for k = 26 was built on the corrected reads, but we only stored k-mers
appearing in at least γ = 3 pools. Due to the error-correction preprocessing step
and the fact that we are discarding k-mers with low pool count, the hash table
was relatively small (about 30GB).
In order to objectively evaluate and compare the performance of our decoding
algorithms, we first had to precisely define the ‘ground truth’ for simulated
reads. An easy choice would have been to consider ‘true’ only the single BAC
from which each read was generated. However, this notion of ground truth is
not satisfactory: for instance, since we can have two or three BACs overlapping
each other in the MTP, reads originating from an overlap region are expected
to be assigned to all the BACs involved. In order to find all the BACs that
contain a read, we mapped all synthetic reads (error-free version) against the
BAC primary sequences using Bowtie [8] in stringent mode (paired-end end-toend alignment with zero mismatches). The top three paired-end hits returned by
Bowtie constituted the ground truth against which we validated the accuracy
of the decoding.
In our experiments we observed that although the majority of the reads
are assigned to 1–3 BACs, due to the repetitive nature of the genome, a small
fraction (≈ 1%) can be correctly assigned to more than 3 BACs. To account
for this, rather than sorting BAC scores and retaining the top 3, we decided to
Mapped to source BAC Precision
Perfect decoding
100.00%
Hashfilter [9]
RBD
CBD
CBD + MRS
99.48% 97.45%
98.05% 97.81%
97.23% 97.74%
96.60% 97.89%
Recall F-score Not decoded
98.11% 49.62% 65.90%
99.28%
97.46%
96.35%
95.58%
98.36%
97.64%
97.04%
96.72%
0.00%
16.25%
14.58%
12.58%
7.09%
Table 1. Accuracy of the decoding algorithms on synthetic reads for the rice genome
(see text for details). All values are an average of 91 pools. Boldface values highlight
the best result in each column (excluding perfect decoding).
assign a read to all BACs whose score was above a certain threshold. We found
that retaining all BACs whose score was at least 0.5Kr gave the best results.
Recall that the score function we are using is the l0 norm, so we are effectively
asking that at least half of the k-mers ‘vote’ for a BAC.
Table 1 summarizes and compares the decoding performance of our algorithms. The first row of the table reports the performance of an ‘ideal’ method
that always assigns each read to its original source BAC. The next four rows
summarize (1) the performance of HashFilter [9] with default parameters; (2)
our read-based decoding (RBD); (3) our cluster-based decoding (CBD); (4) our
cluster-based decoding with mate resolution strategy (CBD + MRS). For all
three versions of the decoding algorithm we used parameters h = ⌊q/2⌋ = 6 and
τ = 1.
To build clusters, we require a minimum overlap of 75 bases between two
reads and a maximum error rate of 0.01 (SGA parameters). The resulting clusters contained on average about 5 reads. Our methods make a decoding decision
if a read (or cluster) contains at least µ = 15 valid k-mers. The columns in
Table 1 report the percentage of reads assigned to the original source BAC, precision (defined as T P/(T P +F P ) where T P is the number of true positive BACs
across all decoded reads; F P and F N are computed similarly), recall (defined
as T P/(T P + F N )), F-score (harmonic mean of precision and recall) and the
percentage of reads that were not decoded. Observe that the highest precision
is achieved by the cluster-based decoding with MRS, and the highest recall is
obtained by Hashfilter. In general, all methods are comparable from the point
of view of decoding precision and recall. In terms of decoding time, once the hash
table is built (≈ 10h on one core), RBD takes on average 14.03s per 1M reads
and CBD takes on average 33.46s per 1M clusters. By comparison, Hashfilter
[9] takes about 30s per 1M reads. These measurements were done on 10 cores of
an Intel Xeon X5660 2.8 GHz server with 12 cores and 192 GB of RAM.
As a more meaningful measure of decoding performance, we assembled the
set of reads assigned by each method to each BAC. We carried out this step
using Velvet [18] for each of the 2,197 BACs, using a range of l-mer from 25
to 79 with an increment of 6, and chose the assembly that achieved the highest
Reads used # of contigs
Perfect decoding (ideal)
Hashfilter [9]
RBD
CBD
CBD + MRS
N50 Sum/size BAC coverage
97.1%
4 136,570
107.4
87.1%
95.0%
96.5%
97.3%
97.0%
24 52,938
20 46,477
22 53,097
11 103,049
93.8
90.0
93.8
97.0
76.2%
81.1%
84.7%
82.9%
Table 2. Assembly results for rice BACs for different decoding algorithms (see text
for details). All values are an average of 2197 BACs. Boldface values highlight the best
result in each column (excluding perfect decoding).
N505 . Table 2 reports the main statistics for the assemblies: percentage of reads
used by Velvet in the assembly, number of contigs (at least 200 bases long) of
the assembly, value of N50, ratio of the sum of all contigs sizes over BAC length,
and the coverage of the BAC primary sequence by the assembly. All reported
values are averages over 2,197 BACs. We observe that our decoding algorithms
lead to superior assemblies than Hashfilter’s. In particular, the N50 and the
average coverage of the original BACs are both very high, and compare favorably
with the statistics for the assembly of perfectly decoded reads.
The discrepancy between similar precision/recall figures but quite different
assembly statistics deserves a comment. First, we acknowledge that the way we
compute precision and recall by averaging T P , F P and F N across all decoded
reads might not be the best way of measuring the accuracy of the decoding.
Taking averages might not accurately reflect mis-assignments at the level of
individual reads. Second, our decoding algorithms makes a better use of the
k-mer frequency information than HashFilter, and, at the same time, takes
advantage of overlap and mate pair information, which is expected to result in
more reads decoded and more accurate assemblies.
5.2
Results on the barley genome
We have also collected experimental results on real sequencing data for the
genome of barley (Hordeum vulgare), which is about 5,300 Mb and at least
95% repetitive. We started from an MTP of about 15,000 BAC clones selected
from a subset of nearly 84,000 gene-enriched BACs for barley (see [9] for more
details). We divided the set of MTP BACs into seven sets of n = 2197 BACs and
pooled each set using the STD parameters defined above. In this manuscript,
we report on one of these seven sets, called HV3 (the average BAC length in
this set is about 116K bases). The 91 pools in HV3 were sequenced on one flow
cell of the Illumina HiSeq2000 by multiplexing 13 pools on each lane. After each
sample was demultiplexed, we quality-trimmed and cleaned the reads of spurious
sequencing adapters and vectors. We ended up with high quality reads of about
5
The N50 is the contig length such that at least half of the total bases of a genome
assembly are contained within contigs of this length or longer.
Reads used # contigs
Hashfilter [9]
RBD
CBD
CBD + MRS
83.6%
85.7%
92.9%
94.3%
N50 Sum/size # obs unigenes % coverage
101 8,190
54 14,419
54 13,482
50 26,842
96.7%
101.0%
94.5%
126.8%
1,433
1,434
1,436
1,434
92.9%
92.4%
92.6%
92.5%
Table 3. Assembly results for barley BACs for different decoding algorithms. All values
are an average of 2197 BACs. Boldface values highlight the best result in each column.
Column “% coverage” refers to the coverage of known unigenes by assembled contigs.
87–89 bases on average. The number of reads in a pool ranged from 4.2M to
10M, for a grand total of 826M reads. We error-corrected and overlap-clustered
the reads using SGA (same parameters as for rice). The average cluster size was
about 26 reads. Computing pairwise overlaps took an average of 217.60s per 1M
reads on 10 cores. The hash table for k = 26 (after discarding k-mers appearing
in fewer than γ = 3 pools) used about 26GB of RAM. After decoding the reads
to their BAC, we obtained an average sequencing depth for one BAC of 409.2×,
382.2× and 412.8× for RBD, CBD and CBD + MRS, respectively. The average
running time was 10.25s per 1M reads for RBD and 82.12s per 1M clusters for
CBD using 10 cores.
The only objective criterion to asses the decoding performance on barley
genome is to assemble the reads BAC-by-BAC and analyze the assembly statistics. We used Velvet with the same l-mer choices as used for rice. Table 3
summarizes the statistics for the highest N50 among those l-mer choices. As
before, rows corresponds to the various decoding methods. Columns show (1)
percentage of reads used by Velvet in the assembly, (2) number of contigs (at
least 200 bases long), (3) value of N50, (4) ratio of the sum of all contigs sizes
over estimated BAC length, (5) the number of barley known unigenes observed
in the assemblies, and (6) the coverage of observed unigenes. Observe that, out
of a total of 1,471 known unigenes expected to be contained in these BACs, a
large fraction are reported by all assemblies. However, cluster-based decoding
appears to generate significantly longer contigs than the other methods.
6
Conclusions
We have presented a novel modeling and decoding approach for pooled sequenced
reads obtained from protocols for de novo genome sequencing, like the one proposed in [9]. Our algorithm is based on the theory of compressed sensing and
uses ideas from the decoding of error-correcting codes. It also effectively exploits
overlap and mate pair information between the sequencing reads. Experimental results on synthetic data from the rice genome as well as real data from
the genome of barley show that our method enables significantly higher quality
assemblies than the previous approach, without incurring higher decoding times.
Acknowledgments
SL and TJC were supported by NSF [DBI-1062301 and DBI-0321756] and by
USDA [2009-65300-05645 and 2006-55606-16722]. MW and ACG were supported
by NSF [CCF-1161233]. HQN and AR were supported by NSF [CCF-1161196].
References
1. S. Alon, F. Vigneault, S. Eminaga, et al. Barcoding bias in high-throughput multiplex sequencing of mirna. Genome Research, 21(9):1506–1511, 2011.
2. A. Amir and O. Zuk. Bacterial community reconstruction using compressed sensing. In RECOMB, pages 1–15, 2011.
3. D. Earl and et al. Assemblathon 1: A competitive assessment of de novo short read
assembly methods. Genome Research, 21(12):2224–2241, Dec. 2011.
4. F. W. Engler, J. Hatfield, W. Nelson, and C. A. Soderlund. Locating sequence on
FPC maps and selecting a minimal tiling path. Genome Research, 13(9):2152–2163,
2003.
5. Y. Erlich, K. Chang, A. Gordon, et al. DNA sudoku – harnessing high-throughput
sequencing for multiplexed specimen analysis. Genome Research, 19(7):1243–1253,
2009.
6. Y. Erlich, A. Gordon, M. Brand, et al. Compressed genotyping. IEEE Transactions
on Information Theory, 56(2):706–723, Apr. 2010.
7. I. Hajirasouliha, F. Hormozdiari, S. C. Sahinalp, and I. Birol. Optimal pooling for
genome re-sequencing with ultra-high-throughput short-read technologies. Bioinformatics, 24(13):i32–i40, 2008.
8. B. Langmead, C. Trapnell, M. Pop, and S. L. Salzberg. Ultrafast and memoryefficient alignment of short DNA sequences to the human genome. Genome Biology,
10(3):R25, 2009.
9. S. Lonardi, D. Duma, M. Alpert, et al. Combinatorial pooling enables selective
sequencing of the barley gene space. PLoS Comput Biol, 9(4):e1003010, 04 2013.
10. H. Q. Ngo, E. Porat, and A. Rudra. Efficiently decodable compressed sensing by
list-recoverable codes and recursion. In STACS, pages 230–241, 2012.
11. S. Prabhu and I. Pe’er. Overlapping pools for high-throughput targeted resequencing. Genome Research, 19(7):1254–1261, 2009.
12. N. Shental, A. Amir, and O. Zuk. Identification of rare alleles and their carriers
using compressed se(que)nsing. Nucleic Acids Research, 38(19):e179–e179, 2010.
13. J. T. Simpson and R. Durbin. Efficient de novo assembly of large genomes using
compressed data structures. Genome Research, 22(3):549–556, 2012.
14. The International Barley Genome Sequencing Consortium. A physical, genetic
and functional sequence assembly of the barley genome. Nature, advance online
publication:in press, 10 2012.
15. N. Thierry-Mieg. A new pooling strategy for high-throughput screening: the shifted
transversal design. BMC Bioinformatics, 7(28), 2006.
16. J. A. Tropp and A. C. Gilbert. Signal recovery from random measurements via
orthogonal matching pursuit. IEEE Trans. Inform. Theory, 53:4655–4666, 2007.
17. J. A. Tropp, A. C. Gilbert, and M. J. Strauss. Algorithms for simultaneous sparse
approximation: part i: Greedy pursuit. Signal Process., 86(3):572–588, 2006.
18. D. Zerbino and E. Birney. Velvet: Algorithms for de novo short read assembly
using de Bruijn graphs. Genome Research, 8(5):821–9, 2008.
| 5cs.CE
|
A Review on Bilevel Optimization: From Classical
to Evolutionary Approaches and Applications
Abstract—Bilevel optimization is defined as a mathematical
program, where an optimization problem contains another optimization problem as a constraint. These problems have received significant attention from the mathematical programming
community. Only limited work exists on bilevel problems using
evolutionary computation techniques; however, recently there has
been an increasing interest due to the proliferation of practical
applications and the potential of evolutionary algorithms in
tackling these problems. This paper provides a comprehensive
review on bilevel optimization from the basic principles to
solution strategies; both classical and evolutionary. A number
of potential application problems are also discussed. To offer
the readers insights on the prominent developments in the field
of bilevel optimization, we have performed an automated textanalysis of an extended list of papers published on bilevel
optimization to date. This paper should motivate evolutionary
computation researchers to pay more attention to this practical
yet challenging area.
Index Terms—Bilevel optimization, Stackelberg Games, Evolutionary Algorithms.
I. I NTRODUCTION
Many large-scale optimization and decision-making processes faced by public and private organizations are hierarchical in the sense that the realized outcome of any solution or
decision taken by the upper level authority (leader) to optimize
their goals is affected by the response of lower level entities
(follower), who will seek to optimize their own outcomes.
Figure 1 illustrates a general bilevel problem solving structure
involving interlinked optimization and decision-making tasks
at both levels. The figure shows that for any given upper
Upper level decision (xu) space
f
Parameter for lower level problem
Upper level
decision vector
Lower Level Objective
arXiv:1705.06270v1 [math.OC] 17 May 2017
Ankur Sinha, Pekka Malo, Kalyanmoy Deb
xu space
( , ) : A feasible bilevel solution
for the upper level optimization problem
Lower level
parametric
optimization
Lower level decision (xl) space
Optimal lower
level response
Fig. 1: A general sketch of a bilevel problem.
level decision vector, there is a corresponding (parametric)
lower level optimization problem to be solved that provides
the rational (optimal) response of the follower for the leader’s
decision. The leader’s decision vector is represented by xu and
the follower’s decision vector is represented by xl . An (xu ,
x∗l ) pair where x∗l is an optimal response to xu represents
a feasible solution to the upper level optimization problem
provided that it also satisfies the constraints in the problem.
Each level has its own objectives and constraints. One aspect
of bilevel problems is that it is not symmetric in terms of two
levels. The upper level decision maker usually has complete
knowledge of the lower level problem, while the lower level
decision maker only observes the decisions of the leader and
then optimizes its own strategies. Interestingly, an incomplete
knowledge about the follower’s optimization problem to the
leader may lead to bilevel optimization problems involving
uncertainties.
It is not uncommon that the objectives of generally profitseeking private agents can well be in conflict with those of
the controlling authority. What makes such bilevel problemsolving tasks highly relevant is that they are typically characterized by very large spillover effects to the economy as well
as the surrounding environment. Given the far-reaching future
impacts of the decisions, it is not surprising that the interest
towards bilevel programming has grown strong especially
among researchers and practitioners dealing with large-scale
public sector decision-making problems. For instance, farmers
often tend to overuse fertilizers to increase the productivity,
which leads to negative externalities such as pollution. In
[179] authors use a bilevel model to design policy measures
to control the overuse of fertilizers and its negative impact
on the environment. Apart from problems concerned with
environmentally sensitive decisions (such as allocation of
mining permits or controlling the use of fertilizers), there
has been widespread interest across a number of fields in
operations research. A good example is homeland security,
where bilevel as well as trilevel optimization frameworks
have been utilized in problems ranging from interdiction of
nuclear-weapons projects to defending critical infrastructure
and solving border security problems [9], [176], [33], [34].
In addition to the public sector challenges, there is abundant
research on bilevel decision-making problems in economics,
logistics, as well as diverse areas of computer science.
The research on decision-making problems with hierarchical
leader-follower structures (bilevel optimization) can be traced
to two roots. The first root is in the domain of game theory,
where Stackelberg [157] used bilevel programming to build
descriptive models of decision behavior and establish game-
theoretic equilibria. The second root is in the domain of
mathematical programming, where the problems appeared
as bilevel optimization problems containing a nested inner
optimization problem as a constraint of an outer optimization
problem [30]. Since then a substantial body of mathematical
literature on bilevel optimization has emerged. Given that the
hierarchical optimization structure may introduce difficulties
such as non-convexity and disconnectedness even for simpler
instances of bilevel optimization, the problems have turned out
to be surprisingly difficult to handle mathematically. Bilevel
programming is known to be strongly NP-hard [80], and it has
been proven that merely evaluating a solution for optimality
is also a NP-hard task [167]. Even in the simplest case of
linear bilevel programs, where the lower level problem has a
unique optimal solution for all the parameters, it is not likely
to find a polynomial algorithm that is capable of solving the
linear bilevel program to global optimality. The proof for the
non-existence of a polynomial time algorithm for linear bilevel
problems can be found in [64].
Due to lack of well-established solution procedures, a
complex practical problem is usually modified into simpler
single level optimization task, which is solved to arrive at
a satisficing instead of an optimal solution. For the complex
bilevel problems, classical methods often fail due to real world
difficulties such as non-linearity and discreteness. Under such
circumstances, evolutionary methods can be useful tools to
offset some of these difficulties. Recent initiatives on bilevel
optimization using evolutionary algorithms suggest that a
coordinated effort on bilevel optimization by the evolutionary community could help make significant progress on this
challenging class of optimization problems (e.g., [152], [145],
[11], [57]).
Figure 2 provides a network map of different themes on
bilevel optimization that have been studied since 1950s. The
network map shows different theoretical and application topics
that have evolved under bilevel optimization. Each link in
the map connects a subtopic with a higher level topic that
are differentiated by font sizes. Subtopics connected with a
link denote an overlap. To provide a more comprehensive
overview on the past as well as recent developments in the field
of bilevel optimization, we have organized this review paper
along the three lines; theory, applications and text-analysis
of the entire bilevel literature body. First, to formalize the
notion of bilevel programming, we begin by introducing a few
central definitions and discuss the differences between optimistic and pessimistic formulations of bilevel problems. Once
the common terminology has been established, we offer an
overview on the algorithms that have been proposed for bilevel
optimization. After a brief coverage of the commonly used
classical approaches (e.g., descent methods, penalty function
methods, and trust region methods), we move on to discuss the
developments in the field of evolutionary computation, discrete
bilevel optimization and multiobjective bilevel optimization.
The method sections are followed by a review on the central
application areas. Finally, we study the research topics, and
the evolution of interest over time. The entire bilevel litera-
ture is divided into topics and a time series analysis across
each topic is performed. The text-analysis performed in this
paper is based on a recently developed non-parametric topic
model [27], [162] for analyzing unstructured information. The
technical details on the automated text-analysis approach are
provided in an appendix. The paper concludes with a brief
discussion on the directions for future research.
II. G ENERAL F ORMULATION AND D EFINITIONS
In this section, we provide a general formulation for bilevel
optimization problem. These problems contain two levels
of optimization tasks where one optimization task is nested
within the other. The outer optimization problem is commonly
referred as the leader’s (upper level) optimization problem and
the inner optimization problem is known as the follower’s
(or lower level) optimization problem. The two levels have
their own objectives and constraints. Correspondingly, there
are also two classes of decision vectors, namely, leader’s
(upper level) decision vectors and follower’s (lower level)
decision vectors. The lower level optimization is a parametric
optimization problem that is solved with respect to the lower
level decision vectors while the upper level decision vectors
act as parameters. The lower level optimization problem is a
constraint to the upper level optimization problem, such that,
only those members are considered feasible that are lower
level optimal and also satisfy the upper level constraints.
A summary of the terminologies and notations used in the
context of bilevel optimization is given in Table I.
Definition 1. For the upper-level objective function F : Rn ×
Rm → R and lower-level objective function f : Rn × Rm →
R, the bilevel problem is given by
“min”
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
xl ∈ argmin{f (xu , xl ) : gj (xu , xl ) ≤ 0, j = 1, . . . , J}
xl ∈XL
Gk (xu , xl ) ≤ 0, k = 1, . . . , K
where Gk : XU × XL → R, k = 1, . . . , K denote the upper
level constraints, and gj : XU × XL → R represent the lower
level constraints, respectively. Equality constraints may also
exist that have been avoided for brevity.
An equivalent formulation of the above problem can be
stated in terms of set-valued mapping (multi-valued function)
as follows:
Definition 2. Let Ψ : Rn ⇒ Rm be a set-valued mapping,
Ψ(xu ) = argmin{f (xu , xl ) : gj (xu , xl ) ≤ 0, j = 1, . . . , J},
xl ∈XL
which represents the constraint defined by the lower-level
optimization problem, i.e. Ψ(xu ) ⊂ XL for every xu ∈ XU .
Then the bilevel optimization problem can be expressed as a
Constraint
Convex
Mixed Integer Program
Condition
Nonlinear Global
Solution Karush-Kuhn-Tucker
Single Level Facility Location
Transport Toll
Freight
Planning
Rail
Transit
Bus
Route
Public
Passenger
Logistics
Combinatorial
Optimality
Fuzzy
Particle Swarm
Operators
Hybrid Multi-objective
Transportation
Theory
Evolutionary
Network
Policy Decision Hierarchy
Pollution Government Firm
Sustainable Tax
Genetic
Stochastic Congestion
Transport Mobile Wireless
Emmission
Environmental Economics
Bilevel
Attacker-Defender
Equilibrium
Hierarchical
Design Optimize
Stackelberg Nash
Security Hub and Spoke
Equilibrium Constraints
Intelligence Robust Design
Structure
Trilevel Adversary Autonomous
Stackelberg Game
Machine Learning
Dynamic
Electricity
Control
Power Grid
System
Design
Parameter Estimation
Transmission
Operation
Error Validation
Architecture
Computing Task
Processor
Design
Supply Chain
Pricing
Product Manufacturer
Bidding
Retailer
Commerce
Profit
Strategic
Distribution Contract
Distribute
Fig. 2: Bilevel network-map showing connections between various applications and theory since 1950s. Each connecting link
represents either a topic connected with a subtopic, or an overlap between two subtopics.
constrained optimization problem as follows:
“min”
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
xl ∈ Ψ(xu )
Gk (xu , xl ) ≤ 0, k = 1, . . . , K
where Ψ can be interpreted as a parameterized rangeconstraint for the lower-level decision vector xl .
In the above two definitions, quotes have been used while
specifying the upper level minimization problem because of
an ambiguity that arises in case of multiple lower level optimal
solutions for any given upper level decision vector. In the
presence of multiple lower level optimal solutions there is
lack of clarity at the upper level as to which optimal solution
from the lower level should be utilized. This ambiguity can be
sorted by defining different positions that may be assumed by
the leader. The two common positions that have been widely
studied are optimistic (weak) position and pessimistic (strong)
position, which we discuss next.
A. Optimistic Position
In an optimistic position, in the presence of multiple lower
level optimal solutions, the leader expects the follower to
choose that solution from the optimal set Ψo (xu ), which leads
to the best objective function value at the upper level. The
choice function of the follower in this case may be defined as
follows:
Ψo (xu ) = argmin{F (xu , xl ) : xl ∈ Ψ(xu )}
xl ∈XL
This formulation assumes some extent of cooperation between
the two players. The bilevel optimization problem under an
optimistic position has been defined below:
min
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
xl = Ψo (xu )
Gk (xu , xl ) ≤ 0, k = 1, . . . , K
Optimistic position is more tractable as compared to the
pessimistic position; therefore, most of the studies handle
TABLE I: Summary of central notations
Category
Notation(s)
Description
Decision vectors
xu ∈ XU
xl ∈ XL
Leader’s (upper level) decision variable and decision space.
Follower’s (lower level) decision variable and decision space.
Objectives
F
f
Leader’s (upper level) objective functions.
Follower’s (lower level) objective functions.
Constraints
Gk , k = 1, . . . , K
gj , j = 1, . . . , J
Leader’s (upper level) constraint functions.
Follower’s (lower level) constraint functions.
Lower level feasible region
Ω : XU ⇒ XL
Ω(xu ) = {xl : gj (xu , xl ) ≤ 0 ∀ j}, represents the lower level
feasible region for any given upper level decision vector
Constraint region (Relaxed feasible set)
Φ = gph Ω
Φ = {(xu , xl ) : Gk (xu , xl ) ≤ 0 ∀ k, gj (xu , xl ) ≤ 0 ∀ j},
represents the region satisfying both upper and lower level constraints
Lower level reaction set
Ψ : XU ⇒ XL
Ψ(xu ) = {xl : xl ∈ argmin{f (xu , xl ) : xl ∈ Ω(xu )}}, represents
xl ∈XL
the lower level optimal solution(s) for an upper level decision vector
I = gph Ψ
I = {(xu , xl ) : (xu , xl ) ∈ Φ, xl ∈ Ψ(xu )}, represents the set
of upper level decision vectors and corresponding lower level optimal
solution(s) belonging to feasible constraint region
Choice function
ψ : XU → XL
ψ(xu ) represents the solution chosen by the follower for any upper
level decision vector. It becomes important in case of multiple lower
level optimal solutions.
Optimal value function
ϕ : XU → R
Inducible region (Feasible set)
ϕ(xu ) = min {f (xu , xl ) : xl ∈ Ω(xu )} represents the minimum
xl ∈XL
lower level function value corresponding to a given upper level decision
vector.
optimistic version of the bilevel optimization problem. The optimistic formulation is guaranteed to have an optimal solutions
under reasonable assumptions of regularity and compactness
that are stated in the theorem below:
Theorem 1. If the functions F, f, Gk and gi are sufficiently
smooth, the constraint region Φ of the bilevel optimization
problem is non-empty and compact, and the MangasarianFromowitz constraint qualification holds at all points, then the
problem is guaranteed to have an optimistic bilevel optimum
provided there exists a feasible solution.
See [81], [127], [106], [107], [61], [59] for further discussion on existence of optimistic bilevel optimum and additional
results on optimality conditions.
B. Pessimistic Position
In a pessimistic position, in the presence of multiple lower
level optimal solutions, the leader optimizes for the worst
case, i.e., she assumes that the follower may choose that
solution from the optimal set which leads to the worst objective
function value at the upper level. Such a worst case choice
function of the follower may be defined as:
Ψp (xu ) = argmax{F (xu , xl ) : xl ∈ Ψ(xu )}
xl
This formulation does not assume any form of cooperation.
The bilevel optimization problem under a pessimistic position
has been defined below:
min
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
xl = Ψp (xu )
Gk (xu , xl ) ≤ 0, k = 1, . . . , K
Pessimistic position is relatively less tractable when compared
to optimistic position. In case of an optimisitc formulation
with a convex lower level problem, it is possible to reduce the
bilevel problem to single level using the variational inequality
corresponding to the lower level problem. However, such a
straightforward single level reduction is not possible in case
of a pessimistic bilevel program. This poses significant challenges in designing methodologies that can handle pessimistic
bilevel problems. For every lower level optimization problem
solved one has to keep track of that lower level optimal
solution that is worst for the upper level. This essentially
makes a pessimistic bilevel optimization a three level task.
The pessimistic formulation is guaranteed to have an optimal
solutions under stronger assumptions, as compared to the
optimistic formulation, that are given below:
Theorem 2. If the functions F, f, Gk and gi are sufficiently
smooth, the constraint region Φ of the bilevel optimization
problem is non-empty and compact, and the set-value mapping, Ψp , is lower semi-continuous for all upper level decision
vectors, then the problem is guaranteed to have a pessimistic
bilevel optimum.
For discussion on existence of pessimistic bilevel optimum
and additional results on optimality conditions, the authors
may refer to [112], [110], [61], [63], [180].
C. Example
Below we provide a simple example of a bilevel optimization problem [72] that arises in case of two firms in a Stackelberg competition. The leader has complete knowledge about
the follower’s inverse-demand function and the cost function,
and desires to maximize it’s own profits by taking into account
the actions of the follower firm. The two firms compete solely
by choosing their production levels that maximize their profits
(Πl and Πf ), and the follower acts only after observing the
actions of the leader. Formally, this model can be presented
as follows:
max
ql ,qf
s.t.
Πl = P (ql , qf )ql − Cl (ql )
qf
(3)
where P (ql , qf ) is the unit price of the goods sold, which
depends on the todal production. The assumption is that at
the optimum, all demand is satisfied. Cl (·) is the cost of
production of the leader and Cf (·) is the cost of production
of the follower. The variables in this model are the production
levels of each firm ql and qf . The leader sets its production
level first, and then the follower chooses its production level
based on the leader’s decision. This simple model assumes
homogeneity of the products manufactured by the firms.
By assuming that the firms produce and sell homogeneous
goods, we may assume a single linear price function for both
firms as an inverse demand function of the form
P (ql , qf ) = α − β(ql , qf ),
(4)
where α, β > 0 are constants. Additionally, since costs often
tend to increase with the amount of production, we assume
convex quadratic cost functions for both firms to be of the
form
C(ql ) =
C(qf ) =
δl ql2 + γl ql + cl ,
δf qf2 + γf qf + cf ,
In this section, we provide a brief overview of the classical
algorithms that have been proposed for bilevel optimization. Given the difficult nature of bilevel problems, it is
not surprising that much of the classical literature considers
bilevel problems that are mathematically well-behaved; i.e.,
contains functions that are linear, quadratic or convex. Strong
assumptions like continuous differentiability and lower semicontinuity are quite common. A significant amount of attention
has been given to linear bilevel optimization problems with
continuous [177], [20] and combinatorial [168] variables. For
more complex bilevel problems, the readers may refer [49],
[10].
A. Single-level Reduction
(1)
qf ∈ argmax{Πf = P (ql , qf )qf − Cf (qf )},(2)
ql , qf ≥ 0,
III. C LASSICAL A PPROACHES
(5)
(6)
where ci denotes the fixed costs of the respective firm, and
δi and γi are positive constants. It is possible to solve the
above model analytically, using the first order conditions of
the lower level problem to reduce it to single level, and then
using the first order conditions of the reduced problem.
The optimal level of production of the leader (ql∗ ) and the
follower (qf∗ ) in terms of the constants of the model is given
as follows:
2(β + δf )(α − γl ) − β(α − γf )
.
(7)
ql∗ =
4(β + δf )(β + δl ) − 2β 2
β 2 (α − γf )
β(α − γl ) −
α − γf
2(β + δf )
qf∗ =
−
.
(8)
2(β + δf ) 4(β + δf )(β + δl ) − 2β 2
When the lower level problem is convex and sufficiently
regular, it is possible to replace the lower level optimization
problem with its Karush-Kuhn-Tucker (KKT) conditions. The
KKT conditions appear as Lagrangian and complementarity
constraints, and reduce the overall bilevel optimization problem to a single-level constrained optimization problem. For
example, the problem in Definition 1 can be reduced to the
following form, when the convexity and regularity conditions
at the lower level are met:
min
xu ∈XU ,xl ∈XL ,λ
F (xu , xl )
subject to
Gk (xu , xl ) ≤ 0, k = 1, . . . , K,
∇xl L(xu , xl , λ) = 0,
gj (xu , xl ) ≤ 0, j = 1, . . . , J,
λj gj (xu , xl ) = 0, j = 1, . . . , J,
λj ≥ 0, j = 1, . . . , J,
where
L(xu , xl , λ) = f (xu , xl ) +
J
X
λj gj (xu , xl ).
j=1
The above formulation, though a single level optimization
task, is not necessarily simple to handle. The Lagrangian
constraints can lead to non-convexities even when suitable
convexity assumptions are made on all the objectives and
constraints in the bilevel formulation. The complementarity
condition, inherently being combinatorial, renders the singlelevel optimization problem as a mixed integer program.
Interestingly, for linear bilevel optimization problems, the
Lagrangian constraint is also linear. Therefore, the singlelevel optimization problem is a mixed integer linear program.
Approaches based on vertex enumeration [25], [46], [163], as
well as branch-and-bound (B&B) [16], [71] have been used
to solve these problems. It is noteworthy that B&B methods
constitute an exponentially slow algorithm with the number of
integer variables. But, B&B approaches have been successfully
applied to single-level reductions of linear-quadratic [17] and
quadratic-quadratic [65], [5] bilevel problems. An extended
KKT approach has also been considered [139] for handling
linear bilevel problems.
B. Descent Methods
In addition to KKT based approaches, a number of descent
methods have been proposed for solving bilevel optimization
problems. A descent direction in bilevel optimization leads
to decrease in upper level function value while keeping the
new point feasible. Given that a point is considered feasible
only if it is lower level optimal, finding the descent direction
can be quite challenging. To resolve the problem, researchers
have investigated ways to approximate the gradient of the
upper level objective [96] as well as considered formulation
of auxiliary programs [137], [167] to determine the direction
of descent.
C. Penalty Function Methods
In penalty function methods the bilevel optimization problem is handled by solving a series of unconstrained optimization problems. The unconstrained problem is generated by
adding a penalty term that measures the extent of violation of
the constraints. The penalty term often requires a parameter
and takes the value zero for feasible points and positive
(minimization) for infeasible points. For bilevel problems, the
first attempt towards using a penalized approach was made by
[1], [2]. They replaced the lower level problem by a penalized
problem; however, the bilevel hierarchy was still maintained
and the reduced problem was still difficult to solve. Later
double penalty method was introduced in [85], where both
upper and lower level objective functions were penalized. The
problem was reduced into a single level task by replacing
the penalized lower level problem with its KKT conditions,
and then solving the single level formulation by penalization.
In a number of studies, the lower level problem is directly
replaced by its KKT conditions and then a penalized approach
is used to solve the single level problem. Few studies where
penalty function approach has been used for linear bilevel
problems are [178], [113]. In [178], the authors convert the
linear bilevel program into a penalized bilinear optimization
problem, and then solve a series of bilinear problems to find
the optimum. In [113], the authors reduce the linear bilevel
program into single level using KKT conditions, and then
append the complementary slackness condition to the upper
level objective function with a penalty. The penalized problem
is then handled using a series of linear programs.
D. Trust-region Methods
In trust-region methods, the algorithms approximate a certain region of the objective function with a model function.
The region is expanded if the approximation is good, otherwise
it is contracted. The first study using trust-region method
to solve non-linear bilevel programs was presented in [109],
where the lower level problem had a convex objective function
and linear constraints. However, no upper level constraints
were considered. Later, a more general idea was proposed
in [117], where the authors locally approximate the bilevel
program with a model involving a linear program at the upper
level and a linear variational inequality at the lower level.
Trust-region and line search ideas have been combined to
approach the bilevel optimum over iterations. Similarly, in
[51], the authors approximate the bilevel program around an
iterate with a model that itself is a linear-quadratic bilevel
program. The authors propose to solve the linear-quadratic
bilevel program using a mixed integer solver after reducing
it to a single level problem using its lower level KKT conditions. Convergence is achieved by sequentially solving linearquadratic bilevel models.
Next, we discuss about evolutionary algorithms for bilevel
optimization. At this point, we would like to refer the readers
to other review papers [50], [58], [166], [91], [151] and books
[18], [141], [61], [161], [62] on bilevel optimization.
IV. E VOLUTIONARY A PPROACHES
A. Nested Methods
Nested evolutionary algorithms are a popular approach
to handle bilevel problems, where lower level optimization
problem is solved corresponding to each and every upper
level member [154]. Though effective, nested strategies are
computationally very expensive and not viable for large scale
bilevel problems. Nested methods in the area of evolutionary
algorithms have been used in primarily two ways. The first
approach has been to use an evolutionary algorithm at the
upper level and a classical algorithm at the lower level,
while the second approach has been to utilize evolutionary
algorithms at both levels. Of course, the choice between two
approaches is determined by the complexity of the lower level
optimization problem.
One of the first evolutionary algorithms for solving bilevel
optimization problems was proposed in the early 1990s.
Mathieu et al. [119] used a nested approach with genetic
algorithm at the upper level, and linear programming at the
lower level. Another nested approach was proposed in [187],
where the upper level was an evolutionary algorithm and the
lower level was solved using Frank-Wolfe algorithm (reduced
gradient method) for every upper level member. The authors
demonstrated that the idea can be effectively utilized to solve
non-convex bilevel optimization problems.
Nested particle swarm optimization (PSO) was used in [104]
to solve bilevel optimization problems. The effectiveness of the
technique was shown on a number of standard test problems
with small number of variables, but the computational expense
of the nested procedure was not reported. A hybrid approach
was proposed in [103], where simplex-based crossover strategy
was used at the upper level, and the lower level was solved
using one of the classical approaches. The authors report the
generations and population sizes required by the algorithm that
can be used to compute the upper level function evaluations,
but they do not explicitly report the total number of lower
level function evaluations, which presumably is high.
Differential evolution (DE) based approaches have also been
used, for instance, in [192], authors used DE at the upper level
and relied on the interior point algorithm at the lower level;
similarly, in [11] authors have used DE at both levels. Authors
have also combined two different specialized evolutionary
algorithms to handle the two levels, for example, in [12]
authors use an ant colony optimization to handle the upper
level and differential evolution to handle the lower level in
a transportation routing problem. Another nested approach
utilizing ant colony algorithm for solving a bilevel model
for production-distribution planning is [37]. Scatter search
algorithms have also been employed for solving productiondistribution planning problems, for instance [40].
Through a number of approaches involving evolutionary
algorithms at one or both levels, the authors have demonstrated
the ability of their methods in solving problems that might
otherwise be difficult to handle using classical bilevel approaches. However, as already stated, most of these approaches
are practically non-scalable. With increasing number of upper
level variables, the number of lower level optimization tasks
required to be solved increases exponentially. Moreover, if the
lower level optimization problem itself is difficult to solve,
numerous instances of such a problem cannot be solved, as
required by these methods.
B. Single-level Reduction
The idea behind single-level reduction, in the context of
evolutionary algorithms, is similar to the the discussions
in Section III-A. A number of researchers in the area of
evolutionary computation have also used the KKT conditions
of the lower level to reduce the bilevel problem into a singlelevel problem. Most often, such an approach is able to solve
problems that adhere to certain regularity conditions at the
lower level because of the requirement of the KKT conditions.
However, as the reduced single-level problem is solved with
an evolutionary algorithm, usually the upper level objective
function and constraints can be more general and not adhering
to such regularities. For instance, one of the earliest papers using such an approach is by Hejazi et al. [83], who reduced the
linear bilevel problem to single-level and then used a genetic
algorithm, where chromosomes emulate the vertex points, to
solve the problem. Wang et al. [174] reduced the bilevel
problem into a single-level optimization problem using KKT
conditions, and then utilized a constraint handling scheme to
successfully solve a number of standard test problems. Their
algorithm was able to handle non-differentiability at the upper
level objective function, but not elsewhere. Later on, Wang et
al. [175] introduced an improved algorithm that was able to
handle non-convex lower level problem and performed better
than the previous approach [174]. However, the number of
function evaluations in both approaches remained quite high
(requiring function evaluations to the tune of 100,000 for 2
to 5 variable bilevel problems). In [172], the authors used
a simplex-based genetic algorithm to solve linear-quadratic
bilevel problems after reducing it to a single level task. More
recently, Jiang et al. [87] reduced the bilevel optimization
problem into a non-linear optimization problem with complementarity constraints, which is sequentially smoothed and
solved with a PSO algorithm. Along similar lines of using
lower level optimality conditions, Li [102] solved a fractional
bilevel optimization problem by utilizing optimality results of
the linear fractional lower level problem. In [170], the authors
embed the chaos search technique in PSO to solve single-level
reduced problem.
C. Metamodeling-based Methods
Metamodeling-based solution methods are commonly used
for optimization problems [171], where actual function evaluations are expensive. A meta-model or surrogate model is an
approximation of the actual model that is relatively quicker
to evaluate. Based on a small sample from the actual model,
a surrogate model can be trained and used subsequently for
optimization. Given that, for complex problems, it is hard
to approximate the entire model with a small set of sample
points, researchers often resort to iterative meta modeling
techniques, where the actual model is approximated locally
during iterations.
Bilevel optimization problems contain an inherent complexity that leads to a requirement of large number of evaluations to solve the problem. Metamodeling, when used with
population-based algorithms, offers a viable means to handle
bilevel optimization problems. In this subsection, we discuss
four ways in which metamodeling can be applied to bilevel
optimization. The discussion related to approximation of the
rational reaction set and lower level optimal value function is
a review of some recent work. However, before starting, we
refer the readers to Figure 3, which provides an understanding
of these two mappings graphically for a hypothetical bilevel
problem. We also provide a brief discussion on approximating
the bilevel problem with an auxiliary problem.
1) Reaction set mapping: One of the approaches to solve
bilevel optimization problems using evolutionary algorithms
would be through iterative approximation of the reaction
set mapping Ψ. If the Ψ-mapping (introduced in Table I)
in a bilevel optimization problem is known, it effectively
reduces the problem to single level optimization. However, this
mapping is seldom available; therefore, the approach could
be to solve the lower level problem for a few upper level
members and then utilize the lower level optimal solutions
and corresponding upper level members to generate an approximate mapping Ψ̂. It is noteworthy that approximating a
set-valued Ψ-mapping offers its own challenges and is not a
straightforward task. Assuming that an approximate mapping,
Ψ̂, can be generated, the following single level optimization
problem can be solved for a few generations of the algorithm
before deciding to further refine the reaction set.
min
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
xl ∈ Ψ̂(xu )
Gk (xu , xl ) ≤ 0, k = 1, . . . , K
Evolutionary algorithms that rely on this idea to solve bilevel
optimization problems are [145], [150], [146], [13]. In [145],
f (x(3)
u ,x l )
ϕ (x u )
f ( xu , xl )
XU
f (x(2)
u ,x l)
XU
f (x(1)
u ,x l )
x(3)
u
x(3)
u
x (2)
u
x (2)
u
x(1)
u
C
(1)
xu
Inset 3
B
A
Inset 1
Inset 1 provides the relationship of the lower level function with
respect to the upper and lower level variables. The surface of the
lower level function is sliced with three planes, wherein first
upper level member has multiple lower level optimal solutions
while other members have unique lower level optimal solution.
Inset 2 provides the rational reaction set of the follower, which
maps follower’s optimal solutions with leader’s decision vectors.
The mapping is set−valued in regions that have multiple lower
level optimal solutions corresponding to leader’s decision vectors.
Inset 3 provides the follower’s optimal value function, which is
the minimum value of follower’s objective function for any given
leader’s decision vector.
XL
XU
x(3)
u
Inset 2
x(2)
u
x(1)
u
Ψ−Mapping
(1)
Ψ(x u )
(2)
Ψ(x u )
(3)
Ψ(x u )
XL
Fig. 3: Graphical representation of rational reaction set (Ψ) and lower level optimal value function (ϕ).
[146] authors have used quadratic approximation to approximate the local reaction set. This helps in saving lower
level optimization calls when the approximation for the local
reaction set is good. In case the approximations generated by
the algorithm are not acceptable, the method defaults to a
nested approach. It is noteworthy that a bilevel algorithm that
uses a surrogate model for reaction set mapping may need not
be limited to quadratic models but other models can also be
used.
2) Optimal lower level value function: Another way to
use metamodeling would be through the approximation of
the optimal value function ϕ. If the ϕ-mapping (introduced
in Table I) is known, the bilevel problem can once again be
reduced to single level optimization problem as follows [186],
min
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
f (xu , xl ) ≤ ϕ(xu )
gj (xu , xl ) ≤ 0, j = 1, . . . , J
Gk (xu , xl ) ≤ 0, k = 1, . . . , K.
However, since the value function is seldom known, one
can attempt to approximate this function using metamodeling
techniques. The optimal value function is a single-valued
mapping; therefore, approximating this function avoids the
complexities associated with set-valued mapping. As described
previously, an approximate mapping ϕ̂, can be generated with
the population members of an evolutionary algorithm and the
following single level optimization problem can be solved with
refinements at every few generations.
min
xu ∈XU ,xl ∈XL
F (xu , xl )
subject to
f (xu , xl ) ≤ ϕ̂(xu )
gj (xu , xl ) ≤ 0, j = 1, . . . , J
Gk (xu , xl ) ≤ 0, k = 1, . . . , K.
An evolutionary approach that relies on this idea can be found
in [149], [144].
3) Bypassing lower level problem: Another way to use a
meta-model in bilevel optimization would be completely bypass the lower level problem, as follows:
min
xu ∈XU
F̂ (xu )
subject to
Ĝk (xu ) ≤ 0, k = 1, . . . , K.
Given that the optimal xl are essentially a function of xu , it
is possible to construct a single level problem by ignoring xl
completely. However, the landscape for such a single level
problem can be highly non-convex, disconnected and nondifferentiable. Advanced metamodeling techniques might be
required to use this approach, which may be beneficial for
certain classes of bilevel problems. A training set for the
metamodel can be constructed by solving few lower level
problems for different xu . Both upper level objective F and
constraint set (Gk ) can then be meta-modeled using xu alone.
Given the complex structure of such a single-level problem, it
might be sensible to create such an approximation locally. We
are currently pursuing such an approach using artificial neural
network as the meta-modeling approach.
4) Auxiliary bilevel meta-model: Building up on the trustregion methods for solving bilevel optimization problems, it is
possible to utilize the population members in an evolutionary
algorithm to formulate auxiliary bilevel problem(s). The auxiliary bilevel problem(s) may be simple enough to be solved
using faster specialized techniques. The population members
could then be updated based on the obtained auxiliary solution(s). For the moment, there does not exist any evolutionary
algorithm based on this idea, but it may be an interesting
direction to pursue in the future.
V. D ISCRETE B ILEVEL O PTIMIZATION
In this section, we would like to discuss the contributions
made towards solving discrete bilevel optimization problems.
The formulation of the bilevel problem remains the same as
described in Definitions 1 and 2, along with one or more
variables at either of the levels being discrete. Presence of
discrete variables can pose a variety of challenges depending
upon, whether the discrete variables are present at upper level,
lower level or both levels. In the classical literature, branchand-bound and branch-and-cut are some of the commonly used
techniques to handle discreteness in variables. Most of the
work on discrete bilevel optimization employs an extension
of these ideas from single-level optimization. To highlight the
kind of complexities induced in the presence of discrete variables, we consider a simple linear bilevel problem described
in [169], [18] to show how the inducible region changes based
on the upper, lower or both level variables being discrete or
continuous.
Consider the following lower level optimization problem
which we use for identifying the inducible region of a bilevel
problem:
min xl
xl
subject to
xu + xl ≤ 2,
−xu + xl ≤ 2,
5xu − 4xl ≤ 10, −5xu − 4xl ≤ 10.
For the above lower level problem there are four possible
scenarios based on the variables being continuous or discrete:
1) Continuous-Continuous bilevel program: Consider xu ∈
R and xl ∈ R
2) Discrete-Continuous bilevel program: Consider xu ∈ Z
and xl ∈ R
Dicrete−Continuous
Continous−Continuous
xl
xl
xu
xu
Continuous−Discrete
Discrete−Discrete
xl
xl
xu
xu
Fig. 4: Inducible region in different cases when the upper and
lower level variables belong to continuous or discrete sets.
3) Discrete-Discrete bilevel program: Consider xu ∈ Z and
xl ∈ Z
4) Continuous-Discrete bilevel program: Consider xu ∈ R
and xl ∈ Z
For each scenario the inducible region can be very different
that has been shown through Figure 4. The figure clearly
demonstrates how a discrete variable at any level of the
problem can lead to a disconnected search space. Of course
there could also be situations where each level has a mix of
continuous and discrete variables.
A. Discrete Bilevel Optimization Survey
One of the early works on discrete bilevel optimization was
by Vicente et al. (1996) [169], which focused on discrete linear
bilevel programs, and analyzed the properties and existence
of the optimal solution for different kinds of discretizations
arising from the upper and lower level variables. The authors
have shown in the paper that certain compactness conditions
guarantee the existence of optimal solution in continouscontinuous linear bilevel programs, discrete-continuous linear
bilevel programs and discrete-discrete linear bilevel programs.
The conditions are equivalent to stating that the inducible
region is non-empty. However, the existence conditions in
the case of continuous-discrete linear bilevel programs are
not straightforward. For instance, the inducible region for the
continuous-discrete linear bilevel problem in Figure 4 is a noncompact set that may lead to non-existence of a bilevel optimal
solution even when the inducible region is non-empty.
A few studies preceded the study by Vicente et al. (1996)
[169]. For instance, in [124] the authors solved mixed integer
linear bilevel problems. The authors pointed out the difficulties
involved in fathoming while solving mixed integer bilevel
problems using traditional branch-and-bound techniques. Certain fathoming rules used in case of mixed integer linear
programming, like fathoming when the relaxed subproblem is
worse than the value of the incumbent or fathoming when the
solution of the relaxed subproblem is feasible for the mixed
integer problem, are not directy applicable to mixed integer
linear bilevel problems. Therefore, the authors proposed a
branch-and-bound approach involving stricter fathoming conditions. However, the algorithm has a nested structure and
is not scalable beyond few integer variables, and to counter
which the authors also proposed some heuristics. This study
was followed by [19], where the same authors solved discrete
linear bilevel programs involving only binary variables using
an implicit enumeration scheme. In this approach, the authors
place a cut, similar to the one used by Bialas and Karwan
[26] (for the continuous linear bilevel program), seeking incremental improvements in the upper level objective function.
A cutting plane method utlizing the Chvátal-Gomory cut
for the continuous-discrete bilevel program was proposed in
[60]. Benders-decomposition based techniques have also been
employed to solve bilevel problems with mixed integers at the
upper level and continuous linear programs at the lower level.
The original problem is decomposed into a master and a slave
problem. Fixing the integer values converts the slave problem
into a bilevel linear program which is solved by KKT based
reduction techniques, and the solution to the slave is utilized
to create a cut for the master problem. The algorithm switches
between master and slave problems until the optimality criteria
is met. Certain studies in this direction are [136], [70], [41].
Despite the attempts made towards algorithm development
for discrete bilevel programs, the research is still open for new
methods and ideas as none of the proposed techniques would
scale well for problems with larger number of variables. Apart
from a few nested approaches and KKT-based single level
reduction approaches, to our best knowledge, there does not
exist any algorithmic study involving evolutionary algorithms
for mixed integer bilevel problems that attempt to solve
the problem efficiently by utilizing its properties. However,
given that the evolutionary approaches are, in particular,
potent for handling difficulties such as discreteness and nondifferentiabilities they offer a significant scope for solving
discrete bilevel optimization problems. Some attempts towards
mixed integer bilevel optimization using evolutionary methods
are [82], [14], [121], [101], [4], [38], [79], [44].
There is no dearth of application problems involving discrete variables. While mixed integer bilevel problems are ubiquitous, even combinatorial bilevel programs find innumerable
applications in the areas of network design, facility location,
hub-and-spoke networks etc. In these areas, these problems are
commonly studied in the context of interdiction, protection,
robust design, competition and supply chain management,
among others. Some of the related applications are highlighted
in Section VII.
VI. M ULTIOBJECTIVE B ILEVEL O PTIMIZATION
In many of the practical problems, a leader and/or the
follower might face multiple objectives. This gives rise to
multiobjective bilevel optimization problems that we define
below.
Definition 3. For the upper-level objective function F : Rn ×
Rm → Rp and lower-level objective function f : Rn × Rm →
Rq , the multi-objective bilevel problem is given by
“min”
xu ∈XU ,xl ∈XL
F (xu , xl ) = (F1 (xu , xl ), . . . , Fp (xu , xl ))
subject to
xl ∈ argmin{f (xu , xl ) = (f1 (xu , xl ), . . . , fq (xu , xl )) :
xl ∈XL
gj (xu , xl ) ≤ 0, j = 1, . . . , J}
Gk (xu , xl ) ≤ 0, k = 1, . . . , K
where Gk : XU × XL → R, k = 1, . . . , K denote the upper
level constraints, and gj : XU × XL → R represent the lower
level constraints, respectively.
The set-valued mapping in this case can be defined as
follows and an equivalent definition can be written as in the
single-objective case:
Ψ(xu ) = argmin{f (xu , xl ) = (f1 (xu , xl ), . . . , fq (xu , xl )) :
xl ∈XL
gj (xu , xl ) ≤ 0, j = 1, . . . , J},
A. Optimistic vs Pessimistic
The optimistic or pessimistic position becomes more prominent in multiobjective bilevel optimization. In the presence of
multiple objectives at the lower level, the set-valued mapping
Ψ(·) normally represents a set of Pareto-optimal solutions
corresponding to any given xu , which we refer as follower’s
Pareto-optimal frontier. A solution to the overall problem (with
optimistic or pessimistic position) is expected to produce a
trade-off frontier for the leader that we refer as the leader’s
Pareto-optimal frontier. From the perspective of the leader, it
becomes important that what kind of position she seeks to
take while solving the problem, as it determines that which
solution(s) from the lower level frontier should be considered
at the upper level.
Though optimistic positions have commonly been studied
in classical [67] and evolutionary [57] literature in the context
of multiobjective bilevel optimization; it is far from realism
to expect that the follower will cooperate to an extent that
she chooses any point from her Pareto-optimal frontier that
is most suitable for the leader. This relies on the assumption
that the follower is indifferent to the entire set of optimal
solutions, and therefore decides to cooperate. The situation
was entirely different in the single-objective case, where, in
case of multiple optimal solutions, all the solutions offered
an equal value to the follower. However, this can not be
assumed in the multiobjective case. Solution to the optimistic
formulation in multiobjective bilevel optimization leads to the
best possible Pareto-optimal frontier that can be achieved by
Follower’s
problem
(1)
f2
for x u
Al
f2
0
Follower’s
problem
(2)
for x u
0
f1
Bl
1
0
Pessimistic
PO front
Au
0
f1
F2
Follower’s
problem
The mining company has a sole objective of maximizing its
profit under the constraints set by the government. In this
scenario, the government would like to have a tax structure
such that it is able to maximize its own revenues in addition
to being able to restrain the mining company from causing
extensive damage to the environment. It is possible for the
leader to optimally regulate the problem in its favour, provided
that it has complete knowledge of the follower’s strategies.
The hierarchical optimization problem in this case can be
formulated as follows:
(3)
for x u
f2
Bu
max
τ,q
0
Optimistic
PO front
Cl
0
f1
s.t.
Cu
q ≥ 0, τ ≥ 0.
0
0
F1
F(q, τ ) = (R, −D)
(
)
π(q) = p(q)q − c(q) − R
q ∈ argmax
π(q) ≥ 0
q
(9)
(10)
(11)
1
Fig. 5: Leader’s Pareto-optimal (PO) frontiers for optimistic
and pessimistic positions. Few follower’s Pareto-optimal (PO)
frontiers are shown (in insets) along with their representations
in the leader’s objective space.
the leader. Similarly, solution to the pessimistic formulation
leads to the worst possible Pareto-optimal frontier at the upper
level.
If the value function or the choice function of the follower is
known to the leader, it provides an information as to what kind
of trade-off is preferred by the follower. A knowledge of such
a function effectively, casually speaking, reduces the lower
level optimization problem into a single-objective optimization
task, where the value function may be directly optimized. The
leader’s Pareto-optimal frontier for such intermediate positions
lies between the optimistic and the pessimistic frontiers. Figure
5 shows the optimistic and pessimistic frontiers for a hypothetical multiobjective bilevel problem with two objectives at
upper and lower levels. Follower’s frontier corresponding to
x1u , x2u and x3u , and her decisions Al , Bl and Cl are shown in
the insets. The corresponding representations of the follower’s
frontier and decisions (Au , Bu and Cu ) in the leader’s space
are also shown.
B. Example
Below, we provide a bilevel optimization problem involving
design of a tax policy [153]. The upper level in this example
is the governemnt that wants to tax the lower level, a mining
company, based on the pollution it causes to the environment.
The government here has two objectives: the first objective is
to maximize the revenues generated by the mining project,
which may include the additional jobs, taxes, etc; and the
second objective is to minimize the harm caused to the environment as a result of mining. Obviously, there is a trade-off
between the two objectives, and the government as a decision
maker needs to choose one of the preferred trade-off solutions.
In (9), the first objective deals with the tax revenue, where
R = τ q; τ is the per unit tax imposed on the mine, and q is the
amount of metal extracted from the ore by the follower. The
second objective denotes the environmental damage caused by
the mine that the government ultimately wants to minimize.
D = kq, where k is the pollution coefficient signifying
the negative impact of extraction on the environment. The
damages are thus linear and scale proportionately with the
amount of gold extracted from the earth since a larger base of
operation implies larger environmental damage.
Equation (10) gives the profit of the mine, where p(q)q
(price function times amount of metal extracted) is the revenue
function, and c(q) is the extraction cost function followed by
the additional tax levied on the mine. The mine is most likely
to be a price taker when it comes to the price of gold and
must base its mining decisions on the possible price paid by
their customers. It would therefore be plausible to replace the
price function for gold in the above equation by a constant.
However, given the assumption that the mine can extract a
large amount of ore, and subsequently gold, at one time, it
might be possible for it to affect the price of gold slightly.
Therefore, we assume the price function to be linear with a
small slope. Extraction cost can be considered to be quadratic.
Thus, we have the following model:
max
τ,q
F (q, τ ) = (τ q, −kq)
(12)
s.t.
π(q) =(α − βq)q−
q ∈ argmax
(δq 2 + γq + φ) − τ q
q
π(q) ≥ 0
(13)
q ≥ 0, τ ≥ 0,
(14)
where α, β, δ, γ, φ are constants, and φ represents the fixed
costs of setting up operations. The above problem can be
solved analytically by taking a weighted sum of squares of
the upper level objectives (wτ q − (1 − w)kq : w ∈ [0, 1]). The
0
−2
Objective 2 (−D)
−4
−6
−8
−10
−12
0
100
200
300
400
Objective 1 (R)
500
600
Fig. 6: Pareto-optimal frontier for the government showing the
trade-off between tax revenues and environmental pollution.
optimal solution to the above problem is given as follows:
α−γ−k
k
+
,
2
2w
w(α − γ) − (1 − w)k
q ∗ (w) =
.
4w(β + δ)
τ ∗ (w) =
(15)
(16)
We assume the parameters as α = 100, β = 1, δ = 1, γ = 1
and φ = 0. By varying the government’s preference weights
(w) in its domain, one can generate the entire Pareto-optimal
solutions for the leader. Note that a very high taxation (or
weight to the environmental objective) may lead to no production at the lower level, for instance, we find that w < 0.01
does not lead to any production. The Pareto frontier generated
using weights 0.01 ≤ w ≤ 1 has been provided in Figure 6
for the above model.
One of the first studies, utilizing an evolutionary approach
for multiobjective bilevel optimization was by Yin [187]. The
study involved multiple objectives at the upper lever, and
a single objective at the lower level. The study suggested
a nested genetic algorithm, and applied it on a transportation planning and management problem. Later Halter and
Mostaghim [77] used a particle swarm optimization (PSO)
based nested strategy to solve a multi-component chemical
system. The lower level problem in their application was
linear for which they used a specialized linear multi-objective
PSO approach. Recently, a hybrid bilevel evolutionary multiobjective optimization algorithm coupled with local search was
proposed in [57] (For earlier versions, refer [54], [143], [56],
[55]). In the paper, the authors handled non-linear as well
as discrete bilevel problems with relatively larger number of
variables. The study also provided a suite of test problems for
bilevel multi-objective optimization.
There has been some work done on decision making aspects
at upper and lower levels. For example, in [142] an optimistic version of multiobjective bilevel optimization, involving
interaction with the upper level decision maker, has been
solved. The approach leads to the most preferred point at the
upper level instead of the entire Pareto-frontier. Since multiobjective bilevel optimization is computationally expensive,
such an approach was justified as it led to enormous savings in
computational expense. Studies that have considered decision
making at the lower level include [147], [152]. In [147], the
authors have replaced the lower level with a value function that
effectively reduces the lower level problem to single-objective
optimization task. In [152], the follower’s value function is
known with uncertainty, and the authors propose a strategy
to handle such problems. Other work related to bilevel multiobjective optimization can be found in [129], [130], [108],
[135], [191].
VII. A PPLICATIONS
C. Multiobjective Bilevel Optimization Survey
There exists a significant amount of work on single objective
bilevel optimization; however, little has been done on multiobjective bilevel optimization primarily because of the computational and decision making complexities that these problems
offer. For results on optimality conditions in multiobjective
bilevel optimization, the readers may refer to [73], [185], [15].
On the methodology side, Eichfelder [66], [67] solved simple
multi-objective bilevel problems using a classical approach.
The lower level problems in these studies have been solved
using a numerical optimization technique, and the upper
level problem is handled using an adaptive exhaustive search
method. This makes the solution procedure computationally
demanding and non-scalable to large-scale problems. In another study, Shi and Xia [140] used -constraint method at
both levels of multi-objective bilevel problem to convert the
problem into an -constraint bilevel problem. The -parameter
is elicited from the decision maker, and the problem is solved
by replacing the lower level constrained optimization problem
with its KKT conditions.
Bilevel optimization commonly appears in many practical
problems. They are often encountered in the fields of economics, transportation, engineering and management, among
others. The following list will provide an insight to the readers
on the relevance of these problems to practice.
1) Toll Setting Problem: Toll-setting problem is essentially
a part of network problems. In this problem, there is an
authority that wants to optimize the tolls for a network
of roads. The authority acts as a leader and the network
users act as followers. Papers on toll-setting problem
and its multi-objective extensions can be found in [32],
[122], [52], [98], [189], [116], [90], [173], [148], [68],
[76], [92]. Bilevel optimization is quite commonly used
in network design problems. Instead of going through
specific problems, we refer the readers to a variety
of applications of bilevel optimization to the area of
network design [100], [21], [115], [184], [188], [43],
[183], [45], [74], [69], [120], [181], [134], [39].
2) Environmental Economics: Bilevel optimization commonly appears in environmental economics, where an
Fig. 7: Interest on bilevel optimization over time.
authority wants to tax an organization or individual that
is polluting the environment as a result of its operations.
Finding an optimal level of tax that offers a compromise
between revenues and pollution results in a bilevel
optimization problem with the regulator as the leader
and the polluting entity as a follower [8], [153], [28],
[29], [179].
3) Chemical Industry: In chemical industries, the chemists
often face a bilevel optimization problem where they
have to decide upon the conditions (state variables and
quantity of reactants) for the reaction to achieve optimal
output. While optimizing the output is the upper level
problem, the lower level appears as an equilibrium
condition, which is an entropy functional minimization
problem. Such applications of bilevel optimization can
be found in [156], [48], [131], [78]
4) Optimal design: Bilevel problems are very common in
structural optimization or optimal shape design. For
instance, in structural optimization one often requires
to minimize the weight or cost of a structure as an
upper level objective with the decision variables as
shape of the structure, choice of materials, amount of
material etc. The constraints at the upper level involve
bounds on displacements, stresses and contact forces
whose values are determined by solving the potential
energy minimization problem at the lower level. The
equilibrium condition in many optimal shape design
problems appears in the form of variational inequalities
which require the overall problem to formulated as a two
level task. For optimal design applications the readers
may refer to [84], [47], [22], [94], [95].
5) Defense applications: Bilevel optimization has a number
of applications in the defense sector [31], for example
attacker-defender Stackelberg games [86], [35], [138],
[126], [9], [132], [133]. Specifically, some recent applications include planning the prepositioning of defensive
missile interceptors to counter an attack threat [33],
interdicting nuclear weapons project [34], homeland
security applications [176], [111], location problems [3].
The bilevel problem, while offending, involves maxi-
Fig. 8: Interest on evolutionary bilevel optimization over time.
mizing the damage caused to the opponent by taking
into account the optimal reactions of the opponent. Conversely, while defending, the bilevel problem involves
minimizing the maximum damage that an attacker can
cause.
6) Facility Location: Facility location problems may take
the form of a Stackelberg game if a firm, while locating
its facility, decides to account the actions of its competitors. For instance, in [97] the authors study the scenario
where a firm enters a market by locating new facilities,
and its competitor reacts by adjusting the attractiveness
of its existing facilities. Another study considers location of logistics distribution centers by minimizing the
planners’ cost at the upper level and customers’ cost
at the lower level [158]. Other applications of bilevel
optimization to facility location problem may be found
in [88], [164], [159], [7], [36], [38], [128], [38], [118],
[41], [114].
7) Inverse Optimal control: Inverse optimal control problems are essentially bilevel in nature [123], [6], [89],
[160] with wide applications in robotics, computer vision, communication theory and remote sensing to name
a few. One of the major challenges in control theory is
deriving the performance index or reward function which
fits best on a given dataset. Such tasks lie in the category
of inverse optimal control theory, where one solicits the
calculation of the cause based on the given result. Such a
requirement necessitates solving a parameter estimation
problem with an optimal control problem.
8) Machine learning: Most of the machine learning and
evolutionary optimization techniques often involve a
number of parameters. A proper choice of these parameters has a substantial effect on the accuracy and
efficiency of the approach. Tuning of these parameters
is often achieved using brute force strategies, such as
grid search and random search. A bilevel formulation
of this problem allows for systematic and more efficient
search when the number of parameters are large. Some
of the approaches that have acknowledged the bilevel
nature of this problem are [23], [24], [155], [105].
9) Principal-agent problems: Principal-agent problem [99]
is a classical problem in economics, where a principal
(leader) sub-contracts a job to an agent (follower).
Given that the agent prefers to act in his own interests rather than those of the principal, it becomes
important for the principal to have an incentive scheme
that aligns the interests of the agent with the principal. Design of such contracts appears as a bilevel
optimization problem. In real life, principal-agent relationships are commonly found in doctor-patient, senior
management-lower management, employer-employee,
corporate board-shareholders and politician-voters scenarios. Studies that the readers may refer to are [165],
[75], [190], [42], [182].
Fig. 9: Topic: Methods.
VIII. I NTEREST OVER TIME
In this section, we perform a text analysis of papers on
bilevel programming (and Stackelberg games) that are indexed
in SCOPUS. To begin with, we analyze the volume of publications every year on bilevel programming since 1950s to
present, and then closely look at the themes within bilevel programming that have contributed to the growth over the years.
The themes were discovered using a non-parametric Bayesian
approach [162], which clusters the documents together based
on similarities. The documents may probabilistically belong
to multiple clusters at the same time.
Figure 7 shows how the interest on bilevel programming has
been growing at a slow pace until early 2000s and then pickedup significantly at the middle of the previous decade. The
studies on bilevel programming using evolutionary algorithms
appeared for the first time during the mid 1990s that took
another decade to pick up to the extent that almost 10% of all
studies on bilevel optimization utilize evolutionary methods.
Figure 7 shows the growth of evolutionary methods in the
context of bilevel optimization from the 1990s to present.
While the early papers on bilevel programming (pre-2000)
were mainly focused on solution methods and optimality
conditions, the growth in the post-2000 period was fueled by
papers on applications of bilevel programming.
To identify the themes that have contributed towards the
literature on bilevel optimization we used topic models. “Topic
models are algorithms for discovering the main themes that
pervade a large and otherwise unstructured collection of
documents.” [27] These models can be used to organize
unstructured collections as well as develop insights from large
text databases that made it suitable for our purposes. The
results from the topic model for the documents retrieved from
SCOPUS are given in Figures 9-24. The figures consist of
themes that have a volume at least to the order of around 1%.
Along with the identification of themes, the analysis helped
in determining the attention received by particular themes at
any point in time.
Each figure contains a word cloud in the inset that describes
the theme and volume of papers as the other inset that
describes the number of papers published on that theme over
the years. Interestingly, we observe that number of papers
Fig. 10: Topic: Optimality conditions.
Fig. 11: Topic: Classical game theory.
Fig. 12: Topic: Network design.
on classical bilevel methods and optimality conditions peaked
during 1995-2000. Since 2000 a number of bilevel applications
picked-up, for example, we see a growth in supply chain
applications, electricity transmission applications, telecommunication applications, facility location applications, railway applications and machine learning applications. Defense applications that appeared to be minimally present before 2000 show
a significant presence later. Network design, optimal design
and business applications do not show a trend but represent
the highest volume of applications on bilevel optimization.
Fig. 13: Topic: Supply chain applications.
Fig. 17: Topic: Business applications.
Fig. 14: Topic: Optimal design applications.
Fig. 18: Topic: Computer architecture and circuit design.
Fig. 15: Topic: Electricity transmission applications.
Fig. 19: Topic: Hierarchical decision making applications.
Fig. 16: Topic: Telecommunication applications.
Fig. 20: Topic: Environment applications.
IX. C ONCLUSIONS AND F UTURE R ESEARCH D IRECTIONS
A. Metamodeling-based Algorithms
In this concluding section, we will raise a few perspectives
that have not yet received much attention but may offer interesting directions for future research. Apart from metamodeling
based techniques to solve bilevel problems, we would like
to highlight the importance of being able to account for
different forms of uncertainties that are often encountered
when solving practical problems. Another interesting direction
is concerned with scalability of bilevel algorithms and ability
to leverage distributed computing platforms to handle large
scale problems.
Though we have already highlighted the importance of
metamodeling-based methods in an earlier section, we have
decided to discuss it once again because of its potential
in solving practical bilevel optimization problems. Bilevel
optimization problems inherit a number of mappings, and
any knowledge about the structure of these mappings can
simplify the solution procedure extensively. In this paper, we
have highlighted approaches that are based on approximating
the reaction set mapping and the optimal lower level value
function mapping. Knowing one of these mappings reduces
B. Multiobjective Bilevel Optimization and Decision Making
Fig. 21: Topic: Facility location applications.
Fig. 22: Topic: Railway applications.
Multiobjective bilevel optimization has received only lukewarm interest from researchers. A number of issues, like
decision interaction between the two levels and uncertainties
in decision making, remain unexplored. It might be of interest
for researchers working in the area of multi-criteria decision
making and multiobjective evolutionary optimization to explore how two levels of decision makers interact to arrive
at a compromising or an equilibrium solution in different
situations. Similarly, the notion of uncertain decision-maker’s
preferences at one or both levels has also not received enough
attention [152]. In the field of Decision Analysis, however,
plenty of research has been carried out to extend traditional
frameworks of decision making such as Expected Utility Theory [125] and Multi-attribute Utility Theory [93] to account
for uncertain preferences on, for instance, the trade-offs among
multiple decision objectives or the risk-attitude. However,
preferential uncertainty in bilevel optimization problems still
requires development of theory as well as methods to account
for decision behavior in a hierarchical setting. The approaches
that have been proposed so far are still very preliminary and
require substantial future research.
C. Bilevel Optimization under Variable Uncertainty
Fig. 23: Topic: Machine learning applications.
Another important research topic is concerned with the
inherent uncertainty of decision variables. This poses several
challenges for the existing more deterministic optimization
frameworks that may fail to find solutions that are robust and
sufficiently close to the optimal solutions. The fact that bilevel
problems have nested optimization tasks makes the search
of robust solutions substantially more challenging compared
with single-level optimization problems. A few preliminary
ideas for handling variable uncertainty have already been
suggested [53], but the required algorithm side innovations
that would make these problems accessible to practitioners
are still missing.
D. Scaling of Evolutionary Bilevel Algorithms
Fig. 24: Topic: Defense applications.
any bilevel optimization problem to a single-level optimization
problem. For solving large scale bilevel problems, one has
to exploit the structure and properties of bilevel problems
that are essentially contained in these mappings. Other ways
of utilizing metamodeling for bilevel problems that we have
discussed are: approximating the bilevel problem by bypassing
the lower level problem completely; and utilizing auxiliary
bilevel models to locally approximate a bilevel optimization
problem. Only few studies in the context of evolutionary
algorithms utilize such a strategies and offer opportunities for
future contributions.
Bilevel problems are well-known for being highly computationally intensive already before considering any types of
uncertainties. One of the promising directions for handling
larger bilevel problems could be the use of the recent distributed computing platforms such as Apache Spark project.
The programming model of Spark is quite different from the
Hadoop MapReduce framework, and it has managed to overcome many of the earlier limitations. In particular, its current
form has turned out to support the use and development of
iterative algorithms quite well. Therefore, it may be interesting
to investigate whether this novel platform will be able to offer
a helping hand for researchers and practitioners who seek to
solve bigger bilevel problems faster.
Though considerable progress has already been made during
the last few years, evolutionary bilevel optimization is still a
relatively young field with numerous opportunities for both
computational as well as theoretical innovations. The growing
availability of algorithms is also opening the field to more
applied research, and we believe that in the future we are
likely to see a considerable amount of novel applications.
X. ACKNOWLEDGMENTS
Ankur Sinha and Pekka Malo would like to acknowledge the
support provided by Liikesivistysrahasto and Helsinki School
of Economics Foundation. K. Deb acknowledges the support
from NSF Beacon Center for the study of evolution in action
at MSU under Cooperative Agreement No. DBI-0939454.
Any opinions, findings, and conclusions or recommendations
expressed in this material are those of the author(s) and
do not necessarily reflect the views of the National Science
Foundation.
R EFERENCES
[1] E. Aiyoshi and K. Shimizu. Hierarchical decentralized systems and
its new solution by a barrier method. IEEE Transactions on Systems,
Man, and Cybernetics, 6:444–449, 1981.
[2] E. Aiyoshi and K. Shimizu. A solution method for the static constrained
Stackelberg problem via penalty method. IEEE Transactions on
Automatic Control, 29:1111–1114, 1984.
[3] Deniz Aksen and Necati Aras. A bilevel fixed charge location model for
facilities under imminent attack. Computers & Operations Research,
39(7):1364–1381, 2012.
[4] Deniz Aksen and Necati Aras. A matheuristic for leader-follower
games involving facility location-protection-interdiction decisions. In
Metaheuristics for Bi-level Optimization, pages 115–151. Springer,
2013.
[5] F. Al-Khayyal, R. Horst, and P. Pardalos. Global optimization of
concave functions subject to quadratic constraints: an application
in nonlinear bilevel programming. Annals of Operations Research,
34:125–147, 1992.
[6] Sebastian Albrecht, K Ramirez-Amaro, Federico Ruiz-Ugalde, David
Weikersdorfer, M Leibold, Michael Ulbrich, and Michael Beetz. Imitating human reaching motions using physically inspired optimization
principles. In Humanoid Robots (Humanoids), 2011 11th IEEE-RAS
International Conference on, pages 602–607. IEEE, 2011.
[7] E. Alekseeva, N. Kochetova, Yu Kochetov, and A. Plyasunov. A
hybrid memetic algorithm for the competitive p-median problem. IFAC
Proceedings Volumes, 42(4):1533–1537, 2009.
[8] Mahyar A. Amouzegar and Khosrow Moshirvaziri. Determining optimal pollution control policies: An application of bilevel programming.
European Journal of Operational Research, 119(1):100–120, 1999.
[9] B. An, F. Ordóñez, M. Tambe, E. Shieh, R. Yang, C. Baldwin,
J. DiRenzo III, K. Moretti, B. Maule, and G. Meyer. A Deployed
Quantal Response-Based Patrol Planning System for the U.S. Coast
Guard. Interfaces, 43(5):400–420, 2013.
[10] G. Anandalingam and T. Friesz. Hierarchical optimization: an introduction. Annals of Operations Research, 34:1–11, 1992.
[11] J. Angelo, E. Krempser, and H. Barbosa. Differential evolution
for bilevel programming. In Proceedings of the 2013 Congress on
Evolutionary Computation (CEC-2013). IEEE Press, 2013.
[12] Jaqueline S. Angelo and Helio J. C. Barbosa. A study on the use
of heuristics to solve a bilevel programming problem. International
Transactions in Operational Research, 2015.
[13] Jaqueline S Angelo, Eduardo Krempser, and Helio JC Barbosa. Differential evolution assisted by a surrogate model for bilevel programming
problems. In Evolutionary Computation (CEC), 2014 IEEE Congress
on, pages 1784–1791. IEEE, 2014.
[14] J. M. Arroyo and F. J. Fernández. A genetic algorithm approach
for the analysis of electric grid interdiction with line switching. In
Intelligent System Applications to Power Systems, 2009. ISAP’09. 15th
International Conference on, pages 1–6. IEEE, 2009.
[15] Bernd Bank, Jürgen Guddat, Diethard Klatte, Bernd Kummer, and
Klaus Tammer. Non-linear parametric optimization. Birkhäuser Basel,
1983.
[16] J. Bard and J. Falk. An explicit solution to the multi-level programming
problem. Computers and Operations Research, 9:77–100, 1982.
[17] J. Bard and J. Moore. A branch and bound algorithm for the bilevel
programming problem. SIAM Journal on Scientific and Statistical
Computing, 11:281–292, 1990.
[18] J. F. Bard. Practical Bilevel Optimization: Algorithms and Applications. The Netherlands: Kluwer, 1998.
[19] Jonathan F. Bard and James T. Moore. An algorithm for the discrete
bilevel programming problem. Naval Research Logistics (NRL),
39(3):419–435, 1992.
[20] O. Ben-Ayed. Bilevel linear programming. Computers and Operations
Research, 20:485–501, 1993.
[21] O. Ben-Ayed, C. Blair, D. Boyce, and L. LeBlanc. Construction of
a real-world bilevel linear programming model of the highway design
problem. Annals of Operations Research, 34:219–254, 1992.
[22] Martin P Bendsoe. Optimization of structural topology, shape, and
material, volume 414. Springer, 1995.
[23] Kristin P. Bennett, Jing Hu, Xiaoyun Ji, Gautam Kunapuli, and JongShi Pang. Model selection via bilevel optimization. In Neural
Networks, 2006. IJCNN’06. International Joint Conference on, pages
1922–1929. IEEE, 2006.
[24] Kristin P. Bennett, Gautam Kunapuli, Jing Hu, and Jong-Shi Pang.
Bilevel optimization and machine learning. In Computational Intelligence: Research Frontiers, pages 25–47. Springer, 2008.
[25] W. Bialas and M. Karwan. Two-level linear programming. Management
Science, 30:1004–1020, 1984.
[26] Wayne F. Bialas and Mark H. Karwan. Two-level linear programming.
Management science, 30(8):1004–1020, 1984.
[27] David M. Blei. Probabilistic topic models. Communications of the
ACM, 55(4):77–84, 2012.
[28] M. Bostian, A. Sinha, Gerald Whittaker, and Bradley Barnhart. Incorporating data envelopment analysis solution methods into bilevel
multi-objective optimization. In 2015 IEEE Congress on Evolutionary
Computation (CEC-2015). IEEE Press, 2015.
[29] Moriah Bostian, Gerald Whittaker, Brad Barnhart, Rolf Färe, and
Shawna Grosskopf. Valuing water quality tradeoffs at different spatial
scales: An integrated approach using bilevel optimization. Water
Resources and Economics, 11:1–12, 2015.
[30] J. Bracken and J. McGill. Mathematical programs with optimization
problems in the constraints. Operations Research, 21:37–44, 1973.
[31] Jerome Bracken and James T. McGill. Defense applications of
mathematical programs with optimization problems in the constraints.
Operations Research, 22(5):1086–1096, 1974.
[32] Luce Brotcorne, Martine Labbe, Patrice Marcotte, and Gilles Savard. A
bilevel model for toll optimization on a multicommodity transportation
network. Transportation Science, 35(4):345–358, 2001.
[33] G. Brown, M. Carlyle, D. Diehl, J. Kline, and K. Wood. A TwoSided Optimization for Theater Ballistic Missile Defense. Operations
Research, 53(5):745–763, 2005.
[34] G. Brown, M. Carlyle, R. C. Harney, E.M. Skroch, and K. Wood. Interdicting a Nuclear-Weapons Project. Operations Research, 57(4):866–
877, 2009.
[35] Gerald Brown, Matthew Carlyle, Javier Salmerón, and Kevin Wood.
Defending critical infrastructure. Interfaces, 36(6):530–544, 2006.
[36] Herminia I. Calvete, Carmen Galé, and José A. Iranzo. An efficient
evolutionary algorithm for the ring star problem. European Journal of
Operational Research, 231(1):22–33, 2013.
[37] Herminia I. Calvete, Carmen Galé, and María-José Oliveros. Bilevel
model for production–distribution planning solved by using ant colony
optimization. Computers & Operations Research, 38(1):320–327,
2011.
[38] José-Fernando Camacho-Vallejo, Álvaro Eduardo Cordero-Franco, and
Rosa G. González-Ramírez. Solving the bilevel facility location
problem under preferences by a stackelberg-evolutionary algorithm.
Mathematical Problems in Engineering, 2014, 2014.
[39] José-Fernando Camacho-Vallejo, Julio Mar-Ortiz, Francisco LópezRamos, and Ricardo Pedraza Rodríguez. A genetic algorithm for
the bi-level topological design of local area networks. PloS one,
10(6):e0128067, 2015.
[40] José-Fernando Camacho-Vallejo, Rafael Muñoz-Sánchez, and José Luis
González-Velarde.
A heuristic algorithm for a supply chain’s
production-distribution planning. Computers & Operations Research,
61:110–121, 2015.
[41] Massimiliano Caramia and Renato Mari. A decomposition approach
to solve a bilevel capacitated facility location problem with equity
constraints. Optimization Letters, pages 1–23, 2015.
[42] Mark Cecchini, Joseph Ecker, Michael Kupferschmid, and Robert
Leitch. Solving nonlinear principal-agent problems using bilevel
programming. European Journal of Operational Research, 230(2):364–
373, 2013.
[43] Halim Ceylan and Michael G. H. Bell. Traffic signal timing optimisation based on genetic algorithm approach, including drivers5. routing.
Transportation Research Part B: Methodological, 38(4):329–342, 2004.
[44] Abir Chaabani, Slim Bechikh, and Lamjed Ben Said. A co-evolutionary
decomposition-based algorithm for bi-level combinatorial optimization.
In 2015 IEEE Congress on Evolutionary Computation (CEC), pages
1659–1666. IEEE, 2015.
[45] Anthony Chen, Juyoung Kim, Seungjae Lee, and Youngchan Kim.
Stochastic multi-objective models for network design problem. Expert
Systems with Applications, 37(2):1608–1619, 2010.
[46] Y. Chen and M. Florian. On the geometry structure of linear bilevel
programs: a dual approach. Technical Report CRT-867, Centre de
Recherche sur les Transports, 1992.
[47] Snorre Christiansen, Michael Patriksson, and Laura Wynter. Stochastic
bilevel programming in structural optimization. Structural and multidisciplinary optimization, 21(5):361–371, 2001.
[48] Peter A. Clark and Arthur W. Westerberg. Bilevel programming for
steady-state chemical process design-i. fundamentals and algorithms.
Computers & Chemical Engineering, 14(1):87–97, 1990.
[49] B. Colson. Mathematical programs with equilibrium constraints and
nonlinear bilevel programming problems. Technical report, Master’s
thesis, Department of Mathematics, FUNDP, Namur, Belgium, 1999.
[50] B. Colson, P. Marcotte, and G. Savard. An overview of bilevel
optimization. Annals of Operational Research, 153:235–256, 2007.
[51] Benoît Colson, Patrice Marcotte, and Gilles Savard. A trust-region
method for nonlinear bilevel programming: algorithm and computational experience. Computational Optimization and Applications,
30(3):211–227, 2005.
[52] Isabelle Constantin and Michael Florian. Optimizing frequencies
in a transit network: a nonlinear bi-level programming approach.
International Transactions in Operational Research, 2(2):149 – 164,
1995.
[53] K. Deb, Z. Lu, and A. Sinha. Handling decision variable uncertainty in
bilevel optimization problems. In 2015 IEEE Congress on Evolutionary
Computation (CEC-2015). IEEE Press, 2015.
[54] K. Deb and A. Sinha. Constructing test problems for bilevel evolutionary multi-objective optimization. In 2009 IEEE Congress on
Evolutionary Computation (CEC-2009), pages 1153–1160. IEEE Press,
2009.
[55] K. Deb and A. Sinha. An evolutionary approach for bilevel multiobjective problems. In Cutting-Edge Research Topics on Multiple Criteria Decision Making, Communications in Computer and Information
Science, volume 35, pages 17–24. Berlin, Germany: Springer, 2009.
[56] K. Deb and A. Sinha. Solving bilevel multi-objective optimization
problems using evolutionary algorithms. In Evolutionary MultiCriterion Optimization (EMO-2009), pages 110–124. Berlin, Germany:
Springer-Verlag, 2009.
[57] K. Deb and A. Sinha. An efficient and accurate solution methodology
for bilevel multi-objective programming problems using a hybrid
evolutionary-local-search algorithm. Evolutionary Computation Journal, 18(3):403–449, 2010.
[58] S. Dempe. Annotated bibliography on bilevel programming and
mathematical programs with equilibrium constraints. Optimization,
52(3):339–359, 2003.
[59] S. Dempe, J. Dutta, and B.S. Mordukhovich. New necessary optimality
conditions in optimistic bilevel programming. Optimization, 56(56):577–604, 2007.
[60] Stephan Dempe. Discrete bilevel optimization problems. Citeseer,
1996.
[61] Stephan Dempe. Foundations of Bilevel Programming. Kluwer
Academic Publishers, Secaucus, NJ, USA, 2002.
[62] Stephan Dempe, Vyacheslav Kalashnikov, Gerardo A. Pérez-Valdés,
and Nataliya Kalashnykova. Bilevel programming problems. Energy
Systems. Springer, Berlin, 2015.
[63] Stephan Dempe, Boris S. Mordukhovich, and Alain Bertrand Zemkoho.
Necessary optimality conditions in pessimistic bilevel programming.
Optimization, 63(4):505–533, 2014.
[64] Xiaotie Deng. Complexity issues in bilevel linear programming. In
Multilevel optimization: Algorithms and applications, pages 149–164.
Springer, 1998.
[65] T. Edmunds and J. Bard. Algorithms for nonlinear bilevel mathematical
programming. IEEE Transactions on Systems, Man, and Cybernetics,
21:83–89, 1991.
[66] G. Eichfelder. Solving nonlinear multiobjective bilevel optimization
problems with coupled upper level constraints. Technical Report
Preprint No. 320, Preprint-Series of the Institute of Applied Mathematics, Univ. Erlangen-Nornberg, Germany, 2007.
[67] Gabriele Eichfelder. Multiobjective bilevel optimization. Mathematical
Programming, 123(2):419–449, June 2010.
[68] Wei Fan. Optimal congestion pricing toll design for revenue maximization: comprehensive numerical results and implications. Canadian
Journal of Civil Engineering, 42(8):544–551, 2015.
[69] Wei Fan and Randy Machemehl. Bi-level optimization model for
public transportation network redesign problem: Accounting for equity
issues. Transportation Research Record: Journal of the Transportation
Research Board, 2263:151–162, 2011.
[70] Pirmin Fontaine and Stefan Minner. Benders decomposition for
discrete–continuous linear bilevel problems with application to traffic
network design. Transportation Research Part B: Methodological,
70:163–172, 2014.
[71] J. Fortuny-Amat and B. McCarl. A representation and economic
interpretation of a two-level programming problem. Journal of the
Operational Research Society, 32:783–792, 1981.
[72] A. Frantsev, A. Sinha, and P. Malo. Finding optimal strategies in
multi-period stackelberg games using an evolutionary framework. In
IFAC Workshop on Control Applications of Optimization (IFAC-2009).
Elsevier, 2012.
[73] N Gadhi and Stephan Dempe. Necessary optimality conditions and a
new approach to multiobjective bilevel optimization problems. Journal
of Optimization Theory and Applications, 155(1):100–114, 2012.
[74] Mariano Gallo, Luca D’Acierno, and Bruno Montella. A meta-heuristic
approach for solving the urban network design problem. European
Journal of Operational Research, 201(1):144–157, 2010.
[75] John E. Garen. Executive compensation and principal-agent theory.
Journal of Political Economy, pages 1175–1199, 1994.
[76] José Luis González-Velarde, José-Fernando Camacho-Vallejo, and
Gabriel Pinto Serrano. A scatter search algorithm for solving a bilevel
optimization model for determining highway tolls. Computación y
Sistemas, 19(1):05–16, 2015.
[77] W. Halter and S. Mostaghim. Bilevel optimization of multi-component
chemical systems using particle swarm optimization. In Proceedings
of World Congress on Computational Intelligence (WCCI-2006), pages
1240–1247, 2006.
[78] Werner Halter and Sanaz Mostaghim. Bilevel optimization of multicomponent chemical systems using particle swarm optimization. In
Evolutionary Computation, 2006. CEC 2006. IEEE Congress on, pages
1240–1247. IEEE, 2006.
[79] Stephanus Daniel Handoko, Lau Hoong Chuin, Abhishek Gupta,
Ong Yew Soon, Heng Chen Kim, and Tan Puay Siew. Solving multivehicle profitable tour problem via knowledge adoption in evolutionary
bi-level programming. In 2015 IEEE Congress on Evolutionary
Computation (CEC), pages 2713–2720. IEEE, 2015.
[80] P. Hansen, B. Jaumard, and G. Savard. New branch-and-bound rules for
linear bilevel programming. SIAM Journal on Scientific and Statistical
Computing, 13:1194–1217, 1992.
[81] P. Harker and J.-S. Pang. Existence of optimal solutions to mathematical programs with equilibrium constraints. Operations Research
Letters, 7:61–64, 1988.
[82] Li Hecheng and Wang Yuping. Exponential distribution-based genetic
algorithm for solving mixed-integer bilevel programming problems.
Journal of Systems Engineering and Electronics, 19(6):1157–1164,
2008.
[83] S. Reza Hejazi, Azizollah Memariani, G. Jahanshahloo, and Mohammad Mehdi Sepehri. Linear bilevel programming solution by genetic
algorithm. Computers & Operations Research, 29(13):1913–1925,
2002.
[84] J. Herskovits, A. Leontiev, G. Dias, and G. Santos. Contact shape
optimization: a bilevel programming approach. Structural and multidisciplinary optimization, 20(3):214–221, 2000.
[85] Y. Ishizuka and E. Aiyoshi. Double penalty method for bilevel
optimization problems. Annals of Operations Research, 34:73–88,
1992.
[86] Eitan Israeli and R. Kevin Wood. Shortest-path network interdiction.
Networks, 40(2):97–111, 2002.
[87] Yan Jiang, Xuyong Li, Chongchao Huang, and Xianing Wu. Application of particle swarm optimization based on chks smoothing
function for solving nonlinear bilevel programming problem. Applied
Mathematics and Computation, 219(9):4332–4339, 2013.
[88] Qin Jin and Shi Feng. Bi-level simulated annealing algorithm for
facility location. Systems Engineering, 2:007, 2007.
[89] Mark Johnson, Navid Aghasadeghi, and Timothy Bretl. Inverse
optimal control for deterministic continuous-time nonlinear systems.
In Decision and Control (CDC), 2013 IEEE 52nd Annual Conference
on, pages 2906–2913. IEEE, 2013.
[90] Vyacheslav Kalashnikov, Fernando Camacho, Ronald Askin, and Nataliya Kalashnykova. Comparison of algorithms for solving a bi-level
toll setting problem. International Journal of Innovative Computing,
Information and Control, 6(8):3529–3549, 2010.
[91] Vyacheslav V. Kalashnikov, Stephan Dempe, Gerardo A. Pérez-Valdés,
Nataliya I. Kalashnykova, and José-Fernando Camacho-Vallejo. Bilevel
programming and applications. Mathematical Problems in Engineering,
2015, 2015.
[92] Vyacheslav V. Kalashnikov, Roberto Carlos Herrera Maldonado, JoséFernando Camacho-Vallejo, and Nataliya I. Kalashnykova. A heuristic
algorithm solving bilevel toll optimization problems. The International
Journal of Logistics Management, 27(1):31–51, 2016.
[93] R. L. Keeney and H. Raiffa. Decisions with Multiple Objectives:
Preferences and Value Tradeoffs. New York: Wiley, 1976.
[94] Michal Kočvara. Topology optimization with displacement constraints:
a bilevel programming approach. Structural optimization, 14(4):256–
263, 1997.
[95] Michal Kocvara and Jifi V. Outrata. On the solution of optimum design
problems with variational inequalities. Recent Advances in Nonsmooth
Optimization, pages 172–192, 1995.
[96] C. Kolstad and L. Lasdon. Derivative evaluation and computational
experience with large bilevel mathematical programs. Journal of
Optimization Theory and Applications, 65:485–499, 1990.
[97] Hande Küçükaydin, Necati Aras, and I Kuban Altınel. Competitive
facility location problem with attractiveness adjustment of the follower:
A bilevel programming model and its solution. European Journal of
Operational Research, 208(3):206–220, 2011.
[98] Martine Labbé, Patrice Marcotte, and Gilles Savard. A bilevel model of
taxation and its application to optimal highway pricing. Management
science, 44(12-part-1):1608–1622, 1998.
[99] Jean-Jacques Laffont and David Martimort. The theory of incentives:
the principal-agent model. Princeton university press, 2009.
[100] L. Leblanc and D. Boyce. A bilevel programming algorithm for
exact solution of the network design problem with user-optimal flows.
Transportation Research, 20 B:259–265, 1986.
[101] François Legillon, Arnaud Liefooghe, and El-Ghazali Talbi. Cobra:
A cooperative coevolutionary algorithm for bi-level optimization. In
2012 IEEE Congress on Evolutionary Computation, pages 1–8. IEEE,
2012.
[102] Hecheng Li. A genetic algorithm using a finite search space for solving
nonlinear/linear fractional bilevel programming problems. Annals of
Operations Research, pages 1–16, 2015.
[103] Hecheng Li and Yuping Wang. A hybrid genetic algorithm for solving
nonlinear bilevel programming problems based on the simplex method.
International Conference on Natural Computation, 4:91–95, 2007.
[104] Xiangyong Li, Peng Tian, and Xiaoping Min. A hierarchical particle swarm optimization for solving bilevel programming problems.
In Leszek Rutkowski, Ryszard Tadeusiewicz, Lotfi A. Zadeh, and
Jacek M. Zurada, editors, Artificial Intelligence and Soft Computing
- ICAISC 2006, volume 4029 of Lecture Notes in Computer Science,
pages 1169–1178. Springer Berlin Heidelberg, 2006.
[105] Jason Liang and Risto Miikkulainen. Evolutionary bilevel optimization
for complex control tasks. In Proceedings of the 17th Annual Genetic
and Evolutionary Computation Conference (GECCO 2015). New York:
ACM Press, 2015.
[106] M. Lignola and J. Morgan. Topological existence and stability for
stackelberg problems. Journal of Optimization Theory and Applications, 84:145–169, 1995.
[107] Maria Beatrice Lignola and Jacqueline Morgan. Existence of solutions
to bilevel variational problems in banach spaces. In Equilibrium
Problems: Nonsmooth Optimization and Variational Inequality Models,
pages 161–174. Springer, 2001.
[108] Mikko Linnala, Elina Madetoja, Henri Ruotsalainen, and Jari Hämäläi-
[109]
[110]
[111]
[112]
[113]
[114]
[115]
[116]
[117]
[118]
[119]
[120]
[121]
[122]
[123]
[124]
[125]
[126]
[127]
[128]
[129]
[130]
[131]
nen. Bi-level optimization for a dynamic multiobjective problem.
Engineering Optimization, 44(2):195–207, 2012.
Guoshan Liu, Jiye Han, and Shouyang Wang. A trust region algorithm
for bilevel programing problems. Chinese science bulletin, 43(10):820–
824, 1998.
P. Loridan and J. Morgan. Weak via strong Stackelberg problems.
Journal of Global Optimization, 8:263–287, 1996.
J. Lowe. Homeland Security: Operations Research Initiatives and
Applications. Interfaces, 36(6):483–485, 2006.
R. Lucchetti, F. Mignanego, and G. Pieri. Existence theorem of
equilibrium points in Stackelberg games with constraints. Optimization,
18:857–866, 1987.
Yibing Lv, Tiesong Hu, Guangmin Wang, and Zhongping Wan. A
penalty function method based on kuhn–tucker condition for solving
linear bilevel programming. Applied Mathematics and Computation,
188(1):808–813, 2007.
Sayuri Maldonado-Pinto, Martha-Selene Casas-Ramírez, and JoséFernando Camacho-Vallejo. Analyzing the performance of a hybrid heuristic for solving a bilevel location problem under different
approaches to tackle the lower level. Mathematical Problems in
Engineering, 2016, 2016.
P. Marcotte and G. Marquis. Efficient implementation of heuristics
for the continuous network design problem. Annals of Operations
Research, 34:163–176, 1992.
Patrice Marcotte, Gilles Savard, and Frédéric Semet. A bilevel
programming approach to the travelling salesman problem. Operations
Research Letters, 32(3):240–248, 2004.
Patrice Marcotte, Gilles Savard, and D. L. Zhu. A trust region
algorithm for nonlinear bilevel programming. Operations research
letters, 29(4):171–179, 2001.
Miroslav Marić, Zorica Stanimirović, Nikola Milenković, and Aleksandar Ðenić. Metaheuristic approaches to solving large-scale bilevel uncapacitated facility location problem with clients’preferences. Yugoslav
Journal of Operations Research ISSN: 0354-0243 EISSN: 2334-6043,
25(3), 2014.
R. Mathieu, L. Pittard, and G. Anandalingam. Genetic algorithm
based approach to bi-level linear programming. Operations Research,
28(1):1–21, 1994.
Mahmoud Mesbah, Majid Sarvi, and Graham Currie. Optimization of
transit priority in the transportation network using a genetic algorithm.
Intelligent Transportation Systems, IEEE Transactions on, 12(3):908–
919, 2011.
Elnaz Miandoabchi and Reza Zanjirani Farahani. Optimizing reserve
capacity of urban road networks in a discrete network design problem.
Advances in Engineering Software, 42(12):1041–1050, 2011.
Athanasios Migdalas. Bilevel programming in traffic planning: Models,
methods and challenge. Journal of Global Optimization, 7(4):381–405,
1995.
Katja Mombaur, Anh Truong, and Jean-Paul Laumond. From human
to humanoid locomotion1.an inverse optimal control approach. Autonomous robots, 28(3):369–383, 2010.
James T. Moore and Jonathan F. Bard. The mixed integer linear bilevel
programming problem. Operations research, 38(5):911–921, 1990.
John Von Neumann and Oskar Morgenstern. Theory of games and
economic behavior. Princeton, NJ. Princeton University Press., 1946.
Jesse R. O’Hanley and Richard L. Church. Designing robust coverage
networks to hedge against worst-case facility losses. European Journal
of Operational Research, 209(1):23–36, 2011.
J. Outrata. Necessary optimality conditions for Stackelberg problems.
Journal of Optimization Theory and Applications, 76:305–320, 1993.
Artem Aleksandrovich Panin, M. G. Pashchenko, and Aleksandr Vladimirovich Plyasunov. Bilevel competitive facility location
and pricing problems. Automation and Remote Control, 75(4):715–727,
2014.
C. O. Pieume, L. P. Fotso, and P. Siarry. Solving bilevel programming
problems with multicriteria optimization techniques. OPSEARCH,
46(2):169–183, 2009.
Surapati Pramanik and Partha Pratim Dey. Bi-level multi-objective
programming problem with fuzzy parameters. International Journal of
Computer Applications, 30(10):13–20, September 2011. Published by
Foundation of Computer Science, New York, USA.
Arvind U. Raghunathan and Lorenz T. Biegler. Mathematical programs
with equilibrium constraints (mpecs) in process engineering. Computers & Chemical Engineering, 27(10):1381–1392, 2003.
[132] P. Ramamoorthy, S. Jayaswal, A. Sinha, and N. Vidyarthi. Hub
interdiction & hub protection problems: Model formulations & exact
solution methods. (No. WP 2016-10-01). Technical report, Indian
Institute of Management Ahmedabad, 2016.
[133] P. Ramamoorthy, S. Jayaswal, A. Sinha, and N. Vidyarthi. Hub-andspoke network design under the risk of interdiction. (No. WP 201705-01). Technical report, Indian Institute of Management Ahmedabad,
2017.
[134] Gang Ren, Zhengfeng Huang, Yang Cheng, Xing Zhao, and Yong
Zhang. An integrated model for evacuation routing and traffic signal optimization with background demand uncertainty. Journal of
Advanced Transportation, 47(1):4–27, 2013.
[135] S. Ruuska and K. Miettinen. Constructing evolutionary algorithms
for bilevel multiobjective optimization. In Evolutionary Computation
(CEC), 2012 IEEE Congress on, pages 1–7, june 2012.
[136] Georges K. Saharidis and Marianthi G. Ierapetritou. Resolution method
for mixed integer bi-level linear problems based on decomposition
technique. Journal of Global Optimization, 44(1):29–51, 2009.
[137] G. Savard and J. Gauvin. The steepest descent direction for the
nonlinear bilevel programming problem. Operations Research Letters,
15:275–282, 1994.
[138] Maria P. Scaparra and Richard L. Church. A bilevel mixed-integer
program for critical infrastructure protection planning. Computers &
Operations Research, 35(6):1905–1923, 2008.
[139] Chenggen Shi, Jie Lu, and Guangquan Zhang. An extended kuhn–
tucker approach for linear bilevel programming. Applied Mathematics
and Computation, 162(1):51–63, 2005.
[140] X. Shi. and H. S. Xia. Model and interactive algorithm of bi-level
multi-objective decision-making with multiple interconnected decision
makers. Journal of Multi-Criteria Decision Analysis, 10(1):27–34,
2001.
[141] K. Shimizu, Y. Ishizuka, and J. F. Bard. Nondifferentiable and two-level
mathematical programming. Dordrecht: Kluwer Academic, 1997.
[142] A. Sinha. Bilevel multi-objective optimization problem solving using
progressively interactive evolutionary algorithm. In Proceedings of
the Sixth International Conference on Evolutionary Multi-Criterion
Optimization (EMO-2011), pages 269–284. Berlin, Germany: SpringerVerlag, 2011.
[143] A. Sinha and K. Deb. Towards understanding evolutionary bilevel
multi-objective optimization algorithm. In IFAC Workshop on Control
Applications of Optimization (IFAC-2009), volume 7. Elsevier, 2009.
[144] A. Sinha, Z. Lu, K. Deb, and P. Malo. Bilevel optimization based on
iterative approximation of mappings. arXiv preprint arXiv:1702.03394,
2017.
[145] A. Sinha, P. Malo, and K. Deb. Efficient evolutionary algorithm for
single-objective bilevel optimization. arXiv preprint arXiv:1303.3901,
2013.
[146] A. Sinha, P. Malo, and K. Deb. An improved bilevel evolutionary
algorithm based on quadratic approximations. In 2014 IEEE Congress
on Evolutionary Computation (CEC-2014), pages 1870–1877. IEEE
Press, 2014.
[147] A. Sinha, P. Malo, and K. Deb. Towards understanding bilevel multiobjective optimization with deterministic lower level decisions. In
Proceedings of the Eighth International Conference on Evolutionary
Multi-Criterion Optimization (EMO-2015). Berlin, Germany: SpringerVerlag, 2015.
[148] A. Sinha, P. Malo, and K. Deb. Transportation policy formulation as a
multi-objective bilevel optimization problem. In 2015 IEEE Congress
on Evolutionary Computation (CEC-2015). IEEE Press, 2015.
[149] A. Sinha, P. Malo, and K. Deb. Solving optimistic bilevel programs by
iteratively approximating lower level optimal value function. In 2016
IEEE Congress on Evolutionary Computation (CEC-2016). IEEE Press,
2016.
[150] A. Sinha, P. Malo, and K. Deb. Evolutionary algorithm for bilevel
optimization using approximations of the lower level optimal solution
mapping. European Journal of Operational Research, 2016 (In press).
[151] A. Sinha, P. Malo, and K. Deb. Evolutionary bilevel optimization: An
introduction and recent advances. In Recent Advances in Evolutionary
Multi-objective Optimization, pages 71–103. Springer, 2017.
[152] A. Sinha, P. Malo, K. Deb, P. Korhonen, and J. Wallenius. Solving
bilevel multi-criterion optimization problems with lower level decision uncertainty. IEEE Transactions on Evolutionary Computation,
20(2):199–217, 2016.
[153] A. Sinha, P. Malo, A. Frantsev, and K. Deb. Multi-objective stackelberg
game between a regulating authority and a mining company: A case
study in environmental economics. In 2013 IEEE Congress on
Evolutionary Computation (CEC-2013). IEEE Press, 2013.
[154] A. Sinha, P. Malo, A. Frantsev, and K. Deb. Finding optimal strategies
in a multi-period multi-leader-follower stackelberg game using an
evolutionary algorithm. Computers & Operations Research, 41:374–
385, 2014.
[155] A. Sinha, P. Malo, P. Xu, and K. Deb. A bilevel optimization approach
to automated parameter tuning. In Proceedings of the 16th Annual
Genetic and Evolutionary Computation Conference (GECCO 2014).
New York: ACM Press, 2014.
[156] W. R. Smith and R. W. Missen. Chemical Reaction Equilibrium
Analysis: Theory and Algorithms. John Wiley & Sons, New York,
1982.
[157] H. Stackelberg. The theory of the market economy. Oxford University
Press, New York, Oxford, 1952.
[158] Huijun Sun, Ziyou Gao, and Jianjun Wu. A bi-level programming
model and solution algorithm for the location of logistics distribution
centers. Applied Mathematical Modelling, 32(4):610–616, 2008.
[159] Huijun Sun, Ziyou Gao, and Jianjun Wu. A bi-level programming
model and solution algorithm for the location of logistics distribution
centers. Applied Mathematical Modelling, 32(4):610 – 616, 2008.
[160] V. Suryan, A. Sinha, , P. Malo, and K. Deb. Handling inverse optimal
control problems using evolutionary bilevel optimization. In 2016
IEEE Congress on Evolutionary Computation (CEC-2016). IEEE Press,
2016.
[161] El-Ghazali Talbi. Metaheuristics for bi-level optimization, volume 482.
Springer, 2013.
[162] Yee Whye Teh, Michael I Jordan, Matthew J Beal, and David M Blei.
Hierarchical dirichlet processes. Journal of the american statistical
association, 101(476), 2006.
[163] H. Tuy, A. Migdalas, and P. Värbrand. A global optimization approach
for the linear two-level program. Journal of Global Optimization, 3:1–
23, 1993.
[164] Takeshi Uno, Hideki Katagiri, and Kosuke Kato. An evolutionary
multi-agent based search method for stackelberg solutions of bilevel
facility location problems. International Journal of Innovative Computing, Information and Control, 4(5):1033–1042, 2008.
[165] A. Van Ackere. The principal/agent paradigm: characterizations and
computations. European Journal of Operations Research, 70:83–103,
1993.
[166] L. Vicente and P. Calamai. Bilevel and multilevel programming: a
bibliography review. Journal of Global Optimization, 5:291–306, 1994.
[167] L. Vicente, G. Savard, and J. Júdice. Descent approaches for quadratic
bilevel programming. Journal of Optimization Theory and Applications, 81:379–399, 1994.
[168] L. Vicente, G. Savard, and J. Júdice. The discrete linear bilevel programming problem. Journal of Optimization Theory and Applications,
89:597–614, 1996.
[169] Luis Vicente, Gilles Savard, and J. Judice. Discrete linear bilevel programming problem. Journal of optimization theory and applications,
89(3):597–614, 1996.
[170] Zhongping Wan, Guangmin Wang, and Bin Sun. A hybrid intelligent
algorithm by combining particle swarm optimization with chaos searching technique for solving nonlinear bilevel programming problems.
Swarm and Evolutionary Computation, 8:26–32, 2013.
[171] G. Gary Wang and S. Shan. Review of metamodeling techniques in
support of engineering design optimization. Journal of Mechanical
Design, 129(4):370–380, 2007.
[172] Guangmin Wang, Zhongping Wan, Xianjia Wang, and Yibing Lv.
Genetic algorithm based on simplex method for solving linear-quadratic
bilevel programming problem. Computers & Mathematics with Applications, 56(10):2550–2555, 2008.
[173] Judith Y. T. Wang, Matthias Ehrgott, Kim N. Dirks, and Abhishek
Gupta. A bilevel multi-objective road pricing model for economic,
environmental and health sustainability. Transportation Research
Procedia, 3:393–402, 2014.
[174] Y. Wang, Y. C. Jiao, and H. Li. An evolutionary algorithm for solving
nonlinear bilevel programming based on a new constraint-handling
scheme. IEEE Transactions on Systems, Man, and Cybernetics, Part
C: Applications and Reviews, 32(2):221–232, 2005.
[175] Yuping Wang, Hong Li, and Chuangyin Dang. A new evolutionary
algorithm for a class of nonlinear bilevel programming problems and
[176]
[177]
[178]
[179]
[180]
[181]
[182]
[183]
[184]
[185]
[186]
[187]
[188]
[189]
[190]
[191]
[192]
its global convergence. INFORMS Journal on Computing, 23(4):618–
629, 2011.
L. Wein. Homeland Security: From Mathematical Models to Policy
Implementation: The 2008 Philip McCord Morse Lecture. Operations
Research, 57(4):801–811, 2009.
U. Wen and S. Hsu. Linear bi-level programming problems - a review.
Journal of the Operational Research Society, 42:125–133, 1991.
D. White and G. Anandalingam. A penalty function approach for
solving bi-level linear programs. Journal of Global Optimization,
3:397–419, 1993.
Gerald Whittaker, Rolf Färe, Shawna Grosskopf, Bradley Barnhart,
Moriah Bostian, George Mueller-Warrant, and Stephen Griffith. Spatial targeting of agri-environmental policy using bilevel evolutionary
optimization. Omega, 2016.
Wolfram Wiesemann, Angelos Tsoukalas, Polyxeni-Margarita Kleniati,
and Berç Rustem. Pessimistic bilevel optimization. SIAM Journal on
Optimization, 23(1):353–380, 2013.
Jiuping Xu, Yan Tu, and Ziqiang Zeng. A nonlinear multiobjective
bilevel model for minimum cost network flow problem in a large-scale
construction project. Mathematical Problems in Engineering, 2012,
2012.
Qing Xu, Dao-li Zhu, and Shan-liang Li. The supply chain optimal contract design under asymmetrical information [j]. Systems EngineeringTheory & Practice, 4:003, 2007.
Tadashi Yamada, Bona Frazila Russ, Jun Castro, and Eiichi Taniguchi.
Designing multimodal freight transport networks: a heuristic approach
and applications. Transportation Science, 43(2):129–143, 2009.
Hai Yang and Michael G. H. Bell. Models and algorithms for road
network design: a review and some new developments. Transport
Reviews, 18(3):257–278, 1998.
Jane J Ye. Necessary optimality conditions for multiobjective bilevel
programs. Mathematics of Operations Research, 36(1):165–184, 2011.
Jane J. Ye and Daoli Zhu. New necessary optimality conditions
for bilevel programs by combining the mpec and value function
approaches. SIAM Journal on Optimization, 20(4):1885–1905, 2010.
Y. Yin. Genetic algorithm based approach for bilevel programming
models. Journal of Transportation Engineering, 126(2):115–120, 2000.
Yafeng Yin. Genetic-algorithms-based approach for bilevel programming models. Journal of Transportation Engineering, 126(2):115–120,
2000.
Yafeng Yin. Multiobjective bilevel optimization for transportation planning and management problems. Journal of advanced transportation,
36(1):93–105, 2002.
G. Zhang, J. Liu, and T. Dillon. Decentralized multi-objective bilevel
decision making with fuzzy demands. Knowledge-Based Systems,
20:495–507, 2007.
T. Zhang, T. Hu, Y. Zheng, and X. Guo. An improved particle swarm
optimization for solving bilevel multiobjective programming problem.
Journal of Applied Mathematics, 2012.
Xiaobo Zhu, Qian Yu, and Xianjia Wang. A hybrid differential
evolution algorithm for solving nonlinear bilevel programming with
linear constraints. In Cognitive Informatics, 2006. ICCI 2006. 5th IEEE
International Conference on, volume 1, pages 126–131. IEEE, 2006.
| 9cs.NE
|
CBMM Memo No. 21
June 14, 2014
Neural tuning size is a key factor underlying
holistic face processing
by
Cheston Tan and Tomaso Poggio
Abstract: Faces are a class of visual stimuli with unique significance, for a variety of reasons.
They are ubiquitous throughout the course of a person’s life, and face recognition is crucial for
daily social interaction. Faces are also unlike any other stimulus class in terms of certain
physical stimulus characteristics. Furthermore, faces have been empirically found to elicit
certain characteristic behavioral phenomena, which are widely held to be evidence of “holistic”
processing of faces. However, little is known about the neural mechanisms underlying such
holistic face processing. In other words, for the processing of faces by the primate visual
system, the input and output characteristics are relatively well known, but the internal neural
computations are not. The main aim of this work is to further the fundamental understanding
of what causes the visual processing of faces to be different from that of objects. In this
computational modeling work, we show that a single factor – “neural tuning size” – is able to
account for three key phenomena that are characteristic of face processing, namely the
Composite Face Effect (CFE), Face Inversion Effect (FIE) and Whole‐Part Effect (WPE). Our
computational proof‐of‐principle provides specific neural tuning properties that correspond to
the poorly‐understood notion of holistic face processing, and connects these neural properties
to psychophysical behavior. Overall, our work provides a unified and parsimonious theoretical
account for the disparate empirical data on face‐specific processing, deepening the
fundamental understanding of face processing.
This work was supported by the Center for Brains, Minds and Machines (CBMM), funded by
NSF STC award CCF‐1231216.
Neural tuning size is a key factor underlying holistic face processing
Authors: Cheston Tana,b,c, Tomaso Poggioa,b
Author affiliations:
a
McGovern Institute for Brain Research, MIT, Cambridge, MA 02139, USA.
b
Department of Brain and Cognitive Sciences, MIT, Cambridge, MA 02139, USA.
c
Institute for Infocomm Research, Singapore 138632, Singapore.
Corresponding author:
Name: Cheston Tan
Address: 1 Fusionopolis Way, #21-01 Connexis (South Tower), Singapore 138632.
Phone: +65 6408 2561
E-mail: [email protected]
Note: this technical report supersedes and replaces a previous report (40): MIT CSAIL
Technical Report 2013-004 / CBCL Memo 311. URI: http://hdl.handle.net/1721.1/77936
1
Introduction
Faces are an important class of visual stimuli with unique significance, and face
processing is a longstanding topic of active study within neuroscience (1, 2), cognitive
psychology (3-5) and artificial intelligence (6-8). Faces are ubiquitous throughout the
course of a person’s life, and face recognition is important for daily social interaction.
Faces are also unlike any other stimulus class in terms of certain defining physical
stimulus characteristics, such as having a very specific configuration of internal parts.
Furthermore, faces have been empirically found to elicit certain characteristic behavioral
phenomena (e.g. the Composite Face Effect), which are widely held to be evidence of
“holistic” processing of faces. However, little is known about the neural mechanisms
underlying such holistic face processing. In other words, for the processing of faces by
the primate visual system, the input and output characteristics are relatively well known
(or are straightforward to measure), as are the neural correlates (1, 4, 9), but the actual
neural computations are not known.
Here, we investigate the nature of the difference between face and (non-face)
object processing at the level of the underlying computational mechanisms, in order to
further the fundamental understanding of what causes the visual processing of faces to be
different from that of objects. In this computational modeling work, we show that a single
factor – “neural tuning size” – is able to account for three key phenomena that are
characteristic of face processing, namely the Composite Face Effect (CFE), Face
Inversion Effect (FIE) and Whole-Part Effect (WPE). Our computational proof-of2
principle provides specific neural tuning properties that correspond to the poorlyunderstood notion of “holistic” face processing, and shows the connection between these
neural properties and empirical psychophysical behavior. Our work predicts that such
tuning properties will be found only in neurons associated with holistic face processing,
e.g. in the macaque Fusiform Face Area (FFA) but not the Occipital Face Area (OFA).
Overall, our results provide a unified and parsimonious theoretical account for the
disparate empirical data on face-specific processing.
Results
Building upon the family of simple, biologically-plausible visual recognition
models (10-14), we found that a single parameter determines whether processing is “facelike” or “object-like”, as gauged by three important face-specific phenomena. The first,
the Composite Face Effect (CFE) (15), is the phenomenon whereby two identical top
halves are sometimes incorrectly perceived as different, when paired with different
bottom halves (Fig. 1a). This effect is ostensibly due to the top and bottom halves of each
composite being perceived “holistically” (together as a whole) when aligned, despite
instructions to ignore the bottom halves. Perception is more accurate when the halves are
misaligned (Fig. 1b). Crucially, this effect occurs only for faces, and is therefore
commonly taken as evidence that face and object processing are qualitatively different
(16-19).
We probed the minimal conditions required to produce – or abolish – such holistic
face processing, and found that a vital factor is the size of the template that specifies the
3
tuning profile and optimal stimulus of each neuron (henceforth termed “tuning size”).
Tuning size is defined in terms of proportion of a whole face covered by a template (see
Methods section), not in terms of number of pixels or degrees of visual angle. When
tuning size is large, even without encompassing the whole face, the Composite Face
Effect is found (Fig. 1c). A single change – reduction of tuning size – abolishes the
Composite Face Effect, i.e. leads to “object-like” processing of faces.
“Holism” is a psychological construct with multiple interpretations and putative
mechanisms (19, 21, 22) for which a consensus has yet to emerge. Our simulation results
argue in favor of one particular interpretation of holism, that it is simply the byproduct of
having large tuning size (18, 23, 24). The Composite Face Effect is found using each
individual model neuron with large tuning size by itself (Fig. 1c inset), even though
tuning size is less than half the whole face, and there is no overlap between receptive
fields to speak of. Conversely, even though the set of neurons with small tuning size
collectively span the whole face, they do not produce the CFE (Fig. 1c).
Since there is nothing qualitatively “whole”, “singular”, “unified”, “global” or
“non-decomposable” about processing that uses large tuning size rather than small, the
term “holistic” may be somewhat misleading (to the extent that it implies a qualitative
difference, an absolute whole, or integration into a single representation). Our results do,
however, indicate that holism is not an all-or-none phenomenon, but it is one that can
vary continuously depending on tuning size (and be modulated by other factors).
In our simulations, tuning size is the sole change between the two conditions
depicted in Fig. 1c (large and small tuning size), which suggests that decisional and
4
attentional factors are not key. Rather, what matters is the amount of “perceptual
integration”, as controlled by tuning size. Additionally, while detection and segmentation
are important processes for accurate face recognition, the absence of explicit mechanisms
for these in our simulations suggest that they are also not key factors relating to holism.
Tuning size also accounts for another key face-specific phenomenon, the Face
Inversion Effect (FIE), whereby upside-down inversion disrupts face processing
significantly more than object processing (25). We found that when tuning size is
reduced, the behavioral effect of inversion is also reduced (Fig. 2), akin to face
processing becoming “object-like”. Our simulations also show that inversion reduces the
mean response of each individual neuron (Fig. 3a), illustrating the neural basis of the
behavioral Face Inversion Effect (26). More importantly, tuning size also controls the
magnitude of the neural FIE (Fig. 3b). Even after taking into account the coverage areas
corresponding to different tuning sizes (Figs. 2e, f) or taking into account the mean
response in the upright condition (Figs. 3c, d), large tuning size still produces
significantly larger behavioral and neural FIEs than small tuning size (see Methods
section for details).
Finally, tuning size also accounts for the Whole-Part Effect (27), a “goldstandard” test of holistic processing (16), like the CFE. In the Whole-Part Effect (WPE),
subjects are first presented a study face (Fig. 4a) to memorize. At test, in the Whole
condition, subjects are presented with two faces (Fig. 4b) that differ only in a localized
region (e.g. eye region), and have to recall and choose the study face. In the Part
condition, the differing localized regions in the Whole condition are cropped (Fig. 4c)
5
and presented instead. Empirically, subjects are significantly more accurate in the Whole
than Part condition. This effect is highly significant for faces, but non-significant or
significantly smaller for non-faces (16). Our simulations show that reduction in tuning
size results in a significantly smaller WPE (Fig. 4e), akin to a shift from “face-like” to
more “object-like” processing.
Discussion
Clearly, our simulations do not capture the full complexity of face processing,
including aspects such as gaze, expression and attentional capture. Also, tuning size is not
necessarily the only difference between face and object processing mechanisms (28-30).
However, we have shown that a change in tuning size alone can account for three
important phenomena commonly thought to be characteristic of face processing,
suggesting that tuning size is a key factor underlying holism.
By changing only tuning size and keeping everything else unchanged (Figs. 1-4),
our simulations are able to sidestep a confound that is unavoidable for empirical
experiments that investigate mechanisms underlying face versus object processing – the
confound of different stimuli. Empirically, face and object stimuli elicit measurable
differences, but do these stem from differences in physical stimulus properties, or from
differences in processing mechanisms? Here, instead of changing the stimuli to produce
measurable differences, we changed only the underlying processing but not the stimuli
(we used only faces throughout), so this confound is avoided.
6
Our results suggest that all three effects stem from a common cause: large tuning
size. This makes intuitive sense for a variety of reasons. Large tuning size is consistent
with the existence of face-selective neurons being tuned to multiple face parts, but not
necessarily the entire face (31, 32). Face recognition algorithms that use features with
such characteristics (large but not necessarily whole-face tuning size) show excellent
performance (6). While face processing in general has a significant innate component
(33-35), visual deprivation during infancy abolishes the Composite Face Effect (36). This
suggests that large tuning size may result from a combination of innate factors and
postnatal visual experience.
Other than large tuning size, another interpretation of holistic processing is that of
“non-linear interaction of parts” (18, 37). Since our model is highly non-linear, our
results and this interpretation are not mutually exclusive. Our results do not preclude
certain non-linearities from also being key factors underlying holistic processing, nor do
they deny that non-linear interactions occur empirically.
Apart from “holistic” processing, faces are also thought to elicit “configural”
processing, but the relationship between these notions is unclear (16, 19, 38). For
instance, the Face Inversion Effect has sometimes been associated with configural rather
than holistic processing of faces. Our results suggest a link between these two concepts,
through the common factor of large tuning size. Because neurons with large tuning size
cover more than individual face parts, they are more sensitive to the configuration of
multiple parts, which is altered by inversion. For this same reason, these neurons are also
more sensitive to information that comes from more distant regions of the face image (in
7
the case of the Composite Face Effect and Whole-Part Effect). The notion of large tuning
size may also be able to account for sensitivity to spacing between face parts. The idea is
that since each neuron’s tuning is specified by a certain face template, any deviation from
that template, such as changing the distance between the eyes, will reduce the neural
response. Note, however, that these neurons are not necessarily more sensitive to spacing
changes than shape changes, consistent with (39).
Methods
Model. The HMAX model (10) simulates hierarchical processing in primate visual
cortex, reflecting the increase in neural tuning complexity and invariance up the
hierarchy. The lowest levels correspond to orientation-selective cells in primary visual
cortex, while the highest levels correspond to face-selective and object-selective cells in
inferotemporal cortex.
We used the model implementation found at http://cbcl.mit.edu/jmutch/cns/. Of
the four model layers, the orientation-selective lower two layers (S1 and C1) contain
model neurons tuned to Gabor patches of various orientations and spatial frequencies; the
parameters have been pre-determined based on prior electrophysiological data. The upper
two layers (S2 and C2) contain model neurons that are tuned during an unsupervised
template-learning process, performed prior to normal model operation. Template-learning
simply involves storing “snapshots” of C1 activity produced in response to some set of
training images. Subsequently, in normal model operation, these snapshots act as
8
templates that new images are matched against. The S2 layer comprises the output of this
template-matching process, and pooling of the S2 model neurons over all image positions
and scales (to produce invariance to these) produces the C2 layer. If training images
consist of faces, then S2 and C2 model neurons are face-selective. All simulations used
1000 C2 model neurons.
Tuning size. The critical independent variable is “tuning size”. Large, medium and small
tuning sizes correspond respectively to S2 tuning templates covering 12x12, 8x8 and 4x4
C1 model neurons, all from the relatively coarse scale 7 (out of 9). At this scale, the
entire face oval corresponds to 17x22 C1 neurons, so each small template is roughly the
size of a face part (e.g. eye, nose), while each large template covers multiple face parts
but not the whole face.
Importantly, “tuning size” is defined as the proportion of a whole face covered by
a template. This is not the same as “size” defined in terms of number of pixels or degrees
of visual angle. In the human and primate visual systems (as well as our model), there
exists some invariance to image scale. Therefore, a particular tuning size (e.g. half a face)
can correspond to a range of physical sizes (in pixels or degrees of visual angle).
Since each template was learnt from a different (random) part of a training image,
even the 1000 small templates (collectively) spanned the whole face – yet they did not
produce a Composite Face Effect (Fig. 1c), thus ruling out some alternative accounts of
mechanisms underlying “holistic processing”.
9
Stimuli. Stimuli were derived from 100 frontal-view male faces from the MPI database
(http://faces.kyb.tuebingen.mpg.de/). Faces were downscaled by 25%, and then ovalcropped to remove outline and external features (e.g. hair). Faces were normalized so that
all had the same pixel-value statistics (mean and variance). Odd-numbered faces were
used for template-learning, even-numbered faces for normal operation. All faces were
upright unless explicitly inverted. Note: in Figs. 1, 2 and 4, backgrounds were cropped
slightly to save space and make face details more apparent.
Composite Face Effect (CFE). Composites were constructed by pairing the top of one
face with the bottom of a different face (with a two-pixel gap). Twenty faces were used;
these were chosen prior to all simulations, for behavioral replication of the CFE in human
subjects (not related to this modeling study).
On each trial, two composites are presented, and their top halves are judged to be
same or different, ignoring the bottom halves. Only trials with identical top halves are
analyzed (16). The Composite Face Effect is defined as a higher hit-rate (i.e. accuracy on
these “same” trials) for misaligned than aligned composites.
To simulate human subjects looking and attending to the top halves, bottom-half
pixel values are multiplied by 0.1, and faces shifted downwards so that the top halves
occupy the center. To simulate subjects comparing composites, if the dissimilarity
between composites (Euclidean distance between the two sets of C2 layer responses) is
below some threshold, the composites are considered “same”. For each condition (e.g.
10
small tuning size), the threshold is independently set so that the aligned, upright hit-rate is
as close to 75% as possible, but results are qualitatively robust to the threshold used.
Fig. 1c: p-values and error bars computed using 1,000 runs of non-parametric
bootstrap (N=20 pairs of composites). Fig. 1c inset: p-values computed using nonparametric Wilcoxon signed-rank test (N=1000 independent C2 features).
Face Inversion Effect. Dissimilarity between two images was defined as the Euclidean
distance between the two sets of C2 layer responses. Fig. 2c shows the mean dissimilarity
between all 1225 (i.e. 50C2) pairs of faces within each condition, averaged over bootstrap
runs. Fig. 2d shows the mean behavioral FIE effect size (i.e. upright – inverted
dissimilarities shown in Fig. 2c) over bootstrap runs. Fig. 2c, d: p-values and error bars
computed using bootstrap runs. In all cases for Fig. 2, 10,000 runs of non-parametric
bootstrap were used.
Figs. 2e and 2f are similar to 2c and 2d, except that for the large and medium
tuning sizes, a random subset of neurons (100 for large, 150 for medium) were used in
each bootstrap run, instead of all 1000 neurons. This was to control for the larger
coverage areas (12x12 vs. 8x8 vs. 4x4 C1 neurons for large, medium and small
respectively, and also to control for the different upright dissimilarities (see black bars in
Fig. 2c); both factors are actually over-compensated for.
Fig. 3a shows the mean response (averaged over all 1000 model neurons) to all 50
faces. Fig. 3b shows the mean neural FIE (i.e. upright – inverted responses shown in Fig.
3a). Fig. 3a, b: p-values computed using non-parametric Wilcoxon signed-rank test.
11
Figs. 3c and 3d were similar to Figs. 3a and 3b, except that only neurons with a
mean response between 0.75 and 0.80 were considered (to avoid ceiling and floor
effects), after which neurons were then randomly chosen so that all conditions (i.e. tuning
sizes) ultimately used the same number of neurons. This was to control for the slightly
different upright responses in Fig. 3a (black bars).
Whole-Part Effect (WPE). Stimuli were constructed by blending the eye region of one
face with the rest of a different face. Stimuli in the Part condition (Fig. 4c) were produced
by cropping out the eye region of stimuli in the Whole condition (Fig. 4b). All possible
trials (3-way combinations of first eye region, second eye region, rest of face) for twenty
original faces (same faces as for CFE) were tested.
The Whole-Part Effect is defined as a higher accuracy for the 2-alternativeforced-choice task in the Whole condition than in the Part condition. To simulate human
subjects looking and attending at the eye regions of the test faces, non-eye-region pixel
values are multiplied by 0.5, and faces shifted downwards so that the eye regions occupy
the center. To simulate subjects choosing between the two test stimuli, the one with the
smaller dissimilarity (Euclidean distance between the C2 layer responses) to the study
stimulus (Fig. 4a) was picked. Fig. 4d, e: p-values and error bars computed using 1,000
runs of non-parametric bootstrap.
12
Acknowledgements
We thank G. Cottrell, R. Desimone, J. DiCarlo, I. Gauthier, N. Kanwisher, J. Leibo, D.
Sheinberg, P. Sinha and the Perceptual Expertise Network for helpful discussions. This
material is based upon work supported by the Center for Brains, Minds and Machines
(CBMM), funded by NSF STC award CCF-1231216. This work was also supported by
A*STAR Joint Council Office (JCO) Visiting Investigatorship Programme (VIP) grant
#1335h00098. This work was also supported by DARPA (IPTO and DSO), National
Science Foundation (NSF-0640097, NSF-0827427), AFSOR-THRL (FA8650-05-C7262). Additional support (T. P.) was provided by: Adobe, Honda Research Institute
USA, KAUST, NEC, Sony and the Eugene McDermott Foundation. C. T. was supported
by fellowships from A*STAR and the McGovern Institute for Brain Research.
13
References
1.
Tsao DY, Freiwald WA, Tootell RBH, Livingstone MS (2006) A cortical region
consisting entirely of face-selective cells. Science 311(5761):670-674.
2.
Desimone R, Albright TD, Gross, CG, Bruce C (1984) Stimulus-selective properties
of inferior temporal neurons in the macaque. J Neurosci 4(8):2051-2062.
3.
Rossion B, et al. (2000) The N170 occipito-temporal component is delayed and
enhanced to inverted faces but not to inverted objects: an electrophysiological
account of face-specific processes in the human brain. Neuroreport 11(1):69-74.
4.
Haxby JV, et al. (1996) Face encoding and recognition in the human brain. Proc
Natl Acad Sci USA 93(2):922-927.
5.
Pitcher D, Walsh V, Yovel G, Duchaine B (2007) TMS evidence for the involvement
of the right occipital face area in early face processing. Curr Biol 17(18):1568-1573.
6.
Viola P, Jones MJ (2004) Robust real-time face detection. Int J Comput Vision
57(2):137-154.
7.
Brunelli R, Poggio T (1993) Face recognition: features versus templates. IEEE
Trans. Pattern Anal. Mach. Intell. 15(10):1042-1052.
8.
Turk M, Pentland A (1991) Eigenfaces for recognition. J Cogn Neurosci 3(1):71-86.
9.
Kanwisher N, McDermott J, Chun MM (1997) The fusiform face area: a module in
human extrastriate cortex specialized for face perception. J Neurosci 17(11):43024311.
10. Riesenhuber M, Poggio T (1999) Hierarchical models of object recognition in
cortex. Nature Neurosci 2(11):1019-1025.
14
11. Cottrell GW, Branson KM, Calder AJ (2002) in Proceedings of the 24th Annual
Conference of the Cognitive Science Society, (Cognitive Science Society, Mahwah,
NJ) pp 238-243.
12. Fukushima K (1980) Neocognitron: A self-organizing neural network model for a
mechanism of pattern recognition unaffected by shift in position. Biol Cybern
36(4):193-202.
13. Perrett DI, Oram MW (1993) Neurophysiology of shape processing. Image Vision
Comput 11(6):317-333.
14. Wallis G, Rolls ET (1997) Invariant face and object recognition in the visual system.
Prog Neurobiol 51(2):167-194.
15. Young AW, Hellawell D, Hay DC (1987) Configurational information in face
perception. Perception 16(6):747-759.
16. McKone E, Robbins R (2011) in The Oxford Handbook of Face Perception, eds
Calder AJ, Rhodes G, Johnson MH, Haxby JV (Oxford Univ. Press, New York,
NY), pp 149-176.
17. Tanaka JW, Gordon I (2011) in The Oxford Handbook of Face Perception, eds
Calder AJ, Rhodes G, Johnson MH, Haxby JV (Oxford Univ. Press, New York, NY)
pp 177-194.
18. Tsao DY, Livingstone MS (2008) Mechanisms of face perception. Annu Rev
Neurosci 31:411-437.
19. Richler JJ, Palmeri TJ, Gauthier I (2012) Meanings, mechanisms, and measures of
holistic processing. Front Psychol, 3:553.
15
20. McKone E, et al. (2013) Importance of the inverted control in measuring holistic
face processing with the composite effect and part-whole effect. Front Psychol 4:33.
21. Maurer D, Le Grand R, Mondloch CJ (2002) The many faces of configural
processing. Trends Cogn Sci 6(6):255-260.
22. Piepers DW, Robbins RA (2012) A review and clarification of the terms “holistic”,
“configural” and “relational” in the face perception literature. Front Psychol, 3:559.
23. Rossion B, Gauthier I (2002) How does the brain process upright and inverted faces?
Behav Cogn Neurosci Rev 1(1):63-75.
24. Farah MJ, Wilson KD, Drain M, Tanaka JN (1998) What is “special” about face
perception? Psychol Rev 105(3):482-498.
25. Yin RK (1969) Looking at upside-down faces. J Exp Psychol 81(1):141-145.
26. Yovel G, Kanwisher N (2005) The neural basis of the behavioral face-inversion
effect. Curr Biol 15(24):2256-2262.
27. Tanaka JW, Farah MJ (1993) Parts and wholes in face recognition. Quart J Exp
Psych 46A:225-245.
28. Wallis G (2013) Toward a unified model of face and object recognition in the human
visual system. Front Psychol 4:497.
29. Riesenhuber M, Wolff BS (2009) Task effect, performance levels, features,
configurations, and holistic face processing: a reply to Rossion. Acta Psychol
132(3):286-292.
30. Wallraven C, Schwaninger A, Bulthoff HH (2005) Learning from humans:
computational modeling of face recognition. Network: Comp Neural Sys 16:401-418
16
31. Freiwald WA, Tsao DY, Livingstone MS (2009) A face feature space in the
macaque temporal lobe. Nature Neurosci 12(9):1187-1196.
32. Perrett DI, Rolls ET, Caan W (1982) Visual neurones responsive to faces in the
monkey temporal cortex. Exp Brain Res 47(3):329-342.
33. Sugita Y (2008) Face perception in monkeys reared with no exposure to faces. Proc
Natl Acad Sci USA 105(1):394-398.
34. Polk TA, Park J, Smith MR, Park DC (2007) Nature versus nurture in ventral visual
cortex: a functional magnetic resonance imaging study of twins. J Neurosci
27(51):13921-13925.
35. Wilmer JB, et al. (2010) Human face recognition ability is specific and highly
heritable. Proc Natl Acad Sci USA 107(11):5238-5241.
36. Le Grand R, Mondloch CJ, Maurer D, Brent HP (2004) Impairment in holistic face
processing following early visual deprivation. Psychol Sci 15(11):762-768.
37. Boremanse A, Norcia AM, Rossion B (2013) An objective signature for visual
binding of face parts in the human brain. J Vis 13(11):1-18.
38. Rossion B (2008) Picture-plane inversion leads to qualitative changes of face
perception. Acta Psychol 128(2):274-289.
39. McKone E, Yovel G (2009) Why does picture-plane inversion sometimes dissociate
perception of features and spacing in faces, and sometimes not? Toward a new
theory of holistic processing. Psychon. B. Rev. 16(5):778-797.
40. Tan C, Poggio T (2013) Faces as a "Model Category" for Visual Object Recognition.
MIT-CSAIL-TR-2013-004 / CBCL-311. URI: http://hdl.handle.net/1721.1/77936
17
Figure Legends
Figure 1. Tuning size determines whether processing is “face-like” or “object-like”,
as gauged by the Composite Face Effect (CFE). a, Aligned composite faces. Top
halves are identical, while bottom halves are different. People sometimes incorrectly
perceive the two identical top halves as different. b, Misaligned composite faces. Human
judgement of the top halves (as being identical) is significantly more accurate for
misaligned than aligned composites (15, 16). This is the CFE, which is only found for
faces. c, Simulations show that the CFE is produced by neurons with large – but not
small – tuning size. Large: p<0.001, small: p=0.85. No CFE was found in either condition
for inverted images (20); large: p=0.39, small: p=0.81. Inset: each individual neuron with
large tuning size can produce the CFE by itself (p<0.001, both with and without
controlling for inverted images). Error bars: SEM.
Figure 2. Tuning size accounts for the behavioral Face Inversion Effect (FIE). a, b,
Illustration of the FIE: dissimilarity between faces is more apparent for upright than
inverted faces (23, 25). c, Simulations show that decrease in dissimilarity (as a result of
inversion) varies with tuning size. d, FIE effect size (upright dissimilarity – inverted
dissimilarity) varies with tuning size. Large vs. medium: p<0.0001. Medium vs. small:
p<0.0001. Neurons with small tuning size show “object-like” processing, i.e. a minimal
inversion effect. e, Dissimilarity after controlling for coverage area of different tuning
18
sizes (see Methods). f, FIE effect size, controlling for coverage area. Large vs. medium:
p=0.023. Medium vs. small: p=0.012. Large vs. small: p=0.0002. Error bars: SEM.
Figure 3. Tuning size accounts for the neural Face Inversion Effect (FIE). a, Mean
individual neuron responses to single faces (as opposed to dissimilarities between pairs of
faces, in Fig. 2) for upright vs. inverted faces. b, FIE effect size (upright response –
inverted response) varies with tuning size. Large vs. medium: p<0.0001. Medium vs.
small: p<0.0001. c, Mean individual neuron responses, after controlling for response in
the upright condition (see Methods). d, FIE effect size, after controlling for response in
the upright condition. Large vs. medium: p<0.0001. Medium vs. small: p<0.0001. Error
bars: SEM.
Figure 4. Tuning size accounts for the Whole-Part Effect (WPE). a, Study face. b,
Whole condition: two choices (differing only in the eye region) are presented, to be
matched to the study face from memory. c, Part condition: two choices (eye regions
cropped from Whole condition) to be matched to the study face. The WPE is defined as a
significantly higher accuracy for the Whole condition than the Part condition (16). d,
Simulation results. Large: p<0.001, small: p<0.001. e, The WPE effect size (Whole
accuracy – Part accuracy) varies with tuning size. Large vs. small: p<0.001. This remains
significant after subtracting effect sizes for inverted images (20); large vs. small:
p<0.001. Error bars: SEM.
19
Figure 1. Tuning size determines whether processing is “face-like” or “object-like”,
as gauged by the Composite Face Effect (CFE).
20
Figure 2. Tuning size accounts for the behavioral Face Inversion Effect (FIE).
21
Figure 3. Tuning size accounts for the neural Face Inversion Effect (FIE).
22
Figure 4. Tuning size accounts for the Whole-Part Effect (WPE).
23
| 9cs.NE
|
arXiv:1701.03856v1 [cs.DM] 14 Jan 2017
The flip Markov chain
for connected regular graphs
∗†
Colin Cooper‡
Martin Dyer‡
Department of Computer Science
King’s College London
London WC2R 2LS, U.K.
School of Computing
University of Leeds
Leeds LS2 9JT, U.K.
[email protected]
[email protected]
Catherine Greenhill§
Andrew Handley¶
School of Mathematics and Statistics
UNSW Australia
Sydney, NSW 2052, Australia
School of Computing
University of Leeds
Leeds LS2 9JT, U.K.
[email protected]
[email protected]
14 January 2017
Abstract
Mahlmann and Schindelhauer (2005) defined a Markov chain which they called kFlipper, and showed that it is irreducible on the set of all connected regular graphs of a
given degree (at least 3). We study the 1-Flipper chain, which we call the flip chain, and
prove that the flip chain converges rapidly to the uniform distribution over connected
2r-regular graphs with n vertices, where n ≥ 8 and r = r(n) ≥ 2. Formally, we prove
that the distribution of the flip chain will be within ε of uniform in total variation
distance after poly(n, r, log(ε−1 )) steps. This polynomial upper bound on the mixing
time is given explicitly, and improves markedly on a previous bound given by Feder et
al. (2006). We achieve this improvement by using a direct two-stage canonical path
construction, which we define in a general setting.
This work has applications to decentralised networks based on random regular connected graphs of even degree, as a self-stablising protocol in which nodes spontaneously
perform random flips in order to repair the network.
∗
An earlier version of this paper appeared as an extended abstract in PODC 2009 [6].
This research was partly performed while the first three authors were visiting the Simons Institute for the
Theory of Computing.
‡
Supported by EPSRC Research Grant EP/M004953/1.
§
Supported by the Australian Research Council Discovery Project DP140101519.
¶
Supported by EPSRC Research Grant EP/D00232X/1.
†
1
1
Introduction
Markov chains which walk on the set of random regular graphs have been well studied. The
switch chain is a very natural Markov chain for sampling random regular graphs. A transition
of the switch chain is called a switch, in which two edges are deleted and replaced with two
other edges, without changing the degree of any vertex. The switch chain was studied by
Kannan et al. [14] for bipartite graphs, and in [4] for regular graphs. Although a switch
changes only a constant number of edges per transition, it is not a local operation since these
edges may be anywhere in the graph.
Random regular graphs have several properties which make them good candidates for
communications networks: with high probability, a random ∆-regular graph has logarithmic
diameter and high connectivity, if ∆ ≥ 3. (See, for example, [20].) Bourassa and Holt [3]
proposed a protocol for a decentralised communications network based on random regular
graphs of even degree. Since then, Markov chains which sample regular graphs using local
operations have been applied to give self-stabilising or healing protocols for decentralised communications networks which have a regular topology. To help the network to recover from
degradation in performance due to arrivals and departures, clients in a decentralised network can spontaneously perform local transformations to re-randomize the network, thereby
recovering desirable properties.
Mahlmann and Schindelhauer [16] observed that the switch operation is unsuitable for
this purpose, since a badly-chosen switch may disconnect a connected network. Furthermore,
since the switch move is non-local there are implementation issues involved in choosing two
random edges in a decentralised network. (If the network is an expander then this problem
can be solved by performing two short random walks in the network and taking the final edge
of each for the next operation [3].)
To overcome these problems, Mahlmann and Schindelhauer proposed an alternative operation which they called a k-Flipper. When k = 1, the 1-Flipper (which we call the flip)
is simply a restricted switch operation, in which the two edges to be switched must be at
distance one apart in the graph. Figure 1 illustrates a switch and a flip.
a
d
b
c
⇒S
a
d
a
d
b
c
b
c
⇒F
a
d
b
c
Figure 1: Left: a switch operation, denoted by ⇒S . Right: a flip, denoted by ⇒F .
By design, the flip operation preserves the degree of all vertices and preserves connectivity.
Indeed, it is the smallest such randomising operation: no edge exchange on fewer than four
vertices preserves the degree distribution, and the only shallower tree than the 3-path is the
3-star, which has no degree-preserving edge exchange. In order to use flips as a self-stabilising
protocol, each peer waits for a random time period using a Poisson clock, after which it
instigates a new flip. This involves communication only with peers at distance at most two
away in the network.
2
We will show that flips have asymptotic behaviour comparable to switches, in that both
operations randomise a network in polynomial time. Simulations also suggest that they
operate equally fast, up to a constant factor (see [11, Section 2.3.6]). A discussion of a
distributed implementation of the flip operation is given in [6].
Our approach will be to relate the mixing time of the flip chain MF to an intermediate
chain MSC , which is the switch chain restricted to connected graphs. In turn, the mixing
time of the connected switch chain MSC will be related to the mixing time of the (standard)
switch chain MS , using the analysis from [4]. Each of these two steps will be performed
using a technique which we call “two-stage direct canonical path construction”, described in
Section 2. Specifically, we prove the following result. This is a corrected version of the bound
given in [6], which failed on certain rare graphs.
Theorem 1.1. For each n ≥ 8 let r = r(n) ≥ 2 be a positive integer such that n ≥ 2r + 1.
The mixing time of the flip Markov chain MF on the set of 2r-regular graphs with n vertices
is at most
480 (2r)35 n15 2rn log(2rn) + log(ε−1 ) ,
of which the two-stage direct construction is responsible for a factor of 480 (2r)12n7 .
Section 3 relates the flip chain and the connected switch chain, while Section 4 relates
the connected switch chain and the switch chain. The proof of Theorem 1.1 can be found
at the end of Section 4. The rest of this section contains some background material, some
helpful graph theory results, precise definitions of the flip and switch Markov chains, and
some definitions regarding canonical paths.
We remark that a “one-stage direct analysis” of the flip chain seems difficult. Furthermore,
the multicommodity flow defined for the switch chain in [4] does not apply directly to the
connected switch chain, since the flow may pass through disconnected graphs, even when
both end-states are connected. It may be possible to circumvent this problem by using an
alternative multicommodity flow argument, but we do not explore that option here.
1.1
Notation, terminology and history
We will use the notation (ab, cd ⇒ ac, bd) to denote the switch operation as illustrated in
Figure 1.
The switch chain MS applies a random switch at each step, and the flip chain MF applies
a random flip at each step. (We define both chains formally in Section 1.4.) The state space
ΩS for the switch chain is the set of all ∆-regular graphs on the vertex set [n] = {1, 2, . . . , n},
while the state space ΩF for the flip chain is the set of all connected graphs in ΩS .
The switch chain is known to be ergodic, and has uniform stationary distribution on ΩS .
Mahlmann and Schindelhauer [16] showed that the flip chain is ergodic on ΩF . They also
proved that the transitions of the flip chain are symmetric, so that the stationary distribution
of the flip chain is uniform on ΩF . (We use a slightly different transition procedure, described
in Section 1.4, which is also symmetric.)
The mixing time of a Markov chain is a measure of how many steps are required before the
Markov chain has a distribution which is close to stationary. Kannan et al. [14] considered
3
the mixing time of the switch Markov chain on “near regular” bipartite graphs. Cooper et
al. [4] gave a polynomial bound on the mixing time τCDG (ε) of the switch chain for all regular
simple graphs. The result of [4] allows the degree ∆ = ∆(n) to grow arbitrarily with n: it is
not restricted to regular graphs of constant degree.
In Section 3 we extend the result of [4] from switches to flips, using a method which we
call two-stage direct canonical path construction. We restrict our attention to even degree,
since working with an odd degree would force the network size to take only even values. So
for the rest of the paper we assume that ∆ = 2r for some r ≥ 2, unless otherwise stated.
It happens that even-degree graphs have some other desirable properties, as will we see in
Section 1.2, which will simplify parts of our argument. However, we see no reason why the
flip chain would fail to be rapidly mixing for odd degrees.
Mahlmann and Schindelhauer [16] do not provide a bound for the mixing time of MF ,
though they do show [16, Lemma 9] that a similar chain with the flip edges at a distance of
Θ(∆2 n2 log ε−1 ) apart will give an ε-approximate expander in time O(∆n). However, these
moves are highly non-local and therefore are unsuitable for the application to self-stabilisation
of decentralised networks.
An upper bound for the mixing time of MF was given by Feder et al. [9] using a comparison
argument with the switch Markov chain, MS . Applying a comparison argument to relate these
chains is a difficult task, since a switch may disconnect a connected graph and a flip cannot.
Feder et al. [9] solve this difficulty by embedding MS into a restricted switch chain which only
walks on connected graphs. They prove that this can be done in such a way that a path in
MS is only polynomially lengthened. Their result relaxes the bound in [4] by O(∆41 n45 ). By
employing a two-stage direct construction, we give a much tighter result, relaxing the bound
on τCDG (ε) by a factor of O(r 12 n7 ) in the case of 2r-regular graphs with
√ r ≥ 2.
2 2
Recently, Allen-Zhu et al. [1] proved that after at most O(n d log n) steps of the flip
chain, the current graph will be a good expander. This is not directly comparable with the
results here, since it does not address the distribution of the output graph. It appears more
closely related to the result of Mahlmann and Schindelhauer [16, Lemma 9].
1.2
Graph-theoretical preliminaries
For further details see, for example, Diestel [8]. Unless otherwise stated, G denotes a simple
graph G = (V, E) with V = {1, 2, . . . , n}. The set of neighbours in G of a vertex v is denoted
by NG (v).
A vertex set S ⊆ V has edge boundary ∂E S = {uv | u ∈ S, v ∈
/ S}, which contains the
edges crossing from inside to outside of S.
A graph G is 2-edge-connected if it is connected and every edge of G lies in a cycle. More
generally, an edge-cut in G is a set F ⊆ E of edges of G such that the graph G − F (which is
G with the edges in F deleted) is disconnected. If |F | = k then F is a k-edge-cut. If vertices
v, w belong to distinct components of G − F then we say that F separates v and w in G.
A cut vertex in a graph G is a vertex v such that deleting v disconnects the graph.
If p is a path from v to w then we say that p is a (v, w)-path. Further, if p = v · · · a · · · b · · · w
then we write p[a : b] to denote the (inclusive) subpath of p with endvertices a and b.
We make use of the following structural properties of 2r-regular graphs with r ≥ 2.
4
Lemma 1.2. Let r ≥ 2 be an integer and let G = (V, E) be a 2r-regular graph on n ≥
max{8, 2r + 1} vertices. Then G satisfies the following properties:
(i) For any S ⊆ V , the edge boundary ∂E S of S is even.
(ii) Every edge of G lies in a cycle; that is, every connected component of G is 2-edgeconnected.
(iii) G has fewer than n/(2r) connected components.
(iv) Let u, v be distinct vertices in the same connected component of G. The number of
2-edge-cuts in G which separate u and v is at most n2 /(15r 2 ).
Proof. For (i), we argue by contradiction. Assume that there is an odd number k = 2k ′ + 1
of edges incident upon some set S. Let G[S] be the subgraph induced by S. Let |S| = s and
suppose that G[S] has m edges. Denote the sum of degrees of the vertices in S by dG (S).
Clearly dG (S) = 2rs. Exactly k edges incident with S have precisely one endvertex in S, so
2m = dG[S] (S) = 2rs − k = 2rs − 2k ′ − 1 = 2(rs − k ′ ) − 1,
but no integer m satisfies this equation. Hence our assumption is false, proving (i).
For (ii), it suffices to observe that every component of G is Eulerian, as all vertices have
even degree. Next, (iii) follows since the smallest 2r-regular graph is K2r+1 , and G can have
at most n/(2r + 1) < n/(2r) components this small. The proof of (iv) is deferred to the
Appendix, as it is somewhat lengthy.
1.3
Canonical paths
We now introduce some Markov chain terminology and describe the canonical path method.
For further details, see for example Jerrum [12] or Sinclair [17].
Let M be a Markov chain with finite state space Ω and transition matrix P . We suppose
that M is ergodic and has unique stationary distribution π. The graph underlying the Markov
chain M is given by G = (Ω, E(M)) where the edge set of G corresponds to (non-loop)
transitions of M; specifically,
E(M) = {xy | x, y ∈ Ω, x 6= y, P (x, y) > 0}
We emphasise that G has no loops.
The mixing time τ (ε) of a Markov chain is given by
τ (ε) = max min {T ≥ 0 | dTV (Pxt , π) ≤ ε for all t ≥ T },
x∈X
where Pxt is the distribution of the random state Xt of the Markov chain after t steps with
initial state x, and dTV is the total variation distance between two probability distributions.
Let the eigenvalues of P be
1 = λ0 > λ1 ≥ λ2 · · · ≥ λN −1 ≥ −1.
5
Then the mixing time satisfies
τ (ε) ≤ (1 − λ∗ )−1 ln(1/π ∗) + ln(ε−1 )
(1.1)
where λ∗ = max{λ1 , |λN −1 |} and π ∗ = minx∈Ω π(x) is the smallest stationary probability.
(See [17, Proposition 1].)
The canonical path method, introduced by Jerrum and Sinclair [13], gives a method for
bounding λ1 . Given any two states x, y ∈ Ω, a “canonical” directed path γxy is defined from
x to y in the graph G underlying the Markov chain, so that each step in the path corresponds
to a transition of the chain. The congestion ρ̄(Γ) of the set Γ = {γxy | x, y ∈ Ω} of canonical
paths is given by
X
π(x)π(y)|γxy |
ρ̄(Γ) = max Q(e)−1
e∈E(M)
x,y∈Ω
e∈γxy
where Q(e) = π(u) P (u, v) when e = uv ∈ E(M). If the congestion along each transition is
low then the chain should mix rapidly. This can be made precise as follows: if λ∗ = λ1 then
(1 − λ1 )−1 ≤ ρ̄(Γ).
(1.2)
(See [17, Theorem 5] or [12] for more details.) In many applications the chain M is made
lazy by replacing its transition matrix P by (I + P )/2: this ensures that the chain has no
negative eigenvalues and hence that λ∗ = λ1 . However, in many cases it is easy to see that
(1 + λN −1 )−1 is smaller than the best-known upper bound on (1 −λ1 )−1 , such as that provided
by the canonical path method (1.2). In this situation, combining (1.1) and (1.2) gives an upper
bound on the mixing time without the need to make the chain lazy.
The multicommodity flow method is a generalisation of the canonical path method. Rather
than send π(x)π(y) units down a single “canonical” path from x to y, a set Pxy of paths is
defined, and the flow from x to y is divided among them. The congestion can be defined
similarly, and low congestion leads to rapid mixing: see Sinclair [17].
Cooper et al. [4] gave a multicommodity flow argument to bound the mixing time of the
switch chain MS . For every pair of distinct states x, y, a set of paths from x to y was defined,
one for each vector (φv )v∈[n] , where φv is a “pairing” of the edges of the symmetric difference
of x and y incident with the vertex v. Then the flow between x and y was shared equally
among all these paths. In our analysis below, we assume that a pairing has been fixed for each
distinct pair of states x, y, leading to one canonical path γxy between them. We will work
with the set of these canonical paths for the switch chain, ΓS = {γxy }. Since our congestion
bounds will be independent on the choice of pairing, the same bound would hold if the original
multicommodity flow from [4] was used.
1.4
The switch and flip Markov chains
Given ∆ ≥ 3 and n ≥ ∆ + 1, the switch chain MS has state space ΩS given by the set of all
d-regular graphs on n vertices. The flip chain MF has state space ΩF given by the set of all
connected ∆-regular graphs on n vertices. We restrict our attention to regular graphs of even
degree, setting ∆ = 2r for some integer r ≥ 2.
6
Let PS (respectively, PF ) denote the transition matrix of the switch (respectively, flip)
Markov chain. The stationary distribution of the switch (respectively, flip) chain is denoted
by πS (respectively, πF ). The graph underlying the Markov chain will be denoted by GS =
(ΩS , E(MS)) and GF = (ΩF , E(MF )), respectively.
The transition procedure for the switch Markov chains is given in Figure 2.
From G ∈ ΩS do
choose non-adjacent distinct edges ab, cd, u.a.r.,
choose a perfect matching M of {a, b, c, d} u.a.r.,
if M ∩ E = ∅ then
delete the edges ab, cd and add the edges of M,
otherwise
do nothing
end if;
end;
Figure 2: The switch Markov chain MS .
The switch chain is irreducible [15, 18]. Furthermore, as PS (x, x) ≥ 1/3 for all x ∈ ΩS ,
the switch chain is aperiodic, and hence ergodic. Using an approach of Diaconis and SaloffCoste [7, p. 702], this also implies that the smallest eigenvalue λN −1 of MS satisfies
(1 + λN −1 )−1 ≤
1
max P (x, x)−1
2 x∈Ω
S
≤ 3/2.
Since the upper bound on (1 − λ1 )−1 proved in [4] is (much) larger than constant, this proves
that we can apply the canonical path method to the switch chain without making the switch
chain lazy.
If x, y are distinct elements of ΩS which differ by a switch then PS (x, y) = 1/(3an,2r ) where
(see [4])
2r
rn
,
(1.3)
−n
an,2r =
2
2
Hence the transition matrix PS is symmetric and the stationary distribution of the switch
chain is uniform over ΩS .
Figure 3 gives the transition procedure of the flip chain. Here we start from a randomly
chosen vertex a and perform a random walk from a of length 3, allowing backtracking. There
are (2r)3 n choices for the resulting walk (a, b, c, d), and we choose one of these uniformly at
random.
Mahlmann and Schindelhauer proved that the flip chain is irreducible on ΩF [16]. For all
states x ∈ ΩF we have P (x, x) ≥ 1/(2r), since in particular the proposed transition is rejected
when c = a. Therefore the flip chain is aperiodic, and hence ergodic. Again, by [7, p. 702],
this implies that the smallest eigenvalue λN −1 of MF satisfies (1 + λN −1 )−1 ≤ r. This in turn
allows us to apply the canonical path method without making the chain lazy, since the upper
bound we obtain on (1 − λ1 )−1 will be (much) bigger than r.
7
From G ∈ ΩF do
choose a vertex a ∈ [n] u.a.r.,
choose a walk (a, b, c, d) of length 3 from a, u.a.r.,
if |{a, b, c, d}| = 4 and E ∩ {ac, bd} = ∅ then
delete the edges ab, cd and add the edges ac, bd,
otherwise
do nothing
end if;
end;
Figure 3: The flip Markov chain MF .
Now suppose that x, y are distinct elements of ΩF which differ by a flip, and that this flip
replaces edges ab, cd with ac, bd. Then If exactly one of the edges bc, ad is present in x, say
bc, then the transition from x to y takes place if and only if the chosen 3-walk is (a, b, c, d) or
(d, c, b, a). Hence in this case
PF (x, y) =
2
= PF (y, x).
(2r)3n
If both of the edges bc and ad are present then either of them could take the role of the “hub
edge” of the flip, and
4
= PF (y, x).
PF (x, y) =
(2r)3n
Hence PF is a symmetric matrix and the stationary distribution of the flip chain is uniform,
as claimed. (Mahlmann and Schindelhauer use a slightly different transition procedure, and
must argue about the number of triangles which contain a given edge in order to show that
their transitions are symmetric: see [16, Lemma 3].)
2
Two-stage direct canonical path construction
Suppose M is an ergodic Markov chain whose mixing time we would like to study, and M′ is
a chain for which a canonical path set Γ′ has been defined, with congestion bounded by ρ̄(Γ′ ).
Let P , π and Ω (respectively, their primed counterparts) be the transition matrix, stationary
distribution and state space of M (respectively, M′ ). Guruswami [10] proved that under mild
conditions on M′ , a bound on the mixing time of M′ implies the existence of an appropriate
set Γ′ with bounded congestion.
We wish to “simulate” transition in M′ using sequences of transitions in M, and showing
that not too many transitions in M can be involved in any one such simulation. Together
with ρ̄(Γ′ ) this results on a bound of the congestion of Γ, and hence on the mixing time of
M. Here we assume that Ω ⊆ Ω′ .
If Ω is a strict subset of Ω′ then canonical paths in Γ′ that begin and end in Ω may travel
through states in Ω′ \ Ω. (This will be the case in Section 4.) To deal with this, we must
8
define a surjection h : Ω′ −→ Ω such that h(u) = u for any u ∈ Ω, and such that the preimage
of any y ∈ Ω is not too large. This means that each state in Ω “stands in” for only a few
states in Ω′ . If Ω = Ω′ then h is the identity map, by definition.
′
Suppose that γwz
∈ Γ′ is a canonical path from w to z in G ′ . Write
′
γwz
= (Z0 , Z1 , . . . , Zj , Zj+1 , . . . , Zℓ ),
with Z0 = w, Zℓ = z and (Zj , Zj+1) ∈ E(M′ ) for all j ∈ [ℓ − 1]. We now define the two-stage
direct canonical path γw,z with respect to the fixed surjection h : Ω′ → Ω.
′
For each (Zj , Zj+1) ∈ γwz
, construct a path
σZj Zj+1 = (X0 , X1 , . . . , Xi , Xi+1 , . . . , Xκ ),
with X0 = h(Zj ) and Xκ = h(Zj+1), such that Xi ∈ Ω for i = 0, . . . , κ and (Xi , Xi+1 ) ∈ E(M)
for i = 0, . . . , κ − 1. We call σZj Zj+1 a (M, M′)-simulation path (or just a simulation path, if
no confusion can arise), since this path simulates a single transition in M′ using edges of M.
These simulation paths are concatenated together to form a canonical path γwz from w to
z in G, as follows:
γwz = (σZ0 Z1 , σZ1 Z2 , . . . , σZℓ−1 Zℓ ).
The following algorithmic interpretation of γwz may be useful as an illustration. Begin by
′
querying γwz
for the first transition (Z0 , Z1 ) to simulate. Beginning from h(Z0 ), perform
′
transitions in the corresponding simulation path until state h(Z1 ) is reached. Now query γwz
for the next transition (Z1 , Z2 ) and simulate that; and so on, until you have simulated all of
′
γwz
, completing the simulation path γwz .
Denote the set of all (M, M′)-simulation paths as
Σ = {σxy | (x, y) ∈ E(M′ )} .
For these simulation paths to give rise to canonical paths for M with low congestion, no
transition in M can feature in too many simulation paths. For each t ∈ E(M), let
Σ(t) = {σ ∈ Σ | t ∈ σ}
be the number of simulation paths containing the transition t. The measures of quality of Σ
are the maximum number of simulation paths using a given transition,
B(Σ) = max |Σ(t)|,
t∈E(M)
and the length of a maximal simulation path,
ℓ(Σ) = max |σ|.
σ∈Σ
Note that the simulation paths all depend on the fixed surjection h, and so the same is true
of the bounds B(Σ) and ℓ(Σ). In particular, the quality of the surjection h is measured by
maxy∈Ω |{z ∈ Ω′ | h(z) = y} and this quantity will be needed when calculating B(Σ) in
particular (as we will see in Section 4).
9
Some chains are ill-suited to simulation. For instance, a transition in M′ with large
capacity might necessarily be simulated in M by a simulation path that contained an edge
with small capacity. Hence we define two quantities that measure the gap between π and π ′ ,
and P and P ′. We define the simulation gap D(M, M′) between M and M′ by
D(M, M′) =
max
uv∈E(M)
zw∈E(M′ )
π ′ (z) P ′ (z, w)
.
π(u) P (u, v)
We combine these ingredients in the following lemma. The key point is that the congestion
across transitions in E(M′ ) (and hence across the simulation paths) is already bounded. If
the number of simulation paths making use of any given transition in E(M) is bounded, and
the graphs are not too incompatible then the congestion of each two-stage direct canonical
path in Γ will also be bounded.
Theorem 2.1. Let M and M′ be two ergodic Markov chains on Ω and Ω′ . Fix a surjection
h : Ω′ → Ω. Let Γ′ be a set of canonical paths in M′ and let Σ be a set of (M, M′)-simulation
paths defined with respect to h. Let D = D(M, M′) be the simulation gap defined above.
Then there exists a set Γ of canonical paths in M whose congestion satisfies
ρ̄(Γ) ≤ D ℓ(Σ) B(Σ) ρ̄(Γ′ ).
Proof. We begin with the definition of congestion of the set of paths Γ for M), and work
towards rewriting all quantities in terms of the set of canonical paths Γ′ for M′ , as follows:
ρ̄(Γ) = max
uv∈E(M)
X
1
π(x)π(y) |γxy |
π(u)P (u, v) x,y∈Ω
uv∈γxy
= max
uv∈E(M)
≤ max
uv∈E(M)
1
π(u)P (u, v)
1
π(u)P (u, v)
X
X
π(x)π(y) |γxy |
(2.1)
X
X
π(x)π(y) |γxy |.
(2.2)
zw∈E(M′ ) x,y∈Ω
σzw ∈Σ(uv) σzw ∈γxy
x,y∈Ω
′
σzw ∈Σ(uv) zw∈γxy
zw∈E(M′ )
Line (2.1) splits the two-stage canonical paths into simulation paths. By a slight abuse of
notation, we write σzw ∈ γxy to mean that σzw is one of the simulation paths which was
concatenated together to make γxy . Here we assume that each edge uv occurs at most once in
any given simulation path σ ∈ Σ. Otherwise, the second use of uv in any σ forms a cycle which
may be pruned: continue pruning until no edge appears more than once in any simulation
path. We also assume that no self-loops appear in any simulation path (these may also be
′
pruned.) Line (2.2) uses the one-to-one correspondence between γxy and γxy
for all x, y ∈ Ω.
′
Note that this correspondence is still one-to-one even when Ω ⊂ Ω , since the endpoints are
identical (as h acts as the identity map on Ω) and each canonical path is determined by its
endpoints.
10
Now we observe that for all uv ∈ E(M) and zw ∈ E(M′ ), by definition of the simulation
gap D,
1
1
≤D ′
.
π(u)P (u, v)
π (z)P ′ (z, w)
Similarly, for all x, y ∈ Ω′ ,
′
|γxy | ≤ ℓ(Σ) |γxy
|.
Therefore, as each summand of (2.2) is nonnegative and Ω ⊆ Ω′ , we obtain
X
X
1
′
π ′ (x)π ′ (y) |γxy
|
ρ̄(Γ) ≤ D ℓ(Σ) max
′ (z)P ′ (z, w)
uv∈E(M)
π
′
′
zw∈E(M )
σzw ∈Σ(uv)
≤ D ℓ(Σ)
max
uv∈E(M)
X
x,y∈Ω
′
zw∈γxy
ρ̄(Γ′ )
zw∈E(M)
σzw ∈Σ(uv)
≤ D ℓ(Σ) B(Σ) ρ̄(Γ′ ),
as required.
We remark that the surjection h does not play much role in the proof of Theorem 2.1,
since it is only needed to construct the simulation paths between two elements of Ω.
3
Analysis of the flip chain
Our aim is to define canonical paths in the flip chain that simulate canonical paths defined
in the switch chain. There are two main differences between the switch and flip chains which
cause difficulties. The state space of the flip chain is restricted to connected graphs, and no
flip can disconnect a connected graph. However, a switch can increase or decrease the number
of components in a graph by one. Hence, from a connected graph, the set of available flips
may be strictly smaller than the set of available switches. Following [9] we overcome these
difficulties in two steps, using an intermediate chain to bridge the gap.
Define the connected switch chain MSC to be the projection of the switch chain MS onto
state space ΩF . A transition of MSC is a connected switch, which is a switch on a connected
graph that preserves connectedness. To be precise, if PSC denotes the transition matrix of
MSC and PS denotes the transition matrix of MS , then for all x 6= y ∈ ΩS ,
(
PS (x, y) if x, y ∈ ΩF ,
(3.1)
PSC (x, y) =
0
otherwise.
Self-loops take up the slack, so for all x ∈ ΩF ,
PSC (x, x) = PS (x, x) +
X
PS (x, y).
y∈ΩS \ΩF
Transitions are symmetric, and so the stationary distribution πSC of MSC is the uniform
stationary distribution on ΩF (that is, πSC = πF ).
11
We proceed as follows. In Section 3.1 we show that a connected switch can be simulated
by sequences of flips, which we will call a long-flip. This gives a set of (MF , MSC )-simulation
paths. Here the surjection h is the identity map, since these chains share the same state
space. We apply Theorem 2.1 to this set of simulation paths.
Then in Section 4 we apply the theorem again to relate MSC and MS . We use a construction of Feder et al. [9] to define a surjection h : ΩS → ΩF . We then define (MSC , MS )simulation paths with respect to this surjection and apply Theorem 2.1. Theorem 1.1 is
proved by combining the results of these two sections.
The two-stage direct method has lower cost, in the sense of relaxation of the mixing time,
than the comparison method used in [9]. Both approaches have at their heart the construction
of paths in MF to simulate transitions in MS , but the two-stage canonical path analysis is
more direct and yields significantly lower mixing time bounds than those obtained in [9].
3.1
Simulating a connected switch: the long-flip
We must define a set of (MF , MSC )-simulation paths, as defined Section 2. The two Markov
chains have the same state space ΩF , and so we use the identity map as the surjection.
Let S = (Z, Z ′ ) be a connected switch which we wish to simulate using flips. Then Z and
Z ′ are connected graphs which differ by exactly four edges, two in Z \ Z ′ and two in Z ′ \ Z.
We say that the two edges ab, cd ∈ Z \ Z ′ have been switched out by S, and the two edges
ac, bd ∈ Z ′ \ Z have been switched in by S. Given an arbitrary labelling of one switch edge
as ab and the other as cd, these four permutations of the vertex labels
( ),
(a b)(c d),
(a c)(b d),
(a d)(b c)
(3.2)
preserve the edges switched in and out by (ab, cd ⇒ ac, bd).
To simulate the connected switch S = (Z, Z ′) we will define a simulation path called a
long-flip, which consists of a sequence of flips
σZZ ′ = (X0 , X1 , . . . , Xi−1 , Xi . . . , Xκ )
with X0 = Z and Xκ = Z ′ , such that (Xi−1 , Xi ) ∈ E(MF) for i = 1, . . . , κ. We define these
simulation paths in this section, and analyse their congestion in Section 3.2.
A long-flip uses a hub path p = (b, p1 , p2 , . . . , pν−1 , c) from b to c. In the simplest case
(in which p does not contain a or d and no internal vertex of p is adjacent to either of these
vertices) we will simply reverse p using flips, which has the effect of deleting the edges ab, cd
and replacing them with ac, bd, as desired. This process is described further in Section 3.1.1
below. The choice of hub path is important, and we now explain how the hub path will be
(deterministically) selected.
Suppose that (ab, cd ⇒ ac, bd) is the connected switch which takes Z to Z ′ , which we wish
to simulate using flips. We say that a path p is a valid hub path for the switch (ab, cd ⇒ ac, bd)
in Z if it satisfies the following (a, b, c, d) path conditions:
(i) p is a (b, c)-path in Z;
(ii) a and d do not lie on p; and
12
(iii) the vertices of p are only adjacent through path edges; that is, there are no edges in Z
of the form pi pj with |i − j| =
6 1 for i, j ∈ {0, . . . , |p|}.
We refer to condition (iii) as the no-shortcut property. Given a path p which satisfies conditions
(i) and (ii), it is easy to find a path with the no-shortcut property, by simply following the
shortcuts. A shortest (b, c)-path is a valid (a, b, c, d) path provided it avoids a and d.
Given Z and the switch edges ab, cd, we perform the following procedure:
• Let p be the lexicographically-least shortest (b, c)-path in Z which avoids a and d, if one
exists;
• If not, let p be the lexicographically-least shortest (a, d)-path in Z which avoids b and
c, if one exists, and apply the vertex relabelling (a b)(c d). (The result of this relabelling
is that now p is the lexicographically-least shortest (b, c)-path in Z which avoids a and
d.)
• If a path p was found, return the path p. Otherwise, return FAIL.
If a path p was found by the above procedure then we will say that we are in Case 1.
Otherwise we are in Case 2.
Suppose that if p is a lexicographically-least shortest path between vertices α and β in a
graph Z. We treat p as a directed path from α to β, with α on the “left” and β on the “right”.
Suppose that vertices γ, δ lie on p, with γ to the “left” of δ. Then the lexicographically-least
shortest (γ, δ)-path in Z is a subpath of p (or else we contradict the choice of p). This
important property, satisfied by hub paths, will be called the recognisability property. It will
enable us to reconstruct the hub path cheaply, as we will see in Section 3.2.
We now define the simulation path for Case 1 and Case 2 separately. Recall that since
flips cannot disconnect a connected graph, the simulation paths always remain within ΩF .
3.1.1
Case 1: the long-flip, and triangle-breaking
We must define a simulation path σZZ ′ for the connected switch from Z to Z ′ given by
(ab, cd ⇒ ac, bd). The path p selected above (the lexicographically-least shortest (b, c)-path
which avoids a and d) is a valid hub path for this switch, and we refer to it as the hub path.
If p has length 1 then this switch is a flip, and so we assume that p has length at least two.
We simulate the connected switch (ab, cd ⇒ ac, bd) by reversing the path p, as shown in
Figure 4. In (A), the switch (ab, cd ⇒ ac, bd) is shown with alternating solid and dashed
edges, with solid edges belong to Z and dashed edges belonging to Z ′ . Write the hub path as
p = (b, p1 , . . . , pν−1 , c), where b = p0 and c = pν .
The reversal of p proceeds in runs, in which successive vertices are moved along the path
one at a time into their final position. The run that moves vertex pi into place is called
the pi -run, during which pi is referred to as the bubble due to the way it percolates along
path p. The first stage of the path reversal is the b-run, which moves b adjacent to d. Next is
the p1 -run, and then the p2 -run, until the pν−2 -run, which puts c adjacent to a and completes
the reversal of p. Specifically, the pi -run consists of the sequence of flips
(a, pi , pi+1 , pi+2 ), (pi+1 , pi , pi+2 , pi+3 ), (pi+2 , pi , pi+3 , pi+4 ), . . . , (pν−1 , pi , c, pi−1 ),
13
(A)
a
b
⇒∗F
p1
p2
d
pν−1
c
p1
p2
p3
⇒F
p1
d
c
⇒∗F
b
d
p2
b
(C)
a
(B)
a
c
pν−1
(D)
a
c
d
pν−1 pν−2
p1
b
Figure 4: A hub path reversed by flips.
where p0 = b, pν = c (and in the case of the b-run, p−1 is understood to denote d). After the pi
run, c has moved one step closer to a on path p. See Figure 4, where the current bubble is
shown as a diamond. The result of the first flip of the b-run is shown in (B), and the situation
at the end of the b-run is shown in (C), At this stage, b is in its final position, adjacent to
d. Finally in (D) we see the graph after all runs are complete, with the path reversed. (Note
that ⇒F denotes a single flip while ⇒∗F denotes a sequence of zero or more flips.)
Since the distance from c to a decreases by one after each run, the total number of flips
used to reverse the path p is 12 ν(ν + 1). The no-shortcut property of hub paths is required so
that no flip in any run is blocked by parallel edges within p.
However, the first flip of the b-run is blocked if p1 is adjacent to a. Before we can reverse
the path p, we must break this triangle [a, b, p1 ] using the following procedure. Choose the
least-labelled vertex x that is adjacent to b but not p1 . The (a, b, c, d) path conditions imply
that p2 is adjacent to p1 , but not to b, so the existence of x follows from 2r-regularity. Then
the triangle [a, b, p1 ] can be removed by performing a flip on the path (x, b, p1 , p2 ), as shown
below.
a
a
x
b
p1
p2
⇒
x
p1
b
p2
Now p1 is still adjacent to a, but it is no longer on the (b, c)-path, and the length of the
(b, c)-path has been reduced by 1. However, it is possible that p2 is also adjacent to a, so we
may still have a triangle. If so, we must repeat the process, as indicated below, successively
shifting b to the right on the (b, c)-path, and reducing its length. To break the triangle [a, b, p2 ]
we flip on (p1 , b, p2 , p3 ).
14
a
a
x
p1
b
p2
p3
⇒
x
p1
p2
b
p3
More generally, if the edges ap1 , . . . , apk are all present but the edge apk+1 is not then the
k’th triangle [a, b, pk ] is broken with the flip (pk−1 , b, pk , pk+1 ). After this flip, b is adjacent to
pk and pk+1 . We call these triangle-breaking flips.
The no-shortcut property of the (b, c)-path guarantees that all the triangle-breaking flips
after the first one are valid. The triangle-breaking process must end when bc becomes an
edge, at the very latest. This implies that k ≤ ν − 1. Furthermore, since a has degree 2r we
also have k ≤ 2r − 1. When the process terminates, the edge apk+1 is absent and bpk+1 is the
first edge on the (b, c)-path. Now the long-flip can be performed on this new (b, c) hub path.
If k = ν − 1 then pk+1 = c and this long-flip is a single flip.
After the path is fully reversed we must undo the triangle removals, by reversing each
triangle-breaking flip, in reverse order. That is, if k triangle-breaking flips were performed,
then performing these flips in order will reinstate these triangles:
(pk−1, pk , b, pk+1), (pk−2 , pk−1 , b, pk ), . . . (p1 , p2 , b, p3 ), (x, p1 , b, p2 ).
(When k = 1, only the last flip is performed.) We call these triangle-restoring flips. These
flips are not blocked because the only edges whose presence could block them are guaranteed
not to exist by the no-shortcut property of p, and by the fact that these edges were removed
by the triangle-breaking flips.
3.1.2
Case 2: The detour flip
In Case 2, we know that there are no paths from b to c in Z which avoid both a and d, and
that there are no paths from a to d in Z which avoid b and c. It follows that Z must have
the structure illustrated in Figure 5. In this figure, each Zij (shown as a grey ellipse) is an
a
Zab
Zac
c
b
Zbd
Zcd
d
Figure 5: The structure of a graph Z in Case 2.
15
induced subgraph of Z, with i, j ∈ Zij . There are no edges from Zij \ {i, j} to Zkℓ \ {k, ℓ}
whenever {i, j} 6= {k, ℓ}, by the assumption that Z is in Case 2. Furthermore, Zab and Zcd
are both connected, since Z is connected and ab, cd ∈ E(Z).
At most one of Zac or Zbd may be empty or disconnected, since Z is connected. A
deterministic procedure will now be described for constructing a 3-path (u, d, w, z) in Z which
will be used for the first flip in the simulation path.
There are two subcases to consider for the choice of w:
• First suppose that, after applying the relabelling (a b)(c d) if necessary, Zac is empty or
disconnected. Since cd is not a bridge in Z, by Lemma 1.2(ii), there exists a cycle in
Zcd which passes through the edge cd. It follows that d has at least one neighbour in
Zcd \ c which is connected to c by a path in Zcd \ cd. Let w be the least-labelled such
neighbour of d in Zcd \ cd.
• Secondly, suppose that both Zac and Zbd are nonempty and connected. (So far, we
have not performed any relabelling.) Then at most one of Zab \ ab, Zcd \ cd may be
disconnected, or else the graph Z ′ will be disconnected, a contradiction. Applying the
relabelling (a c)(b d) if necessary, we can assume that Zcd \ cd is connected. Hence there
is a path from c to d in Zcd \ cd, so d has a neighbour in Zcd \ c which is joined to c by
a path in Zcd \ cd. Let w be the least-labelled such neighbour of d in Zcd \ c.
In both subcases, we have established the existence of a suitable vertex w. Next, observe
that since (in both the above subcases) Zbd is nonempty and connected, there is a path from
b to d in Zbd of length at least two (as bd 6∈ E(Z)). This implies that d has a neighbour
in Zbd which is joined to b by a path in Zbd \ d. Let u be the least-labelled such neighbour
of d in Zbd \ b. Finally, since all neighbours of w belong to Zcd and u 6∈ Zcd , it follows that
{u, w} ⊆ NZ (d) \ NZ (w). Since Z is regular, this implies that
|NZ (w) \ NZ (d)| = |NZ (d) \ NZ (w)| ≥ 2.
Hence there is some vertex other than d in NZ (w) \ NZ (d). Let z be the least-labelled such
vertex. Then z ∈ Zcd \ {c, d}, since c ∈ NZ (d) and every neighbour of w belongs to Zcd .
Now we can define the simulation path σZZ ′ from X0 = Z to Xκ = Z ′ . The first step in
the simulation path is the flip on the path (u, d, w, z), giving the graph X1 . We call this the
detour flip, as it will give us a way to avoid d. See the first line of Figure 6.
After the detour flip, there is a path from b to u in X1 \ {a, d}, by choice of u, and there is
a path from c to w in X1 \ {a, d}, by choice of w. It follows that there exists a (b, c)-path in
X1 which avoids {a, d}: let p be the lexicographically-least shortest such path. Note that p
necessarily contains the edge uw. (See X1 in Figure 6: the path p is indicated by the dotted
lines together with the edge from u to w.) We proceed as in Case 1, simulating the switch
(ab, cd ⇒ ac, bd) from X1 using a long-flip with hub path p. At the end of the long-flip the
path p is reversed and we reach the graph Xκ−1 shown at the bottom left of Figure 6.
For the final step of the simulation path, we reverse the detour flip by performing a flip
on the path (u, w, d, z). This produces the final graph Xκ = Z ′ shown at the bottom right of
Figure 6, completing the description of the simulation path in Case 2.
16
a
a
d
b
u
⇒F
c
w
a
c
b
u
z
X0
⇒∗F
d
u
z
a
⇒F
b
Xκ−1
c
X1
d
w
w
c
z
d
w
u
Xκ
b
z
Figure 6: Case 2: the simulation path
3.2
Bounding the congestion
The previous section defines a set ΣF of (MF, MSC )-simulation paths. First, we prove an
upper bound on the maximum length ℓ(ΣF ) of a simulation path in ΣF .
Lemma 3.1. If r ≥ 2 then ℓ(ΣF ) ≤ 21 n2 .
Proof. Since hub paths are simple and do not include a or d, every hub path contains at
most n − 2 vertices. Therefore, at most 21 (n − 2)(n − 3) path-reversing flips will be required.
Furthermore, we may have up to 2r − 1 triangle-breaking flips, and up to 2r − 1 trianglerestoring flips. Additionally, if we are in Case 2 then we also have a detour flip and its inverse.
Therefore
(n − 2)(n − 3)
ℓ(ΣF ) ≤
+ 4r,
2
and this is bounded above by 21 n2 whenever r ≥ 2, since n ≥ 2r + 1.
For the remainder of this subsection we work towards an upper bound on B(ΣF ), the
maximum number of simulation paths containing a given flip. A given flip t = (X, X ′ ) could
perform one of a number of roles in a simulation path for a connected switch S = (Z, Z ′):
• a triangle-breaking flip;
• a triangle-restoring flip;
• a flip on the path reversal (that is, part of the long-flip);
• in Case 2, the detour flip (u, w, d, z);
• in Case 2, the inverse detour flip (u, d, w, z).
We bound the number of simulation paths that could use t in each of these roles separately,
in Lemmas 3.2–3.4 below. Summing over these bounds over all roles yields an upper bound
17
for B(ΣF ). In our exposition, we attempt to uniquely determine a given simulation path σZZ ′
which contains t, by “guessing” from a number of alternatives. Since the number of possible
alternatives is precisely the number of simulation paths σZZ ′ that use t, this approach gives
the desired bound.
The simulation path is determined once we have identified the switch (ab, cd ⇒ ac, bd)
and either Z or Z ′ (since each of these can be obtained from the other using the switch or
its inverse). Observe that in order to uniquely identify the switch it is not enough to just
identify the two edges e1 , e2 which are switched out (or to just identify the two edges that
are switched in). This follows since there are up to two possible switches on these edges: if
e1 = ab and e2 = cd then we could switch in either the edges ac, bd or the edges ad, bc.
We pay a factor of 2 to distinguish between these two possibilities, and will describe this as
orienting one of the switch edges.
Finally, suppose that the current flip t = (X, X ′ ) switches out the edges αβ, γδ and
replaces them by αγ, βδ. If both αδ and βγ are present in X (and hence, both present in X ′ )
then either one of these could have been the middle edge of the 3-path chosen for the flip. In
this case, if we wish to identify β, say, then there are 4 possibilities. However, if we know a
priori that the edge αδ is not present in X, say, then we can guess β from two possibilities.
Lemma 3.2. If t = (X, X ′ ) is a detour flip then we can guess the connected switch S = (Z, Z ′)
from at most 8r 2n possibilities. The same bound holds if t is an inverse detour flip.
Proof. First suppose that t is a detour flip. This is the first flip in the simulation path, so
Z = X.
• The flip is on the 3-path (u, w, d, z), and uz 6∈ E(X) by construction, so we can guess
d from 2 possibilities.
• Next, we can guess c ∈ NX (d) from at most 2r possibilities.
• Finally, guess the edge ab ∈ E(X) and orient it, from 2rn possibilities.
From the switch edges and Z we can obtain Z ′ . Hence we can guess the switch S from at
most 8r 2n possibilities overall when t is a detour flip. The same argument holds when t is an
inverse detour flip, replacing X and Z by X ′ and Z ′ and exchanging the roles of b and c.
Lemma 3.3. Suppose that t = (X, X ′ ) is a triangle-breaking flip. Then we can guess the
connected switch S = (Z, Z ′) from at most 4(2r)8n possibilities. The same bound holds if t is
a triangle-restoring flip.
Proof. Suppose that t is a triangle-breaking flip.
• One of the vertices involved in the flip t is b, so we can guess it from 4 alternatives. (If
t is not the first triangle-breaking flip then by the no-shortcuts property, we can guess
b from 2 alternatives. But in the first triangle-breaking flip xp2 may be an edge of Z,
in which case there are 4 possibilities for b.)
• We can guess a by adjacency to b, from 2r possibilities.
• We can guess the edge cd ∈ E(X) and orient it, from 2rn possibilities.
18
We now know the switch vertices, so recovering Z will give us the rest of the path. There
may have been some other triangle-breaking flips which took place before t, and if we are in
Case 2 then there was also a detour flip at the start of the simulation path. We continue our
accounting below.
• First suppose that we are in Case 1, and that t is the kth triangle-breaking flip which
has been performed (so far). Then k ≥ 1. Let p′ = (x, p1 , . . . , pk , b) denote the path containing vertices which were removed from the hub path by the first k triangle-breaking
flips. Now (a, p1 , x) is a 2-path in X, so we can guess x and p1 from 2r(2r − 1) ≤ (2r)2
alternatives.
• Next, pk belongs to NX (b) \ a, so we can guess pk from at most 2r alternatives.
(Here we allow the possibility that p1 = pk , which covers the case that t is the first
triangle-breaking flip.) Then we can reconstruct p′ using the fact that (p1 , . . . , pk ) is
the lexicographically-least shortest (p1 , pk )-path in X \ {a, d}, by the recognisability
property. Using p′ , we can reverse these k triangle-breaking flips to obtain Z from X.
• Finally, we must guess whether we are in Case 1 (1 possibility) or Case 2. If we are
in Case 2 then we must also reverse the detour flip before we obtain Z. Since u and
w are neighbours of d and z is a neighbour of w, in this step we guess from at most
1 + (2r)2 (2r − 1) < (2r)3 possibilities.
Overall, we can guess the switch S from a total of at most 4(2r)8 n possibilities, when t is a
triangle-breaking flip.
The argument is similar when t is a triangle-restoring flip:
• First, suppose that we are in Case 1. Guess b, then guess d from the neighbourhood of
b, then guess and orient the edge ac ∈ E(X). This gives the switch edges, guessed from
at most 4 (2r)3 alternatives.
• Recover p′ = (x, p1 , . . . , pk , b) as in the triangle-breaking case, from at most (2r)3 alternatives. Now we can Z ′ uniquely by performing the remaining triangle-restoring flips
to move b past p1 .
• Again, there at most (2r)3 possibilities for deciding whether or not we are in Case 2 and
if so, guessing the inverse detour flip and applying it to produce Z ′ .
Hence there are at most 4(2r)8 n possibilities for the connected switch S when t is a trianglerestoring flip.
Lemma 3.4. Suppose that t = (X, X ′ ) is a flip on the path-reversal. We can guess the
connected switch S = (Z, Z ′) from at most 4(2r)8 ((2r)2 − 1) n3 possibilities.
Proof. Suppose that t is a flip on the path-reversal with respect to the hub path p =
(b, p1 , . . . , pν−1 , c), where b = p0 and c = pν . Then t is part of a pi run, for some i. The
path-reversal may have been preceded by some triangle-breaking flips, and (if we are in Case
2) a detour flip. We need to guess the switch edges and reconstruct Z.
First suppose that we are in Case 1. The typical situation is shown in Figure 7, which
shows that graph X ′ produced by the flip t. The current bubble pi is shown as a diamond,
and the flip t has just been performed on the 3-path (pj , pi , pj+1, pj+2 ), moving pi past pj+1 .
19
a
x
p1
p2
pk
′
d
pj
pj+1
pi
pj+2
a
p[a′ : pj+1 ]
p[pj+2 : c]
c
c′
pk+1
b
rev p[pk+1 : c′ ]
Figure 7: Reconstruction of Z from X ′ , the flip edges and some guessed information.
• The current bubble pi is one of the four vertices involved in the flip, so it can be guessed
from four possibilities.
• Next we guess a, c and d from n3 alternatives.
• We recover b as follows: if the current run is the b-run then b is the bubble; otherwise,
b is already adjacent to d, and we can guess it from the 2r neighbours of d. This gives
2r + 1 possibilities for b.
• Vertices a′ and c′ can be guessed from at most 2r alternatives each, by their adjacency
to a and c.
Now we know the switch (ab, cd ⇒ ac, bd), and it remains to recover Z. In order to do that, we
must recover the path p, reverse the path-reversal flips and then reverse the triangle-breaking
flips, if any.
Observe that having identified pi , the identities of pj+1 and pj+2 can be inferred. To see
this, note that pi pj+2 is one of the two edges switched in by the flip t, which identifies pj+2 ,
while pj+1 pj+2 is one of the two edges switched out by the flip t, which identifies pj+1 .
Next, observe that a′ = pi+1 , and so the subpath pi+1 · · · pj+1 is the lexicographically-least
shortest (pi+1 , pj+1 )-path in X ′ \ {a, d}, by the recognisability property. This holds since the
flips made on the path-reversal so far, between Z and X ′ , have not altered this section of the
original path. Similarly, the subpath pj+2 · · · c can be recovered as the lexicographically-least
shortest (pj+2 , c)-path in X ′ \ {a, d}.
If the current bubble is b then c′ = c. Otherwise, observe that c′ = pi−1 , since c′ is the
most recent bubble that successfully completed its run. We continue our accounting below.
• Guess pk+1 from NX ′ (b) \ d, from at most 2r − 1 possibilities. Reconstruct the subpath
pk+1 · · · pi−1 as the lexicographically-least shortest (pk+1 , pi−1 )-path in X ′ \ {a, d}.
• By now we have recovered all sections of the path between a′ and b as shown in Figure 7,
and we can reverse all path-reversing flips. Now there may have been no trianglerestoring flips, in which case we have constructed Z, or else we may need to reverse the
triangle-breaking flips as explained earlier, after guessing the identities of x, p1 and pk−1
from at most (2r)2(2r −1) possibilities. This gives a factor of at most 1+(2r)2(2r −1) ≤
(2r)3 .
• Finally, multiply by 1 + (2r)2 (2r − 1) ≤ (2r)3 to account for the fact that we may be in
Case 1 or Case 2, as explained in Lemma 3.3.
20
Overall, we have guessed the switch S from at most
4(2r + 1)(2r − 1)(2r)8 n3 = 4(2r)8 (2r)2 − 1 n3
possibilities, when t is a flip on a path-reversal.
Now we apply the two-stage direct canonical path theorem to our (MF , MSC )-simulation
paths. Let ρ̄(ΓSC ) denote the congestion of some set of canonical paths for the connected
switch MSC . We will define such a set of canonical paths in Section 4 below.
Lemma 3.5. The (MF , MSC )-simulation paths defined in Section 3.1 define a set of canonical
paths for MF with congestion ρ̄(ΓF ) which satisfies
ρ̄(ΓF ) ≤ 8 (2r)11n4 ρ̄(ΓSC ).
Proof. Firstly, we bound the maximum number B(ΣF ) of simulation paths which contain a
given flip t, by adding together the number of possibilities from each of the five roles that t
may play. Using Lemmas 3.2– 3.4, we obtain
B(ΣF ) ≤ 4(2r)8 (2r)2 − 1 n3 + 8(2r)8n + 16r 2n
≤ 4 (2r)10 − (2r)8 + 2(2r)6 + 1 n3
≤ 4(2r)10 n3 ,
since n ≥ 2r ≥ 4. Next, we calculate the simulation gap DF = D(MF , MSC ). Since MF and
MSC share the same state space ΩF , and both have uniform stationary distribution,
DF =
max
uv∈E(MF )
zw∈E(MSC )
PSC (z, w)
|ΩF | PSC (z, w)
=
.
|ΩF | PF(u, v)
PF (u, v)
Let uv ∈ E(MF) and zw ∈ E(MSC ). Then u 6= v, as self-loop transitions are not included
in E(MF ) or E(MSC ). Hence
1
(2r)3 n
≤
PF (u, v)
2
and
PSC (z, w) =
1
2
≤ 2 2
3an,2r
r n
using (1.3), since n ≥ 8. Therefore DF ≤ 8r/n.
Substituting these quantities into Theorem 2.1, using Lemma 3.1, we obtain
ρ̄(ΓF ) ≤ DF ℓ(ΣF ) B(ΣF ) ρ̄(ΓSC )
≤ 8 (2r)11 n4 ρ̄(ΓSC ),
as claimed.
21
4
Disconnected graphs
Next, we must define a set of (MSC , MS )-simulation paths and apply the two-stage direct
canonical path method. The process is similar to the approach introduced by Feder et al. [9]:
the chief difference is in the analysis.
We begin by defining a surjection h : ΩS → ΩF such that maxG∈ΩF |h−1 (G)| is polynomially
bounded. Using this we construct short paths in MSC to correspond to single transitions in
MS . In [9] this step is responsible for a factor of O(r 34 n36 ) in the overall mixing time bound. In
contrast, by using the two-stage direct method we construct canonical paths with a relaxation
of only a factor of 120rn3 .
First we must define the surjection h. Let GS ∈ ΩS and let {H1 , . . . , Hk } be the set of
components in GS . Let vi the vertex in Hi of highest label, for i = 1, . . . , k. By relabelling
the components if necessary, we assume that
v1 < v2 < · · · < vk .
We call each vi an entry vertex. By construction, vk is the highest-labelled vertex in GS .
In each component we will identify a bridge edge ei = vi vi′ , where vi′ is the neighbour of vi
with the highest label. We call edge vi′ an exit vertex. The graph G = h(GS ) is formed from
GS by removing the bridge edges e1 , . . . , ek and replacing them with chain edges e′1 , . . . , e′k ,
where ei = vi′ vi+1 for 1 ≤ i ≤ k − 1, and e′k = vk′ v1 . We call e′k the loopback edge and we
call this procedure chaining. (When discussing the chain, arithmetic on indices is performed
′
with v1′ .)
cyclically; that is, we identify vk+1 with v1 and vk+1
By construction we have vi′ < vi < vi+1 for i = 1, . . . , k − 1. Hence the “head” vi+1 of the
chain edge e′i is greater than the “tail” vi′ , except for possibly the loopback edge: v1 may or
may not be greater than vk′ . We prove that G = H(GS ) is connected in Lemma 4.1 below.
An example of a chained graph G is shown in Figure 8 in the case k = 3. Here G = H(GS )
for a graph GS with three components H1 , H2 , H3 . The shaded black vertices are the entry
vertices v1 , v2 and v3 , and the white vertices are the exit vertices v1′ , v2′ and v3′ . After chaining,
H1 contains two cut edges (shown as dashed edges), because e1 = v1 v1′ was part of a cut cycle
in H1 . Similarly, H3 contains one cut edge. Note that these dashed edges are not chain edges.
The loopback edge is e′3 and v1 < v2 < v3 .
Lemma 4.1. Let h be the map defined above. Then h(GS ) is connected for all GS ∈ ΩS , and
h : ΩS → ΩF is a surjection.
Proof. Let GS ∈ ΩS and suppose G = h(GS ). Lemma 1.2(ii) implies that each connected
component Hi remains at least 1-connected after removal of bridge edges. The chain edges
then connect these components in a ring, forming a connected graph. Additionally, h preserves
degrees since each vertex vi , vi′ is adjacent to exactly one bridge edge that is removed and
exactly one chain edge that is added. Hence, G ∈ ΩF . Finally, observe that h(G) = G
whenever G is connected, which proves that h is a surjection since ΩF ⊂ ΩS .
Lemma 4.2. Let h : ΩS → ΩF be defined as above. Then
max |h−1 (G)| ≤ 2rn.
G∈ΩF
22
e′3
v1′
v1
e′1
v2 v2′
H2
H1
e′2
v3′
v3
H3
Figure 8: A connected graph h(GS ) ∈ ΩF obtained by chaining a disconnected graph GS ∈ ΩS
with three components H1 , H2 , H3 .
Proof. Let GF ∈ ΩF be given. First we prove that given an edge e = {v, w} of GF and an
orientation (v, w) of e, there is a unique disconnected graph GS ∈ h−1 (GF ) such that when
the chaining procedure is performed with input GS (producing GF ), the loopback edge e′k
equals e and v1 = v. We recover GS from GF and the oriented edge e using the unchaining
algorithm which we now describe.
To identify e′1 given e′k and v1 , begin by searching the graph GF \ e′k , starting from v1 ,
until a vertex is reached with a label greater than v1 . By construction, this vertex must be
v2 , as v1 has the greatest label of any vertex in H1 . The edge traversed to reach v2 must be
e′1 . Repeat this process until the chain edges e′1 , . . . , e′k−1 have been discovered. Then GS is
recovered by deleting all chain edges (including the loopback edge) from GF and replacing
them with the bridge edges e1 , . . . , ek .
For future reference, note that this unchaining algorithm will still correctly recover GS so
long as vi′ is any neighbour of vi in GS (that is, vi′ does not need to be the neighbour with
largest label).
Hence the number of choices for the oriented loopback edge gives an upper bound on
|h−1 (GS )| − 1. We use the naive estimate of 2r(n − 1) as an upper bound on this quantity,
since the loopback edge could be any oriented edge which does not originate in vk (the
highest-labelled vertex in GF ). Since GF is the only connected graph in h−1 (GF ), we obtain
|h−1 (GS )| ≤ 2r(n − 1) + 1 ≤ 2rn. (It should be possible to obtain a tighter bound with a
little more work, though we do not attempt that here.)
4.1
Simulating general switches
Now we must construct a set of (MSC , MS )-simulation paths ΣSC with respect to the surjection h defined above. Given a switch S = (W, W ′ ) ∈ E(MS ), with W, W ′ ∈ ΩS , we must
simulate it by a path σW W ′ in ΓSC , consisting only of connected switches. We then give an
upper bound on the congestion parameters for the set ΣSC of simulation paths, and apply
Theorem 2.1. The main difficulty is that switches in MS may create or merge components,
whereas connected switches in MSC cannot.
In this section we refer to transitions in E(MS ) as general switches, to distinguish them
from connected switches. We say that a general switch S = (W, W ′ ) is
23
• neutral if W and W ′ have the same number of components;
• disconnecting if W ′ has one more component than W ;
• reconnecting if W ′ has one fewer component than W .
The general switch S = (ab, cd ⇒ ac, bd) is disconnecting if ab, cd form a 2-edge-cut in the
same component of W and ac, bd do not. If S is a reconnecting switch then ab and cd below
to distinct components of W . Finally, S is a neutral switch if ab and cd both belong to the
same component of W , but do not form a 2-edge-cut, or if ab and cd form a 2-edge-cut in a
component of W , and ac and bd form another 2-edge-cut in the same component.
To simulate the general switch S = (W, W ′ ), we must define a path
σW,W ′ = (Z0 , Z1, . . . , Zq ),
in E(MSC ), where Z0 = h(W ) and Zq = h(W ′ ), such that Zi ∈ ΩF for i = 0, . . . , q and
(Zi , Zi+1 ) ∈ E(MSC ) for i = 0, . . . , q − 1. Here are the kinds of connected switch that we will
need in our simulation paths.
• Suppose that a bridge edge vj vj′ of W is one of the edges to be removed by the switch S.
This is a problem, since all bridge edges have been deleted and replaced by chain edges
during the chaining process used to construct Z0 = h(W ). To deal with this, we first
perform a connected switch called a bridge change switch. The bridge change switch
reinstates the edge ej e′j (so that the desired switch S can be performed) and changes the
choice of exit vertex (previously e′j ) in Hj . Importantly, the resulting chain structure
can still be unwound using the algorithm in Lemma 4.2. Specifically, to reinstate the
′
bridge edge ej = vj vj′ in component Hj , let vj∗ ∈ NW (vj ) \ vj−1
be the highest-labelled
∗
neighbour of vj in W [Hj ] and perform the connected switch (vj vj , vj′ vj+1 ⇒ vj vj′ , vj∗ vj+1 ).
(Clearly vj∗ 6= vj′ since the edge vj vj′ is absent in Z0 , and vj+1 vj∗ is not an edge of Z0 due
to the component structure of W , so this connected switch is valid.) The new chain
edge is vj∗ vj+1 and the new bridge edge is vj vj∗ . We denote the bridge change switch by
B∆. See Figure 9.
vj′
vj
′
vj+1
vj+1
vj∗
Hj
Hj+1
Figure 9: A bridge change in component Hj .
• We will need a connected switch to simulate the general switch S. We denote this switch
as NS, DS or RS if S is neutral, disconnecting or reconnecting, respectively, and call it
the neutral (respectively, disconnecting or reconnecting) simulation flip. It removes the
same two edges that S removes, and inserts the same two edges that S inserts.
24
• Let Z be the current graph after the DS, RS or NS has been performed. We say that
component Hj is healthy in Z if there is no neighbour of the entry vertex vj in Z with
an label higher than the exit vertex vj′ . We must ensure that all components of Z are
healthy before performing the unchaining algorithm, or else we lose the guarantee that
Z = h(W ′ ).
A component Hj can only become unhealthy if some switch along the simulation path
involves the entry vertex vj and introduces a new neighbour of vj , say vj∗ , which has a
higher label than the (current) exit vertex vj′ . In this case, vj∗ also has the highest label
among all current neighbours of vj . (In fact, the switch that makes Hj unhealthy will
be the switch which simulates S, namely, the DS, RS or NS.) Note, if there has been a
bridge change in component Hj at the start of the simulation path, then vj′ might not
be the exit vertex originally produced in Hj by the chaining procedure. But this will
not cause any problems, as we will see.
We can make Hj healthy by performing the bridge change switch (vj vj∗ , vj′ vj+1 ⇒
vj vj′ , vj∗ vj+1), precisely as described above. (Now the main purpose of this switch is
not to reinstate an edge, but to ensure the correct vertex of Hj becomes the exit vertex
before unchaining.) When this bridge change is performed at the end of the simulation
path we will refer to it as a bridge rectification switch, denoted BR.
• If S is a disconnecting switch then performing this switch will produce a disconnected
graph (and destroy the chain structure) by splitting Hj into two components, one of
which is not chained. In order to avoid this, we perform a disconnected housekeeping
switch, denoted DHK, before the disconnecting simulation flip (DS). The DHK is defined
in detail in Section 4.1.2 below.
• Similarly, if S is a reconnecting switch then performing this switch destroys the chain
structure, by merging Hi and Hj together, forming one component which is incident with
four chain edges instead of two. To fix the chain structure we perform a reconnecting
housekeeping switch, denoted RHK, after the reconnecting simulation flip (RS). More
detail is given in Section 4.1.3 below.
We now define simulation paths for the three types of general switch S = (W, W ′ ). The
paths are summarised in Table 1, with optional steps denoted by square brackets. Here
“optional” means that these steps may or may not appear in the simulation path for a general
switch of that type. However, given a particular switch S, the definition of the simulation
path complete determines whether or not these “optional” switches are required.
Neutral switch
Disconnecting switch
Reconnecting switch
−→
−→
−→
[B∆] NS [BR]
[B∆] DHK DS [BR] [BR]
[B∆] [B∆] RS RHK [BR]
Table 1: Simulation path for a general switch S ∈ ΩS using connected switches.
25
4.1.1
The simulation path for a neutral switch
Suppose that the general switch S is a neutral switch. The simulation path for S is formed
using the following procedure.
• Let Hi be the component of W that contains all vertices involved in S.
• If the bridge edge vi vi′ is switched away by S then perform a bridge change.
• After the bridge change, if any, perform a connected switch NS which switches in and
out the the edges specified by S.
• If a bridge edge is unhealthy after this switch then perform a bridge rectification.
4.1.2
The simulation path for a disconnecting switch
Suppose that the general switch S is a disconnecting switch. The simulation path for S is
formed using the following procedure.
• Let Hj contain the 2-edge-cut ab, cd and let B1 , B2 be the subgraphs of Hj separated
by this cut. Without loss of generality, suppose that B1 contains vj .
• If a bridge change is required in Hj then perform the bridge change. In that case, vj vj′
is one of the edges forming the 2-edge-cut, so all other neighbours of vj must lie in B1 .
Therefore the new exit vertex vj∗ also belongs to B1 .
• After the bridge change, if any, notice that switching in and out the edges specified
by S would split Hj into two components B1 and B2 . Of these, B1 is already in the
correct place in the chain, but B2 is not connected to the chain at all. This is not
allowed, since we are restricted to connected switches. Therefore we must first perform
the disconnected housekeeping switch (DHK) to place B2 into the correct position in the
chain. This switch is described in more detail below.
• After the DHK has been performed, we perform the disconnecting switch specified by
S. (This is now safe, as the DHK ensures that the resulting graph is not disconnected.)
• Up to two bridge edges may now be unhealthy. If so, then perform up to two bridge
rectification switches. For definiteness, if two BR switches are required, perform the
one with the lower component index first.
It remains to specify the disconnected housekeeping switch DHK. See Figure 10. Let v+ be
′
′
the vertex in B2 with the highest label and let v+
be its greatest neighbour. Then e+ = v+ v+
must become the bridge edge of B2 . Find i such that vi < v+ < vi+1 , so B2 belongs in the
chain between Hi and Hi+1 in the chain. (Note that vk is the highest-labelled vertex in the
graph, so v+ ≤ vk , but it is possible that v+ < v1 , in which case B2 must become the first
component in the chain and vk′ v+ must become the new loopback edge.) Then DHK is the
′
′
switch (v+ v+
, vi′ vi+1 ⇒ v+ vi′ , v+
vi+1 ), which inserts B2 into the correct position in the chain
as required.
26
vi
vi′
Hi
′
vi+1
vi+1
vj
a
vj′
c
b
d
Hi+1
Hj
v+
′
v+
Figure 10: The situation just before a disconnecting housekeeping switch (DHK).
4.1.3
The simulation path for a reconnecting switch
Suppose that the general switch S is a reconnecting switch. The simulation path for S is
defined using the following procedure.
• Suppose that ab ∈ Hi and cd ∈ Hj where, without loss of generality vi < vj . Up to two
bridge changes may be necessary, as this switch involves two components. If both are
necessary, perform the bridge change in Hi first.
• Then perform the reconnecting switch RS that switches in and out the edges specified
by S. At this point, the component formed by linking Hi and Hj is incident with four
chain edges, so the chain structure is not correct. In effect, this new component is
occuping both position i and position j in the chain. Since vi < vj , vertex vj must be
the entry vertex of the new component, so the correct position for this component is
between Hj−1 and Hj+1.
′
′
• The reconnecting housekeeping switch (RHK) (vi vi−1
, vi′ vi+1 ⇒ vi vi′ , vi−1
vi+1 ) restores
the chain by placing the new component in position between Hj−1 and Hj+1 , linking
Hi−1 directly to Hi+1 (which is now the next component in the chain) and reinstating
′
the bridge edge vi vi′ . Here vi′ and vi−1
refer to endvertices of the current chain edges,
after the at most two bridge changes have been performed. (These may differ from the
endvertices of the original chain edges, before the bridge rectification steps.)
• Finally, at most one bridge rectification may be required to ensure that the single bridge
edge vj vj′ is healthy.
In the special case when the reconnecting switch S = (W, W ′ ) results in a connected graph
W ′ , we must have i = 1 and j = 2: here the reconnecting housekeeping switch reinstates both
bridge edges, and no bridge rectification is required.
27
4.2
Bounding the congestion
Let us calculate upper bounds for B(ΣSC ) and ℓ(ΣSC ). From Table 1, the longest possible
simulation path consists of five connected switches, so
ℓ(ΣSC ) = 5.
(4.1)
We spend the rest of this subsection bounding B(ΣSC ).
Each connected switch t = (Z, Z ′) can play a number of roles in a simulation path. We
consider each kind of general switch separately (neutral, disconnecting, reconnecting). As in
Section 3.2, the number of simulation paths (of a given kind) containing t is equal to the
amount of extra information required to identify a general switch S of this kind uniquely,
given t (and assuming that t belongs to the simulation path for S): we speak of “guessing”
this information. Again, it is enough to identify one of W or W ′ , together with the edges
switched in/out by S, and then the other graph (W ′ or W ) can be inferred.
4.2.1
Neutral switches
A connected switch t = (Z, Z ′) in MSC could take on three possible roles when simulating a
neutral switch S = (W, W ′ ). We consider each separately below.
(B∆) First, suppose that the connected switch t is a bridge change in a simulation path for
a neutral (general) switch S.
• Since t is a bridge change, one of the edges which is switched in by t is also an edge
which is switched out by S. Guess which one, from 2 alternatives.
• The other edge which is switched out by S is an edge of Z ′ , so this edge and its
orientation can be guessed from fewer than 2rn alternatives.
• Also guess whether W is connected and if not, guess the oriented loopback edge from at
most 2rn alternatives (as in the proof of Lemma 4.2). If W is connected then W = Z.
Otherwise, we must apply the unchaining algorithm from Lemma 4.2 to the graph Z
with respect to the chosen oriented loopback edge. This produces the graph W , and
then using the switch edges we can determine W ′ .
In all we guessed S from 2 · (2rn)2 = 8r 2 n2 alternatives.
(NS) Next, suppose that the connected switch t is a neutral simulating switch for a neutral
(general) switch S.
• The edges switched in/out by t are precisely the edges switched in/out by S. This gives
us the vertices {a, b, c, d} involved in the switch. Guess the (current) oriented loopback
edge in Z from at most 2rn alternatives, and perform the unchaining algorithm to reveal
the chain structure of Z. In particular, we now know the entry vertex and (current)
exit vertex of every component.
28
• If t was not preceded by a bridge change then Z = Z0 . Otherwise, a bridge change
has been performed before t, which implies that Z = Z1 . In this case, one of the edges
switched out by t was originally a bridge edge in W , and hence is incident with vi for
some component Hi of W . Since S is a neutral switch, all vertices involved in the switch
belong to the same component Hi , and hence vi has the largest label among a, b, c, d.
Without loss of generality, if a = vi then b = vi′ , while our knowledge of the chain
structure of Z gives us the current exit vertex vi∗ . Hence we can uniquely reverse the
bridge change switch to determine Z0 . Therefore we can find Z0 by guessing between
just two alternatives, namely, whether or not a bridge change was performed.
From Z0 we obtain W by deleting the chain edges and reinstating the bridge edges
(taking care to adjust the exit vertex of Hi if a bridge change has been performed).
Since we know W and the switch edges, we can determine W ′ .
We have guessed S from 4rn alternatives, when t is a neutral simulating switch. (The above
analysis holds even if the switch S replaces one 2-edge-cut in W by another.)
(BR) Finally, suppose that t is a bridge rectification switch in a simulation path for a neutral
(general) switch S. Then t is the final switch in the simulation path.
• Guess the oriented loopback edge in Z ′ , from at most 2rn alternatives, and perform the
unchaining algorithm on Z ′ to produce W ′ . (Again, this includes the possibility that
W ′ is connected, in which case W ′ = Z ′ .)
• Since t is a bridge rectification, the switch S switches in an edge which is incident with
the entry vertex vi of Hi , giving it a new neighbour with a higher label than the current
exit vertex in Hi . This edge is one of the two edges switched out by t: guess which one,
from 2 alternatives.
• The other edge which is switched in by S, together with its orientation, can be guessed
from the (oriented) edges of Z: at most 2rn possibilities. Now W can be inferred from
W ′ and the switch edges.
We have guessed S from at most 2rn · 2 · 2rn = 8r 2 n2 alternatives.
Adding the contributions from these three roles, a connected switch t may be part of at
most
8r 2 n2 + 4rn + 8r 2n2 ≤ (8r 2 + 1)n2
(4.2)
simulation paths for neutral general switches.
Before we move on to consider disconnecting and reconnecting switches, we make a couple
of comments. The analysis of the neutral simulation switch (NS) shows that once the edges
of the general switch S have been guessed, we have enough information to decide whether
a bridge rectification switch is required and if so, its specification. Since these bridge rectification switches do not affect the number of choices for S in each subcase below, we only
mention them when t itself is playing the role of a bridge rectification, and otherwise we omit
them from our accounting.
29
In every situation we must guess the oriented loopback edge, so that we can perform
the unchaining algorithm. In some cases we must also be careful to specify which graph is
given as input to the unchaining algorithm. In particular, we must not perform the unchaining algorithm immediately after a disconnecting housekeeping switch or immediately before
a reconnecting housekeeping switch, since in those graphs the chaining structure has been
temporarily compromised.
4.2.2
Disconnecting switches
A connected switch t = (Z, Z ′) in MSC could take on five possible roles when simulating a
disconnecting switch S = (W, W ′ ). We consider each separately below.
(B∆) If t is a bridge change then we can guess S from at most 8r 2n2 alternatives, as explained
in Section 4.2.1.
(DHK) Next, suppose that t is a disconnecting housekeeping switch.
• Guess the (current) oriented loopback edge as an edge of Z from at most 2rn possibilities.
Perform the unchaining algorithm on Z to reveal the chain structure of Z. Again, this
tells us the entry and (current) exit vertex of every component of W .
• Decide whether or not a bridge change was performed before t, from 2 possibilities. As
described in Section 4.2.1, we can uniquely determine the bridge change from the chain
structure of Z, so in both cases we can find Z0 (note that Z = Z0 if there was no bridge
change, and Z = Z1 if a bridge change was performed before t). By reversing the bridge
change if necessary we obtain Z0 , and then from our knowledge of the chain structure
we can determine W . (If a bridge change was performed then we must use the original
exit vertex for that component: this vertex is revealed by reversing the bridge change.)
Since t is a disconnected housekeeping switch, it switches out a chain edge vj′ vj+1 , for
′
which is entirely contained within a component Hi , for some
some j, and an edge v+ v+
i 6= j. Hence we can distinguish between these two edges of t and identify which is the
′
chain edge and which is v+ v+
, also determining i and j. Furthermore, we can identify
′
v+ uniquely as it has a higher label than v+
.
• Let vi be the entry vertex in Hi . A disconnected housekeeping switch is only performed
when the disconnected simulation flip (DS) representing S would switch out the two
edges of a 2-edge-cut separating v+ and vi , say. By Lemma 1.2(iv), we can guess which
2-edge-cut is switched out by S from at most n2 /(15r 2 ) possibilities. The orientation
of the switch S is uniquely determined, since there is only one possibility which would
disconnect Hi . Now we know W and the switch edges, we can determine W ′ .
We have guessed S from at most 2rn · 2 · n2 /(15r 2) = 4n3 /(15r) possibilities.
(DS) Now suppose that t is a disconnecting simulation switch. The edges switched in/out by
t are precisely the edges switched in/out by S, so we do not need to guess these. Note that
Z does not have a valid chain structure, but Z ′ does.
30
• We know that W ′ is not connected, so we guess the oriented loopback edge from the
(oriented) edges of Z ′ : at most 2rn possibilities. Perform the unchaining algorithm on
Z ′ , revealing the chain structure of Z ′ . Suppose that Hi and Hj are the two components
of Z ′ which contain an edge which was switched in by t, and suppose that i < j.
Then the disconnecting housekeeping switch (DHK) is uniquely determined: it is the
′
′
switch (vi vi′ , vi−1
vi+1 ⇒ vi vi−1
, vi′ vi+1 ).
• Guess whether or not a bridge change was performed before the disconnecting housekeeping switch: if so, it is uniquely specified (as explained earlier), so we guess from 2
possibilities. Reverse this bridge change to obtain Z0 and hence find W . From W and
the switch edges we can determine W ′ .
We have guessed the general switch S from at most 4rn possibilities.
(The first BR) Now suppose that t is the first bridge rectification switch in the simulation
path for a disconnecting (general) switch S. There may or may not be a second bridge
rectification switch after t. The calculations are similar to the bridge rectification case in
Section 4.2.1.
• Guess the oriented loopback edge in Z ′ , from at most 2rn alternatives, and perform the
unchaining algorithm to reveal the chain structure of Z ′ . (Note, we know that W ′ is
disconnected when S is a disconnecting switch.)
• Since t is a bridge rectification switch, one of the edges switched in by S is vi vi∗ , where
vi is the entry vertex of some component Hi and vi∗ has a higher label than the current
exit vertex vi′ in that component. This edge is then switched out by t. (It is possible
that this condition is satisfied by both edges of S, but if so, we know that the first bridge
rectification will act on the component Hi with the lower index first.) Guess which edge
of t is vi vi∗ , for a factor of 2.
• Guess the other edge switched out by S, and the orientation of the switch, from at most
2rn possibilities. (The other edge switched out by S must lie in a distinct component
from Hi , but for an upper bound we ignore this.) Whether or not a second bridge
change will be needed after t is completely determined by the switch edges and the
chain structure, as described in Section 4.2.1. If a second bridge rectification is needed,
perform it to produce Zq . From Zq we may determine W ′ using our knowledge of the
chain structure of Z ′ (and adjusting the exit vertex of the component in which the
second bridge rectification was performed, if necessary). From W ′ and the switch edges
we may deduce W .
We have guessed S from at most 2rn · 2 · 2rn = 8r 2 n2 alternatives.
(The second BR) Here we assume that one bridge rectification has been already performed
to produce Z. The connected switch t is the last one in the simulation path.
• Again, we guess the oriented loopback edge for Z ′ , from at most 2rn possibilities. Apply
the unchaining algorithm to produce W ′ .
31
• Choose which edge switched out by t is an edge switched in by S, out of 2 possibilities.
The endvertex of this edge with the higher label is vi , for some i.
• Since this is the second bridge rectification, the other edge switched in by S was vj vj∗
for some other component Hj , and this edge has been made into the bridge edge in Hj ,
using the first bridge rectification. Therefore we can identify and orient this edge once
we guess Hj , from at most n/r possibilities (using Lemma 1.2(iii)). This determines the
edges switched in by S. From W ′ and the switch edges, we can determine W .
We have guessed S from at most 2rn · 2 · n/r = 4n2 possibilities.
Summing over these five roles, a connected switch t can be included in at most
8r 2 n2 + 4n3 /(15r) + 4rn + 8r 2 n2 + 4n2 ≤ 9rn3
(4.3)
simulation paths for disconnecting general switches. (The upper bound follows since 2 ≤ r ≤
n/2.)
4.2.3
Reconnecting switches
A connected switch t = (Z, Z ′) in MSC could take on give possible roles when simulation a
reconnecting (general) switch S = (W, W ′ ). We consider each separately below.
(The first B∆) When t is the first bridge change we can guess S from at most 8r 2 n2
alternatives, using arguments very similar to those given in Section 4.2.1.
(The second B∆) Now suppose that t is the second of two bridge changes in the simulation
path for some reconnecting (general) switch S.
• Guess the oriented loopback edge in Z, from at most 2rn possibilities. Perform the
unchaining algorithm to reveal the chain structure of Z.
• Since t is a bridge change, one of the edges switched in by t is an edge which will be
switched out by S. Choose which, from 2 possibilities. This edge is incident with the
entry vertex vj of some component Hj .
• Since t is the second bridge change, the other edge which will be switched out by S
must be incident with the entry vertex of some other component Hi , with i < j. (It has
been put back into Z by the first bridge change.) By Lemma 1.2(iii) there are at most
n/(2r) components, so we choose one to determine vi .
• Choose a non-chain edge incident with vi and orient it, from 2(2r − 1) possibilities.
This determines the edges switched in and out by S. From these edges and W , we may
determine W ′.
We have guessed S from at most 2rn · 2 · n/(2r) · 2(2r − 1) < 8rn2 possibilities.
(RS) Suppose that t is a reconnecting simulation switch. The edges switched in/out by t are
precisely the edges switched in/out by S.
32
• Guess the oriented loopback edge for Z from at most 2rn possibilities. Perform the
unchaining algorithm on Z to reveal the chain structure of Z.
• Up to two bridge changes may have been performed before t, corresponding to the up to
two edges switched out by S which are incident with entry vertices in some component of
Z. If two bridge changes have been performed then their order is uniquely determined.
As argued in Section 4.2.1, we must just decide whether or not a bridge change has
been performed in each component, and then the rest is specified, so there are 22 = 4
possiblities for the bridge changes before t, including the possibility that there were
none. Once the bridge change switches are known, they can be reversed, which reveals
W . Together with the switch edges, this determines W ′ and hence S.
We have guessed S from at most 8rn possibilities.
(RHK) Suppose that t is a reconnecting housekeeping switch. Then Z ′ has a valid chain
structure, though Z does not.
• Guess the oriented loopback edge for Z ′ from at most 2rn possibilities. Perform the
unchaining algorithm on Z ′ to reveal the chain structure of Z ′ .
• The switch t involves entry vertices vi and vj , where i < j. Since t is a reconnecting
housekeeping switch, the edges switched in by S form a 2-edge-cut in Z ′ which separate
vi and vj . By Lemma 1.2(iv) we can guess this 2-edge-cut from at most n2 /(15r 2)
possibilities. This specifies the edges switched out by S. The orientation of the switch
is determined by the fact that one edge switched out by S is contained in Hi and the
other is contained in Hj .
• Finally, we must decide whether or not a bridge rectification is needed to produce W ′ ,
giving 2 possibilities. If a bridge rectification is needed then it is uniquely determined:
perform it to obtain W ′ , and then W can be obtained using the switch edges.
We have guessed S from a total of at most 2rn · n2 /(15r 2 ) · 2 = 4n3 /(15r) possibilities.
(BR) Suppose that t is a bridge rectification for a reconnecting (general) switch S. Then t
is the final switch in the simulation path. In particular, W ′ is disconnected, since otherwise
the reconnecting housekeeping switch produces W ′ , and no BR or unchaining is necessary.
• Guess the oriented loopback edge and perform the unchaining algorithm in Z ′ to produce
W ′ , from at most 2rn alternatives.
• One of the edges switched in by S is an edge which was switched out by t: choose one,
from 2 possibilities, and call it e.
• Let eb be the other edge which is switched in by S. Recall that the edges removed by the
′
RHK are ei−1 = vi vi−1
and ei = vi′ vi+1 , where e′i , e′i−1 are the chain edges in the graph
Zq obtained after the (at most two) bridge changes were performed. (See Section 4.1.3.)
Let Zq+1 be the result of performing the RS simulating S, starting from Zq . Since the
edges switched in by the RS must be disjoint from the edges of Zq , we conclude that
eb 6∈ {e′i−1 , e′i }: the three edges b
e, e′i−1 , e′i are distinct and all belong to Zq+1 . Next,
33
the RHK (Zq+1 , Zq+2) removes precisely two edges, namely e′i−1 , e′i , and hence eb is still
present in Z = Zq+2 . Therefore we can guess and orient b
e from among the 2rn oriented
edges of Z. (In fact, the two edges switched in by S form a 2-edge-cut in Z, but we do
not use that fact here.) Now we know W ′ and the switch edges, we can determine W .
We have guessed S from at most 2rn · 2 · 2rn = 8r 2 n2 possibilities.
Summing the contribution from these five roles, a connected switch t can be included in
at most
8r 2 n2 + 8rn2 + 8rn + 4n3 /(15r) + 8r 2 n2 ≤ 11rn3
(4.4)
simulation paths for reconnecting general switches. (The upper bound follows since 2 ≤ r ≤
n/2.)
4.3
Completing the analysis
Now we apply the two-stage direct canonical path construction to our set of (MSC , MS )simulation paths. Here ρ̄(ΓS ) denotes the congestion of any set of canonical paths, or multicommodity flow, for the switch chain. We will use the bound given in [4].
Lemma 4.3. The (MSC , MS)-simulation paths defined in Section 4.1 define a set of canonical
paths for MSC with congestion ρ̄(ΓSC ) which satisfies
ρ̄(ΓSC ) ≤ 120 rn3 ρ̄(ΓS ).
Proof. First, by adding together (4.2) – (4.4), the maximum number B(ΣSC ) of simulation
paths containing a given connected switch satisfies
B(ΣSC ) ≤ (8r 2 + 1)n2 + 9rn3 + 11rn3 ≤ 24rn3 ,
using the fact that r ≥ 2 and n ≥ 2r + 1 ≥ 5. Next we calculate the simulation gap
DSC = D(MSC , MS). Recall that for all uv ∈ E(MSC ) and zw ∈ E(MS ) we have u 6= v and
z 6= w, and hence
1
PSC (u, v) = PS (z, w) =
,
3an,r
by definition of both chains. The state space of MS is ΩS and the state space of MSC is ΩF ,
with ΩF ⊆ ΩF . Since both MS and MSC have uniform stationary distribution,
DSC =
PS (z, w)
|ΩF | PS (z, w)
≤ max
= 1.
uv∈E(MSC ) PSC (u, v)
uv∈E(MSC ) |ΩS | PSC (u, v)
max
zw∈E(MS )
zw∈E(MS )
Substituting these values and (4.1) into Theorem 2.1 gives
ρ̄(ΓSC ) ≤ DSC ℓ(ΣSC ) B(ΣSC ) ρ̄(ΓS )
≤ 120 rn3 ρ̄(ΓS ),
as required.
34
Finally, we may prove Theorem 1.1.
Proof of Theorem 1.1. Combining the bounds of Lemma 3.5 and Lemma 4.3 gives
ρ̄(ΓF )
≤ 480 (2r)12n7 .
ρ̄(ΓS )
(4.5)
The upper bound on the mixing time of the switch chain given in [4, 5] has (an upper bound
on) ρ̄(ΓS ) as a factor. Therefore, multiplying this bound by the right hand side of (4.5), we
conclude that the mixing time of the flip chain is at most
480 (2r)35 n15 2rn log(2rn) + log(ε−1 ) ,
completing the proof of Theorem 1.1. (This uses the fact that |ΩS | ≤ (rn)rn , which follows
from known asymptotic enumeration results [2].)
It is unlikely that the bound of Theorem 1.1 is tight. Experimental evidence was presented
in an earlier version of this work [6] which provides support for the conjecture that the true
mixing time for the flip chain is O(n log n), when r is constant. Perhaps O(rn log n) is a
reasonable conjecture when r grows with n.
A
Maximising 2-edge-cuts separating two vertices
We now present the deferred proof of Lemma 1.2(iv). Our aim here is to present a simple
upper bound on Λ(u, v) which holds for all relevant values of n and r, though our proof
establishes tighter bounds than the one given in the statement of Lemma 1.2(iv).
Proof of Lemma 1.2(iv). Let Λ(u, v) denote the number of 2-edge-cuts in G which separate
u and v. By considering the connected component of G which contains u, if necessary, we
may assume that G is connected. Consider the binary relation ∼ on V defined by w1 ∼ w2
if and only if there are three edge-disjoint paths between w1 and w2 in G. Then ∼ is an
equivalence relation [19] which partitions the vertex set V into equivalence classes U0 , . . . , Uk .
For j = 0, . . . , k let Hj = G[Uj ] be the subgraph of G induced by Uj . Each Hj is either
a maximal 3-edge-connected induced subgraph of G, or a single vertex. Without loss of
generality, suppose that u ∈ H0 and v ∈ Hk .
Note that the number of edges from Hi to Hj is either 0, 1 or 2 for all i 6= j, by construction.
e of G by replacing each Hj by a single vertex hj , which we
Define the node-link multigraph G
call a node, and replacing each edge from a vertex of Hi to a vertex of Hj by an edge from
hi to hj , which we call a link. In particular, if there is a 2-edge-cut from Hi to Hj in G then
e Each node has even degree in G,
e by Lemma 1.2(i).
the link hi hj has multiplicity 2 in G.
e belongs to a cycle (possibly a 2-cycle, which is a double link),
Furthermore, every link in G
e must be edge-disjoint, or the corresponding edge of
by Lemma 1.2(ii), and these cycles in G
e is a planar multigraph (with edge multiplicity
G cannot be part of a 2-edge-cut. Therefore G
at most two) which has a tree-like structure, as illustrated in Figure 11: the black squares
e A node of G
e which belongs to more than one cycle is a join node.
represent the nodes of G.
35
hk
h0
e
Figure 11: The node-link graph G
e have degree two.
The join nodes each have degree 4, and all other nodes of G
Since we seek an upper bound on Λ(u, v) for a given number of vertices n, we may assume
e lie on some path from h0 to hk in G.
e (This corresponds to trimming all
that all nodes of G
e which do not lie on the unique “path” from h0 to hk in G,
e viewed as a tree.)
“branches” of G
This only removes 2-edge-cuts of G which do not separate u and v. Hence for an upper bound,
e consists of ℓ cycles C1 , . . . , Cℓ ,
by relabelling the nodes if necessary, we can assume that G
where h0 belongs only to C1 and hk belongs only to Cℓ , such that Cj and Cj+1 intersect at a
single join node hj , for j = 1, . . . , ℓ − 1, while all other cycles are disjoint. This situation is
illustrated in Figure 12. (In particular, h0 and hk are not join nodes.)
C1
h0
C2
C3
h1
h2
C4
hk
h3
Figure 12: The node-link graph after trimming: an example with ℓ = 4 cycles.
Each cycle Cj can be considered as the disjoint union of two paths between hj−1 and hj ,
or between hℓ−1 and hk if j = ℓ. Denote by aj , a′j the lengths of these two paths around
e which
Cj . Any 2-edge-cut which separates u from v in G corresponds to a pair of links in G
both belong to some cycle Cj , with one link from each of the two paths which comprise Cj .
P
Hence Λ(u, v) = ℓj=1 aj a′j . If the length aj + a′j of each cycle is fixed then the quantity aj a′j
P
is maximised when |aj − a′j | ≤ 1. Therefore Λ(u, v) ≤ ℓi=1 s2i , where sj = |Cj |/2 ≥ 1. (This
upper bound holds
P evenwhen some cycles have odd length.) Now, the total number of nodes
ℓ
e
in G is N = 2
j=1 sj − (ℓ − 1), by inclusion-exclusion. Thus, for a given number of nodes
N and a given value of ℓ, the bound S is maximised by setting sj = 1 for j = 2, . . . , ℓ, and
s1 = (N − ℓ + 1)/2. Therefore
(N − ℓ + 1)2
+ ℓ − 1.
Λ(u, v) ≤ T =
4
(A.1)
First suppose that r ≥ 3. Then every subgraph Hj must contain a vertex which is only
36
e has degree at most 4. Therefore each
adjacent to other vertices of Hj , since every node in G
e Note that
Hj must contain at least 2r + 1 vertices, so there are N ≤ n/(2r + 1) nodes in G.
dT /dℓ = −ℓ + 1 which equals zero when ℓ = 1 and is negative when ℓ > 1. So T is largest
when ℓ = 1, giving Λ(u, v) ≤ T ≤ N 2 /4 ≤ n2 /(4r + 2)2 , as claimed. An extremal graph G
is shown in Figure 13: each grey block is K2r+1 minus an edge. This construction gives an
extremal example for any positive n = 0 mod (4r + 2), so this upper bound is asymptotically
tight. Certainly the weaker bound Λ(u, v) ≤ n2 /(15r 2 ) also holds when r ≥ 3.
v
u
Figure 13: Extremal graph when r ≥ 3.
Now suppose that r = 2 and consider the induced subgraphs H0 , . . . , Hk . Again we may
e then Hj may be as
assume that s2 = · · · = sℓ = 1, giving (A.1). If hj is a join node of G
e has degree 4. We call such a join node
small as a single vertex, since each join node of G
hj a singleton. Otherwise, if Hj contains a vertex whose neighbourhood is contained within
Hj then Hj contains at least 2r + 1 = 5 vertices. In particular, this must hold for hk and
for all nodes in C1 except h1 . A final option (only available when r = 2) is that a join node
hj corresponds to to a subgraph Hj which is isomorphic to K4 , and hence has precisely four
vertices: two of these vertices are joined to a vertex in Hj−1 and two are joined to Hj+1 (or
Hk , if j = ℓ − 1). Here we use the fact that any part Uj of the partition of V with more than
one vertex must have at least 4 vertices, as G[Uj ] must be 3-edge-connected. Since G has no
repeated vertices, at most every second join node can be a singleton, with the remaining join
nodes contributing at least 4 vertices to G.
It follows that Λ(u, v) ≤ T = s21 + ℓ − 1 and
(
10s1 + 5(ℓ − 1)/2 if ℓ is odd,
ℓ−1
ℓ−1
+4
=
n ≥ 10s1 +
2
2
10s1 + 5ℓ/2 − r
if ℓ is even.
For an upper bound on Λ(u, v) we take equality the above expression for n and assume that ℓ
is even, leading to Λ(u, v) ≤ T = s21 − 4s1 + 2(n + 4)/5. (If ℓ is odd then only the term which
is independent of s1 changes.) Hence dT /ds1 is negative when s1 = 1, zero when s1 = 2 and
positive for s1 ≥ 3. It follows that for a given value of n, the maximum possible value of T
occurs when s1 = 1 or when s1 is as large as possible.
When s1 = 1 we have Λ(u, v) ≤ ℓ ≤ 2(n − 6)/5, at least when ℓ is even. A graph G on
n vertices which meets this bound can be constructed whenever n ≥ 11 and n = 1 mod 5.
The example shown in Figure 14 has n = 21 vertices and ℓ = 6 2-edge-cuts between u and v.
Since 2(n − 6)/5 ≤ n2 /60 for all values of n, this establishes the required bound.
37
u
v
Figure 14: Maximum number of separating 2-edge-cuts for n = 21
Next, consider the subcase when s1 is as large as possible, for a fixed value of n. This is
achieved by setting ℓ = 1, so there are no join nodes. Then every Hj has at least 5 vertices, so
for an upper bound we take s1 = n/10 and Λ(u, v) ≤ s21 = n2 /100 < n2 /60. This completes
the proof.
References
[1] Z. Allen-Zhu, A. Bhaskara, S. Lattanzi, V. Mirrokni and L. Orecchia, Expanders via
local edge flips, in Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete
Algorithms, New York–Philadelpha, 2016, pp. 259–269.
[2] E. Bender and E.R. Canfield, The asymptotic number of labelled graphs with given
degree sequences, Journal of Combinatorial Theory (Series A) 24 (1978), 296–307.
[3] V. Bourassa and F. Holt. Swan: Small-world wide area networks. In Proceedings of
International Conference on Advances in Infrastructure (SSGRR 2003w), L’Aquila, Italy,
2003. Paper #64.
[4] C. Cooper, M. Dyer, and C. Greenhill. Sampling regular graphs and a peer-to-peer
network. Comb. Probab. Comput., 16(4):557–593, 2007.
[5] C. Cooper, M.Dyer and C. Greenhill, Corrigendum: Sampling regular graphs and a
peer-to-peer network. arXiv:1203.6111
[6] C. Cooper, M. Dyer, and A. Handley. The flip Markov chain and a randomising P2P
protocol. In Proceedings of the 28th ACM Symposium on Principles of Distributed Computing, ACM Press, New York, 2009.
[7] P. Diaconis and L. Saloff-Coste. Comparison theorems for reversible Markov chains.
Annals of Applied Probability, 3:696–730, 1993.
[8] R. Diestel, Graph Theory (4th edn.), Springer, Berlin, 2010.
[9] T. Feder, A. Guetz, M. Mihail, and A. Saberi. A local switch Markov chain on given
degree graphs with application in connectivity of peer-to-peer networks. FOCS, 2006.
[10] V. Guruswami. Rapidly mixing Markov chains: A comparison of techniques (a survey).
Unpublished manuscript, 2000. arXiv:1603.01512
38
[11] A.J. Handley, Self-organising, self-randomising peer-to-peer networks, Ph.D. thesis, University of Leeds, 2013.
[12] M. Jerrum. Counting, Sampling and Integrating: Algorithms and Complexity. Birkhäuser,
2003.
[13] M. Jerrum and A. Sinclair.
18(6):1149–1178, 1989.
Approximating the permanent.
SIAM J. Comput.,
[14] R. Kannan, P. Tetali, and S. Vempala. Simple markov-chain algorithms for generating
bipartite graphs and tournaments. Random Struct. Algorithms, 14(4):293–308, 1999.
[15] J. Petersen, Die Theorie der regulären graphs, Acta Mathematica 15 (1891), 193–220.
[16] P. Mahlmann and C. Schindelhauer. Peer-to-peer networks based on random transformations of connected regular undirected graphs. In SPAA ’05: Proceedings of the 17th
Annual ACM Symposium on Parallelism in algorithms and architectures, pages 155–164,
New York, NY, USA, 2005. ACM.
[17] A. Sinclair. Improved bounds for mixing rates of Markov chains and multicommodity
flow. Combinatorics, Probability and Computing, 1:351–370, 1992.
[18] R. Taylor, Constrained switching in graphs, in Combinatorial Mathematics VIII, Springer
Lecture Notes in Mathematics vol. 884, 1981, pp.314–336.
[19] W.T. Tutte, Connectivity in Graphs, University of Toronto Press, Toronto, 1966.
[20] N. C. Wormald. Models of random regular graphs. In J.D. Lamb and D.A. Preece, editors,
Survey in Combinatorics, London Mathematical Society Lecture Notes 276. Cambridge
University Press, Cambridge, 1999.
39
| 8cs.DS
|
lpopt: A Rule Optimization Tool for Answer Set
Programming
arXiv:1608.05675v2 [cs.LO] 23 Aug 2016
Manuel Bichler, Michael Morak, and Stefan Woltran
TU Wien, Vienna, Austria
{surname}@dbai.tuwien.ac.at
Abstract. State-of-the-art answer set programming (ASP) solvers rely on a program called a grounder to convert non-ground programs containing variables into
variable-free, propositional programs. The size of this grounding depends heavily
on the size of the non-ground rules, and thus, reducing the size of such rules is
a promising approach to improve solving performance. To this end, in this paper we announce lpopt, a tool that decomposes large logic programming rules
into smaller rules that are easier to handle for current solvers. The tool is specifically tailored to handle the standard syntax of the ASP language (ASP-Core)
and makes it easier for users to write efficient and intuitive ASP programs, which
would otherwise often require significant hand-tuning by expert ASP engineers.
It is based on an idea proposed by Morak and Woltran (2012) that we extend significantly in order to handle the full ASP syntax, including complex constructs
like aggregates, weak constraints, and arithmetic expressions. We present the algorithm, the theoretical foundations on how to treat these constructs, as well as
an experimental evaluation showing the viability of our approach.
1
Introduction
Answer set programming (ASP) [15,17,8,13] is a well-established logic programming
paradigm based on the stable model semantics of logic programs. Its main advantage
is an intuitive, declarative language, and the fact that, generally, each answer set of a
given logic program describes a valid answer to the original question. Moreover, ASP
solvers—see e.g. [14,1,12,2]—have made huge strides in efficiency.
A logic program usually consists of a set of logical implications by which new facts
can be inferred from existing ones, and a set of facts that represent the concrete input
instance. Logic programming in general, and ASP in particular, have also gained popularity because of their intuitive, declarative syntax. The following example illustrates
this:
Example 1. The following rule naturally expresses the fact that two people are relatives
of the same generation up to second cousin if they share a great-grandparent.
uptosecondcousin(X, Y) :parent(X, PX), parent(PX, GPX),
parent(GPX, GGP), parent(GPY, GGP),
parent(PY, GPY), parent(Y, PY), X != Y.
Rules written in an intuitive fashion, like the one in the above example, are usually
larger than strictly necessary. Unfortunately, the use of large rules causes problems for
current ASP solvers since the input program is grounded first (i.e. all the variables in
each rule are replaced by all possible, valid combinations of constants). This grounding step generally requires exponential time for rules of arbitrary size. In practice, the
grounding time can thus become prohibitively large. Also, the ASP solver is usually
quicker in evaluating the program if the grounding size remains small.
In order to increase solving performance, we could therefore split the rule in Example 1 up into several smaller ones by hand, keeping track of grandparents and greatgrandparents in separate predicates, and then writing a smaller version of the second
cousin rule. While this is comparatively easy to do for this example, this can become
very tedious if the rules become even more complex and larger, maybe also involving
negation or arithmetic expressions. However, since current ASP grounders and solvers
become increasingly slower with larger rules, and noting the fact that ASP programs
often need expert hand-tuning to perform well in practice, this represents a significant
entry barrier and contradicts the fact that logic programs should be fully declarative: in
a perfect world, the concrete formulation should not have an impact on the runtime. In
addition, to minimize solver runtime in general, it is therefore one of our goals to enable logic programs to be written in an intuitive, fully declarative way without having
to think about various technical encoding optimizations.
To this end, in this paper we propose the lpopt tool that automatically optimizes
and rewrites large logic programming rules into multiple smaller ones in order to improve solving performance. This tool, based on an idea proposed for very simple ASP
programs in [18], uses the concept of tree decompositions of rules to split them into
smaller chunks. Intuitively, via a tree decomposition joins in the body of a rule are arranged into a tree-like form. Joins that belong together are then split off into a separate
rule, only keeping the join result in a temporary atom. We then extend the algorithm
to handle the entire standardized ASP language [10], and also introduce new optimizations for complex language constructs such as weak constraints, arithmetic expressions,
and aggregates.
The main contributions of this paper are therefore as follows:
– we extend, on a theoretical basis, the lpopt algorithm proposed in [18] to the
full syntax of the ASP language according to the ASP-Core-2 language specification [10];
– we establish how to treat complex constructs like aggregates, and propose an adaptation of the decomposition approach so that it can split up large aggregate expressions into multiple smaller rules and expressions, further reducing the grounding
size;
– we implement the lpopt algorithm in C++, yielding the lpopt tool for automated
logic program optimization, and give an overview of how this tool is used in practice; and
– we perform an experimental evaluation of the tool on the encodings and instances
used in the fifth Answer Set Programming Competition which show the benefit of
our approach, even for encodings already heavily hand-optimized by ASP experts.
2
Preliminaries
General Definitions. We define two pairwise disjoint countably infinite sets of symbols:
a set C of constants and a set V of variables. Different constants represent different
values (unique name assumption). By X we denote sequences (or, with slight notational
abuse, sets) of variables X1 , . . . , Xk with k > 0. For brevity, let [n] = {1, . . . , n}, for
any integer n > 1.
A (relational) schema S is a (finite) set of relational symbols (or predicates). We
write p/n for the fact that p is an n-ary predicate. A term is a constant or variable.
An atomic formula a over S (called S-atom) has the form p(t), where p ∈ S and t
is a sequence of terms. An S-literal is either an S-atom (i.e. a positive literal), or an
S-atom preceded by the negation symbol “¬” (i.e. a negative literal). For a literal `,
we write dom(`) for the set of its terms, and var (`) for its variables. This notation
naturally extends to sets of literals. For brevity, we will treat conjunctions of literals as
sets. For a domain C ⊆ C, a (total or two-valued) S-interpretation I is a set of S-atoms
containing only constants from C such that, for every S-atom p(a) ∈ I, p(a) is true,
and otherwise false. When obvious from the context, we will omit the schema-prefix.
A substitution from a set of literals L to a set of literals L0 is a mapping s : C∪V →
C ∪ V that is defined on dom(L), is the identity on C, and p(t1 , . . . , tn ) ∈ L (resp.
¬p(t1 , . . . , tn ) ∈ L) implies p(s(t1 ), . . . , s(tn )) ∈ L0 (resp., ¬p(s(t1 ), . . . , s(tn )) ∈
L0 ).
Answer Set Programming (ASP). A logic programming rule is a universally quantified
reverse first-order implication of the form
H(X, Y) ← B + (X, Y, Z, W) ∧ B − (X, Z),
where H (the head), resp. B + (the positive body), is a disjunction, resp. conjunction, of atoms, and B − (the negative body) is a conjunction of negative literals, each
over terms from C ∪ V. For a rule π, let H (π), B + (π), and B − (π) denote the set
of atoms occurring in the head, the positive, and the negative body, respectively. Let
B (π) = B + (π) ∪ B − (π). A rule π where H (π) = ∅ is called a constraint. Substitutions naturally extend to rules. We focus on safe rules where every variable in the rule
occurs in the positive body. A rule is called ground if all its terms are constants. The
grounding of a rule π w.r.t. a domain C ⊆ C is the set of rules ground C (π) = {s(π) |
s is a substitution, mapping var (π) to elements from C}.
A logic program Π is a finite set of logic programming rules. The schema of a
program Π, denoted sch(Π), is the set of predicates appearing in Π. The active domain
of Π, denoted adom(Π), with adom(Π) ⊂ C, is the set of constants appearing in
Π. A program Π is ground if all its rules S
are ground. The grounding of a program
Π is the ground program ground (Π) = π∈Π ground adom(Π) (π). The (GelfondLifschitz) reduct of a ground program Π w.r.t. an interpretation I is the ground program
Π I = {H (π) ← B + (π) | π ∈ Π, B − (π) ∩ I = ∅}.
A sch(Π)-interpretation I is a (classical) model of a ground program Π, denoted
I Π if, for every ground rule π ∈ Π, it holds that I ∩ B + (π) = ∅ or I ∩ (H (π) ∪
B − (π)) 6= ∅, that is, I satisfies π. I is a stable model (or answer set) of Π, denoted I s
Π if, in addition, there is no J ⊂ I such that J Π I , that is, I is subset-minimal w.r.t.
the reduct Π I . The set of answer sets of Π, denoted AS (Π), are defined as AS (Π) =
{I | I is a sch(Π)-interpretation, and I s Π}. For a non-ground program Π, we
define AS (Π) = AS (ground (Π)). When referring to the fact that a logic program is
intended to be interpreted under the answer set semantics, we often refer to it as an ASP
program.
Tree Decompositions. A tree decomposition of a graph G = (V, E) is a pair T =
(T, χ), where T is a rooted tree and χ is a labelling function over nodes t of T , with
χ(t) ⊆ V called the bag of t, such that the following holds: (i) for each v ∈ V , there
exists a node t in T , such that v ∈ χ(t); (ii) for each {v, w} ∈ E, there exists a node t
in T , such that {v, w} ⊆ χ(t); and (iii) for all nodes r, s, and t in T , such that s lies on
the path from r to t, we have χ(r) ∩ χ(t) ⊆ χ(s). The width of a tree decomposition
is defined as the cardinality of its largest bag minus one. The treewidth of a graph G,
denoted by tw (G), is the minimum width over all tree decompositions of G. To decide
whether a graph has treewidth at most k is NP-complete [3]. For an arbitrary but fixed
k however, this problem can be solved (and a tree decomposition constructed) in linear
time [6].
Given a non-ground logic programming rule π, we let its Gaifman graph Gπ =
(var (π), E) such that there is an edge (X, Y ) in E iff variables X and Y occur together
in the head or in a body atom of π. We refer to a tree decomposition of Gπ as a tree
decomposition of rule π. The treewidth of rule π is the treewidth of Gπ .
3
Rule Decomposition
This section lays out the theoretical foundations for our rule decomposition approach.
First, we recall the algorithm from [18], and then describe how it can be extended to
handle three of the main extensions of the ASP language, namely arithmetic expressions, aggregates, and weak constraints (i.e. optimization statements), as defined in the
ASP-Core language standard [10].
As demonstrated in Example 1, rules that are intuitive to write and read are not necessarily the most efficient ones to evaluate in practice. ASP solvers generally struggle
with rules that contain many variables since they rely on a grounder-solver approach:
first, the grounding of a logic program is computed by a grounder. As per the definition
in Section 2, the size of the grounding can, in the worst case, be exponential in the
number of variables. For large rules, the grounding step can already take a prohibitively
large amount of time. However, the solver is also adversely affected by this blowup. In
practice, this leads to long runtimes and sometimes the inability of the ASP system to
solve a given instance. This also contributes to the fact that, while the syntax of ASP is
fully declarative, writing efficient encodings still takes expert knowledge.
It is therefore desirable to have a way to automatically rewrite such large rules into
a more efficient representation. One way to do this is the rule decomposition approach,
first proposed in [18], which we will briefly recall next.
3.1
Decomposition of Simple Rules
Generally speaking, the approach in [18] computes the tree decomposition of a rule,
and then splits the rule up into multiple, smaller rules according to this decomposition.
While in the worst case this decomposition may not change the rule at all, in practice it
is often the case that large rules can be split up very well. For instance, the large rule in
Example 1 will be amenable for such a decomposition.
Let us briefly recall the algorithm from [18] which we will refer to as the lpopt
algorithm. For a given rule π, the algorithm works as follows:
1. Compute a tree decomposition T = (T, χ) of π with minimal width where all
variables occurring in the head of π are contained in its root node bag.
2. For each node n, let temp n be a fresh predicate, and the same for each variable X
in π and predicate dom X . Let Yn = χ(n) ∩ χ(pn ), where pn is the parent node
of n. For the root node root, let temp root be the entire head of π, and, accordingly,
Yroot = var (H (π)). Now, for a node n, generate the following rule:
temp n (Yn ) ←
{a ∈ B (π)
∪ {dom X (X)
| var (a) ⊆ χ(n)}
| a ∈ B − (π), X ∈ var (a), var (a) ⊆ χ(n),
6 ∃b ∈ B + (π) : var (b) ⊆ χ(n), X ∈ var (b)}
∪ {temp m (Ym ) | m is a child of n}.
3. For each X ∈ var (B − (π)), for which a domain predicate dom is needed to
guarantee safety of a rule generated above, pick an atom a ∈ B + (π), such that
X ∈ var (a) and generate a rule
domX (X) ← a.
Step 3 is needed because splitting up a rule may make it unsafe. In order to remedy
this, a domain predicate is generated for each unsafe variable that arises due to the rule
splitting in step 2. The following example illustrates how the algorithm works.
Example 2. Given the rule
π = h(X, W ) ← e(X, Y ), e(Y, Z), ¬e(Z, W ), e(W, X),
a tree decomposition of π could look as follows (note that we write in each bag of the
tree decomposition not just the variables as per definition but also all literals of rule π
over these variables which is a more intuitive notation):
h(X, W ), e(X, Y ), e(W, X)
e(Y, Z), ¬e(Z, W )
Applying the lpopt algorithm to π with the tree decomposition above yields the following set of rules lpopt(π):
domW (W ) ← e(W, X),
temp(Y, W ) ← e(Y, Z), ¬e(Z, W ), domW (W ), and
h(X, W ) ← e(X, Y ), e(W, X), temp(Y, W ),
where temp is a fresh predicate not appearing anywhere else.
Let Π be a logic program. When the above algorithm is applied to all rules in
Π, resulting in a logic program lpopt(Π) as stated in [18], the answer sets of Π are
preserved in the following way: when all temporary atoms are removed, each answer
set of lpopt(Π) coincides with exactly one answer set from the original program Π.
Furthermore, the size of the grounding now no longer depends on the rule size. In fact,
it now only depends on the rule treewidth as the following result states:
Theorem 1 ([18]). The size of ground (lpopt(Π)) is bounded by O(2k · n), where n is
the size of Π, and k is the maximal treewidth of the rules in Π.
The above theorem implies that the size of the grounding of a program Π, after
optimization via the lpopt algorithm, is no longer exponential in the size of Π, but only
in the treewidth of its rules. As [18] demonstrates, this decomposition approach already
has a significant impact on the size of the grounding in practical instances.
However, the ASP language standard [10] extends the ASP language with other
useful constructs that the lpopt algorithm proposed in [18] cannot handle. These include
arithmetic expressions, aggregates, and weak constraints. Looking at concrete, practical
instances of ASP programs, e.g. the encodings used in recent ASP competitions [11], a
large majority use such constructs. In the following, we will therefore extend the lpopt
algorithm to be able to treat them in a similar way.
3.2
Treating Arithmetic Expressions
Arithmetic expressions are atoms of the form X = ϕ(Y), that is, an equality with
one variable (or constant number) X on the left-hand side, and an expression ϕ on the
right-hand side, where ϕ is any mathematical expression built using the variables from
Y, constant numbers, and the arithmetic connectives “+,” “-,” “*,” and “/.” In addition
to the positive and negative body, a rule π may also contain a set of such arithmetic
expressions describing a relationship between variables with the obvious meaning.
Clearly, in order to adapt the rule decomposition approach to this it is easy to extend the definition of the graph representation of π to simply contain a clique between
all variables occurring together in an arithmetic expression. The lpopt algorithm then
works as described above up to step 2. However, a problem may arise when, in step 3
of the lpopt algorithm, a domain predicate domX (X) is to be generated. Consider the
following example:
Example 3. Let π be the rule a(X) ← ¬b(X, Y ), c(Y ), d (Z), X = Z + Z. A simple
decomposition according to the lpopt algorithm may lead to the following rules:
temp(X) ← ¬b(X, Y ), c(Y ), domX (X), and
a(X) ← d (Z), X = Z + Z, temp(X).
It remains to define the domain predicate domX . According to the original definition of
lpopt, we would get
domX (X) ← X = Z + Z
which is unsafe.
The conditions for safety of rules with arithmetic expressions are defined in the ASP
language specification [10]. As Example 3 shows, in order for such expressions to work
with the lpopt algorithm a more general approach to defining the domain predicates
is needed in step 3. In fact, instead of choosing a single atom from the rule body to
generate the domain predicate, in general a set of atoms and arithmetic expressions must
be chosen. It is easy to see that if a rule π is safe then, for each variable X ∈ B (π),
there is a set A of (positive) atoms and arithmetic expressions in the body of π that
makes that variable safe. In step 3 of the lpopt algorithm, for a variable X we now
choose such a set AX of body elements in a greedy fashion as follows: let S = {X} the
set of variables that we need to make safe. For each variable S ∈ S, pick a (positive)
atom from B (π) that makes S safe, add it to AX , and remove S from S. If no such
atom exists in the body of π, greedily add the smallest arithmetic expression S = ϕ(Y)
in B (π) to AX and let S = S \ {S} ∪ Y. Repeat this process until S is empty. Since
π itself is safe and finite in size, the above procedure necessarily terminates. Finally,
generate the rule domX (X) ← AX . It is easy to see that this rule is safe and describes
the possible domain of variable X as required. Note also that this rule can not be split
up futher as removing any single element of the rule would make it unsafe.
Example 4. A correct domain predicate for Example 3 would be defined as follows:
domX (X) ← X = Z + Z, d (Z).
This ensures the proper safety of all rules generated by the lpopt algorithm.
Note that the rule generated in Example 4 repeats most of the atoms that the second
rule generated in Example 3 already contains. It is not immediately obvious how such
situations can be remedied in general. Investigating this issue is part of ongoing work.
3.3
Treating Weak Constraints
As defined in [10], a weak constraint π[k : t] is a constraint π annotated with a term k
representing a weight and a sequence of terms t occurring in π. The intended meaning
is that each answer set I is annotated by a total weight w(I), which is the sum over
all k for each tuple of constants c that realize t in I and satisfy the body of π. Such
a weak constraint can easily be decomposed by replacing π[k : t] with the rule π 0 =
temp(k, t) ← B (π), where temp is a fresh predicate, and the weak constraint ⊥ ←
temp(k, t)[k : t]. Finally, the lpopt algorithm is then applied to rule π 0 . This allows our
rule decomposition approach also to be applied in an optimization context (i.e. where
the task for the solver is to find optimal answer sets w.r.t. their weight).
3.4
Treating Aggregate Expressions
An aggregate expression, as defined in [10], is an expression of the form
t 4 #agg{t : ϕ(X)},
where t is a term; 4 ∈ {<, 6, =, 6=, >, >} is a builtin relation; agg is one of sum,
count, max , and min; t = ht1 , . . . , tn i is a sequence of terms; and ϕ(X) is a set of
literals, arithmetic expressions, and aggregate expressions, called the aggregate body.
Aggregates may appear in rule bodies, or recursively inside other aggregates, with the
following semantic meaning: Given an interpretation I, for each valid substitution s
such that s(ϕ(X)) ⊆ I, take the tuple of constants s(t). Let us denote this set with T .
Now, execute the aggregate function on T as follows: for #count, calculate |T |; for
#sum, calculate Σt∈T t1 , where t1 is the first term in t; for #max and #min, take
the maximum and minimum term appearing in the first position of each tuple in T ,
respectively. Finally, an aggregate expression is true if the relation 4 between term t
and the result of the aggregate function is fulfilled.
Extending the lpopt algorithm to aggregate expressions is again straightforward:
The rule graph Gπ = (V, E) of a rule π containing aggregate expressions is defined as
follows: Let V be the set of variables occurring in π outside of aggregate expressions.
Let E be as before and, in addition, add, for each aggregate expression e, a clique
between all variables var (e) ∩ V to E. Intuitively, the rule graph should contain, for
each aggregate expression, a clique between all variables that appear in the aggregate
and somewhere else in the rule. Variables appearing only in aggregates are in a sense
“local” and are therefore not of interest when decomposing the rule.
While the above transformation is straightforward, we can, however, go one step
further and also decompose the inside elements of an aggregate expression. To this
end, let t 4 #agg{t : ϕ(X, Y)} be an aggregate expression occurring in some
rule π, where X are variables that occur either in t or somewhere else in π, and
Y are variables occurring inside the aggregate only. Replace the aggregate expression with t 4 #agg{t : ψ(X, Z), temp(t, Z)}, and furthermore, generate a rule
temp(t, Z) ← ψ(Y), ψ dom (Y), for some fresh predicate temp. Here, ψ contains all
those atoms from ϕ that contain a variable from X, and ψ contains the rest. ψ dom contains domain predicates generated like in step 3 of the lpopt algorithm, as needed to
make the temporary rule safe. The temporary rule can then be decomposed via lpopt.
This is best illustrated by an example:
Example 5. Let π be the following logic programming rule, saying that a vertex is
“good” if it has at least two neighbours that, themselves, have a red neighbour:
good (X) ← vertex (X), 2 6 #count{Y : edge(X, Y ), edge(Y, Z), red (Z)}.
According to the above approach, the rule can now be split up as follows. Firstly, the
aggregate is replaced:
good (X) ← vertex (X), 2 6 #count{Y : edge(X, Y ), temp(Y )},
and furthermore, a temporary rule is created as follows:
temp(Y ) ← edge(Y, Z), red (Z).
The latter rule is now amenable for decomposition via the lpopt algorithm.
Note that the above approach allows us to decompose, to a degree, even the insides
of an aggregate, which, for large aggregate bodies, can lead to a further significant
reduction in the grounding size.
3.5
Correctness
The correctness of the above extensions to the original algorithm follows by the same
arguments that prove the correctness of the original algorithm proposed in [18], and
trivially from the construction for arithmetic expressions and safety. For the latter, note
that for domain predicates of a variable X we explicitly select a set of atoms that make
the variable safe, and that such a set always exists, since the original rule is safe. For
the former two (namely weak constraints and aggregate expressions), the only thing
that needs to be examined is the first step: replacing (part of) the body with a temporary
predicate. But correctness of this is easy to see. Instead of performing all joins within the
weak constraint or aggregate, we perform the join in a new, separate rule and project
only relevant variables into a temporary predicate. The weak constraint or aggregate
then only needs to consider this temporary predicate since, by construction, all other
variables not projected into the temporary predicate do not play a role w.r.t. optimization
or aggregation. Finally, the original algorithm from [18] extended to handle arithmetic
expressions, for which correctness has already been established, is then applied to this
new, separate rule.
3.6
Further Language Extensions
The ASP-Core language specification [10], as well as the gringo grounder1 , allow further constructs like variable pooling, aggregates with multiple bodies, or with upper
and lower bounds in the same expression, in addition to various extensions that amount
to syntactic sugar. These constructs make the above explanations unnecessarily more
tedious. However, from a theoretical point of view, all of these additional constructs
can be normalized to one of the forms discussed in the previous subsections. Furthermore, as we shall see in the next section, we have implemented the lpopt algorithm to
directly treat all standard ASP language constructs and certain other additions, like variable pooling. More details about this general approach, and the exact, but more tedious,
algorithm details, can be found in [4].
4
Implementation
A full implementation of the algorithm and its extensions described in Section 3 is now
available in the form of the lpopt tool, available with relevant documentation and
examples at http://dbai.tuwien.ac.at/proj/lpopt. The following gives
a quick outline of how to use the tool.
lpopt accepts as its input any form of ASP program that follows the ASP input
language specification laid out in [10]. The output of the program in its default configuration is a decomposed program that also follows this specification. In addition, the tool
guarantees that no language construct is introduced in the output that was not previously
present in the input (cf. Section 3). Therefore, for example, a program without aggregates will not contain any aggregates as a result of rule decomposition. The following
is a description of the parameters of the tool:
1
http://potassco.sourceforge.net
Usage: lpopt [-idbt] [-s seed] [-f file] [-h alg] [-l file]
-d
dumb: do not perform optimization
-b
print verbose and benchmark information
-t
perform only tree decomposition step
-i
ignore head variables when decomposing
-h alg
decomposition algorithm, one of {mcs, mf, miw (def)}
-s seed
initialize random number generator with seed.
-f file
the file to read from (default is stdin)
-l file
output infos (treewidth) to file
In what follows, we will briefly describe the most important features of the tool.
Tree Decomposition Heuristics. As stated in Section 2, computing an optimal tree decomposition w.r.t. width is an NP-hard problem. We thus make use of several heuristic
algorithms, namely the maximum cardinality search (mcs), minimum fill (mf), and minimum induced width (miw) approaches described in [7], that yield tree decompositions
that provide good upper bounds on the treewidth (i.e. on an optimal decomposition).
It turns out that in practice, since rules in ASP programs are usually not overly large,
these heuristics come close to, and often even yield, an optimal tree decomposition for
rules. The heuristic algorithm to use for decomposition can be selected using the -h
command line parameter. Since these heuristic approaches rely to some degree on randomization, a seed for the pseudo-random number generator can be passed along with
the -s command line parameter.
Measuring the Treewidth of Rules. Theorem 1 allows us to calculate an upper bound
on the size of the grounding of the input program. In order to do this, the maximal
treewidth of any rule in an ASP program must be known. The -l switch of the lpopt
tool allows this to be calculated. It forces the tool to perform tree decompositions on
all rules inside an input ASP program, simply outputting the maximal treewidth (or,
more accurately, an upper bound; see above) over all of them into the given file, and
then exiting. Clearly, when a single ASP rule is given as input, this switch will output a
treewidth upper bound of that single rule.
Recommended Usage
Assuming that a file enc.lp contains the encoding of a problem as an ASP program
and that a file instance.db contains a set of ground facts representing a problem
instance, the recommended usage of the tool is as follows:
cat enc.lp instance.db | lpopt | grounder | solver
In the above command, grounder and solver are programs for grounding and for
solving, respectively. One established solver that we will use in the next section for
our experimental evaluation is clasp [14]. If clasp is used as a solver together with
the lpopt tool, we generally recommend the use of the --sat-prepro flag, which
often speeds up the solving process substantially for decomposed rules generated by
lpopt (by considering the fact that the truth values of all temporary atoms generated
by lpopt are determined exactly by the rule body, and need never be guessed).
5
Experimental Evaluation
We have tested our lpopt tool and benchmarked the performance of grounding and
solving of programs preprocessed with lpopt against non-preprocessed ones. All
benchmarks were made on the instance sets of the fifth answer set programming competition 2014 2 , which, for most problem classes, provides two encodings, one from 2013,
and one from 2014. The benchmarks have been run on a 3.5GHz AMD Opteron Processor 6308 with 192 GB of RAM to its disposal. We used the potassco software suite3 ,
namely gringo verison 4.5.3 as the grounder and clasp version 3.1.3 as the solver. A
timeout of 300 seconds was set for solving, and 1000 seconds for grounding. Furthermore, as suggested in the previous section, clasp was called with the --sat-prepro
flag enabled. In this paper, we will survey the most important results.
Remark. One central aim of our tool is to improve solving performance for handwritten encodings by non-experts of ASP. In the spirit of a truly declarative language,
it shouldn’t matter how an encoding is written as long as it is correct (i.e. w.r.t. runtime, there should not be a difference between “good” and “bad” encodings). In this respect, the ASP competition does not offer an optimal benchmark set since all encodings
are extensively hand-tuned by ASP experts. However, as to the best of our knowledge
there is no better-suited comprehensive benchmark set available, we will show that even
for these extensively hand-tuned ASP competition encodings our tool can still find decompositions that decrease grounding size and improve solving performance. However,
there are also encodings that are so perfectly hand-tuned that only trivial optimizations
are possible with the current version of lpopt.
Results. Let us first note that the runtime of lpopt itself, for all encodings in the
benchmark set, was always less than what can be accurately measured on a computer
system today. Applying our rule decomposition algorithm thus comes virtually for free
for hand-written encodings. Out of the 49 encodings provided by the ASP competition,
lpopt was able to syntactically rewrite 41 which indicates that, as mentioned above,
even extensively hand-tuned programs can be further decomposed in an automated manner. The remaining eight encodings contained rules that were so small that no further
decomposition was possible (i.e. their Gaifman graph was a clique of usually 3-4 nodes)
and thus the output of lpopt was the original, unmodified encoding in these cases. In
27 of the 41 encodings rewritten by lpopt, the decompositions were trivial and had
no significant impact on the solving performance. This is due to the fact that only rules
that were already very small (and thus did not contribute much to the grounding size in
the first place) could be decomposed. In five cases out of the 41 rewritten encodings, we
noticed a decrease in solving performance (see the paragraph on limitations of lpopt
below for an explanation) and in the remaining seven cases, the lpopt rewriting was
able to speed up the solving process with substantial improvements in three of these
seven. Two of those were the stable marriage problem encoding of 2013, and the permutation pattern matching encoding of 2014 which we will take a closer look at below.
Full benchmark results for the entire dataset can be found in [4].
2
3
https://www.mat.unical.it/aspcomp2014/
http://potassco.sourceforge.net
(a)
(b)
Fig. 1: Benchmark results for the stable marriage 2013 instances. The horizontal axis
represents the individual test instances, sorted by runtime without rule decomposition.
As can be seen in Figure 1, both grounding and solving time decrease dramatically. Notice that the grounding time is, in general, directly correlated with the size
of the respective grounding. With lpopt preprocessing, the grounding size decreases
dramatically by a factor of up to 65. The grounder is thirty times faster when using preprocessing, and the solver about three times. This is because of the following constraint
in the encoding that can be decomposed very well:
:- match(M,W1), manAssignsScore(M,W,Smw), W1!=W,
manAssignsScore(M,W1,Smw1), Smw>Smw1, match(M1,W),
womanAssignsScore(W,M,Swm), womanAssignsScore(W,M1,Swm1),
Swm>=Swm1.
The constraint rule above is quite intuitive to read: There cannot be a man M and
a woman W , such that they would both be better off if they were matched together,
instead of being matched as they are (that is, to W 1 and M 1, respectively). It encodes, precisely and straightforwardly, the condition of a stable marriage. The 2014
encoding splits this rule up, making the encoding much harder to understand. However,
with lpopt preprocessing, the grounding and solving performance matches that of the
hand-tuned 2014 encoding. This again illustrates that the lpopt algorithm allows for
efficient processing of rules written by non-experts that are not explicitly hand-tuned.
(a)
(b)
Fig. 2: Benchmark results for permutation pattern matching 2014. The horizontal axis
represents the individual test instances, sorted by runtime without rule decomposition.
A second example of lpopt’s capabilities is the permutation pattern matching
problem illustrated in Figure 2. The grounding time of the largest instance is 980 seconds without preprocessing and 17 seconds with preprocessing. This instance was also
impossible to solve within the timeout window of 300 seconds without lpopt preprocessing, but finishing within 88 seconds when lpopt was run first.
Other Use Cases. lpopt has also been employed in other works that illustrate its performance benefits. In particular, several solvers for other formalisms rely on a rewriting
to ASP in order to solve the original problem. Such rewritings can easily lead to the generation of large rules that current ASP solving systems are generally unable to handle.
For example, in [16] ASP rewritings for several problems from the abstract argumentation domain, proposed in [9], are implemented. In Section 4.6 of the thesis, the performance benefits of lpopt are clearly demonstated for these rewritings. Interestingly,
these rewritings also make heavy use of aggregates which goes to show that lpopt
also handles these constructs well. Another example is [5], where multiple rewritings
for ΣP2 and ΣP3 -hard problems are proposed and then benchmarked, again showcasing
that without lpopt these rewritings could not be solved by current ASP solvers in all
but the most simple cases.
Limitations. However, we also want to point out some limitations of the lpopt algorithm. When a domain predicate is used by the algorithm, the selection of atoms that
generate this domain predicate is at the moment essentially random, since the greedy
selection depends on the order of the atoms appearing in the rule. This approach, as
discussed in Section 3, may thus not pick an optimal set of atoms. However, it depends
on this selection how many ground rules this domain predicate rule will generate when
passed to the grounder. Therefore, it may at the moment be the case that the increased
grounding size caused by the domain predicate rules may destroy any benefit caused by
splitting up the main rule. This is precisely what caused the increase in solving time for
the five encodings out of 49 that lpopt was able to rewrite but where solving performance deteriorated. Clearly, this begs the question of what the best strategy is to select
atoms to generate domain predicates. This is part of ongoing work.
6
Conclusions
In this paper, we present an algorithm, based on a prototype from [18], that allows the
decomposition of large logic programming rules into smaller ones that current state-ofthe-art answer set programming solvers are better equipped to handle. Our implementation handles the entire ASP-Core-2 language [10]. Benchmark results show that in practice, even for extensively hand-tuned ASP programs, our rule decomposition algorithm
can improve solving performance significantly. Future work will include implementing this approach directly into state-of-the-art grounders like the gringo grounder used
in our benchmarks, as well as further refining the algorithm w.r.t. selection of domain
predicate atoms, as discussed at the end of Section 5.
Acknowledgments. Funded by the Austrian Science Fund (FWF): Y698, P25607.
References
1. Alviano, M., Dodaro, C., Faber, W., Leone, N., Ricca, F.: WASP: A native ASP solver based
on constraint learning. In: Proc. LPNMR. pp. 54–66 (2013), http://dx.doi.org/10.
1007/978-3-642-40564-8_6
2. Alviano, M., Faber, W., Leone, N., Perri, S., Pfeifer, G., Terracina, G.: The disjunctive datalog system DLV. In: Datalog Reloaded. Revised Selected Papers. pp. 282–301 (2010),
http://dx.doi.org/10.1007/978-3-642-24206-9_17
3. Arnborg, S., Corneil, D.G., Proskurowski, A.: Complexity of finding embeddings in a k-tree.
SIAM J. Algeb. Discr. Meth. 8(2), 277–284 (1987), http://dx.doi.org/10.1137/
0608024
4. Bichler, M.: Optimizing non-ground answer set programs via rule decomposition. BSc Thesis, TU Wien. http://dbai.tuwien.ac.at/proj/lpopt (2015)
5. Bichler, M., Morak, M., Woltran, S.: The power of non-ground rules in answer set programming. In: Proc. ICLP. To appear (2016), https://arxiv.org/abs/1608.01856
6. Bodlaender, H.L.: A linear-time algorithm for finding tree-decompositions of small
treewidth. SIAM J. Comput. 25(6), 1305–1317 (1996), http://dx.doi.org/10.
1137/S0097539793251219
7. Bodlaender, H.L., Koster, A.M.C.A.: Treewidth computations i. upper bounds. Inf. Comput.
208(3), 259–275 (2010), http://dx.doi.org/10.1016/j.ic.2009.03.008
8. Brewka, G., Eiter, T., Truszczynski, M.: Answer set programming at a glance. Commun. ACM 54(12), 92–103 (2011), http://doi.acm.org/10.1145/2043174.
2043195
9. Brewka, G., Woltran, S.: GRAPPA: A semantical framework for graph-based argument
processing. In: Proc. ECAI. pp. 153–158 (2014), http://dx.doi.org/10.3233/
978-1-61499-419-0-153
10. Calimeri, F., Faber, W., Gebser, M., Ianni, G., Kaminski, R., Krennwallner, T., Leone, N.,
Ricca, F., Schaub, T.: ASP-Core-2 Input Language Format v2.03c (2015), https://www.
mat.unical.it/aspcomp2013/ASPStandardization, accessed: 2016-06-27
11. Calimeri, F., Gebser, M., Maratea, M., Ricca, F.: Design and results of the fifth answer set
programming competition. Artif. Intell. 231, 151–181 (2016), http://dx.doi.org/
10.1016/j.artint.2015.09.008
12. Elkabani, I., Pontelli, E., Son, T.C.: SmodelsA - A system for computing answer sets of
logic programs with aggregates. In: Proc. LPNMR. pp. 427–431 (2005), http://dx.
doi.org/10.1007/11546207_40
13. Gebser, M., Kaminski, R., Kaufmann, B., Schaub, T.: Answer Set Solving in Practice. Synthesis Lectures on Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers (2012), http://dx.doi.org/10.2200/S00457ED1V01Y201211AIM019
14. Gebser, M., Kaufmann, B., Schaub, T.: Conflict-driven answer set solving: From theory to practice. Artif. Intell. 187, 52–89 (2012), http://dx.doi.org/10.1016/j.
artint.2012.04.001
15. Gelfond, M., Lifschitz, V.: The stable model semantics for logic programming. In: Proc.
ICLP/SLP. pp. 1070–1080 (1988)
16. Heißenberger, G.: A System for Advanced Graphical Argumentation Formalisms. Master’s
thesis, TU Wien (2016), www.dbai.tuwien.ac.at/proj/adf/grappavis/
17. Marek, V.W., Truszczyński, M.: Stable Models and an Alternative Logic Programming
Paradigm. In: The Logic Programming Paradigm – A 25-Year Perspective, pp. 375–398.
Springer (1999)
18. Morak, M., Woltran, S.: Preprocessing of complex non-ground rules in answer set programming. In: Proc. ICLP. pp. 247–258 (2012), http://dx.doi.org/10.4230/LIPIcs.
ICLP.2012.247
| 2cs.AI
|
Under consideration for publication in Theory and Practice of Logic Programming
arXiv:1507.06576v2 [cs.PL] 14 Feb 2018
Abstract Gringo
MARTIN GEBSER∗
Aalto University, HIIT, Finland
University of Potsdam, Germany
[email protected]
AMELIA HARRISON†
Univeristy of Texas at Austin, USA
[email protected]
ROLAND KAMINSKI∗
University of Potsdam, Germany
[email protected]
VLADIMIR LIFSCHITZ†
Univeristy of Texas at Austin, USA
[email protected]
TORSTEN SCHAUB∗‡
University of Potsdam, Germany
INRIA Rennes, France
[email protected]
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
This paper defines the syntax and semantics of the input language of the ASP grounder GRINGO.
The definition covers several constructs that were not discussed in earlier work on the semantics of
that language, including intervals, pools, division of integers, aggregates with non-numeric values,
and lparse-style aggregate expressions. The definition is abstract in the sense that it disregards some
details related to representing programs by strings of ASCII characters. It serves as a specification
for GRINGO from Version 4.5 on.
This is a corrected version of the paper published in Theory and Practice of Logic Programming,
Volume 15, Issue 04-05 (the special issue on the 31st International Conference on Logic Programming).
∗ Supported by AoF (grant 251170) and DFG (grants SCHA 550/8 and 550/9).
† Partially supported by the National Science Foundation under Grant IIS-1422455.
‡ Affiliated with Simon Fraser University, Canada, and IIIS Griffith University, Australia.
Gebser et al.
% place queens on the chess board
{ q(1..n,1..n) }.
% exactly 1 queen per row/column
:- X = 1..n, not #count{ Y : q(X,Y) } = 1.
:- Y = 1..n, not #count{ X : q(X,Y) } = 1.
% pre-calculate the diagonals
d1(X,Y,X-Y+n) :- X = 1..n, Y = 1..n.
d2(X,Y,X+Y-1) :- X = 1..n, Y = 1..n.
% at most one queen per diagonal
:- D = 1..n*2-1, 2 { q(X,Y) : d1(X,Y,D) }.
:- D = 1..n*2-1, 2 { q(X,Y) : d2(X,Y,D) }.
Table 1. An ASP solution to the n-queens problem.
1 Introduction
Version 4.0 of the ASP grounder GRINGO was released in March of 2013.1 Harrison et al.
(2014) defined the semantics of a subset of its input language in terms of stable models of
infinitary propositional formulas (Truszczynski 2012).
That subset does not include, however, several constructs that are frequently used in ASP
programs. One such construct is integer intervals. Take, for instance, the ASP solution
to the n-queens problem shown in Table 1. (It is similar to one of the solutions in the
language of Version 3 presented by Gebser et al. (2011).) Intervals are used in each rule of
this program. To include intervals, we have to modify the semantics from Harrison et al.
(2014) in two ways. First, we have to say that an arithmetic term denotes, generally, a
finite set of integers, not a single integer. (And it is not necessarily a set of consecutive
integers, because the language of GRINGO allows us to write (1..3)*2, for instance.
This expression denotes the set {2, 4, 6}.) Second, in the presence of intervals we cannot
treat a choice rule {A} as shorthand for the disjunctive rule A ; not A as proposed by
2
Ferraris and Lifschitz (2005). Indeed, the first rule of the program in Table 1 has 2n stable
models; the rule
q(1..n,1..n) ; not q(1..n,1..n)
has only 2 stable models.
Another feature of GRINGO not covered by Harrison et al. (2014), which is somewhat
similar to integer intervals, is pooling. Pooling is used, for instance, in the head of the rule
p(X;Y) :- q(X,Y).
(Note that a semicolon, not a comma, separates X from Y in the head.) This rule has the
1
http://potassco.sourceforge.net/
Abstract Gringo
same meaning as the pair of rules
p(X) :- q(X,Y).
p(Y) :- q(X,Y).
Pooling is often used to abbreviate a set of facts. For instance, instead of
p(a,5). p(b,10). p(c,12).
we can write p(a,5;b,10;c,12). In this paper, we talk about “pools”—groups of
terms such as a,5;b,10;c,12.
Yet another limitation of the proposal from Harrison et al. (2014) is related to the difference between “dlv-style” aggregates, such as
not #count{ Y : q(X,Y) } = 1
(1)
in the second rule of the program shown in Table 1, and “lparse-style” aggregates, such as
2 { q(X,Y) : d2(X,Y,D) }
(2)
in the last rule of the program. Both expressions have to do with counting. Syntactically,
the difference is that in expression (1) both the name of the aggregate (#count) and the
binary relation applied to the result of counting and a constant (=) are shown explicitly;
in (2), the fact that the constant 2 occurs on the left, in the lower bound position, tells us
that the relation ≤ is applied to that number and to the result of counting. More importantly,
there is a difference between the kinds of objects that we count. In case of expression (1)
we count, for a given value of X, the values of the variable Y such that q(X,Y) belongs to
the stable model. In case of (2) we count, for a given value of D, the atoms q(X,Y) that
belong to the stable model and satisfy an additional condition: d2(X,Y,D) belongs to the
model as well. Thus the atom in front of the colon in (2) plays two roles: it tells us what to
count, and it gives a condition on the stable model that needs to be checked.
The language studied by Harrison et al. (2014) does not include lparse-style aggregates.
The easiest way to add such aggregates is to treat them as abbreviations. For instance, (2)
can be viewed as shorthand for the dlv-style expression
2 <= #count{ q(X,Y) : q(X,Y), d2(X,Y,D) }.
(In this paper we adopt a more elaborate translation that allows us to accommodate negated
atoms in front of the colon.) In this expression, the first occurrence of q(X,Y) is syntactically a term, and the second is an atom. Thus treating (2) as an abbreviation depends on the
possibility of using the same symbol as a function and as a predicate. This is customary in
Prolog, but not in first-order logic, and this was not allowed by Harrison et al. (2014).
Our goal is to define the syntax and semantics of the language AG (short for Abstract Gringo )—a large subset of the input language of GRINGO that includes the features mentioned previously and a few other constructs not described by Harrison et al.
Gebser et al.
(2014). This is similar to the work that has led to the definition of the ASP Core language
(Calimeri et al. 2012).2,3
The semantics of AG defined in this paper serves a specification for GRINGO from Version 4.5 on. It can be used to prove the correctness of programs written in its input language.
As an example, in the electronic appendix we prove the correctness of the program shown
in Table 1.
AG is abstract in the sense that its definition disregards some details related to representing programs by strings of ASCII characters. For example, semicolons are used in the
input language of GRINGO in at least three ways: to denote disjunction in the head of a
rule, conjunction in the body, and pooling within an atom. In AG, three different symbols
play these different roles. The richer alphabet of AG makes it easier to define the semantics
of the language and to reason about ASP programs.
2 Syntax of AG
2.1 Symbols and Terms
We assume that five sets of symbols are selected: numerals, symbolic constants, negated
constants, variables, and aggregate names. We assume that a 1–1 correspondence between
the set of symbolic constants and the set of negated constants is chosen. For every symbolic constant p, the corresponding negated constant will be called its strong negation and
denoted by p̃. Further, we assume that these sets do not contain the symbols
−
+
×
inf
6=
=
⊥
,
;
:
(
..
(3)
sup
<
not
/
(4)
≤
>
∧
)
≥
∨
{
(5)
←
}
(6)
h
i
(7)
and that they are pairwise disjoint. All these symbols together form the alphabet of AG,
and AG rules will be defined as strings over this alphabet.
When a symbol is represented in ASCII, its type is determined by its first two characters.
For instance, a numeral starts with a digit or - followed by a digit. A symbolic constant
starts with a lower-case letter. A negated constant starts with - followed by a lower-case
letter. A variable starts with an upper-case letter, and an aggregate name starts with #. (The
2
3
Syntactically, AG is essentially an extension of ASP Core. But it does not include extra-logical constructs, such
as weak constraints and queries. The semantics of aggregates in AG is based on the approach of Ferraris (2005)
and thus is not equivalent to the semantics of aggregates in ASP Core when aggregates are used recursively
in the presence of negation. Among the language constructs that are not in ASP Core, in AG we find pooling,
intervals, and conditional literals. (These constructs originally appeared in the input language of LPARSE, but
in AG they are more general; for instance, interval bounds may contain variables, and restrictions to “domain
predicates” have disappeared.) Unlike ASP Core, AG supports aggregates in rule heads; see Section 3.
The definition of ASP Core does not refer to infinitary objects, such as infinitary propositional formulas used
in this paper. But it appears that infinitary objects of some kind will be required to correct the oversight in
(Calimeri et al. 2012, Section 2.2)—the set inst({e1 ; . . . ; en }), included in the body of a rule in the process of
instantiation, can be infinite.
Abstract Gringo
strings #false, #inf, and #sup, which represent ⊥, inf, and sup, also start with #.)
The symbols h and i (which are used to indicate the boundaries of a tuple within a term)
correspond to the ASCII characters ( and ).4 Each of the symbols (3)–(7) except for ∧
and ∨ has a unique ASCII representation; the symbols ∧ and ∨ can be represented by
semicolons and in some cases also by commas.
We assume that a 1–1 correspondence between the set of numerals and the set Z of
integers is chosen. For every integer n, the corresponding numeral will be denoted by n.
Terms are defined recursively, as follows:
•
•
•
•
all numerals, symbolic constants, and variables are terms;
if f is a symbolic constant and t is a tuple5 of terms then f (t) is a term;
if t1 and t2 are terms and ⋆ is one of the symbols (3) then (t1 ⋆ t2 ) is a term;
if t is a tuple of terms then hti is a term.
In a term of the form f () the parentheses can be dropped, so that every symbolic constant
can be viewed as a term. In a term of the form (t1 ⋆ t2 ) we will drop the parentheses when
it should not lead to confusion. A term of the form (0 − t) can be abbreviated as −t.
A term, or a tuple of terms, is precomputed if it contains neither variables nor symbols (3). We assume a total order on precomputed terms such that inf is its least element,
sup is its greatest element, and, for any integers m and n, m ≤ n iff m ≤ n.
We assume that for each aggregate name α a function α
b is chosen that maps every
6
set of tuples of precomputed terms to a precomputed term. The AG counterparts of the
aggregates implemented in Version 4.5 of GRINGO are defined below using the following
terminology. If the first member of a tuple t of precomputed terms is a numeral n then we
say that the integer n is the weight of t; if t is empty or its first member is not an integer
then the weight of t is 0. For any set T of tuples of precomputed terms,
• c\
ount(T ) is the numeral corresponding to the cardinality of T if T is finite, and sup
otherwise;
• sum(T
d ) is the numeral corresponding to the sum of the weights of all tuples in T
if T contains finitely many tuples with non-zero weights, and 0 otherwise;
\ ) is the numeral corresponding to the sum of the weights of all tuples in T
• sum+(T
whose weights are positive if T contains finitely many such tuples, and sup otherwise;
d ) is sup if T is empty, the least element of the set consisting of the first
• min(T
elements of the tuples in T if T is a finite non-empty set, and inf if T is infinite;
• max
d (T ) is inf if T is empty, the greatest element of the set consisting of the first
elements of the tuples in T if T is a finite non-empty set, and sup if T is infinite.
4
5
6
When an AG term representing a tuple of length 1, such as hai, is represented in ASCII, a comma is appended
to the tuple: (a,) .
In this paper, when we refer to a tuple of syntactic objects, we mean that the tuple may be empty and that its
members are separated by commas.
This understanding of α
b is different from that given by Harrison et al. (2014, Section 3.3). There, α
b is understood as a function that maps tuples of precomputed terms to elements of Z ∪ {∞, −∞}.
Gebser et al.
2.2 Atoms, Literals, and Choice Expressions
A pool is an expression of the form t1 ; . . . ; tn where n ≥ 1 and each ti is a tuple of terms.7
In particular, every tuple of terms is a pool.
An atom is a string of one of the forms p(P ), p̃(P ) where p is a symbolic constant and
P is a pool. In an atom of the form p() or p̃() the parentheses can be dropped, so that all
symbolic constants and all negated constants can be viewed as atoms.
For any atom A, the strings
A
not A
not not A
(8)
are symbolic literals.8 An arithmetic literal is a string of the form t1 ≺ t2 where t1 , t2 are
terms and ≺ is one of the symbols (5).
A conditional literal is a string of the form H : L where H is a symbolic or arithmetic
literal or the symbol ⊥ and L is a tuple of symbolic or arithmetic literals. If L is empty
then we will drop the colon, so that every symbolic or arithmetic literal can be viewed as a
conditional literal.9
An aggregate atom is a string of one of the forms
α{t1 : L1 ; . . . ; tn : Ln } ≺ s
s ≺ α{t1 : L1 ; . . . ; tn : Ln }
s1 ≺1 α{t1 : L1 ; . . . ; tn : Ln } ≺2 s2
(9)
(10)
(11)
(n ≥ 0), where
• α is an aggregate name,
• each ti is a tuple of terms,
• each Li is a tuple of symbolic or arithmetic literals (if Li is empty and ti is nonempty
then the preceding colon may be dropped),
• each of ≺, ≺1 , ≺2 is one of the symbols (5),
• each of s, s1 , s2 is a term.
For any aggregate atom A, the strings (8) are aggregate literals.
A literal is a conditional literal or an aggregate literal.
A choice expression is a string of the form {A} where A is an atom.
2.3 Rules and Programs
A rule is a string of the form
H1 ∨ · · · ∨ Hk ← B 1 ∧ · · · ∧ B m
7
8
9
(12)
This form of pooling is less general than what is allowed in the input language of GRINGO. For instance, f (a; b)
is neither a term nor a pool.
Semantically, the status of “double negations” in AG is the same as in logic programs with nested expressions
(Lifschitz et al. 1999), where conjunction, disjunction, and negation can be nested arbitrarily. Dropping a double negation may change the meaning of a rule. For instance, the one-rule program p ← not not p has two
stable models ∅, {p} (see Section 4.1); the latter will disappear if we drop not not .
In the input language of GRINGO, dropping the colon when L is empty is required.
Abstract Gringo
or of the form
C ← B1 ∧ · · · ∧ Bm
(13)
(k, m ≥ 0), where each Hi is a symbolic or arithmetic literal,10 C is a choice expression,
and each Bj is a literal. The expression B1 ∧· · ·∧Bm is the body of the rule; H1 ∨· · ·∨Hk
is the head of (12); C is the head of (13). If the body of a rule is empty and the head is not
then the arrow can be dropped.
For instance, here are the first five rules of the program from Table 1 written in the syntax
of AG:
R1
R2
R3
R4
R5
:
:
:
:
:
{ q(1 .. n, 1 .. n) },
← X = 1 .. n ∧ not count {Y : q(X, Y )}
← Y = 1 .. n ∧ not count{X : q(X, Y )}
d1 (X, Y, X − Y + n) ← X = 1 .. n ∧ Y
d2 (X, Y, X + Y − 1) ← X = 1 .. n ∧ Y
= 1,
= 1,
= 1 .. n,
= 1 .. n.
The other two rules use abbreviations introduced in the next section.
A program is a finite set of rules.
3 Abbreviations
Let C be an expression of the form
s1 ≺1 α{t1 : L1 : L1 ; . . . ; tn : Ln : Ln } ≺2 s2
(14)
(n ≥ 0), where each Li is a symbolic literal of one of the forms
p(t)
not p(t)
not not p(t)
(15)
(p is a symbolic or negated constant and t is a tuple of terms) and α, ti , Li , ≺1 , ≺2 , s1 ,
and s2 are as in the definition of an aggregate atom. Then a string of the form
C ← B1 ∧ · · · ∧ Bm
(16)
(m ≥ 0), where each Bj is a literal, is shorthand for the set of rules consisting of the rule
← B1 ∧ · · · ∧ Bm ∧ not s1 ≺1 α{t1 : L1 , L1 ; . . . ; tn : Ln , Ln } ≺2 s2
(17)
and, for each Li in (14) such that Li is an atom, the rule
{Li } ← B1 ∧ · · · ∧ Bm ∧ Ci
(18)
where Ci is the conjunction of the members of Li .
In both (17) and (18), the conjunction sign shown after Bm should be dropped if m = 0;
in (18) it should also be dropped if Ci is empty. The parts s1 ≺1 and ≺2 s2 in (14) are
optional; if one of them is missing then it is dropped from (17) as well; if both are missing
then rule (17) is dropped from the set altogether. If m = 0 in (16) then the arrow can be
dropped.
The term representations of literals (15) are the tuples
0, p(t)
10
1, p(t)
2, p(t)
In the input language of GRINGO, Hi may be any conditional literal.
Gebser et al.
of terms. (Each of them is indeed a tuple of terms, because p(t) can be viewed as a term.)
Also viewed as an abbreviation is any expression of the form
s1 {L1 : L1 ; . . . ; Ln : Ln } s2
(19)
(n > 0), where s1 , s2 are terms, each Li is a symbolic literal of one of the forms (15)
that does not contain .. , and each Li is a tuple of symbolic or arithmetic literals.11 Such
an expression is understood differently depending on whether it occurs in the head or the
body of a rule. In the head of a rule, (19) is understood as shorthand for an expression of
the form (14):
s1 ≤ count{t1 : L1 : L1 ; . . . ; tn : Ln : Ln } ≤ s2
(20)
where ti is the term representation of Li . If either or both of the terms s1 , s2 are missing,
the abbreviation is understood in a similar way. (Note that choice expressions that do not
contain .. are expressions of the form (19) where both s1 and s2 are missing, n = 1, L1 is
of the form p(t), and L1 is empty. In this case, we do not view (19) as an abbreviation.)
In the body of a rule (19) is understood as shorthand for the aggregate atom
s1 ≤ count{t1 : L1 , L1 ; . . . ; tn : Ln , Ln } ≤ s2
where ti is the term representation of Li .12 If either of the terms s1 , s2 in (19) is missing,
the abbreviation is understood in a similar way.
These abbreviations can be used, for instance, to represent the last two rules of the
program from Table 1 in the syntax of AG:
← D = 1 .. n ∗ 2 − 1 ∧ 2 {q(X, Y ) : d1(X, Y, D)},
← D = 1 .. n ∗ 2 − 1 ∧ 2 {q(X, Y ) : d2(X, Y, D)}.
Written out in full, these expressions become
R6 :
R7 :
← D = 1 .. n ∗ 2 − 1 ∧ 2 ≤ count{0, q(X, Y ) : q(X, Y ), d1(X, Y, D)},
← D = 1 .. n ∗ 2 − 1 ∧ 2 ≤ count{0, q(X, Y ) : q(X, Y ), d2(X, Y, D)}.
4 Semantics of AG
We will define the semantics of AG using a syntactic transformation τ . The function τ
converts rules into infinitary formulas formed from atoms of the form p(t) or p̃(t), where p
is a symbolic constant, and t is a tuple of precomputed terms. Then the stable models of
a program will be defined in terms of stable model semantics of infinitary formulas in the
sense of Truszczynski (2012), which is reviewed below.
4.1 Review: Infinitary Propositional Formulas
Let σ be a propositional signature, that is, a set of propositional atoms. The sets F0 , F1 , . . .
are defined as follows:
• F0 = σ,
11
12
To be precise, if Li is empty then the colon after Li is dropped.
If Li is empty then the comma after Li in this expression should be dropped.
Abstract Gringo
• Fi+1 is obtained from Fi by adding expressions H∧ and H∨ for all subsets H of Fi ,
and expressions F → G for all F, G ∈ Fi .
S∞
The elements of i=0 Fi are called (infinitary) formulas over σ.
In an infinitary formula, the symbols ⊤ and ⊥ are understood as abbreviations for ∅∧
and ∅∨ respectively; ¬F stands for F → ⊥, and F ↔ G stands for (F → G) ∧ (G → F ).
Subsets of a signature σ will also be called its interpretations. The satisfaction relation
between an interpretation and a formula is defined recursively as follows:
•
•
•
•
For every atom p from σ, I |= p if p ∈ I.
I |= H∧ if for every formula F in H, I |= F .
I |= H∨ if there is a formula F in H such that I |= F .
I |= F → G if I 6|= F or I |= G.
We say that an interpretation satisfies a set H of formulas, or is a model of H, if it satisfies
every formula in H. Two sets of formulas are equivalent if they have the same models.
The reduct F I of a formula F w.r.t. an interpretation I is defined as follows:
•
•
•
•
For p ∈ σ, pI = ⊥ if I 6|= p; otherwise pI = p.
(H∧ )I = {GI | G ∈ H}∧ .
(H∨ )I = {GI | G ∈ H}∨ .
(G → H)I = ⊥ if I 6|= G → H; otherwise (G → H)I = GI → H I .
An interpretation I is a stable model of a set H of formulas if it is minimal w.r.t. set
inclusion among the interpretations satisfying the reducts of all formulas from H.
For instance, if I = ∅ then
(¬¬p → p)I = (¬¬p)I → pI = ⊥ → ⊥;
if I = {p} then
(¬¬p → p)I = (¬¬p)I → pI = ¬(¬p)I → p = ¬⊥ → p.
In both cases, I is a minimal model of the reduct. Consequently, both ∅ and {p} are stable
models of {¬¬p → p}.
4.2 Semantics of Terms and Pools
A term is ground if it does not contain variables. The definition of “ground” for pools,
symbolic literals, and arithmetic literals is the same. Semantically, every ground term t
represents a finite set of precomputed terms [t], which is defined recursively:
• if t is a numeral or a symbolic constant then [t] is {t};
• if t is f (t1 , . . . , tn ) then [t] is the set of terms f (r1 , . . . , rn ) for all r1 ∈ [t1 ], . . . ,
rn ∈ [tn ];
• if t is (t1 + t2 ) then [t] is the set of numerals n1 + n2 for all integers n1 , n2 such
that n1 ∈ [t1 ] and n2 ∈ [t2 ]; similarly when t is (t1 − t2 ) or (t1 × t2 );
• if t is (t1 /t2 ) then [t] is the set of numerals ⌊n1 /n2 ⌋ for all integers n1 , n2 such that
n1 ∈ [t1 ], n2 ∈ [t2 ], and n2 6= 0;
Gebser et al.
• if t is (t1 .. t2 ) then [t] is the set of numerals m for all integers m such that, for some
integers n1 , n2 ,
n1 ∈ [t1 ],
n2 ∈ [t2 ],
n1 ≤ m ≤ n2 ;
• if t is ht1 , . . . , tn i then [t] is the set of terms hr1 , . . . , rn i for all r1 ∈ [t1 ], . . . ,
rn ∈ [tn ].
This definition is extended to an arbitrary ground pool P ; [P ] is a finite set of precomputed
tuples:
• if P is a tuple t1 , . . . , tn of terms (n 6= 1) then [P ] is the set of tuples r1 , . . . , rn for
all r1 ∈ [t1 ], . . . , rn ∈ [tn ];
• if P is a pool t1 ; . . . ; tn (n > 1) then [P ] is [t1 ] ∪ · · · ∪ [tn ].
For instance, [1..n, 1..n] is the set {i, j : 1 ≤ i, j ≤ n}.
It is clear that if a ground term t contains neither symbolic constants nor the symbols h
and i then every element of [t] is a numeral. If a tuple t of ground terms is precomputed
then [t] is {t}. The set [t] can be empty. For example, [1..0] = [1/0] = [1 + a] = ∅.
About a tuple of terms that does not contain .. we say that it is interval-free. It is clear
that if a tuple t of ground terms is interval-free then the cardinality of the set [t] is at most 1.
4.3 Semantics of Arithmetic and Symbolic Literals
For any ground (symbolic or arithmetic) literal L we will define two translations, τ∧ L and
τ∨ L. The specific translation function applied to an occurrence of a symbolic or arithmetic
literal in a rule depends on the context, as we will see in the following sections.
We will first consider symbolic literals. For any ground atom A,
• if A is p(P ) then τ∧ A is the conjunction of atoms p(t) over all tuples t in [P ], and
τ∨ A is the disjunction of these atoms;
• if A is p̃(P ) then τ∧ A is the conjunction of atoms p̃(t) over all tuples t in [P ], and
τ∨ A is the disjunction of these atoms;
• τ∧ (not A) is ¬τ∨ A, and τ∨ (not A) is ¬τ∧ A;
• τ∧ (not not A) is ¬¬τ∧ A, and τ∨ (not not A) is ¬¬τ∨ A.
The definitions of τ∧ and τ∨ for arithmetic literals are as follows:
• τ∧ (t1 ≺ t2 ) is ⊤ if the relation ≺ holds between the terms r1 and r2 for all r1 ∈ [t1 ]
and r2 ∈ [t2 ], and ⊥ otherwise;
• τ∨ (t1 ≺ t2 ) is ⊤ if the relation ≺ holds between the terms r1 and r2 for some r1 , r2
such that r1 ∈ [t1 ] and r2 ∈ [t2 ], and ⊥ otherwise.
For instance, τ∨ p(2..4) is p(2) ∨ p(3) ∨ p(4), and τ∨ (2 = 2..4) is ⊤.
For any tuple L of ground literals, τ∨ L stands for the conjunction of the formulas τ∨ L
for all members L of L. The expressions τ∧ ⊥ and τ∨ ⊥ both stand for ⊥.
It is clear that if A has the form p(t) or p̃(t), where t is a tuple of precomputed terms,
then each of the formulas τ∧ A and τ∨ A is A.
Abstract Gringo
4.4 Semantics of Choice Expressions
The result of applying τ to a choice expression {p(P )} is the conjunction of the formulas p(t) ∨ ¬p(t) over all tuples t in [P ]. Similarly, the result of applying τ to a choice
expression {p̃(P )} is the conjunction of the formulas p̃(t) ∨ ¬p̃(t) over all tuples t in [P ].
For instance, the result of applying τ to rule R1 (see Section 2.3) is
^
(21)
q(i, j) ∨ ¬q(i, j) .
1≤i,j≤n
4.5 Global Variables
About a variable we say that it is global
• in a conditional literal H : L, if it occurs in H but does not occur in L;
• in an aggregate literal A, not A, or not not A, where A is of one of the forms (9)–
(11), if it occurs in s, s1 , or s2 ;
• in a rule (12), if it is global in at least one of the expressions Hi , Bj ;
• in a rule (13), if it occurs in C or is global in at least one of the expressions Bj .
An instance of a rule R is any rule that can be obtained from R by substituting precomputed terms for all global variables.13 A literal or a rule is closed if it has no global
variables. It is clear that any instance of a rule is closed.
For example, X is global in the rule R2 from Section 2.3, so that the instances of R2 are
rules of the form
← r = 1 .. n ∧ not count{Y : q(r, Y )} = 1
for all precomputed terms r. The variables X and Y are global in R4 ; instances of R4 are
d1 (r, s, r − s + n) ← r = 1 .. n ∧ s = 1 .. n
for all precomputed terms r and s.
4.6 Semantics of Conditional Literals
If t is a term, x is a tuple of distinct variables, and r is a tuple of terms of the same length
as x, then the term obtained from t by substituting r for x will be denoted by txr . Similar
notation will be used for the result of substituting r for x in expressions of other kinds,
such as literals and tuples of literals.
The result of applying τ to a closed conditional literal H : L is the conjunction of the
formulas
τ∨ (Lxr ) → τ∨ (Hrx )
where x is the list of variables occurring in H : L, over all tuples r of precomputed terms
of the same length as x.
13
This definition differs slightly from that given by Harrison et al. (2014, Section 3.3). There, substitutions that
yield symbolic constants in the scope of arithmetical operators do not form instances. In a similar way, we
treat variables in conditional literals and aggregate literals (Sections 4.6 and 4.7) differently than how they are
treated by Harrison et al. (2014).
Gebser et al.
For instance, the result of applying τ to the arithmetic literal r = 1..n, where r is
a precomputed term, is τ∨ (ǫ) → τ∨ (r = 1..n), where ǫ is the tuple of length 0. The
antecedent of this implication is ⊤. The consequent is ⊤ if r is one of the numerals 1, . . . , n
and ⊥ otherwise.
4.7 Semantics of Aggregate Literals
In this section, the semantics of ground aggregates proposed by Ferraris (2005, Section 4.1)
is adapted to closed aggregate literals. Let E be a closed aggregate atom of one of the
forms (9)–(11), and let xi be the list of variables occurring in ti : Li (1 ≤ i ≤ n). By Ai
we denote the set of tuples r of precomputed terms of the same length as xi . By A we
denote the set {(i, r) : i ∈ {1, . . . , n}, r ∈ Ai }.
Let ∆ be a subset of A. Then by [∆] we denote the union of the sets [(ti )xr i ] for all pairs
(i, r) ∈ ∆. We say that ∆ justifies E with respect to a precomputed term14 t if
• E is of the form (9) and the relation ≺ holds between α
b[∆] and t, or
• E is of the form (10) and the relation ≺ holds between t and α
b[∆].
We say that ∆ justifies E with respect to a pair t1 , t2 of precomputed terms if E is of the
form (11), the relation ≺1 holds between t1 and α
b[∆], and the relation ≺2 holds between
α
b[∆] and t2 .
If t is a precomputed term, and E is of form (9) or (10), we define τt E as the conjunction
of the implications
_
^
(22)
τ∨ ((Li )xr i )
τ∨ ((Li )xr i ) →
(i,r)∈A\∆
(i,r)∈∆
over all sets ∆ that do not justify E with respect to t. If t1 , t2 is a pair of precomputed
terms, and E is of form (11), we define τt1 ,t2 E as the conjunction (22) over all sets ∆ that
do not justify E with respect to t1 , t2 .
For instance, if E is count{p(X) : p(X)} > 0 then τø0 E is the (conjunction containing
the single) implication expressing that p(r) holds for at least one precomputed term r:
_
⊤ →
p(r).
r
For a closed aggregate atom E of form (9) or (10),
• by τ E we denote the disjunction of formulas τt E over all terms t in [s];
• by τ (not E) we denote the disjunction of formulas ¬τt E over all terms t in [s]; and
• by τ (not not E) we denote the disjunction of formulas ¬¬τt E over all terms t
in [s].
It is clear that if [s] is a singleton set {t}, then τ E is (the disjunction containing only) τt E.
For a closed aggregate atom E of form (11),
14
This definition of the semantics of aggregates is more complicated than that published in the original version
of this document. There, a set ∆ either justifies an aggregate atom or not, without reference to a particular precomputed term t. The version here corrects a discrepancy between the semantics and the behavior of GRINGO
in the case when s represents a non-singleton set.
Abstract Gringo
• by τ E we denote the disjunction of formulas τt1 ,t2 E over all pairs of precomputed
terms t1 , t2 such that t1 in [s1 ] and t2 in [s2 ];
• by τ (not E) we denote the disjunction of formulas ¬τt1 ,t2 E over all pairs of precomputed terms t1 , t2 such that t1 in [s1 ] and t2 in [s2 ]; and
• by τ (not not E) we denote the disjunction of formulas ¬¬τt1 ,t2 E over all pairs of
precomputed terms t1 , t2 such that t1 in [s1 ] and t2 in [s2 ].
4.8 Semantics of Rules and Programs
For any rule R of form (12), τ R stands for the set of the formulas
τ B1 ∧ · · · ∧ τ Bm → τ∧ H1 ∨ · · · ∨ τ∧ Hk
for all instances (12) of R. For a rule of form (13), τ R stands for the set of the formulas
τ B1 ∧ · · · ∧ τ Bm → τ C
for all instances (13) of R. For any program Π, τ Π stands for the union of the sets τ R for
all rules R of Π.
A stable model of a program Π is any stable model of τ Π (in the sense of Section 4.1)
that does not contain any pair of atoms of the form p(t), p̃(t).
5 Simplifying τ Π
When we investigate the stable models of an AG program, it is often useful to simplify the
formulas obtained by applying transformation τ to its rules. By simplifying an infinitary
propositional formula we mean turning it into a strongly equivalent formula that has simpler syntactic structure. The definition of strong equivalence, introduced by Lifschitz et al.
(2001), is extended to infinitary formulas by Harrison et al. (2015). Corollary 1 from that
paper shows that the stable models of an infinitary formula are not affected by simplifying
its parts.
Proofs of the theorems stated in this section are outlined in the electronic appendix.
5.1 Monotone and Anti-Monotone Aggregate Atoms
When a rule contains aggregate atoms, we can sometimes simplify the implications (22) in
the corresponding infinitary formula using the theorems on monotone and anti-monotone
aggregates from Harrison et al. (2014, Section 6.1). The monotonicity or non-monotonicity
of an aggregate atom (9) can sometimes be established simply by looking at its aggregate
name α and its relation symbol ≺. If α is one of the symbols count, sum+, max, then (9)
is monotone when ≺ is < or ≤, and anti-monotone when ≺ is > or ≥. It is the other way
around if α is min.
Our semantics of aggregates is somewhat different from that adopted by Harrison et al.
(2014, Section 3.5), as explained in Footnote 6 (and also in view of the difference in the
treatment of variables discussed in Footnote 13, and the modification to the definition of
“justifies” explained in Footnote 14). Nevertheless, the statements and proofs of the two
Gebser et al.
theorems mentioned above remain essentially the same in the framework of AG. The theorems show that the antecedent in (22) can be dropped if E is monotone, and that the
consequent can be replaced by ⊥ if E is anti-monotone. These simplifications produce
strongly equivalent formulas.
5.2 Eliminating Equality from Aggregate Atoms
If ≺ in an aggregate atom (9) is = then the following theorem15 can be useful, in combination with the facts reviewed in Section 5.1:
Theorem 1
If E is a closed aggregate atom of the form
α{t1 : L1 ; . . . ; tn : Ln } = s,
E≤ is
α{t1 : L1 ; . . . ; tn : Ln } ≤ s,
and E≥ is
α{t1 : L1 ; . . . ; tn : Ln } ≥ s,
then for any precomputed term t, τt E is strongly equivalent to τt E≤ ∧ τt E≥ .
5.3 Properties of Counting
For any set S, by |S| we denote the cardinality of S if S is finite, and ∞ otherwise.
Theorem 2
For any closed aggregate atom E of the form
count{t1 : L1 ; . . . ; tn : Ln } ≥ m
where m is an integer and each ti is interval-free, τ E is strongly equivalent to
_
^
τ∨ ((Li )xr i ).
∆⊆A
|[∆]|=m
(23)
(i,r)∈∆
Theorem 3
For any closed aggregate atom E of the form
count{t1 : L1 ; . . . ; tn : Ln } ≤ m
where m is an integer and each ti is interval-free, τ E is strongly equivalent to
^
^
τ∨ ((Li )xr i ).
¬
∆⊆A
|[∆]|=m+1
15
(24)
(i,r)∈∆
The statement and proof of this theorem have been modified with respect to the original version of this paper
in accordance with the change in the definition of “justifies” described in Footnote 14.
Abstract Gringo
Without the assumption that each ti is interval-free the assertions of the theorems would
be incorrect. For instance, if E is count{1..2 : p} ≥ 1 then τ E is ⊤ → p, and (23) is ⊥.
In the special (but common) case when E has the form count{x : L} ≥ m, where
x is a tuple of variables and each variable occurring in L occurs also in x, the condition
|[∆]| = m in (23) can be replaced by |∆| = m. Indeed, in this case ∆ and [∆] have the
same cardinality because [∆] is the set of tuples r of terms such that (1, r) ∈ ∆. Similarly,
the condition |[∆]| = m + 1 in (24) can be replaced by |∆| = m + 1 if E has the form
count{x : L} ≤ m.
6 Conclusion
We proposed a definition of stable models for programs in the language AG and stated
a few theorems that facilitate reasoning about them. This definition can be viewed as a
specification for the answer set system CLINGO (see Footnote 1) and other systems with
the same input language. If such a system terminates given the ASCII representation of an
AG program Π as input, and produces neither error messages nor warnings, then its output
is expected to represent the stable models of Π.
Acknowledgements
We are grateful to the anonymous referees for useful comments.
Gebser et al.
References
C ALIMERI , F., FABER , W., G EBSER , M., I ANNI , G., K AMINSKI , R., K RENNWALLNER , T.,
L EONE , N., R ICCA , F., AND S CHAUB , T. 2012. ASP-Core-2: Input language format. Available
at https://www.mat.unical.it/aspcomp2013/files/ASP-CORE-2.0.pdf.
F ERRARIS , P. 2005. Answer sets for propositional theories. In Proceedings of International Conference on Logic Programming and Nonmonotonic Reasoning (LPNMR). 119–131.
F ERRARIS , P. AND L IFSCHITZ , V. 2005. Weight constraints as nested expressions. Theory and
Practice of Logic Programming 5, 1–2, 45–74.
G EBSER , M., K AMINSKI , R., K AUFMANN , B., AND S CHAUB , T. 2011. Challenges in answer
set solving. In Logic programming, knowledge representation, and nonmonotonic reasoning.
Springer, 74–90.
H ARRISON , A., L IFSCHITZ , V., P EARCE , D., AND VALVERDE , A. 2015. Infinitary equilibrium
logic and strong equivalence. In Proceedings of International Conference on Logic Programming
and Nonmonotonic Reasoning (LPNMR). http://www.cs.utexas.edu/users/vl/
papers/iel lpnmr.pdf; to appear.
H ARRISON , A., L IFSCHITZ , V., AND YANG , F. 2014. The semantics of Gringo and infinitary
propositional formulas. In Proceedings of International Conference on Principles of Knowledge
Representation and Reasoning (KR).
L IFSCHITZ , V., P EARCE , D., AND VALVERDE , A. 2001. Strongly equivalent logic programs. ACM
Transactions on Computational Logic 2, 526–541.
L IFSCHITZ , V., TANG , L. R., AND T URNER , H. 1999. Nested expressions in logic programs. Annals
of Mathematics and Artificial Intelligence 25, 369–389.
T RUSZCZYNSKI , M. 2012. Connecting first-order ASP and the logic FO(ID) through reducts. In
Correct Reasoning: Essays on Logic-Based AI in Honor of Vladimir Lifschitz, E. Erdem, J. Lee,
Y. Lierler, and D. Pearce, Eds. Springer, 543–559.
Under consideration for publication in Theory and Practice of Logic Programming
Online appendix for the paper
arXiv:1507.06576v2 [cs.PL] 14 Feb 2018
Abstract Gringo
published in Theory and Practice of Logic Programming
MARTIN GEBSER
Aalto University, HIIT, Finland
University of Potsdam, Germany
[email protected]
AMELIA HARRISON
Univeristy of Texas at Austin, USA
[email protected]
ROLAND KAMINSKI
University of Potsdam, Germany
[email protected]
VLADIMIR LIFSCHITZ
Univeristy of Texas at Austin, USA
[email protected]
TORSTEN SCHAUB
University of Potsdam, Germany
INRIA Rennes, France
[email protected]
Proofs of Theorems 1–3
Proof of Theorem 1
We will show that for any term t, the set of conjunctive terms of τt E is the union of the
sets of conjunctive terms of τt E≤ and τt E≥ . For any subset ∆ of A,
iff
iff
iff
iff
iff
iff
(22) is a conjunctive term of τt E
∆ does not justify E with respect to t
α
b[∆] 6= t
α
b[∆] < t or α
b[∆] > t
∆ does not justify E≥ with respect to t or ∆ does not justify E≤ with respect to t
(22) is a conjunctive term of τt E≥ or of τt E≤
(22) is a conjunctive term of τt E≤ ∧ τt E≥ .
Gebser, Harrison, Kaminski, Lifschitz, and Schaub
Proof of Theorem 2
Since [øm] is the singleton set {øm}, τ E is τøm E. Since E is monotone, the antecedent
of (22) can be dropped (Section 5.1), so that τøm E is strongly equivalent to
^
_
(25)
τ∨ ((Li )xr i ).
∆⊆A
|[∆]|<m
(i,r)∈A\∆
To derive (25) from (23) in HT∞ , assume (23). We will reason by cases, with one case
corresponding to each disjunctive term
^
(26)
τ∨ ((Li )xr i )
(i,r)∈∆
′
of (23). Let ∆ be a subset of A such that |[∆′ ]| < m. We will show that the conjunctive
term of (25) corresponding to ∆′ can be derived from (26). Since
|[∆′ ]| < m = |[∆]|,
(27)
there exists a pair (i, r) that is an element of ∆ but not an element of ∆′ . Indeed, if
∆ ⊆ ∆′ then [∆] ⊆ [∆′ ], which contradicts (27). Since (i, r) ∈ ∆, from (26) we can
derive τ∨ ((Li )xr i ). Since (i, r) ∈ A \ ∆′ , we can further derive
_
τ∨ ((Li )xr i ).
(i,r)∈A\∆′
It follows that each conjunctive term of (25) can be derived from (26).
We will prove by induction on m that (23) can be derived from (25) in HT∞ . Base case:
when m = 0 the disjunctive term of (23) corresponding to the empty ∆ is ⊤. Inductive
step: assume that (23) can be derived from (25), and assume
^
_
(28)
τ∨ ((Li )xr i ).
∆⊆A
|[∆]|<m+1
(i,r)∈A\∆
From (28) we can derive (25), and consequently (23). Now we reason by cases, with one
case corresponding to each disjunctive term of (23). Assume
^
(29)
τ∨ ((Li )xr i )
(i,r)∈Σ
where Σ is a subset of A such that |[Σ]| = m. Consider the set
Σ′ = {(i, r) : [(ti )xr i ] ⊆ [Σ]}.
By the definition of [Σ], for any (i, r) ∈ Σ, [(ti )xr i ] ⊆ [Σ]. So Σ ⊆ Σ′ . It follows that
[Σ] ⊆ [Σ′ ]. On the other hand,
[
[
[(ti )xr i ] =
[(ti )xr i ] ⊆ [Σ].
[Σ′ ] =
(i,r)∈Σ′
x
(i,r) : [(ti )r i ]⊆[Σ]
Consequently [Σ] = [Σ′ ], and |[Σ′ ]| = |[Σ]| = m. From (28),
_
τ∨ ((Li )xr i ).
(30)
(i,r)∈A\Σ′
Again, we reason by cases, with one case corresponding to each disjunctive term of (30).
Online appendix
x
x
Assume τ∨ ((Lj )s j ), where (j, s) ∈ A \ Σ′ . Combining assumption (29) and τ∨ ((Lj )s j ),
we derive
^
(31)
τ∨ ((Li )xr i ).
(i,r)∈Σ∪{(j,s)}
Consider the set [Σ ∪ {(j, s)}], that is,
x
[Σ] ∪ [(tj )s j ].
(32)
x
Recall that the cardinality of [Σ] is m. Since tj is interval-free, the cardinality of [(tj )s j ]
is at most 1. Furthermore, since (j, s) 6∈ Σ′ it follows that
x
[(tj )s j ] 6⊆ [Σ],
x
so that [(tj )s j ] is nonempty. Consequently, the set is a singleton, and therefore [Σ] is disjoint from it. It follows that the cardinality of (32) is m + 1. So from (31) we can derive
_
^
τ∨ ((Li )xr i ).
∆⊆A
|[∆]|=m+1
(i,r)∈∆
Proof of Theorem 3
Since [øm] is the singleton set {øm}, τ E is τøm E. Since the consequent of (22) can be
replaced in this case by ⊥, τøm E is strongly equivalent to
^
^
(33)
τ∨ ((Li )xr i ).
¬
∆⊆A
|[∆]|>m
(i,r)∈∆
Every conjunctive term of (24) is a conjunctive term of (33). To derive (33) from (24),
consider a set ∆ such that |[∆]| > m. Let f (i, r) stand for the set [(ti )xr i ]. Since each ti
is interval-free, this set is either empty or a singleton. Let s1 , . . . , sm+1 be m + 1 distinct
elements of [∆]. Choose elements (i1 , r1 ), . . . , (im+1 , rm+1 ) of ∆ such that each sk belongs to f (ik , rk ), and let ∆′ be {(i1 , r1 ), . . . , (im+1 , rm+1 )}. The cardinality of [∆′ ] is
at least m + 1, because this set includes s1 , . . . , sm+1 . On the other hand, it is at most
m + 1, because this set is the union of m + 1 sets of cardinality at most 1. Consequently,
|[∆′ ]| = m + 1. From (24) we can conclude in HT∞ that
^
(34)
τ∨ ((Li )xr i ).
¬
(i,r)∈∆′
Then the conjunctive term
¬
^
τ∨ ((Li )xr i )
(i,r)∈∆
of (33) follows, because ∆′ ⊆ ∆.
Correctness of the n-Queens Program
In this section, we prove the correctness of the program K, consisting of rules R1 , . . . , R7
(Sections 2.3 and 3).
The n-queens problem involves placing n queens on an n × n chess board such that no
two queens threaten each other. We will represent squares by pairs of integers (i, j) where
Gebser, Harrison, Kaminski, Lifschitz, and Schaub
1 ≤ i, j ≤ n. Two squares (i1 , j1 ) and (i2 , j2 ) are said to be in the same row if i1 = i2 ; in
the same column if j1 = j2 ; and in the same diagonal if |i1 − i2 | = |j1 − j2 |. A set Q of n
squares is a solution to the n-queens problem if no two elements of Q are in the same row,
in the same column, or in the same diagonal.
For any stable model I of K, by QI we denote the set of pairs (i, j) such that q(øi, øj) ∈
I.
Theorem 4
For each stable model I of K, QI is a solution to the n-queens problem. Furthermore, for
each solution Q to the n-queens problem there is exactly one stable model I of K such
that QI = Q.
Review: Supported Models and Constraints
We start by reviewing two familiar facts that will be useful in proving Theorem 4.
An infinitary program is a conjunction of (possibly infinitely many) infinitary formulas
of the form G → A, where A is an atom. We say that an interpretation I is supported
by an infinitary program Π if each atom A from I is the consequent of a conjunctive
term G → A of Π such that I |= G. Lifschitz and Yang (2013) give a condition, “tightness
on an interpretation,” under which the stable models of an infinitary program are identical
to its supported models. Proposition 1 below gives a simpler condition of this kind that is
sufficient for our purposes.
We say that an atom A occurs nonnegated in a formula F if
• F is A, or
• F is of the form H∧ or H∨ and A occurs nonnegated in at least one element of H,
or
• F is of the form G → H, where H is different from ⊥, and A occurs nonnegated in
G or in H.
It is clear, for instance, that no atom occurs nonnegated in a formula of the form ¬F .
The positive dependency graph of an infinitary program Π is the directed graph containing a vertex for each atom occuring in Π, and an edge from A to B for every conjunctive
term G → A of Π and every atom B that occurs nonnegated in G. We say that an infinitary
program Π is extratight if the positive dependency graph of Π contains no infinite paths.
The following fact is immediate from (Lifschitz and Yang 2013, Lemma 2).
Proposition 1
For any model I of an extratight infinitary program Π, I is stable iff I is supported by Π.
A constraint is an infinitary formula of the form ¬F (which is shorthand for F → ⊥).
The following theorem is a straightforward generalization of Proposition 4 from (Ferraris and Lifschitz 2005).
Proposition 2
Let H1 be a set of infinitary formulas and H2 be a set of constraints. A set I of atoms is a
stable model of H1 ∪ H2 iff I is a stable model of H1 and satisfies all formulas in H2 .
Online appendix
Proof
Case 1: Every formula in H1 ∪ H2 is satisfied by I. For each formula ¬F in H2 , I does
not satisfy F . So the reduct of each formula in H2 w.r.t. I is ¬⊥. It follows that the set
of reducts of all formulas in H1 ∪ H2 is satisfied by the same interpretations as the set of
reducts of all formulas in H1 . Consequently, I is minimal among the sets satisfying the
reducts of all formulas from H1 ∪ H2 iff it is minimal among the sets satisfying the reducts
of all formulas from H1 . Case 2: Some formula F in H1 ∪ H2 is not satisfied by I. Then I
is not a stable model of H1 ∪ H2 . If F ∈ H1 then I is not a stable model of H1 . Otherwise,
it is not true that I satisfies all formulas in H2 .
Proof of Theorem 4
To simplify notation, we will identify each set Q of squares with the set of atoms q(øi, øj)
where (i, j) ∈ Q. By Dn we denote the set of atoms of the forms ıd1(øi, øj, øi − j + n)
and ıd2(øi, øj, øi + j − 1) for all i, j from {1, . . . , n}. Recall that the rules of the program
K are denoted by R1 , . . . , R7 .
Lemma 1
A set of atoms is a stable model of
τ R1 ∪ τ R4 ∪ τ R5
(35)
iff it is of the form Q ∪ Dn where Q is a set of squares.
Proof
We can turn (35) into a strongly equivalent infinitary program as follows. The result of
applying τ to R1 is (21). Each conjunctive term in this formula is strongly equivalent to
¬¬q(øi, øj) → q(øi, øj).
(36)
The set τ R4 is strongly equivalent to the set of formulas
⊤ → d1(øi, øj, øi − j + n)
(37)
(1 ≤ i, j ≤ n). (We take into account that τ (øi = ø1..øn) is equivalent to ⊤ if 1 ≤ i ≤ n
and to ⊥ otherwise, and similarly for j.) Similarly, τ R5 is strongly equivalent to the set of
formulas
⊤ → d2(øi, øj, øi + j − 1)
(38)
(1 ≤ i, j ≤ n). Consequently, (35) is strongly equivalent to the conjunction H of formulas
(36)–(38). It is easy to check that H is an extratight infinitary program, so that by Proposition 1 its stable models are identical to its supported models. A set I of atoms is a model
of H iff Dn ⊆ I. Furthermore, I is supported iff every element of I has the form q(øi, øj)
or is an element of Dn . Consequently, supported models of H are sets of the form Q ∪ Dn
where Q is a set of squares.
Lemma 2
A set I of atoms is a stable model of τ K iff it has the form Q ∪ Dn , where Q is a solution
to the n-queens problem.
Gebser, Harrison, Kaminski, Lifschitz, and Schaub
Proof
Let H1 be (35) and H2 be
τ R2 ∪ τ R3 ∪ τ R6 ∪ τ R7 .
All formulas in H2 are constraints. Consequently, by Proposition 2, I is a stable model
of τ K iff it is a stable model of H1 and satisfies all formulas in H2 . By Lemma 1, I is a
stable model of H1 iff it is of the form Q ∪ Dn , where Q is a set of squares. It remains to
show that a set I of the form Q ∪ Dn satisfies all formulas in H2 iff Q is a solution to the
n-queens problem. Specifically, we will show that for any set I of the form Q ∪ Dn
(i) I satisfies τ R2 iff for all i ∈ {1, . . . , n}, I contains exactly one atom of the form
q(øi, øj);
(ii) I satisfies τ R3 iff for all j ∈ {1, . . . , n}, I contains exactly one atom of the form
q(øi, øj);
(iii) I satisfies τ R6 ∪ τ R7 iff no two squares in I are in the same diagonal.
To prove (i), note first that τ R2 is equivalent to the set of formulas
¬¬ (τ (ıcount{Y : q(øi, Y )} = ø1))
(1 ≤ i ≤ n). Let E be the aggregate atom above. Since [ø1] is a singleton set, τ E is the
same as τø1 E. By Theorem 1, this set is strongly equivalent to the set of formulas
¬¬ (τø1 (ıcount{Y : q(øi, Y )} ≤ ø1) ∧ τø1 (ıcount{Y : q(øi, Y )} ≥ ø1)) .
(39)
Again note that the result of applying τ to first aggregate atom in (39) is the same as the
result of applying τø1 . Then by Theorem 3 and the comment at the end of Section 5.3, τ
applied to this aggregate atom is strongly equivalent to
^
^
q(øi, r).
¬
∆⊆A
|∆|=2
(1,r)∈∆
This formula can be written as
^
Σ⊆P
|Σ|=2
¬
^
q(øi, r),
r∈Σ
where P is the set of precomputed terms. It is easy to see that I satisfies this formula iff it
contains at most one atom of the form q(øi, r). On the other hand, by Theorem 2, the result
of applying τ to the second aggregate atom in (39) is strongly equivalent to
_
^
q(øi, r).
∆⊆A
|∆|=1
(1,r)∈∆
Similar reasoning shows that I satisfies this formula iff it contains at least one atom of the
form q(øi, r). Since I = Q ∪ Dn , r in this atom is one of ø1, . . . , øn.
Claim (ii) is proved in a similar way.
To prove (iii), note first that two squares (øi1 , øj1 ), (øi2 , øj2 ) are in the same diagonal
iff there exists a k ∈ {1, . . . , 2n − 1} such that
d1(øi1 , øj1 , øk), d1(øi2 , øj2 , øk) ∈ Dn
(40)
Online appendix
or
d2(øi1 , øj1 , øk), d2(øi2 , øj2 , øk) ∈ Dn .
(41)
We will show that a set I of the form Q ∪ Dn does not satisfy τ R6 iff there exists a k such
that (40) holds for two distinct elements q(øi1 , øj1 ), q(øi2 , øj2 ) ∈ Q, and that it does not
satisfy τ R7 iff there exists a k such that (41) holds for such two elements. The result of
applying τ to R6 is strongly equivalent to the set of formulas
¬τ (2 ≤ ıcount{ø0, q(X, Y ) : q(X, Y ), d1(X, Y, øk)})
(42)
(1 ≤ k ≤ 2n − 1). Formula (42) is identical to
¬τ (ıcount{X, Y : q(X, Y ), d1(X, Y, øk)} ≥ 2).
In view of Theorem 2, it follows that it is strongly equivalent to
_
^
¬
(q(r, s) ∧ d1(r, s, øk))
∆⊆A
|∆|=2
(1,(r,s))∈∆
(1 ≤ k ≤ 2n − 1). This formula can be written as
_
^
(q(r, s) ∧ d1(r, s, øk)).
¬
Σ⊆P ×P
|Σ|=2
(43)
(r,s)∈Σ
For any set Q of squares,
iff
iff
Q ∪ Dn does not satisfy (43)
there exist two distinct pairs (r1 , s1 ), (r2 , s2 ) from P × P such that
q(r1 , s1 ), q(r2 , s2 ) ∈ Q and d1(r1 , s1 , øk), d1(r2 , s2 , øk) ∈ Dn
there exist two distinct squares (øi1 , øj1 ), (øi2 , øj2 ) ∈ Q such that (40) holds.
The claim about (41) is proved in a similar way.
Theorem 4 is immediate from the lemma.
References
F ERRARIS , P. AND L IFSCHITZ , V. 2005. Mathematical foundations of answer set programming. In
We Will Show Them! Essays in Honour of Dov Gabbay. King’s College Publications, 615–664.
L IFSCHITZ , V. AND YANG , F. 2013. Lloyd-Topor completion and general stable models. Theory
and Practice of Logic Programming 13, 4–5.
| 6cs.PL
|
Streaming Periodicity with Mismatches∗
Funda Ergün†
Elena Grigorescu‡
Erfan Sadeqi Azer§
Samson Zhou▽
arXiv:1708.04381v1 [cs.DS] 15 Aug 2017
August 16, 2017
Abstract
We study the problem of finding all k-periods of a length-n string S, presented as a data
stream. S is said to have k-period p if its prefix of length n − p differs from its suffix of length
n − p in at most k locations.
We give a one-pass streaming algorithm that computes the k-periods of a string S using
poly(k, log n) bits of space, for k-periods of length at most n2 . We also present a two-pass
streaming algorithm that computes k-periods of S using poly(k, log n) bits of space, regardless
of period length. We complement these results with comparable lower bounds.
1
Introduction
In this paper we are interested in finding (possibly imperfect) periodic trends in sequences given
as streams. Informally, a sequence is said to be periodic if it consists of repetitions of a block of
characters; e.g., abcabcabc consists of repetitions of abc, of length 3, and thus has period 3. The
study of periodic patterns in sequences is valuable in fields such as string algorithms, time series
data mining, and computational biology. The question of finding the smallest period of a string is a
fundamental building block for many string algorithms, especially in pattern matching, such as the
classic Knuth-Morris-Pratt [KMP77b] algorithm. The general technique for many pattern matching
algorithms is to find the periods of prefixes of the pattern in a preprocessing stage, then use them
as a guide for ruling out locations where the pattern cannot occur, thus improving efficiency.
While finding exact periods is fundamental to pattern matching, in real life, it is unrealistic to
expect data to be perfectly periodic. In this paper, we assume that even when there is a fixed period,
data might subtly change over time. In particular, we might see mismatches, defined as locations in
the sequence where a block is not the same as the previous block. For instance, while abababababab
is perfectly periodic, abababadadad contains one mismatch where ab becomes (and stays) ad. This
model captures periodic events that undergo permanent modifications over time (e.g., statistics
that remain generally cyclic but experience infrequent permanent changes or errors). We consider
∗
A preliminary version of this paper is to appear in the Proceedings of the 21st International Workshop on
Randomization and Computation (RANDOM 2017)
†
School of Informatics and Computing, Indiana University, Bloomington, IN. Research supported by NSF CCF1619081. Email: [email protected].
‡
Department of Computer Science, Purdue University, West Lafayette, IN. Research supported by NSF CCF1649515. Email: [email protected].
§
School of Informatics and Computing, Indiana University, Bloomington, IN. Email: [email protected].
▽
Department of Computer Science, Purdue University, West Lafayette, IN. Research supported by NSF CCF1649515. Email: [email protected].
1
our problem in the streaming setting, where the input is received in a sequential manner, and is
processed using sublinear space.
Our problem
generalizes exact periodicity studied in [EJS10], where the authors give a one-pass,
O log2 n -space algorithm for finding the smallest exact period of stream S of length n, when the
period is at most n/2, as well as a linear space lower bound when the period is longer than n/2.
They use two standard and equivalent definitions of periodicity: S has period p if it is of the form
B ℓ B ′ where B is a block of length p that appears ℓ ≥ 1 times in a row, and B ′ is a prefix of B.
For instance, abcabcabcab has period 3 where B = abc, and B ′ = ab. Equivalently, the length
n − p prefix of S is identical to its length n − p suffix. These definitions imply that at most k
of the repeating blocks differ from the preceding ones. According to this definition, for instance,
abcabdabdae is 2-periodic with period 3, with the mismatches occurring at positions 6 and 11.
In order to allow mismatches in S while looking for periodicity in small space, we utilize the
fingerprint data structure introduced for pattern matching with mismatches by [PP09, CFP+ 16].
Ideally, one would hope to combine results from [EJS10] and [CFP+ 16] to readily obtain an algorithm for detecting k-periodicity. Unfortunately, reasonably direct combinations of these techniques
do not seem to work. This is due to the fact that, in the presence of mismatches, the essential
structural properties of periods break down. For instance, in the exact setting, if S has periods
p and q, it must also have period r, where r is any positive multiple of p or q. It must also have
period d = gcd(p, q). These are not necessarily true when there are mismatches; as an example
consider the following.
Example 1 S = aaaaba has only one mismatch where S[i] 6= S[i + 2] (over all non range-violating
values of i); likewise where S[i] 6= S[i + 3], thus S is 1-periodic with periods 2 and 3. S is not
1-periodic with period 1 = gcd (2, 3) as it has two mismatches where S[i] 6= S[i + 1].
In the exact setting the smallest period t determines the entire structure of S as all other periods
must be multiples of t. This property does not necessarily hold when we allow mismatches, thus
the smallest period does not carry as much information as in the exact case. Similarly, overlaps
of a pattern with itself in S exhibits a much less well-defined periodic structure in the presence of
mismatches. This makes it much harder to achieve the fundamental space reduction achievable in
exact periodicity computation, where this kind of structure is crucially exploited.
1.1
Our Results
Given the structural challenges introduced by the presence of mismatches, we first focus on understanding the unique structural properties of k-periods and the relationship between the period p,
and the number of mismatches k (See Theorem 9). This understanding gives us tools for “compressing” our data into sublinear space. We proceed to present the following on a given stream S
of length n:
(1) a two-pass streaming algorithm that computes all k-periods of S using O k4 log9 n space,
regardless of period length (see Section 4)
(2) a one-pass streaming
algorithm that computes all k-periods of length at most n/2 of S using
O k4 log9 n space (see Section 5)
(3) a lower bound that any one-pass streaming algorithm that computes all k-periods of S requires
Ω(n) space (see Section 7)
2
√
(4) a lower bound that for k = o( n) with k > 2, any one-pass streaming algorithm that computes
all k-periods of S with probability at least 1 − 1/n requires Ω(k log n) space, even under the
promise that the k-periods are of length at most n/2. (see Section 7)
Given the above results, it is trivial to modify the algorithms to return, rather than all k-periods,
the smallest, largest, or any particular k-period of S.
1.2
Related Work
Our work extends two natural directions in sublinear algorithms for strings: on one hand the study
of the repetitive structure of long strings, and on the other hand the notion of approximate matching
of patterns, in which the algorithm can detect a pattern even when some of it got corrupted.
In the first line of work, Ergün et al. [EJS10] initiate the study of streaming algorithms for
detecting the period of a string, using poly(log n) bits of space. Indyk et al. [IKM00] also studied
mining periodic patterns in streams, [EAE06] studied periodicity in time-series databases and online
data, and Crouch and McGregor [CM11] study periodicity via linear sketches. [EMS10] and [LN11]
studied the problem of distinguishing periodic strings from aperiodic ones in the property testing
model of sublinear-time computation. Furthermore, [AEL10] studied approximate periodicity in
RAM model under the Hamming and swap distance metrics.
The pattern matching literature is a vast area (see [AG97] for a survey) with many variants.
Following the pattern matching streaming algorithm of Porat and Porat [PP09], Clifford et al.
[CFP+ 16] recently show improved streaming algorithms for the k-mismatch problem, as well as
offline and online variants. We adapt the use of sketches from [CFP+ 16] though there are some
other works with different sketches for strings ([AGMP13], [CEPR09], [RS16] and [PL07]). [CJPS13]
also showed several lower bounds for online pattern matching problem.
This line of work is also related to the detection of other natural patterns in strings, such as
palindromes or near palindromes. Ergün et al. [BEMS14] initiate the study of this problem and give
sublinear-space algorithms, while [GMSU16] show lower bounds. In recent work, [GSZ17] extend
this problem to finding near-palindromes (i.e., palindromes with possibly a few corrupted entries).
Many ideas used in these sublinear algorithms stem from related work in the classical offline
model. The well-known KMP algorithm [KMP77a] initially used periodic structures to search for
patterns within a text. Galil et al.[GS83] later improved the space performance of this pattern
matching algorithm. Recently, [Gaw13] also used the properties of periodic strings for pattern
matching when the strings are compressed. These interesting properties have allowed several algorithms to satisfy some non-trivial requirements of respective models (see [GKP16], [CFP+ 15] for
example).
2
Preliminaries
We assume our input is a stream S[1, . . . , n] of length |S| = n over some alphabet Σ. The ith
character of S is denoted S[i], and the substring between locations i and j (inclusive) S[i, j]. Two
strings S, T ∈ Σn are said to have a mismatch at index i if S[i] 6= T [i], and their Hamming distance
is the number of such mismatches, denoted HAM (S, T ) = {i | S[i] 6= T [i]} . We denote the
concatenation of S and T by S ◦ T .
S is said to have period p if S[x] = S[x + p] for all 1 ≤ x ≤ n − p; more succinctly, if
S[1, n−p] = S[p+1, n]. In general, we say S has k-period p (i.e., S has period p with k mismatches)
3
if S[x] = S[x + p] for all but at most k valid indices x. Equivalently, S has k-period p if and only
if HAM (S[1, n − p], S[p + 1, n]) ≤ k.
Observation 2 If p is a k-period of S, then at most k of the sequence of substrings S[1, p], S[p +
1, 2p], S[2p + 1, 3p], . . . can differ from the previous substring in the sequence.
When obvious from the context, given k-period p, we denote as a mismatch a position i for which
S[i] 6= S[i + p].
Example 3 The string S = aaaaaabbccd has 3-period equal to 1, since S[i] = S[i + 1] for all valid
locations i except mismatches at i = 6, 8, 10. On the other hand, S = abcabcadcabc has 2-period
equal to 3 since S[i] = S[i + 3] for all valid i except mismatches i = 5, 8.
The following observation notes that the number of mismatches between two strings is an upper
bound on the number of mismatches between their prefixes of equal length.
Observation 4 If p is a k-period of S, then for any x ≤ n − p, the number of mismatches between
S[1, x] and S[p + 1, p + x] is at most k.
Given two integers x and y, we denote their greatest common divisor by gcd (x, y).
We repeatedly use data structures and subroutines that use Karp-Rabin fingerprints. For more
about the properties of Karp-Rabin fingerprints see [KR87], but for our purposes, the following
suffice:
+ 16]) Given two strings S and T of length n, there exists a data structure that
Theorem 5 ([CFP
uses O k log6 n bits of space, and outputs whether HAM (S, T ) > k or HAM (S, T ) ≤ k, along with
the set of locations of the mismatches in the latter case.
From here, we use the term fingerprint to refer to this data structure.
2.1
The k-Mismatch Algorithm
For our string-matching tasks, we utilize an algorithm from [CFP+ 16], whose parameters are given
in Theorem 6. For us, string matching is a tool rather than a goal; as a result, we require additional
properties from the algorithm that are not obvious at first glance. In Corollary 7 we consider these
properties. Throughout our algorithms and proofs, we frequently refer to this algorithm as the
k-Mismatch Algorithm.
Theorem 6 ([CFP+ 16]) Given a pattern P of length ℓ, a text T of length n and some mismatch
threshold k, there exists an algorithm that, with probability 1 − n12 , outputs all indices i such that
HAM (T [i, i + ℓ − 1], P ) ≤ k using O k2 log8 n bits of space.
Whereas the pattern in the k-Mismatch Algorithm is given in advance and can be preprocessed
before the text, in our case the pattern is a prefix of the text, and the algorithm must return
any matches of this pattern, starting possibly at location 2, well within the original occurrence
of the pattern itself. (Consider text ‘abcdabcdabcdabcd’ and the pattern ‘abcdabcd,’ the first six
characters of the text. The first match starts at location 4, but the algorithm does not finish
reading the full pattern until it has read location 6.) To eliminate a potential problem due to this
requirement, we make modifications so that the algorithm can search for all matches in S of a prefix
of S.
4
Corollary 7 Given a string S and an index x, there exists an algorithm which, with
probability
8
1
2
1− n2 , outputs all indices i where HAM (S[1, x], S[i + 1, i + x]) ≤ k using O k log n bits of space.
Proof : We claim that the algorithm of Theorem 6 can be arranged and modified to output all
such indices i. We need to input S[1, x] as the pattern and S[2, n] as the text for this algorithm.
Thus, it suffices to argue that the data structure for the pattern is built in an online fashion.
That is, after reading each symbol of the pattern, the data structure corresponding to the prefix
of the pattern that has already been read is updated and ready to use. Moreover, the process of
building the data structure for the text should not depend on the pattern. The only dependency
between these two processes can be that they need to use the same randomness. Therefore, the
algorithm only needs to decide the randomness before starting to process the input and share it
between processes.
The algorithm of Theorem 6 has a few components, explained in the proof of Theorem 1.2 in
[CFP+ 16]. Here, we go through these components and explain how they satisfy the conditions we
mentioned.
The main data structure for this algorithm is also used in Theorem 5. In this data structure,
each symbol is partitioned to various subpatterns determined by the index of the symbol along with
predetermined random primes. Each subpattern is then fed to a dictionary matching algorithm.
The dictionary entries are exactly the subpatterns of the original patterns and thus can be updated
online.
The algorithm also needs to consider run-length encoding for each of these subpatterns in case
they are highly periodic. It is clear that run-length encoding can be done independently for the
pattern and the text.
Finally the approximation algorithm (Theorem 1.3 of [CFP+ 16]) uses a similar data structure to
Theorem 5, but with different magnitudes for primes. Thus, the entire algorithm can be modified
to run in an online fashion.
✷
3
Our Approach
Our approach to find all the k-periods of S is to first determine a set T of candidate k-periods,
which is guaranteed to be a superset of all the true k-periods. We first describe the algorithm to
find the k-period in two passes. In the first pass, we let T be the set of indices π that satisfy
HAM (S[1, x], S[π + 1, π + x]) ≤ k,
for some appropriate value of x that we specify later. Note that by Observation 4, all k-periods must
satisfy the above inequality. We show that even though T may be linear in size, we can succinctly
represent T by adding a few additional indices into T . We then show how to use the compressed
version of T during the second pass to verify the candidates and output the true k-periods of S.
This strategy does not work if we are allowed only one pass; by the time we discover a candidate
k-period p, it may be too late for us to start collecting the extra data needed to verify p (in the
two-pass version this is not a problem, as the extra pass allows us to go back to the start of S
and any needed data). We approach this problem by utilizing a trick from [EJS10] of identifying
candidate periods p using non-uniform criteria depending on the value of p. Using this idea, once
a candidate period is found, it is not too late to verify that it is a true k-period, and the data can
still be compressed into sublinear size.
5
Perhaps the biggest hidden challenge in the above approach is due to the major structural differences between exactly periodic and k-periodic strings; k-periodic strings show much less structure
than exactly periodic strings. As a result, incremental adaptations of existing techniques on periodic strings do not yield corresponding schemes for k-periodic strings. In order to achieve small
space, one needs to explore the weaker structural properties of k-periodic streams. A large part of
the effort in this work is in formalizing said structure (see Section 6), culminating in Theorem 17
and its proof, as well as exploring its application to our algorithms.
To show lower bounds for randomized algorithms finding the smallest k-period, we use a strategy
similar to that in [EJS10], using a reduction from the Augmented Index Problem. To show lower
bounds for randomized algorithms finding the smallest k-period given the promise that the smallest
k-period is at most n2 , we use Yao’s Principle [Yao77].
4
Two-Pass Algorithm to Compute k-Periods
In this section, we provide a two-pass, O k4 log9 n -space algorithm to output all k-periods of S.
The general approach is to first identify a superset of the k-periods of S, based on the self-similarity
of S, detected via the k-Mismatch algorithm of [CFP+ 16] as a black box. Unfortunately, while this
tool allows us to match parts of S to each other, we get only incomplete information about possible
periods, and this information is not readily stored in small space due to insufficient structure.
We explore the structure of periods with mismatches in order to come up with a technique that
massages our data into a form that can be compressed in small space, and is easily uncompressed.
During the second pass, we go over S as well as the compressed data to verify the candidate periods.
We consider two classes of periods by their length, and run two separate algorithms in parallel.
The first algorithm identifies all k-periods p with p ≤ n2 , while the second algorithm identifies all
k-periods p with p > n2 .
4.1
Finding small k-periods
Our algorithm for finding periods of length at most n/2 proceeds in two passes. In the first pass,
we identify a set T of candidate k-periods, and formulate its compressed representation, T C . In
the second pass, we recover each index from T C and verify whether or not it is a k-period. We
need T and T C to satisfy four properties.
(1) All true k-periods (likely accompanied by some candidate k-periods that are false positives)
are in T .
(2) T C can be stored in sublinear space.
(3) T can be fully recovered from T C in small space.
(4) The verification process in the second pass weeds out those candidates that are not true
periods in sublinear space.
We now describe our approach and show how it satisfies the above properties.
6
4.2
Pass 1: Property 1.
We crucially observe that any k-period p must satisfy the requirement
HAM (S[1, x], S[p + 1, p + x]) ≤ k
for all x ≤ n − p, and specifically for x = n2 . This observation allows us to refer to indices as
periods, as the index p + 1 where the requirement is satisfied corresponds to (possible) k-period
p. For the remainder of this algorithm, we set x = n2 , and designate the indices p + 1 that satisfy
the requirement with x = n2 as candidate k-periods; collectively these indices serve as T . Since
satisfying this requirement is necessary but not sufficient for a candidate to be a real k-period,
Property 1 follows.
4.3
Pass 1: Property 2.
Observe that T could be linear in size, so we cannot store each index explicitly. We observe that if
our indices followed an arithmetic progression, they could be kept implicitly in very succinct format
(as is the case where there are no mismatches). Unfortunately, due to the presence of mismatches in
S, such a regular structure does not happen. However, we show that it is still possible to implicitly
add a small number of extra indices to our candidates and end up with an arithmetic series and
allow for succinct representation. Our algorithm produces several such series, and represents each
one in terms of its first index and the increment between consecutive terms, obtaining T C from T ,
with the details given below.
C
T into T , we partition [1, x] into the 2mk + 2 disjoint intervals Hj =
h In order to compress
(j+1)x
, where m = log n. The goal is, possibly through the addition of extra
+ 1, 2(mk+1)
candidates, to represent the candidates in each interval as a single arithmetic series. This series
will be represented by its first term, as well as the increment between its consecutive terms, πj .
As each new candidate arrives, we update πj (except for the first update, πj never increases,
and it may shrink by an integer factor). Throughout the process, we maintain the invariant, by
updating πj , that the arithmetic sequence represented in Hj contains all candidates in Hj output
by the k-Mismatch algorithm. Then it is clear that T C and {πj } take sublinear space, satisfying
Property 2.
jx
2(mk+1)
4.4
Pass 1: Property 3.
It remains to describe how to update πj . The first time we see two candidates in Hj , we set πj to
be the increment between the candidates (before, it is set to -1). Each subsequent time we see a
new candidate index in the interval Hj , we update πj to be the greatest common divisor of πj and
the increment between the candidate and the smallest index in T ∩ Hj , which is kept explicitly.
For instance, if our first candidate index is 10, and afterwards we receive 22, 26, 32 (assume the
interval ends at 35), our πj values over time are -1, 12, 4, 2. Ultimately, the candidates that we will
be checking in Pass 2 will be 10, 12, 14, 16, 18, . . . , 34. For another example, see Figure 1. We
now need to show that the above invariant is maintained throughout the algorithm. To do this, we
show that any k-period p ∈ Hj is an increment of some multiple of πj away from the smallest index
in T ∩ Hj . Then, if we insert implicitly into T all indices in Hj whose distance from the smallest
index in T ∩ Hj is a multiple of πj , we will guarantee that any k-period in Hj will be included in
T.
7
H1
H2
H3
H4
H5
...
S:
1 π1
π2
π3
π4 = −1
π5
Fig. 1: Observe that all dots in each interval are equally spaced after the first. These dots represent
T c : the black dots represent T , while the white dots are added to convert the irregularly spaced
black dots into regularly spaced dot sequences.
We now show that any k-period p is implicitly represented in, and can be recovered from T C
and the values {πj } at the end of the first pass.
Lemma 8 If p <
n
2
is a k-period and p ∈ Hj , then p can be recovered from T C and πj .
Proof :
Since p ∈ Hj is a k-period, then it satisfies HAM (S[1, n − p], S[p + 1, n]) ≤ k. More
specifically, i = p satisfies
h ni h
i
n
HAM S 1,
, S i + 1, + i ≤ k
2
2
and will be reported by the k-Mismatch Algorithm. If there is no other index in T C ∩ Hj , then p
will be inserted into T C in the first pass, so p can clearly be recovered from T C .
On the other hand, if there is another index q in T C ∩ Hj , then πj will be updated to be a
divisor of the pairwise distances. Hence, the increment p − q is a multiple of πj . Any change that
might later happen to πj will be due to a gcd operation, and thus, will reduce it by a factor by at
least 2. Thus, p − q will remain a multiple of the final value of πj , and p will be recovered at the
end of the first pass as a member of T .
✷
Thus Property 3 is satisfied. The first pass algorithm in full appears below.
(To determine any k-period p with p ≤ n2 ):
First pass:
(1) Initialize πj = −1 for each 0 ≤ j < 2k log n + 2.
(2) Initialize T C = ∅.
(3) For each index i such that (using the k-Mismatch algorithm)
h ni h
i
n
HAM S 1,
, S i + 1, + i ≤ k
2
2
h
(j+1)n
jn
+
1,
For the integer j for which i is in the interval Hj = 4(k log
n+1)
4(k log n+1) :
(a) If there exists no candidate t ∈ T C in the interval Hj , then add i to T C .
(b) Otherwise, let t be the smallest candidate in T C and either πj = −1 or πj > 0. If
πj = −1, then set πj = i − t. Otherwise, set πj = gcd (πj , i − t).
8
4.5
Pass 2: Property 4.
Our task in the second pass is to verify whether each candidate recovered from T C and {πj } is actually a k-period or not. Thus, we must simultaneously check whether HAM (S[1, n − p], S[p + 1, n]) ≤
k for each candidate p, without using linear space. Fortunately, Theorem 9 states that at most
32k2 log n + 1 unique fingerprints for substrings of length πj are sufficient to recover the fingerprints
of both S[1, n − p] and S[p + 1, n] for any p ∈ Hj .
Before detailing, we first state a structural property, whose proof we defer to Section 6. This
property states that the greatest common divisor of the pairwise difference of any candidate kperiods within Hj must be a (32k2 log n + 1)-period.
Theorem 9 For some 0 ≤ j < 2mk + 2, let
Ij = {i ∈ Hj | HAM (S[1, x], S[i + 1, i + x]) ≤ k} .
For any p1 < . . . < pm ∈ I, the greatest common divisor d of p2 − p1 , p3 − p1 . . . , pm − p1 satisfies
HAM (S[1, x], S[d + 1, d + x]) ≤ 32mk 2 + 1.
Observe that πj is exactly d. Moreover, each time the value of πj changes, it gets divided by an
integer factor at least equal to 2, ending up finally as a positive integer. Since πj ≤ n, this change
can occur at most log n times, and so m ≤ log n. We now show that we can verify all candidates
in sublinear space.
Lemma 10 Let pi be a candidate k-period for a string S, with p1 < p2 < . . . < pm all contained
within Hj . Given the fingerprints of S[1, n − p1 ] and S[p1 + 1, n], we can determine whether or not
S has k-period pi for any 1 ≤ i ≤ m by storing at most 32k2 log n + 1 additional fingerprints.
Proof : Consider a decomposition of S into substrings wi of length pi , so that S = w1 ◦w2 ◦w3 ◦. . ..
Note that each index i for which wi 6= wi+1 corresponds with at least one mismatch. It follows
from Observation 2 that there exist at most k indices i for which wi 6= wi+1 . Thus, recording
the fingerprints and locations of these indices i suffice to determine whether or not there are k
mismatches for candidate period pi .
By Theorem 9, the greatest common divisor of the difference between each term in I is a
(32k2 log n + 1)-period πj . Thus, S can be decomposed S = v ◦ v1 ◦ v2 ◦ v3 ◦ . . . so that v has length
p1 , and each substring vi has length πj . It follows from Observation 2 that there exist at most
32k2 log n + 1 indices i for which vi 6= vi+1 . Therefore, recording the fingerprints and locations of
these indices i allow us to recover the fingerprint of S[1, n − pi ] from the fingerprint of S[1, n − pi−1 ],
since pi − pi−1 is a multiple of πj . Similarly, we can recover the fingerprint of S[pi + 1, n] from the
fingerprint of S[pi−1 + 1, n]. Hence, we can confirm whether or not pi is a k-period.
✷
The second pass algorithm in full follows.
(To determine all the k-periods p with p ≤ n2 ):
Second pass:
(1) For each t such that t ∈ T C :
(a) Let j be the integer for which t is in the interval Hj =
h
jn
4(k log n+1)
+ 1, 4(k(j+1)n
log n+1)
(b) If πj > 0, then record up to 32k2 log n + 1 unique fingerprints of length πj and of
9
length t, starting from t.
(c) Otherwise, record up to 32k2 log n + 1 unique fingerprints of length t, starting from
t.
(d) Check if HAM (S[1, n − t], S[t + 1, n]) ≤ k and return t if this is true.
h
(j+1)n
jn
(2) For each t which is in interval Hj = 4(k log
+
1,
n+1)
4(k log n+1) for some integer j:
If there exists an index in T C ∩ Hj whose distance from t is a multiple of πj , then check
if HAM (S[1, n − t], S[t + 1, n]) ≤ k and return t if this is true.
This proves Property 4. Next, we show the correctness of the algorithm for small k-periods.
Lemma 11 For any k-period p ≤ n2 , the algorithm outputs p.
Proof : Since the intervals {Hj } cover 1, n2 , then p ∈ Hj for some j. It follows from Lemma 8
that after the first pass, p can be recovered from T and πj . Thus, the second pass tests whether or
not p is a k-period. By Lemma 10, the algorithm outputs p, as desired.
✷
4.6
Finding large k-periods
As in the previous discussion, we would like to pick candidate periods during our first pass. However,
if a k-period p satisfies p > n2 , then clearly it will no longer satisfy
h ni h
n i
, S p + 1, p +
≤ k,
HAM S 1,
2
2
as p + n2 > n, and S p + n2 is undefined. Instead, recall that HAM (S[1, x] = S[p + 1, p + x]) ≤ k
for all x ≤ n − p. Ideally, when choosing candidate periods p based on their satisfying this formula,
we would like to use as large an x as possible without exceeding n−p, but we cannot do this without
knowing the value of p. Instead, [EJS10] observes we can try exponentially decreasing values of x:
we run log n instances of the algorithm sequentially, with x = n2 , n4 , . . ., since one of these values
of x must be the largest one that does not lead to an illegal index of S. Therefore, the desired
instance produces p, while all other instances do not.
(To determine a k-period p if p > n2 ):
First pass:
(m)
(1) Initialize πj
= −1 for each 0 ≤ j < 2k log n + 2 and 0 ≤ m ≤ log n.
(2) Initialize TmC = ∅.
(3) For each index i, let r be the largest m such that n2 + n4 + . . . +
k-Mismatch algorithm, check whether
h ni h
n i
HAM S 1, r , S i + 1, i + r
≤ k.
2
2
10
n
2r
≤ i. Using the
If so, let R =
n
2
+
n
4
+ ... +
n
2r −1
and j be the integer for which i is in the interval
nj
n(j + 1)
= R + r+1
+ 1, R + r+1
2 (k log n + 1)
2 (k log n + 1)
(r)
Hj
(r)
(a) If there exists no candidate t ∈ TrC in the interval Hj , then add i to TrC .
(r)
(r)
(b) Otherwise, let t be the smallest candidate in TrC and either πj = −1 or πj
(r)
(r)
(r)
(r)
If πj = −1, then set πj = i − t. Otherwise, set πj = gcd πj , i − t .
> 0.
This partition of [1, n] into the disjoint intervals 1, n2 , n2 + 1, n2 + n4 , . . . guarantees that any
(r)
k-period p is contained in one of these intervals. Moreover, the intervals {Hj } partition
hn n
n n
ni
+ + . . . + r−1 , + . . . + r ,
2
4
2
2
2
(r)
and so p can be recovered from TrC and {πj }. We now present the algorithm for the second-pass
to find all k-periods p for which p > n2 .
Second pass:
(1) For each t and any r such that t ∈ TrC :
(a) Let R =
n
2
+
n
4
+ ... +
(r)
Hj
n
2r−1
= R+
and j be the integer for which t is in the interval
n(j + 1)
nj
+ 1, R + r+1
2r+1 (k log n + 1)
2 (k log n + 1)
(r)
(r)
(b) If πj > 0, then record up to 32k2 log n + 1 unique fingerprints of length πj
of length t, starting from t.
and
(c) Otherwise, record up to 32k2 log n + 1 unique fingerprints of length t, starting from
t.
(d) Check if HAM (S[1, n − t], S[t + 1, n]) ≤ k and return t if this is true.
h
(r)
n(j+1)
for some
+
1,
R
+
(2) For each t which is in interval Hj = R + 2r+1 (knj
log n+1)
2r+1 (k log n+1)
integer j:
(r)
(r)
(a) If there exists an index in TrC ∩ Hj whose distance from t is a multiple of πj ,
then check if HAM (S[1, n − t], S[t + 1, n]) ≤ k and return t if this is true.
Since correctness follows from the same arguments as the case where p ≤ n2 , it remains to analyze
the space complexity of our algorithm.
Theorem 12 There
exists a two-pass algorithm that outputs all the k-periods of a given string
9
4
using O k log n space.
Proof : In the first pass, for each Tm , we maintain a k-Mismatch algorithm which requires
O k2 log8 n bits of space, as in Corollary 7. Since 1 ≤ m ≤ log n, we require O k2 log9 n bits
11
of space in total. In the second pass, we keep up to
O k2 log n fingerprints for any set of indices
in Tm . Each fingerprint requires space O k log6 n and there may be O (k log n) indices
in Tm for
each 1 ≤ m ≤ log n, for a total of O k4 log7 n bits of space. Thus, O k4 log9 n bits of space
suffice for both passes.
✷
5
One-Pass Algorithm to Compute k-Periods
We now give a one-pass algorithm that outputs all the k-periods smaller than n2 . Similar to
two-pass algorithm, we have two processes running in parallel. The first process handles all the
k-periods p with p ≤ n4 , while the second process handles the k-periods p with p > n4 . Both
processes are designed again based on the crucial observation that all the k-periods p must satisfy
n
HAM (S[1, x], S[p + 1,p + x]) ≤ k for all x ≤ n − p. In the first process,
nwe
set x = 2 and find all
n
indices i such that S i + 1, i + 2 has at most k mismatches from S 1, 2 .
The second process cannot use the same approach, because the k-Mismatch Algorithm reports
that index i is a candidate after reading position n2 + i, at which point we have already passed n − i.
This means that the fingerprint of S[1, n − i] cannot be built. For example, see Figure 2.
S[1, n − i]
Recognizes i is candidate
S:
1
n−i
i+1
i+
n
2
n
S i + 1, i + n2
Fig. 2: When i is recognized as a candidate, the algorithm has already passed n − i and cannot
build S[1, n − i].
Thus, for a fixed p in the second process, if we set x to be the largest power of two which does
not exceed n − 2p, the k-mismatch algorithm could report p. However, we cannot do this without
knowing the value of p.
Building off the ideas in [EJS10], we run log n instances of the algorithm in parallel, with
x = 1, 2, 4, . . ., then one of these values of x must correspond to the instance of k-mismatch
algorithm that recognizes p and reports it for later verification.
5.1
Finding small k-periods
We consider all the k-periods p with p ≤ n4 for this subsection. Run the k-Mismatch algorithm to
find
n
o
h ni h
n
n i
T = i i ≤ , HAM S 1,
, S i + 1, i +
≤k .
4
2
2
Upon finding an index i ∈ T , the algorithm uses the fingerprint for S i + 1, i + n2 to continue building S[i+1, n]. Simultaneously, it builds S[1, n−i], and checks whether HAM (S[1, n − i], S[i + 1, n]) ≤
k. The algorithm identifies that i ∈ T upon reading character i + n2 − 1. Since i ≤ n4 , then
i + n2 − 1 < 3n
4 ≤ n − i. Thus, the algorithm can identify i in time to build S[1, n − i]. By
Theorem 9, these entries can be computed from a sequence of compressed fingerprints.
12
5.2
Finding large k-periods
Now, consider all the k-periods p with
1 ≤ m ≤ log n − 1, define
n
4
< p ≤
n
2.
Let Im =
n
2
− 2m + 1, n2 − 2m−1 and for
Tm = {i |i ∈ Im , HAM (S[1, 2m ], S[i + 1, i + 2m ]) ≤ k} .
Let πm be a k-period of S[1, 2m ]. We first consider the case where πm ≥
m
where πm < 24 .
2m
4
and then the case
Observation 13 [CFP+ 16] If p is a k-period for S[1, n/2], then each i such that
h ni h
n i k
, S i + 1, i +
≤
HAM S 1,
2
2
2
must be at least p symbols apart.
m
By Observation 13, if πm ≥ 24 , then |Tm | ≤ 4. Moreover, we can detect whether i ∈ Tm by
index n2 − 2m−1 + 2m . On the other hand, n − i ≥ n2 + 2m + 1, and so we can properly build
S[1, n − i].
m
Now, suppose πm < 24 . Since Tm may be linear in size, we use the same trick to obtain a
succinct representation, whose properties satisfy those in Section 4, while including a few additional
indices. Let S[2m + 1, 2m+1 ] = w1 w2 . . . wt w′ , where each wi has length πm and for 0 ≤ d ≤ 3k, let
xd be the largest index such that S[1, 2m ] ◦ w1 ◦ w2 ◦ · · · ◦ wx has d-period πm .
Let Tm = i1 , i2 , . . . , ir in increasing order. Let S ir + 2m + 1, n2 + 2m = v1 v2 . . . vs v ′ , where
each vi has length πm and let y be the largest index such that S[ir + 1, ir + 2m ] ◦ v1 ◦ v2 ◦ · · · ◦ vy
has 3k-period πm .
2. Moreover, by
If y = s, then at most k of the substrings vi can
be unique by Observation
2
′
storing the fingerprints and positions of O k log n substrings, as well as v , we can
n recover the
fingerprint of each S[n−ij+1, n−ij ] by Lemma
10.
Thus,
we
keep
the
fingerprint
of
S
2 + 1, n − ir ,
n
and can construct the fingerprint of each S 2 + 1, n − ij
On the other hand if y 6= s, then for each ij , let ∆ be the number of indices z such that
ij ≤ z ≤ ir and S[z] 6= S[z + πm ]. That is, ∆ = |{z|ij ≤ z ≤ ir , S[z] 6= S[z + πm ]}|. Since πm is
a k-period of S[1, 2m ], HAM (S[1, 2m ], S[ij + 1, ij + 2m ]) ≤ k, and each mismatch between S[1, 2m ]
and S[ij + 1, ij + 2m ] can cause up to two indices z such that S[z] 6= S[z + πm ], then it follows
that 0 ≤ ∆ ≤ 3k. Then if y + |r − j| =
6 x3k−∆ , then ij ∈
/ Tm , since x3k−∆ is the largest index with
(3k − ∆)-period πm , while y is the largest index with 3k-period πm .
Thus, for each 0 ≤ ∆ ≤ 2k, there is at most one
6 x2k+∆ . Again
n index j with
y + |r − j| =
the
fingerprint
of
S
by
storing
the
fingerprints and
+
1,
n
−
i
by Lemma 10, we can compute
j
2
2
positions of O k log n substrings.
Computing each xd requires determining πm and the fingerprint of S[2m − πm + 1, 2m ]. Since
m
πm ≤ 24 , the algorithm determines πm by position πm + 2m < 2m − πm + 1. Thus, the algorithm
knows πm in time to start creating the fingerprint of S[2m − πm + 1, 2m ].
To compute y, we compute the fingerprint of S[ir + 1, ir + πm ]. We then compute the fingerprint
of each non-overlapping substring of length πm starting from ir + πm , and compare the fingerprint
to the previous fingerprint. We only record the fingerprint of the most recent substring, but keep
a running count of the number of mismatches.
13
Theorem 14 There exists a one-pass
algorithm that outputs all the k-periods p of a given string
9
n
4
with p ≤ 2 , and uses O k log n bits of space.
Proof : The process for small k-periods uses O k2 log8 n bits of space determining T . Verifying
whether an index in T is actually a k-period requires the fingerprints of O k2 log n substrings,
each using O k log6 n bits of space (Theorem 5). This adds up to a total of O k3 log7 n bits of
space.
The process for large k-periods has log n parallel
instances of the k-Mismatch algorithm to
9
2
compute Tm for 1 ≤ m ≤ log n, using O k log n bits of space. To reconstruct the fingerprint
of S[1, n − i] for each i ∈ Tm the algorithm needs to store the fingerprints
of at most O k2 log n
6
5) and
unique substrings (Lemma 10). Each fingerprint uses O k log n bits of space (Theorem
9
4
there can be up to O (k
log n) indices in Tm . This adds up to a total of O k log n bits of space.
Thus, O k4 log9 n bits of space suffice for both processes.
✷
6
Structural Properties of k-Periodic Strings
In this section, we detail the necessary steps in proving Theorem 9.
Reminder of Theorem 9. For some 0 ≤ j < 2mk + 2, let
Ij = {i ∈ Hj | HAM (S[1, x], S[i + 1, i + x]) ≤ k} .
For any p1 < . . . < pm ∈ I, the greatest common divisor d of p2 − p1 , p3 − p1 . . . , pm − p1 satisfies
HAM (S[1, x], S[d + 1, d + x]) ≤ 32mk 2 + 1.
We first show Theorem 17, which assumes there are only two candidate k-periods and both are
small. We then relax these conditions and prove Theorem 28, which does not restrict the number
of candidate k-periods, but still assumes that their magnitudes are small. Theorem 9 considers all
candidate k-periods in some interval. We use the fact that the difference between these candidates
is small, thus meeting the conditions of Theorem 28, although with an increase in the number of
mismatches.
To show that the greatest common divisor d of any two reasonably small candidates p < q for
k-periods is also a (16k2 +1)-period (Theorem 17), we consider the cases where either all candidates
are less than (2k + 1)d (Lemma 18) or some candidate is at least (2k + 1)d (Lemma 19).
In the first case, where all candidate period are less than (2k + 1)d, we partition the string into
disjoint intervals of a certain length, followed by partitioning the intervals further into congruence
classes. We show in Lemma 16 that any partition which contains an index i such that S[i] 6= S[i+d]
must also contain an index j which is a mismatch from some symbol p or q distance away. Since
there are at most 2k indices j, we can then bound the number of such partitions, and then extract
an upper bound on the number of such indices i.
In the second case, where some candidate is at least (2k + 1)d, our argument relies on forming
a grid (such as in Figure 3) where adjacent points are indices which either differ by p or q. We
include 2k + 1 rows and columns in this grid. Since dq ≥ 2k + 1, then no index in S is represented by
multiple points in the grid. We call an edge between adjacent points “bad” if the two corresponding
indices form a mismatch.
14
Observation 15 S[i] 6= S[i + d] only if each path between i and i + d contains a bad edge.
Our grid contains at most 2k bad edges, since p and q are both k-periods, and each index is
represented at most once. We then show that for all but at most (16k 2 + 1) indices i, there exists
a path between indices i and i + d that avoids bad edges. Therefore, there are at most (16k2 + 1)
indices i such that S[i] 6= S[i + d], which shows that d is an (16k 2 + 1)-period.
Before proving Lemma 18, we first show that given integers i, p, q, we can repeatedly hop by
distance p or q, starting from i, ending at i + gcd (p, q), all the while staying in a “small” interval.
Lemma 16 Suppose p < q are two positive integers with gcd (p, q) = d. Let i be an integer such
that 1 ≤ i ≤ p + q − d. Then there exists a sequence of integers i = t0 , . . . , tm = i+ d where |ti − ti+1 |
is either p or q, and 1 ≤ ti < p + q. Furthermore, each integer is congruent to i (mod d). In other
words, any interval of length p + q which contains indices i, i + d such that S[i] 6= S[i + d] also
contains an index j such that either S[j] 6= S[j + p] or S[j] 6= S[j + q].
Proof : Since d is the greatest common divisor of p and q, then there exist integers a, b such that
ap + bq = d. Suppose a > 0. Then consider the sequence ti = ti−1 + p if 1 ≤ ti−1 ≤ q. Otherwise, if
ti−1 > q, let ti = ti−1 − q. Then clearly, each |ti − ti+1 | is either p or q, and 1 ≤ ti < p + q. That is,
each ti either increases the coefficient of p by one, or decreases the coefficient of q by one. Thus, at
the last time the coefficient of p is a, ti = ap + bq = d, since any other coefficient of q would cause
either ti > q or ti < 1. Hence, terminating the sequence at this step produces the desired output,
and a similar argument follows if b > 0 instead of a > 0. Since p ≡ q ≡ 0 (mod d), then all integers
in these sequence are congruent to i (mod d).
✷
We now prove that the greatest common divisor d of any two reasonably small candidates p, q for
k-periods is also a (16k2 + 1)-period.
n
o
x
, HAM (S[1, x], S[i + 1, i + x]) ≤ k . For
Theorem 17 For any 1 ≤ x ≤ n2 , let I = i i ≤ 4k+2
any two p, q ∈ I with p < q, their greatest common divisor, d = gcd (p, q) satisfies
HAM (S[1, x], S[d + 1, d + x]) ≤ (16k2 + 1).
We now proceed to the proof of Theorem 17 for the case q < (2k + 1)d.
Lemma 18 Theorem 17 holds when q < (2k + 1)d.
Proof : If x ≤ 16k2 , then clearly there are at most 16k2 indices i such that S[i] 6= S[i + d], and
so d is a (16k2 + 1)-period. Otherwise, suppose x > 16k2 + 1, and by way of contradiction, that
there are at least 16k2 + 1 indices i such that S[i] 6= S[i + d].
i
p+q h
3(p+q)
,
:
I
=
1,
,
p
+
q
+
1,
Consider the following two classes of intervals of length p+q
1
2
2
h
i
h
i2 h
i
3(p+q)
2(p + q) + 1, 5(p+q)
, . . . and I2 = p+q
+ 1, 2(p + q) , 5(p+q)
+ 1, 3(p + q) ,
2
2 + 1, p + q ,
2
2
. . .. If there are at least 16k2 + 1 indices i such that S[i] 6= S[i + d], then either I1 or I2 contains
at least 8k2 + 1 of these indices.
Suppose I1 has at least 8k2 + 1 indices i such that S[i] 6= S[i + d]. Now, consider the disjoint
intervals of length p + q: [1, p + q], [p + q + 1, 2(p + q)], [2(p + q) + 1, 3(p + q)], . . .. Furthermore,
for each of these intervals, consider the congruence classes modulo d. Since x > 16k2 + 1 and each
2q
of these congruence classes within an intervals have p+q
d < d ≤ 2(2k) = 4k indices, then S[1, x]
certainly contains at least 2k + 1 of these congruence classes.
15
If I1 has at least 8k2 + 1 indices i such that S[i] 6= S[i + d] and each congruence class within an
interval contains less than 4k indices, then there are at least 2k + 1 congruence classes containing
such an index i. Because each of these indices occur within I1 , it follows that both i and i + d
are contained within the interval (and therefore, the same congruence class). By Lemma 16, each
congruence class within an interval containing indices i and i + d S[i] 6= S[i + d] also contains an
index j such that either S[j] 6= S[j + p] or S[j] 6= S[j + q]. Since there are at least 2k + 1 congruence
classes within intervals, then there are at least 2k + 1 such indices j. This either contradicts that
there are at most k indices j such that S[j] 6= S[j + p] or that there are at most k indices j such
that S[j] 6= S[j + q].
The proof for the case where I2 has at least 8k2 + 1 indices i such that S[i] 6= S[i + d] is
symmetric.
✷
The following lemma considers the case where at least one of candidate periods p or q is at least
(2k + 1)d. Without loss of generality, assume q ≥ (2k + 1)d. We form a grid, such as in Figure 3,
where adjacent points in the grid correspond to indices which either differ by p or q. An edge
between adjacent points is “bad” if the two corresponding indices form a mismatch. Otherwise, we
call an edge an “good”.
From Observation 15, S[i] 6= S[i + d] only if each path between i and i + d contains a bad edge.
Thus, if S[i] 6= S[i + d], then the point in the grid corresponding to i must be contained in some
region whose boundary is formed by bad edges. We partition the indices into congruence classes
modulo d, count the number of mismatches in each class, and aggregate the results.
That is, in a particular congruence class, we assume p is a k1 -period, and q is a k2 -period, where
k1 , k2 ≤ k. Then the grid contains at most k1 + k2 bad edges, which bounds the perimeter of the
regions. From this, we deduce a generous bound of (16k1 k2 +1) on the number of points inside these
regions, which is equivalent to the number of indices i such that S[i] 6= S[i + d] in the congruence
class. We then aggregate over all congruence classes to show that d is a (16k2 + 1)-period.
Lemma 19 Let p ≤ q and k be positive integers with q ≥ (2k + 1)d and let d = gcd (p, q). Given
a string S and an integer 0 ≤ m < d, let there be k1 > 0 indices i ≡ m (mod d) such that
S[i] 6= S[i + p] and k2 > 0 indices i ≡ m (mod d), not necessarily disjoint, such that S[i] 6= S[i + q]
and k1 , k2 ≤ k. If d = gcd (p, q), then there exist at most k1 k2 indices i ≡ m (mod d) such that
S[i] 6= S[i + d].
Proof : Consider a pair of indices (i, i+d) with S[i] 6= S[i+d] in congruence class m (mod d). We
ultimately want to build a grid of “large” size around i, but this may result in illegal indices if i is
too small or too large. Therefore, we first consider the case where k(p + q) ≤ i ≤ x − k(p + q), where
we can place i in the center of the grid. We then describe a similar argument with modifications
for i < k(p + q) or i > x − k(p + q), when we must place i near the periphery of the grid.
Given index i with k(p + q) ≤ i ≤ x − k(p + q), we define a (2k + 1)-grid centered at i on a subset
of indices of S[1, x] as follows: the node at the center of the grid is i and for any node j, the nodes
j +p, j +q, j −p and j −q are the top, right, bottom and left neighbors of j, respectively. We include
(2k + 1) rows and columns in this grid, so that i is the intersection of the middle row and the middle
column. See Figure 3 for example of such a grid. Note that since k(p + q) ≤ i ≤ x − k(p + q), all
points in the grid correspond to indices of S.
Claim 20 The points in a (2k + 1)-grid centered at i correspond to distinct indeces in S.
16
i+p
i−q i
i+q
i−p
Fig. 3: The dashed lines are good edges and the solid lines are bad edges. Note that it is impossible
to go from an isolated (light) node to one outside the the enclosed region (i.e., to a dark node)
without traversing through a bad edge. The total number of enclosed edges can be at most k2 if
the number of bad edges is at most 4k.
Proof :
Suppose, by way of contradiction, there exists some index j which is represented by
multiple points in the grid. That is, j = i+a1 p+b1 q = i+a2 p+b2 q with a1 6= a2 . Since d = gcd (p, q),
there exist integers r, s with p = rd, q = sd, and gcd (r, s) = 1. Then (a1 − a2 )p = (b2 − b1 )q so
(a1 − a2 )r = (b2 − b1 )s. Because gcd (r, s) = 1, it follows that (a1 − a2 ) is divisible by s = dq ≥ 2k + 1.
Therefore, |a1 − a2 | ≥ 2k + 1, and so a1 and a2 are at least 2k + 1 columns apart. However, this
contradicts both points being in the grid, since the grid contains exactly 2k + 1 columns.
✷
Claim 21 In each (2k + 1)-grid there exist at least k + 1 rows and k + 1 columns in the grid that
do not contain any bad edge.
Proof : Since HAM (S[1, x], S[α + 1, α + x]) ≤ k, for α = p, q, there are at most k indices i for
which S[i] 6= S[i+p] or S[i] 6= S[i+q]. By Claim 20, each index is represented at most once. Hence,
there are at most k vertical bad edges and at most k horizontal bad edges in this grid. Because
the grid contains 2k + 1 rows and columns, then there exist at least k + 1 rows and columns in the
grid that do not contain any bad edge.
✷
We say that a row with no bad edges in a (2k + 1)-grid is a no-change row. We define a no-change
column similarly.
Claim 22 Suppose the following hold:
(1) There exists a path avoiding bad edges between i and a no-change row or column in a (2k + 1)grid containing i.
(2) There exists a path avoiding bad edges between i + d and a no-change row or column in a
(2k + 1)-grid containing i + d.
17
Then there exists a path between i and i + d avoiding bad edges.
Proof :
Consider the two (2k + 1)-grids centered at i and i + q. By Claim 21, there are at
least k + 1 no-change rows in each grid, but the two grids overlap in 2k + 1 rows. Thus, some
no-change row in the grid centered at i must also be a no-change row in the grid centered at i + q.
Similarly, some no-change column in the grid centered at i must also be a no-change column in
the grid centered at i + p. These common no-change rows and columns allow traversal between
grids, as we can freely traverse between any no-change rows and columns while avoiding bad edges.
Thus, if we can traverse from i to any no-change row in the first grid, we can ultimately reach any
no-change row in the final grid containing i + d while avoiding all bad edges. Finally, if we can
traverse between i + d and any no-change row in the final grid, then there exists a path between i
and i + d without any bad edges.
✷
By the contrapositive of Claim 22, it follows that if S[i] 6= S[i + d], then either the (2k + 1)-grid
centered at i or the (2k + 1)-grid centered at i + d, has no path that avoids bad edges from the
center of the grid to a no-change row or no-change column. Suppose without loss of generality that
all paths from i to a no-change row/column within the (2k + 1)-grid centered at i has some bad
edge. Define an enclosed region containing i within the (2k + 1)-grid centered at i to be the set of
points reachable from i on paths containing only good edges. See Figure 3 for an example.
Thus, to bound the number of indices i such that S[i] 6= S[i+ d], it suffices to bound the number
of points enclosed in such regions, which are themselves all contained within (2k + 1)-grids.
We next argue that the number of unique indices that can be enclosed with k1 vertical edges
and k2 horizontal edges is at most k12k2 , even on an extended grid with no boundaries and multiple
vertices which correspond to the same index.
Lemma 23 The number of mismatches (i, i + d), for k(p + q) ≤ i ≤ x − k(p + q), is at most
Proof :
k1 k2
2 .
The proof follows from the following observations.
Observation 24 Two sets of indices represented by the grid points in two enclosed regions are
either identical or completely disjoint.
Henceforth, we consider only one representative for each enclosed region.
Observation 25 If two enclosed regions consist of sets of grid points I1 and I2 , respectively, then
there is a way to enclose at least |I1 | + |I2 | many points in the grid using at most the same number
of edges as the number of edges bounding the two enclosed regions, but such that these edges form
only one enclosed region. Moreover, the new enclosed region can be made convex, and in particular
a rectangle.
Specifically, the rectangle contains at most k1 vertical bad edges and at most k2 horizontal bad
edges. Because the total area (defined as the number of grid points) of regions enclosed by at
most k1 vertical bad edges and at most k2 horizontal bad edges is at most k14k2 , then the number of
enclosed nodes cannot exceed k14k2 . Therefore, the number of (i, i+ d) mismatches is at most double
the number of enclosed nodes (if i is enclosed, both (i, i + d) and (i − d, i) may be mismatches), and
the proof is complete for the case k(p + q) ≤ i ≤ x − k(p + q). Refer to Figure 3 for an example. ✷
We now describe the cases where i < k(p + q) and i > x − k(p + q). The problem with the above
grid for these values of i is that many points in the grid either have value less than 0 or greater
18
than x. These points correspond to illegal indices, as S[j] for j < 0 or j > x is nonsensical. Hence,
we simply change the construction so that we still use 2k + 1 rows and columns in total, but that i
appears in the bottom left corner for i < k(p + q). On the other hand, if i > x − k(p + q), then we
construct our grid so that i appears in the top right corner. Once again, since q ≥ (2k + 1)d and
the grid contains 2k + 1 rows and columns, then each index appears at most once inside the grid.
In both cases, the boundary of the grid serves to help enclose an area containing i. Thus, any
node can be enclosed by a combination of the boundary of the grid and a number of bad edges.
However, the boundary of the grid can be at most half of the entire perimeter of an enclosed region.
The remaining half of the perimeter consists of at most k1 vertical bad edges and k2 horizontal bad
edges, and so the entire area is at most k1 k2 . Then the number of enclosed nodes is at most k1 k2 .
Again, the number of (i, i + d) mismatches is at most double the number of enclosed nodes:
Lemma 26 The number of mismatches (i, i + d), for i < k(p + q) or i > x − k(p + q), is at most
2k1 k2 .
See Figure 4 for example.
✷
i + 2p
i+p
i
i + q i + 2q
Fig. 4: The dashed lines are good edges and the solid lines are bad edges. Part of the boundary of
the enclosed points is induced by the boundary of the grid. The total area of the enclosed regions
is at most k2 if the perimeter of the bad edges is at most 2k.
We now complete the proof of Theorem 17 by aggregating each congruence class handled in
Lemma 19.
Recall that we have two cases: q < (2k + 1)d and q ≥ (2k + 1)d.
Proof of Theorem 17:
Lemma 18 handles the first case.
In the second case, we observe that the indices in each congruence class modulo d do not
interfere with each other. In other words, the indices i, i + d, i + p and i + q are all in the same
congruence class modulo d, as are all nodes and edges in the grids containing i and i + d. Thus,
points in an enclosed region for one congruence class modulo d cannot be in an enclosed region
(m)
for a different congruence class modulo d. Now, for 0 ≤ m < d, let k1 be the number of indices
(m)
i ≡ m (mod d) such that S[i] 6= S[i + p] and let k2 be the number of indices i ≡ m (mod d)
19
such that S[i] 6= S[i + q]. By Observation 25, the number of enclosed points for a congruence class
(m)
(m)
modulo d is at most the number of points inside a rectangle with length k1 and width k2 . Since
P (m)
P (m)
k1 ≤ k, the sum of the lengths of the rectangles is at most k. Similarly,
k2 ≤ k implies
that the sum of the lengths of the rectangles is at most k.
To aggregate all these points across all congruence classes modulo d, we finally observe that a
possibly larger enclosed number of points can be obtained if the edges from different congruence
classes are all uniquely mapped into just one congruence class, and hence we may assume without
loss of generality that all bad edges occur in the same congruence class. The following observation
essentially finishes the proof.
Observation 27 The total number of enclosed points across all congruence classes is at most the
number of points inside a square with length and width k, i.e. k2 .
It follows that the total number of indices i such that S[i] 6= S[i + d] is at most k2 , which finishes
✷
the proof of Theorem 17.
We generalize Theorem 17 by showing that the greatest common divisor of any m ≥ 2 reasonably
small candidates for k-periods is also a (2mk2 + 1)-period. We emphasize that it is sufficient for
m ≤ log n, since the greatest common divisor can change at most log n times.
o
n
x
, HAM (S[1, x], S[i + 1, i + x]) ≤ k . For any p1 , . . . , pm ∈ I,
Theorem 28 Let I = i i ≤ 2(mk+1)
their greatest common divisor, d = gcd (p1 , . . . , pm ) satisfies
HAM (S[1, x], S[d + 1, d + x]) ≤ 8mk 2 + 1.
Proof : Observe that it no longer holds that the pairwise greatest common divisor between two
candidates pi and pj is d. However, it suffices to consider δ = gcd (p1 , pm ). If pδm < 2k + 1, then the
proof reduces to that of Lemma 18. Notably, the 2k+1 intervals [1, p1 +pm ], [p1 +pm +1, 2(p1 +pm )],
. . . each consist of δ disjoint congruence classes. We must modify 4(2k)(2k) + 1 to 4(2k)(mk) + 1
to apply the Pigeonhole Principle with mk mismatched indices instead of 2k mismatched indices.
Otherwise, the proof is similar to that of Lemma 19, as follows.
Whereas for two candidate k-periods p1 , p2 we represented the indices of the stream as points
in a grid, here, for a number of m candidate k-periods we represent the indices of the stream as
points in an m-dimensional hypergrid. Here too, we reduce the problem to counting points inside
an enclosed region within an extended m-dimensional hypergrid (instead of grid).
As before, an enclosed region containing i is the set of points on the grid reachable from i on
paths containing only good edges. A point is a boundary point of an enclosed region if it is in the
enclosed region and is incident to a bad edge (if no bad edges are incident to a point in the region,
then the point must be in the interior of the region). We sometimes denote by boundary edges
the bad edges adjacent to boundary points. There may again be several disjoint regions enclosing
points, but like in Observation 25, the number of points enclosed by a fixed number of edges is
maximized within a continguous “convex” set:
Observation 29 If two enclosed regions consist of sets of hypergrid points I1 and I2 , respectively,
then there is a way to enclose at least |I1 | + |I2 | many points in the grid using at most the same
number of boundary edges as the number of edges bounding the two enclosed regions, but such that
these edges form only one enclosed region. Moreover, the new enclosed region can be made convex,
and in particular a hyperrectangle.
20
Recall that we will use the number of points enclosed within such regions as an upper bound for
the number of pairs of indices i, i + d that have different values in the stream S, since it is necessary
that such an enclosed region exists in order to cause the existence of the mismatched pair.
For the sake of building up intuition, first consider the case m = 3. Note that there are at most
2(3k) = 6k bad edges in total (if, as before, we may run into the boundary of the hypergrid). Thus,
the number of boundary points of the cube that forms the enclosed region is 6k. For an illustration,
see Figure 5. Since a cube with at most 6k boundary points has volume at most k3/2 , it follows
Fig. 5: An enclosed region and the bad edges incident to the surface.
that the number of enclosed points is at most k3/2 ≤ k2 , which is what we aimed for.
For general m, as before, we may assume without loss of generality that all the bad edges
are in the same congruence class modulo d. Since there are m candidate k-periods, the total
number of bad edges is at most mk. Similar to Figure 4, at most another mk points can be on the
boundary of the worst-case hyperrectangle implied in Observation 29, due to the boundary of the
hypergrid, corresponding to illegal indices of S. Thus, there are at most 2mk boundary points of
the hyperrectangle.
In particular, we may assume that all sides have the same length, and thus the hyperrectagle
is isomorphic to the hypergrid [ℓ]m , for some integer ℓ, such that the number of boundary points is
2mk. More specifically, a boundary point x must have some coordinate i such that either xi = 1
or xi = ℓ. Therefore, there are 2mℓm−1 = 2mk points on the boundary, blocking every path from
points in the interior of the enclosed region to points outside the region. Since such a hyperrectangle
encloses ℓm = km/(m−1) ≤ k2 many points, it follows that that the number of indices i such that
S[i] 6= S[i + d], is again at most k 2 , which completes the proof of the general case.
✷
Finally, we show that the distance between any m candidate k-periods that are reasonably close to
each other must be a (32mk2 + 1)-period. This relaxes the constraints of Theorem 28.
Reminder of Theorem 9. For some 0 ≤ j < 2mk + 2, let
Ij = {i ∈ Hj | HAM (S[1, x], S[i + 1, i + x]) ≤ k} .
For any p1 < . . . < pm ∈ I, the greatest common divisor d of p2 − p1 , p3 − p1 . . . , pm − p1 satisfies
HAM (S[1, x], S[d + 1, d + x]) ≤ 32mk 2 + 1.
Note that p2 − p1 , p3 − p1 , . . . , pm − p1 are in I, where
x
I = i i≤
, HAM (S[1, x], S[i + 1, i + x]) ≤ 2k .
2(mk + 1)
Proof of Theorem 9:
21
Then by Theorem 28, their greatest common divisor d satisfies
HAM (S[1, x], S[d + 1, d + x]) ≤ 8m(2k)2 + 1 = 32mk2 + 1.
✷
7
7.1
Lower Bounds
Lower Bounds for General Periods
Recall the following variant of the Augmented Indexing Problem, denoted INDn,δ , where Alice is
given a string S ∈ Σn . Bob is given an index i ∈ [n], as well as S[1, i − 1], and must output S[i]
correctly with probability at least 1 − δ.
Lemma 30 [MNSW95] The one-way communication complexity of INDn,δ is Ω((1 − δ)n log |Σ|).
Theorem 31 Any one-pass streaming algorithm which computes the smallest k-period of an input
string S requires Ω(n) space.
Proof :
Consider the following communication game between Alice and Bob, who are given
strings A and B respectively. Both A and B have length n, and the goal is to compute the smallest
k-period of a◦b. Then we show that any one-way protocol which successfully computes the smallest
k-period of a◦b requires Ω(n) communication by a reduction from the augmented indexing problem.
Suppose Alice gets a string S ∈ {0, 1}n , while Bob gets an index i ∈ [n − 1] and S[1, i − 1]. Let
u be the binary negation of S[1], i.e., u = 1 − S[1]. Then Alice sets A = (S[1])k (S[2])k . . . (S[n])k
and Bob sets B = uk(n−i) ◦ (S[1])k (S[2])k . . . (S[i − 1])k ◦ 1k so that both A and B have length kn.
Moreover, the smallest k-period of A ◦ B is k(2n − i) if and only if S[i] = 1.
✷
7.2
Lower Bounds for Small Periods
√
We now show that for k = o( n), even given the promise that the smallest k-period is at most n2 ,
any randomized algorithm which computes the smallest k-period with probability at least 1 − n1
requires Ω(k log n) space. By Yao’s Minimax Principle [Yao77], it suffices to show a distribution
n
bits of memory fails with
over inputs such that every deterministic algorithm using less than k log
6
1
probability at least n .
Define an infinite string 11 01 12 02 13 03 . . ., as in [GMSU16], and let ν be the prefix of length n4 .
Let X be the set of binary strings of length n4 at Hamming distance k2 from ν. Given x ∈ X, let
Yx be the set of binary strings of length n4 with either HAM (x, y) = k2 or HAM (x, y) = k2 + 1. We
pick (x, y) uniformly at random from (X, Yx ).
n
bits
Theorem 32 Given an input x ◦ y, any deterministic algorithm D that uses less than k log
6
k
k
of memory cannot correctly output whether HAM (x, y) = 2 or HAM (x, y) > 2 with probability at
√
least 1 − n1 , for k = o( n).
22
n k/2
n k/4
.
By
Stirling’s
approximation,
|X|
≥
Proof :
Note that |X| = n/4
≥
for
2k
4
k/2
√
k = o( n).
k log n
n
Because D uses less than k log
bits of memory, then D has at most 2 6 = nk/6 unique
6
k/4
′
memory configurations. Since |X| ≥ n4
, then there are at least 21 (|X| − nk/6 ) ≥ |X|
4 pairs x, x
′
such that D has the same configuration after reading x and x . We show that D errs on a significant
fraction of these pairs x, x′ .
Let I be the positions where either x or x′ differ from ν, so that k2 + 1 ≤ |I| ≤ k. Observe that
if HAM (x, y) = k2 , but x and y do not differ in any positions of I, then HAM (x′ , y) > k2 . Recall
that D has the same configuration after reading x and x′ , so then D has the same configuration
after reading x ◦ y and x′ ◦ y. But since HAM (x, y) = k2 and HAM (x′ , y) > k2 , then the output of
D is incorrect for either x ◦ y or x′ ◦ y.
For each pair (x, x′ ), there are n/4−|I|
≥ n/4−k
such y with HAM (x, y) = k2 , but x and y do
k/2
k/2
n/4−k
not differ in any positions of I. Hence, there are |X|
strings S(x, y) for which D errs. Recall
4
k/2
n/4
k
k
+
that y satisfies either HAM (x, y) = 2 or HAM (x, y) = 2 + 1 so that there are |X| n/4
k/2+1
k/2
strings x ◦ y in total. Thus, the probability of error is at least
|X|
|X| n/4−k
4
k/2
n/4
k/2
+
1
= 4 ·
n/4
k/2+1
n/4−k
k/2
n/4+1
k/2+1
=
(k/2 + 1) (n/4 − 3k/2 + 1) . . . (n/4 − k)
4
(n/4 − k/2 + 1) . . . (n/4 + 1)
k/2
k/2 + 1 n/4 − 3k/2 + 1 k/2
k+2
k
≥
=
1−
n+4
n/4 − k/2 + 1
2n + 8
n/4 − k/2 + 1
2
k
1
k+2
1−
≥
≥
2n + 8
n/2 − k + 2
n
√
where the last line holds for large n, from Bernoulli’s Inequality and k = o( n).
√
Lemma 33 For k = o( n), any k-period of the string S(x, y) = x ◦ y ◦ x ◦ x is at least
✷
n
4.
Proof : We show that
if p < n4 , k > 2, and n > 4(18k + 1)(18k + 2), then
p nstronger result that
√
|{z|S[z] 6= S[z + p]}| > 8 > k, for k = o( n).
Let T = ν ◦ ν ◦ x ◦ x and for each z, consider T [z] and T [z + p]. For each j > 0, some position
z + p in 12j 02j 12j+1 02j+1 in the second ν corresponds with a mismatch in z. Since HAM (x, ν) = k2
and HAM (x, y) ≤ k2 + 1, then HAM S 1, n2 , T 1, n2 ≤ 3k
2 + 1. Each mismatch between S and
T can cause at most
two indices z for whichn T [z] 6= T [z + p] but S[z] = S[z + p]. Thus, by setting
+
1
+ 2k, we have that for 4 > (12k + 1)(12k + 2), there are at least 6k indices z
j = 6k > 2 3k
2
for which T [z] 6= T [z + p], and thus at least 2k indices for which S[z] 6= S[z + p].
✷
Corollary 34 If HAM (x, y) = k2 , then the string S(x, y) = x ◦ y ◦ x ◦ x has period
hand, if HAM (x, y) = k2 + 1, then S(x, y) has period greater than n4 .
n
4.
On the other
√
Theorem 35 For k = o( n) with k > 2, any one-pass streaming algorithm which computes the
smallest k-period of an input string S with probability at least 1 − n1 requires Ω(k log n) space, even
under the promise that the k-period is at most n2 .
23
n
bits of memory cannot distinguish
Proof : By Theorem 32, any algorithm using less than k log
6
k
k
between HAM (x, y) = 2 and HAM (x, y) = 2 + 1 with probability at least 1 − 1/n. Thus, no
algorithm can distinguish whether the period of S(x, y) is n4 with probability at least 1 − 1/n while
n
bits of memory.
✷
using less than k log
6
References
[AEL10]
Amihood Amir, Estrella Eisenberg, and Avivit Levy. Approximate periodicity. Algorithms and Computation, pages 25–36, 2010. 1.2
[AG97]
Alberto Apostolico and Zvi Galil, editors. Pattern Matching Algorithms. Oxford University Press, Oxford, UK, 1997. 1.2
[AGMP13] Alexandr Andoni, Assaf Goldberger, Andrew McGregor, and Ely Porat. Homomorphic
fingerprints under misalignments: sketching edit and shift distances. In Proceedings of
the 45th annual ACM symposium on Theory of computing, pages 931–940, 2013. 1.2
[BEMS14] Petra Berenbrink, Funda Ergün, Frederik Mallmann-Trenn, and Erfan Sadeqi Azer.
Palindrome recognition in the streaming model. In 31st International Symposium on
Theoretical Aspects of Computer Science (STACS), pages 149–161, 2014. 1.2
[CEPR09] Raphaël Clifford, Klim Efremenko, Ely Porat, and Amir Rothschild. From coding
theory to efficient pattern matching. In Proceedings of the 20th Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 778–784, 2009. 1.2
[CFP+ 15]
Raphaël Clifford, Allyx Fontaine, Ely Porat, Benjamin Sach, and Tatiana A.
Starikovskaya. Dictionary matching in a stream. In Algorithms - ESA 23rd Annual
European Symposium, Proceedings, pages 361–372, 2015. 1.2
[CFP+ 16]
Raphaël Clifford, Allyx Fontaine, Ely Porat, Benjamin Sach, and Tatiana A.
Starikovskaya. The k -mismatch problem revisited. In Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA, pages 2039–2052, 2016.
1, 1.2, 5, 2.1, 6, 2.1, 4, 13
[CJPS13]
Raphaël Clifford, Markus Jalsenius, Ely Porat, and Benjamin Sach. Space lower bounds
for online pattern matching. Theoretical Computer Science, 483:68–74, 2013. 1.2
[CM11]
Michael S. Crouch and Andrew McGregor. Periodicity and cyclic shifts via linear
sketches. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques - 14th International Workshop, APPROX, and 15th International Workshop, RANDOM. Proceedings, pages 158–170, 2011. 1.2
[EAE06]
Mohamed G. Elfeky, Walid G. Aref, and Ahmed K. Elmagarmid. STAGGER: periodicity mining of data streams using expanding sliding windows. In Proceedings of the
6th IEEE International Conference on Data Mining (ICDM), pages 188–199, 2006. 1.2
24
[EJS10]
Funda Ergün, Hossein Jowhari, and Mert Saglam. Periodicity in streams. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, 13th International Workshop, APPROX 2010, and 14th International Workshop, RANDOM 2010. Proceedings, pages 545–559, 2010. 1, 1.2, 3, 4.6, 5
[EMS10]
Funda Ergün, S. Muthukrishnan, and Süleyman Cenk Sahinalp. Periodicity testing
with sublinear samples and space. ACM Trans. Algorithms, 6(2):43:1–43:14, 2010. 1.2
[Gaw13]
Pawel Gawrychowski. Optimal pattern matching in lzw compressed strings. ACM
Transactions on Algorithms (TALG), 9(3):25, 2013. 1.2
[GKP16]
Shay Golan, Tsvi Kopelowitz, and Ely Porat. Streaming Pattern Matching with d
Wildcards. In 24th Annual European Symposium on Algorithms (ESA), pages 44:1–
44:16, 2016. 1.2
[GMSU16] Pawel Gawrychowski, Oleg Merkurev, Arseny M. Shur, and Przemyslaw Uznanski.
Tight tradeoffs for real-time approximation of longest palindromes in streams. In 27th
Annual Symposium on Combinatorial Pattern Matching, CPM, pages 18:1–18:13, 2016.
1.2, 7.2
[GS83]
Zvi Galil and Joel Seiferas. Time-space-optimal string matching. Journal of Computer
and System Sciences, 26(3):280–294, 1983. 1.2
[GSZ17]
Elena Grigorescu, Erfan Sadeqi Azer, and Samson Zhou. Streaming for aibohphobes:
Longest palindrome with mismatches. CoRR, abs/1705.01887, 2017. 1.2
[IKM00]
Piotr Indyk, Nick Koudas, and S. Muthukrishnan. Identifying representative trends in
massive time series data sets using sketches. In VLDB, Proceedings of 26th International
Conference on Very Large Data Bases, pages 363–372, 2000. 1.2
[KMP77a] Donald E Knuth, James H Morris, Jr, and Vaughan R Pratt. Fast pattern matching
in strings. SIAM journal on computing, 6(2):323–350, 1977. 1.2
[KMP77b] Donald E. Knuth, James H. Morris Jr., and Vaughan R. Pratt. Fast pattern matching
in strings. SIAM J. Comput., 6(2):323–350, 1977. 1
[KR87]
Richard M. Karp and Michael O. Rabin. Efficient randomized pattern-matching algorithms. IBM Journal of Research and Development, 31(2):249–260, 1987. 2
[LN11]
Oded Lachish and Ilan Newman. Testing periodicity. Algorithmica, 60(2):401–420,
2011. 1.2
[MNSW95] Peter Bro Miltersen, Noam Nisan, Shmuel Safra, and Avi Wigderson. On data structures and asymmetric communication complexity. In Proceedings of the 27th Annual
ACM Symposium on Theory of Computing, pages 103–111, 1995. 30
[PL07]
Ely Porat and Ohad Lipsky. Improved sketching of hamming distance with error
correcting. In Annual Symposium on Combinatorial Pattern Matching, pages 173–182,
2007. 1.2
25
[PP09]
Benny Porat and Ely Porat. Exact and approximate pattern matching in the streaming
model. In 50th Annual IEEE Symposium on Foundations of Computer Science, FOCS,
pages 315–323, 2009. 1, 1.2
[RS16]
Jakub Radoszewski and Tatiana Starikovskaya. Streaming k-mismatch with data recovery and applications. arXiv preprint arXiv:1607.05626, 2016. 1.2
[Yao77]
Andrew Chi-Chih Yao. Probabilistic computations: Toward a unified measure of complexity (extended abstract). In 18th Annual Symposium on Foundations of Computer
Science, FOCS, pages 222–227, 1977. 3, 7.2
26
| 8cs.DS
|
1
Energy Harvesting Communications Using
Dual Alternating Batteries
Rajshekhar Vishweshwar Bhat, Mehul Motani, Chandra R Murthy and Rahul
arXiv:1801.03813v2 [cs.IT] 15 Jan 2018
Vaze
Abstract
We consider an energy harvesting transmitter equipped with two batteries having finite storage
capacities, communicating over an additive white Gaussian channel. The work is motivated by an
observation that many practical batteries, when repeatedly charged after being partially discharged, suffer
from degradation in the usable capacity. The capacity can be recovered by completely discharging the
battery before charging it fully again. Hence, in this work, we impose the constraint that a battery must
be charged (discharged) only after it is fully discharged (charged). Our goal is to maximize the long-term
average throughput with non-causal and causal knowledge of the energy arrivals, which we assume to
be Bernoulli. We propose two sub-optimal policies and obtain an upper bound on the performance gap
(G) from the optimal long-term average throughput that is achieved with infinite capacity batteries. We
find that G remains constant as the amount of energy harvested per arrival increases. Numerically, we
also find that G decreases with the battery capacity faster than the inverse of the square root of the
battery capacity.
I. I NTRODUCTION
Recently, energy harvesting (EH) from natural and man-made sources for powering electronic
devices has been gaining tremendous interest. In EH systems, energy arrives randomly and
the harvested energy needs to be stored in batteries for reliable system operation. For efficient
management of the harvested energy, several architectures have been proposed [1], [2]. When
Rajshekhar Vishweshwar Bhat and Mehul Motani are with the Department of Electrical and Computer Engineering,
National University of Singapore, Singapore (e-mails: [email protected], [email protected]). Chandra R. Murthy is
with the Dept. of ECE, Indian Institute of Science, Bangalore 560012, India (e-mail: [email protected]). Rahul Vaze is with
the School of Technology and Computer Science, Tata Institute of Fundamental Research, Mumbai 400005, India (e-mail:
[email protected]).
2
the system is equipped with a single battery, a dual-path architecture, wherein the harvested
energy is simultaneously used to charge the battery and power the load directly, is shown to
perform better than other architectures [2], [3]. However, in some cases, when directly powering
the load from EH sources is infeasible, the performance with the single battery can significantly
degrade because it is not possible to simultaneously save energy in the battery and power the
load. Hence, energy harvesting is suspended while the battery energy is being used. In such
cases, a two-battery solution can potentially far outperform the single battery solution [3].1
In this work, we consider an EH transmitter equipped with two batteries having finite storage
capacities. In our two-battery architecture, the batteries are charged and discharged alternately.
That is, when a battery, referred to as the working battery, is being discharged, the harvested
energy is stored in the other battery, referred to as the charging battery, as charging and discharging a battery simultaneously is not possible [4], [5]. In this case, one must switch the role of the
batteries when the charging battery becomes full or when the working battery becomes empty, to
avoid energy overflow and blank (no-transmission) periods, respectively. Depending on the energy
arrival and consumption rates, these two events may not always occur simultaneously. This leads
to charging (discharging) of batteries after discharging (charging) them partially. In practice, such
a charging and discharging pattern affects the usable capacity of batteries. For instance, the usable
capacity of many Nickel based batteries [6] and some Li-ion [7] batteries reduces at a significant
rate with the number of charge/discharge cycles when a battery is repeatedly charged after a
partial discharge. This phenomenon, referred to as the memory effect or voltage depression, can
be avoided by discharging (charging) the battery completely before fully charging (discharging)
it again [8]. In this work, we do not focus on modeling the voltage depression, but rather, we
charge and discharge the batteries such that the voltage depression does not occur. With the
above constraint, our goal in the work is to maximize the long-term average throughput.
In the EH communications literature, the impact of various battery limitations, such as inefficiency, finite capacity, internal resistance and leakage have been considered [4], [9], [10].
However, the impact of charging and discharging dynamics on the usable capacity of batteries
has been scarcely studied. The authors in [10] characterize the interplay between the battery
1
Practical batteries cannot be simultaneously charged and discharged. The works which assume the batteries can be
simultaneously charged and discharged, implicitly assume either the dual-path architecture or a two-battery framework where
one of the batteries acts as an intermediate storage device.
3
a1
EH
BI
S1
Source
a2
S2
b1
BII
Transmitter
b2
Fig. 1: The system model adopted in the work. The node is equipped with two finite capacity batteries BI and BII each with capacity of B J.
At any given time, the EH source and the transmitter will be connected to a1 (a2 ) and b2 (b1 ), respectively.
discharge policy and the irreversible degradation (aging) of the storage capacity. We note that
this phenomenon is different from voltage depression. Further, in [11], [12], the authors indirectly
control the battery degradation by constraining the number of charge and discharge cycles per
unit time. The work in this paper is related to [13] which considers an EH transmitter equipped
with a single battery. In [13] a Bernoulli energy arrival model assumed, where either an energy
packet with energy amount equal to the battery capacity arrives or no energy arrives in a slot.
This implies, whenever a packet of energy arrives, the battery fills up completely. In this case,
when a fresh energy packet arrives, the residual energy in the battery can be thought to be
discarded instantaneously before replenishing it with the freshly arrived energy. Hence, we see
that [13] implicitly accounts for the constraint considered in this paper. Further, we note that [13]
implicitly assumes a two-battery solution. Our work generalizes [13] by accounting for the case
when multiple energy arrivals are required to fill a battery completely. The main contributions
of the paper are:
•
We obtain optimal power allocations with non-causal (offline) and causal (online) knowledge
of energy arrivals. For the online case, we note, discarding the residual energy when a fresh
arrival occurs, as in [13], may be sub-optimal. We present a dynamic programming based
policy with a better performance which optimally utilizes the remaining energy in the battery
when a fresh energy packet arrives.
•
We also propose two sub-optimal online policies for the finite capacity battery case. We show
that the performance gap between the optimal policy for an ideal system equipped with infinite
capacity batteries and the proposed policies with finite capacity batteries decays faster than
the inverse of the square root of the battery capacity.
II. S YSTEM M ODEL
We consider a transmitter equipped with two batteries, BI and BII (see Fig. 1). We assume
that each battery has the storage capacity of B J. The harvested energy is stored in one of the
batteries before using it for the transmission [1]. Since it is not feasible to simultaneously charge
4
and discharge a battery [4], [5], we operate the system as follows. When BI is being charged
from the EH source, the transmitter draws power from BII and vice versa. Hence, at any point
in time, the EH source and the transmitter will be connected to a1 (a2 ) and b2 (b1 ), respectively,
as shown in Fig. 1. Based on the discussion in the introduction, we switch the role of batteries
only when the working battery is completely discharged and the charging battery is completely
charged.
We consider a time-slotted system with unit slot length. Energy arrivals in each slot are
independent and identically distributed (i.i.d.) Bernoulli random variables with parameter p,
where B/NB , NB ∈ Z+ , units of energy is harvested per arrival. That is, the amount of energy
harvested in slot i is,
Ei =
B
NB
0
w.p. p,
w.p. 1 − p,
(1)
where w.p. stands for with probability. Let Biw and Bic denote the amount of energy stored in
the working and the charging battery at the start of slot i, respectively. We assume B1w = B and
B1c = 0. Let the transmit power in slot i be Pi . Then,
c
Bi+1
= min(Bic + Ei , B), i = 1, 2, . . .
(2)
w
Bi+1
= max(Biw − Pi , 0), i = 1, 2, . . .
(3)
The min and max in the above equations capture the facts that the battery energy cannot exceed
its capacity or become negative, respectively. The communication is over an additive white
Gaussian noise channel having unit power spectral density. We assume the throughput with the
constant transmit power of P W in a slot is given by
1
2
log(1 + P ) bits-per-second (bps). All
the logarithms are to the base 2.
III. P ROBLEM F ORMULATION
We now formulate the long-term average throughput maximization problem subject to the
constraint that the working (charging) battery must be discharged (charged) completely before
switching the roles of the batteries. Since the harvested energy arrives randomly, the switching
instants will be random. Due to the above constraint, the switching instants are the renewal
instants of the battery state processes. Hence, we use the terms ‘switching’ and ‘renewal’ instants
interchangeably. Let Lk denote the length (in slots) of the k th renewal period. Let the number of
slots required for the charging battery to accumulate B units of energy in the k th renewal period
5
be Ck which is less than or equal to Lk by our design. Since the system is reset after a renewal
instant and because the energy arrivals and allocations are independent across renewal intervals,
L1 , L2 , . . . form a sequence of i.i.d. random variables. The k th renewal occurs in Skth slot, where,
Sk = Sk−1 + Lk ,
k = 2, 3, . . . ,
(4)
where we define S1 , 1. Further, due to the constraint described above, we must have,
BSwk = B, and BSc k = 0,
∀ k = 1, 2, . . .
We now define the long-term average throughput as follows.
" k
#
X1
1
log(1 + Pi )
T , lim E
k→∞ k
2
i=1
(5)
(6)
where the expectation is over all the possible sequences of energy arrivals. Our objective is
to find the optimal power policy, P1 , P2 , . . ., that maximizes the long-term throughput, i.e., the
optimization problem of interest is given by,
maximize T
(7a)
subject to Pk ≥ 0, k = 1, 2, . . .
(7b)
Pk ,k=1,2,...
(5),
S
i −1
X
Pk ≤ B, i = 2, 3, . . . ,
(7c)
k=Si−1
where (7c) is because the maximum total amount of energy that can be consumed in a renewal
period is B.
IV. O PTIMAL S OLUTION
In this section, we obtain the long-term throughputs under various cases. Before we proceed,
we note the following. Firstly, by renewal-reward theorem [14], the long-term average throughput
in (6) is equal to,
E
hP
L
1
i=1 2
log(1 + Pi )
i
,
(8)
E[L]
where L is the length of the renewal period, whose distribution is identical to L1 , L2 , . . . ,. The
T =
expectation is with respect to the random variable L. Secondly, the random variable C, which
is the number of slots required to accumulate B units of energy, has the negative binomial
distribution given by,
n−1
pNB (1 − p)n−NB ,
Prob(C = n) , qn =
n − NB
for n ∈ {NB , NB + 1, . . .} and E(C) = NB /p.
(9)
6
A. An Upper-Bound on the Long-term Average Throughput
When no constraints are present, and the nodes are equipped with infinite capacity batteries,
it is optimal to utilize ǫ smaller amount of power per slot on average, compared to the average
harvesting rate, where ǫ > 0 can be arbitrarily small. Further, by the concavity of the rate
function, it is optimal to utilize an equal amount of energy µ = Bp/NB in every slot, resulting
in the long-term average throughput [15]
1
log(1 + µ)
(10)
2
The subscript ‘ub’ indicates that this performance is an upper bound on the performance with
Tub =
any other constraints.
B. Optimal Offline Policy
In the offline policy, we assume that the number of slots required to completely charge the
battery is known at the start of the current renewal instant, i.e., the realization of Ck+1 is known
at start of slot Sk for k ∈ {1, 2, . . .}. Due to the concavity of
1
2
log(1 + P ) in P , it is optimal
to transmit with a constant power over the Ck+1 slots. Hence, when Ck+1 is known, we can
discharge the working battery such that it gets emptied in the same slot in which the charging
battery fills up. Hence, we have Lk+1 = Ck+1 and we transmit at the constant power of B/Lk+1
in the k th renewal. Hence, from (8) and (9), the optimal long-term throughput is given by
∞
B
p X nqn
.
(11)
log 1 +
Toff =
NB n=N 2
n
B
It is easy to numerically compute the value of Toff .
C. Optimal Online Policy
We now consider the online case with only causal knowledge of renewal instants. Since the
precise time when the charging battery will get full is unknown, power is allocated based only
on the distribution of energy arrivals. To obtain the optimal online policy, we adopt a dynamic
programming framework. We now define the relevant quantities.
a) State space: The state of the system is defined by 3-tuples s , (b1 , b2 , α), where b1 and
b2 are the amounts of energy stored in batteries BI and BII , respectively, and α is an indicator
variable defined as follows.
α=
1 if BI is the working battery,
I
0 if B is the charging battery.
(12)
7
By design, we note that when α = 1 (α = 0), BII is the charging (working) battery. The state
space is given by
S = {(b1 , b2 , α) : 0 ≤ b1 , b2 ≤ B, α ∈ {0, 1}}
(13)
b) Action space and reward: The action space the system can take in state s = (b1 , b2 , α) ∈
S is given by
A(s) = {P : 0 ≤ P ≤ b1 α + b2 (1 − α)}
(14)
The constraint on P in (14) is due to the fact that P units are drawn from the working battery.
The reward when the system takes action P ∈ A(s) in state s ∈ S is given by
1
2
log(1 + P ).
c) State transition probability matrix: The next state s′ , (b′1 , b′2 , α′ ) ∈ S when the system
takes action P ∈ A(s) in state s = (b1 , b2 , α) ∈ S is given by
(b′1 , b′2 ) =
(b1 − P α, b2 − P (1 − α))
w.p. 1 − p,
(15)
(b1 − P α + E(1 − α), b2 − P (1 − α) + Eα) w.p. p,
where we have accounted for the fact that, in a slot, the event that the transmitter harvests and
stores E units of energy in the charging battery occurs with probability p. As the role of the
batteries must be switched when the working battery becomes empty and the charging battery
becomes full, we have,
1 − α if ((b1 , b2 ) = (0, B)) or ((b1 , b2 ) = (B, 0)),
′
α =
α
otherwise.
(16)
From (15) and (16), we can easily construct the probability transition matrix, q(s′ |P, s) for all
s ∈ S and P ∈ A(s).
d) Optimal value function: We consider K slots for the optimization. We obtain the optimal
value function Vk (s) in slot k ∈ {1, . . . , K} and state s ∈ S by solving the following Bellman
equation.
X
1
q(s′ |P, s)Vk+1(s′ )}
Vk (s) = max { log(1 + P ) +
P ∈A(s) 2
s′ ∈S
(17)
for k = K, K − 1, . . . , 1, where VK+1(s) , 0. The optimal online throughput is then given by
VK (s)
K→∞
K
Ton = lim
(18)
8
We note that the optimal online throughput, Ton , can also be obtained by solving the Bellman
equation in the infinite horizon case with the discount factor arbitrarily close to one. Since the
reward is bounded and the state space is finite, there exists an optimal stationary deterministic
policy for (17), i.e., there exists a unique optimal action P ∗ (s) in state s independent of slot
indices [16]. Hence, it suffices to search only in the set of all stationary deterministic policies.
D. Proposed Suboptimal Online Policy
In this policy, we assume the energy remaining in the working battery is discarded once
the charging battery accumulates B units of energy. On the other hand, if the working battery
gets completely discharged before the charging battery is full, the transmitter waits without
transmission till the charging battery gets full. This assumption implies that the renewal instant
is the same as the instant when the charging battery accumulates B units of energy. Hence,
the renewal length, L is distributed identically as C in (9). Now, we obtain the optimal power
allocation. Our analysis proceeds along the lines in [13], [17]. Let P̃i be the transmit power in
slot i ∈ {1, . . . , L} after a renewal, i.e., P̃i , PSk +i−1 for k = 1, 2, . . . ,. Then, the throughput
in (8) can be re-written as
Tprop
n
∞
p X X1
log(1 + P̃i ),
qn
=
NB n=1 i=1 2
=
∞
X
i=1
where Fi (p, NB ) , (p/NB )
P∞
n=i qn .
1
Fi (p, NB ) log(1 + P̃i ),
2
From (7), (8) and (19), to maximize the long-term average
throughput in the online case, we need to solve the following optimization problem.
∞
X
1
Fi (p, NB ) log(1 + P̃i )
maximize
2
P̃i ,i=1,2,...
i=1
subject to
∞
X
(19)
P̃i ≤ B, P̃i ≥ 0, i = 1, 2, . . .
(20a)
(20b)
i=1
Clearly, (20) is a convex optimization problem. Hence, Karush-Kuhn-Tucker (KKT) conditions
are necessary and sufficient for optimality. Based on the KKT conditions, we obtain the optimal
power allocations in the following theorem.
Lemma 1. Let
Pn
i=1 Fi (p, NB )
≤ Fn (p, NB ) .
N = max n :
B+n
(21)
9
For k = 1, 2, . . ., the optimal transmit power PS∗k +i−1 is given by PS∗k +i−1 = P̃i∗ , where
(B+N )p
P
− 1 for i = 1, . . . , NB ,
NB Ni=1 Fi (p,NB )
(B+N )Fi (p,NB )
P̃i∗ =
PN
−1
for i = NB + 1, . . . , N,
i=1 Fi (p,NB )
0
i > N.
(22)
Proof. See Appendix A [18].
We note the probability that the charging battery fills up in less than NB slots after a renewal
is zero. Hence, it is intuitive that we must consume the available energy at the constant rate
in the first NB slots after a renewal, as suggested by (22). Further, when NB = 1, we recover
Theorem 1 of [13]. Since we are discarding the remaining energy in the working battery, the
∗
optimal solution in Lemma 1 is a sub-optimal online policy and clearly, we have, Tprop
≤ Tub ,
∗
where Tprop
is obtained by substituting P̃i∗ from (22) into (19).
E. Simplified Suboptimal Online Policy
While preserving the structure of the policy in Lemma 1, we propose another policy where
the power allocation in the ith slot after a renewal instant is given by
P̃i∗ = BFi (p, NB ) =
∞
Bp X
qn .
NB n=i
(23)
We discard the energy remaining in the working battery when the charging battery becomes full.
Let Tsimp denote the long-term average throughput obtained in this strategy. We now have the
following result.
Theorem 2. The long-term average throughput with the power allocation policy in (23) satisfies
Tub ≥ Tprop ≥ Tsimp ≥ Tub − G(NB ),
where G(NB ) = maxp − 2NpB
P∞
i=1 (
Proof. See Appendix B in [18].
P∞
n=i qn ) log (
(24)
P∞
n=i qn ) .
We make the following remarks on the above theorem.
•
Firstly, G(1) ≈ 0.72 and we recover Proposition 3 of [13].
•
Further, numerically, we can show that maxNB G(NB ) ≈ 0.72. Hence, the long-term average
throughputs in the simplified and the proposed suboptimal policies are at most 0.72 bits away
10
7
·10−2
Average Throughput, T (bps)
6.5
6
5.5
5
Tub
Toff
Ton
Tprop
Tsimp
4.5
4
3.5
2
4
6
8
10
12
B J
14
16
18
20
22
Fig. 2: Variation of the optimal long-term average throughput with B for p = 0.1, E = 1 and with B = NB .
from the upper bound Tub for any value of p, B and NB , under the Bernoulli energy harvesting
model.
•
Finally, when the parameters E and p are fixed, NB increases proportionately with B, as
NB = B/E. Based on numerical analysis, we find the bound G(NB ) decreases monotonically
at a rate faster than the inverse of the square root of NB and B.
F. Constant Power Policy
In this policy, we consume the energy available in the battery at a constant rate of B/(⌊NB p−1 ⌋)
as long as the battery is not empty, i.e., the power allocation in the ith slot after a renewal instant
is given by
P̃i =
B
,
⌊NB p−1 ⌋
for i = 1, . . . , ⌊NB p−1 ⌋.
(25)
We discard the remaining energy in the working battery when the charging battery becomes
full. We denote the long-term average throughput of the policy by Tconst . Several variants of this
policy have been considered in the literature [13] and references therein, and they have been
shown to perform competitively with optimal policies. Specifically, the version proposed in [19]
has been shown to approach Tub asymptotically with B. Hence, we compare suboptimal online
policies with the constant power policy.
11
Average Throughput, T (bps)
4
3
2
1
Tub
Toff
Tprop
Tsimp
Tconst
Tlb = Tub − 0.35
0
−1
101
102
103
E
Fig. 3: Variation of the optimal long-term average throughput with E for B = ENB , NB = 4 and p = 0.1.
V. N UMERICAL R ESULTS
In Fig. 2, we plot the variation of the long-term average throughput with the battery capacity
in the optimal offline, optimal online, proposed suboptimal and simplified suboptimal online
policies. We note as the battery capacity increases, the performance gap between offline and
online policies decreases. Further, the performance gap of the proposed suboptimal online policy
from the upper bound, Tub tends to zero as B increases much faster than the simplified suboptimal
policy.
In Fig. 3, we plot the variation of the long-term average throughput with the energy harvested
per arrival in optimal offline, proposed suboptimal online, simplified suboptimal online and
constant power policies, along with upper and lower bounds from (24). Although the performance
of the constant power policy is slightly better than the simplified suboptimal policy for smaller
values of E, the proposed suboptimal policy always performs better than the constant power
policy. Further, as E increases, the performance of the constant power policy diverges while the
offline, proposed and the simplified suboptimal policies are always within 0.35 bits from the
upper bound. We note G(NB )|NB =4 ≈ 0.35.
In Fig. 4, we plot the variation of the long-term average throughputs with the energy arrival
probability, p, for a fixed mean harvesting rate of µ. Note that the lower the value of p, the higher
is the amount of energy harvested per arrival. In this case, the long-term average throughputs
are low as the optimal power allocation may not be constant across the slots. For the higher
12
Average Throughput, T (bps)
0.5
0.45
0.4
0.35
Toff
Tprop
Tsimp
0.3
0.25 −3
10
10−2
10−1
100
Probability, p
Fig. 4: Variation of the optimal long-term average throughput with p for a fixed mean harvesting rate µ = 1, for B = 1000.
values of p, the harvested energy arrives more uniformly and the power allocation can be nearly
constant, and, hence, the performance is better.
VI. C ONCLUSIONS
We considered an energy harvesting transmitter equipped with two batteries. In order to avoid
recoverable capacity degradation of the batteries, we charge (discharge) a battery only when it is
completely discharged (charged). With this constraint, we obtained long-term average throughputs
in offline and online policies. We proposed two online sub-optimal policies, and found that their
performance gap from the unconstrained optimal performance decreases with the battery capacity
faster than the inverse of the square root of the battery capacity.
A PPENDIX
A. Proof of Lemma 1
The Lagrangian of (20) is given by
!
∞
∞
∞
X
X
X
1
L=−
Fi (p, NB ) log(1 + P̃i ) + λ
P̃i − B −
µi P̃i ,
2
i=1
i=1
i=1
(26)
where λ, µi ≥ 0 are Lagrange multipliers. Differentiating L with respect to P̃i , we have the
following stationarity condition.
Fi (p, NB )
∂L
+ λ − µi = 0,
=−
∂ P̃i
2 ln 2(1 + P̃i )
i = 1, 2, . . .
(27)
13
Further, the complementary slackness conditions, λ
P
∞
i=1
P̃i − B = 0 and µi P̃i = 0, must be
satisfied in the optimal solution. Hence, whenever P̃i > 0, we must have, µi = 0. From (27),
we get,
P̃i =
Fi (p, NB )
− 1,
2λ ln 2
i = 1, 2, . . . ,
(28)
where the optimal λ can be found from (20b), as shown below.
Solving for the optimal λ: Due to the total energy constraint, we have,
N
N
X
X
Fi (p, NB )
−1 ,
B=
P̃i =
2λ ln 2
i=1
i=1
(29)
N
1 X
=
Fi (p, NB ) − N,
2λ ln 2 i=1
PN
Fi (p, NB )
∗
=⇒ λ = i=1
,
2 ln 2(B + N)
(30)
(31)
where N is the last slot where P̃i > 0. From (28) and the expression for λ∗ , we find that N is
the largest n that satisfies
Pn
Fi (p, NB )
≤ Fn (p, NB )
B+n
i=1
(32)
Hence, from (28),
(B + N)Fi (p, NB )
P̃i = PN
− 1,
F
(p,
N
)
i
B
i=1
i = 1, 2, . . . , N
(33)
The transmit power in the rest of the slots are zero. Further, noting that Fi (p, NB ) = p/NB for
i = 1, . . . , NB , we obtain (22).
14
B. Proof of Theorem 2
Tsimp =
=
a
≥
E
E
E
hP
L
1
i=1 2
log(1 + P̃i )
E(L)
hP
L
log(1 +
hP
L
log(1 +
1
i=1 2
1
i=1 2
i
Bp
NB
E(L)
,
P∞
Bp
)
NB
n=i qn )
+
i
PL
1
i=1 2
,
log(
P∞
n=i qn )
i
,
E(L)
h
i
i
hP
P∞
L 1
E L 21 log(1 + NBpB )
q
)
log
(
E
n=i n
i=1 2
=
+
,
E(L)
E(L)
!
l
∞
∞
X
p X X1
Bp
1
ql
)+
log
qn ,
= log(1 +
2
NB
NB l=1 i=1 2
n=i
!
!
∞
∞
∞
X
p X X
Bp
1
1
)+
qn
log
= log(1 +
qn ,
2
NB
NB i=1 n=i
2
n=i
≥
1
log(1 + µ) − G(NB )
2
(34)
where (a) is because log(1 + αx) ≥ log(1 + x) + log(α) for 0 ≤ α ≤ 1 and we note 0 ≤
P∞
P∞
P∞ P∞
1
n=i qn )).
n=i qn ≤ 1, and we define G(NB ) , maxp (−p/NB
i=1 (
n=i qn ) 2 log (
R EFERENCES
[1] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes: Survey and implications,” IEEE Commun. Surveys
Tuts., vol. 13, no. 3, pp. 443–461, Third 2011.
[2] S. Bandyopadhyay et al., “Platform architecture for solar, thermal, and vibration energy combining with MPPT and single
inductor,” IEEE J. Solid-State Circuits, vol. 47, no. 9, pp. 2199–2215, Sept 2012.
[3] R. V. Bhat, M. Motani, and T. J. Lim, “Dual-path architecture for energy harvesting transmitters with battery discharge
constraints,” in IEEE Globecom, Dec 2015, pp. 1–6.
[4] ——, “Energy harvesting communication using finite-capacity batteries with internal resistance,” IEEE Trans. Wireless
Commun., vol. 16, no. 5, pp. 2822–2834, May 2017.
[5] S. Luo, R. Zhang, and T. J. Lim, “Optimal save-then-transmit protocol for energy harvesting wireless transmitters,” IEEE
Trans. Wireless Commun., vol. 12, no. 3, pp. 1196–1207, March 2013.
[6] R. A. Huggins, “Mechanism of the memory effect in nickel electrodes,” Solid State Ionics, vol. 177, no. 26, pp. 2643 –
2646, 2006.
[7] T. Sasaki, Y. Ukyo, and P. Novák, “Memory effect in a lithium-ion battery,” Nature Materials, vol. 12, no. 6, pp. 569–575,
Jun 2013, article.
[8] G. M. Siddesh, G. C. Deka et al., Cyber-Physical Systems: A Computational Perspective. Chapman & Hall/CRC, 2015.
15
[9] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi et al., “Energy harvesting wireless communications: A review of
recent advances,” IEEE J. Sel. Areas Commun., vol. 33, no. 3, pp. 360–381, March 2015.
[10] N. Michelusi, L. Badia, R. Carli et al., “Energy management policies for harvesting-based wireless sensor devices with
battery degradation,” IEEE Trans. Commun., vol. 61, no. 12, pp. 4934–4947, December 2013.
[11] X. Wu, J. Yang, and J. Wu, “Optimal energy management for energy harvesting transmitters under battery usage constraint,”
in IEEE ISIT, July 2016, pp. 475–479.
[12] J. Yang and J. Wu, “Optimal transmission for energy harvesting nodes under battery size and usage constraints,” in IEEE
ISIT, June 2017, pp. 819–823.
[13] D. Shaviv and A. Ozgur, “Universally near optimal online power control for energy harvesting nodes,” IEEE J. Sel. Areas
Commun., vol. 34, no. 12, pp. 3620–3631, Dec 2016.
[14] R. G. Gallager, Stochastic Processes: Theory for Applications.
Cambridge University Press, 2014.
[15] O. Ozel and S. Ulukus, “Achieving AWGN capacity under stochastic energy harvesting,” IEEE Trans. Inf. Theory, vol. 58,
no. 10, pp. 6471–6483, Oct 2012.
[16] A. Aprem, C. R. Murthy, and N. B. Mehta, “Transmit power control policies for energy harvesting sensors with
retransmissions,” IEEE J. Sel. Topics Signal Process., vol. 7, no. 5, pp. 895–906, Oct 2013.
[17] S. Satpathi, R. Nagda, and R. Vaze, “Optimal offline and competitive online strategies for transmitter-receiver energy
harvesting,” IEEE Trans. Inf. Theory, vol. 62, no. 8, pp. 4674–4695, Aug 2016.
[18] R. V. Bhat, M. Motani, C. R. Murthy, and R. Vaze, “Energy harvesting communications using dual alternating batteries.”
[Online]. Available: http://arxiv.org/abs/1801.03813
[19] R. Srivastava and C. E. Koksal, “Basic performance limits and tradeoffs in energy-harvesting sensor nodes with finite data
and energy storage,” IEEE/ACM Trans. Netw, vol. 21, no. 4, pp. 1049–1062, Aug 2013.
| 7cs.IT
|
Neural Block Sampling
arXiv:1708.06040v3 [cs.AI] 14 Dec 2017
Tongzhou Wang
EECS, UC Berkeley
Yi Wu
EECS, UC Berkeley
Abstract
Efficient Monte Carlo inference often requires
manual construction of model-specific proposals. We propose an approach to automated proposal construction by training neural networks
to provide fast approximations to block Gibbs
conditionals. The learned proposals generalize to occurrences of common structural motifs
both within a given model and across different
models, allowing for the construction of a library of learned inference primitives that can
accelerate inference on unseen models with no
model-specific training required. We explore
several applications including open-universe
Gaussian mixture models, in which our learned
proposals outperform a hand-tuned sampler,
and a real-world named entity recognition task,
in which our sampler’s ability to escape local modes yields higher final F1 scores than
single-site Gibbs.
1
INTRODUCTION
Model-based probabilistic inference is a highly successful paradigm for machine learning, with applications to
tasks as diverse as movie recommendation (Stern et al.,
2009), visual scene perception (Kulkarni et al., 2015),
music transcription (Berg-Kirkpatrick et al., 2014), and
monitoring nuclear tests (Moore and Russell, 2017),
among many others. People learn and plan using mental models, and indeed the entire enterprise of modern
science can be viewed as constructing a sophisticated hierarchy of models of physical, mental, and social phenomena. Probabilistic programming provides a formal
representation of models as sample-generating programs,
promising the ability to explore a rich range of models.
Preliminary work.
David A. Moore
EECS, UC Berkeley
Stuart J. Russell
EECS, UC Berkeley
This requires us to perform efficient inference in novel
models, motivating the development of black-box inference techniques.
Unfortunately, generic inference methods such as singlesite Gibbs sampling (Spiegelhalter et al., 1996) often perform poorly, suffering from slow mixing and being stuck
in local optima. Effective real-world inference often
requires block proposals that update multiple variables
together to overcome near-deterministic and long-range
dependence structures. However, computing exact Gibbs
proposals for large blocks quickly becomes intractable
(approaching the difficulty of posterior inference), and in
practice it is common to invest significant effort in handengineering proposals specific to a particular model.
In this work, we propose to learn tractable block samplers, in the form of approximate Gibbs proposals, that
can then be reused both within a given model and across
models containing similar structural motifs. Recent work
has recognized that a wide range of models can be represented as compositions of simple components (Grosse
et al., 2012), and that domain-specific models may still
reuse general structural motifs such as chains, grids,
rings, or trees (Kemp and Tenenbaum, 2008). By learning flexible samplers, we can improve inference not
only within a specific model but even on previously
unseen models containing similar structures, with no
additional training required. In contrast to techniques
that compile inference procedures specific to a given
model (Stuhlmüller et al., 2013; Le et al., 2017; Song
et al., 2017), learning inference artifacts that generalize
to novel models is valuable in allowing model builders to
quickly explore a wide range of possible models.
We explore the application of our approach to a wide
range of models. On grid-structured models from a
UAI inference competition, our learned proposal significantly outperforms Gibbs sampling even given no
model-specific training. For open-universe Gaussian
mixture models, we show that a simple learned block
proposal yields performance comparable to a modelspecific hand-tuned sampler, and generalizes to models
more than those it was trained on. We additionally ap-
Neural Block Sampling
ply our method to a named entity recognition (NER)
task, showing that not only do our learned block proposals mix effectively, the ability to escape local modes
yields higher-quality solutions than the standard Gibbs
sampling approach.
2
RELATED WORK
There have been great interests in using learned, feedforward inference networks to generate approximate posteriors. Variational autoencoders (VAE) train an inference network jointly with the parameters of the forward
model to maximize a variational lower bound (Kingma
and Welling, 2013). Within the VAE framework, Burda
et al. (2015) and Gu et al. (2015) utilize another neural
network as an adaptive proposal distribution to improve
the convergence of variational inference. However, the
use of a parametric variational distribution means they
typically have limited capacity to represent complex, potentially multimodal posteriors, such as those incorporating discrete variables or structural uncertainty.
A related line of work has developed data-driven proposals for importance samplers (Paige and Wood, 2016;
Le et al., 2017; Ritchie et al., 2016), training an inference network from prior samples which is then used as
a proposal given observed evidence. In particular, Le
et al. (2017) generalize the framework to probabilistic
programming, and is able to automatically generate and
train a neural proposal network given an arbitrary model
described in a probabilistic program. Our approach differs in that we focus on MCMC inference, allowing modular proposals for subsets of model variables that may
depend on latent quantities, and exploit recurring substructure, and that generalize to new models containing
analogous structures with no additional training.
Several approaches have been proposed for adaptive
block sampling, in which sets of variables exhibiting
strong correlations are identified dynamically during inference, so that costly joint sampling is used only for
blocks where it is likely to be beneficial (Venugopal and
Gogate, 2013; Turek et al., 2016). This is largely complementary to our current approach, which assumes the
set of blocks (structural motifs) is given and attempts to
learn fast approximate proposals.
There are exciting recent works that train model-specific
MCMC proposals with machine learning techniques.
Using adversarial training, Song et al. (2017) directly
optimize the similarity between posterior values and
proposed values from a symmetric MCMC proposal
parametrized by a special neural network architecture.
Considering stochastic inverses of graphical models, Inverse MCMC by Stuhlmüller et al. (2013) learns Gibbs-
like proposals from training density estimators using either prior or posterior data. However, both works have
limitations on applicable models and require bootstrapping as well as model-specific training using global information, e.g., samples containing all variables. This
work differ in that our approach is simpler and more scalable, requiring only local information and generating local proposals that may be reused both within and across
different models. Section 4.1.2 explores this comparison
with Inverse MCMC empirically.
Taking a much broader view, the approach in this work
of learning an approximate local update scheme can be
seen as related to approximate message passing (Ross
et al., 2011; Heess et al., 2013), and to recent advances in
learning gradient-based optimizers for continuous objectives (Andrychowicz et al., 2016; Li and Malik, 2017).
3
NEURAL BLOCK MCMC
In this work, we propose a framework of using a neural
network to approximate the Gibbs proposal for a block of
variables. Each learned proposal is specific to a particular block structure and a conditioning set corresponding
to an approximate Markov blanket. Together we refer
to these components as a structural motif. Crucially our
proposals do not fix the model parameters, which are instead provided as network input, so that the same trained
network may be reused to perform inference on novel
models with parameterizations not previously observed.
Our inference networks are parameterized as mixture
density networks (Bishop, 1994), and trained to minimize the KL divergence between the true posterior conditional and the approximate proposal, given prior samples
generated from the model. The approximate proposals
are then accepted or rejected following the MetropolisHastings (MH) rule (Andrieu et al., 2003), so that we
maintain the correct stationary distribution even though
the proposals are approximate. The following sections
describe our approach in more detail.
3.1
Background
Although our approach generalizes to arbitrary probabilistic programs, for simplicity we focus on models
represented as factor graphs. A model consists of a
set of variables V represented as the nodes of a graph
G = (V, E), along with a set of factors specifying a joint
probability distribution pΨ (V ) described by parameters
Ψ. In particular, this paper focuses primarily on directed
models, in which the factors Ψ speicifies the conditional
probability distributions of each variable given its parents. In undirected models, such as the CRFs in Sec. 4.3,
the factors are arbitrary functions associated with cliques
Tongzhou Wang, Yi Wu, David A. Moore, Stuart J. Russell
in the graph (Koller and Friedman, 2009).
Given observations of a set of evidence variables, inference attempts to compute (by way of drawing samples)
the conditional distribution on the remaining variables.
A standard approach is Gibbs sampling (Spiegelhalter
et al., 1996; Andrieu et al., 2003), in which each variable vi is successively resampled from its conditional
distribution p(vi |V¬i ) given all other variables V¬i in the
graph. In most cases this conditional in fact depends
only on a subset of V¬i , known as the Markov blanket,
MB(vi ) ⊆ V¬i . Each Gibbs update can be viewed as a
MH proposal that is accepted by construction, thus inheriting the MH guarantee that the limiting distribution
of the sampling process is the desired posterior conditional distribution (Andrieu et al., 2003).
In models with tight coupling between adjacent variables, proposals that only resample a single variable at
a time will tend to mix very slowly. In many cases it is
necessary to resample multiple variables simultaneously,
i.e., a block proposal. Block proposals can yield much
faster mixing per step, but each step is much slower; the
cost of computing and storing the block conditional distribution is generally exponential in the size of the block,
becoming intractable for large blocks. This motivates the
approach in this paper, in which we train fast, feedforward neural networks to approximate block proposals at
much lower computational cost.
3.2
Structural Motifs in Graphical Models
We associate each learned proposal with a structural motif that determines the shape of the network inputs and
outputs. In general, structural motifs can be arbitrary
subgraphs, but we are more interested in motifs that represent interesting conditional structure between two sets
of variables, the block proposed variables B and the conditioning variables C. A given motif can have multiple
instantiations with a model, or even across models. As
a concrete example, Fig. 1 shows two instantiations of
a structural motif of six consecutive variables in a chain
model. In each instantiation, we want to approximate
the conditional distribution of two middle variables given
neighboring four.
We now formalize the notion of a structural motif.
Definition. A structural motif (B, C) (or motif in short)
is an (abstract) graph with nodes partitioned into two
components, B and C, and a parameterized joint distribution p(B, C) whose factorization is consistent with the
graph structure. This specifies the functional form of the
conditional p(B|C), but not the specific parameters.
A motif may have many instantiations within a particular
graphical model as well as across different models.
Figure 1: Two instantiations of a structural motif in a directed
chain of length 7. The motif consists of two consecutive variables and their Markov blanket of four neighboring variables.
Each instantiation is separated into block proposed variables
Bi (white) and conditioning variables Ci (shaded).
Definition. For a graphical model (G = (V, E), Ψ), an
instantiation (Bi , Ci , Ψi ) of a motif (B, C) includes
1. a subset of the model variables (Bi , Ci ) ⊆ V such
that the induced subgraph on (Bi , Ci ) is isomorphic
to the motif (B, C) with the partition preserved by
the isomorphism (so nodes in B are mapped to Bi ,
and C to Ci ), and
2. a subset of model parameters Ψi ⊆ Ψ required
to specify the joint distribution pΨi (B, C) on motif
variables.
We would typically define a structural motif by first picking out a block of variables B to jointly sample, and
then selecting a conditioning set C. Intuitively, the natural choice for a conditioning set is the Markov blanket,
C = MB(B). However, this is not a fixed requirement,
and C could be either a subset or superset of it (or neither). We might deliberately choose to use some alternate
conditioning set C, e.g., a subset of the Markov blanket to gain a faster proposal, or a superset with the idea
of learning longer-range structure. More fundamentally,
however, Markov blankets depend on the larger graph
structure might not be consistent across instantiations of
a given motif (e.g., if one instantiation has additional
edges connecting Bi to other model variables not in Ci ).
Allowing C to represent a generic conditioning set leaves
us with greater flexibility in instantiating motifs.
Formally, our goal is to learn a Gibbs-like block proposal
q(Bi |Ci ; Ψi ) for all instantiations of a structural motif
{(Bi , Ci , Ψi )}i that is close to the true conditional in the
sense that ∀i, ∀ci ∈ supp(Ci ),
q(Bi ; ci , Ψi ) ≈ pΨ (Bi |Ci = ci )
(1)
This provides another view of this approximation problem. If we choose the motif to have complex structures
in each instantiation, the conditionals pΨ (Bi |Ci = ci )
can often be quite different for different i, and thus difficult to approximate. Therefore, choosing what is a structural motif represents a trade-off between generality of
the proposal and easiness to approximate. While our approach works for any structural motif complying with the
above definition, we suggest using common structures
as motifs, such as chain of certain length as in Fig. 1.
Neural Block Sampling
In principle, recurring motifs could be automatically detected, but in this work, we focus on hand-identified
commons structures.
3.3
Parameterizing Neural Block Proposals
We choose mixture density networks (MDN) (Bishop,
1994) as our proposal parametrization. An MDN is
a form of neural network whose outputs parametrize a
mixture distribution, where in each mixture component
the variables are uncorrelated. Given a sufficiently large
network and number of mixture components, MDNs can
represent arbitrary joint distributions.
In our case, a neural block proposal is a function
qθ parametrized by a MDN with weights θ. The
function qθ represents proposals for a structural motif
{(Bi , Ci , Ψi )}i by taking in current values of Ci and local parameters Ψi and outputting a distribution over Bi .
Then, the goal becomes to optimize θ so that qθ is close
to the true conditional.
In the network output, mixture weights are represented
explicitly. Within each mixture component, distributions
of bounded discrete variables are directly represented as
probability tables, and distributions of continuous variables are represented as isotropic Gaussians with mean
and variance. To avoid degenerate proposals, we threshold the variance of each Gaussian component to be at
least 0.00001.
3.4
Algorithm 1 Neural Block MCMC
Input: Graphical model (G, Ψ), observations y,
motifs {(B (m) , C (m) , Ψ(m) )}m , and
(m)
(m)
(m)
instantiations {(Bi , Ci , Ψi )}i,m in (G, Ψ).
(m)
(m)
(m)
1: for motif with instantiation {(Bi , Ci , Ψi )}i do
2:
if exists proposal trained for this motif then
3:
q (m) ←− trained neural block proposal
4:
else
(m)
5:
Initialize neural block proposal qθ
(m)
6:
Train qθ using prior samples from (G, Ψ)
7:
end if
8: end for
9: x ←− initialize state
10: for t in 1 . . . M do
(m)
(m)
(m)
(m)
11:
Propose x0 ←− qθ on some Bi , Ci , Ψi
12:
Accept or reject according to MH rule
13: end for
loss as the mean instantiation loss:
L(θ) =
=−
ij ∼ Unif{1, . . . , N }
To optimize our MDN , we use the Kullback-Leibler divergence D(pΨ (Bi |Ci ) k qθ (Bi ; Ci , Ψi )) as the measure of closeness between our proposal and the true conditional in Eq. 1. In order to minimize this divergence
across all instantiations (Bi , Ci , Ψj ) and all possible values in supp(Ci ), our total loss will be a linear combination of expected divergences of all instantiations.
In particular, for an instantiation (Bi , Ci , Ψi ), the expected divergence over Ci ’s prior is
pΨ (Bi |Ci )
]]
qθ (Bi ; Ci , Ψi )
= −EBi ,Ci [log qθ (Bi ; Ci , Ψi )] + constant
= ECi [EBi |Ci [log
(2)
(3)
(4)
The second term is a constant independent of θ. Therefore, we define the loss for (Bi , Ci , Ψi ) as
Li (θ) = −EBi ,Ci [log qθ (Bi ; Ci , Ψi )]
N
1 X
EBi ,Ci [log qθ (Bi ; Ci , Ψi )]
N i=1
(5)
In this work, given N instantiations, we define the overall
(6)
(7)
Since the loss L is an average of expected values over priors, we train our neural block proposals using minibatch
SGD. For each training sample in a batch of size K, we
randomly select a motif instantiation, and generate sample values from the prior defined for that instantiation.
Specifically, for batch index j ∈ {1, . . . , K},
Training Neural Block Proposals
ECi [D(pΨ (Bi |Ci ) k qθ (Bi ; Ci , Ψi ))]
N
1 X
Li (θ)
N i=1
bj , cj ∼ pΨ (Bij , Cij )
K
1 X
L̂(θ) = −
log qθ (bj ; cj , Ψij )
K j=1
3.5
(8)
(9)
(10)
Neural Block MCMC
Neural Block MCMC procedure is outlined in Algorithm 1. It is worth pointing out that our framework allows a great amount of flexibility. One may be only interested in a good proposal for a specific part of a particular
model. Then a neural block proposal can be trained with
the underlying motif instantiated only in that part. In
other cases, in order to learn a general proposal, say for
all grid models, one can work with a grid-shaped motif
instantiated in every possible grid model, and extend the
training procedure described in Sec. 3.4 by modifying
Eq. 8 to match an arbitrary distribution over all instantiations. In experiment Sec. 4.1, we use this approach to
train a proposal for all binary-valued grid models. Therefore, it is potentially possible to store a library of neural
block proposals trained on common motifs to speed up
inference on previously unseen models.
Tongzhou Wang, Yi Wu, David A. Moore, Stuart J. Russell
Figure 2: Motif for general grid
models. Conditioning variables
(shaded) form the Markov blanket of proposed variables (white).
Dashed gray arrows show possible
but irrelevant dependencies.
4
Figure 3: Sample runs with single-site Gibbs, Neural Block MCMC, and block Gibbs with true
conditionals. For each model, we compute 10 random initializations and run three algorithms
for 1500s on each one. Plots show average error for each algorithm. Epochs plots are cut off
at 500 epochs to better show the comparison because true block Gibbs finishes far less epochs
than other two within given time.
EXPERIMENTS
In this section, we evaluate our method of learning
neural block proposals against single-site Gibbs sampler as well as several model-specific MCMC methods. In all experiments, the parameterizing MDNs
have elu activation and two hidden layers each of size
λ max {input size, output size}, where 4 ≤ λ ≤ 5, and
output the proposal distribution as a mixture of 4 ≤ m ≤
16 components. Details on the architecture for each experiment are available in the supplementary material.
4.1
Grid Models
We start with a common structural motif in graphical
models, grids. In this section, we focus on binary-valued
grid models of all sorts for their relative easiness to directly compute posteriors. Specifically, to test the performance of MCMC algorithms, we compare the estimated
posterior marginals P̂ against true posterior marginals P
computed using IJGP (Mateescu et al., 2010). Then, for
each inference task with N variables, we calculated the
PN
error N1 i=1 P̂ (Xi = 1) − P (Xi = 1) as the mean
absolute deviation of marginal probabilities.
4.1.1
General Binary-Valued Grid Models
Neural block proposals have the potential to achieve
good performance on models with shared structural motifs without extra training. In our first experiment, we
consider the motif in Fig. 2, which is instantiated in arbitrary binary-valued grid Bayesian networks (BN). The
neural block proposal takes in CPTs of all 9 + 14 = 23
variables as well as the current assignments of 14 conditioning variables in Markov blanket.
In order for the proposal to generalize, we need to consider motif instantiations in all possible binary-valued
grid models, as argued in Sec. 3.5. We thus train the proposal on random grid BNs generated by sampling each
CPT entry i.i.d. from the following mixed distribution:
Figure 4: Performance comparison on 180 grid models from
UAI 2008 inference competition. Each mark represents error
integrals for both single-site Gibbs and our method in a single
run over 1200s inference.
[0, 1]
[1, 0]
Dirichlet([0.5, 0.5])
w.p.
w.p.
w.p.
1
40
1
40
19
20
(11)
After training, the proposal is completely fixed. No
model specific optimization is done in inference time.
We evaluate the performance of the trained neural block
proposal on all 180 grid BNs up to 500 nodes from UAI
20081 inference competition. In each epoch, for each
latent variable, we try to identify and propose the block
as in Fig. 6 with the variable located at center. If this is
not possible, e.g., the variable is at boundaries or close to
evidence, single-site Gibbs resampling is used instead.
Figure 4 shows the performance of both Neural Block
MCMC and singles-site Gibbs in terms of error integrated over time for all 180 models. The models are divided into three classes, grid-50, grid-75 and grid-90, according to the percentage of deterministic relations. Our
Neural Block MCMC significantly outperforms Gibbs
sampler in nearly every model. We notice that the improvement is less significant as the percentage of deter1
http://graphmod.ics.uci.edu/uai08
Neural Block Sampling
Figure 5: Small version of the
triangle grid model in experiment
Sec. 4.1.2. Evidence nodes (shaded)
are at bottom layer. The actual network has 15 layers and 120 nodes.
Figure 6: Motif for the triangle model in
Fig. 5. Conditioning variables (shaded)
form the Markov blanket of proposed
variables (white). Dashed gray arrows
are possible irrelevant dependencies.
ministic relations increases. This is largely due to the
fact that the above proposal structure in Fig. 2 can only
easily handle dependency among the 9 proposed nodes.
We expect an increased block size to yield stronger performance on models with many deterministic relations.
To investigate the behavior of Neural Block MCMC in
detail, we compare it against single-site Gibbs, and exact block Gibbs with the same proposal block, on several
grid models with different percentages of deterministic
relations. Figure 3 shows their performance w.r.t. both
time and epochs on two models. Single-site Gibbs performs worst on both models since it gets stuck very
quickly in local modes. Between the two block proposal MCMC methods, Neural Block MCMC is more
performing in error w.r.t. time due to shorter computational time. However, because the neural block proposal
is only an approximate of the true block Gibbs proposal,
it is worse in terms of error w.r.t. epochs, as expected.
Detailed comparisons on other models are available in
supplementary material.
In summary, our experiment results show that neural
block proposals can achieve significantly faster and better mixing, while using much less computation overhead
than calculating the exact Gibbs block proposal.
4.1.2
Comparison with Inverse MCMC
Neural block proposals can also be used modelspecifically by training only on a particular model. In this
subsection, we demonstrate that our method can achieve
comparable performance with a more complex taskspecific MCMC method, Inverse MCMC (Stuhlmüller
et al., 2013).
Figure 5 illustrates the triangle grid network we use in
this experiment, which is identical to what Stuhlmüller
et al. (2013) used to evaluate Inverse MCMC. For our
method, we chose the motif shown in Fig. 6. The underlying MDN takes in assignments of conditioning variables and all relevant CPTs, then outputs a block proposal of 13 variables. The proposal is trained on all in-
Figure 7: Error w.r.t. epochs on the triangle
model in Fig. 5. Semitransparent line show single
MCMC runs. Opaque lines show averages over
10 MCMC runs for each algorithm. Numbers in
parentheses are the amounts of training data.
stantiations in this triangle model using prior samples.
Inverse MCMC is an algorithm that builds auxiliary data
structures offline to speed up inference. Given an inference task, it computes and trains an inverse graph for
each latent variable where the latent variable is at bottom
and evidence variables are at top. These graphs are then
used in MCMC procedures. In this experiment, we run
Inverse MCMC with frequency density estimator trained
with posterior samples, proposal block size up to 20 and
Gibbs proposals precomputed, following the original approach of Stuhlmüller et al. (2013).
It is difficult to compare these two methods w.r.t. time.
While both methods require offline training, Inverse
MCMC needs to train inverse graphs from scratch if the
set of evidence nodes changes, yet Neural Block MCMC
only needs one-time training for different inference tasks
on this model. In this experiment setting, for each inference epoch, both methods propose about 10.5 values on
average per latent variable. Figure 7 shows a more meaningful comparison of error w.r.t. epochs among singlesite Gibbs, Neural Block MCMC, and Inverse MCMC
with different amount of training data. The learned neural block proposal, trained using 104 samples, is able to
achieve comparable performance with Inverse MCMC,
which is trained using 105 samples and builds modelspecific data structure (inverse graphs).
4.2
Gaussian Mixture Model with Unknown
Number of Components
We next consider open-universe Gaussian mixture models (GMMs), in which the number of mixture components is unknown, subject to a prior. Similarly to Dirichlet process GMMs, these are typically treated with handdesigned model-specific split-merge MCMC algorithms.
Specifically, we study the performance of Neural Block
MCMC and split-merge Gibbs on a GMM shown in
Fig. 9. In this setting, n points x = {xi }i=1,...,n are observed from the GMM with unknown number of active
Tongzhou Wang, Yi Wu, David A. Moore, Stuart J. Russell
Figure 8: All except bottom right: Average log likelihoods of MCMC algorithms run on 200 inference tasks for a total of 600s in
various GMMs. Bottom right: Trace plots of M over 12 runs from initiations with different M values on a GMM with m = 12,
n = 90. Neural Block MCMC explores the sample space significantly faster than Gibbs with SDDS.
mixtures M ∼ Unif{1, 2, . . . , m}. Each point comes
uniformly randomly from one of the M active mixtures.
Our task is to infer the posterior of mixture means µ =
{µj }j=1,...,M , their indicators showing whether active or
not v = {vj }j=1,...,M , and the labels z = {zi }i=1,...,n ,
where zi is the mixture index xi comes from. Formally,
the model can be written as:
M ∼ Unif{1, 2, . . . , m}
µj ∼ N (0, σµ2 I)
v|M ∼ Unif{a ∈ {0, 1}m :
zi |v ∼ Unif{j : vj = 1}
2
xi |zi , µ ∼ N (µz , σ I)
i
j = 1, . . . , m
X
j
aj = M }
i = 1, . . . , n
i = 1, . . . , n,
where m, n, σµ2 and σ 2 are model parameters. Notice that
M is completely dependent
Pof v, so in this experiment,
we always calculate M = j vj instead of sampling M .
The GMM’s many nearly-deterministic relations, e.g.,
p(vj = 0, zi = j) = 0, can cause vanilla single-site
Gibbs stuck in local optima and unable to efficiently
explore the state space by jumping across different M
values. To solve such issues, split-merge MCMC algorithms, such as Restricted Gibbs split-merge (RGSM)
(Jain and Neal, 2004) and Smart-Dumb/Dumb-Smart
(SDDS) (Wang and Russell, 2015), use hand-designed
model-specific MCMC moves that split and merge mixture components.
For Neural Block MCMC framework, it is possible to
deal with such nearly-deterministic relations with a proposal block including all of z, µ and v. However, doing
so would require a long training time of a large network
and cause the resulting proposal not able to generalize to
GMMs of various sizes.
Instead, we consider the
M
model with z variables
collapsed and train a proµj
...
v
posal qθ for two arbij = 1, . . . , m
trary mixtures (µi , vi ) and
(µj , vj ) conditioned on all
zi
other variables (excluding
...
z). The inference task is
xi
indeed equivalent to first
i = 1, . . . , n
sampling µ, v from the Figure 9: GMM with uncollapsed model p(µ, v|x) known number of components
and then sampling z from M , component means µ, and
N observed points x with
p(z|µ, v, x). We modify cluster labels z . We i reprei
the algorithm such that the sent the open-universe model
proposal from qθ is ac- in truncated form, where each
cepted or rejected by the vj determines whether the
P cluster is active, so that
MH rule on the collapsed jth
j vj = M deterministically.
model. Afterwards, z is
resampled from p(z|µ, v, x). This approach is less sensitive to different n values and leads to good performance
in variously sized GMMs, as will be shown below. More
details on this are in the supplementary material.
In training, we notice that such mixture models have
symmetries that must be broken before used as input to
the neural network (Nishihara et al., 2013). In particular, the mixtures {(vj , µj )}j can be permuted in m! ways
and the points {(zi , xi )}i in n! ways. Following a similar procedure by Le et al. (2017), we sort these values
according the first principal component of x, and also
feed the first principal component vector into the MDN.
Using prior samples, we are able to train a neural block
proposal for the mentioned structural motif. In inference,
we randomly choose two clusters to propose at each proposal.
Neural Block Sampling
Although our proposal is trained on a GMM with a specific number of mixtures m = 8 and number of points
n = 60, we also apply it on GMMs with larger m and n
by randomly selecting 8 mixtures and 60 points for each
proposal. Figure 8 shows how the Neural Block MCMC
performs on GMM of various sizes, compared against
split-merge Gibbs with SDDS. In particular, we notice
that as model gets larger, Gibbs with SDDS mixes more
slowly, while Neural Block MCMC still mixes fairly
fast and outperforms Gibbs with SDDS. Bottom right
of Fig. 8 shows a trace plot of M for both algorithms
over multiple runs on the same observation. Gibbs with
SDDS takes a long time to find a high likelihood explanation and fails to explore other possible ones efficiently.
Neural Block MCMC, on the other hand, mixes quickly
among the possible explanations.
4.3
Named Entity Recognition (NER) Tagging
Named entity recognition (NER) is the task of inferring
named entity tags for words in natural language sentences. One way to tackle NER is to train a conditional
random field (CRF) model representing the joint distribution of tags and word features (Liang et al., 2008). In
particular, the model contains weights between word features and tags as well as high order factors of consecutive
tags. For each test sentence, we can build a chain Markov
random field (MRF) containing only the tags variables
using extracted word features and learned CRF model,
and then apply MCMC methods like single-site Gibbs
to sample the NER tags. For this experiment, we use
a dataset of 17494 sentences taken from CoNLL-2003
Shared Task2 . The CRF model is trained with AdaGrad
(Duchi et al., 2011) through 10 sweeps over the training
dataset.
Our goal is to train good neural block proposals for the
chain MRFs built for test sentences. In order to experiment with different block sizes, we train three proposals,
each for a motif of two, three, or four consecutive proposed tag variables and their Markov blanket. With each
proposal, the MDN takes in both the local MRF parameters and assignments of Markov blanket variables, then
outputs the proposal as a mixture of 4 components. Due
to the difficulty in generating natural language sentences,
we reuse the training dataset for CRF model to train neural block proposals.
We then evaluate the learned neural block proposals on
the previously unseen test dataset of 3453 sentences. Figure 10 plots the performance of Neural Block MCMC
and single-site Gibbs w.r.t. both time and epochs on
the entire test dataset. As block size grows larger,
learned proposal takes more time to mix. But even2
http://www.cnts.ua.ac.be/conll2003/ner/
Figure 10: Average F1 scores and average log likelihoods over
entire test dataset. In each epoch, all variables in every test
MRF is proposed roughly once for all algorithms. F1 scores
are measured using states with highest likelihood seen over
Markov chain traces. To better show comparison, epoch plots
are cut off at 500 epochs and time plots at 12850s. Log likelihoods shown don’t include normalization constant.
tually, block proposals generally achieve better performance than single-site Gibbs in terms of both F1 scores
and log likelihoods. Therefore, as shown in the figure, a mixed proposal of single-site Gibbs and neural
block proposals can achieve better mixing without slowing down much. As an interesting observation, Neural
Block MCMC sometimes achieves higher F1 scores even
before passing single-site Gibbs in log likelihood, implying that log likelihood is at best an imperfect proxy for
performance on this task.
5
CONCLUSION
This paper proposes and explores the (to our knowledge) novel idea of training neural nets to approximate
block Gibbs proposals. Our proposals are trained offline
and can be applied directly to novel models given only
a common set of structural motifs. Experiments show
that the neural block sampling approach can help overcome bad local modes comparing with single-site Gibbs
sampling and achieve comparable performance against
model-specialized methods.
In the current stage, our framework requires the user to
manually detect common structural motifs and choose
where and how to apply the pretrained block sampler.
It will be a very interesting direction to investigate, when
given a library of trained block proposals, how an inference system can automatically detect the common structural motifs and (adaptively) apply appropriate samplers
to help convergence for real-world applications.
Tongzhou Wang, Yi Wu, David A. Moore, Stuart J. Russell
References
Christophe Andrieu, Nando De Freitas, Arnaud Doucet, and
Michael I Jordan. An introduction to MCMC for machine
learning. Machine learning, 50(1):5–43, 2003.
Marcin Andrychowicz, Misha Denil, Sergio Gómez,
Matthew W Hoffman, David Pfau, Tom Schaul, and
Nando de Freitas. Learning to learn by gradient descent
by gradient descent. In D. D. Lee, M. Sugiyama, U. V.
Luxburg, I. Guyon, and R. Garnett, editors, Advances
in Neural Information Processing Systems 29, pages
3981–3989. Curran Associates, Inc., 2016.
Taylor Berg-Kirkpatrick, Jacob Andreas, and Dan Klein. Unsupervised transcription of piano music. In Advances in neural
information processing systems, pages 1538–1546, 2014.
Christopher M Bishop. Mixture density networks. 1994.
Yuri Burda, Roger Grosse, and Ruslan Salakhutdinov.
Importance weighted autoencoders.
arXiv preprint
arXiv:1509.00519, 2015.
John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–
2159, 2011.
Roger Grosse, Ruslan R Salakhutdinov, William T Freeman,
and Joshua B Tenenbaum. Exploiting compositionality to
explore a large space of model structures. In 28th Conference on Uncertainly in Artificial Intelligence, pages 15–17.
AUAI Press, 2012.
Shixiang Gu, Zoubin Ghahramani, and Richard E Turner. Neural adaptive sequential Monte Carlo. In Advances in Neural
Information Processing Systems, pages 2629–2637, 2015.
Nicolas Heess, Daniel Tarlow, and John Winn. Learning to pass
expectation propagation messages. In Advances in Neural
Information Processing Systems, pages 3219–3227, 2013.
Sonia Jain and Radford M Neal. A split-merge Markov chain
Monte Carlo procedure for the Dirichlet process mixture
model. Journal of Computational and Graphical Statistics,
13(1):158–182, 2004.
Charles Kemp and Joshua B Tenenbaum. The discovery of
structural form. Proceedings of the National Academy of
Sciences, 105(31):10687–10692, 2008.
Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. arXiv preprint arXiv:1312.6114, 2013.
Daphne Koller and Nir Friedman. Probabilistic graphical models: principles and techniques. MIT press, 2009.
Tejas D Kulkarni, Pushmeet Kohli, Joshua B Tenenbaum, and
Vikash Mansinghka. Picture: A probabilistic programming language for scene perception. In Proceedings of the
ieee conference on computer vision and pattern recognition,
pages 4390–4399, 2015.
Tuan Anh Le, Atlm Gne Baydin, and Frank Wood. Inference
Compilation and Universal Probabilistic Programming. In
20th International Conference on Artificial Intelligence and
Statistics, Fort Lauderdale, FL, USA, 2017.
Ke Li and Jitendra Malik. Learning to optimize neural nets.
arXiv preprint arXiv:1703.00441, 2017.
Percy Liang, Hal Daumé III, and Dan Klein. Structure compilation: trading structure for features. In Proceedings of the
25th international conference on Machine learning, pages
592–599. ACM, 2008.
Robert Mateescu, Kalev Kask, Vibhav Gogate, and Rina
Dechter. Join-graph propagation algorithms. Journal of Artificial Intelligence Research, 37(1):279–328, 2010.
David A. Moore and Stuart J. Russell. Signal-based Bayesian
seismic monitoring. Artificial Intelligence and Statistics
(AISTATS), April 2017.
Robert Nishihara, Thomas Minka, and Daniel Tarlow. Detecting parameter symmetries in probabilistic models. arXiv
preprint arXiv:1312.5386, 2013.
B. Paige and F. Wood. Inference networks for sequential Monte
Carlo in graphical models. In Proceedings of the 33rd International Conference on Machine Learning, volume 48 of
JMLR, 2016.
Daniel Ritchie, Anna Thomas, Pat Hanrahan, and Noah Goodman. Neurally-guided procedural models: Amortized inference for procedural graphics programs using neural networks. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon,
and R. Garnett, editors, Advances in Neural Information
Processing Systems 29, pages 622–630. Curran Associates,
Inc., 2016.
Stephane Ross, Daniel Munoz, Martial Hebert, and J Andrew Bagnell. Learning message-passing inference machines for structured prediction. In Computer Vision and
Pattern Recognition (CVPR), 2011 IEEE Conference on,
pages 2737–2744. IEEE, 2011.
Jiaming Song, Shengjia Zhao, and Stefano Ermon. A-NICEMC: Adversarial training for MCMC. arXiv preprint
arXiv:1706.07561, 2017.
David J Spiegelhalter, Andrew Thomas, Nicky G Best, Wally
Gilks, and D Lunn. BUGS: Bayesian inference using Gibbs
sampling. Version 0.5,(version ii) http://www. mrc-bsu. cam.
ac. uk/bugs, 19, 1996.
David H Stern, Ralf Herbrich, and Thore Graepel. Matchbox:
large scale online Bayesian recommendations. In Proceedings of the 18th international conference on World wide web,
pages 111–120. ACM, 2009.
Andreas Stuhlmüller, Jacob Taylor, and Noah Goodman.
Learning stochastic inverses. In Neural Information Processing Systems, 2013.
Daniel Turek, Perry de Valpine, Christopher J Paciorek, Clifford Anderson-Bergman, et al. Automated parameter blocking for efficient Markov chain Monte Carlo sampling.
Bayesian Analysis, 2016.
Deepak Venugopal and Vibhav Gogate. Dynamic blocking and
collapsing for Gibbs sampling. In Uncertainty in Artificial
Intelligence, page 664. Citeseer, 2013.
Wei Wang and Stuart J Russell. A smart-dumb/dumb-smart
algorithm for efficient split-merge MCMC. In UAI, pages
902–911, 2015.
Supplementary Materials: Neural Block Sampling
6
EXPERIMENT DETAILS
As mentioned in main paper Sec. 4, parameterizing MDNs in all experiments have elu activation and two hidden layers each of size
λ max {input size, output size}, where 4 ≤ λ ≤ 5
depending on the task, and output the proposal distribution as a mixture of 4 ≤ m ≤ 16 components.
6.1
Grid Models
For directed binary-valued grid models, we use higher
amount of MDN mixtures than other experiments because more variables are proposed and general discrete
BNs can have highly multi-modal conditionals.
6.1.1
General Binary-Valued Grid Models
Architecture The underlying MDN has 106-480-480120 network structure, mapping the CPTs of all 23 motif
variables and 14 conditioning variable values to the proposal distribution of 9 proposed variables as a mixture of
12 components.
Additional Sample Runs We provide additional sample runs on four grid models with various percentages of
deterministic relations in Fig. 1.
6.1.2
Comparison with Inverse MCMC
Architecture The underlying MDN has 161-11201120-224 network structure, mapping the CPTs of all 29
motif variables and 16 conditioning variable values to the
proposal distribution of 13 proposed variables as a mixture of 16 components.
6.2
Gaussian Mixture Model with Unknown
Number of Components
Architecture Our neural block proposal is trained using the GMM with m = 8 max mixtures and n = 60 data
points. It proposes two mixture components (µj , vj )s
through underlying MDN of 156-624-624-36 network
structure. The MDN’s inputs include 60 observed points
x = {xi }i , 8 component means µ = {µj }j and component active indicators v = {vj }j with values for the
two proposed mixtures replaced by zeros. Orders for
Figure 1: Additional sample runs with single-site Gibbs, neural
block MCMC, and block Gibbs with true conditionals on UAI
2008 grid models. These results are obtained in same setting as
Fig. 3 of main paper. For each model, we compute 10 random
initializations and run three algorithms for 1500s on each one.
Plots show average error for each algorithm. Epochs plots are
cut off at 500 epochs to better show the comparison. “grid-k”
represents that the model has k% deterministic relations.
Neural Block Sampling
x, µ and v are such that they are sorted along the first
principle component of x to break symmetry. In addition, the inputs also contain the principle component and
the indicators of which components are being proposed.
The MDN outputs a proposal distribution over the two
(µj , vj )s as a mixture of 4 MDN components.
Proposal and Inference with z Collapsed In order to
avoid nearly-deterministic relations, e.g., p(vj = 0, zi =
j) = 0, and still train a general proposal unconstrained
by n, we choose to consider the collapsed model without
z. We first experiment on the intuitive approach which
adds a resampling step for z in the proposal. At each
proposal step, trained proposal qθ is first used to propose new mixtures µ0 and v0 , and then z is proposed
from p(z|µ0 , v0 , x). The MH rule is applied lastly to
either accept or reject all proposed values. While this
method gives good performance in small models, it suffers greatly from low acceptance ratio as n, the number of observed points z, grows large. Therefore, we
eventually choose the approach described in main paper Sec. 4.2, i.e., applying the MH rule on the collapsed
model with z resampled from p(z|µ, v, x) afterwards.
Since the acceptance ratio no longer depends on n, this
approach behaves much more scalable than the first one
in our experiments. It outperforms SDDS split-merge
MCMC in GMMs of various sizes, as shown in Fig. 8 of
main paper.
6.3
NER Tagging
Architecture The underlying MDN two hidden layers
of size 4×max {input size, output size}, with output size
varying according to the number of proposed variables.
It maps local CRF parameters of all motif variables and
conditioning variable values to the NER tag proposal for
consecutive proposed variables as a mixture of 4 components.
| 2cs.AI
|
arXiv:1506.00034v2 [cs.IT] 4 Nov 2017
Bracketing Numbers of Convex Functions on Polytopes
Charles R. Doss ∗
School of Statistics, University of Minnesota
Abstract
We study bracketing numbers for spaces of bounded convex functions in the
Lp norms. We impose no Lipschitz constraint. Previous results gave bounds
when the domain of the functions is a hyperrectangle. We extend these results to the case wherein the domain is a polytope. Bracketing numbers are
crucial quantities for understanding asymptotic behavior for many statistical
nonparametric estimators. Our results are of interest in particular in many
multidimensional estimation problems based on convexity shape constraints.
1
Introduction and Motivation
To quantify the size of an infinite dimensional set, the pioneering work of Kolmogorov and Tihomirov
(1961) studied the so-called metric entropy of the set, which is the logarithm of the
metric covering number of the set. In this paper, we are interested in a related quantity, the bracketing entropy for a class of functions, which serves a similar purpose as
metric entropy. Metric or bracketing entropies quantify the amount of information
it takes to approximate any element of a set with a given accuracy ǫ > 0. This
quantity is important in many areas of statistics and information theory; in particular, the asymptotic behavior of empirical processes and thus of many statistical
estimators is fundamentally tied to the entropy of related classes of functions under
consideration (Dudley, 1978).
Let F be a set of functions on some space X and let d be a metric on F. Given
a pair of functions l, u on X , a bracket [l, u] is the set of all functions f : X → R with
l ≤ f ≤ u pointwise. For ǫ > 0, we say [l, u] is an ǫ-bracket (for d) if d(l, u) ≤ ǫ.
Then the ǫ-bracketing number of F, denoted N[ ] (ǫ, F, d), is the smallest integer
N such that there exist ǫ-brackets [li , ui ], i = 1, . . . , N , such that for all f ∈ F,
f ∈ [li , ui ] for some i. (We do not actually force li , ui ∈ F.) The bracketing entropy
is the logarithm of the bracketing number. Like metric entropies, bracketing entropies are fundamentally tied to rates of convergence of certain estimators (see e.g.,
Birgé and Massart (1993), van der Vaart and Wellner (1996), van de Geer (2000)).
In this paper, we study the bracketing entropy of classes of convex functions. Our interest is motivated by the study of nonparametric estimation of functions satisfying
∗
Charles R. Doss, 224 Church St. SE, Minneapolis, MN 55455. Email: [email protected]
1
convexity restrictions, such as the least-squares estimator of a convex or concave regression function on Rd (e.g., Seijo and Sen (2011), Guntuboyina and Sen (2015)),
possibly in the high dimensional setting (Xu et al., 2016), or estimators of a logconcave or s-concave density (e.g., Seregin and Wellner (2010), Koenker and Mizera
(2010), Kim and Samworth (2016), Doss and Wellner (2016a,b,c), among others).
Entropy bounds, of the metric or bracketing type, are directly relevant for studying
asymptotic behavior of estimators in these contexts.
Fix the dimension d ∈ {2, 3, . . .}. Let D ⊂ Rd be a convex set, let v1 , . . . , vd ∈
Rd , be linearly independent vectors, let B, Γ1 , . . . , Γd be positive reals, and let
v = (v1 , . . . , vd ) and Γ = (Γ1 , . . . , Γd ). For f : D → R, let Lp,D (f ) ≡ Lp (f ) =
1/p
R
p
for 1 ≤ p < ∞, and let L∞ (f ) = supx∈D |f (x)|. We will let C
D f (x) dx
with various arguments denote different classes of convex functions. We let C ≡ Cd
be the class of convex functions on Rd , where we consider all convex functions
f to be defined
on all of Rd and to take the value ∞ off of its effective domain
dom(f ) := x ∈ Rd : f (x) < ∞ (Rockafellar, 1970). (This approach does not affect bracketing numbers.) For a function f and a set D ⊂ Rd , we will use the
notation f : D → R to mean that dom(f ) = D and we let Cd (D) ≡ C (D) be the
class of convex functions on Rd with dom(f ) = D. Then we let
C (D, B, Γ, v) := {f ∈ C (D) : L∞ (f ) ≤ B, |f (x + λvi ) − f (x)| ≤ Γi |λ| if x, x + λvi ∈ D}
(1)
be the class of convex functions on D satisfying uniform boundedness and Lipschitz
constraints given by B and Γ. When {v1 , . . . , vn } Q
is the standard basis of Rd , we
just write C (D, B, Γ). If D is the hyperrectangle di=1 [ai , bi ] (with ai < bi ), then
Bronshtein (1976) and Dudley (1984) (chapter 8) show that if 0 < ǫ ≤ ǫ0 , for some
ǫ0 > 0, then
Y
d
[ai , bi ], B, Γ , L∞ ≤ Cǫ−d/2
log N ǫ, C
(2)
i=1
for a constant C ≡ CD,B,Γ . Here, N (ǫ, F, ρ) is the ǫ-covering number of F in the
metric ρ, which is defined to be the smallest number of balls of ρ-radius ǫ that cover
F, and log N (ǫ, F, ρ) is the corresponding metric entropy of F, discussed in the first
paragraph of this paper.
One would like to use (2) in the study of asymptotic properties of the statistical
estimators discussed above. Unfortunately, the function classes that arise in those
problems generally do not include Lipschitz constraints, and so the class C (D, B, Γ)
is not of immediate use. Furthermore, it turns out that without Lipschitz constraints, the L∞ covering or bracketing numbers are not bounded. Thus, instead
of of using the L∞ distance, we may consider using the Lp distances, 1 ≤ p < ∞.
Let C (D, B) be the class of convex functions on D with uniform bound B and
no Lipschitz constraints. Then Dryanov (2009) and Guntuboyina and Sen (2013)
found bounds when d = 1 and d > 1, respectively, for metric entropies of C (D, B):
they showed that log N (ǫ, C (D, B) , Lp ) . ǫ−d/2 , again with D a hyperrectangle and
1 ≤ p < ∞. Here . means ≤ up to a constant which does not depend on ǫ (but
2
does depend on D, B, and p). The d = 1 case (from Dryanov (2009)) was the fundamental building block in computing global rates of convergence of the univariate
log-concave and s-concave MLEs in Doss and Wellner (2016a). In the corresponding
statistical problems when d > 1, the domain of the functions under consideration
is not restricted to be a hyperrectangle but rather may be an arbitrary convex set
D. Thus the results of Guntuboyina and Sen (2013) are not immediately applicable, and there is need for results on more general convex domains D with a more
complicated boundary and no Lipschitz constraints.
In this paper we are indeed able to generalize the results of Guntuboyina and Sen
(2013) considerably by finding bracketing entropies for all (convex) polytopes D,
attaining the bound
log N[ ] (ǫ, C (D, B) , Lp ) . ǫ−d/2
(3)
with 1 ≤ p < ∞, D a polytope, and 0 < B < ∞. Note that we work with
bracketing entropy rather than metric entropy. Bracketing entropies are larger than
metric entropies for the Lp (1 ≤ p ≤ ∞) metrics (van der Vaart and Wellner, 1996),
so our bracketing entropy bounds imply metric entropy bounds of the same order.
Along the way, we also generalize the results of Bronshtein (1976) to bound the L∞
bracketing numbers of C (D, B, Γ) when D is arbitrary. One of the benefits of our
method is its constructive nature. We initially study only simple polytopes (defined
in Subsection 3.2) and in that case attempt to keep track of how constants depend
on D.
During the course of the development of this paper, we became aware of the
related work Gao and Wellner (2017), which was developed simultaneously and separately from our paper. In Gao and Wellner (2017), the authors demonstrate in
their Theorem 1.6 that if D is a sphere then (3) fails when p(d − 1)/2 > d/2. This
shows that if D is not a polytope the situation may be more complicated than when
D is a polytope. They also find upper bounds of order ǫ−d/2 when D is a polytope.
Their methods are quite different than ours and in particular they do not explicitly
construct their bracketing set but rather rely on an an algebraic relation (see their
function g(·, ·) in their Section 2.5); our method on the other hand is explicitly constructive. We provide some discussion of how our constants can differ from those
of Gao and Wellner (2017) in Example 3.1. In particular, our constants depend on
the volume (measured in the appropriate dimension) of the faces of the polytope D,
which is an interesting phenomenon.
This paper is organized as follows. In Section 2 we prove bounds for bracketing
entropy of classes of convex functions with Lipschitz bounds, using the L∞ metric.
We use these to prove our main result, Theorem 3.1, for the bracketing entropy of
classes of convex functions without Lipschitz bounds in the Lp metrics, 1 ≤ p < ∞,
which we do in Section 3. We defer some of the details of the proofs to Section 4.
There is a notation index at the end of the document.
3
2
Bracketing with Lipschitz Constraints
If we have sets Di ⊂ Rd , i = 1, . . . , M , for M ∈ N, and D ⊆ ∪M
i=1 Di then for ǫi > 0,
N[ ]
X
M
i=1
ǫpi
1/p
, C (D, 1) , Lp
≤
M
Y
i=1
N[ ] (ǫi , C (D, 1) |Di , Lp ) ,
(4)
where, for a class of functions F and a set G, we let F|G denote the class {f |G : f ∈ F}
where f |G is the restriction of f to the set G. We will apply (4) to a cover of D
by sets G with the property that C (D, 1) |G ⊆ C (G, 1, Γ) for some bounded vector Γ, so that we can apply bracketing results for classes of convex functions with
Lipschitz bounds. Thus, in this section, we develop the needed bracketing results
for such Lipschitz classes, for arbitrary convex domains D. Recall the definition
of C (D, B, Γ, v) and C (D, B, Γ) from (1). When we have Lipschitz constraints on
convex functions, we will see that the situation for forming brackets for C (D, 1, Γ)
with D ⊆ [0, 1]d is essentially the same as for forming brackets for C [0, 1]d , 1, Γ .
P
For y, z ∈ Rd let hy, zi := di=1 yi zi , let kzk2 := hz, zi, and for two sets C, D ⊂ Rd ,
define the Hausdorff distance between them by
!
lH (C, D) := max
sup inf kx − yk, sup inf kx − yk .
x∈D y∈C
y∈C x∈D
Let Bd (0, R) := x ∈ Rd : kxk ≤ R , and let
Kd (R) = {D : D is a closed, convex, nonempty set, D ⊆ Bd (0, R)} .
Then Bronshtein (1976) found entropy estimates in the Hausdorff distance for Kd+1 (R),
d ≥ 1 (see also Dudley (1999), chapter 8). These entropy bounds for classes of
convex sets are the main tool for Bronshtein (1976)’s metric entropy bounds for
classes of convex functions, and they will also be the main tool in our bracketing entropy bounds for classes of convex functions with Lipschitz constraints.
The connection between convex sets and convex functions is, of course, found by
considering the epigraphs of the convex functions. If, for B > 0 and a convex
function f defined on a convex set D, we define the (truncated) epigraph VB (f )
by VB (f ) := {(x1 , . . . , xd , xd+1 ) : (x1 , . . . , xd ) ∈ D, f (x1 , . . . , xd ) ≤ xd+1 ≤ B} , then
VB (f ) lies in Kd+1 (R) for some R > 0.
Theorem 2.1 (Bronshtein (1976)). For any R > 0 and any integer d ≥ 1, there
exists a real number cd > 0 such that for all ǫ > 0 there isan ǫ-cover of Kd+1 (R) in
the Hausdorff distance of cardinality not larger than exp cd (R/ǫ)d/2 .
Proof. Bronshtein (1976) proves that log N (ǫ, Kd+1 (R), lH ) ≤ cd,1 (R/ǫ)d/2 for all
0 < ǫ ≤ ǫ0,d R, where N (ǫ, Kd+1 (R), lH ) is the ǫ-covering number of Kd+1 (R) in
the lH metric. We note here only that this can be easily extended to all 0 <
ǫ as follows. By re-scaling, we can take R = 1. Then taking any ǫ ≥ 1, note
4
that the neighborhood of radius ǫ with center given by Bd (0, 1) covers Kd+1 (1), so
log N (ǫ, Kd+1 (1), lH ) = 0 ≤ cd,1 · 1. Then for ǫ0,d ≤ ǫ ≤ 1, by non-increasingness of
ǫ 7→ log N (ǫ, Kd+1 (1), lH ),
−d/2
log N (ǫ, Kd+1 (1), lH ) ≤ cd,1 ǫ0,d
−d/2
≤ cd,1 ǫ0,d ǫ−d/2
−d/2
so the theorem holds with cd := cd,1 ǫ0,d .
The following lemma extends Lemma 4.1 from (Guntuboyina and Sen, 2013) to
the case where G is not necessarily equal to [0, 1]d ; the lemma connects the Hausdorff
distance on sets of epigraphs of Lipschitz functions to the supremum distance for
those functions.
Lemma 2.1. Let G ⊆ [0, 1]d be any convex set and B, Γ1 , . . . , Γd > 0. Let f and g
be elements of C (G, B, (Γ1 , . . . , Γd )). Then
kf − gk∞
1/2
d
X
2
Γi
≤ lH (VB (f ), VB (g)) 1 +
i=1
Proof. For ease of notation, let ρ = lH (VB (f ), VB (g)). Fix x ∈ G and suppose
f (x) < g(x), without loss of generality. Now, (x, f (x)) ∈ VB (f ) and, since f (x) <
g(x), (x, f (x)) is outside the convex epigraph VB (g). Let (x′ , y ′ ) be the projection of
(x, f (x)) onto the convex set VB (g) given by Hilbert’s projection theorem (Rudin,
1987). Then k(x′ , y ′ ) − (x, f (x))k ≤ ρ, and (x′ , y ′ ) = (x′ , g(x′ )) since (x′ , y ′ ) is on
the boundary of VB (g). (Note (x, f (x)) is orthogonal to an element (x′ , B) ∈ VB (g)
only if x′ = x, and since f (x) < g(x) ≤ B if (x′ , y ′ ) = (x, B) then by convexity of g
this forces g ≡ B, and g(x) = B in particular.) Thus
q
0 ≤ g(x)−f (x) = g(x)−g(x′ )+g(x′ )−f (x) ≤ kx−x′ k Γ21 + · · · + Γ2d +|g(x′ )−f (x)|,
since |g(x) − g(x′ )| = |g(x1 , . . . , xd ) − g(x1 , . . . , xd−1 , x′d ) + · · · + g(x1 , x′2 , . . . , x′d ) −
g(x′1 , . . . , x′d )| which is bounded above by
q
′
′
′
|xd − xd |Γd + · · · + |x1 − x1 |Γ1 ≤ kx − x k Γ21 + · · · + Γ2d
by the Cauchy-Schwarz inequality. Thus, again by Cauchy-Schwarz,
X 1/2
0 ≤ g(x) − f (x) ≤ ρ 1 +
Γ2i
,
i
as desired.
Lemma 2.1 is needed for the proof of Theorem 2.2, below. Theorem
Qd 3.2 from
(Guntuboyina and Sen, 2013) gives the result of Theorem 2.2 when D = i=1 [ai , bi ];
we now extend it to the case of a general D. When we consider convex functions
without Lipschitz constraints, we will partition D into sets that are contained in
parallelotopes and apply Theorem 2.2 to those sets.
5
Q
Theorem 2.2. Let ai < bi and let D ⊂ di=1 [ai , bi ] be a convex set. Let Γ =
(Γ1 , . . . , Γd ) and 0 < B, Γ1 , . . . , Γd < ∞. Then there exists a positive constant
c ≡ cd such that
(5)
log N[ ] ǫ Vold (D)1/p , C (D, B, Γ) , Lp ≤ log N[ ] (ǫ, C (D, B, Γ) , L∞ )
d/2
d
X
−d/2
(6)
Γi (bi − ai )
≤ cǫ
B+
i=1
for ǫ > 0 and p ≥ 1.
Here, Vold (D) is d-dimensional volume (Lebesgue measure) of the set D.
Proof. The first inequality of the theorem is elementary. We will show the second
inequality. Fix ǫ > 0. First we note the following scaling relationship. For f ∈
C (D, B, Γ) we can define f˜ : D̃ → R, whereD̃ ⊆ [0, 1]d , by f˜(t1 , . . . , td ) = f (a1 +
t1 (b1 − a1 ), . . . , ad + td (bd − ad )). Then f˜ ∈ C D̃, B, (Γ1 (b1 − a1 ), . . . , Γd (bd − ad )) .
This shows that
N[ ] ǫ, C D̃, B, (Γ1 (b1 − a1 ), . . . , Γd (bd − ad )) , L∞
(7)
= N[ ] (ǫ, C (D, B, (Γ1 , . . . , Γd )) , L∞ ) .
d
Thus, we now let ai = 0 and b
i = 1 and consider a convex domain D̃ ⊂ [0, 1] .
√
Therefore it is clear that if f ∈ C D̃, B then VB (f ) ∈ Kd+1 ( d + B 2 ). We can pick
q
√
an ǫ/ 4 1 + Γ21 + · · · + Γ2d -cover in Hausdorff distance of Kd+1 ( d + B 2 ) of Ñ
elements V1 , . . . , VÑ , with a bound on Ñ as given by Theorem 2.1. We can convert
theseqsets to functions by picking f1 , . . . , fN ,with N ≤ Ñ , such that
lH (VB (fi ), Vi ) ≤
ǫ/(4
1 + Γ21 + · · · + Γ2d ), if such an fi ∈ C D̃, B, (Γ1 , . . . , Γd ) exists. (If no such
fi exists this does not affect the proof; we just have N < Ñ then.) Then from
Lemma 2.1, [fi −ǫ/2, fi +ǫ/2] form an ǫ-L∞ bracketing set for C D̃, B, (Γ1 , . . . , Γd ) .
Thus, by Theorem 2.1, for some positive c ≡ cd ,
log N[ ]
q
d/2
(d + B 2 )(1 + Γ21 + · · · + Γ2d )
.
ǫ, C D̃, B, (Γ1 , . . . , Γd ) , L∞ ≤ c
ǫ
Using (7), we see that
d/2
q
P
(d + B 2 )(1 + i Γ2i (bi − ai )2 )
.
log N[ ] (ǫ, C (D, B, (Γ1 , . . . , Γd )) , L∞ ) ≤ c
ǫ
(8)
6
It is immediate that the left side of (8) equals
ǫ
B Γ1
Γd
log N[ ]
, C D, ,
,...,
, L∞
A
A
A
A
for any A > 0, so that the right side of (8) is bounded above by
q
(dA2 + B 2 ) 1 +
c
P
2
i Γi (bi
ǫ
− ai )2 /A2
for all A > 0. We pick
2
A =
s
B2
Pd
2
i=1 Γi (bi
d
− ai )2
d/2
,
which yields
log N[ ] (ǫ, C (D, B, (Γ1 , . . . , Γd )) , L∞ ) ≤ c
Since
sX
i
Γ2i (bi
− ai
)2
≤
X
i
B+
q P
d/2
d i Γ2i (bi − ai )2
.
ǫ
s X
Γi (bi − ai ) ≤ d
Γ2i (bi − ai )2
i
(by the triangle and Cauchy-Schwarz inequalities), we are done showing the second
inequality of the theorem.
3
Bracketing without Lipschitz Constraints
In the previous section we bounded bracketing entropy for classes of functions with
Lipschitz constraints. In this section we remove those Lipschitz constraints. With
Lipschitz constraints we could consider arbitrary domains D, but without the Lipschitz constraints we need more restrictions: now we will take D to be a simple
polytope (defined below). We now define notation and assumptions we will use for
the remainder of the document.
3.1
Notation and Terminology
We will consider only the case d ≥ 2 since the result when d = 1 is given in Dryanov
(2009). Recall that for a convex set G, a set F ⊂ G is a face of G if F is either ∅
(the empty set), G, or if F = G ∩ H for some supporting hyperplane H (Rockafellar,
1970) of G. A set F ⊂ G is a facet of G if F is a (d − 1)-dimensional face (see e.g.,
Grünbaum (1967)). We will focus on simple polytopes first (see Assumption 1). A
7
simple polytope is one in which all (d − k)-dimensional faces (abbreviated “(d − k)faces”) of D have exactly k incident facets for k ∈ {0, . . . , d}. The simple polytopes
are dense in the class of all polytopes in the Hausdorff distance (page 82 of Grünbaum
(1967)). Any convex polytope can be triangulated into O(n⌈d/2⌉ ) simplices (which
are simple polytopes) if the polytope has n vertices (see e.g. Dey and Pach (1998)),
and so we can translate our theorem into a result for a general polytope D; see
Corollary 3.1. For two sets A and B let A + B := {a + b : a ∈ A, b ∈ B}. For a
vector v ∈ Rd , we let [0, v] := {λv : λ ∈ [0, 1]}. For a set G, let d+ (x, G, e) :=
inf {K ≥ 0 : (x + Ke) ∩ G 6= ∅} (which may in general be infinite). For a point x, a
set H, and a unit vector v, let
d(x, H, v) := inf {|k| : x + kv ∈ H} = min d+ (x, H, v), d+ (x, H, −v)
be the distance from x to H along the vector v, and for a set E, let d(E, H, v) :=
inf x∈E d(x, H, v). We let ∂G be the boundary of G in Rd and we let ∂r G be the
relative boundary of G, the set difference between the closure of G and the relative
interior of G (e.g., page 44 of Rockafellar (1970)). Let Vold−k (G) be the (d − k)dimensional volume of G (and, in particular, Vol0 (G) is the number of elements in
G).1 For a, b ∈ R, we let a ∨ b be the maximum of a and b, and a ∧ b be the minimum
of a and b. For two vectors e, v ∈ Rd and a linear subspace V of Rd , we write e ⊥ v
if he, vi = 0, we write e ⊥ V if e ⊥ v for all v ∈ V , and we let V ⊥ be the orthogonal
complement linear subspace of V in Rd .
3.2
Definitions and Assumptions
In what follows, we will assume
D is a polytope, meaning that for some N ∈ N,
that
d : hv , xi ≥ p
D = ∩N
E
where
E
:=
x
∈
R
j
j
j are halfspaces with inner normal
j=1 j
unit vectors
vj such that vi 6= vj if i 6= j, and where pj ∈ R, for j = 1, . . . , N . Let
Hj := x ∈ Rd : hx, vj i = pj be the corresponding hyperplanes and let Fj := Hj ∩D
be the corresponding facets of D. For
n k ∈ {0, . . . , d}, we will define Jk to index
o the
k
˜
(d − k)-faces of D. First let Jk := (j1 , . . . , jk ) ∈ {1, . . . , N } : j1 < · · · < jk , and
for j ∈ J˜k , let
Gj = ∩kα=1 Hjα ∩ D if k 6= 0,
and let
Gj = D if k = 0.
n
o
Now let J0 = {1}, and for k ∈ {1, . . . , d}, let Jk := j ∈ J˜k : Gj 6= ∅ . The face Gj ,
j ∈ Jk , is (d − k)-dimensional and Hj1 ∩ D, . . . , Hjk ∩ D are the only facets of D
containing Gj , by Theorem 12.14 of Brøndsted (1983). Thus, by John’s theorem,
Theorem 4.1 (John (1948), see also Ball (1992) or Ball (1997)), there exists xj ∈ Gj
1
Technically, Vold−k is known as the so-called (d − k)-dimensional Hausdorff measure. We will
only need the (d − k)-dimensional volume of polytopes contained in affine spaces, and in such cases
the definition is straightforward (and only requires Lebesgue measure).
8
such that Gj − xj contains a (d − k)-dimensional ellipsoid Aj − xj of maximal
(d − k)-dimensional volume and such that
Aj − xj ⊂ Gj − xj ⊂ d(Aj − xj ).
(9)
Let γj,α /2 := d+ (xj , ∂r Aj , eα ) be the radius of Aj in the direction ej,α , where
ej,k+1 , . . . , ej,d are the orthonormal unit vectors given by the axes of the ellipsoid
Aj − xj . Let Ej := span {ej,k+1 , . . . , ej,d } be the linear space containing Gj − xj .
Let A be an integer and u a positive real number, and let
0 = δ0 < δ1 < · · · < δA < u < δA+1 < δA+2 = ∞
(10)
be a sequence. This sequence as well as A and u will be specified in greater detail
later. For k ∈ {1, . . . , d}, let Ik := {0, . . . , A}k , and let I0 := {A}. For k ∈ {1, . . . , d},
i = (i1 , . . . , ik ) ∈ Ik , and j = (j1 , . . . , jk ) ∈ Jk let
Gi,j := {x ∈ D : δiα ≤ d(x, Hjα , vjα ) ≤ δiα +1 for α = 1, . . . , N }
(11)
where in the previous display for α > k we let iα = A + 1 and jα take on the
values in {1, . . . , N } \ {j1 , . . . , jk } (in any order). For the k = 0 case, let GA,1 :=
{x ∈ D : d(x, ∂D) ≥ u} . These sets are not parallelotopes,
since for α > k, δiα +1 =
∞. However, for any x ∈ Gj , (Gi,j − x) ∩ span vj1 , . . . , vjβ , for β ≤ k, is contained
in a β-dimensional parallelotope by construction; this will be S
used S
to understand
d
the volume of Gi,j . We will eventually define u such that D ⊂ k=0 j∈Jk ,i∈Ik Gi,j
(see Lemma 3.1).
The setup for our first main results is summarized in the following assumption.
Assumption 1. Let d ≥ 2, let the definitions of the above Subsection 3.2 hold, and
let D ⊂ Rd be a simple convex polytope.
Additionally, define the support function for a convex set D to be, for x ∈ Rd
with kxk = 1, h(D, x) := maxd∈D hd, xi . Then the width function is, for kuk = 1,
w(D, u) := h(D, u) + h(D, −u), which gives the distance between supporting hyperplanes of D with inner normal vectors u and −u, respectively, and let diam(D) :=
supkuk=1 w(D, u) be the diameter of D.
3.3
Main Results
We want to bound the slope of functions f ∈ C (D, 1) |Gi,j , so that we can apply
bracketing bounds on convex function classes with Lipschitz bounds. Note that
each Gi,j is distance δiα in the direction of vjα from Hjα , which means that if
f ∈ C (D, 1) |Gi,j then f has Lipschitz constant bounded by 2/δiα along the direction
vjα . However, the vectors vjα are not orthonormal, so the distance from Gi,j along
vjα to a hyperplane other than Hjα may be smaller than δiα .
Note that if P ⊂ R ⊂ Rd where R is a hyperrectangle and P is a parallelotope
defined by vectors v1 , . . . , vd , then if A is a linear map with v1 , . . . , vd as its eigenvectors (thus rescaling P ), then AR will not necessarily still be a hyperrectangle,
9
i.e. its axes may no longer be orthogonal. Thus, we cannot argue by simple scaling
arguments that bracketing numbers for P scale with the lengths along the vectors
vi .
For each Gi,j we will find an orthonormal basis such that Gi,j is contained in
a rectangle R whose axes are given by the basis and whose lengths along those
axes (i.e., widths) are bounded by a constant times the width of one of the normal
vectors vjα . Furthermore, the distance from R along each basis vector to ∂D will
be bounded by the distance from Gi,j along vjα to Hjα . This will give us control
of both the Lipschitz parameters and the widths corresponding to the basis, and
thus control of the bracketing number for classes of convex functions. The following
proposition constructs a basis and gives control for the basis elements in span {Gj }.
For the basis elements perpendicular to span {Gj }, control is given by Lemma 4.2
and Lemma 4.3 in Section 4.
Proposition 3.1. Let Assumption 1 hold for a convex polytope D. For each k ∈
{0, . . . , d}, i ∈ Ik , j ∈ Jk , and each Gi,j , there is an orthornormal basis ei,j ≡
e := (e1 , . . . , ed ) of Rd such that for any f ∈ C (D, B) |Gi,j , f has Lipschitz constant
2B/δiα in the direction eα , where δiα = δA+1 if k + 1 ≤ α ≤ d. Furthermore, there
exists a permutation π of (1, . . . , k) such that for α = 1, . . . , k, ei,j,α ≡ eα satisfies
E
o
D
o
n
n
eα ∈ span vjπ(1) , . . . , vjπ(α) , eα ⊥ span vjπ(1) , . . . , vjπ(α−1) , and eα , vjπ(α) > 0,
(12)
o
n
and and for α ∈ {k + 1, . . . , d}, eα ⊥ span vjπ(1) , . . . , vjπ(k) =: V . In particular,
we may take ek+1 , . . . , ed to be the orthonormal unit axis vectors of Aj −xj as defined
on page 9. Thus it is immediate that neither V nor V ⊥ depend on i.
Proof. Without loss of generality, for ease of notation we assume in this proof that
jβ = β for β = 1, . . . , k, and then that
δi1 ≤ δi2 ≤ · · · ≤ δik ≤ δik+1 = · · · = δiN ,
where we let iα = A + 1 for k < α ≤ N . That is, we assume that H1 , . . . , Hk are the
nearest hyperplanes to Gi,j , in order of increasing distance; we then take π to be
the identity. To define the orthonormal basis vectors, we will use a Gram-Schmidt
orthonormalization, proceeding according to increasing distances from Gi,j to the
hyperplanes Hj . Define e1 := v1 and for 1 < j ≤ k, define ej inductively by
ej ∈ span {v1 , . . . , vj } , ej ⊥ span {v1 , . . . , vj−1 } , hej , vj i > 0, and kej k = 1.
Let ek+1 , . . . , ed be orthonormal unit vectors given by the axes of the ellipsoid Aj −xj .
Note that these vectors form an orthonormal basis of span {v1 , . . . , vk }⊥ because
span {ek+1 , . . . , ed } = span(Gj − xj ) is perpendicular to span {v1 , . . . , vk } by definition. For α ∈ {1, . . . , k}, for any x ∈ Gi,j , since d(x, Hα , v) achieves its minimum
10
when v is vα ,
d(x, Hα , eα ) ≥ d(x, Hα , vα ) ≥ δiα ,
d(x, Hj , eα ) ≥ d(x, Hj , vj ) ≥ δij ≥ δiα , for all N ≥ j > α, and
d(x, Hj , eα ) = ∞ > δiα for j < α,
since eα ⊥ span {v1 , . . . , vα−1 }. Similarly, for α ∈ {k + 1, . . . , d},
d(x, Hj , eα ) ≥ d(x, Hj , vj ) ≥ δA+1 , for all N ≥ j ≥ k + 1, and
d(x, Hj , eα ) = ∞ > δA+1 for j ≤ k,
since eα ⊥ span {v1 , . . . , vk }. Thus, we have d(Gi,j , Hj , eα ) ≥ δiα for α ∈ {1, . . . , d}
and for j ∈ {1, . . . , N }. That is, we have shown
d(Gi,j , ∂D, eα ) ≥ δiα for all α ∈ {1, . . . , d} .
(13)
Thus, if f ∈ C (D, B) |Gi,j , then for any x ∈ Gi,j , let z1 = x−γ1 eα and z2 = x+γ2 eα ,
γ1 , γ2 > 0, both be elements of ∂Gi,j , so that by convexity we have
f (z1 ) − f (z1 − δiα eα )
f (x + keα ) − f (x)
2B
f (z2 + δiα eα ) − f (z2 )
−2B
≤
≤
≤
,
≤
δiα
δiα
k
δiα
δiα
using (13). Thus, f satisfies a Lipschitz constraint in the direction of eα .
The next lemma is necessary for us to be able to apply (4). To state it, we
first define some constants. For k ∈ {1, . . . , d}, let di,j,k := d(Ei , Fj ) where Ei ,
i = 1, . . . , Nk , is a (d − k)-face and Fj , j = 1, . . . , N , is a facet. Then let
rD := min {di,j,k : di,j,k 6= 0, k ∈ {1, . . . , d}} > 0.
(14)
Let
2 (p+2)
u ≡ uD := rD /2 ∧ 2−2(p+1)
∧
min
min
d+ (xj , ∂r Gj , e)
Lk,2
(15)
D
E
(16)
k∈{1,...,d−1} j∈Jk ,e∈Ej
where for k ∈ {1, . . . , d − 1},
Lk,2 := 1 ∨ max
max
j∈Jk i∈{1,...,N }\j
k
X
γ=1
D
f˜j,γ , vi
f˜j,γ , vjγ
E,
where f˜j,γ are defined in Proposition 4.1, and Ej is defined on page 9.
Lemma 3.1. Under Assumption 1, with u given in (15), we have
D⊂
d
[
[
k=0 j∈Jk ,i∈Ik
11
Gi,j .
Proof. Fix x ∈ D. We need to show that there are no more than d facets F such that
d(x, F ) < u. If d(x, ∂D) ≥ u then x ∈ GA,1 (corresponding to k = 0), so we assume
d(x, ∂D) < u. Then let kx := max {k ∈ {1, . . . , d} : d(x, G) < u, some (d − k)-face G}
and let Gx be any (d − kx )-face such that the minimum is attained. Now for any
facet F , if d(x, F ) < u then we also have d(Gx , F ) < 2u ≤ rD . But this contradicts
the definition of rD unless d(Gx , F ) = 0. Because Gx is nonempty, Gx = Gj for
some j ∈ Jkx (rather than j ∈ J˜kx \ Jkx ). The distance from x to the boundary
of Gx is no smaller than u, because otherwise we would contradict the maximality
defining kx since the boundary is given by (d − (kx + 1))-faces. Thus the distance
from x to any facet intersecting but not containing Gx is no smaller than u. Furthermore because D is simple, there are exactly kx ≤ d facets containing Gx ; and
we have shown that the distance to every facet excluding these kx is no smaller than
u. Thus, Gx is unique and x lies in Gi,j for some i ∈ Ikx .
The next lemma combines Lemma 4.2 and Lemma 4.3 with Theorem 2.2. The
statement depends on the constants Lk,1 , k ∈ {1, . . . , d}, and Lj,4 , j ∈ Jk . These
depend only on D and are defined in (42) and (24).
Lemma 3.2. Let Assumption 1 hold. Fix k ∈ {1, . . . , d}, i ∈ Ik , j ∈ Jk . Then for
ǫ > 0,
log N[ ] ǫ Vold (Gi,j )1/p , C (D, 1) |Gi,j , Lp
!d/2
d
(17)
2
X
8L
ρ
2d
δ
k,1 j,α
iα +1
≤ cd ǫ−d/2 1 +
.
max
+
Lj,4 α=1,...,k δiα
u
α=k+1
Proof. Let
Γi :=
2
2
2
2
,...,
, ,··· ,
d(Gi,j , ∂D, e1 )
d(Gi,j , ∂D, ek ) u
u
(18)
where ei,j,α ≡ eα , α = 1, . . . , d, is given by Proposition 3.1. Then
C (D, 1) |Gi,j ⊂ C (Gi,j , 1, Γi , e)
(19)
where e = (e1 , . . . , ed ). Let f˜jγ be given by Lemma 4.1 applied to the k linearly
independent unit normal vectors vj1 , . . . , vjk , and (as in that lemma, with “dβ ” given
by (δiγ +1 − δiγ )), let
E
D
(20)
fi,j,jγ ≡ fjγ := (δiγ +1 − δiγ )f˜jγ / f˜jγ , vjγ .
P
Let Pi,j := kγ=1 [0, fjγ ], where [0, v] := {λv : λ ∈ [0, 1]}. By Lemma 4.2, Pi,j ⊂
Pk
α=1 [0, γα eα ] where γα are given by the lemma. Thus by (50), for some x ∈ Gi,j ,
Gi,j ⊂ x +
k
X
[0, γα eα ] +
α=1
d
X
α=k+1
12
[−2Lk,1 ρj,α eα , 2Lk,1 ρj,α eα ] .
(21)
Now, using (19), we apply Theorem 2.2 to see
log N[ ] ǫ Vold (Gi,j )1/p , C (D, 1) |Gi,j , Lp
≤ cd ǫ−d/2
k
X
d
X
8Lk,1 ρj,α
2γα
1+
+
d(Gi,j , ∂D, eα )
u
α=1
α=k+1
!d/2
(22)
Now by applying (52), (58), and (59) with v = eα , we see that
δiβ +1
2d minβ=1,...,k
2d diam(Gi,j , eα )
2γα
≤
≤
d(Gi,j , ∂D, eα )
d(Gi,j , ∂D, eα )
minβ=1,...,k
D
eα ,vjβ
δiβ
D
eα ,vjβ
E
E
≤
δiβ +1
2d
max
Lj,4 β=1,...,k δiβ
(23)
where
Lj,4 := min
e1 ,...,ed
D min E
vjβ : vjβ ,eα >0
eα vjβ .
(24)
(We can restrict to vjβ such that vjβ , eα > 0 in the definition of Lj,4 because the
numerator in (23) is finite.) Thus (22) is bounded above by
cd ǫ
−d/2
d
X
δiβ +1
8Lk,1 ρj,α
2d2
1+
+
max
Lj,4 β=1,...,k δiβ
u
α=k+1
!d/2
.
Now we present our main theorem. It gives a bracketing entropy of order ǫ−d/2
when D is a fixed simple polytope. Its proof relies on embedding Gi,j in a set Ri,j
(defined in (49)) which is a set-sum of a parallelotope and a hyperrectangle with
axes given by Proposition 3.1. We need to control the distance of Gi,j to ∂D, and
we need to control the size of Ri,j in terms of the widths along its axes. Then we
can use the results of Section 2 on Ri,j and thus on Gi,j . We defer some statements
and proofs of needed facts about Gi,j and Ri,j until Section 4.
Q
Theorem 3.1. Let Assumption 1 hold for a convex polytope D ⊆ di=1 [ai , bi ]. Fix
p ≥ 1. Then for all ǫ > 0,
log N[ ] (ǫ, C (D, B) , Lp ) ≤ Sǫ
−d/2
1/p d/2
Y
d
,
(bi − ai )
B
i=1
where S is a constant depending only on d and D.
The form of the constant S is given in the proof of the theorem.
13
(25)
Proof. Fix ǫ > 0. First, we will reduce to the case
D ⊂ [0, 1]d and B = 1 by a
Qwhere
d
scaling argument. Let C be an affine map from i=1 [ai , bi ] to [0, 1], where D̃
is the
image of D, and assume we have a bracketing cover [˜l1 , ũ1 ], . . . , [˜lN , ũN ] of C D̃, 1 .
Let li := B ˜li ◦ C and similarly for ui , so that [l1 , u1 ], . . . , [lN , uN ] form brackets for
C (D, B). Their Lpp size is
Z
Z
d
Y
p
p
p
˜
(bi − ai )dx.
(ui (x) − li (x)) dx = B
(ũi (x) − li (x))
D
D̃
Thus,
N[ ] ǫB
Y
d
bi − a i
1/p
, C (D, B) , Lp
so apply the theorem with η = ǫ/B
Q
d
bi − ai
≤ N[ ] ǫ, C D̃, 1 , Lp ,
1/p
for ǫ. Note that the constant S
depends on D̃, the version of D normalized to lie in [0, 1]d .
We now assume D ⊂ [0, 1]d and B = 1. We specify the sequence in (10) and
ai,k ≡ ai > 0, which will govern the Lp -sizes of our brackets on Gi,j , as follows. Let
)
(
p + 1 i−1
log ǫ
for i = 1, . . . , A,
and
δ0 = 0.
(26)
δi := exp p
p+2
Note that this implicitly defines A, by (10) and (15). For k ∈ {1, . . . d} and i ∈ Ik ,
we will let a(i1 ,...,ik ) = 2 if iα = 0 for any α ∈ {1, . . . , k}, and otherwise we let
a(i1 ,...,ik ) :=
k
Y
β=1
aiβ :=
k
Y
ǫ
1/k
β=1
For the k = 0 case, let aA := ǫ/u. Let
X
d
X
a=
(p + 1)iβ −2
log ǫ .
exp −p
(p + 2)iβ −1
1/p
api Vold (Gi,j )
k=0 j∈Jk ,i∈Ik
.
(27)
Then since D ⊂ ∪dk=0 ∪j∈Jk ,i∈Ik Gi,j by Lemma 3.1, as in (4),
log N[ ] (a, C (D, 1) , Lp ) ≤
d X X
X
k=0 j∈Jk i∈Ik
log N[ ] ai Vold (Gi,j )1/p , C (D, 1) |Gi,j , Lp .
(28)
First, consider the case k ∈ {1, . . . , d} and compute the sum over Ik for a fixed
j ∈ Jk . We use the trivial bracket [−1, 1] for any Gi,j where iα = 0 for any
α ∈ {1, . . . , k}. Otherwise apply Lemma 3.2 which shows us that the sum over the
remaining terms in (28) is bounded by
!d/2
A
d
A
X
X
X
8Lk,1 ρj,α
2d2
δiα +1
−d/2
.
(29)
cd ai
1+
···
max
+
Lj,4 α=1,...,k δiα
u
i1 =1
ik =1
α=k+1
14
Since Lk,1 ≥ 1 and u ≤ ρj,α by (15) for all k, i, j and α = k + 1, . . . , d, we have
Pd
P
Q
8ρj,α Lk,1
2ρ
2ρ
= 4Lk,1 dα=k+1 uj,α ≤ 4Lk,1 dα=k+1 uj,α (using the fact that
α=k+1
u
Q
for a, b ≥ 2, ab ≥ a + b). We also bound maxα=1,...,k 2δiα +1 /δiα ≤ kα=1 2δiα +1 /δiα
since 2δiα +1 /δiα > 2. Thus (29) is bounded above by
2
cd d
L−1
j,4
d−k+2
1+2
Lk,1
d
Y
ρj,α
u
α=k+1
!d/2
A
X
i1 =1
···
A
X
−d/2
ai
k
Y
2δi
α=1
ik =1
α +1
δiα
d/2
, (30)
which is
cd d2 L−1
j,4
d−k+2
1+2
d
Y
ρjα
u
Lk,1
α=k+1
!d/2
A
X
i1 =1
···
A Y
k
X
2δiβ +1 d/2
ik =1 β=1
δiβ aiβ
.
(31)
Note that when k = d we take the product over an empty set to be 1. For i =
1, . . . , A, let
q
ζi := ǫ1/k δi+1 /(δi ai ),
(32)
d/2
P
PA Q k
2δiβ +1
so that A
·
·
·
equals
i1 =1
ik =1
β=1 δi ai
β
A
X
i1 =1
···
A
X
2kd/2 ǫ−d/2
ik =1
k
Y
ζidβ = 2kd/2 ǫ−d/2
Bu :=
A
X
i=1
A
X
i1 =1
β=1
where, for 0 < ǫ ≤ 1
by Lemma 3.3.
β
ζid1
A
X
i2 =1
= ǫ−d/2 2kd/2 Buk
ζid2 · · ·
A
X
ζidk
ik =1
ζid ≤ 2ud/(2(p+1)(p+2)) ,
(33)
d/2
Q
ρ
Next, we will relate the term 1 + 2d−k+2 Lk,1 dα=k+1 ujα
to Vold−k (Gj ).
Recall that Aj is the ellipsoid defined in (9) which has diameter in the eα direction
given by γj,α . By (9), ρj,α ≤ dγj,α . The volume of Aj is Vold−k (Aj ) =
Q
d−k Γ((d−k)/2+1)
d
(d−k)/2 /Γ((d − k)/2 + 1). Thus, letting C := (2d)
,
d
α=k+1 γj,α /2 π
π (d−k)/2
we have
d
Y
ρj,α ≤ Cd Vold−k (Aj ) ≤ Cd Vold−k (Gj ).
α=k+1
Thus we have shown that (31) is bounded above by
d/2
d−k+2
−(d−k)
kd/2
1
+
2
L
u
C
Vol
(G
)
Buk · ǫ−d/2 .
cd d2 L−1
2
k,1
d
d−k
j
j,4
15
(34)
Therefore, letting c̃d,k := cd d2 2k 2kd/2 , we have shown that
X
log N[ ] ai Vold (Gi,j )1/p , C (D, 1) |Gi,j , Lp
i∈Ik
d/2
d−k+2
−(d−k)
kd/2(p+1)(p+2)
1
+
2
L
u
C
Vol
(G
)
ǫ−d/2 .
≤ L−1
c̃
u
k,1
d
d−k
j
j,4 d,k
(35)
Display (35) holds for k ∈ {1, . . . , d}. When k = 0, recalling aA = ǫ/u, we have
(36)
log N[ ] aA Vold (GA,1 )1/p , C (D, 1) |GA,1 , Lp ≤ cd (u + 2d)d/2 ǫ−d/2
by Theorem 2.2 since C (D, 1) |GA,1 ⊂ C GA,1 , 1, u2 1 where 1 ∈ Rd is the vector of
all 1’s. Then, combining (36) and (35), the cardinality of the collection of brackets
covering the entire domain D is given by summing over j ∈ Jk and k ∈ {0, . . . , d}.
We have computed the cardinality of the brackets. Now we bound their size.
Let Ik0 be the subset of i ∈ Ik such that some iα is 0, and let Ik+ := Ik \ Ik0 . We have
ap ≤ apA Vold (D) +
d
X
X
2p Vold (Gi,j )
k=0 j∈Jk ,i∈Ik0
d
k
X
X
X pY
δiα +1 − δiα
d−k
E
D
+
Vold−k (Gj )
(2Lk,1 )
ai
˜α , vj
f
+
α=1
k=1
j∈Jk
α
i∈I
(37)
k
P
P
by Proposition 4.1 with f˜α ≡ f˜j,α defined there. Note that the term dk=0 j∈Jk ,i∈I 0 2p Vold (Gi,j )
k
is bounded above by 2p ǫp Vold−1 (∂D) (recall δ1 = ǫp ). Fixing k ∈ {1, . . . d}, we have
X
Vold−k (Gj )
X
api
i∈Ik+
j∈Jk
k
k
A Y
A
Y
X
X
X
δiα +1 − δiα
E ≤
D
apiα δiα +1
Vold−k (Gj )Lkj,3
···
˜
fα , vjα
α=1
ik =1 α=1
i1 =1
j∈Jk
≤
X
Vold−k (Gj )Lkj,3
A
X
i1 =1
j∈Jk
api1 δi1 +1 · · ·
A
X
apik δik +1 .
ik =1
E
D
where Lj,3 := maxα∈{1,...,k} 1/ f˜j,α , vjα . We have
A
X
α=1
where Au ≤ 2u
X
j∈Jk
1/(p+1)2
Vold−k (Gj )Lkj,3
apα δα+1 = ǫp/k
A
X
ζα2 =: ǫp/k Au ,
(38)
α=1
by Lemma 3.3. Thus
! A
A
X
X
X
Vold−k (Gj )Lkj,3 ,
apik δik +1 ≤ ǫp Aku
api1 δi1 +1 · · ·
i1 =0
ik =0
16
j∈Jk
2/d
so by (37) a ≤ SD,s ǫ where
2/d
SD,s :=
d
Vol (D)
1/p
X
X
d
p
d−k k
k
+
2
Vol
(∂D)
+
.
(2L
)
A
Vol
(G
)L
d−1
k,1
d−k
j
u
j,3
up
k=1
j∈Jk
(39)
We have thus bounded the bracketing entropy when D ⊂ [0, 1]d and B = 1.
Thus, by the scaling at the beginning of the proof, for any convex polytope D ⊂
Q
1/p
Qd
d
[a
,
b
]
and
any
B
>
0,
we
have
shown
for
0
<
ǫ
≤
B
b
−
a
that
i
i
i
i
i=1
i=1
log N[ ]
2/d
ǫSD̃,s , C
!1/p !d/2
d
Y
(bi − ai )
(D, B) , Lp ≤ SD̃,c ǫ−d/2 B
i=1
where
d/2
SD̃,c := cd (u+2d)
+
d X
X
k=1 j∈Jk
Letting δ :=
2/d
SD̃,s ǫ,
d/2
d−k+2
−(d−k)
kd/2(p+1)(p+2)
1
+
2
L
u
C
Vol
(G
)
.
L−1
c̃
u
k,1
d
d−k
j
d,k
j,4
(40)
we have shown that
log N[ ] (δ, C (D, B) , Lp ) ≤ SD̃,c SD̃,s δ
2/d
for 0 < δ ≤ SD̃,s B
Qd
i=1 (bi
−d/2
!1/p !d/2
d
Y
(bi − ai )
B
(41)
i=1
− ai )1/p . (Note that the constants SD̃,s and SD̃,c should
d , D̃.)
be calculated using the rescaling of D that lies in [0, 1]Q
Finally, we can extend from requiring δ ≤ SD,s B di=1 (bi − ai )1/p to allowing
any δ > 0 just as in the proof of Theorem 2.1, at the slight cost of increasing the
constant on the right hand side of (41).
Lemma 3.3. For any γ ≥ 1, 0 < ǫ ≤ 1, with ζi given in (32), and with A and u
given by (10) and (26), we have
A
X
α=1
2
ζαγ ≤ 2uγ/(2(p+1) ) .
Proof. Notice that ζα ≤ 1. We have, for α = 2, . . . , A,
(
)
p + 1 α−1
−p log ǫ
ζα
= exp
ζα+1
2(p + 1)2 (p + 2) p + 2
(
)
p + 1 A−1
−p log ǫ
≥ exp
2(p + 1)2 (p + 2) p + 2
− log u
≥ exp
=: R.
2(p + 1)2 (p + 2)
17
γ
Then, ζαγ (Rγ − 1) ≤ ζαγ Rγ − (Rζα−1 )γ so ζαγ ≤ (Rγ /(Rγ − 1)) (ζαγ − ζα−1
) and thus
A
X
α=1
ζαγ
≤
ζ1γ
A
Rγ X γ
Rγ
Rγ
γ
γ
γ
+ γ
(ζA
− ζ1γ ) ≤ γ
ζA
(ζα − ζα−1
) = ζ1γ + γ
R − 1 α=2
R −1
R −1
γ
and ζA
≤ uγ/(2(p+1)(p+2)) . Since u ≤ exp −2(p + 1)2 (p + 2) log 2 by its definition
(15), R ≥ 2 so Rγ /(Rγ − 1) ≤ 2 for any γ ≥ 1.
For any convex D and convex subset D̃ ⊂ D, note that C (D, 1) |D̃ ⊂ C D̃, 1 .
Thus by covering any convex polytope D by simple polytopes Di ⊂ D, we can bound
N[ ] (ǫ, C (D, B) , Lp ) by applying Theorem 3.1 repeatedly to C (Di , 1) and using (4).
A cover of D can be attained by, for instance, subdividing D into simple polytopes
(Lee, 1997), such as simplices. The constant in the bound then depends on the
subdivision of D.
Q
Corollary 3.1. Fix d ≥ 1 and p ≥ 1. Let D ⊆ di=1 [ai , bi ] be any convex polytope.
Then for ǫ > 0,
log N[ ] (ǫ, C (D, B) , Lp ) ≤ Cd,D ǫ−d/2 B
!1/p !d/2
d
Y
(bi − ai )
.
i=1
Proof. By the same scaling argument as in the proof of Theorem 3.1 we may assume
[ai , bi ] = [0, 1] and B = 1. The d = 1 case is given by Dryanov (2009). Any convex
polytope D can be triangulated into d-dimensional simplices (see e.g. Dey and Pach
(1998), Rothschild and Straus (1985)). We are done by applying Theorem 3.1 to
each of those simplices, by (4).
Example 3.1. Let D = [0, 1]d be the unit hyperrectangle in Rd . Then f˜j,γ =
ej,γ = vjγ for all k ∈ {1, . . . , d}, j ∈ Jk . Thus d+ (xj , ∂r Gj , e) (appearing in (15))
is no smaller than 1/2 for any e ∈ Ej . Since | hvj , vi i | is 1 if i = j or if vi and vj
correspond to opposite facets, and otherwise is 0, we see that Lk,2 is 1. We can
2
see that rD ≥ 1/2. Thus u = 2−2(p+1) (p+2) . Since eα , vjβ is 0 or 1 we see that
Lj,4 (from (24)) is 1. We can also see that Lk,1 (from (42)) is 1. This is because
for j ∈ {1, . . . , N } \ j and i ∈ j, then hvi , vj i is 0 unless vi and vj correspond to
opposing facets in which case hvi , vj i = −1. Thus there are no j ∈ {1, . . . , N } \ j
and i ∈ j such that hvi , vj i > 0, so the final maximum in (42) is taken over an empty
set and thus Lk,1 is 1. Thus, the constants SD,s and SD,c ((39) and (40)) depend
on D only through the (d − k)-volumes of the faces Gj of D. This phenomenon is
somewhat different than that of Gao and Wellner (2017) where the bound depended
on the triangulation of the polytope D.
When D is not a hyperrectangle, the constants Lk,1 , Lk,2 , and Lj,4 may obfuscate
the computations somewhat. And, of course, if D is not simple, then some partition
or triangulation of of D is necessary, as in Corollary 3.1.
18
4
Properties of Gi,j
In this section we show how to embed the domains Gi,j , which partition D, into
hyperrectangles. We used this in the proof of Theorem 3.1 so we could apply Theorem 2.2. Theorem 2.2 says that the bracketing entropy of convex functions on domain D with Lipschitz constraints along directions e1 , . . . , ek depends on w(D, ei )
(since that gives the maximum “rise” in “rise over run”). In our proof of Theorem 3.1 we partitioned D into sets related to parallelotopes. Thus we will study
these parallelotopes. We know the width of Gi,j in the directions vjα , which are
δiα +1 − δiα , by definition.
P
A polytope P is a d-parallelotope if P = di=1 [ai , bi ] for vectors ai , bi ∈ Rd , where
for all i, [ai , bi ] is not parallel to the affine hull of [aj , bj ] for any j 6= i (Grünbaum
(1967) page 56). We will rely on the following representation for a k-dimensional
parallelotope.
Lemma 4.1. Let k be a positive integer and let P := ∩kβ=1 Ẽβ be a parallelotope
where Ẽβ := x ∈ Rk : 0 ≤ hx, vβ i ≤ dβ for k linearly independent normal unit
vectors vβ . Let Hβ0 := x ∈ Rk : hx, vβ i = 0 . Let f˜β be the unit vector lying in
D
E
0
k
˜
H̃ with fβ , vβ > 0, for β = 1, . . . , k. Then 0 is a vertex of P and we
∩
γ=1,γ6=β
β
can write
P =
k
X
[0, fβ ]
β=1
D
E
where fβ := dβ f˜β / f˜β , vβ , [0, fβ ] = {λfβ : λ ∈ [0, 1]}.
Proof. Since the vectors vβ are unique, ∩kβ=1 Hβ0 = 0 and the intersection of any k −1
n o
of the hyperplanes H 0 gives a 1-dimensional space, span f˜β . A k-dimensional
β
parallelotope can be written as the set-sum of the k intervals emanating from the
vertex, each given by the intersection of k − 1 of the hyperplanes Hβ0 . See page
56 of Grünbaum (1967). Note that fβ satisfy hfβ , vβ i = dβ so that fβ ∈ H̃β+ :=
x ∈ Rk : hx, vβ i = dβ ; thus the k intervals are given by [0, fβ ], β = 1, . . . , k.
Note the vector f˜β can be written as (I − Q)vβ where I is the identity projection in Rk and Q is the projection onto span {v1 , . . . , vβ−1 , vβ+1 , . . . , vk }. The next
proposition uses Lemma 4.1 to bound the widths of Gi,j , in certain directions, in
terms of the width of Gj in those directions. We will need the following constant
(depending on D). For k ∈ {1, . . . , d − 1}, let
Lk,1 := 1 ∨ max max
max
j∈Jk kek=1 j∈{1,...,N }\j; he,vj i<0
e∈Ej
hvi ,vj i>0, some i∈j
h−e, vj i−1 ,
(42)
where Ej := span {ej,k+1 , . . . , ej,d } from Proposition 3.1, and we abuse notation as
convenient to treat j as if it were a set rather than a vector. We also (arbitrarily)
define Ld,1 := 1, for ease of presentation later on.
19
Proposition 4.1. For each k ∈ {1, . . . , d − 1}, i ∈ Ik , j ∈ Jk , and each Gi,j , and
the basis e ≡ ei,j from Proposition 3.1, for α ∈ {k + 1, . . . , d}, we have
w(Gi,j , eα ) ≤ 2Lk,1 w(Gj , eα ).
(43)
D
E
Then for k ∈ {1, . . . , d}, let f˜α ≡ f˜j,α be the unit vector with f˜α , vjα > 0 lying in
span {vj1 , . . . , vjk }∩ ∩kγ=1,γ6=α Hj0γ , α = 1, . . . , k, where Hj0γ := y ∈ Rd : y, vjγ = 0 .
Then for k ∈ {1, . . . , d}, we have
Vold (Gi,j ) ≤ (2Lk,1 )d−k Vold−k (Gj ) ·
k
Y
δiα +1 − δiα
D
E
˜
f
,
v
α=1
α jα
(44)
where Lk,1 is given by (42) for k ∈ {1, . . . , d − 1} (and we set Ld,1 := 1 arbitrarily).
Proof. Fix k ∈ {1, . . . , d − 1}. Let x ≡ xj ∈ Gj (from (9)). Let fjγ be as given in
P
(20). Let Pi,j := kγ=1 [0, fjγ ]. We will show that Gi,j is contained in the set-sum
P
of a hyperrectangle and Pi,j . To begin with let Gi,j ∋ z = x + kγ=1 fj∗γ where
f ∗ = dj f˜j where
jγ
γ
γ
E
E
D
D
0 ≤ djγ ≤ (δiγ +1 − δiγ )/ f˜jγ , vjγ ≤ u/ f˜jγ , vjγ .
(45)
Take an arbitrary e ∈ span {ek+1 , . . . , ed } with kek = 1. Let λz,e := d+ (z, ∂Gi,j , e)
and let j give the corresponding facet of Gi,j that x+λz,e e hits, so that hz + λz,e e, vj i =
pj + u for some j ∈
/ j (abusing notation to treat j as if were a set rather than a
vector). Note that this means
he, vj i < 0.
(46)
DP
E
k
∗
If
γ=1 fjγ , vj ≤ 0 then
λz,e ≤ d+ (x, ∂Gi,j , e).
(47)
E
D
Thus if (47) does not hold then fj∗γ , vj > 0 for some γ ∈ {1, . . . , k}, so hvjα , vj i > 0
for some α ∈ {1, . . . , k}. Now, since hz + λz,e e, vj i = pj + u, we have
λz,e
pj + u − hz, vj i
≤
=
he, vj i
hx, vj i − pj + u
hf˜γ ,vj i
γ=1 hf˜γ ,vj i
γ
Pk
h−e, vj i
Now
hx, vj i − pj ≤ d(x, Hj ) ≤ d+ (x, ∂r Gj , e)
since Hj is the closest hyperplane to x in the direction e. Recall the definition of
Lk,1 in (42). Now, by (15) and the definition of Lk,2 (16), we have shown
λz,e ≤ 2Lk,1 d+ (x, ∂r Gj , e),
20
(48)
by (46) and (47). This means that
(Gi,j − z) ∩ span {ek+1 , . . . , ed } ⊂ 2Lk,1 (Gj − x)
so we can conclude that w(Gi,j −z, eα ) ≤ 2Lk,1 w(Gj , eα ) and w(Gi,j , eα ) ≤ 2Lk,1 w(Gj , eα )
since hz, eα i = 0 for all djγ given by the range (45), α = k + 1, . . . , d, for k =
1, . . . , d − 1.
Let ρj,α := w(Gj , eα ). Then let
Ri,j := Pi,j +
d
X
[−2Lk,1 ρj,α eα , 2Lk,1 ρj,α eα ] .
(49)
α=k+1
Then for any x ∈ Gi,j such that hx, vjα i = pjα + δiα for α ∈ {1, . . . , k}, we have
shown
Gi,j ⊂ x + Ri,j .
(50)
It then also follows that
d−k
Vold (Gi,j ) ≤ (2Lk,1 )
Vold−k (Gj ) · Volk
k
X
α=1
!
[0, fjα ] .
(51)
Since of parallelotopes
with given axis
lengths, the one with largest volume is the
P
Q
δiα +1 −δiα
k
k
hyperrectangle, Volk
α=1 [0, fjα ] ≤
α=1 hf˜j ,vj i , and so we have shown (44)
α
α
P
k
(with this bound on Volk
α=1 [0, fjα ] being all that is needed in the k = d
case).
The previous proposition controls the width and volume of Gi,j in directions
lying in span {Gj }. Next we control width, volume, and also distance to ∂D in
directions perpendicular to span {Gj }.
P
Lemma 4.2. Let P := kα=1 [0, fα ] be a parallelotope in Rk where f1 , . . . , fk are
k linearly independent vectors. Then there exists an orthonormal basis of Rk ,
e1 , . . . , ek ∈ Rk and γ1 , . . . , γk ∈ R, such that
P ⊂
k
X
[0, γα eα ]
where
α=1
|γα | ≤ k diam(P, eα ).
(52)
Proof. We will construct a permutation π of {1, . . . , k} and inductively define e1 , . . . , ek
based on the sequence fπ(1) , . . . , fπ(k) . Let e1 := fπ(1) /kfπ(1) k where kfπ(1) k is maximal over {kfα k}kα=1 . Now let Qj−1 be the projection of Rk onto span {e1 , . . . , ej−1 }
⊥
⊥
⊥
and let Q⊥
j−1 be the projection onto span {e1 , . . . , ej−1 } . Then let ej := Qj−1 fπ(j) /kQj−1 fπ(j) k
where π(j) ∈ {1, . . . , k} \ {π(1), . . . , π(j − 1)} is defined so that kQ⊥
j−1 fπ(j) k is maxiPj
mal. Let Pj := α=1 [0, fπ(j) ]. Now, diam(Pj , eα ) is given by the valueof hx − y, eα i
such that x, y ∈ Pj and hx − y, eα i is maximal. Since fπ(j) ∈
/ span fπ(i) i6=j , we
21
also have that ej ∈
/ span fπ(i)
i6=j
. Thus for i ≥ j, diam(Pi , ej ) ≤ fπ(j) , ej and so
in fact diam(Pi , ej ) = diam(P, ej ) = fπ(j) , ej .
Now we prove by induction that
Pj ⊂
j
X
[0, γj,α eα ]
(53)
α=1
where 0 ≤ γj,α ≤ j diam(Pj , eα ) = j diam(P, eα ). The statement is immediate for
j = 1. Thus let 1 < j ≤ k and assume the induction hypothesis holds for j − 1.
Then for 1 < i ≤ j
ei , fπ(j)
⊥
≤ kQ⊥
i fπ(j) k ≤ kQi fπ(i) k =
ei , fπ(i)
= diam(P, ei )
(54)
where the first inequality is because ei ∈ span {e1 , . . . , ei−1 }⊥ , and the next inequality and equality are by the definition of ei . Also, (54) is immediately verifiable for
i = 1.
Now, we can write
fπ(j) = λj,1 e1 + · · · + λj,j ej
(55)
where |λj,i | ≤ diam(P, ei ) by (54). For any x ∈ Pj = Pj−1 + [0, fπ(j) ], we can write
x=
j−1
X
ηj−1,α eα + ηfπ(j)
(56)
α=1
where 0 ≤ η ≤ 1 and |ηj−1,α | ≤ (j − 1) diam(P, eα ) by the induction hypothesis.
Thus (56) equals
j−1
X
(ηj−1,α + ηλj,α ) eα + ηλj,j ej ,
α=1
and |ηj−1,α + ηλj,α | ≤ (j − 1) diam(P, eα ) + diam(P, eα ) for α ≤ j − 1 and |λj,j | =
diam(P, ej ), so the induction hypothesis is shown.
To state the next lemma we make the following definitions. For a set D ⊂ Rd
and a unit vector v, let
diam(D, v) :=
sup
x,y∈D
x−y∈span{v}
kx − yk.
(57)
Lemma 4.3. Let Assumption 1 hold. Let k ∈ {1, . . . , d}, i ∈ Ik , and j ∈ Jk . Then
for any v ∈ span {vj1 , . . . , vjk },
diam(Gi,j , v) ≤
min
α∈{1,...,k}
d(Gi,j , ∂D, v) ≥
δiα +1
,
|hv, vjα i|
max
α∈{1,...,k}
22
and
δiα
.
|h−v, vjα i|
(58)
(59)
Proof. Fix k ∈ {1, . . . , d}, i ∈ Ik , j ∈ Jk . Fix v ∈ span {vj1 , . . . , vjk } with kvk = 1,
fix α ∈ {1, . . . , k}. Since diam(Gi,j , v) = diam(Gi,j , −v), we restrict attention to v
such that
h−v, vjα i ≥ 0.
(60)
We will upper bound diam(Gi,j , v). Consider x, y ∈ Gi,j such that x − y ∈ span {v}.
In particular, assume without loss of generality that x − y = λv for λ ≥ 0. Since
x, y ∈ Gi,j , hy, vjα i ≤ pjα + δiα +1 and pjα + δiα ≤ hx, vjα i; thus δiα − δiα +1 ≤ pjα +
δiα − hy, vjα i ≤ λ hv, vjα i. Since h−v, vjα i ≥ 0, we have λ ≤ (δiα +1 − δiα )/ h−v, vjα i.
Thus we see diam(Gi,j , −v) = diam(Gi,j , v) ≤ (δiα +1 − δiα )/| hv, vjα i |. This holds
for all α ∈ {1, . . . , k}, so for any ṽ ∈ span {vj1 , . . . , vjk } (where we do not assume
hṽ, vjα i ≥ 0)
δiα +1 − δiα
.
(61)
diam(Gi,j , ṽ) ≤ min
α∈{1,...,k} |hṽ, vjα i|
Next we take v as above and now lower bound d(Gi,j , ∂D, v). Fix α ∈ {1, . . . , k}.
We begin by considering d(Gi,j , Hjα , v). Again, since d(Gi,j , ∂D, v) = d(Gi,j , ∂D, −v),
we can and do assume (60) holds. Fix x ∈ Gi,j . Consider λ > 0 such that
x + λv ∈ Hjα . Then λ hv, vjα i = pjα − hx, vjα i ≤ −δiα since hx, vjα i ≥ δiα + pjα , and
so λ ≥ δiα / h−v, vjα i. This shows for any β ∈ {1, . . . , k} that
d(Gi,j , ∪kα=1 Fjα , v) ≥
δiα
.
α∈{1,...,k} |hv, vjα i|
min
(62)
To complete the proof, note for j ∈ {1, . . . , N } \ j, that
d(x, Fj , v) ≥ u = u min |hv, vjα i|−1
α=1,...,k
which is larger than the right hand side of (62).
Appendix
Theorem 4.1 (John’s theorem, John (1948); Theorem 13.4.1 Matoušek (2002)).
Let K ⊂ Rd be a bounded closed convex body with nonempty interior. Then there
exists an ellipsoid E of maximal volume such that E ⊆ K ⊆ nE.
List of Notation
N[ ] (ǫ, F, d)
Lp
C(·, . . . , ·)
k·k
D
∂; ∂r
bracketing number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
R p
1/p
if 1 ≤ p < ∞ or supx |f (x)| if p = ∞ . . . 2
Lp (f ) is
f (x)dx
convex function classes . . . . .P
....................................2
Euclidean distance: kzk2 := di=1 zi2 for z ∈ Rd . . . . . . . . . . . . . . . . . 4
d-dimensional convex set; d-dimensional convex polytope . . . . 4, 8
boundary; relative boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
23
d+ (x, G, v)
d(x, H, v)
d(E, H, v)
N
Ei
Hi
Fi
vi
Gj
Gi,j
Aj , xj
γj
δi , A, u
Jk
Ik
Ej ; e ≡ ei,j
h(D, x)
w(D)
................................................................ 8
................................................................ 8
................................................................ 8
number of halfspaces defining D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8
halfspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
hyperplanes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
facets of convex polytope D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
(inner) normal vectors for Hi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8
faces of polytope D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
................................................................ 8
................................................................ 9
................................................................ 9
................................................................ 9
................................................................ 8
................................................................ 9
See Proposition 3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9, 10
support function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
width . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
24
References
Ball, K. (1992). Ellipsoids of maximal volume in convex bodies. Geometriae
Dedicata, 41 241–250.
Ball, K. (1997). An elementary introduction to modern convex geometry. In
Flavors of Geometry (S. Levy, ed.), vol. 31. Cambridge University Press, 1–58.
Birgé, L. and Massart, P. (1993). Rates of convergence for minimum contrast
estimators. Probability Theory and Related Fields, 97 113–150.
Brøndsted, A. (1983). An Introduction to Convex Polytopes, vol. 90 of Graduate
Texts in Mathematics. Springer-Verlag, New York-Berlin.
Bronshtein, E. M. (1976). ǫ-entropy of convex sets and functions. Siberian
Mathematical Journal, 17 393–398.
Dey, T. K. and Pach, J. (1998). Extremal problems for geometric hypergraphs.
Discrete & Computational Geometry, 19 473–484.
Doss, C. R. and Wellner, J. A. (2016a). Global rates of convergence of the MLEs
of log-concave and s-concave densities. The Annals of Statistics, 44 954–981.
Doss, C. R. and Wellner, J. A. (2016b). Inference for the mode of a log-concave
density. arXiv:1611.10348.
Doss, C. R. and Wellner, J. A. (2016c). Mode-constrained estimation of a
log-concave density. arxiv:1611.10335v2.
Dryanov, D. (2009). Kolmogorov entropy for classes of convex functions. Constructive Approximation, 30 137–153.
Dudley, R. M. (1978). Central limit theorems for empirical measures. The Annals
of Probability, 6 899–929.
Dudley, R. M. (1984). A course on empirical processes. In École d’été de probabilités de Saint-Flour, XII—1982, vol. 1097 of Lecture Notes in Math. Springer,
Berlin, 1–142.
Dudley, R. M. (1999). Uniform Central Limit Theorems, vol. 63 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge.
Gao, F. and Wellner, J. A. (2017). Entropy of convex functions on Rd . Constructive Approximation, 65 181–28.
Grünbaum, B. (1967). Convex Polytopes. With the cooperation of Victor Klee,
M. A. Perles and G. C. Shephard. Pure and Applied Mathematics, Vol. 16, Interscience Publishers John Wiley & Sons, Inc., New York.
25
Guntuboyina, A. and Sen, B. (2013). Covering numbers for convex functions.
IEEE Transactions on Information Theory, 59 1957–1965.
Guntuboyina, A. and Sen, B. (2015). Global risk bounds and adaptation in
univariate convex regression. Probability Theory and Related Fields, 163 379–
411.
John, F. (1948). Extremum problems with inequalities as subsidiary conditions.
In Studies and Essays Presented to R. Courant on his 60th Birthday, January 8,
1948. Interscience Publishers, Inc., New York, N. Y., 187–204.
Kim, A. K. H. and Samworth, R. J. (2016). Global rates of convergence in
log-concave density estimation. The Annals of Statistics, 44 2756–2779.
Koenker, R. and Mizera, I. (2010). Quasi-concave density estimation. The
Annals of Statistics, 38 2998–3027.
Kolmogorov, A. N. and Tihomirov, V. M. (1961). ε-entropy and ε-capacity of
sets in function spaces. AMS Translations: Series 2, 17 277–364.
Lee, C. W. (1997). Subdivisions and triangulations of polytopes. In Handbook of
discrete and computational geometry. CRC, Boca Raton, FL, 271–290.
Matoušek, J. (2002). Lectures on discrete geometry, vol. 212 of Graduate Texts in
Mathematics. Springer-Verlag, New York, New York, NY.
Rockafellar, R. T. (1970). Convex Analysis. Princeton Landmarks in Mathematics, Princeton University Press, Princeton, NJ.
Rothschild, B. L. and Straus, E. G. (1985). On triangulations of the convex
hull of n points. Combinatorica, 5 167–179.
Rudin, W. (1987). Real and Complex Analysis. 3rd ed. McGraw-Hill Book Co.,
New York.
Seijo, E. and Sen, B. (2011). Nonparametric least squares estimation of a multivariate convex regression function. The Annals of Statistics, 39 1633–1657.
Seregin, A. and Wellner, J. A. (2010). Nonparametric estimation of multivariate convex-transformed densities. The Annals of Statistics, 38 3751–3781. With
supplementary material available online.
van de Geer, S. A. (2000). Empirical Processes in M-Estimation. Cambridge
University Press.
van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and
Empirical Processes. Springer Series in Statistics, Springer-Verlag, New York.
Xu, M., Chen, M. and Lafferty, J. (2016). Faithful variable screening for highdimensional convex regression. The Annals of Statistics, 44 2624–2660.
26
| 10math.ST
|
1
An Evolutionary Game for User Access Mode
Selection in Fog Radio Access Networks
arXiv:1709.05288v1 [cs.IT] 14 Sep 2017
Shi Yan, Mugen Peng, Senior Member, IEEE, Munzali Ahmed Abana, and Wenbo Wang
Key Laboratory of Universal Wireless Communication, Ministry of Education
Beijing University of Posts and Telecommunications, Beijing, 100876, China
Abstract—The fog radio access network (F-RAN) is a promising paradigm for the fifth generation wireless communication
systems to provide high spectral efficiency and energy efficiency.
Characterizing users to select an appropriate communication
mode among fog access point (F-AP), and device-to-device
(D2D) in F-RANs is critical for performance optimization. Using
evolutionary game theory, we investigate the dynamics of user
access mode selection in F-RANs. Specifically, the competition
among groups of potential users space is formulated as a dynamic
evolutionary game, and the evolutionary equilibrium is the
solution to this game. Stochastic geometry tool is used to derive
the proposals’ payoff expressions for both F-AP and D2D users
by taking into account the different nodes locations, cache sizes
as well as the delay cost. The analytical results obtained from
the game model are evaluated via simulations, which show that
the evolutionary game based access mode selection algorithm can
reach a much higher payoff than the max rate based algorithm.
I. INTRODUCTION
To achieve the goals of the fifth generation (5G) mobile
wireless systems [1] and alleviate the existing challenges in
cloud radio access networks (C-RANs) [2]∼ [4], the fog radio
access network (F-RAN) has been proposed as a new network
architecture by incorporating of fog computing, edge storage
and centralized cloud computing into radio access networks
[5] [6]. Fog computing, which is similar to edge computing,
is first proposed by Cisco [7]. It extends cloud computing
based services to the edge of the network. In F-RANs, services
cannot only be executed in a centralized unit such as the
base band unit (BBU) pool in C-RAN, but also can be
hosted at smart terminal devices which are closer to the users.
Meanwhile, through the user-centric adaptive techniques such
as device-to-device (D2D), distributed coordination, and largescale centralized cooperation, users do not have to connect
to the centralized cloud computing unit to complete the data
transmission, which will improve both spectrum and energy
efficiencies as well as relieve the load of fronthaul and alleviate
the burden of BBU pool. In order to execute the above, the
traditional access point (AP) is evolved to the fog access point
(F-AP) through equipped with a certain caching and sufficient
computing capabilities to execute the local cooperative signal
processing in the physical layer.
Extensive studies have been done to study the user access
mode schemes in C-RANs and heterogeneous cloud radio
access networks (H-CRANs). The performance of uplink
transmissions under different user access modes in C-RANs
have been investigated in [8], and the impact of the remote
radio heads (RRHs) intensity, the number of antennas and
the pathloss exponent on the performance is characterized.
By using the stochastic geometry tool, the authors in [9]
have provided the probabilistic characterization of the signal
interference ratio (SIR) distribution and the ergordic rate at a
randomly located user in a downlink H-CRAN system. In [10],
the successful access probability has been derived for C-RAN
RRH cluster mode, and based on the analytical result, two
coalition game based cluster algorithms have been designed.
And in our previous study [11], the ergodic rates of two tier
F-RANs under different access modes are characterized, and
a mode selection mechanism is presented.
However, most of literatures focused on the mode selection
schemes based on the user received SIR and distance, whereas
little attention has been paid to study the impact of cache
and fronthaul delay. The fronthaul limitations between RRHs
and cloud are known to impose a formidable bottleneck
to the system performances and a remarkable challenge to
block the commercial practices [12]. Consequently, taking
advantage of the fog-computing to switch the content cache
or data process to network edge devices is the key to increase
both of the spectral and energy efficiency, as well as relieve
the burden of fronthaul in cloud computing based network
architectures. D2D communications as one of the network
edge communication technologies which has the ability of
support content delivery and sharing is more and more paid
extensive attention. Existing research on D2D networks mode
selection mainly focused on underlaid cellular networks with
fixed location model [13] ∼ [15]. Different from the traditional
underlaid cellular networks, the main challenge to design a
user access mode selection algorithm in F-RANs is that the
F-APs and D2D users are often deployed randomly and the
network performance of F-RANs may drastically deteriorate
with the increasing number of users who select the access
mode with more fronthaul load.
In this paper, we present an adaptive user access mode selection algorithm for downlink F-RAN by taking into account the
different nodes locations, cache sizes as well as the fronthaul
delay cost. The main contributions of this paper are threefolds.
• We characterize the ergodic rate and the coverage probability of both D2D and F-AP modes in F-RANs. We
investigate the impacts of the user node density, and
the quality of service (QoS) constrains and the intra-tier
interference. The closed-form expressions are presented
in some special cases, which can make the analysis not
only tractable, but also flexible.
2
Based on the above results obtained from the stochastic
geometry analysis, an evolutionary game theoretic approach is presented to solve the problem of user access
mode selection, while taking into account different access
service requirements and fronthaul delay cost.
• A distributed algorithm is proposed to reach the evolutionary equilibrium, and its performance is compared with
that of the max rate based user access mode selection
scheme. The accuracy of the stochastic geometry analytical results are validated by Monte Carlo simulation
results.
The remainder of this paper is organized as follows. Section
II describes the system model, cache model, delay model
as well as the user access modes in a downlink F-RAN. In
Section III, the dynamic behaviors of potential F-AP users
are modeled as an evolutionary game. And we present the
results of the coverage probability and ergodic rate in F-RAN
system for both D2D and F-AP modes. While in Section IV,
we propose a distributed protocol to converge the evolutionary
equilibrium, and the evolutionary stable strategy (ESS) is
analyzed. Simulation results provided in Section V prove that
the performance comparisons of different user access modes
with different system parameters and the robustness of our
protocol. Finally, Section VI concludes the paper.
•
II. SYSTEM MODEL
A. Radio Access Network Model
As shown in Fig.1, a F-RAN downlink system is considered
in this paper, where a group of F-APs are deployed according
to a two-dimensional PPP Φf with density of λf in a disc
plane D2 . Each F-AP is assumed as single antenna configuration with a fixed transmit power Pf , and interfaced to
the cloud computing layer through a wired fronthaul network
which is composed of links connecting F-APs with several
gateways. Then, the signals over gateways will be large-scale
processed in the cloud network server via dedicated fiber
[16]. The locations of gateways are distributed according to
a homogeneous PPP Φg with intensity λg .
The spatial distribution of users is modeled as an independent PPP Φu distribution with constant intensity λu . We
set p ∼ (0, 1] as the probability that a user can direct
transmit its local cache content to its nearby content require
user (i.e.support D2D mode). Then, the location distribution
of the D2D transmit users can be denoted as a thinning
homogeneous Φtu with the density of λtu = pλu . Meanwhile
according to Marking Theorem, the distribution of content
require users follows a stationary PPP Φru with the density
of λru = (1 − p)λu .
In this paper, we focus on the user-centric access mode
selection for content require users via the above F-RAN
system within two phases, namely cache pre-fetching and
access mode selection. The total number of content require
users can be estimated as Nru = A(D2 )(1 − p)λu , where
A(D2 ) is the area of the disc plane D2 .
B. Cache Pre-Fetching Phase
In cache pre-fetching phase, the F-APs will cache content
from the central cloud server file library which stored N
Core Network
Backhaul Link
Cloud Cache
Fronthaul Link
Signal Link
Gateways
...
F-APs with Local Cache
x Df Content Require
User
Ld
x Ff
Ld
F-AP Mode
D2D
Transmit User
D2D
Transmit User
D2D Mode
xdD
Fig. 1. Downlink transmission scenario
contents with uncoded cache strategies that may be requested
by users V = {v1 , v2 , ..., vN }. Then, the D2D transmit users
will cache part of the content from F-APs via wireless link.
The pre-fetching content is limited by the local caching storage
space size of F-APs and D2D transmit users, which denote by
Cf and Cd (Cd < Cf < N ), respectively, and the pre-fetching
content in Cf and Cd is assumed to be constant across access
mode selection phase.
Prior studies have been found that users are mostly interested in downloading the most popular video contents [17],
which leads to only a small portion of the N contents are
frequently accessed by the majority of users. Therefore, in
this paper we assume the F-APs and D2D transmit users only
cache the most popularity content, and the demand probability
can be modeled as the following Zipf distribution
1/iσ
,
fi (σ, N ) = PN
σ
k=1 1/k
(1)
PN
where
i=1 fi (σ, N ) = 1, and the video content with a
smaller index has a larger probability of being requested by
users, i.e. fi (σ, N ) > fj (σ, N ), if i < j. Zipf exponent σ > 0
controls the relative popularity of files, and with the larger
σ the caching storage has a fewer of popular video contents
accounting for the majority of the requests.
Next, we define the content caching probability as the
probability of an event that a user U can find the requested
file V in its accessed node, i.e., pxc = Pr(V ∈ Cx ). By setting
the caching storages in F-APs and D2D transmit users to only
cache the most popularly requested video contents, the content
caching probabilities of each F-AP and D2D transmit user can
be respectively obtained as
pF
c = Pr(V ∈ Cf ) =
pD
c = Pr(V ∈ Cd ) =
XC f
i=1
XCd
i=1
fi (σ, N ),
(2)
fi (σ, N ).
(3)
3
C. Access Mode Selection Phase
In access mode selection phase, a content require user can
select a communication mode independently according to the
perceived performance, distance, local cache and the video
streaming traffic delay cost. Let U → X signify that a desired
content require user U is associated to a node located at
X, and kXk denotes the distance between U and X. And
we can categorize the content require users into two types:
potential D2D mode users and potential F-AP mode users, the
definitions of which are given as follows:
• Potential D2D mode Users: A content require user is a
potential D2D mode user if it can successfully obtain the
requested contents from another D2D transmit user in a
known location within a distance threshold Ld .
• Potential F-AP mode Users: A content require user is a
potential F-AP mode user if the user cannot find a D2D
transmit user within Ld , or the content require user can
find a D2D transmit user near it but the requested content
V is not cached in that D2D transmit user. Thus, U will
try to access its nearest F-AP node.
It’s worth noting that all potential D2D mode users can
select both the D2D mode and F-AP mode, while each of the
potential F-AP users only can select the F-AP mode to access
its nearest F-AP. More specifically, if a potential D2D mode
user selects D2D mode, a communication link is established
between the tagged content require user and its service D2D
transmit user which has a known location. On the other hand,
when F-AP mode is selected and the requested content V is
cached in the F-AP node, the potential D2D mode user or
potential F-AP user can download data directly from F-AP. If
user cannot find the requested content in its nearby F-AP, then
user will download its data through the fronthaul in the same
way as traditional C-RANs.
Let N (D) = Nru pd and N (F ) = Nru (1 − pd ) denote
the number of potential D2D mode and F-AP mode users,
respectively, where pd denotes the probability of at least one
D2D transmit user located in distance threshold Ld meanwhile
has the requested content V , which can be given as
2
pd = 1 − exp(−πλtu pD
c Ld ).
(4)
Proof. By using the property of 2-D Poisson process, the
probability distribution of the nodes number m in a circle
area πl2 with radius limit l can be derived as
m
2
λX πl2 e−λX πl
2
.
(5)
Pr Φ πl = m =
(m)!
Let l = Ld , λX = pD
c λtu and m = 0. Then, we have the
probability of none D2D transmit user has the the requested
video content V within the distance limit Ld . Therefore, (4)
can be given as the probability of complementary events.
During the access mode selection phase, we consider two
kinds of delay, cache processing delay and average F-AP
fronthaul packet delay. The cache processing delay mainly
comes from the cache data processing time, and in this work
we assume that the processing delay in D2D transmit users
and F-APs as constants δd and δf , respectively.
On the other hand, the F-AP fronthaul packet delay is
defined as the time needed for a packet to be transmitted from
fronthaul to a F-AP. It depends on the size of the packet and
the number of F-APs associated with the gateway [19]. In
wired scenario, Gamma distribution is usually used to model
the distribution of the delay in routers or switches [20] [21].
Then, the fronthual packet delay is expressed as
Tf ∼ Gamma(
λf
k, a + bµ),
λg
(6)
where the first and second terms of Gamma function represent
the effect of number of connecting nodes and packet size on
delay, respectively. a, µ and k are constants, representing the
processing capability of the nodes, and b is the packet size.
Therefore, we can directly obtain the mean F-AP fronthaul
packet delay from the expectation of Gamma distribution as
Tf =
λf
k(a + bµ).
λg
(7)
And the total delay φi in access mode i can be expressed
as follows
φd = pD
c δd ,
(8)
F
φf = pF
c δf + (1 − pc )Tf .
III. EVOLUTIONARY GAME APPROACH
ACCESS MODE SELECTION
FOR
USER
In this section, the evolutionary game formulation for user
access mode selection is given, and the replicator dynamics is
used to model the strategy adaptation process.
A. Formulation of the Evolutionary Game
We define the evolutionary game as follows:
• Players: Each user who can choose among multiple
access mode is a player of the game. In this paper, the
players are all of the potential D2D mode users. Note
that the potential F-AP mode users are not involved in the
game since F-AP mode is the only access mode available
to these users.
• Strategies: The set of strategies is the selection of a communication mode. The set of strategies for the potential
D2D mode players is D = {D, F }, and there is obvious
only one strategy possible for the potential F-AP players
F = {F }.
• Payoff: The payoff quantifies the performance satisfaction level of a potential player which depends on the
ergodic rate rate as well as the delay cost. The payoff
function for a (n) mode potential player selected access
mode i is defined as follows:
(n)
πi (x) =
=
p i Bi
(a)
Σa∈A(n) ni
p i Bi
(n)
Ci
− qi φi ,
(n)
C
(a) i
Σa∈A(n) N (a) xi
(9)
− qi φi ,
where x denotes the vector of the proportion of users choosing
(n)
denotes the utility function
different access modes. Ci
measuring the achieved performance in (n) mode choosing
4
mode i, which will be further discussed in Section IV. Bi
(a)
(a)
denotes the bandwidth in mode i. xi = ni /N (a) is the
proportion of individuals of potential users in (a) choosing
(a)
strategy i. ni is the number of potential users in (a) mode
choosing mode i, A(n) is the set of subareas in mode n,
for the potential D2D mode users, this set can be defined as
A(n) = {D, F } since both D2D and F-AP modes are available
for these users. pi is the coefficient of linear pricing function
used by mode i to charge a user.
Moreover, since the F-AP fronthaul bandwidth is limited, it
will lead to negative impact on their payoff if too many users
select F-AP mode to receive their data through fronthaul link.
On the other hand, the processing delay cost of both of D2D
and F-AP modes will also increase with the growth of the
users at the same time. In order to evaluate these impacts,
we define the cost price coefficient qd and qf to reflect the
influence of the proportion of mode selection.
h
i
qd = c1 exp(c2 N (D) x(D) ) − 1 ,
d
i
h
qf = c3 exp c4 N (F ) + N (D) x(D) − 1 ,
f
(10)
where c1 , c2 , c3 and c4 are positive constants.
In order to measure the users’ performance utility functions
presented in the evolutionary game formulation (9), in the
next subsection, we use stochastic geometry tool to derive the
coverage probability and ergodic rate for both D2D and F-AP
modes by taking into account the different nodes locations,
the SIR QoS constrains and the interference.
Next, if U is served by a single nearest F-AP in F-AP mode,
the SIR is given by
In this subsection, we derive the coverage probability
and ergodic rate for F-RAN with two different user access
modes. The ergodic rate in F-RAN is defined as Ri =
E[ln (1 + SIR(U → Xi )) |SIR(U→ Xi ) > Ti ], where the unit
of the ergodic rate is in terms of nats/s/Hz, Ti is the QoS
constraint of access mode i.
In this paper, we limit our attention to the interferencelimited scenario since the interference is much larger than the
noise. Path loss is represented by a standard power law path
−α
loss function kXk , where α > 2 is the path loss exponent.
Then, if U is served by a D2D transmit user at a fixed
distance of kXf k to U , the received SIR at the desired user
is given by
SIR(U → Xd ) = γd =
Pd hd kXd k−α
,
Id,ru + If,ru
(11)
where hd characterizes the flat Rayleigh channel fading between the potential D2D mode user and its associated D2D
−α
transmit user, and kXd k
denotes the path loss. Id,ru =
P
−α
P
g
r
is
the
intra-tier
interference from other D2D
i∈Φtu d i i
transmit users, gi ∼ exp(1) and ri−α denote the exponentially
distributed fading power over the Rayleigh fading channel and
path lossP
from other D2D transmit users to U , respectively.
Id,ru = j∈Φf Pf gj lj−α denotes inter-tier interference from
F-APs, the definition of gj and lj−α are similar to that in Id,ru .
(12)
P
′
′ −α
, and Id,f u
=
If,f u
=
j∈Φf /f Pf gj lj
−α
denote the intra-tier interference from
j ′ ∈Φtu Pd gj ′ rj ′
other F-APs and inter-tier interference from D2D transmit
users, respectively, and the definitions of hf , gi′ and li′ are
given in (11).
1.D2D mode
The coverage probability of a user in D2D mode can be
denoted as
where
P
!
−α
Pd hd kXd k
Pd (Td , α, kXd k) = Pr
≥ Td
Itu,ru + If,du
!
α2 !
2
Pf
2
D
α
= exp −πkXd k pc λtu +
λf C (α) Td ,
Pd
(13)
where C (α) = 2π csc (2π/α)/α.
Proof. See Appendix A.
Similarly, the achievable ergodic rate can be derived as
Rd = E [ln (1 + γd )|γd ≥ Td ]
(a)
α
≈ ln(Td )Pd (Td , α kXd k) −
2
"
2
α
B. Performance Analysis for Different User Access Modes
Pf hf kXf k−α
,
Id,f u + If,f u
SIR(U → Xf ) = γf =
·Ei −Td πkXd k
2
pD
c λtu
+
Pf
Pd
2
λf
!
#
(14)
C (α) ,
where (a) follows inR the high SIR conditions ln (1 + γd ) →
∞
ln (γd ), Ei[s] = − −s e−t /tdt is the exponential integral
function.
Proof. See Appendix B.
2.F-AP mode
In F-AP mode, the desired user will try to access its
nearest F-AP Xf to download the requested content V . The
probability density function (PDF) of the distance between Xf
and U can be derived by using a similar way as (4)
∂ (1 − Pr (No F − AP closer than rm ))
fkXf k (rf ) =
∂rf
2
∂ 1 − exp −πλf rf
2
= 2πλf rf e−πλf rf .
=
∂rf
(15)
Thus, the coverage probability of F-AP mode can be calculated as
1
Pf (Tf , α) =
2/α , (16)
Pd Tf
C
(α)
1 + ρ (Tf , α) + λλtu
Pf
f
where ρ (Tf , α) =
R∞
−2
Tf α
Proof. See Appendix C.
2/α
Tf
dv.
1+v α/2
5
And the ergodic rate for F-AP mode can be given as
Rf = E [ln (1 + γf )|γf ≥ Tf ]
Z ∞
≈
Pf (eθ , α)dθ + ln(Tf )Pf (Tf , α).
IV. EVOLUTIONARY EQUILIBRIUM
STABILITY ANALYSIS
(17)
ln(Tf )
Special Case: Path loss exponent for F-AP to user link is
4 (α = 4), and SIR threshold Tf > 1
An approximated closed-form approximate expression can
be derived in this special case, and we give the ergodic rate
in the following lemma 1.
Lemma 1. The ergodic rate for nearest F-AP mode with α=
4 and Tf > 1 can be expressed as
Rf α=4 = E [ln (1 + γf )|γf ≥ Tf ]
4
2 ln(Tf )
q + p
q
≈ p
Pd
Pd
λtu
π Tf 1 + λf
π Tf 1 + λλtu
(18)
Pf
Pf
f
2(2 + ln(Tf ))
q .
= p
Pd
π Tf 1 + λλtu
Pf
f
Proof. See Appendix D.
C. Replicator Dynamics
Evolutionary game theory differs from classical game theory
by focusing more on the dynamics of strategy change as
influenced not solely by the quality of the various competing
strategies, but by the effect of the frequency with which those
various competing strategies are found in the population [25]
[26]. In a dynamic evolutionary game, the replicator dynamic
is a simple model of evolution and prestige-biased learning in
games [27].
We consider an evolutionary game of user access mode
selection where the group of potential users in (a) can choose
among the available wireless access modes i. The game is
repeated, and the user adopts a strategy that gives a higher
payoff which is determined by the payoff matrix and the
proportion of each strategy in the population. The speed of
the user in observing and adapting the access mode selection
is controlled by parameter ρ > 0.
For a small period of time, the rate of strategy is governed
by the replicator dynamics, which is defined as follows:
(n)
(n)
(n)
πi (x) − π (n) (x) ,
ẋi = ρxi
(19)
where π (n) (x) is the average payoff of the users in potential
group (n) which is computed as
AND
In this section, the evolutionary equilibrium is considered to
be the solution of the formulated access mode selection game
and we also analyzed its stability. An evolutionary equilibrium
is a fixed point where there is no change in proportion of
players choosing different strategies. In other words, since the
(n)
rate of strategy adaptation is zero (i.e., ẋi = 0), there is no
user who deviates to gain a higher payoff [28].
(n)∗
, which is
To evaluate the stability at the fixed point ẋi
(n)
obtained by solving ẋi = 0, the eigenvalues of the Jacobian
matrix corresponding to the replicator dynamics need to be
evaluated. The fixed point is stable if all eigenvalues have a
negative real part [30].
For the mode selection game in this paper, the replicator
dynamics can be expressed as follows:
(D)
(D) (D)
(D) (D)
πf (x) − xf πf (x) − xd πd (x)
(a)
(D)
(D)
(D)
(D)
= σx(D)
1
−
x
π
(x)
−
1
−
x
π
(x)
f
f
f
f
d
(D)
(D)
(D)
(D)
= σxf (1 − xf ) πf (x) − πd (x) ,
(21)
(D)
(D)
where (a) follows that xf + xd = 1.
Hence, we could derive the evolution algorithm as described
in Algorithm 1. Also, the stability of the equilibrium is
clarified in Theorem 2.
(D)
ẋf
(D)
= σxf
Algorithm 1 Evolution Algorithm
1: Initialize Each user selects to get access to a access mode in
random.
(n)
2: Step 1 The payoff of users accessing i mode πi
is calculated
according to (9). This payoff information is sent to the F-APs.
3: Step 2 The average payoff of the content require users π (n) is
calculated according to (20) in F-APs, and is broadcast to all
potential D2D mode users.
4: Step 3 For a user accessing mode i, if its payoff is less than the
(n)
average payoff, i.e., πi < π(n) , it would randomly switch to
(n)
(n)
another access mode j, where j 6= i and πj > πi
5: Repeat from Step 1 to Step 3 until convergence.
Theorem 2. For a potential D2D mode user with a stable
equilibrium point, the interior evolutionary equilibrium in
game is asymptotically stable.
Proof. See Appendix E.
(20)
A. Comparison with Max Rate Based Access Mode Selection
Algorithm
Based on this replicator dynamics of the users in potential group (n), the number of users choosing access mode
i increases if their payoff is above the average payoff. It
is impossible for a user to choose access mode k, which
provides a lower payoff than the current payoff. This replicator
(n)
dynamics satisfies the condition of Σi ẋi = 1. In the next
section, we will use above expressions to analyze the stability
of the access mode selection game.
For comparison purpose, we consider a max rate based
access mode selection algorithm under the same system model.
Max rate based access mode selection algorithm needs to
sort both the D2D mode and F-AP mode rate list based on
QuickSort algorithms to have ordered FAP and MBS rate lists,
and then control each user to access to its node with max rate.
We derive the max rate based access mode selectio algorithm
as described in Algorithm 2.
(n) (n)
π (n) (x) = Σi xi πi (x).
6
Algorithm 2 Max Rate Based User Access Mode Selection
Algorithm
1: Initialize Each user selects to get access to a access mode in
TABLE I
SIMULATION PARAMETERS
Parameters
Value
Description
2: for u = 1, 2, 3, ..., Nu do.
3: Step 1 Find the max rate D2D transmit user with the required
N
1000
Number of video contents
content Vu of the desired user Uu within a radius threshold Lf ,
B(Uu , Lf ) ∩ Φtu = {X1 , X2 , ..., XD }.
4: if B(Uu , Lf ) ∩ Φtu = φ.
5:
Rdmax = 0, go to Step 2.
6: end if
7: Calculate the D2D X1 rate RX1 from (14).
8: Set Rdmax = RX1 .
9:
for i = 2, 3, ..., D do.
10:
if RXi > Rdmax .
11:
Max Rate D2D transmit user Rdmax = RXi .
12:
end if
13:
end for
14: Step 2 Find the max rate F-AP of the desired user Uu in Φf =
{Y1 , Y2 , ..., YF }.
15: Calculate the F-AP Y1 rate RY1 from (17).
16: Set Rfmax = RY1 .
17:
for j = 2, 3, ..., F do.
18:
if RYj > Rfmax .
19:
Max Rate F-AP node Rfmax = RYj .
20:
end if
21:
end for
22: Step 3
23:
if Rdmax > Rfmax .
24:
User Uu select D2D mode.
25:
else
26:
User Uu select F-AP mode.
27:
end if
28: end for
Ld
30m
D2D distance threshold
Cd
80
Cache size of D2D transmit users
Cf
100 ∼ 500
Cache size of F-APs
V. NUMERICAL RESULTS
In this section, the accuracy of the above ergodic rate
expressions and the performance of the proposed algorithm
are evaluated by using Matlab with Monte Carlo simulation
method. The simulation parameters are listed as follows in
Table I, and the simulation network topology is shown in Fig.
3.
Bd
300MHz
Bandwidth of D2D users
Bf
100MHz
Bandwidth of F-APs
Pd
13dBm
Transmit power of D2D transmit user
Pf
23dBm
Transmit power of F-AP
λru
λtu
λf
λg
α
6 × 10
−5
Intensity of content require users
5 × 10
−5
Intensity of D2D transmit users
3 × 10
−5
Intensity of F-APs
2 × 10
−6
Intensity of gateways
4 [29]
Path loss exponent
1.8
Simulation Rf
1.7
Analytical Rf
Simulation Rd
1.6
Analytical Rd
1.5
Ergodic Rate (bps/Hz)
random.
1.4
1.3
1.2
1.1
1
0.9
0.8
2
3
4
5
6
7
SIR Threshold (dB)
8
9
10
Fig. 3. Ergodic rate of D2D mode and F-AP mode versus different SIR
thresholds.
Fig. 2. Poisson distributed content require users, D2D transmit users and
F-APs. The F-APs coverage boundaries are shown and form a Voronoi
tessellation.
Fig. 4 shows the ergodic rate achieved by the D2D mode and
F-AP mode in F-RAN with the varying SIR QoS thresholds
T . The analytical results closely match with the corresponding
simulation results, which validates our analysis in Section III.
It can be observed that the ergodic rate of both D2D and FAP user access mode decreases as the SIR threshold increases.
This is because the larger T suggests that the user is more strict
in the quality of SIR, which leads to ergodic rate decreasing.
The proportion and the achieved average payoff per user accessing different mode during the evolution process are shown
in Fig. 5 and Fig. 6. It can be seen that the proposed user
access mode selection algorithm can reach the evolutionary
equilibrium very fast, within less than 15 iterations. More
specifically, as can be seen from Fig. 5, we could find that
after several iterations all curves with different proportion at
7
evolutionary game based access mode selection algorithm can
reach a much higher payoff than the max rate based algorithm.
8
7
Average Payoff Per User
6
5
4
Max Rate Access
(0)=0.75
x(D)
d
3
x(D)
(0)=0.5
d
x(D)
(0)=0.25
d
2
1
0
5
10
15
20
25
30
Iteration
35
40
45
50
Fig. 4. Average payoff per user with the user access mode selection algorithm.
1
0.9
Proportion of Each Population
0.8
A PPENDIX
A. Proof of Pd (Td , α, kXd k)
According to the definition of coverage probability, we have
!
Pd hd kXd k−α
≥ Td
Pd (Td , α, kXd k) = Pr
Id,ru + If,ru
α
Td kXd k
(Id,ru + If,ru )
= Pr hd ≥
Pd
α
(a)
Td kXd k
= E exp −
(Id,ru + If,ru )
Pd
(b)
Td kXd kα
Td kXd kα
= LId,ru
LId,ru
Pd
Pd
!
α2 !
2
Pf
2
D
α
= exp −πkXd k pc λtu +
λf C (α) Td ,
Pd
(22)
where (a) follows from the Laplace transform of hd ∼ exp(1)
and the independence of Id,ru and If,ru [22] [23]. (b) follows
from letting s = Td kXd kα /Pd in the Laplace transforms of
Id,ru and If,ru , which are given by
0.7
x(D)
(0)=0.75
d
0.6
0.5
x(D)
(0)=0.25
f
(D)
xd (0)=0.5
0.4
x(D)
(0)=0.5
f
LId,ru (s) = EId,ru
Y
LIf,ru (s) = EIf,ru
Y
j∈Φf
0.2
−sPd gi ri−α
#
.
(23)
exp −sPf gj lj−α .
(24)
exp
i∈Φtu
0.3
Using the independence of the fading random variables in
LId,ru (s) ,we have
0.1
0
"
0
5
10
15
20
25
Iteration
30
35
40
45
50
Fig. 5. Proportion of user with the user access mode selection algorithm.
the initial point at last achieve a much better average payoff
than only consider max rate based user access mode selection
algorithm.
On the other hand, the trajectories of the proportion of users
selecting each access mode are illustrated in Fig. 6. It can
be observed that the proportions eventually converge to the
equilibrium after changing in different directions. At the same
time, those curves about payoff finally converge to the network
average payoff.
VI. CONCLUSION
In this paper, an evolutionary game theory based algorithm
to solve the problem of user access mode selection in downlink
F-RAN is proposed. Stochastic geometry is used to derive the
payoff expressions by taking into account the different nodes
locations, interferes as well as the delay cost. The evolutionary
equilibrium has been considered to be the stable solution
for which all users receive identical payoff from accessing
different access modes. The simulation results show that the
LId,ru (s) = EId,ru
"
"
Y
i∈Φtu
exp
−sPd gi ri−α
#
#
1
=
1 + sPd ri−α
i∈Φtu
Z
(b)
1
= exp −pD
dr
1
−
λ
i
c tu
1 + sPd ri−α
r2
Z ∞ Z ∞
2
(c)
(−t(1+r α ))
α
= exp −2πpD
λ
(sP
)
r
e
dtdr
,
tu
d
c
0
0
(d)
2/α
= exp −s2/α C (α) pD
,
c λtu Pd
(25)
where (a) follows gi ∼ exp(1), (b) follows from probability
generating functional (PGFL) of PPP [24] and, (c) results from
algebraic manipulation after converting from Cartesian to polar
coordinates and (d) follows from using some properties of
Gamma function, and LIf,ru (s) can be obtained in a similar
way.
(a)
Y
B. Proof of Rd
For a positive continuous random variable A, we can use
the following formula for computing its expectation
8
D. Proof of lemma 1
E [A |A ≥ W ]
Z ∞
Z ∞Z t
=
tfA (t)dt =
fA (t)dadt
W
W
0
Z WZ ∞
Z ∞Z ∞
=
fA (t)dtda +
fA (t)dtda
0
W
Z ∞ W a
= W Pr (A ≥ W ) +
Pr (A ≥ a)da .
W
{z
}
|
we first drive the coverage probability of F-AP mode with
α = 4 and Tf > 1, which can be denoted as
Pfα=4 (Tf ) =
(26)
=
(a)
S
≈
Next, we focus on the second term of (26), after changing
variables with W = ln(Td ), A = ln(1 + γd ) and a = θd the
expression of this term can be given as
S=
Z∞
Pr
ln(Td )
Z∞
≈
=
ln(Td )
Z∞
ln(Td )
LId,ru
!
Pd hd kXf k−α
θd
> e − 1 dθd
Id,ru + If,ru
α
eθd kXf k
Pd
LIf,ru
α
eθd kXd k
Pd
dθd
C. Proof of Pf (Tf , α)
According to the definition of coverage probability, we have
!
−α
Pf hf kXf k
≥ Tf
Pf (Tf , α) = Pr
Id,ru + If,ru
Z ∞
Tf rfα
(If,ru + Id,ru ) fkXf k (rf ) drd
=
Pr hf ≥
Pf
Z0 ∞
Tf rfα
Tf rfα
(a)
=
LId,ru
fkXf k (rf ) drf
LIf,ru
Pf
Pf
Z 0∞
(b)
=
exp −πλf rf2 ρ (Tf , α)
0
2!
2
Pd Tf α
2
· exp −πλtu rf C (α)
2πλf rf e−πλf rf drf
Pf
1 + ρ (Tf , α) +
(α)
2/α ,
+
πλtu
2λf
1
p
p π
1 + Tf 2 − 1/ Tf +
λtu
λf
q
Pd
Pf
q
πλtu
2λf
,
Pd Tf
Pd
Pd Tf
Pf
q
(29)
Pd Tf
Pf
For the system to have a stable equilibrium point, all
eigenvalues of the Jacobian of the system of equations should
have a negative real part [30]. As we have one system equation,
an equivalence to this condition is that the Jacobian should be
negative definite. To this end, we first denote by f the right
hand side of (21). Accordingly, we have
(D)
Pd Tf
Pf
1
1+v 2 dv
q
E. Proof of Theorem 2
dxd
1
p R∞
1 + Tf 1/√Tf
2
= p
π Tf 1 +
df
λtu
λf C
(4)
(27)
2
α
2
= − Ei −Tdα πkXd k βC (α) ,
2
2/α
where β = pD
λ
+
(P
/P
)
λ
, and the proof is
tu
f
d
f
c
finished.
1
λtu
λf C
1 + ρ (Tf , 4) +
where (a) follows the property of the inverse trigonometric
functions that arctan(A) ≈ A if A is smaller than 1, i.e.,
Tf ≥ 1.
Then, substituting (29) into (26) with W = ln(Tf ) and we
obtain the result.
2θd
2
exp −πkXd k βC (α) e α dθd
=
1
(28)
where (a) follows the setting of hf ∼ exp(1) and the
independence between inter-tier interference Id,ru and intratier interference If,ru , equation (b) follows the definition of
the Laplace transform for Id,ru , If,ru , where LId,ru and LIf,ru
can be derived by using a similar way as (25), and the proof
is finished.
(D)
σxf
=
(D)
1 − xf
(D)
dπf
(x)
(D)
dxf
(D)
−
dπd (x)
(D)
dx
f
(D)
(D)
(D)
+σ 1 − 2xd
πf (x) − πd (x)
!
(a)
(b)
(30)
(D)
(D)
According to the definition, πf (x) = πd (x) at the
(D)
equilibrium point. Hence, (b) = 0. Then we calculate
(D)
and
dπd
(x)
(D)
dxf
(x)
(D)
dxf
(a) as (31) and (31) on the top of next page
(D)
It is obvious that for any xf
(D)
dπf (x)
(D)
dxf
(D)
dπf (x)
(D)
dxf
dπf
<
0,and
−
(D)
dπd (x)
(D)
dxf
(D)
dπd (x)
(D)
dxf
>
< 0, and
>
0, we have
0. Hence this proves
df
(D)
dxf
(D)
is strictly negative
(D)
for any non-zero value of xf . Therefore, ẋf evaluated at
any interior equilibrium point is negative. This completes the
proof.
R EFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, et al., “What will 5G be?”, IEEE J.
Sel. Areas Commun., vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[2] China Mobile Res. Inst., “C-RAN The Road Towards Green RAN”,
Beijing, China, Oct. 2011, White Paper, ver. 2.5.
[3] M. Peng, Y. Sun, X. Li, Z. Mao, and C. Wang, “Recent advances in
cloud radio access networks: System architectures, key techniques, and
open issues,” IEEE Commun. Surveys Tuts., vol. 18, no. 3, pp. 22822308, Aug. 2016.
9
(D)
dπf
(x)
(D)
dxf
=
d
(D)
dπd (x)
(D)
dxf
=
(D)
pf Bf
(D) Cf
N(F) +N(D) xf
d
(D)
dxf
(D)
pd Bd
(D) C
N(D) (1−xf ) d
(D)
dxf
− qf φf
− qd φd
(D)
= −
N (D) pf Bf Cf
(D)
N (F ) + N (D) xf
(D)
(F )
(D)
+ N (D) xf
2 − c3 c4 N φf exp c4 N
(31)
(D)
=
N (F ) pd Bd Cd
(D)
(D)
(D)
N
(1
−
x
)
+
c
c
N
φ
exp
c
2
1 2
d
2
f
(D)
N (D) (1 − xf )
[4] M. Peng, Y. Li, J. Jiang, J. Li, and C. Wang, “Heterogeneous cloud radio
access networks: A new perspective for enhancing spectral and energy
efficiencies”, IEEE Wireless Commun., vol. 21, no. 6, pp. 126-135, Dec.
2014.
[5] M. Peng, S. Yan, K. Zhang, and C. Wang, “Fog Computing based Radio
Access Networks: Issues and Challenges”, IEEE Networks, vol.30, no.
4, pp. 46-53, Jul.. 2016.
[6] M. Peng, K. Zhang, “Recent Advances in Fog Radio Access Networks:
Performance Analysis and Radio Resource Allocation,” IEEE Access,
vol. 4, pp. 5003-5009, Sep. 2016.
[7] Cisco technology radar trends: Fog computing. [Online]. Available:
http://www.cisco.com/web
/solutions/trends/tech−radar/fogcomputing.html
[8] M. Peng, S. Yan, H. V. Poor, “Ergodic capacity analysis of remote
radio head associations in cloud radio access networks”, IEEE Wireless
Commun. Letters, vol. 3, no. 4, pp. 365–368, Aug. 2014.
[9] M. Abana, M. Peng, Z. Zhao, et al., “Coverage and Rate Analysis in
Heterogeneous Cloud Radio Access Networks With Device-to-Device
Communication”, IEEE Access, vol. 4, pp. 2357–2370, Jun. 2016.
[10] Z. Zhao, M. Peng, Z. Ding, et al., “Cluster Formation in Cloud-Radio
Access Networks: Performance Analysis and Algorithms Design”, IEEE
ICC 2015, London, Jun. 2015.
[11] S. Yan, M.Peng and W. Wang “User Access Mode Selection in Fog
Computing Based Radio Access Networks”, IEEE ICC 2016, Kuala
Lumpur, Jun. 2016.
[12] M. Peng, C. Wang, V. Lau, and H. V. Poor, “Fronthaul-constrained
cloud radio access networks: Insights and challenges,” IEEE Wireless
Commun., vol. 22, no. 2, pp. 152-160, Apr. 2015.
[13] D. Feng, L. Lu, G.Y. Li, et al., “Device-to-Device Communications
Underlaying Cellular Networks”, IEEE Trans. Commun., vol. 61, no.
8, pp. 3541 - 3551, Jul. 2013.
[14] M. Peng, Y. Li, T. Q. S. Quek, and C. Wang, “Device-to-device underlaid
cellular networks under Rician fading channels,” IEEE Trans. Wireless
Commun., vol. 13, no. 8, pp. 4247-4259, Aug. 2014.
[15] Y. Li, D. Jin, J. Yuan, Z. Han, “Coalitional Games for Resource Allocation in the Device-to-Device Uplink Underlaying Cellular Networks”,
IEEE Trans. Wireless Commun., vol. 13, no. 7, pp. 3965 - 3977, May.
2014.
[16] D. Chen, T. Quek, and M. Kountouris, “Backhauling in Heterogeneous
Cellular Networks: Modeling and Tradeoffs”, IEEE Trans. Wireless
Commun., vol. 14, no. 6, pp. 3194–3206, Jun. 2015.
(32)
[17] L. Breslau, C. Pao, F. Li, et al., “Web Caching and Zipf-like Distributions: Evidence and Implications”, Infocom 99’, pp. 126–134, Mar.
1999.
[18] C. Yang, Y. Yao, Z. Chen, et al., “Analysis on Cache-enabled Wireless
Heterogeneous Networks”, IEEE Trans. Wireless Commun., vol.15, no.
1, pp. 131–145, Jan. 2016.
[19] G. Zhang, T. Q. S. Quek, M. Kountouris, et al., “Fundamentals of
Heterogeneous Backhaul Design Analysis and Optimization ”, IEEE
Trans. Commun., vol. 64, no.2, pp. 876–889, Feb. 2016.
[20] G. Hooghiemstra and P. Van Mieghem, “Delay distribution on fixed
Internet paths”, Delft University of Technology, Tech. Rep. 20011020,
2001.
[21] K. Papagiannaki,S. Moon, C. Fraleigh, et al., “Measurements and
analysis of single-hop delay on an IP backbone networks”, IEEE J.
Sel. Areas Commun., vol. 21, no. 6, pp. 908–921, Aug. 2003.
[22] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage
and rate in cellular networks”, IEEE Trans. Commun., vol. 59, no. 11,
pp. 3122–3134, Nov. 2011.
[23] F. Baccelli, B. Blaszczyszyn, and P. Uhlethaler, “Stochastic analysis of
spatial and opportunistic ALOHA”, IEEE J. Sel. Areas Commun., vol.
27, no. 7, pp. 1105–1119, Sep. 2009.
[24] D. Stoyan, W. Kendall, and J. Mecke, “Stochastic Geometry and Its
Applications”, 2nd ed. John Wiley and Sons, 1996.
[25] D. Easley, J. Kleinberg, Networks, Crowds, and Markets: Reasoning
About a Highly Connected World, Cambridge University Press, 2010.
[26] H. Tembine, E. Altman, R. El-Azouzi, et al.,“Evolutionary Games in
Wireless Networks”, IEEE Trans. Systems, Man, and Cybernetics, Part
B , vol. 40, no. 3, pp. 634–646, Jun. 2010.
[27] D. Niyato, E. Hossain, and Z. Han, “Dynamics of multiple-seller
and multiple-buyer spectrum trading in cognitive radio networks: a
gametheoretic modeling approach”, IEEE Trans. Mobile Computing, vol.
8, no. 8, pp. 1009–1022, Aug. 2009.
[28] D. Niyato and E. Hossain, “Dynamics of network selection in heterogeneous wireless networks: An evolutionary game approach”, IEEE Trans.
Veh. Technol., vol. 58, no. 4, pp. 2008–2017, May 2009.
[29] W. Cheung, T. Quek, and M. Kountouris, “Throughput Optimization,
Spectrum Allocation, and Access Control in Two-Tier Femtocell Networks”, IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 561–574, Apr.
2012.
[30] Y. A. Kuznetsov, Elements of Applied Bifurcation Theory., Berlin,
Germany: Springer-Verlag, 2004.
| 7cs.IT
|
1
A Distributed and Privacy-Aware Speed Advisory
System for Optimising Conventional and Electric
Vehicles Networks
arXiv:1506.01436v2 [cs.SY] 20 Aug 2015
Mingming Liu, Rodrigo H. Ordóñez-Hurtado, Fabian Wirth, Yingqi Gu, Emanuele Crisostomi and Robert Shorten
Abstract—One of the key ideas to make Intelligent Transportation Systems (ITS) work effectively is to deploy advanced
communication and cooperative control technologies among the
vehicles and road infrastructures. In this spirit, we propose a
consensus-based distributed speed advisory system that optimally
determines a recommended common speed for a given area in
order that the group emissions, or group battery consumptions,
are minimised. Our algorithms achieve this in a privacy-aware
manner; namely, individual vehicles do not reveal in-vehicle
information to other vehicles or to infrastructure. A mobility
simulator is used to illustrate the efficacy of the algorithm, and
hardware-in-the-loop tests involving a real vehicle are given to
illustrate user acceptability and ease of the deployment.
Index Terms—Speed advisory systems, Distributed algorithms,
Optimisation.
vehicles to take into account their differences (e.g., vehicle
type, vehicle age, fuel mode, load, desired time of arrival).
However, asking different cars to drive at different speeds
does not make sense in any practical scenario. On the other
hand, groups of cars following a common speed occurs
frequently in practice. For example, roughly speaking, in
highway driving cars tend to follow a given speed limit
where possible. This tendency is increased in situations
where intelligent tempomats are deployed. Other situations
where common cars follow common speeds include dedicated
lanes, and special zones in cities. Finally, generic benefits of
following a common speed include, reduced emissions (due
to less frequent accelerations/decelerations), reduced energy
consumption, increased throughput, and increased safety
[9]–[12].
I. I NTRODUCTION
At present, Intelligent Speed Advisory (ISA) systems, as
a part of Advanced Driver Assistance Systems (ADASs),
have become a fundamental part of Intelligent Transportation
Systems (ITS). Such systems offer many potential benefits,
including improved vehicle and pedestrian safety, better
utilisation of the road network, and reduced emissions.
Recently, many papers have appeared on this topic
reflecting the problem from the viewpoint of road operators,
infrastructure providers, and transportation solution providers
[1]–[8].
In this paper, we consider the design of a speed advisory
system (SAS) making use of vehicle-to-vehicle/infrastructure
(V2X) technologies. Our starting point is the observation that
different vehicle classes are designed to operate optimally at
different vehicle speeds and at different loading conditions.
Thus, a recommended speed, or speed limit, may be optimal
for one vehicle and not for others. One possible way to
handle this is by recommending a different speed to different
M. Liu and R. Ordóñez-Hurtado, joint first authors, are with the
Hamilton Institute, Maynooth University, Maynooth, Ireland (e-mail:
[email protected]; [email protected]).
F. Wirth is with the Faculty of Computer Science and Mathematics,
University of Passau, Germany (e-mail: [email protected]).
Y. Gu, M. Liu, R. Ordóñez-Hurtado, R. Shorten are with the School of
Electrical, Electronic and Communications Engineering, University College
Dublin, Ireland (e-mail: [email protected])
E. Crisostomi is with the Department of Energy, Systems, Territory and Constructions Engineering, University of Pisa, Italy (e-mail:
[email protected]).
R. Shorten is also with IBM Research, Dublin, Ireland (e-mail: [email protected]).
In this paper, we consider the problem of recommending a
common speed to a set of cars. Our ISA will be constructed
with an instantaneous optimality goal in mind. For example,
we may wish to minimise the instantaneous group emissions,
or the instantaneous group energy consumption. Specifically,
we are interested in addressing the following questions:
(i) Firstly, we are interested in designing an ISA system to
recommend the same speed to a network of conventional
vehicles (i.e., with an Internal Combustion Engine (ICE))
travelling along an extra-urban route (e.g., a highway),
connected via a V2X communication system, such that
the total emission is minimised;
(ii) Secondly, we are interested in designing an ISA system
to recommend the same speed to a network of EVs
travelling in an urban network (e.g., the city centre),
connected via a V2X communication system, such that
the total battery consumption is minimised.
We shall show that these problems can be solved in a
privacy preserving manner without a large communication
burden on the vehicle-to-vehicle (V2V) network. Furthermore,
we believe that algorithms addressing these problems will
be of use when cars are asked to follow a constant speed
to achieve certain goals. These include, transit through
environmental and safety zones, highway driving with
adaptive speed-limits (up to some maximum deviation from
the nominal speed limit), and eco-driving for fleets of
autonomous vehicles travelling in special lanes.
2
This paper is an extended version of the idea presented in
[13]. In order to allow vehicles to collaboratively compute the
optimal recommended speed, we shall further assume they are
equipped with V2X technologies, can exchange information
with their neighbours, and can exchange limited information
with the infrastructure. We shall show that one can design,
using very simple ideas, an effective SAS in a manner that
preserves the privacy of individual vehicles. Extensive simulations, including hardware-in-the-loop (HIL) tests using real
vehicles, are given to demonstrate the efficacy of our approach.
recommendations are made. Let si (k) be the recommended
speed of the vehicle i ∈ N := {1, 2, ..., N } calculated at
time instant k. Thus, the vector of recommended speeds for
T
all vehicles is given by s (k) := [s1 (k) , s2 (k) , ..., sN (k)],
where the superscript T represents the transposition of the
vector. Note that between two consecutive time instants
(k, k + 1), the recommended speeds are constant while the
driving speeds are time-varying real-valued variables. We
denote by Nki the set of neighbours of vehicle i at time instant
k, i.e., those vehicles which can successfully broadcast their
recommended speeds to vehicle i.
II. R ELATED WORK
In this section, we give a brief review of some related
work. First note that a detailed review of this topic is given
in [14]. Conventional systems are described in [5]–[7], [15],
[16]. These papers describe various aspects of the ISA design
process, which includes the design of driver display systems,
the incorporation of external environmental information, and
the algorithmic aspects of speed and distance recommendations. Recently, there has been a strong trend to also include
traffic density information; references [14], [17]–[21] describe
work in this direction. In these works, density information is
included in the procedure via loop detectors or via explicit
density estimation using V2V technology. The differentiating
feature of the approach followed in this paper is that density
and composition of the vehicle fleet is also used, but in an
implicit manner as part of the optimisation algorithm. Finally,
we note that there is a huge body of work on cooperative
control of vehicles and its connection to consensus algorithms
[16], [22]–[24]. It is important to note that we are designing a
SAS and not a cooperative control system. This distinction is
important as it allows us to ignore string stability effects which
are a fundamental limitation of many cooperative control
architectures [25]–[28].
In addition, we assume that each vehicle i can evaluate a
function fi that determines its average emissions, in the ICE
case study, or its average energy consumption in the EV case,
were it to be travelling at the recommended speed si (k). Such
functions are typically convex functions of the vehicle speed,
as it will be shown in more detail in the following sections.
We shall further assume that these functions are continuously
differentiable and with a Lipschitz continuous first derivative
fi0 which is assumed with positive bounded growth rate, i.e.,
(i)
0 < dmin ≤
fi0 (a) − fi0 (b)
≤ d(i)
max ,
a−b
(1)
for all a, b ∈ R such that a6=b, and suitable positive constants
(i)
(i)
dmin , dmax . A schematic diagram of the above is illustrated
in Fig. 1. In this context, we consider the following problem.
Problem 1: Design an ISA system for a network of vehicles
connected via V2X communication systems, in order to
recommend a common speed that minimises the amount of
emissions, or the total energy consumption, of the whole fleet
of vehicles.
Base station
III. M ODEL AND A LGORITHM
A. Problem Statement
Our objective is to develop an algorithm that converges
to a common recommended speed for a fleet of vehicles
driving along the same stretch of a road (e.g., a highway),
or in the same area (e.g., in the city centre). Let N denote
the total number of vehicles in the fleet, and let us assume
that they all are able to receive the ISA broadcast signal.
For this purpose, each vehicle is equipped with a specific
communication device (e.g., a mobile phone with access to
WiFi/3G networks) so that they are able to receive/transmit
messages from/to either nearby vehicles or available road
infrastructure (e.g. a base station). We assume that each
vehicle can communicate a limited amount of information
with the infrastructure, the infrastructure can broadcast
information to the entire network of cars, and each vehicle
can send a broadcast signal to its neighbours.
For convenience, we assume that all vehicles have
access to a common clock (for example, a GPS clock).
Let k ∈ {1, 2, 3, ...} be a discrete-time instant in which
new information from vehicles is collected and new speed
Measurement of
global values
Feedback of
global values
Vehicles
Network of vehicles
Fig. 1: Schematic diagram of the framework [29].
The optimisation problem that needs to be solved in order
to address Problem 1 can be formulated as follows:
X
min
fj (sj ) ,
s∈RN
(2)
j∈N
s.t. si = sj , ∀i 6= j ∈ N.
3
The above is an optimised consensus problem and can
be solved in a variety of ways (for example using ADMM
[30]–[32]). Our focus in this present work is not to construct
a fully distributed solution to this problem, but rather to
construct a partially distributed solution which allows rapid
convergence to the optimum, without requiring the vehicles
to exchange information that reveals individual cost functions
to other vehicles. This is the privacy preserving component
of our problem statement.
Comment: Note that (2) involves the recommended speeds sj
and not the actual speeds at which the vehicles are travelling
in reality. This implies, among other things, that the optimal
solution of (2) is the same, disregarding whether vehicles
follow in reality to the recommended speed or not.
Comment: In the following, we shall not solve (2) in a
single step. Instead, we shall propose an iterative algorithm
that at each step yields individual recommended speeds that
will eventually converge to the same value on the consensus
constraints.
To solve (2) we use the iterative feedback scheme
has a solution if and only if there exists a y ∗ ∈ R satisfying
N
X
fj0 (y ∗ ) = 0.
(4)
j=1
In this case, y ∗ is unique by strict convexity, and the unique
optimal point of (2) is given by
s∗ := y ∗ e ∈ RN .
(5)
In order to obtain convergence of (3) we select a feedback
signal
N
X
G (s (k)) = −µ
fj0 (sj (k)) ,
(6)
j=1
and we obtain the dynamical system
s(k + 1) = P (k)s(k) − µ
N
X
fj0 (sj (k))e, µ ∈ R.
(7)
j=1
In [29] it is shown that if {P (k)}k∈N is a uniformly strongly
ergodic sequence2 and µ is chosen according to
−1
N
X
,
d(j)
0 < µ < 2
(8)
max
j=1
s (k + 1) = P (k) s (k) + G (s (k)) e,
(3)
where {P (k)} ∈ RN ×N is a sequence of row-stochastic
matrices1 , e ∈ RN is a column vector with all entries
equal to 1, and G : RN 7→ R is a continuous function
with some assumptions to satisfy as we shall see in
Theorem 1. Algorithms of this type were proposed and
studied among others in [33]–[35]; the principal theoretical
contribution here is to extend this framework to a new class
of optimisation problems and to give conditions guaranteeing
their convergence.
Remark: Eq. (3) describes a consensus algorithm with an
input. In the first part of the right-hand side of (3), P (k)
creates pressure on s(k) so that all components of s(k)
achieve a common value. The second part of (3) describes the
regulation constraint that the nonlinear function of s(k) must
satisfy. Thus, we can use Eq. (3) to force all components
of s(k) to achieve a common value and at the same time
to satisfy a nonlinear constraint. The choice of G thus
determines which constraint is satisfied when s(k) reaches
the equilibrium. In principle, many choices of G can be used.
In this paper, we choose G to solve an optimisation problem.
The principal mathematical considerations in deploying this
algorithm are to determine the mathematical conditions on
G so that (3) actually converges to the desired value. While
this is not a primary concern of this paper, rather we are
interested in the applications of this, in the remainder of this
section we shall give a brief overview of such considerations.
We will require that (2) has a unique solution. Note that, it
follows from elementary optimisation theory that if all the fi ’s
are strictly convex functions, then the optimisation problem (2)
1 Square
matrices with non-negative real entries, and rows summing to 1.
then (7) is uniformly globally asymptotically stable at the
unique optimal point s∗ = y ∗ e of (2). Such systems were
studied in [33] and formally analysed in [29]. For completeness, we formally state the relevant results from these works
as a theorem (Theorem 1). An overview of the proof is given
in the appendix, and the interested reader may refer to [29]
for details.
Theorem 1 ( [29]) Consider the optimisation problem (2)
and the optimisation algorithm (3), and the associated Lure
system
y (k + 1) = h (y (k)) ,
(9)
h (y) := y + G (ye) .
If G is defined by (6) and the condition (8) holds, then the
following assertions hold:
(i) There exists a unique, globally asymptotically stable
fixed point y ∗ ∈ R of the Lure system (9).
(ii) The fixed point y ∗ of (i) satisfies the optimality condition (4), and thus y ∗ e ∈ RN is the unique optimal point
for the optimisation problem (2).
(iii) If, in addition, {P (k)}k∈N ⊂ RN ×N is a strongly
ergodic sequence of row-stochastic matrices, then y ∗ e is
a globally asymptotically stable fixed point for system (3).
An outline of the proof can be found in Appendix A. To
apply the Theorem 1 to solve the optimisation problem we
proceed as follows. For each k we define the P (k) as
P
1 − j∈Nki ηj , if j = i,
Pi,j (k) =
(10)
ηj ,
if j ∈ Nki , ,
0,
otherwise.
2 That is, for every k ∈ N the sequence P (k ), P (k + 1)P (k ), . . .,
0
0
0
0
P (k0 + `) · · · P (k0 ), . . . converges to a rank one matrix. See [29] for further
details.
4
where i, j are the entries’ indexes of the matrix P (k), and
ηj ∈ R is a weighting
factor. For example, a convenient choice
ηj is N i1 +1 ∈ 0, N 1−1 , where |•| denotes cardinality,
| k|
giving rise to an equal weight factor for all elements in the
reference speed vector s (k).
The assumption of uniform strong ergodicity holds if the
neighborhood graph associated with the problem has suitable
connectedness properties. If sufficiently many cars travel on a
given area, it is reasonable to expect that this graph is strongly
connected at most time instances. Weaker assumptions are
possible but we do not discuss them here for reasons of
space; see [36] for possible assumptions in this context.
Now, we propose the Optimal Decentralised Consensus
Algorithm for solving (2) as shown in Algorithm 1. The
underlying assumption here is that at all time instants all cars
communicate their value fj0 (sj (k)) to the base station, which
reports the aggregate sum back to all cars. This is precisely
the privacy preserving aspect of the algorithm, as cars do not
have to reveal their cost functions to anyone. Also implicit
information as derivatives of the cost function at certain speeds
is only revealed to the base station but not to any other agent
involved in the system.
Algorithm 1 Optimal Decentralised Consensus Algorithm
for k = 1, 2, 3, .. do
for each i ∈ N doP
Get F̃ (k) =
fj0 (sj (k)) from the base station.
j∈N
i
4:
Get sj (k) from all
Pneighbours j ∈ Nk .
5:
Do qi (k) = ηi ·
(sj (k) − si (k)).
1:
2:
3:
j∈Nki
6:
7:
8:
Do si (k + 1) = si (k) + qi (k) − µ · F̃ (k).
end for
end for
the highway where the SAS is active; in doing so, we
provide a discussion about the rate of convergence of
Algorithm 1, and also evaluate what happens when some
vehicles decide not to follow the recommended speed.
3) Finally, we give a HIL emulation with a real vehicle
travelling on a real road, embedded into a emulated
network with a fixed number of simulated vehicles.
In all the above experiments, we evaluate the performance
of Algorithm 1 for η = 0.001 and µ = 0.01 (a convergent
set-up), and all the vehicles are also subject to the driving
constraints arising from the SUMO simulation (e.g., in terms
of acceleration/deceleration profiles).
A. Cost funtions to represent pollution emission in ICE vehicles
We shall adopt the average-speed model proposed in [38] to
model each cost function fi in function of the average speed
s as
a + bsi + cs2 + ds3 + es4 + fs5 + gs6
, (11)
fi = k
s
where a, b, c, d, e, f, g, k ∈ R are used to specify different
levels of emissions by different classes of vehicles. In addition,
we use the emission types from [38] shown in Table I and Fig.
2, corresponding to petrol cars/minibuses with up to 2.5 tons of
gross vehicle mass. Also, we use the following vehicle types:
•
•
•
•
Type
Type
Type
Type
accel.
accel.
accel.
accel.
2.15
1.22
1.75
2.45
m/s2 ,
m/s2 ,
m/s2 ,
m/s2 ,
decel.
decel.
decel.
decel.
5.5
5.0
6.1
6.1
m/s2 ,
m/s2 ,
m/s2 ,
m/s2 ,
length
length
length
length
4.54
4.51
4.45
4.48
m.
m.
m.
m.
TABLE I: Emission factors for some CO2 emission types
reported in [38]. Here, {e, f, g} = 0 and k = 1.
Type
R007
R014
R021
R040
IV. A S PEED A DVISORY S YSTEM FOR C ONVENTIONAL
V EHICLES
1000
CO2 emissions [g/km]
Comment: We note that in any practical implementation of the
previous algorithm, the recommended speed may be bounded
above and below by the road operator (i.e., should the optimal
solution computed by Algorithm 1 exceed speed limits, or be
unreasonably low).
In this section, we evaluate the performance of Algorithm 1
using SUMO [37]. SUMO is a microscopic, open source road
traffic simulator, which is frequently used for validation and
prediction purposes in the ITS community. First, we evaluate
it using conventional simulations, and then use a real vehicle
embedded into a hardware-in-the-loop (HIL) emulation. In
particular, we perform the following experimental activities:
1) First, we compare the pollution emission of a fixed fleet
of vehicles (with initial non-optimal speeds) before and
after Algorithm 1 is turned on.
2) We then provide simulation results which are dynamic
in nature, where vehicles enter and leave a section of
1:
2:
3:
4:
a
2.2606E+3
2.5324E+3
3.7473E+3
1.2988E+3
b
3.1583E+1
6.8842E+1
1.0571E+2
2.0203E+2
R007
c
2.9263E-1
-4.3167E-1
-8.5270E-1
-1.5597E-0
R014
R021
d
3.0199E-3
6.6776E-3
1.0318E-2
1.2264E-2
R040
800
600
400
200
0
0
20
40
60
80
100
120
140
Average speed [km/h]
Fig. 2: Curves for the CO2 emission types in Table I. Black
circles mark the minimum point on each curve.
5
Evolution of the recommended speeds
B. SUMO Simulations with a Fixed Number of Vehicles
90
Speed [km/h]
In this experiment we consider 40 vehicles travelling along
a highway. The set-up for this set of experiments is as follows.
• Road: A straight 5 km long highway with 4 lanes.
• Duration of the simulation: 600 s.
• Algorithm sampling interval: ∆T = 1s.
• Switch-on time: the algorithm is activated at time 300 s.
80
70
60
50
200
In the second experiment (see Fig. 4), we assume that the
vehicles are initially driving at random speeds around the
optimal one. Then, once the SAS is activated, the vehicles
change their speed to eventually converge to the optimal
one. Significant CO2 savings can be noticed in this example
again; however, carbon savings here are less than in the first
experiment, since now some vehicles were already driving at
(or very close to) the optimal speed.
Evolution of the recommended speeds
Speed [km/h]
90
80
70
60
50
200
250
300
350
400
450
500
550
600
500
550
600
Time [s]
Total CO2 emissions
CO2 emissions [g/km]
9000
8950
8900
8850
8800
200
250
300
350
400
450
Time [s]
Fig. 3: Simulation results for the static case, before and after
the activation of the algorithm at time step 300 s. Setup: 32
vehicles of emission type R007 and 8 vehicles of emission
type R021, with uniform distribution of vehicle types.
C. SUMO Simulations with a Dynamic Number of Vehicles
We now consider a dynamic scenario, with a time-varying
number of vehicles. To do this we partition the highway into
300
350
400
450
500
550
600
500
550
600
Time [s]
Total CO2 emissions
5440
CO2 emissions [g/km]
The experimental results are given in Fig. 3 and Fig. 4. As
we can see from Fig. 3, making a small change in a common
non-optimal initial speed yields a significant improvement
in CO2 emissions, as about 172 g of CO2 are saved every
kilometer. If this result is integrated over a whole road
network, and integrated over a longer time horizon (e.g., a
year), significant improvements can be observed. For instance,
the savings are comparable to those that can be achieved if
a EURO-3 vehicle (e.g., R039) is substituted by a EURO-4
vehicle (e.g., R040), where the savings are around 10 g/km.
250
5420
5400
5380
5360
200
250
300
350
400
450
Time [s]
Fig. 4: Simulation results for the static case, before and after
the activation of the algorithm at time step 300 s. Setup:
uniform distribution of emission types in R014, R021 and
R040, with uniform distribution of vehicle types.
three consecutive sections L1, L2 and L3. We then proceed
as follows. First, vehicles enter the uncontrolled section L1,
with constant speed (randomly chosen in a given range).
After completing L1, vehicles enter the section L2 where the
SAS is always active, so the vehicles iteratively update their
recommended speed using Algorithm 1. After completing L2,
they finally enter section L3 where they travel freely. The
experiments are setup as follows.
• Road: three consecutive straight edges:
– L1: 5 km long highway with 4 lanes, uncontrolled;
– L2: 5 km long highway with 4 lanes, ISA controlled;
– L3: 5 km long highway with 4 lanes, uncontrolled.
• Total number of cars: 650, with uniform distribution of
both emission types among R014, R021 and R040, and
vehicle types.
• Vehicular flow entering L1: one new car every 2 seconds
until simulation time 1300 s.
• Length of simulation: 3010 s.
• Window size for the calculation of the moving average
(MA) of CO2 emissions for visualisation purposes: 500
time steps.
• Travelling speeds for cars on L1 are randomly chosen
with uniform distribution in 3 scenarios:
– Case 1, constant speeds in (80, 100) km/h.
– Case 2, constant speeds in (60, 80) km/h.
– Case 3, constant speeds in (40, 60) km/h.
Note that even though this is a dynamic situation, the
vehicle density on each part of the road becomes almost
constant after certain time. A sample of simulation results
is given in Fig. 5, which reveals what might be expected
from the initial experiments. Namely, the further vehicles are
6
Moving average (MA) of CO2 emissions per section
CO2 emissions [g]
1500
MA emissions L1
MA emissions L2
1000
500
0
0
500
1000
1500
2000
2500
Time [s]
Fig. 5: Simulation results for the dynamic case: MA of CO2
emissions on L1 and L2, for initial speeds on L1 in Case 3.
To generalise the previous results, we conducted one
hundred random experiments for each of the three cases
described above. In each experiment, we collected the
simulation data of the total CO2 emission generation on each
section of the highway from SUMO. Table II summarises the
aggregated results of this exercise, and clearly demonstrates
the benefits of the ISA.
TABLE II: General results: total emissions per lane.
of neighbours. Convergence rate of algorithms of graphs
is a well-studied topic. The interested readers are referred
to [39], [40] for the favours of results available. In this
paper, we just give a brief empirical illustration. As we
shall see, roughly speaking, the larger is the communication
range of the vehicles, the larger will be the number of
neighbouring vehicles, and (on average) the earlier is the
convergence to the steady-state value. In particular, Fig. 6
shows how convergence of a single vehicle is affected by its
communication range. Note that convergence occurs after a
few steps (less than three minutes) when the communication
range is of the order of hundreds of meters. On the other side,
convergence is very slow when the communication range is
of only 15 meters, where it is frequent to get an empty set of
neighbouring vehicles. Finally, observe that the steady-state
value of the optimal speed is not constant as the optimal
speed depends on the vehicles currently driving along the
road section L2, and thus the types of vehicles (and in turn
the optimal speed) change continuously.
Evolution of the recommended speed
72
r=15 m
71
Speed [km/h]
away from the optimal speed, the more is to be gained by
deploying the ISA.
r=100 m
r=400 m
r=1000 m
70
69
68
67
66
65
Improvement
Mean
σ
3.40%
0.07
0.69%
0.03
7.94%
0.16
∗: Sum of emissions at every time step (i.e. time integration). Mean: average
of 100 different measurements. σ: standard deviation.
Comment: Note that the solution we have obtained is optimal
for the environment and for the collective, e.g., in terms of the
reduction of overall emissions. However, the solution might be
unfair for some single users who would be recommended to
drive at a different speed than originally desired. One way to
improve fairness could be to decrease road taxes for virtuous
vehicles, to compensate them from the inconvenience caused
by the dirty vehicles in terms of recommended average speeds.
Comment: Note that it is clearly the case that the design
parameters of the algorithm have the potential to affect
emission savings. For example, the speed of convergence of
the algorithm affects the rate of which the emissions are saved.
Following the previous comments, convergence issues are
briefly discussed with the support of Fig. 6. As we have
mentioned in the previous paragraph, the recommended
speed is updated in an iterative fashion, and it eventually
converges to a steady-state value after a number of steps. The
time of convergence is mainly affected by the topology of
the communication graph and, in particular, by the number
0
50
100
150
200
250
Time [s]
Observed number of neighbours
80
Number of neighbours
Case
1
2
3
Total Emissions∗ [tonnes]
L1 (Uncontrolled)
L2 (Controlled)
Mean
σ
Mean
σ
1.5199
0.00114
1.4684
0.0002
1.4751
0.0004
1.4648
0.0001
1.5945
0.0028
1.4679
0.0001
r=15 m
r=100 m
r=400 m
r=1000 m
60
40
20
0
0
50
100
150
200
250
Time [s]
Fig. 6: Evolution of the recommended speed (top) and number
of neighbours (bottom) of a target vehicle, as functions of the
radius r of its communication range.
As has been previously remarked, convergence is for the
recommended speeds, and thus it is independent on whether
vehicles do follow the recommended speed or not. However,
not following the recommended speed usually decreases the
benefits of the proposed procedure in terms of CO2 savings.
Accordingly, Fig. 7 illustrates how CO2 savings are affected
by the percentage of vehicles that decide not to follow the
recommended speed.
D. SUMO-based hardware-in-the-loop (HIL) emulation
Finally, to give a feeling to a real driver of how this system
might function, we now describe a HIL implementation of the
7
Total CO2 emissions on L2
one lane, the vehicles are released sequentially from the same
starting point. The avatar representing the Prius departs in
the seventh position. Once the connection between the Prius
and the workstation is established, the position and speed
of the Prius’ avatar are updated using real-time information
from the Prius via the OBD-II adaptor. From the point of
view of the ISA algorithm, the Prius is regarded as a normal
agent in the SUMO simulation, i.e. treated just like any other
simulated vehicle.
2.658
CO2 emissions [Tonnes]
2.656
2.654
2.652
2.65
2.648
2.646
2.644
2.642
2.64
0
20
40
60
80
100
Percent of vehicles ignoring the recommended speed [%]
Fig. 7: Total CO2 emissions as a function of the percentage
of vehicles not following the recommendations: emissions are
maximum when all vehicles do not follow the recommended
speed, and minimum when all they do. Mean (x) and standard
deviation (I) are calculated based on five different experiments.
algorithm. Specifically, we use a SUMO-based HIL emulation
platform that was developed at the Hamilton Institute [41],
[42]. This emulation platform uses SUMO to simulate a real
environment and generate virtual cars, along with a dedicated
communication architecture supported by TraCI (a Python
script implementing a TPC-based client/server architecture) to
provide on-line access to SUMO, a smartphone connected to
the 3G network and running the plug-in SumoEmbed (designed
for use with Torque Pro [43]), and a OBD-II adaptor [44] to
embed a real car into the simulation, as shown in Fig. 8. The
idea then is to allow a person to drive a real vehicle on a real
street circuit, to experience being connected to a network of
(virtual) vehicles driving along a virtual environment based
on the physical street circuit. Specifically, we performed this
experiment by driving a Toyota Prius on a single-lane street
circuit in the North Campus of the Maynooth University, while
the Prius is embedded into the HIL emulation and represented
by an avatar which interacts with the avatars of 29 other
virtual (simulated) vehicles driving along the same stretch of
(emulated) road.
The consensus algorithm for the proposed ISA system is
embedded in the main component Python script. Thus, once
the respective recommended speeds are calculated, they are
sent to the vehicles through the server component, via the
cellular network to the smartphone in the case of the Prius,
and via TraCI commands in the case of the other vehicles
in the simulation. Note here that the driver behaviour is
different for a simulated vehicle compared to the case of the
Prius: while we force each simulated vehicle to follow the
recommended speed as far as possible3 , the Prius’ driver is
allowed to either follow or ignore the speed recommendation
(displayed on the smartphone’s screen) as desired.
The HIL experiment is setup as follows.
• Length of the experiment: 600 s, of which the ISA
algorithm is only engaged at around time 300 s;
• Total number of cars: 30, with uniform distribution of
emission types between R014 and R040, and uniform
distribution of vehicle types, with a maximum speed of
100 km/h.
• The sampling time interval ∆T for collecting new
information and updating the recommendations is 1 s.
Results of the experiment are depicted in Fig.9. From Fig.9
it can be observed that the total emissions oscillate around an
average value of 4323 g/km once all the 30 vehicles are added
to the simulation, and also that they start reducing once the
ISA algorithm is switched on, up to a final average value of
4253 g/km. Finally, as can be observed in Fig.9 (bottom), the
recommended speed can be roughly followed (on average) by
the driver subject to the physical constraints on the real street
circuit (e.g. traffic calming devices), and the discretisation step
for the visualisation of the recommended speed.
V. A S PEED A DVISORY S YSTEM FOR E LECTRIC V EHICLES
Fig. 8: Schematic of the SUMO-based HIL emulation platform.
The experiment begins when the simulation is started on
the workstation and the server component of the Python
script waits for a call from the OBD-II connected smartphone
in the real vehicle. Since the selected street circuit only has
In this section, we extend our previous discussion to the
special case of a fleet of EVs. The motivation for doing
so is the recent interest in EVs as a cleaner alternative
to conventional, more polluting, ICE vehicles. In recent
years, some cities have decided to close the city centres to
normal traffic, only allowing the transit to specific categories
of low (or zero) polluting vehicles; see for instance [45].
Similarly, in many cities all around the world, the urban
public transportation fleet has been restricted to electrically
3 Concerning mainly the interaction between vehicles and the design parameters for the simulated cars such as acceleration, deceleration, car following
model or driver information.
8
CO2 emissions per km
CO2 emissions [g/km]
4340
Overall emissions
4320
4300
•
4280
4260
4240
250
300
350
400
450
500
Time [s]
Evolution of the speeds for the Prius
90
Speed [km/h]
80
70
60
50
Prius’ recommended speed
Prius’ average driving speed
40
30
250
300
350
400
450
•
induction motor, gears, etc); in some cases, these power
losses have been modelled as a third-order polynomial,
whose parameters have been obtained by fitting some
experimental data (see [49]).
Tires: the power required to overcome the rolling distance
depends on the weight of the vehicle (and thus, on the
number of passengers as well), and is proportional to the
speed of the vehicle.
Ancillary systems: this category includes all other
electrical loads in the vehicle, such as HVAC systems,
external lights, audio system, battery cooling systems,
etc. Here, the power consumption does not depend on
the speed of the vehicle and can be represented by a
constant term that depends on external factors (e.g.,
weather conditions) and personal choices (desired indoor
temperature, volume of the radio, etc). According to
experimental evaluations [49], the power losses due
to ancillary services usually vary between 0.2 and 2.2 kW.
500
Time [s]
Fig. 9: Evolution of the overall CO2 emissions (top) and of the
speeds related with the Prius (bottom) for the HIL simulation.
The algorithm was turned on around time 300 s. Average speed
calculated with a window size of 20 time steps.
driven vehicles, as for the case of fleets of urban electric
buses; see for instance the cases of São Paulo [46] in Latin
America, Louisville in the US [47], or Wien in Europe [48].
Clearly, the previously designed Algorithm I needs to be
adapted when applied to a fleet of EVs concerning different
aspects. First, in terms of the field of application, EVs are
typically used for short distances due to their reduced driving
ranges, and thus are most likely deployed in city centres;
second, the cost functions should not consider instantaneous
emissions that, in the case of EVs, can be considered equal to
zero. Accordingly, in the following we show how the previous
framework can be adapted to the case of a fleet of EVs, and
now the objective is to maximise the energy efficiency of the
fleet of cars or, in other words, to extend their driving range.
A. Cost functions to represent energy consumption in EVs
Most of the discussion here follows the reference [49],
where the ranges of EVs are reported for different brands and
under different driving cycles. Power consumption in an EV
driving at a steady-state speed (along a flat road) is caused
mainly by four sources:
• Aerodynamics power losses: they are proportional to
the cube of the speed of the EV, and depend on other
parameters typical of a single vehicle such as its frontal
area and the drag coefficient (which in turn depends on
the shape of the vehicle).
• Drivetrain losses: they result from the process of converting energy in the battery into torque at the wheels
of the car. Their computation is not simple, as losses
might occur at different levels (in the inverter, in the
Thus, by summing up all the previous terms, the power
consumption Pcons can be represented as a function of the
speed v as
α0
Pcons
=
+ α1 + α2 v + α3 v 2 ,
(12)
v
v
where the left hand side is divided by the speed in order to
obtain an indication of energy consumption per km, expressed
in kWh/km. Such a unit of measurement is usually employed
in energy-efficiency evaluations, and we shall assume that
every single EV will use (12) as its personal cost function fi .
Accordingly, Fig. 10a shows a possible relationship between
speed and power consumption, obtained using data from
Tesla Roadster and assuming a low power consumption for
ancillary services of 0.56 kW (i.e., assuming air conditioning
switched off). As can be noted from Fig. 10a, there is a large
energy consumption at large speeds due to the fact that power
increases with the cube of the speed for aerodynamic reasons;
however, it is also large for low speeds, due to the fact that
travel times increase and, accordingly, constant power required
by ancillary services demands more energy than the same
services delivered with high speeds.
B. Experimental results
According to the previous discussion, we now assume that
the objective is to infer the optimal speed that the ISA system
should broadcast to a fleet of EVs travelling in a given area
of a city (e.g., in the city centre). For this purpose, we assume
that a fleet of 100 vehicles travels in the city centre for an
hour, and following the next steps:
• In the first 20 minutes, the vehicles travel at the optimal
speed calculated from Algorithm 1.
• In the second 20 minutes, they travel at a speed below
the optimal speed.
• In the last 20 minutes, they travel at a speed above the
optimal speed.
In the first stage we assume that the communication graph
among the EVs changes in a random way, i.e., at each time
9
50
Speed [km/h]
step an EV receives information from a subset of vehicles
belonging to the fleet. This is a simplifying assumption that
can be justified by assuming that in principle all vehicles
might communicate to all the other vehicles (i.e., they are
relatively close), but some communications might fail due
to obstacles, shadowing effects, external noise, etc. Besides,
in the two last stages we assume that the change of speed
occurs almost instantaneously, since there is no requirement
to iteratively compute an optimal speed.
40
30
20
10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time [h]
(a) Evolution of the vehicles’ speeds.
14
Energy losses [kWh/km]
We tuned our parameters in Algorithm 1 as η = µ = 0.001,
and we simulate different cost functions for each EVs by
assuming a random number of people inside each car (between
1 and 5 people) with an average weight of 80 kg, and
by assuming a different consumption from ancillary services
within the typical range of [0.2, 2.2] kW. The curves of the
cost functions used in our experiment are shown in Fig. 10b.
The evolution of the speeds of the EVs are shown in Fig. 11a,
while the average energy consumption is shown in Fig. 11b.
13
12
11
10
0
0.2
0.4
0.6
0.8
1
Time [h]
Energy losses [Wh/km]
600
(b) Evolution of the overall energy loss.
500
Fig. 11: Simulation results for the network of EVs: Algorithm
1 is applied until time 0.33 h, and then two different speeds
(below and above the optimal one) are suggested in [0.33, 0.66]
h and [0.66, 1] h, respectively.
400
300
200
100
0
0
20
40
60
80
100
120
140
Speed [km/h]
(a) An individual cost function.
Energy losses [Wh/km]
600
500
400
the acceptability of the algorithm. Finally, the algorithm has
been implemented in a real production vehicle embedded into
a HIL emulation using nothing more than a smartphone and
a commercially available OBD-II plug-in. In future work we
intend to extend the HIL simulation to include more vehicles
(and thus, more drivers) to better test the compliance of people
to follow the recommended speed.
300
A PPENDIX
O UTLINE OF PROOF OF T HEOREM 1
200
100
0
0
20
40
60
80
100
120
140
Speed [km/h]
(b) All the cost functions overlapped. Black circles mark the minimum
point on each curve.
Fig. 10: Curves for the cost functions used in the experiment.
All of them were chosen convex.
As is shown in Fig. 11b, a simple ISA can be used for the
case of a fleet of EVs just as in the case of conventional cars
(for what regards the mathematical background) by simply
adapting the cost functions to the new application of interest.
VI. C ONCLUSION
In this paper we present a new ISA system. The system
is based on a solution to an optimised consensus problem.
We show that the ISA can be implemented in a privacy
preserving fashion, in a manner that accounts for vehicle
density and composition, and in a manner that is provably
convergent. Simulations are given to illustrate the efficacy and
In this section, we give an outline of the proof for the
claims in Subsection III-A, in which we largely rely on the
results obtained in [29].
In Theorem 1, statement (i) is a consequence of the Banach
contraction theorem. It is a straightforward calculation to
show that the bounds (8) ensure that the function h defining
the Lure system (9) is in fact a global strict contraction on R.
Statement (ii) then follows directly from the definition of h:
if h(y ∗ ) = y ∗ then G(y ∗ e) = 0 and by (6) this is equivalent
to the optimality condition (4). The global optimality of y ∗ e
for the optimisation problem (2) of this optimal point then
follows as (4) is the standard first order necessary condition
for optimality, and because strict convexity of the cost
functions implies this condition is also sufficient. Uniqueness
is a further consequence of strict optimality.
It therefore remains to show that statement 1 (iii) holds. To
this end we recall the following two lemmas from [29].
10
Lemma 2 ( [29]) Let {P (k)}k∈N be a sequence of rowstochastic matrices. If {y (k)}k∈N is a solution of the Lure
system (9) then {y (k) e}k∈N is a solution of (3).
Lemma 3 ( [29]) Let {P (k)}k∈N be a strongly ergodic sequence of row-stochastic matrices, and suppose that G : Rn →
R is continuous and satisfies the following conditions:
(i) there exists an ε > 0 such that G satisfies a Lipschitz
condition with constant L > 0 on the set
Bε (E) := {x ∈ Rn : dist (x, E) ≤ ε} ,
where dist (x, E) := inf {kx − zk : z ∈ E} is the distance of a vector x ∈ Rn to the consensus set E :=
span {e}; and
(ii) there exists constants β, γ > 0 such that
|h (y)| ≤ |y| − γ when |y| ≥ β,
where h (y) = y + G (ye).
Then, every trajectory of (3) is bounded.
It is easy to see that G as defined in (6) satisifies the
conditions of Lemma 3. Indeed, G is even globally Lipschitz
continuous because of the Lipschitz continuity assumption (1).
Furthermore, as h is a strict contraction on R with fixed point
y ∗ , we may denote the contraction constant of h by 0 < c < 1
and obtain for any y ∈ R such that
|h (y)| ≤ |h (y) −y ∗ | + |y ∗ |
≤ c |y| + (1+c) |y ∗ |
≤ c |y−y ∗ | + |y ∗ |
= |y| − (1−c) |y| + (1+c) |y ∗ | ,
from which it is easy to derive constants β and γ.
Finally, if every trajectory of (3) is bounded, then every
trajectory has a nonempty bounded ω-limit set. Because of the
averaging property of stochastic matrices and the assumption
of uniform strong ergodicity, this ω-limit set is a subset of
the span of e. By part (i) of the theorem, the Lure system
has a globally asymptotically stable fixed point. Lemma 2 on
the other hand ensures that on span {e} the trajectories of (3)
and (9) (multiplied by e) coincide. It follows that restricted to
span {e}, the optimisation algorithm (3) has only one ω-limit
set, namely y ∗ e. It then follows from a continuity argument
that y ∗ e is a globally asymptotically stable fixed point of (3).
R EFERENCES
[1] E. Giordano, R. Frank, G. Pau, and M. Gerla, “CORNER: A realistic
urban propagation model for VANET,” in 7th IEEE Int. Conf. on
Wireless On-demand Network Systems and Services, 2010, pp. 57–60.
[2] G. Yan, S. Olariu, and M. C. Weigle, “Providing VANET security
through active position detection,” Computer Communications, vol. 31,
no. 12, pp. 2883–2897, 2008.
[3] A. Buchenscheit, F. Schaub, F. Kargl, and M. Weber, “A VANET-based
emergency vehicle warning system,” in IEEE Vehicular Networking
Conference (VNC2009), 2009, pp. 1–8.
[4] O. Carsten, F. Lai, K. Chorlton, P. Goodman, D. Carslaw, and S. Hess,
“Speed limit adherence and its effect on road safety and climate changefinal report,” Tech. rep., University of Leeds - Institute for Transport
Studies, Oct 2008.
[5] E. Adell, A. Varhelyi, M. Alonso, and J. Plaza, “Developing humanmachine interaction components for a driver assistance system for safe
speed and safe distance,” IET Intelligent Transport Systems, vol. 2, no. 1,
pp. 1–14, 2008.
[6] N. Hounsell, B. Shrestha, J. Piao, and M. McDonald, “Review of urban
traffic management and the impacts of new vehicle technologies,” IET
Intelligent Transport Systems, vol. 3, no. 4, pp. 419–428, 2009.
[7] N. Tradisauskas, J. Juhl, H. Lahrmann, and C. S. Jensen, “Map matching
for intelligent speed adaptation,” IET Intelligent Transport Systems,
vol. 3, no. 1, pp. 57–66, 2009.
[8] M. Barth and B. Kanok. (2010) ”Intelligent” Ways to Cut Transportation’s CO2 Emissions. University of California Transportation Center.
[Online]. Available: http://www.uctc.net/research/briefs/PB-2010-03.pdf
[9] D. Helbing and B. A. Huberman, “Coherent moving states in highway
traffic,” Nature, vol. 396, no. 6713, pp. 738–740, 1998.
[10] M. Barth, K. Boriboonsomsin, and G. Wu, “The potential role of vehicle
automation in reducing traffic-related energy and emissions,” in 2nd
IEEE International Conference on Connected Vehicles and Expo (ICCVE
2013), 2013, pp. 604–605.
[11] M. M. Waldrop, “Autonomous vehicles: No drivers required.” Nature,
vol. 518, no. 7537, pp. 20–23, 2015.
[12] P. Fernandes and U. Nunes, “Platooning of autonomous vehicles with
intervehicle communications in SUMO traffic simulator,” in 13th IEEE
International Conference onIntelligent Transportation Systems (ITSC
2010), 2010, pp. 1313–1318.
[13] Y. Gu, M. Liu, E. Crisostomi, and R. Shorten, “Highway speed limits,
optimised consensus, and intelligent speed advisory systems,” in 3rd
IEEE International Conference on Connected Vehicles and Expo (ICCVE
2014), 2014.
[14] R. H. Ordóñez-Hurtado, W. M. Griggs, K. Massow, and R. N. Shorten,
“Intelligent speed advising based on cooperative traffic scenario determination,” in Optimization and Optimal Control in Automotive Systems.
Springer, 2014, pp. 77–92.
[15] R. Gallen, N. Hautiere, A. Cord, and S. Glaser, “Supporting drivers in
keeping safe speed in adverse weather conditions by mitigating the risk
level,” IEEE Transactions on Intelligent Transportation Systems, vol. 14,
no. 4, pp. 1558–1571, 2013.
[16] R. M. Murray, “Recent research in cooperative control of multivehicle
systems,” Journal of Dynamic Systems, Measurement and Control, vol.
129, no. 5, pp. 571–583, 2007.
[17] F.-Y. Wang, “Parallel control and management for intelligent transportation systems: Concepts, architectures, and applications,” IEEE
Transactions on Intelligent Transportation Systems, vol. 11, no. 3, pp.
630–638, 2010.
[18] L. Garelli, C. Casetti, C. Chiasserini, and M. Fiore, “MobSampling: V2V
communications for traffic density estimation,” in 73rd IEEE Vehicular
Technology Conference (VTC Spring), 2011, pp. 1–5.
[19] V. Tyagi, S. Kalyanaraman, and R. Krishnapuram, “Vehicular traffic
density state estimation based on cumulative road acoustics,” IEEE
Transactions on Intelligent Transportation Systems, vol. 13, no. 3, pp.
1156–1166, 2012.
[20] W. J. Schakel and B. van Arem, “Improving traffic flow efficiency
by in-car advice on lane, speed, and headway,” IEEE Transactions on
Intelligent Transportation Systems, vol. 15, no. 4, pp. 1597–1606, 2014.
[21] R. Ordóñez-Hurtado, W. Griggs, K. Massow, and R. Shorten, “Evaluation of a new intelligent speed advisory system using hardware-in-theloop simulation,” in 2nd International Conference on Connected Vehicles
and Expo (ICCVE 2013), Dec 2013, pp. 945–946.
[22] S. Kato, S. Tsugawa, K. Tokuda, T. Matsui, and H. Fujii, “Vehicle
control algorithms for cooperative driving with automated vehicles
and intervehicle communications,” IEEE Transactions on Intelligent
Transportation Systems, vol. 3, no. 3, pp. 155–161, 2002.
[23] M. C. Choy, D. Srinivasan, and R. L. Cheu, “Cooperative, hybrid agent
architecture for real-time traffic signal control,” IEEE Transactions on
Systems, Man and Cybernetics, Part A: Systems and Humans, vol. 33,
no. 5, pp. 597–607, 2003.
[24] F. Flemisch, J. Kelsch, C. Löper, A. Schieben, J. Schindler, and
M. Heesen, “Cooperative control and active interfaces for vehicle
assistance and automation,” in FISITA World automotive congress, 2008.
[25] J. Hedrick, M. Tomizuka, and P. Varaiya, “Control issues in automated
highway systems,” IEEE Control Systems, vol. 14, no. 6, pp. 21–32,
1994.
[26] D. Swaroop and J. Hedrick, “String stability of interconnected systems,”
IEEE Transactions on Automatic Control, vol. 41, no. 3, pp. 349–357,
1996.
[27] S. Klinge and R. Middleton, “Time headway requirements for string
stability of homogeneous linear unidirectionally connected systems,” in
48th IEEE Conference on Decision and Control, 2009 held jointly with
the 2009 28th Chinese Control Conference (CDC/CCC 2009), Dec 2009,
pp. 1992–1997.
11
[28] S. Knorn and R. Middleton, “Stability of two-dimensional linear systems
with singularities on the stability boundary using LMIs,” IEEE Transactions on Automatic Control, vol. 58, no. 10, pp. 2579–2590, 2013.
[29] M. Liu, F. Wirth, M. Corless, and R. Shorten, “On global convergence
of consensus with nonlinear feedback, the Lure problem, and some
applications,” submitted to IEEE Transactions on Automatic Control,
2014.
[30] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction method
of multipliers,” Foundations and Trends R in Machine Learning, vol. 3,
no. 1, pp. 1–122, 2011.
[31] W. Shi, Q. Ling, K. Yuan, G. Wu, and W. Yin, “On the linear
convergence of the ADMM in decentralized consensus optimization,”
IEEE Transactions on Signal Processing, vol. 62, no. 7, 2014.
[32] T. Erseghe, D. Zennaro, E. Dall’Anese, and L. Vangelista, “Fast consensus by the alternating direction multipliers method,” IEEE Transactions
on Signal Processing, vol. 59, no. 11, pp. 5523–5537, 2011.
[33] F. Knorn, “Topics in cooperative control,” Ph.D. dissertation, National
University of Ireland Maynooth, Maynooth, Ireland, 2011.
[34] F. Knorn, R. Stanojevic, M. Corless, and R. Shorten, “A framework for
decentralised feedback connectivity control with application to sensor
networks,” International Journal of Control, vol. 82, no. 11, pp. 2095–
2114, 2009.
[35] F. Knorn, M. Corless, and R. Shorten, “Results in cooperative control
and implicit consensus,” International Journal of Control, vol. 84, no. 3,
pp. 476–495, 2011.
[36] L. Moreau, “Stability of multiagent systems with time-dependent communication links,” IEEE Transactions on Automatic Control, vol. 50,
no. 2, pp. 169–182, 2005.
[37] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz, “SUMOSimulation of Urban MObility,” in The Third International Conference
on Advances in System Simulation (SIMUL 2011), Barcelona, 2011.
[38] P. Boulter, T. Barlow, and I. McCrae, “Emission factors 2009: Report
3-exhaust emission factors for road vehicles in the United Kingdom,”
TRL Report PPR356. TRL Limited, Wokingham, 2009.
[39] R. Olfati-Saber, A. Fax, and R. M. Murray, “Consensus and cooperation
in networked multi-agent systems,” Proceedings of the IEEE, vol. 95,
no. 1, pp. 215–233, 2007.
[40] A. Olshevsky and J. N. Tsitsiklis, “Convergence speed in distributed
consensus and averaging,” SIAM Journal on Control and Optimization,
vol. 48, no. 1, pp. 33–55, 2009.
[41] W. M. Griggs and R. N. Shorten, “Embedding real vehicles in SUMO
for large-scale ITS scenario emulation,” in 2nd IEEE International
Conference on Connected Vehicles and Expo (ICCVE 2013), 2013, pp.
962–963.
[42] W. M. Griggs, R. H. Ordóñez-Hurtado, E. Crisostomi, F. Häusler,
K. Massow, and R. N. Shorten, “A Large-Scale SUMO-Based Emulation
Platform,” 2014, accepted in the IEEE Transactions on Intelligent
Transportation Systems.
[43] I. Hawkins. Torque pro. Available from Google Play. Last accessed
website on 23 July, 2013. [Online]. Available: https://play.google.com/
store/apps/details?id=org.prowl.torque
[44] PLX Devices Inc. California, USA. Last accessed website on 23 July,
2013. [Online]. Available: http://www.plxdevices.com
[45] http://gis.uba.de/website/umweltzonen/umweltzonen.php.
[46] http://cities-today.com/2014/01/sao-paulo-to-introduce-its-first-fleet-offully-electric-buses/.
[47] http://cleantechnica.com/2015/01/16/louisville-gets-first-fully-electriczero-emissions-city-buses/.
[48] http://www.siemens.com/innovation/en/home/pictures-of-thefuture/mobility-and-motors/electric-mobility-electric-buses.html.
[49] R. Van Haaren, “Assessment of electric cars’ range requirements and
usage patterns based on driving behavior recorded in the national
household travel survey of 2009,” in A Study conducted as part of the
Solar Journey USA project, 2011.
Mingming Liu received his double B.E. degrees
in Electronic Engineering with first class honours
from National University of Ireland Maynooth (now
Maynooth University) and Changzhou University in
2011. He is currently working towards a Ph.D. degree at the Hamilton Institute, Maynooth University
and he has also held a research assistant (pending viva) position with University College Dublin,
working with Prof. R. Shorten, since June 2015.
His current research interests are nonlinear system
dynamics, distributed control techniques, modelling
and optimisation in the context of smart grid and smart transportation systems.
Rodrigo H. Ordóñez-Hurtado received his degree in Engineering in Industrial Automatica from
the University of Cauca, Colombia, in 2005. He
commenced his Ph.D. in Electrical Engineering at
the University of Chile in March 2008, and between April-November 2012, he was an intern at
the Hamilton Institute, a multi-disciplinary research
centre established at the Maynooth University. In
November 2012, he received his Ph.D. degree from
the University of Chile, and between December
2012 - May 2015 he had a full-time postdoctoral
position at the Hamilton Institute working with Professor R. Shorten and his
research group. Currently, he has held a part-time postdoctoral position in
both Maynooth Unversity and University College Dublin since June 2015,
also working with Professor R. Shorten and his research group. Rodrigo’s
interests include robust adaptive systems (control and identification), stability
of switched systems, swarm intelligence, large-scale systems and intelligent
transportation systems. His focus is on applications to the mining industry
and transportation systems.
Fabian Wirth received his PhD from the Institute of
Dynamical Systems at the University of Bremen in
1995. He has since held positions in Bremen, at the
Centre Automatique et Systèmes of Ecole des Mines,
the Hamilton Institute at NUI Maynooth, Ireland, the
University of Würzburg and IBM Research Ireland.
He now holds the chair for Dynamical Systems at
the University of Passau in Germany. His current
interests include stability theory, queueing networks,
switched systems and large scale networks with
applications to networked systems and in the domain
of smart cities.
Yingqi Gu received her B.E. degree in Electronic
Engineering (first class honours) from Maynooth
University in 2013. She obtained her M.Sc degree
in Signal Processing and Communications at School
of Engineering, University of Edinburgh in 2014.
From February 2015, she commenced her Ph.D.
degree in the University College Dublin with Prof.
Robert Shorten. Her current research interests are
modelling, simulation and optimisation of intelligent
transportation systems.
12
Emanuele Crisostomi received the B.S. degree in
computer science engineering, the M.S. degree in automatic control, and the Ph.D. degree in automatics,
robotics, and bioengineering, from the University of
Pisa, Italy, in 2002, 2005, and 2009, respectively. He
is currently an Assistant Professor of electrical engineering with the Department of Energy, Systems,
Territory and Constructions Engineering, University
of Pisa. His research interests include control and optimization of large-scale systems, with applications
to smart grids and green mobility networks.
Robert Shorten graduated from University College
Dublin (UCD) in 1990 with a First Class Honours B.E. degree in Electronic Engineering. He was
awarded a Ph.D. in 1996, also from UCD, while
based at Daimler-Benz Research in Berlin, Germany.
From 1993 to 1996, Prof. Shorten was the holder of
a Marie Curie Fellowship at Daimler-Benz Research
to conduct research in the area of smart gearbox
systems. Following a brief spell at the Center for
Systems Science, Yale University, working with Professor K. S. Narendra, Prof. Shorten returned to
Ireland as the holder of a European Presidency Fellowship in 1997. Prof.
Shorten is a co-founder of the Hamilton Institute at NUI Maynooth, where
he was a full Professor until March 2013. He was also a Visiting Professor at
TU Berlin from 2011-2012. Professor Shorten is currently a senior research
manager at IBM Research Ireland. Prof. Shorten’s research spans a number
of areas. He has been active in computer networking, automotive research,
collaborative mobility (including smart transportation and electric vehicles),
as well as basic control theory and linear algebra. His main field of theoretical
research has been the study of hybrid dynamical systems.
| 3cs.SY
|
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
HARRY CRANE AND WALTER DEMPSEY
Abstract. Exchangeable models for countable vertex-labeled graphs cannot replicate the large sample behaviors of sparsity and power law degree
distribution observed in many network datasets. Out of this mathematical
impossibility emerges the question of how network data can be modeled in
a way that reflects known empirical behaviors and respects basic statistical
principles. We address this question by observing that edges, not vertices,
act as the statistical units in networks constructed from interaction data,
making a theory of edge-labeled networks more natural for many applications. In this context we introduce the concept of edge exchangeability,
which unlike its vertex exchangeable counterpart admits models for networks with sparse and/or power law structure. Our characterization of
edge exchangeable networks gives rise to a class of nonparametric models,
akin to graphon models in the vertex exchangeable setting. Within this
class, we identify a tractable family of distributions with a clear interpretation and suitable theoretical properties, whose significance in estimation,
prediction, and testing we demonstrate.
1. Introduction
Statistical network analysis is hamstrung by the lack of an inferential
framework that both facilitates sound conclusions from and replicates empirical properties of network data. Of the basic challenges facing network
analysis, Robins and Morris (2007) wrote,
A good [network] model needs to be both estimable from data and
a reasonable representation of that data[...]. Models that cannot
be estimated from the data, that cannot reproduce the data to some
adequate extent, models with implausible or empirically unsound
assumptions, [...] cannot be considered complete. (Robins and
Morris, 2007, p. 169)
Above all, this quote emphasizes that sound modeling precedes, and is
a crucial part of, methodological progress in the study of networks. The
importance of modeling is further underscored by the fact that much of
the current body of statistical theory and methods for networks has been
developed for a select class of tractable models, such as graphon models
Date: October 21, 2016.
Key words and phrases. interaction data; edge-labeled network; sparse network; power
law distribution; exchangeable random graph; scale-free network.
H. Crane is partially supported by NSF grants CNS-1523785 and CAREER DMS-1554092.
R code and links to datasets publicly available at https://github.com/wdempsey/e2.
1
2
HARRY CRANE AND WALTER DEMPSEY
Figure 1. (a) A network structure derived from some process
of interactions. Neither vertices nor edges come equipped
with labels. (b) Network obtained by labeling vertices of
network in Panel (a). (c) Network obtained by labeling
edges of network in Panel (a).
(Bickel and Chen, 2009; Gao et al., 2015; Klopp et al., 2016; Wolfe and
Olhede, 2014) and the stochastic blockmodel (Gao et al., 2016; Zhao et al.,
2011), which were not originally intended to handle complex networks and
tend to fit basic features of modern datasets rather poorly.
Here we initiate a framework for modeling and analyzing network data
constructed from interaction processes, such as communications (Klimt
and Yang, 2004; Opsahl and Panzarasa, 2009), collaborations (Barabási and
Albert, 1999; Newman, 2001), and relationships (McAuley and Leskovec,
2012; Zachary, 1977). In this context, we assume the relevant information
for the interactions is captured by an unlabeled network structure, as in
Figure 1(a), with the vertices corresponding to elements of a population and
edges representing interactions among adjacent vertices.
The convention in statistics, and network science more broadly, is to represent the structure in Figure 1(a) as a graph with labeled vertices, as in Figure
1(b). Out of this convention emerges the tendency—on grounds of logic
as well as expediency—to assume an exchangeable model, which assigns
equal probability to any two graphs that are equivalent up to relabeling
vertices; see Figure 2.
On one hand, exchangeability seems to correct for any arbitrariness in
labeling the vertices as well as narrow the analysis to a tractable class of
models. On the other hand, exchangeable random graph models cannot
replicate the widely observed empirical properties of sparsity and power
law degree distribution observed in many modern network datasets (Abello
et al., 1998; Barabási and Albert, 1999; Faloutsos et al., 1999; Jeong et al., 2001).
Paraphrasing a well known outcome from probability and combinatorics
(Diaconis and Janson, 2008; Lovász and Szegedy, 2006):
A realization from an exchangeable model for countable graphs is
sparse if and only if it has no edges.
Together, these observations prompt a foundational question of network
modeling:
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
3
Figure 2. Two different ways to label vertices of the same unlabeled network. Vertex exchangeable models assign equal
probability to both outcomes.
Is there a notion of probabilistic symmetry whose ergodic measures [...] describe useful statistical models for sparse graphs with
network properties? (Orbanz and Roy, 2015, p. 459)
We address this question as part of our broader development of edge
exchangeable network models, which are most appropriate when the edges are
the statistical units, as they are for the interaction processes we study. We
define this setup more precisely in Sections 2-3 and go on to establish basic
properties of edge exchangeable models throughout Sections 4-8.
2. Interaction data
Definition 2.1 (Interaction data). For a set P, we write fin(P) to denote the set
of all finite (ordered) multisets of P. An interaction process for a population P
is a correspondence I : I → fin(P) between a set I indexing interactions and finite
multisets of P.
The interaction processes in Definition 2.1 are the primary objects of our
study, for which we assume no additional information, such as covariates.
Incorporating covariates is notoriously difficult in network analysis, with
only a limited number of successful attempts to date (Airoldi et al, 2011;
Latouche et al, 2015; Mariadassou et al, 2010; Sweet, 2015; Tallberg, 2004;
Zhang et al, 2015). Butts (2008) and Perry and Wolfe (2013) discuss elements
of social theory in the context of interaction networks.
In Definition 2.1, each I(i) consists of the individuals involved in interaction i ∈ I. The ordering within I(i) captures directional interactions present
in some applications, for example, the sending of emails from one account
to a list of recipients. This setup contains the case in which interactions
are undirected, which we obtain by specifying a canonical ordering of the
elements in P and listing interactions according to that ordering.
Definition 2.1 captures the structure of many interaction datasets studied
in the networks literature. In the actors collaboration network (Barabási
and Albert, 1999; Rossi and Ahmed, 2015), for example, I indexes a set of
movies, with each movie i ∈ I corresponding to an interaction involving
the set of actors I(i) in its cast. In the Enron network (Klimt and Yang,
4
HARRY CRANE AND WALTER DEMPSEY
2004), interactions correspond to emails exchanged among Enron employees.
Crane and Dempsey (2016, Section B) list other common interaction datasets.
2.1. Network representation of interaction data. Though not inherently
structured as networks, interaction datasets are often depicted as in Figure
1. An interaction process I : I → fin(P) is commonly represented by a graph
GI = (P, E) with vertex set P and edge set E ⊆ P × P defined by
(1)
(v, v0 ) ∈ E if and only if {v, v0 } ⊆ I(i) for some i ∈ I;
that is, there is an edge between v and v0 in GI if v and v0 both participate
in at least one common interaction in I.
The representation in (1) may disregard important features of the data,
such as interactions involving more than two individuals and repeated
interactions among the same set of individuals. In the actors dataset, for
example, there are often more than two actors in each movie, nothing
precludes the same set of actors from being cast together in more than one
movie, and actors sometimes play more than one role in the same film.
These features of I : I → fin(P) could be more faithfully represented by
a hypergraph HI : fin(P) → {0, 1, . . .}, where
(2)
HI (A) = #{i ∈ I : I(i) = A},
A ∈ fin(P),
records the number of interactions corresponding to each multiset A. This
representation, however, presents another practical problem, as most statistical models are tailored specifically to graphs, with little or no easy extension
to handle hyperedges with multiplicity. See Kivelä et al. (2014) for some
discussion of hypergraph models in the context of multilayer networks.
2.2. Statistical units. The convention of representing network data as a
graph or hypergraph with labeled vertices implicitly, and in most cases
unintentionally, treats the vertices as the statistical units for the intended
application. To be clear, the mere assignment of labels to the observed
vertices, as in Figure 1(b), presents no immediate issue. The trouble arises
only when the data are assumed to be a partial observation of a larger
population process, as is often the case.
In this situation, the units are inherently tied to sampling, in that they
represent the smallest elements on which observations are taken. The act
of labeling vertices and treating the observed data as a graph G = (V, E), as
in (1), is consistent with the assumption that G is a partial observation of
a population graph G0 = (V 0 , E0 ) with V 0 ⊇ V and E0 ⊆ V 0 × V 0 , so that the
data G = G0 |V corresponds to the restriction of G0 to vertices V. In this way,
the observed data G, even if not explicitly stated or intended, is assumed to
contain information about all the edges between sampled vertices.
For interaction data, however, the interactions, and therefore edges, comprise the fundamental units of observation. For example, we may observe
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
5
the actors dataset by sampling movies without replacement from the Internet Movie Database1 (IMDB). Or, in the Enron dataset, we may observe only
those emails exchanged during a specific period of time, which constitutes a
sample from the collection of all emails exchanged in the company’s history.
In either case, the data result from a partial observation of the interactions.
For this purpose, we assume a population process I : N → fin(P) of a
potentially infinite collection of interactions indexed by the positive integers
N = {1, 2, . . .}. From I, we observe the restriction I|S : S → fin(P), s 7→ I(s),
to some finite sample S ⊂ N. The apparent role of interactions, and therefore
edges, as the statistical units suggests the alternative representation as an
edge-labeled network, as in Figure 1(c).
2.3. Edge-labeled networks. The foregoing discussion stresses the difference between the population process, as a collection of interactions I : N →
fin(P), and its representation as so-called network data in Figure 1.
Let I : N → fin(P) be an interaction process. Any bijection ρ : P → P0
induces an action fin(P) → fin(P0 ) by the map
(3)
s = (s1 , . . . , sr ) 7→ ρs = (ρ(s1 ), . . . , ρ(sr )) ∈ fin(P0 ).
With this, any bijection ρ : P → P0 acts on I : N → fin(P) by composition
of maps, I 7→ ρI, with ρI : N → fin(P0 ) given by
(ρI)(i) = ρ(I(i)),
(4)
i ∈ N,
as defined in (3). From this, we define the edge-labeled network induced by
I : S → fin(P) as the equivalence class
(5)
[
EI =
{I0 : S → fin(P0 ) : ρI = I0 for some bijection ρ : P → P0 }.
P0 :#P0 =#P
The equivalence class EI corresponds to an edge-labeled network structure,
as in Figure 1(c). Note that EI does not depend on the specific set P, and so
we may disregard P, or treat it implicitly as P = N, in our discussion. We
write ES for the set of networks with edges labeled in S ⊆ N.
For any S0 ⊆ S, we define the restriction of E ∈ ES to ES0 by E|S0 , the edgelabeled network obtained by removing any edges labeled in S \ S0 . If E = EI
for some interaction process I : S → fin(P), then E|S0 is the edge-labeled
network induced by the restricted process I|S0 : S0 → fin(P), s 7→ I(s).
Remark 2.2. For clarity we reserve the term graph to specifically refer to a vertexlabeled structure, such as the objects given in (1), (2), and Figure 1(b). We use the
term network for the generic unlabeled structure in Figure 1(a) and edge-labeled
network for the object defined in (5) and shown in Figure 1(c).
1http://www.imdb.com/
6
HARRY CRANE AND WALTER DEMPSEY
3. Network properties
For any edge-labeled network E, we define v(E) as the number of nonisolated vertices in E, that is, the number of vertices involved in at least
one of the edges of E. We also define Mk (E) as the number of k-ary edges
in E for each k ≥ 1, Nk (E) as the number of vertices that appear exactly
k times in E, and d(E) = (dk (E))k≥1 as the degree distribution of E, where
dk (E) = Nk (E)/v(E) is the proportion of vertices with degree k in E. For
example, the edge-labeled network E in Figure 1(c) has v(E) = 6, e(E) = 6,
Mk (E) = 6 for k = 2, Mk (E) = 0 for k , 2, and d(E) = (3/6, 1/6, 1/6, 1/6, 0, . . .).
Notice that these statistics do not depend on the edge labels assigned to E.
Definition 3.1 (Sparsity and power law degree distribution). Let (En )n≥1 be
a sequence of edge-labeled networks for which e(En ) → ∞ as n → ∞. The sequence
(En )n≥1 is sparse if
(6)
lim sup
n→∞
e(En )
= 0,
v(En )m• (En )
where m• (En ) = e(En
k≥1 kMk (En ) is the average arity of the edges in En .
The sequence (En )n≥1 exhibits a power law degree distribution if for some
γ > 1 the degree distributions (d(En ))n≥1 satisfy dk (En ) ∼ `(k)k−γ as n → ∞ for
all large k for some slowly varying function `(x), that is, limx→∞ `(tx)/`(x) = 1
for all t > 0, where an ∼ bn indicates that an /bn → 1 as n → ∞.
)−1
P
Remark 3.2. The slowly varying function `(k) makes the definition of power
law distribution robust to finite sample behavior and only affects the shape of the
distribution. It does not alter the tail behavior.
Our definition of sparsity in (6) refines the usual notion of sparsity for
vertex-labeled graphs, as defined, for example, by Nesetril and Ossona de
Mendez (2012). In the more familiar case when every interaction involves
exactly 2 vertices, m• (En ) = 2 for all n ≥ 1 and (En )n≥1 is sparse if
lim sup
n→∞
e(En )
= 0.
v(En )2
4. Edge exchangeable network models
Given E ∈ ES , for S ⊆ N, and a permutation σ : S → S, we write Eσ
to denote the edge-labeled network obtained by relabeling the edges of E
according to σ. Figure 3 demonstrates this operation visually.
More precisely, given an edge-labeled network E based on interaction
data I : S → fin(P) and a permutation σ : S → S, Eσ is the edge-labeled
network induced by Iσ : S → fin(P) with Iσ (i) = I(σ−1 (i)) for each i ∈ S.
(It is clear from (5) that the definition of Eσ does not depend on which
representative of the equivalence class EI is chosen when defining Eσ .)
Note the distinction between the action ρI defined in (4) and Iσ given here.
Whereas the action I 7→ ρI in (4) reassigns labels to the population through
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
7
Figure 3. Relabeling and restriction operations for edgelabeled networks. Panel (a) shows a network with edges
labeled 1, . . . , 6. Panel (b) shows the network from (a) with
edges relabeled according to permutation (1354)(26). Panel
(c) shows the restriction of the network in (a) to the edges
labeled in 1, . . . , 4.
a bijection ρ : P → P0 , the operation I 7→ Iσ corresponds to relabeling
interactions and E 7→ Eσ to relabeling edges.
We often write Y to denote a random edge-labeled network.
Definition 4.1 (Edge exchangeable network). A random edge-labeled network
Y ∈ ES is edge exchangeable if Yσ =D Y for all permutations σ : S → S, where
=D denotes equality in distribution.
Edge exchangeable models assign the same probability to all edge-labeled
networks that are isomorphic up to relabeling, as shown in Figure 4.
The assumption of edge exchangeability has a natural interpretation
in terms of sampling in that the interactions that determine an observed
edge-labeled network E ∈ ES are assumed to be representative of a larger
population of interactions I : N → fin(P). Contrast this with the assumption of exchangeability for vertex-labeled graphs, in which case sampled
vertices are assumed to be a representative sample of all vertices.
4.1. Characterization of edge exchangeable networks. We describe here
a special case a more general de Finetti-type representation theorem of
edge exchangeable networks, which we defer to Crane and Dempsey (2016,
Theorem A.2). The case below is most relevant to statistical applications.
To construct an edge exchangeable random network Y ∈ EN , we first
choose any f = ( fs )s∈fin(N) in the fin(N)-simplex,
X
F1 =
(
f
)
:
f
≥
0
and
f
=
1
.
s s∈fin(N)
s
s
s∈fin(N)
Each f ∈ F1 determines a distribution on finite multisets of N by
(7)
pr(X = s | f ) = fs ,
s ∈ fin(N),
which in turn determines an edge exchangeable network in EN as follows.
8
HARRY CRANE AND WALTER DEMPSEY
Figure 4. Two edge labelings of the network data from Figure 1. An edge exchangeable model assigns equal probability
to both networks. Vertices are not labeled and, therefore, vertex labels play no role in the exchangeability condition.
Let X1 , X2 , . . . be independent, identically distributed (i.i.d.) random multisets drawn from (7). Given X1 , X2 , . . ., define the random interaction process
X : N → fin(N) by X(i) = Xi and write EX to denote the corresponding
edge-labeled network obtained through (5), whose distribution we denote
by f . See Figure 5(a) for an illustration.
With probability 1, each s ∈ fin(N) occurs either zero or infinitely many
times in the interaction process X : N → fin(N) corresponding to X1 , X2 , . . .
chosen i.i.d. from (7), and the resulting edge-labeled network EX is edge
exchangeable. This is a special case of the more generic edge exchangeable
processes, for which interactions can occur zero, one, or infinitely many
times. We call an interaction that occurs only once a blip, and we call
an interaction process and its induced edge-labeled network blip-free if
it contains no blips. We state here the blip-free version of our general
representation theorem (Crane and Dempsey, 2016, Theorem A.2).
Theorem 4.2. Let Y ∈ EN be an edge exchangeable network that is blip-free with
probability 1. Then there exists a probability measure φ on F1 such that Y ∼ φ ,
where
Z
(8)
φ (·) =
f (·)φ(d f ).
F1
That is, every blip-free edge exchangeable network Y ∈ EN can be generated by
first sampling f ∼ φ and, given f , putting Y = EX with X : N → fin(N) defined
by X(i) = Xi for X1 , X2 , . . . i.i.d. from (7).
Remark 4.3. The measure φ in Theorem 4.2 is not unique. Uniqueness results
from a more technical treatment; see Crane and Dempsey (2016, Theorem A.2).
Corollary 4.4. The vertices in any blip-free edge exchangeable network Y ∈ EN
arrive in size-biased order weighted by the relative frequency of their occurrence.
According to Corollary 4.4, the assumption of edge exchangeability is
incompatible with vertex exchangeability, under which the vertices may be
interpreted as arriving in exchangeable random order. Edge exchangeable
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
9
Figure 5. Illustration of the generic generating process for
edge exchangeable networks described in (7) and Theorem
4.2. (a) Network representation of the interaction process
X1 = {2, 4}, X2 = {1, 2}, X3 = {1, 5}, X4 = {6, 9}, X5 = {2, 6},
X6 = {2, 6}. (b) Edge-labeled network obtained by removing
vertex labels from the network in Panel (a).
networks, therefore, do not admit a graphon representation in the sense of
Lovász and Szegedy (2006). But the generic construction of edge exchangeable networks in (7) and Theorem 4.2 suggests the following analog to the
graphon theory in terms of elements f ∈ F1 .
4.2. Nonparametric edge exchangeable models. Here we highlight the
tractable nonparametric subclass of vertex components models, for which we
construct f = ( fs )s∈fin(N) by specifying a probability distribution ν = {νk }k≥1
on the positive integers, choosing a sequence W = (Wi )i≥1 at random from
the infinite simplex
X
∆1 =
(s
,
s
,
.
.
.)
:
s
=
1
,
2
1
i
i≥1
and then defining
(9)
f(s1 ,...,sk ) = νk
k
Y
Ws j ,
(s1 , . . . , sk ) ∈ fin(N).
j=1
Under (9), we interpret the distribution in (7) as first choosing the size of a
random multiset s according to ν and, given #s = k, choosing the k elements
i.i.d. from (Wi )i≥1 . (Recall from Definition 2.1 that fin(N) consists of ordered
multisets. We recover the unordered case by writing each {s1 , . . . , sk } so that
s1 ≤ s2 ≤ · · · ≤ sk and multiplying (9) by the number of distinct ways to
order the elements of s, which can be computed explicitly.)
Stick-breaking representations, see, for example, Sethuraman (1994), offer
a computationally tractable way to fit the vertex components model to data.
In this construction, we can generate the sequence X1 , X2 , . . . of interactions
simultaneously with the vertex components W = (Wi )i≥1 .
Let {ϕi }i≥1 be a collection of probability densities on [0, 1]. We first
choose the size Kn ∼ ν of the nth directed interaction Xn = (Xn,1 , . . . , Xn,Kn )
10
HARRY CRANE AND WALTER DEMPSEY
independently for each n = 1, 2, . . .. We begin by choosing K1 ∼ ν, putting
X1,1 = 1, and sampling W1 ∼ ϕ1 . We continue inductively for each n =
1, 2, . . . as follows. Given Kn and k < Kn , we define Xn+1,≤k as the set of all
X1 , . . . , Xn and Xn+1,1 , . . . , Xn+1,k up to the kth chosen element of the (n + 1)st
interaction. Given Xn+1,≤k and W1 , . . . , WVn , where Vn = max(Xn+1,≤k ) is the
largest vertex label assigned among Xn+1,≤k , we choose Xn+1,k+1 , provided
Kn+1 ≥ k + 1, according to
(10)
(
W,
r = 1, . . . , Vn ,
P rn
pr(Xn+1,k+1 = r | Xn+1,≤k , W1 , . . . , WVn ) =
1 − Vj=1
W j , r = Vn + 1.
P n
If Xn+1,k+1 = Vn + 1, then we choose WVn +1 ∼ ϕn+1 (·/(1 − Vj=1
W j )). (The
PVn
division by 1 − j=1 W j represents the normalization by the remaining
length of the stick in the stick-breaking construction.) We continue to draw
Xn+1,k+2 , . . . , Xn+1,Kn+1 as in (10).
From the sequence X1 , X2 , . . . constructed above, we define Xn : [n] →
fin(N) as interaction process constructed by putting Xn (i) = Xi and defining
Yn = EXn to be the edge-labeled network induced by Xn for each n ∈ N. We
compute the joint density of Yn and W by
(11)
pr(Yn = E, (W1 , . . . , Wv(E) ) ∈ (dwi )1≤i≤v(E) ; {ϕi }i≥1 ) =
j−1
v(E)
Y
X
Dn (j)−1
wj
w
1 −
dw1 · · · dwv(E) ,
=
wi ϕ j
j
P
j−1
w
1
−
i=1
j=1
i=1 i
where Dn ( j) is the number of times the vertex corresponding to weight w j
P
appears in E and 0i=1 wi = 0.
Below we study a two parameter family of models corresponding to the
vertex components model in (11) with W generated from a Dirichlet process
(Ferguson, 1973). We explain this connection further in Section 5.2.
5. The Hollywood model
Though not limited to any specific application, we phrase the following
Hollywood model in the terms of the actors collaboration network to aid
comprehension. In this description, vertices correspond to actors and each
edge contains the set of actors involved in the corresponding movie.
Let ν = {νk }k≥1 be a probability distribution on the positive integers and
let (α, θ) satisfy either
• 0 < α < 1 and θ > −α (infinite population) or
• α < 0 and θ = −kα for some k = 1, 2, . . . (finite population).
We generate a sequence of interactions X1 , X2 , . . . as follows.
Given X1 , . . . , Xn−1 , we choose the number of roles in the next movie
independently according to Kn ∼ ν. Given Kn = k, we choose the k actors
Xn,1 , . . . , Xn,k in order of their prominence, first filling the lead role, then
the second lead role, and so on until all k roles are filled. Let Vn ( j) be the
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
11
number of unique actors seen in all the movies through the (j − 1)st actor
cast in movie n. (Thus, Vn (1) is the number of unique actors appearing
in movies 1, . . . , n − 1.) For each i = 1, . . . , Vn ( j), we write Dn,j (i) to denote
the number of roles for which the actor labeled i has been cast up to and
including the (j − 1)st role of movie n. (Note that an actor may play more
than one role in a given movie.) The actor Xn,j cast in the jth role of movie n
is chosen randomly among the actors labeled 1, . . . , Vn ( j) and a previously
unseen actor, labeled Vn ( j) + 1, according to
(12)
(
Dn, j (i) − α, i = 1, . . . , Vn ( j),
pr(Xn, j = i | X1 , . . . , Xn−1 , Xn,1 , . . . , Xn, j−1 ) ∝
θ + αVn (j), i = Vn (j) + 1.
We update according to (12) until all k roles of movie n have been filled.
Each interaction X1 , X2 , . . . records the order in which actors are selected
in each movie. From X1 , X2 , . . ., we define Xn : [n] → fin(N) by Xn (i) = Xi
for i = 1, . . . , n, and we write Yn = EXn for the edge-labeled network induced
by Xn as defined in (5). We call the resulting sequence (Yn )n≥1 the Hollywood
process with parameter (α, θ, ν), which determines a family of distributions
on EN , called the Hollywood model.
The Hollywood model has a closed form expression for random edgelabeled networks of each finite size n ≥ 1 given by
pr(Yn = E; α, θ, ν) =
∞
Y M (E)
(θ/α)↑v(E) Y
(13)
νk k αv(E)
exp{Nk (E) log((1 − α)↑(k−1) )},
=
↑m(E)
θ
k≥1
k=2
where E ∈ E[n] , v(E) is the number of non-isolated vertices in E, (Nk (E))k≥1
gives the number of vertices with
P each degree k ≥ 1, Mk (E) is the number
of k-ary edges in E, m(E) = k≥1 kMk (E) is the total degree of E, and
x↑j = x(x + 1) · · · (x + j − 1) is the ascending factorial function.
Though defined for edge-labeled networks with directed edges, the Hollywood model determines a distribution on undirected edge-labeled networks by ignoring the edge orientations. In this case, the probability of
an undirected network E∗ is given by the expression in (13) multiplied by
a combinatorial factor C(E∗ ) that counts the number of directed networks
corresponding to E∗ . While we have no closed form expression for C(E∗ ),
the quantity plays no role in our inferences because it does not depend on
the parameters (α, θ, ν), as we discuss further in Section 6.1.
5.1. Interpretation of parameters and finite population model. The split
parameter space of the Hollywood model enables both bounded and unbounded population sizes. The region 0 < α < 1 and θ > −α gives rise to a
sequence (Yn )n≥1 for which v(Yn ) → ∞ almost surely (a.s.) as n → ∞, as is
reasonable to assume for the actors, Enron, and Wikipedia networks. The
Karate Club dataset, on the other hand, is known to have exactly thirty-four
12
HARRY CRANE AND WALTER DEMPSEY
club members but no limit on the number of interactions between individuals. The range α < 0 and θ = −kα accommodates this case by describing an
edge exchangeable sequence (Yn )n≥1 for which v(Yn ) → k a.s. as n → ∞.
By (12), α > 0 increases the probability of observing previously unseen
vertices but decreases the probability of observing a vertex again after
its initial occurrence. The range α < 0 has the opposite effect. Thus, α
values near 1 make it more likely that new edges involve previously unseen
vertices, but less likely that previously seen vertices occur in future edges.
On the other hand, α < 0 corresponds to a finite population size, so that
each newly observed vertex decreases the number of unseen vertices and
increases the probability that future edges involve previously seen vertices.
In the 0 < α < 1 regime, larger values of θ increase the probability of seeing
previously unobserved vertices in new edges, but the effect of θ diminishes
as n → ∞. In Section 5.3, we see that 0 < α < 1 is directly related to the
sparsity and power law behavior of the sequence (Yn )n≥1 .
5.2. Connection to vertex components model. Crane (2016b) previously
noted a connection between the Hollywood model and the Ewens–Pitman
two parameter family of distributions on set partitions (Ewens, 1972; Pitman,
2005). The two models coincide in the unary setting of the Hollywood model,
that is, the (α, θ, ν) case when ν1 = 1. In this sense, the Hollywood model
may be viewed as a natural refinement of the Ewens–Pitman distribution
and Chinese restaurant process, which enjoys wide relevance throughout
statistics, mathematics, and applied science (Crane, 2016a).
For α < 0 and θ = −kα, the Hollywood model corresponds to the vertex
components model with W = (W1 , . . . , Wk ) chosen from the symmetric
Dirichlet distribution with parameter (α, . . . , α) on the (k − 1)-simplex.
For 0 < α < 1 and θ > −α, the Hollywood model is a special case of
the vertex components model with W = (Wi )i≥1 chosen from the Griffiths–
Engen–McCloskey (GEM) distribution with parameter (α, θ) on ∆1 ; see
Feng (2010), also Crane (2016a), for further details on the GEM distribution
and its relationship to the Poisson–Dirichlet distribution. Alternatively, the
Hollywood model with 0 < α < 1 and θ > −α can be constructed by noting
the stick-breaking construction of the GEM distribution and taking ϕ j to be
the density of the Beta distribution with parameter (1 − α, θ + jα) for each
j ≥ 1 (Pitman, 2005). We recover (13) by marginalizing over W in (11).
5.3. Statistical properties of the Hollywood model. The close connection between the Hollywood model and the Poisson–Dirichlet distribution
through the vertex components interpretation of Section 5.2 addresses the
main questions posed in Section 1. Theorems 5.2 and 5.3, in particular,
follow from the corresponding power law behavior of the Ewens–Pitman
process (Crane, 2016a; Pitman, 2005).
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
13
Theorem 5.1. The Hollywood model with parameter (α, θ, ν) determines an edge
exchangeable probability distribution on EN for all (α, θ, ν) in the parameter space
of the model.
Theorem 5.2. Let (Yn )n≥1 obey the Hollywood process with parameter (α, θ, ν).
For each n ≥ 1, let pn (k) = Nk (Yn )/v(Yn ), k ≥ 1, be the empirical degree
distribution of Yn , where Nk (Yn ) is the number of vertices with degree k ≥ 1 and
v(Yn ) is the number of non-isolated vertices in Yn . Then, for every k ≥ 1,
pn (k) ∼ αk−(α+1) /Γ(1 − α)
a.s.
as n → ∞,
R∞
where Γ(t) = 0 xt−1 e−x dx is the gamma function. That is, (Yn )n≥1 has a power
law degree distribution with exponent γ = 1 + α ∈ (1, 2).
Theorem 5.3. Let (Yn )n≥1 obey the Hollywood process with parameter (α, θ, ν)
for 0 < α < 1 and θ > −α. Then the expected number of vertices in Yn satisfies
(14)
E(v(Yn )) ∼
Γ(θ + 1)
(µn)α
αΓ(θ + α)
as n → ∞,
P
where µ = k≥1 kνk is the mean edge arity. Furthermore, if 1/µ < α < 1, then
(Yn )n≥1 is almost surely sparse in the sense of (6).
The power law behavior in the range 1 < γ < 2 of the Hollywood model
complements that of the preferential attachment model (Chung and Lu,
2006). Some authors (Barabási and Albert, 1999) suggest that γ > 2 is more
prevalent in datasets that exhibit power law, but more recent work by Crane
and Dempsey (2015) demonstrates empirically that the range 1 < γ < 2 of
the Hollywood model is common for interaction networks.
For a more general class of edge exchangeable models that admits sparsity
and power law, we define the α-diversity, 0 < α < 1, of an edge exchangeable
process (Yn )n≥1 by
(15)
Sα = lim v(Yn )/e(Yn )α ,
n→∞
provided the limit exists and is strictly positive and finite almost surely.
Following Pitman (2005, Lemma 3.11), an edge exchangeable process (Yn )n≥1
from the vertex components model constructed from ν = {νk }k≥1 and W =
(Wi )i≥1 has α-diversity if and only if the reordered sequence W(1) ≥ W(2) ≥
· · · ≥ 0 satisfies W(i) ∼ Zi−1/α a.s. as i → ∞ for a random variable
0 < Z < ∞.
P
In this case, (Yn )n≥1 is sparse for 1/µ < α < 1, with µ = k≥1 kνk , and has
power law degree distribution with γ = α + 1.
5.4. Projecting to a network without multiple edges. Though multiple
edges occur naturally in networks constructed from interaction processes,
most network models are designed to handle only simple graphs, and many
network datasets record only a single edge to indicate the occurrence of
some positive number of interactions, as in (1). The fact that many network
datasets are obtained by thresholding edge multiplicities is often ignored
14
HARRY CRANE AND WALTER DEMPSEY
during data analysis; and the significance of this action on inference is
underappreciated in the broader literature.
We define the (t, c)-projection of an interaction process I : N → fin(P) by
t,c
HI : fin(P) → {0, 1}, where
(
1, t({i ∈ N : I(i) = A}) > c,
t,c
(16)
HI (A) =
A ∈ fin(P),
0,
otherwise,
for some thresholding function t and cutoff value c ≥ 0. We define the
standard projection HI∗ by taking t(A) = #A and c = 0 in (16).
We call (Yn )n≥1 a binary Hollywood process with parameter (α, θ) if it follows
the Hollywood process with parameter (α, θ, ν) having ν2 = 1.
Theorem 5.4 (Sparsity of projected network data). Let (Yn )n≥1 obey the binary
Hollywood process with parameter (α, θ). For each n ≥ 1, let Hnt,c be the (t, c)projection of Yn by applying (16) for any c ≥ 0 and t(A) = #A, the cardinality map.
Then the sequence (Hnt,c )n≥1 is sparse almost surely for all 0 < α < 1 and θ > −α.
Theorem 5.4 makes clear that projecting not only discards information
but may also alter the assumed behavior of the data. By Theorems 5.3 and
5.4, (Yn )n≥1 generated from the binary Hollywood model is sparse only for
1/2 < α < 1, while the projected network after applying (16) is sparse for all
0 < α < 1. We discuss these implications further in Section 6.4.
On the other hand, Figure 6 suggests that the power law behavior of
the Hollywood model, as established in Theorem 5.2, might be preserved
under projection. Whether this phenomenon is real or perceived, there
remains no logical justification for projecting interaction data to a simple
graph, especially when this operation makes the otherwise easy practice
of parameter estimation intractable, as we discuss further surrounding
(21) below. We also note that thresholding an edge exchangeable network
according to (16) does not preserve edge exchangeability.
6. Inference from edge exchangeable models
6.1. Maximum likelihood estimation. Given edge-labeled network data
Yn with n edges, the log-likelihood `(α, θ, ν; Yn ) based on (13) with parameter (α, θ, ν) satisfies
(17)
v(Y
∞
n )−1
X
X
`(α, θ, ν; Yn ) =
Mk (Yn ) log νk + v(Yn ) log(α) +
log(θ/α + j)−
j=0
k=1
−
m(Y
Xn )
j=1
log(θ + j − 1) +
∞
X
k=2
Nk (Yn )
k−2
X
log(1 − α + j),
j=0
where Mk (Yn ) is the number of edges in Yn with exactly k vertices,
Nk (Yn )
P∞
is the number of vertices in Yn with degree k, and m(Yn ) = k=1 kMk (Yn ) is
the total degree of Yn .
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
*
*
* * ** *
* * *
** * ** *
* * ******** ******* ***
* * ** *******************************
0
1
2
3
4
log proportion
γ = 1.67
(A2)
−6 −5 −4 −3 −2
−9 −8 −7
*
(A1)
*
−11
log proportion
*
5
*
γ = 1.67
*
** * *
* ** *
* ** *
* ***
*
* * ******* * * * ** * ** *
* ***********************************
0
1
2
log degree
5
2
4
6
−2
−4
*
*
−6
log proportion
**
γ = 1.25
**
**
***
** * *
*******
* ***********
******************************************** **
********************************************************
(B2)
*
−8
−4
*
−6
log proportion
4
*
(B1)
*
−10 −8
3
log degree
*
0
15
0
log degree
γ = 1.25
****
*
**
*******
************** **** ** * * **
*******************************************************************
2
4
6
log degree
Figure 6. Simulated degree distribution of networks and
their standard projection. Plots show the degree distribution
of network obtained from (A1) the binary Hollywood model
with (α, θ) = (0.67, 1), (B1) the binary Hollywood model with
(α, θ) = (0.25, 1), (A2) the standard projection of the network
in (A1), and (B2) the standard projection of the network in
(B1). The line with slope −γ in (A1) and (B1) indicates the
true power law based on Theorem 5.2. The line with slope
−γ in (A2) and (B2) is the conjectured power law based on
visual evidence and the connection to Theorem 5.2.
Maximum likelihood estimation of ν returns the empirical distribution
ν̂MLE = {ν̂k }k≥1 , where ν̂k = Mk (Yn )/n for each k = 1, 2, . . .. We then estimate
α and θ by iterating between the score functions
v(Y
∞ k−2
n )−1
X
∂`(α, θ, ν; Yn ) v(Yn )
−θ/α2 X X Nk (Yn )
(18)
=
+
−
=0
α
θ/α + j
1−α+ j
∂α
and
k=2 j=0
j=0
v(Y
m(Y
n )−1
n )−1
X
X
∂`(α, θ, ν; Yn )
1/α
1
(19)
=
−
= 0.
θ/α + j
θ+ j
∂θ
j=0
j=0
We encounter no convergence issues when iterating between (18) and (19)
for our applications below.
The distribution in (13), and therefore the log-likelihood in (17), applies
to the oriented network generated by the Hollywood model with parameter
(α, θ, ν). With Yn∗ denoting the undirected edge-labeled network obtained
16
HARRY CRANE AND WALTER DEMPSEY
by removing the orientations from Yn , for each n ≥ 1, we have
pr(Yn∗ = E∗ ; α, θ, ν) =
∞
↑v(E∗ ) Y
Y M (E∗ )
∗
v(E∗ ) (θ/α)
k
(20) C(E )
νk
exp{Nk (E∗ ) log((1 − α)↑(k−1) )},
α
∗
θ↑m(E )
k≥1
k=2
where C(E∗ ) is the combinatorial factor that counts the number distinct ways
to orient the edges of E∗ to obtain a directed network. The log-likelihood
based on Yn∗ satisfies
˜ θ, ν; Yn∗ ) = `(α, θ, ν; Yn ) + log C(Yn∗ ),
`(α,
where `(α, θ, ν; Yn ) is the log-likelihood from (17) with Yn taken to be any
oriented edge-labeled network whose edges agree with Yn∗ . Thus, the score
functions based on Yn∗ are just as in (18) and (19) and maximum likelihood
estimation for (α, θ, ν) can be performed without issue.
Problems arise, however, when projecting multiple edges according to
the operation in (16). For definiteness, suppose Hn∗ is the standard projection
from the binary Hollywood model with parameter (α, θ) on E[n] as in (13).
The log-likelihood `∗ (α, θ, ν; Hn∗ ) in this case satisfies
exp{`∗ (α, θ, ν; Hn∗ )} =
∞
X Y
(θ/α)↑v(E) Y
Mk (E)
v(E)
α
exp{Nk (E) log((1 − α)↑(k−1) )}
=
νk
↑m(E)
θ
E≥Hn∗ k≥1
k=2
∗
∗
∞
Y M (H∗ ) αv(Hn ) (θ/α)↑v(Hn ) Y
νk k n
=
exp{Nk (Hn∗ ) log(1 − α)↑(k−1) } ×
↑m(Hn∗ )
θ
k≥1
k=2
∗)
Q
M
(E)−M
(H
k
k
n
X
k≥1 νk
×
exp{(Nk (E) − Nk (Hn∗ )) log(1 − α)↑(k−1) },
∗ ))↑(m(E)−m(Hn∗ ))
(θ
+
m(H
n
E≥H∗
n
Hn∗
where E ≥
indicates that E ∈ E[n] is an edge-labeled network whose
standard projection is Hn∗ . The log-likelihood has the form
(21)
`∗ (α, θ, ν; Hn∗ ) = log C∗n (Hn∗ ; α, θ, ν) + `(α, θ, ν; Hn∗ ),
so that without a manageable expression for C∗n (Hn∗ ; α, θ, ν), maximum likelihood estimation based on projected network data Hn∗ is intractable.
6.2. Application to Wikipedia voting and Karate Club networks. Table 1
shows the maximum likelihood estimates for the binary Hollywood model
fit to the Wikipedia voting and Karate Club networks described in Crane
and Dempsey (2016, Section B). Together these examples cover both regimes
in the parameter space of the Hollywood model.
The Karate Club network consists of interactions among thirty-four club
members, warranting the choice of k = 34 when fitting the model with
parameters α < 0 and θ = −34α. The Wikipedia network, by contrast, has
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
17
Maximum likelihood estimates
network
α̂MLE std. error∗ θ̂MLE std. error∗
Wikipedia
0.37
0.01
183
14.76
0.47
61.3
16.04
Karate Club -1.80
Table 1. Maximum likelihood estimates and standard errors
for (α, θ) from binary Hollywood model fit to Wikipedia voting and Karate Club networks. ∗ Standard errors estimated
by the Cramér–Rao lower bound, which we have verified as
a good estimate for the standard error based on simulation.
no upper limit on the number of vertices, and so we fit the model under
regime 0 < α < 1 and θ > −α.
The large standard error for maximum likelihood estimates of θ agrees
with what is known about estimation of the mutation rate in Ewens’s
sampling formula (Crane, 2016a; Ewens, 1972): although θ̂MLE → θ almost
surely as the sample size grows, it converges at a rate on the order of log(n),
rendering it inconsistent for practical purposes. The estimate of α is of
greater interest in the applications we envision because of its relationship to
the power law behavior, cf. Theorem 5.2. When α < 0 and the total number
of vertices is taken to be known and finite k = 1, 2, . . ., then θ = −kα and we
obtain θ̂MLE = kα̂MLE . In this case, we have s.e.(θ̂MLE ) = k × s.e.(α̂MLE ), with
s.e.(·) shorthand for the standard error.
By Theorem 5.2, the maximum likelihood estimate α̂MLE = 0.37 for the
Wikipedia dataset implies an estimated power law of γ̂MLE = 1.37, which
is reasonably close to the estimated power law exponent of γ̂YULE = 1.44
obtained by fitting degree distribution of the network to the Yule model,
(22)
pr(K = k; γ) = (γ − 1)Γ(γ)/Γ(k + γ),
k ≥ 1,
for γ > 1. The likelihood surface is continuous in α, θ, except at α = 0,
and estimation experienced no instability or convergence issues in multiple
iterations from different starting points.
6.3. Prediction using growth dynamics. The growth dynamics of edge exchangeable models by the sequential process of edge addition facilitates
predictive inferences in networks generated by a process of repeated interactions. In the setting of Section 5, we can predict the next interaction,
based on an observed network Yn , from the update probabilities in (12) with
(α, θ, ν) = (α̂MLE , θ̂MLE , ν̂MLE ) given by the maximum likelihood estimates
obtained from (17).
For a concrete application, we consider the actors collaboration network
from Barabási and Albert (1999). For (α, θ, ν) in the parameter space of the
Hollywood model, we compute the predictive probability, conditional on
Yn , that the next movie contains an actor not previously observed in the
sample. A straightforward calculation based on (12) and the law of total
18
HARRY CRANE AND WALTER DEMPSEY
Maximum likelihood estimates
k
ν̂k
k
ν̂k
1 0.081 7 0.057
2 0.071 8 0.054
3 0.065 9 0.046
4 0.072 10 0.275
5 0.062 11 0.161
6 0.059
Table 2. Maximum likelihood estimates of the movie size
distribution for the actors collaboration network. The mean
number of actors in each movie is 7.136.
probability gives
(M − αN)↑k
,
(θ + M)↑k
k≥1
P
P
where N = k≥1 Nk (Yn ) is the total degree of Yn and M = k≥1 kNk (Yn ) is
the total number of roles in Yn . (To see the calculation in (23), we note that
(M − αN)↑k /(θ + M)↑k is the conditional probability that the next movie does
not feature a new actor, given that it has k actors in its cast. Summing over
all k ≥ 1 gives the total probability that the next movie does feature a new
actor, yielding the rightmost term in (23). The probability that a new actor
appears follows by taking the complementary probability.)
Fitting the model to the actors collaboration dataset yields α̂MLE = 0.66
(s.e. 6.8 × 10−4 ) and θ̂MLE = 4.21 (s.e. 2.86), with ν̂MLE given in Table 2. Since
α̂MLE = 0.66 lies in the range 1/µ < α < 1, with µ = 7.136 computed from
Table 2, Theorem 5.3 suggests that the actors collaboration network is sparse.
This estimate also agrees exactly with the estimated fit γ̂YULE = 1.66 if the
degree distribution is fit directly to the Yule distribution in (22).
The estimated predictive probability based on these maximum likelihood
estimates is 0.78. We check the accuracy of this estimate by data-splitting
cross validation based on samples of 2, 000 movies from the total collection
of about 200, 000. For each iteration j = 1, 2, . . ., we sample 2, 000 movies
uniformly at random from the 200, 000 to obtain Y2000 . We then compute
the probability in (23) based on the estimates α̂MLE , θ̂MLE , and ν̂MLE from
Y2000 and compare this estimated probability to the empirical probability
computed as the proportion of the 198, 000 unsampled movies for which
there appears an actor not among the 2, 000 movies in the sample Y2000 . The
mean relative error between the fitted probability (23) and the empirical
probability for 100 iterations was −0.003 with a standard deviation of 0.002.
(23)
pr(new actor in next movie | Yn ; α, θ, ν) = 1 −
X
νk
6.4. Tests for sparsity and power law. As asymptotic network properties,
sparsity and power law cannot be verified with certainty based on any finite
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
19
amount of data. Statistical tests for sparsity and/or power law based on
finite sample data, therefore, require that the finite sample models faithfully
represent the properties exhibited by the population network.
Our discussion in Section 5.4 advises caution when testing for asymptotic
properties based on thresholded network data. If the parameter space Π
partitions as Π = Π0 ∪ Π1 such that π ∈ Π1 parameterizes sparse networks
and π ∈ Π0 parameterizes non-sparse networks, then H0 : π ∈ Π0 versus
H1 : π ∈ Π1 might yield a valid test for sparsity. Alternative interpretations
of the partition Π = Π0 ∪ Π1 , however, may provide a more parsimonious
conclusion. For example, under the binary Hollywood model of Section
5, the region 1/2 < α < 1 corresponds to sparse networks while α <
1/2 parameterizes dense networks, with α < 0 corresponding to the case
where the number of vertices stays bounded and finite. By Theorem 5.4, if
the network is thresholded by the standard projection, then the projected
network is sparse a.s. for all α > 0. In this case, the partition Π = {α <
0} ∪ {0 ≤ α ≤ 1} offers the more parsimonious interpretation according to
whether the population of vertices is finite (α < 0) or infinite (0 ≤ α ≤ 1)
rather than sparse (0 ≤ α ≤ 1) or not sparse (α < 0).
A particular consequence of this is on display when testing for sparsity
in the US airport dataset (Opsahl, 2011), which is built from the flight map
between all US airports in 2010. For each pair of airports, there are as
many edges as there were seats on all flights between those airports. The
edge weights range over several orders of magnitude from a minimum of
1 to several hundred thousand, making the projection in (16) particularly
deleterious to the data structure. For the binary Hollywood model fit to
this dataset, there are two possible maximum likelihood estimates, both
with log-likelihood of about −8.21 × 10−9 : if α > 0 we get α̂MLE,1 = 0.13 (s.e.
0.003) and θ̂MLE,1 = 0.08 (s.e. 0.134), and if α < 0 we get α̂MLE,2 = −0.11 (s.e.
0.003) by taking k = 1574 the number of airports in the sample. The estimate
α̂MLE,1 = 0.13 is consistent with γ̂YULE = 1.11 obtained by fitting (22) to the
degree distribution of the airport network. The choice of regime α < 0 or
α > 0 is a matter of whether we believe the population of airports under
consideration is fixed and finite or potentially unbounded, a choice which
is closely tied to the desired hypothesis test.
Technically, it makes sense to test for sparsity only if the population is
infinite, in which case α > 0 and we can test H0 : 0 < α ≤ 1/2 against
H1 : 1/2 < α < 1. From the estimate α̂MLE,1 = 0.13, we cannot reject the
hypothesis H0 : α < 1/2. But if testing based on the projected network, then
the above test is degenerate since the ranges for which the population is
infinite and for which the population is sparse coincide, per Theorem 5.4.
By contrast, Caron and Fox (2014, Section 7.2) conclude that the US airport
network is sparse based on a 99% credible interval of [0.099, 0.181] for an
analogous parameter to α in our case, but that analysis appears to be based
on the standard projection.
20
HARRY CRANE AND WALTER DEMPSEY
7. Discussion of other approaches
The above discussion offers ample reason to prefer the edge exchangeable
approach over other common models for network datasets constructed from
interaction processes. The observation that the interactions, and therefore
edges, comprise the units leads naturally to the alternative representation
by edge-labeled networks. Beyond treating the units appropriately, representing the data by an edge-labeled network frees us of the limitations of
the conventional representation by vertex-labeled graphs. The Hollywood
model, in particular, is computationally tractable, performs admirably in
several real data examples, and replicates the key features of sparsity and
power law degree distribution of primary interest in the modern network
science literature. The nonparametric vertex components model in Section
4.2 offers an even larger class of models with likely computational benefits
due to its stick-breaking construction.
To round out the discussion, we compare the performance of the Hollywood model to that of the exponential random graph model (ERGM)
(Holland and Leinhardt, 1981) and stochastic blockmodel (SBM) (Holland
et al., 1983; Snijders and Nowicki, 1997) fit to the actors collaboration network. We then discuss some treatments of sparse networks by other authors.
7.1. Empirical comparison. We compare the performance of the Hollywood model (HW), stochastic blockmodel (SBM), Erdős–Rényi model (ER),
and exponential random graph model (ERGM) on the actors collaboration
dataset (Barabási and Albert, 1999), which consists of approximately 200, 000
movies. We have already fit HW to the full movie dataset in Section 6.3, but
we could not compare to other methods on the full dataset because the best
available software for SBM2 and ERGM3 are computationally inefficient for
networks of even moderate size.
For comparison, we subsample 250 movies uniformly at random from
the database of 200, 000 movies. Since SBM, ER, and ERGM are models for
graphs without multiple edges, we treat the de facto population network as
the graph G250 induced by these 250 movies as in (1). The maximum likelihood estimate (MLE) under SBM estimates 21 blocks and log-likelihood of
−32, 818 and MLE under ER returns a log-likelihood of −42, 924. The ERGM
with sufficient statistics given by edge and triangle density, respectively,
ε(G) =
2
e(G)
v(G)(v(G) − 1)
and τ(G) =
6
∆(G),
v(G)(v(G) − 1)(v(G) − 2)
where ∆(G) is the number of triangles in G, returns a model degeneracy
error.
We also fit ERGM and ER to the induced graph (1) obtained from a
1, 000 movie subsample. In this case, the fitted parameters for ERGM are
2https://cran.r-project.org/web/packages/blockmodels/blockmodels.pdf
3https://cran.r-project.org/web/packages/ergm/ergm.pdf
*
(B)
5
2
3
4
5
4
*
*
3
*
*
2
log(proportions)
4
0
1
* **
***
*****
******
******
*****
*****
********
********
*********
***********
*************** *
** *************
* *****************
*********************** *
*
*
* *
*
*
1
**
*
6
8
*
2
log(proportions)
(A)
21
* *
*
*
6
*
*
0
*
0
*
6
*
10
12
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
0.0
0.5
1.0
log(degree)
1.5
2.0
2.5
3.0
log(degree)
8
*
*
(C)
(D)
*
3
2
*
*
* *
*
*
1.0
1.5
log(degree)
2.0
2.5
6
* **
3.0
*
*
****
*
**
* **
** *
*
***
0
0
0.5
*
*
*
*
0.0
*
4
log(proportions)
4
*
*
* * * **
2
*
*
*
*
* *
1
log(proportions)
5
6
*
0
1
2
3
*
4
log(degree)
Figure 7. Comparison of fitted degree distribution (- - -) to
empirical degree distribution (*) for (A) Hollywood model
(compared to full dataset), (B) Erdős–Rényi model and (C)
stochastic blockmodel (compared to 250 movie sample), and
(D) ERGM (red) and ER (black) (compared to 1, 000 movie
sample).
β̂ε = −12.26 (s.e. 0.083) and β̂τ = 5.24 (s.e. 0.019) with log-likelihood of
−190, 630, which improves over Erdős–Rényi (log-likelihood: −208, 702).
By contrast, we fit the Hollywood model to both the 250 and 1, 000 movie
subsamples and obtain a maximum log-likelihood of −476 and −26, 320,
respectively.
Figure 7 compares the goodness of fit for the degree distribution for HW,
SBM, ER, and ERGM. The fitted degree distribution for HW in (A) is based
on the estimate obtained from the sample of 250 movies compared to the
degree distribution of the complete population network on 200, 000. In (B)
and (C), we compare the fitted ER and SBM to the degree distribution of
subsampled network G250 induced by (1) and in (D) we compare the fitted
ER and ERGM to the degree distribution of G1000 based on 1, 000 movies
through (1). In this sense, the comparison is generous to SBM, ER, and ERGM
since those are compared only to the data to which they were fit while HW is
evaluated based on how well it detects the population structure for 200, 000
edges based on a sample of 250 movies. Even with this disadvantage, the
Hollywood model gives a much better fit than ER and SBM, while the fit
under ERGM is very close to the observed degree distribution and shows
strong evidence of overfitting. Finally, the Hollywood model converged to
the MLE without problem while SBM and ERGM experienced convergence
issues and long run times.
7.2. Further discussion.
22
HARRY CRANE AND WALTER DEMPSEY
7.2.1. Preferential attachment models. The most well known model for sparse,
power law networks is the preferential attachment model of Barabási and
Albert (1999), which successfully puts forward a generating mechanism for
sparse networks that evolve by vertex growth and have a power law degree
distribution of exponent γ > 2. The model, however, seems suitable as
a statistical model only for specially structured datasets, such as citation
networks (Simon, 1955; De Solla Price, 1965). Otherwise, if the vertex
ordering is unobserved or there is no natural ordering of the vertices, then
these models lack basic logical properties, namely label equivariance, to
ensure sound inferences. Kolaczyk (2009, Section 6.4.3) also discusses some
practical issues with fitting the preferential attachment model to data.
7.2.2. Graphon models. Graphon models grow out of the theory of dense
graph limits (Lovász and Szegedy, 2006) and are intimately related to
the Aldous–Hoover theory of partially exchangeable arrays (Aldous, 1981;
Hoover, 1979). As we mentioned in Section 1, traditional exchangeable
random graph models with vertex set N cannot reproduce sparsity.
Probabilists often address this issue by studying the graphon under a
sparse regime, an approach brought into the statistics literature by Bickel and
Chen (2009). With φ : [0, 1] × [0, 1] →R [0, 1] symmetric in its arguments and
{ρn }n≥1 satisfying ρn → ∞ and ρ−1
n n [0,1]×[0,1] φ(u, v)du dv = O(1) as n → ∞,
Bickel and Chen (2009) define the distribution of a graph Gn = ([n], En ) with
vertices labeled in [n] by taking U1 , . . . , Un i.i.d. Uniform[0, 1] and putting
(24)
pr({i, j} ∈ En | U1 , . . . , Un ) = ρ−1
n φ(Ui , U j ),
1 ≤ i < j ≤ n,
conditionally independently for all 1 ≤ i < j ≤ n. Under (24), every Gn is
exchangeable, and collectively the sequence (Gn )n=1,2,... is sparse in the sense
of Definition 3.1. But the formulation does not automatically correspond to a
generating process for a population network since the marginal distributions
in (24) are not logically related through any specified sampling mechanism.
Without a direct link between samples of different size, the meaning of
the parameter φ varies with sample size and there is no logical way to relate
parameter estimates φ̂n based on a sample of size n to a single parameter φ
for the population. In this case, the interpretation of estimated parameters
is obscured and the significance of asymptotic statements for finite sample
approximations, as derived, for example, by Borgs et al. (2015) and Wolfe
and Olhede (2014), is unclear. Shalizi and Rinaldo (2013) discuss related
issues for estimation from exponential random graph models. Crane and
Dempsey (2015) discuss the effects of sampling more generally.
7.2.3. Completely random measures and graphex models. Caron and Fox (2014),
and later Veitch and Roy (2015) and Borgs et al. (2016), studied a class of network models derived from exchangeable point processes X on [0, ∞) × [0, ∞).
In this setting, each (x, x0 ) ∈ X determines an edge between vertices corresponding to points x, x0 . The point process X is assumed to be exchangeable
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
23
in the sense that its distribution is invariant under the joint action on
[0, ∞) × [0, ∞) by measure preserving transformations of [0, ∞).
In this context, a sequence of network data is obtained by defining a graph
Gt for each t ≥ 0, where Gt is derived from the restriction of X to [0, t] × [0, t]
by only including in Gt those vertices that are labeled in [0, t] and which
are not isolated in the restriction. Veitch and Roy (2015) adapt a theorem
of Kallenberg (1990) to obtain a generic graphex representation of the class of
random graphs derived from exchangeable point processes in this way.
Graphex models allow for sparse, power law behavior, and their description in terms of the exchangeable point process X does invoke some
invariance principle in the construction. It is important to stress, however,
that the notion of exchangeability applies to the generating point process
X, and not directly to the associated process of graphs (Gt )t≥0 . In fact, the
graphs (Gt )t≥0 induced from the process are not equipped with any labels,
and so it is not yet clear what exchangeability of the point process implies
for the induced graph sequence, or how the construction of Gt by restricting
to [0, t] and removing isolated vertices relates to the manner in which real
world networks are observed. At the time of this writing, graphex models
are still under development with some of these questions in mind.
8. Proof of main theorems
We equip EN with the topology and Borel σ-field induced by the metric
d(E, E0 ) = 1/(1 + sup{n ∈ N : E|[n] = E0 |[n] }),
E, E0 ∈ EN ,
with the convention 1/∞ = 0.
8.1. Proof of Theorem 5.1. Let (Yn )n∈N be a realization of the Hollywood
process with parameter (α, θ, ν) from Section 5. Then the distribution of
Yn is given by (13) for each n ∈ N, and these distributions are consistent
with respect to the restriction operation by consequence of the sequential
construction in (12). The distribution in (13) depends on Yn only through
v(Yn ), Nk (Yn ), and Mk (Yn ), that is, the number of vertices, number of
vertices of degree k ≥ 1, and the number of k-ary edges for each k ≥ 1, all of
which are invariant under relabeling edges. Edge exchangeability follows.
8.2. Proof of Theorem 5.2. Let (Yn )n≥1 be a realization from the Hollywood
process with parameter (α, θ, ν). The power law behavior is apparent by
the following connection between (Yn )n≥1 and the two parameter Ewens–
Pitman distribution. In particular, the Ewens–Pitman distribution corresponds to the special case of the Hollywood model with 0 < α < 1, θ > −α,
and ν1 = 1. In this case, there is exactly one vertex incident to every edge
and the network data corresponds to a partition of the edges. The number of
blocks in the Ewens–Pitman partition corresponds to the number of vertices
in the unary Hollywood process. Let Nk (Yn ) be the number of vertices
with degree k in Yn and let v(Yn ) be the total number of vertices in Yn . By
Pitman (2005, Lemma 3.11), we know that Nk (Yn )/v(Yn ) → α(1 − α)↑(k−1) /k!
24
HARRY CRANE AND WALTER DEMPSEY
a.s. for every k ≥ 1 as n → ∞. Thus, the sequence of degree distributions {dn }n≥1 = {(Nk (Yn )/v(Yn ))k≥1 }n≥1 converges a.s. in the total variation
topology to the distribution given by pα (k) = α(1 − α)↑(k−1) /k!.
In the general Hollywood process with arbitrary distribution ν, the degree
sequences of (Yn )n≥1 coincide
with a random subsequence of {dn }n≥1 indexed
P
by {Kr }r≥1 for Kr = rj=1 κ j , r = 1, 2, . . ., where κ1 , κ2 , . . . are i.i.d. from
ν. Thus, the degree distributions of (Yn )n≥1 correspond to {dKr }r≥1 , which
is a subsequence of the a.s. converging sequence {dn }n≥1 and, therefore,
must have the same a.s. limit. The proof is completed by noting that
α(1 − α)↑(k−1) /k! ∼ αk−(α+1) /Γ(1 − α) as k → ∞, which corresponds to the
power law with exponent α + 1.
8.3. Proof of Theorem 5.3. We once again exploit the connection to the
Ewens–Pitman distribution from the proof of Theorem 5.2. Let Nn be the
number of vertices in the unary Hollywood process with parameter (α, θ)
satisfying 0 < α < 1 and θ > −α. Pitman (2005, Theorem 3.8) shows that
n−α Nn → Sα a.s., where Sα is a strictly positive and finite random variable.
The sequence (Nn )n≥1 then satisfies Nn ∼ nα Sα a.s. as n → ∞, so that
E(Nn ) ∼
Γ(θ + 1) α
n
αΓ(θ + α)
as n → ∞,
by Pitman (2005, Theorem 3.8).
Let (Yn )n≥1 be the Hollywood process with parameter (α, θ, ν) for arbitrary
distribution ν. Then (v(YPn ))n≥1 is a random subsequence of (Nn )n≥1 given
by (NKr )r≥1 , where Kr = rj=1 κ j for κ1 , κ2 , . . . i.i.d. from ν. It follows from
the above argument that v(Yn ) = NKn ∼ Knα Sα a.s. as n → ∞. By the strong
law P
of large numbers, n−1 Kn ∼ µ a.s. and Knα ∼ (µn)α a.s. as n → ∞, where
µ = k≥1 kνk ; whence,
E(v(Yn )) ∼
Γ(θ + 1)
(µn)α
αΓ(θ + α)
as n → ∞.
To establish sparsity of (Yn )n≥1 , we consider (n−1 v(Yn )m• (Yn ) )n≥1 , where
m• (Yn ) is the average total degree in Yn . We must identify values of
0 < α < 1 for which lim infn→∞ n−1 v(Yn )m• (Yn ) = +∞. By the strong law
of large numbers, m• (Yn ) → µ a.s. as n → ∞. By the above discussion,
v(Yn ) ∼ (µn)α Sα a.s. as n → ∞ for a strictly positive and finite random
variable Sα . It follows that n−1 v(Yn )m• (Yn ) ∼ n−1 (µn)µα Sα a.s. as n → ∞,
which goes to infinity as long as µα > 1. Thus, (Yn )n≥1 is sparse with
probability 1 provided 1/µ < α < 1, for all θ > −α.
8.4. Proof of Theorem 5.4. Let (Yn )n≥1 be the binary Hollywood process
with parameter 0 < α < 1 and θ > −α. For each n ≥ 1, let Hnt,c be the
projection obtained by applying (16) to Yn . To establish sparsity of (Hnt,c )n≥1 ,
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
we must show
25
e(Hnt,c )
= 0 a.s.,
v(Hnt,c )2
for which it suffices to establish sparsity under the standard projection Hn∗ ,
since e(Hnt,c ) ≤ e(Hn∗ ) and v(Hnt,c ) = v(Hn∗ ).
For each k ≥ 1 and n ≥ 1, we write Nk (Yn ) and Nk (Hn∗ ) to denote the
number of vertices with degree k in Yn and Hn∗ , respectively. We also write
e(Yn ) and e(Hn∗ ) to denote the number of edges in Yn and Hn∗ , respectively,
and v(Yn ) = v(Hn∗ ) to denote the number of vertices in Yn and Hn∗ . Since
the projection operation reduces multiple occurrences of the same edge to a
single edge, the degree of each vertex in Hn∗ can be no larger than v(Hn∗ ) and
X
X
(25)
e(Hn∗ ) =
kNk (Hn∗ ) ≤
(k ∧ v(Hn∗ ))Nk (Yn ).
lim sup
n→∞
k≥1
k≥1
For every K ≥ 1, (25) implies
lim sup
n→∞
e(Hn∗ )
v(Hn∗ )2
≤ lim sup
n→∞
≤ lim sup
n→∞
≤
(26)
K
X
k=1
∞
K
X
X
k ∧ v(Hn∗ ) Nk (Yn )
Nk (Yn )
+
lim
sup
∗
∗
v(Hn ) v(Hn )
v(Hn∗ )
n→∞
k=K+1
k=1
K
X
k ∧ v(H∗ )
n
v(Hn∗ )
k=1
lim sup
n→∞
∞
X
Nk (Yn )
+ lim sup
v(Hn∗ )
n→∞
k ∧ v(Hn∗ )
+ lim sup
v(Hn∗ )
n→∞
k=K+1
∞
X
k=K+1
Nk (Yn )
.
v(Hn∗ )
n−α v(Hn∗ )
Pitman (2005, Corollary 3.9) implies that
→ Sα a.s., where Sα
is a strictly positive, finite random variable; thus, v(Hn∗ ) → ∞ a.s. and
lim supn→∞ (k ∧ v(Hn∗ ))/v(Hn∗ ) → 0 a.s. for every k = 1, . . . , K, so that (26)
implies
∞
X
e(Hn∗ )
Nk (Yn )
≤ lim sup
(27)
lim sup
for all K ≥ 1.
∗
2
v(Hn∗ )
n→∞ v(Hn )
n→∞
k=K+1
By Pitman (2005, Lemma 3.11),
Nk (Yn )
= pα (k) = αΓ(k − α)/(k!Γ(1 − α)) for every k ≥ 1 a.s.,
v(Hn∗ )
implying that for every ε, δ > 0 and k ≥ 1 there exists R = R(ε, δ, k) such that
lim
n→∞
pr(|v(Hn∗ )−1 Nk (Yn ) − pα (k)| < ε for all n > R) ≥ 1 − δ.
For any K ≥ 1, we choose R∗ = max1≤k≤K R(ε/K, δ/K, k) so that
pr(|v(Hn∗ )−1 Nk (Yn ) − pα (k)| < ε/K for all n > R∗ , for all k = 1, . . . , K) ≥ 1 − δ
and
∞
∞
X
X
∗
−1
∗
pr
v(Hn ) Nk (Yn ) −
pα (k) < ε for all n > R ≥ 1 − δ.
k=K+1
k=K+1
26
HARRY CRANE AND WALTER DEMPSEY
We combine this with (27) and the tail calculation pα (> K) =
Γ(K + 1 − α)/(Γ(K + 1)Γ(1 − α)) to observe
P∞
k=K+1 pα (k)
=
pr(lim sup e(Hn∗ )/v(Hn∗ )2 > pα (> K) + ε) ≤
n→∞
∞
X
Nk (Yn )/v(Hn∗ ) > pα (> K) + ε
≤ pr lim sup
n→∞ k=K+1
∞
∞
X
X
≤ pr
v(Hn∗ )−1 Nk (Yn ) −
pα (k) > ε for some n > R∗
k=K+1
k=K+1
≤ δ
for every K ≥ 1. Since pα (> K) ↓ 0 as K → ∞, it follows that
pr(lim sup e(G∗n )/v(Hn∗ )2 > ε) ≤ δ
for all ε, δ > 0
n→∞
and (Hn∗ )n≥1 is sparse a.s.
References
[1] J. Abello, A. Buchsbaum, and J. Westbrook. A functional approach to external
graph algorithms. Proceedings of the 6th European Symposium on Algorithms, pages
332–343, 1998.
[2] E.M. Airoldi, D.S. Choi, and P.J. Wolfe. Confidence sets for network structure.
Statistical Analysis and Data Mining, 4(5):461–469, 2011.
[3] D. J. Aldous. Representations for partially exchangeable arrays of random variables.
J. Multivariate Anal., 11(4):581–598, 1981. ISSN 0047-259X. doi: 10.1016/0047259X(81)90099-3. URL http://dx.doi.org/10.1016/0047-259X(81)90099-3.
[4] A.-L. Barabási and R. Albert. Emergence of scaling in random networks. Science,
286(5439):509–512, 1999. ISSN 0036-8075. doi: 10.1126/science.286.5439.509. URL
http://dx.doi.org/10.1126/science.286.5439.509.
[5] P. Bickel and A. Chen. A nonparametric view of network models and Newman–
Girvan and other modularities. Proceedings of the National Academy of Sciences of
the United States of America, 106(50):21068–21073, 2009.
[6] C. Borgs, J.T. Chayes, H. Cohn, and S. Ganguly. Consistent nonparametric estimation for heavy-tailed sparse graphs. arXiv:1508.06675, 2015.
[7] C. Borgs, J.T. Chayes, H. Cohn, and N. Holden. Sparse exchangeable graphs and
their limits via graphon processes. arXiv:1601.07134, 2016.
[8] C.T. Butts. A relational event framework for social action. Sociol. Methodol. 38,
155–200, 2008.
[9] F. Caron and E.B. Fox. Sparse graphs using exchangeable random measures.
Accessed at arXiv:1401.1137v3, 2014.
[10] F. Chung and L. Lu. Complex graphs and networks, volume 107 of CBMS Regional
Conference Series in Mathematics. Published for the Conference Board of the
Mathematical Sciences, Washington, DC, 2006. ISBN 978-0-8218-3657-6; 0-82183657-9.
[11] H. Crane and W. Dempsey. Supplementary material for edge exchangeable models
for network data. 2016.
EDGE EXCHANGEABLE MODELS FOR NETWORK DATA
27
[12] H. Crane. The ubiquitous Ewens sampling formula (with discussion and a rejoinder
by the author). Statistical Science, 31(1):1–39, 2016a.
[13] H. Crane. Rejoinder: The ubiquitous Ewens sampling formula. Statistical Science,
31(1):37–39, 2016b.
[14] H. Crane and W. Dempsey. A framework for statistical network modeling. 2015.
[15] H. Crane and W. Dempsey. Atypical scaling behavior persists in real world interaction networks. Unpublished manuscript, July 2015.
[16] D. J. De Solla Price. Networks of Scientific Papers. Science, 149(3683):510–515, 1965.
[17] P. Diaconis and S. Janson. Graph limits and exchangeable random graphs. Rend.
Mat. Appl. (7), 28(1):33–61, 2008. ISSN 1120-7183.
[18] W. J. Ewens. The sampling theory of selectively neutral alleles. Theoret. Population
Biology, 3:87–112, 1972. ISSN 0040-5809.
[19] M. Faloutsos, P. Faloutsos, and C. Faloutsos. On power-law relationships of the
Internet topology. ACM Comp. Comm. Review, 29, 1999.
[20] S. Feng. The Poisson-Dirichlet Distribution and Related Topics. Probability and its
Applications. Springer-Verlag, Berlin, 2010.
[21] T. Ferguson. A Bayesian Analysis of Some Nonparametric Problems. The Annals of
Statistics, 1(2):209–230, 1973.
[22] C. Gao, Lu. Y., and H.H. Zhou. Rate optimal graphon estimation. Annals of Statistics,
43(6):2624–2652, 2015.
[23] C. Gao, Ma. Z., A.Y. Zhang, and H.H. Zhou. Community Detection in DegreeCorrected Block Models. arXiv:1607.06993, 2016.
[24] P.W. Holland and S. Leinhardt. An exponential family of probability distributions
for directed graphs. Journal of the American Statistical Association, 33–65, 1981.
[25] P.W. Holland, K.B. Laskey, and S. Leinhardt. Stochastic blockmodels: First steps.
Social Networks, 5(2):109–137, 1983.
[26] D.N. Hoover. Relations on Probability Spaces and Arrays of Random Variables.
Preprint, Institute for Advanced Studies, 1979.
[27] H. Jeong, S.P. Mason, A.-L. Barabási, and Z.N. Oltvai. Lethality and centrality in
protein networks. Nature, 411:41, 2001.
[28] O. Kallenberg. Exchangeable random measures in the plane. Journal of Theoretical
Probability, 3(1):81–136, 1990.
[29] M. Kivelä, A. Arenas, M. Barthelemy, J.P. Gleeson, Y. Moreno, and M.A. Porter.
Multilayer Networks. Journal of Complex Networks, 2(3):203–271, 2014.
[30] B. Klimt and Y. Yang. Introducing the Enron corpus. CEAS, 2004.
[31] O. Klopp, A.B. Tsybakov, and N. Verzelen. Oracle inequalities for network models
and sparse graphon estimation. arXiv:1507.04118, 2016.
[32] E. D. Kolaczyk. Statistical analysis of network data. Springer, New York, 2009.
[33] P. Latouche, S. Robin, and S. Ouadah. Goodness of fit of logistic models for random
graphs. arXiv:1508.00286, 2015.
[34] L. Lovász and B. Szegedy. Limits of dense graph sequences. J. Comb. Th. B, 96:
933–957, 2006.
[35] M. Mariadassou, S. Robin, and C. Vacher. Uncovering latent structure in valued
graphs: a variational approach. Annals of Applied Statistics, 715–742, 2010.
[36] J. J. McAuley and J. Leskovec. Learning to discover social circles in ego networks.
In Neural Information Processing Systems, pages 539–547, 2012.
[37] J. Nesetril and P. Ossona de Mendez. Sparsity: Graphs, Structures, and Algorithms.
Springer:Heidelberg, 2012.
28
HARRY CRANE AND WALTER DEMPSEY
[38] M.E.J. Newman. The structure of scientific collaboration networks. Proceedings of
the National Academy of Sciences, 98:404–409, 2001.
[39] T. Opsahl. Why Anchorage is not (that) important: Binary ties and Sample selection.
Available at http://wp.me/poFcY-Vw, 2011.
[40] T. Opsahl and P. Panzarasa. Clustering in weighted networks. Social Networks, 31:
155–163, 2009.
[41] P. Orbanz and D.M. Roy. Bayesian Models of Graphs, Arrays and Other Exchangeable Random Structures. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 37(2):437–461, 2015.
[42] P.O. Perry and P.J. Wolfe. Point process modelling for directed interaction networks.
Journal of the Royal Statistical Society: Series B, 75(5):821-849, 2013.
[43] J. Pitman. Combinatorial stochastic processes, volume 1875 of Lecture Notes in Mathematics, 2005.
[44] G. Robins and M. Morris. Editorial: Advances in exponential random graph (p∗ )
models. Social Networks, 29:169–172, 2007.
[45] R. A. Rossi and N. K. Ahmed. The network data repository with interactive graph
analytics and visualization. In Proceedings of the Twenty-Ninth AAAI Conference
on Artificial Intelligence, 2015. URL http://networkrepository.com.
[46] J. Sethuraman. A constructive definition of Dirichlet priors. Statistica Sinica, 4:
639–650, 1994.
[47] C.R. Shalizi and A. Rinaldo. Consistency under subsampling of exponential random
graph models. Annals of Statistics, 41:508–535, 2013.
[48] H.A. Simon. On a class of skew distribution functions. Biometrika, 42:425–440, 1955.
[49] T.A.B. Snijders and K. Nowicki. Estimation and prediction for stochastic blockmodels for graphs with latent block structure. Journal of Classification, 14(1):75–100,
1997.
[50] T.M. Sweet. Incorporating Covariates Into Stochastic Blockmodels. Journal of
Educational and Behavioral Statistics, 1076998615606110, 2015.
[51] C. Tallberg. A Bayesian approach to modeling stochastic block- structures with
covariates. Journal of Mathematical Sociology, 29(1):1–23, 2004.
[52] V. Veitch and D. Roy. The class of random graphs arising from exchangeable random
measures. arXiv:1512.03099 accessed at http://arxiv.org/pdf/1512.03099v1.pdf, 2015.
[53] P.J. Wolfe and S.C. Olhede. Nonparametric graphon estimation. Available at
arXiv:1309.5936, 2014.
[54] W.W. Zachary. An information flow model for conflict and fission in small groups.
Journal of Anthropological Research, 33:452–473, 1977.
[55] Y. Zhang, E. Levina, and J. Zhu. Community detection in networks with node
features. arXiv:1509.01173, 2015.
[56] Y. Zhao, E. Levina, and J. Zhu. On consistency of community detection in networks.
Annals of Statistics, 40(4):2266–2292, 2011.
Department of Statistics & Biostatistics, Rutgers University, 110 Frelinghuysen Avenue, Piscataway, NJ 08854, USA
E-mail address: [email protected]
URL: http://stat.rutgers.edu/home/hcrane
Department of Statistics, University of Michigan, 1085 S. University Ave, Ann Arbor,
MI 48109, USA
E-mail address: [email protected]
SUPPLEMENTARY MATERIAL: EDGE EXCHANGEABLE MODELS
FOR NETWORK DATA
HARRY CRANE AND WALTER DEMPSEY
Abstract. This is the supplement to Crane and Dempsey (2016). We
characterize edge exchangeable network models and describe additional
examples of interaction datasets.
Appendix A. Characterization of edge exchangeable networks
Our representation in Crane and Dempsey (2016, Theorem 4.2) applies
only to the special case of edge exchangeable networks that are blip-free. We
prove here a more general, complete characterization for all edge exchangeable networks in the binary case for which each edge is undirected and
involves exactly two vertices. The argument for directed networks with
edges of any finite arity is analogous but more technical. Recall the notation
for edge-labeled networks introduced in Crane and Dempsey (2016).
Let fin2 (P) ⊂ fin(P) be the set of all size 2 multisets of P. Given a binary
edge-labeled network E with e(E) = n, we call S : [n] → fin2 (N) a selection
function for E if ES = E, where ES is as defined by Crane and Dempsey
(2016, Equation (5)); that is, S is an interaction process whose induced edgelabeled network agrees with E. We think of the selection function as a way
of labeling the vertices of the edge-labeled network E.
Selection functions S, S0 : [n] → fin2 (N) are equivalent, written S ≡ S0 ,
if they correspond to the same edge-labeled network, that is, ES = ES0 .
To every edge-labeled network E with n edges we associate a canonical
selection function SE : [n] → fin2 (N) defined by labeling the vertices in order
of appearance, as follows.
We initialize by putting SE (1) = {1, 1} if the edge labeled 1 is a self loop
and otherwise we put SE (1) = {1, 2} to indicate an edge between two distinct
vertices. Given SE (1), . . . , SE (i − 1), we define SE (i) = {v1 (i), v2 (i)} by choosing
v1 (i) ≤ v2 (i) to be the smallest vertex labels consistent with the structure of
E|[i] . In this way, v j (i) = r coincides with a previously observed vertex label
if one of the vertices involved in the ith interaction corresponds to the vertex
labeled r in the previous interactions SE (1), . . . , SE (i − 1). See Figures 1(b)
and 1(c) for an illustration.
Date: October 19, 2016.
H. Crane is partially supported by NSF grants CNS-1523785 and CAREER DMS-1554092.
1
2
HARRY CRANE AND WALTER DEMPSEY
Figure 1. Illustration of the generic process of generating
edge exchangeable networks in Section A. (a) Network representation of interaction process X1 = {2, 4}, X2 = {1, 2},
X3 = {1, 5}, X4 = {6, 9}, X5 = {2, 6}, X6 = {2, 6}. (b) Edgelabeled network EX obtained by removing vertex labels from
the vertex-edge-labeled network in Panel (a). (c) Vertex labeling of the network in Panel (b) according to its canonical
selection function.
The fin2 (N)-simplex consists of all ( f{i, P
j} ) j≥i≥−1 such that f{i, j} ≥ 0 for all
j ≥ i ≥ −1, f{−1,i} = 0 for all i , 0, and j≥i≥−1 f{i, j} = 1. (The labels −1
and 0 have a special status in the construction, as we shall see.) For any
f = ( f{i, j} ) j≥i≥−1 in the fin2 (N)-simplex and i ∈ N, we define
(i)
f• =
∞
X
f{i,j}
j=0
as the sum of masses involving element i.
Every f = ( f{i, j} ) j≥i≥−1 in the fin2 (N)-simplex determines a probability
distribution on edge-labeled networks, denoted f , as follows. Let X1 , X2 , . . .
be i.i.d. random pairs {i, j} with
(1)
pr(Xk = {i, j} | f ) = f{i,j} ,
j ≥ i ≥ −1.
Given X1 , X2 , . . ., we define the selection function X : N → fin2 (Z), where
Z = {. . . , −1, 0, 1, . . .}, as follows. We initialize with m0 = 0. For n ≥ 1,
suppose mn−1 = z ≤ 0. If Xn contains no 0s, then we define X(n) = Xn
and update mn = mn−1 . If Xn = {0, j} for some j ≥ 1, then we put X(n) =
{z − 1, j} and update mn = z − 1. Otherwise, if Xn = {0, 0}, then we put
X(n) = {z − 1, z − 1} and update mn = z − 1; and if Xn = {−1, 0}, then we
put X(n) = {z − 1, z − 2} and update mn = z − 2. (Thus, events with −1 or
0 involve vertices that appear once and never again. These are the “blips”
that we previously ruled out in Crane and Dempsey (2016, Theorem 4.2).)
We define Y = EX ∼ f to be the edge-labeled network induced by X. See
Figure 1 for an illustration of this procedure.
SUPPLEMENTARY MATERIAL
3
Proposition A.1. The edge-labeled network Y = EX corresponding to X1 , X2 , . . .
i.i.d. from (1) is edge exchangeable for all f in the fin2 (N)-simplex.
For identifiability, we define the ranked reordering of f by f ↓ = ( f{i,↓ j} ) j≥i≥−1 ,
↓
the element of the fin2 (N)-simplex obtained by putting f{−1,0}
= f{−1,0} ,
(i)
(i+1)
↓
f{0,0}
= f{0,0} , and reordering elements 1, 2, . . . so that f• ≥ f•
(i)
f•
for all i ≥ 1
(i+1)
f•
and then breaking ties
=
by declaring that ( f{i,i} , f{i,i+2} , . . .) comes
before ( f{i+1,i+1} , f{i+1,i+2} , . . .) in the lexicographic ordering. We write F ↓ to
denote the space of rank reordered elements of the fin2 (N)-simplex.
As the vertex labels other than −1 and 0 are inconsequential, it is clear that
f and f 0 determine the same distribution for any f, f 0 for which f ↓ = f 0↓ .
For any edge-labeled network E, we write |E|↓ ∈ F ↓ to denote its signature, if
it exists, as follows. Let SE : N → fin2 (N) be the canonical selection function
for E. For every {i, j} ∈ fin2 (N), j ≥ i ≥ 1, we define
n
X
−1
f{i,j} (E) = lim n
1{SE (k) = {i, j}} and
n→∞
(i)
f• (E) = lim n−1
n→∞
k=1
n
X
1{i ∈ SE (k)},
k=1
if the limits exist. We also define
∞
X
(i)
f{0,i} (E) = f• −
f{i,j} (E),
i ≥ 1,
j=1
∞
n X
X
X
−
f{0,0} (E) = lim n−1
1{S
(k)
=
{`,
`}}
f{i,i} (E),
E
n→∞
f{−1,0} (E) = lim n
−1
n→∞
k=1 `≥1
n X
X
k=1 `,r≥1:`,r
and
i=1
X
1{SE (k) = {`, r}} −
f{i,j} (E).
j>i≥1
Provided each of the above limiting frequencies exists, we define |E| =
( f{i,j} (E)) j≥i≥−1 and |E|↓ = ( f{i,j} (E))↓j≥i≥−1 .
(Note the role of f{0,0} (E) and f{−1,0} (E) for recording the residual proportion
of loops and edges, respectively, that do not contribute to the limiting
frequencies f{i,j} (E) for a given j ≥ i ≥ 1. For example, the interaction
process I : N → fin2 (N) given by I(i) = {i, i} for each i ≥ 1 corresponds to
the edge-labeled network E = EI for which every edge is a loop at a distinct
vertex. For any given i ≥ 1, f{i,i} (E) = 0 and we have f{0,0} (E) = 1. Conversely,
for f = ( f{i,j} ) j≥i≥−1 with f{0,0} = 1, the sequence X1 , X2 , . . . i.i.d. from (1) yields
{0, 0}, {0, 0}, . . . with probability 1, whose associated interaction process has
X(1) = {−1, −1}, X(2) = {−2, −2}, and X(n) = {−n, −n}, for each n ≥ 1, with
associated edge-labeled network having all edges corresponding to a loop
at a distinct vertex.)
4
HARRY CRANE AND WALTER DEMPSEY
Theorem A.2. Let Y ∈ EN be an edge exchangeable network. Then there exists a
unique probability measure φ on F ↓ such that Y ∼ φ , where
Z
φ (·) =
(2)
F↓
f (·)φ(d f ).
That is, every edge exchangeable network Y can be generated by first sampling
f ∼ φ and, given f , putting Y = EX for X : N → fin2 (Z) constructed from
X1 , X2 , . . . i.i.d. according to (1).
Theorem 4.2 of Crane and Dempsey (2016) follows as a corollary to
Theorem A.2 by ruling out blips, that is, by confining φ to the subset of
f ∈ fin2 (N) for which f{i,0} = 0 for all i ≥ −1.
A.1. Proof of Theorem A.2. We equip EN with the product-discrete topology induced by the metric
dEN (E, E0 ) = 1/(1 + sup{n ∈ N : E|[n] = E0 |[n] }),
E, E0 ∈ EN ,
with convention 1/∞ = 0, and F ↓ with the topology induced by
dF ↓ ( f, f 0 ) =
X
| f{i,j} − f{i,0 j} |,
f, f 0 ∈ F ↓ .
j≥i≥−1
We then work with the respective Borel σ-fields induced by these topologies.
Let Y ∈ EN be an edge exchangeable random network, let SY : N →
fin2 (N) be its canonical selection function, and let ξ1 , ξ2 , . . . be an i.i.d.
sequence of Uniform[0, 1] random variables which are independent of Y.
Given Y and (ξi )i≥1 , we define Z : N → fin2 ([0, 1]) by Z(n) = {ξi , ξ j } on the
event SY (n) = {i, j}, for n ≥ 1.
By independence of Y and (ξi )i≥1 and edge exchangeability of Y, (Z(n))n≥1
is an exchangeable sequence taking values in the Polish space fin2 ([0, 1]). By
de Finetti’s theorem, see, for example, Aldous (1985), there exists a unique
measure µ on the space P(fin2 ([0, 1])) of probability measures on fin2 ([0, 1])
such that Z =D Z∗ = (Z∗ (n))n≥1 with
Z
pr(Z ∈ ·) =
∗
m∞ (·)µ(dm),
P(fin2 ([0,1]))
where m∞ denotes the infinite product measure of m. In particular, there
exists a random measure ν on fin2 ([0, 1]) such that
pr(Z ∈ · | ν) = ν∞
a.s.
SUPPLEMENTARY MATERIAL
5
Given ν, we define
f{i,j} = ν({{ξi , ξ j }}),
i, j ≥ 1,
(i)
f• = ν({{x, y} ∈ fin2 ([0, 1]) : ξi ∈ {x, y}}),
∞
X
(i)
f{0,i} = f• −
f{i,j} , i ≥ 1,
i ≥ 1,
j=1
f{0,0} = ν({{u, u} ∈ fin2 ([0, 1])}) −
∞
X
f{i,i} ,
and
i=1
f{−1,0} = ν({{u, v} ∈ fin2 ([0, 1]) : v , u}) −
X
f{i, j} .
j>i≥1
By construction ( f{i,j} ) j≥i≥−1 is in the fin2 (N)-simplex and, therefore, f ↓ ∈ F ↓ .
(Note that F ↓ is a subset of the fin2 (N)-simplex and f 7→ f ↓ is measurable
with respect to the Borel σ-field induced by the metric dF ↓ (·, ·) given above.)
Given ν, we let (Z0 , S0 ) be an i.i.d. copy of (Z, SY ) and let Y0 = ES0 be the
edge-labeled network induced by S0 . We complete the proof by showing
pr(Y0 ∈ · | ν) = f ↓ , for f ↓ as defined above from ν.
(i)
First, let A = {i ∈ N : f• > 0} and ξA = {ξi : i ∈ A}. It follows that
pr(Z0 (1) ∩ ξA = ∅ | ν) = f{0,0} + f{−1,0}
and
pr(Z0 (1) ∩ ξA = {ξi } | ν) = f{0,i} .
By exchangeability, i < A implies ξi appears at most once in Z with probability 1. We further have that
pr(Z01 ∩ ξA = ∅ and Z0 (1) = {u, u} for some u ∈ [0, 1] | ν) = f{0,0}
and
pr(Z (1) ∩ ξA = ∅ and Z (1) = {u, v} for u , v | ν) = f{−1,0} .
0
0
Now, define X0 : N → fin2 (N ∪{−1, 0}) and the random selection function
SX0 : N → fin2 (Z) as follows. Let m0 = 0. For n ≥ 1, suppose mn−1 = z ≤ 0.
If Z0 (n) ∩ ξA = {ξi , ξ j } for some i, j ∈ N, then put X(n) = SX (n) = {i, j}. If
Z0 (n) ∩ ξA = {ξi } for some i ≥ 1, then put X(n) = {0, i}, SX (n) = {z − 1, i},
and mn = z − 1. If Z0 (n) ∩ ξA = ∅ and Z0 (n) = {u, u} for some u ∈ [0, 1], put
X(n) = {0, 0}, SX (n) = {z − 1, z − 1}, and mn = z − 1. And if Z0 (n) ∩ ξA = ∅
and Z0 (n) = {u, v} for u , v, put X(n) = {−1, 0}, SX (n) = {z − 1, z − 2}, and
mn = z − 2.
By construction, we have SX0 ≡ S0 a.s. and, given f , X0 is conditionally
i.i.d. from distribution (1). The integral representation in (2) follows by de
Finetti’s theorem, completing the proof.
Appendix B. Description of network datasets
Below we describe and provide references for some interaction datasets
discussed in Crane and Dempsey (2016).
6
HARRY CRANE AND WALTER DEMPSEY
• Actors collaboration (Barabási and Albert, 1999): Network built from
collaborations among actors in a given sample of movies. Each edge
connects the actors who played a role in the corresponding movie.
• Enron email corpus (Klimt and Yang, 2004): Network built from a
corpus of about 500,000 emails. Vertices are employees in the Enron
Corporation with a directed edge from vertex i to j for each email sent
from i to j. Some versions of this dataset project edge multiplicities
and ignore edge direction so that for each pair of vertices i and j
there is an undirected edge between i and j if at least one email was
exchanged between the two.
• Karate Club (Zachary, 1977): Network built from social interactions
among 34 members of a karate club. Vertices are the members of the
club and an edge between i and j corresponds to a social interaction
between the two. The network exhibits no vertex sampling or growth
since it is assumed all club members have been observed.
• Wikipedia voting (Leskovec et al., 2010): The Wikipedia voting
network represents voting behavior for elections to the administrator
role in Wikipedia. Vertices are Wikipedia users and a directed edge
points from i to j if user i voted for user j.
• US Airport (Colizza et al., 2007; Opsahl, 2011): The network from
Opsahl (2011) is built from the flight map between all US airports
in 2010. A directed edge from i to j indicates that a flight was
scheduled from airport i to airport j in 2010. Edges are weighted by
the number of seats on the scheduled flights. The network grows as
a consequence of additional flights between airports.
• Co-authorship (Newman, 2001): Network built from co-authorship
of preprints on the Condensed Matter section of arXiv between 1995
and 1999. Vertices are of two types, authors and papers, and edges
only exist between vertices of a different type. An edge between i
(author) and j (paper) indicates that i is an author on paper j. The
data is more succinctly represented by an edge-labeled network
as in Crane and Dempsey (2016) by associating each article to an
interaction involving all of its authors.
• UC Irvine (Opsahl and Panzarasa, 2009): Network built from UC
Irvine online community. Vertices are active members of the community and a directed edge from i to j indicates that a message was
sent from user i to user j.
• Political blogs (Adamic and Glance, 2005): Network built from
hyperlinks between political blogs. Vertices are websites (blogs)
with a directed edge from i to j for every hyperlink from website i
to website j. Sampling is similar to the Facebook network.
SUPPLEMENTARY MATERIAL
7
References
[1] L. A. Adamic and N. Glance. The Political Blogosphere and the 2004 US
Election: Divided They Blog. In Proceedings of the 3rd international workshop
on Link discovery, pages 36–43, 2005.
[2] D. J. Aldous. Exchangeability and related topics. In École d’été de probabilités
de Saint-Flour, XIII—1983, volume 1117 of Lecture Notes in Math., pages
1–198. Springer, Berlin, 1985.
[3] A.-L. Barabási and R. Albert. Emergence of scaling in random networks.
Science, 286(5439):509–512, 1999. ISSN 0036-8075. doi: 10.1126/science.286.
5439.509. URL http://dx.doi.org/10.1126/science.286.5439.509.
[4] V. Colizza, R. Pastor-Satorras, and A. Vespignani. Reaction diffusion processes and metapopulation models in heterogeneous networks. Nature
Physics, 3:276–282, 2007.
[5] H. Crane and W. Dempsey. Edge exchangeable models for network data.
2016.
[6] B. Klimt and Y. Yang. Introducing the Enron corpus. CEAS, 2004.
[7] J. Leskovec, D. Huttenlocher, and J. Kleinberg. Signed networks in social
media. CHI, 2010.
[8] J.J. McAuley and Jure Leskovec. Learning to discover social circles in ego
networks. In Neural Information Processing Systems, pages 539–547, 2012.
[9] M.E.J. Newman. The structure of scientific collaboration networks. Proceedings of the National Academy of Sciences, 98:404–409, 2001.
[10] T. Opsahl. Why Anchorage is not (that) important: Binary ties and Sample
selection. Available at http://wp.me/poFcY-Vw, 2011.
[11] T. Opsahl and P. Panzarasa. Clustering in weighted networks. Social Networks, 31:155–163, 2009.
[12] D. Watts and S. Strogatz. Collective dynamics of ‘small-world’ networks.
Nature, 393:440–442, 1998.
[13] W.W. Zachary. An information flow model for conflict and fission in small
groups. Journal of Anthropological Research, 33:452–473, 1977.
Department of Statistics & Biostatistics, Rutgers University, 110 Frelinghuysen Avenue, Piscataway, NJ 08854, USA
E-mail address: [email protected]
URL: http://stat.rutgers.edu/home/hcrane
Department of Statistics, University of Michigan, 1085 S. University Ave, Ann Arbor,
MI 48109, USA
E-mail address: [email protected]
| 10math.ST
|
Learning to Attend, Copy, and Generate
for Session-Based Query Suggestion
Mostafa Dehghani∗
University of Amsterdam
[email protected]
Enrique Alfonseca
arXiv:1708.03418v4 [cs.IR] 13 Nov 2017
Google Research
[email protected]
ABSTRACT
Users try to articulate their complex information needs during search
sessions by reformulating their queries. To make this process more
effective, search engines provide related queries to help users in
specifying the information need in their search process. In this paper
we propose a customized sequence-to-sequence model for sessionbased query suggestion. In our model, we employ a query-aware
attention mechanism to capture the structure of the session context.
This enables us to control the scope of the session from which we
infer the suggested next query, which helps not only handle the noisy
data but also automatically detect session boundaries. Furthermore
we observe that, based on the user query reformulation behavior,
within a single session a large portion of query terms is retained from
the previously submitted queries and consists of mostly infrequent
or unseen terms that are usually not included in the vocabulary. We
therefore empower the decoder of our model to access the source
words from the session context during decoding by incorporating
a copy mechanism. Moreover, we propose evaluation metrics to
assess the quality of the generative models for query suggestion. We
conduct an extensive set of experiments and analysis. The results
suggest that our model outperforms the baselines both in terms of
the generating queries and scoring candidate queries for the task of
query suggestion.
KEYWORDS
Sequence to Sequence Model, Query Suggestion, Query-Aware Attention, Copy Mechanism
1
INTRODUCTION
Users interact with search engines during search sessions and try
to direct their search by submitting a sequence of queries. Based
on these interactions, search engines provide a prominent feature,
in which they assist their users to formulate their queries to better
∗ Work done while interning at Google Research.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
CIKM’17, November 6–10, 2017, Singapore.
© 2017 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ISBN 978-1-4503-4918-5/17/11. . . $15.00
DOI: https://doi.org/10.1145/3132847.3133010
Sascha Rothe
Google Research
[email protected]
Pascal Fleury
Google Research
[email protected]
represent their intent during Web search by providing suggestions
for the next query.
Query suggestion might address the need for disambiguation of
the user queries to make the direction of the search more clear for
both, the user and the search engine. It might help users by providing a precise and succinct query when they are not familiar with
the specific terminology or when they lack understanding of the
internal vocabulary and structures in order to be able to formulate
an effective query. It has been shown that in general, query suggestion accelerates search satisfaction by either diving deeper into
the current search direction or by moving to a different aspect of a
search task [33, 40].
There has been a lot of research on the task of query suggestion and
similar tasks like query auto-completion. A large body of methods
leverages the idea of the “wisdom of crowds” by analyzing the search
logs to use either query co-occurrences [22, 33] in the search logs,
or document clicks information [2, 29, 35]. However, co-occurrence
based models suffer from data sparsity and lack of coverage for rare
or unseen queries. On the other hand, considering the previously
issued queries in the session, i.e context queries, and their order as
a sequence of attempts for finding relevant information is of crucial
for providing an effective suggestion. Dealing with these highly
diverse sessions makes using co-occurrence based model almost
impossible [6, 19, 42].
Sessions are driven by query reformulations and users modifying
existing queries in order to pursue new search results. Taking the
structure of the context queries into account is important as query
suggestion is well tightened to the understanding of query reformulation behaviors. A good query suggestion system should be able
to reproduce natural reformulation patterns from users. There are
several patterns in query reformulation like term addition, removal,
and retention [12, 41]. It has been shown that retained terms make
up a large proportion of query reformulation in search sessions. For
example, an average of 62% of the terms in a query are retained from
their preceding queries [41]. More than 39% of the users repeat at
least one term from their previous query [25]. On the other hand
retained terms are clearly core terms indicating the user’s information need, hence, they are usually discriminative terms and entities.
Based on statistics from the AOL query log [34], more than 67% of
the retained terms in the sessions are from the bottom 10% of terms
ordered by their frequency.
The recent success of sequence-to-sequence (seq2seq) models
in which recurrent neural networks (RNNs) both read and freely
generate text makes it possible to generate the next query by reading
the previously issued queries in the session [42]. Although generic
Time step #1
Time step #2
Time step #3
q1 bob dylan
q1 bob dylan
q1 bob dylan
q2 forever young dylan
q2 forever young dylan
q2 forever young dylan
q3 dylan photo
q3 dylan photo
q3 dylan photo
Cop:
bob dylan photo
Gen:
bob <OOV> bio
copy
generate
Cop:
bob dylan photo
Gen:
bob <OOV> bio
copy
generate
Cop:
bob dylan photo
Gen:
bob <OOV> bio
copy
generate
Figure 1: Example of generating a suggestion query given the previous queries in the session. The suggestion query is generated
during three time steps. The heatmap indicates the attention, red for query-level attention and blue for word-level attention.
The pie chart shows if the network decides to copy or to generate.
seq2seq models are promising in generating text, they have some
shortcomings in the task of query suggestion. The first problem of
directly employing the generic seq2seq model for the task of query
suggestion is that it considers the input data as a sequence of words,
ignoring the query level information. To address this, Sordoni et al.
[42] proposed a context-aware seq2seq model in which they use a
hierarchical architecture to encode the previously issued queries in
the session and generate the most likely sequence of words as the
next query. The second shortage of a generic word-based seq2seq
model is that it’s unable to deal with out-of-vocabulary words (OOV).
Besides, these models are less likely to generate terms with very low
frequency [14]. This makes them unable to effectively model term
retention, which is the most common reformulation patterns for the
next query generation.
In this paper, we present an architecture that addresses these two
issues in the context of session-based query suggestion. We augment
the standard seq2seq model with query-aware attention mechanism
enabling the model to Attend to the promising scope of the session
for generating the next query. Furthermore, we incorporate the copy
mechanism by adding a copier component which lets the decoder
Copy terms from the session context that improves the performance
by modeling the term retention and handling of OOVs. The model
still has the ability to Generate new words through a generator
component. Our model, which we are going to call ACG in the rest
of the paper, is trained in a multi-objective learning process.
Figure 1 illustrates an example of the output of our model as the
suggestion for the next query, given the previously submitted queries
in a session. This example session is composed of three queries: bob
dylan 7−→ forever young dylan 7−→ dylan photo, which were submitted sequentially. Our model outputs the sequence of the words bob,
dylan, and bio. At each time step, the heatmap of the query level attention (red) and word level attention (blue) is illustrated. Furthermore,
the output of the copier, of the generator, and the probability of the
network deciding to copy a term from the previous queries or to generate a new term is given for each time step. At time step #1, the first
query in the session and in this query, word bob has the highest attention. The outputs of both copier and generator are the same, but the
network decides to copy the term bob (probably from the first query).
At time step #2, dylan is an OOV. So the output of the generator is
the hOOV i token and based on the learned attentions, the network
decides to copy dylan from queries in the session. At time step #3, the
last query in the session and in this session term photo has the highest
attention, and the network decides to generate the new term bio.
Besides proposing a seq2seq model which learns to effectively
attend, copy and generate for the task of session-based query suggestion, we introduce new metrics for evaluating the output of generative models for the task of query suggestion. We train and evaluate
ACG on the AOL query log data and compare it to the state-of-the-art
models both in terms of the ability to discriminate and the ability to
generate. The results suggest that ACG as a discriminative model is
able to effectively score good candidates and as a generative model
generates better queries compared to the baseline models. In the
following, we first explain our model in detail. Then we describe the
evaluation paradigm we use in this paper. Afterward, we present our
results followed by analysis and discussions. In the end, we review
some related work and conclude the paper.
2
OUR PROPOSED MODEL
In this section, we first describe the seq2seq model with attention
as one of the baselines and as the base model we build our model
upon. We then introduce the general architecture of our model and
explain the query-aware attention and copy mechanism as two main
components employed in our model.
2.1
Seq2seq with Attention
As the base model, we employ seq2seq (seq2seq) with attention,
proposed by Bahdanau et al. [3] for the task of machine translation,
which is able to jointly learn the translation and the alignment. In
general, the seq2seq model is an RNN-based encoder-decoder in
which the encoder learns the representation for the source sequence,
and the decoder generates the target sequence.
The encoder is a bidirectional recurrent neural network (RNN)
that reads the input sequence X = [x 1 , x 2 , ... , x n ] in the left-toright direction in the RNN forward pass. It creates sequence of
→
− →
−
→
−
→
−
→
−
hidden states, [ h 1 , h 2 ,..., h n ], where h i = RNN(x i , h i−1 ) is a dynamic function for which we can use for example an LSTM [20] or
a GRU [7]. The RNN backward pass reads X in the reverse direction,
←
−
←
−
i.e. h i = RNN(x i , h i+1 ) resulting in a sequence of hidden states
←
− ←
−
←
−
[ h n , h n−1 ,..., h 1 ].
The forward and backward states for each time step are concatenated to create the encoder hidden states [h 1 ,h 2 ,...,hn ], in which
→
− ←
−
hi = [ h i ; h i ] is the encoded information from the corresponding
token x i in the source sequence X , taking it’s surrounding tokens
into consideration. The encoded hidden state are then summarized
using a function ϕ to generate a fixed length vector which is called
context vector: c =ϕ(h 1 ,h 2 ,...,hn ).
Encoder
Decoder
bio
dylan
query-aware attention
query level
attention
switch gate
switch gate
<UNK>
dylan
photo
bio
copier
word level
attention
copier
embedding
layer
. . .
…
<q>
bob
</q>
dylan
bob
dylan
photo
</q>
…
Figure 2: General schema of our proposed model for query suggestion.
The decoder is a unidirectional RNN with hidden states st which
unfolds the context vector to the target sequence. During the decoding process, in order to predict the target token at each time step t,
the attention mechanism is used to dynamically change the context
vector c t . To do so, for each token x i in the source sequence, we
compute a weight ai that determines the contribution of hi , to the
t th token in the target sequence:
lt,i =η(st −1 ,hi )
(1)
exp(lt,i )
at,i = Ín
j exp(l t, j )
(2)
where η is a mapping (usually a multilayer perceptron (MLP) or a
bilinear function), st −1 is the previous decoder’s hidden state. This
mapping will give us the logits l which are normalized with a softmax
layer to get the final attention weights a. The context vector c t is
calculated as follow:
n
Õ
c t = at,i hi
(3)
i
In the decoder, the hidden state is computed based on the previous
state st −1 , the context vector c t , and the previous output token yt −1 :
st = RNN(st −1 ,yt −1 ,c t )
(4)
An output projection layer is used to compute the conditional distribution over the vocabulary:
p(yt |y <t ,X ) = fo (st )
(5)
In the seq2seq model encoder and decoder are jointly trained to
maximize the conditional log-likelihood of the Yi , given X i .
2.2
Overview of Proposed Changes
Our proposed model is an extended seq2seq model in which we
employ a hierarchical attention mechanism to create a query-aware
attention. We will show in the experiments that this enables us to
not only control the scope of the queries in the session from which
we generate the next query, but also to automatically detect session
boundaries. Furthermore, regarding the fact that on average 62% of
terms of a query in a session come from the previously submitted
queries in the same session [41] and they are mostly rare terms, we
empower our model to access the source words from the session
context during decoding by incorporating a copy mechanism. A
general schema of the proposed model is depicted in Figure 2. In the
following, we describe these two functionalities in detail.
2.3
Query-Aware Attention
We already described the attention mechanism which is in charge
of assigning weights to the hidden states of the encoder. We will call
this the word-level encoder and word-level attention mechanism
in the remainder of this paper. We now add a bidirectional RNN
query-level encoder, that reads the input sequence Q = [q 1 ,q 2 ,...,qm ],
whereas q j is the encoding of query j (blue component in Figure 2).
To compute q j we employ the fact that at word-level the input is
linearized by concatenating all queries, each followed by a special
token h/qi. Let K ⊂ {1,...,n} be the indices of these special tokens,
i.e. the inputs {x k |k ∈ K } corresponds to h/qi. Note that |K | = m.
We annotate the elements in K so that the following holds true:
ka <kb ⇔a <b. We now define the encoding of query j as:
→
−
q j = id( h k j )
(6)
with id being the identity function in our experiments. However, this
function can also be replaced by a small MLP. The query-level encoder
is again a bidirectional RNN with the sequence of hidden states of
−д ,→
−
→
−
→
−
the forward RNN being [→
1 д 2 ,..., д m ], where дj = RNN(q j , д j−1 ),
i.e. the query encoding q j is the input to the query-level encoder.
Analog the backward RNN hidden states [←
д−m , ←
д−m−1 ,..., ←
д−1 ] are
computed by ←
д−j = RNN(q j ,←
д−j+1 )
The forward and backward states for each time step are again
concatenated to create the encoder hidden states [д1 ,д2 ,...,дm ], in
−д ;←
−
which дi = [→
i д i ].
We now define the query-level attention weights over the query
encoding:
q
lt, j =η(st −1 ,дj ,yt −1 )
q
at, j =
q
exp(lt, j )
Ín
q
i exp(l t,i )
(7)
(8)
To get the final query-aware attention weights (red component
in Figure 2) we multiply the word-level attention weight aw (previ-
ously a in Equation 2) with the corresponding query-level attention
weights and renormalize it as shown below:
q
aw
t,i a t, j
at,i = Ín w q ,
i 0 a t,i 0 a t, j 0
(9)
whereas j and j 0 is the smallest j, respectively j 0 , so that i ≤ k j , respectively i 0 ≤ k j0 , holds true. Finally, the context vector is computed
as stated in Equation 3.
2.4
2.5
Copy and Generate for Decoding
As already mentioned, we use an output projection layer to compute
the conditional distribution over the vocabulary. This allows us to
generate words, that are part of our vocabulary (yellow components
in Figure 2):
p(yt |y <t ,X ,generate) = fo (st )
(10)
However, a large vocabulary will slow down the training process and
increase the model size. Additionally, some entities or numbers (e.g.
telephone numbers) will not be part of even a very big vocabulary.
We therefore employ a copy mechanism (green component in
Figure 2). This copy mechanism (also know as Pointer Network [46])
computes a probability distribution over the input sequence, not
by using a fixed length output projection matrix but by using the
hidden states of the encoder:
p
lt,i =η(st ,hi )
p(yt =x i |y <t ,X ,copy) =
(11)
p
exp(lt,i )
Ín
p
j=1 exp(l t, j )
(12)
Note, that this equation is very similar to the attention mechanism.
Instead of using the final weights as attention weights, we use them
as probabilities for copying the underlying word. In order to be
able to handle words that have to be generated, i.e. are not in the
input, we define x 0 to be the unknown token (hU N Ki) and change
the equation above as follows:
p
lt,0 =η(st ,e(x 0 ))
p
lt,i =η(st ,hi )
p(yt =x i |y <t ,X ,copy) =
(13)
i >0
p
exp(lt,i )
Ín
p
j=0 exp(l t, j )
(14)
(15)
with e being an embedding function. We also include a switch gate,
that decides whether the generated or the copied word is used as the
final prediction of the decoder (grey component in Figure 2). The
switch gate has to make a binary decision and is defined as follows:
p(copy) =σ (w T st )
p(generate) = 1−p(copy)
In the following step we again produce the top b tokens, this time
for each of the tokens of the first step. We store them together with
the probability of predicting the sequence y1 ,y2 :
y21 ,p(y21 |y11 ) ,..., y2b ,p(y2b |y11 ) ,..., y21 ,p(y21 |y1b ) ,..., y2b ,p(y2b |y1b )
(20)
We then drop all but the b most likely decoding paths, which will
be use to produce the next token.
(16)
(17)
whereas w is a weight vector and σ is the sigmoid function, i.e.:
1
σ (x) =
(18)
1−exp(−x)
In other words, we have a fully-connected layer with an output size
1 and the sigmoid as the activation function.
During decoding, we perform a beam search of size b. At each
decoding step t we produce the top b tokens and store them together
with their probabilities:
y11 ,p(y11 ) ,..., y1b ,p(y1b )
(19)
Multi-Objective Training
To train our model we define two losses. The loss of the generator
is the averaged cross entropy between our predicted probability distribution p and the target probability distribution q, i.e. the one-hot
encoding of the target token:
1
1 Õ
lossgenerate =
pv logqv
(21)
H (p,q) =
|V |
|V |
v ∈V
with V being the vocabulary and |V | its size. Similarly we define
the loss of the copier as being the cross-entropy averaged over the
length of the input:
1 Õ
1
H (p,q) =
px logq x
(22)
losscopy =
|X |
|X |
x ∈X
In order to avoid producing hU N Ki, i.e. the output of copier when
the target is not in the source sequence, and hOOV i, i.e. the output
of generator when the target is not in the vocabulary, we introduce
a set of rules defining if the switch gate should favor the copier or
the generator at a given time step:
(1) target copier is hU N Ki and target generator is not hOOV i:
the switch gate shall choose generation (t switch = 0).
(2) target copier is not hU N Ki and target generator is hOOV i:
the switch gate shall choose copying (t switch = 1).
(3) target copier is hU N Ki and target generator is hOOV i: the
switch gate shall choose generation (t switch = 0).
(4) target copier is not hU N Ki and target generator is not
hOOV i: the switch gate shall choose copying (t switch = 1).
While the first two rules are an obvious choice, the last two rules
derive from the fact that we want to choose a target label for the
switch gate to copy as much as possible from the input and let the
generator handle the rest. Note that the loss of the generator will be 0
if the target token is hOOV i and the generator predicted hOOV i. The
same applies to the copier choosing hU N Ki. Now, based on these
rules we can define the loss of the switch gate to be the following:
lossswitch = (p(copy)−t switch )2
(23)
During the backward pass of the back-propagation algorithm, we
update the parameters of the network with respect to the losses in
three separate steps. In step one, we use the gradient calculated from
losscopy to update all parameters of the network, except those just
belonging to the switch gate and the generator, by freezing these
components (yellow and gray components in Figure 2). In step two,
we propagate the gradients calculated from lossgenerate to update all
parameters of the network except the parameters of the switch gate
and the copier (green and gray components in Figure 2). In the last
stage, we update the parameters of the network using the gradients
from lossswitch , while the parameters of copy and generator (green
and yellow components in Figure 2) are fixed. We empirically found
that updating parameters in separate stages instead of combining
losses as one loss leads to better results.
3
EVALUATION
With the increase of interest in generative models and in particular
neural network based models, we believe automatic evaluations
have to be studied as part of this research.
In this paper, we evaluate the performance of our proposed model
and the baselines using two paradigms. In the first paradigm, similar to the previous studies, we assess the ability of the model as a
discriminative algorithm. While in the second one, we introduce
metrics to investigate the quality of the model in terms of generating
data as a step toward evaluation of the generative models for the
query suggestion task.
3.1
Evaluation based on Discrimination
In this paradigm, similar to [31, 33, 36, 42] we evaluate our model as a
feature to score the candidate queries and use it within the learningto-rank (L2R) framework for ranking candidate suggestions. For the
sake of a fair comparison, we replicate the experimental setups in the
paper by Sordoni et al. [42] by generating candidates using the cooccurrence based model they used and also extract the set of features
they employed to be used in an L2R method as the BaseRanker. We
compare the performance of the BaseRanker with a similar ranker
that gets the score from our model as an additional feature.
The score of a query q = [y1 ,...,yn ] given the context X is calculated based on the output of the generator, the copier, and the switch
gate for each term in q as follows:
n
Ö
p(q|X ) =
p(generate|y <t ,X ) p(yt |y <t ,X ,generate)
(24)
t =1
+p(copy|y <t ,X ) p(yt |y <t ,X ,copy)
We use the mean reciprocal rank (MRR) to measure the quality of
the ranking.
3.2
Evaluation based on Generation
Besides the aforementioned evaluation paradigm, we also evaluate
the quality of our model’s output as a generative model. Loosely
speaking, we try to evaluate how similar the generated query is to
the target query taken from ground truth. To do so, we introduce
the following metrics.
Word Overlap Based Query Similarity. As a word overlap based
metrics, we consider Position Independent Word Error Rate (denoted
as PER), which is the minimum number of word insertions and deletions necessary to transform the generated query into the target
query by neglecting the words order, normalized by the length of
the target query.
Embedding Based Query Similarity. Word overlap based metrics
are not efficient in assessing the quality of the generative models as
they may generate queries that are as effective as the target query
without sharing any term with it. For example consider the target
query city hall phone number and a system that generated query municipality contact information, where the score is zero in terms of PER.
To tackle this issue, we can use soft matching metrics. In this paper,
we used an embedding-based query similarity metric (denoted as
simemb ) to assess the semantic matching of the target and generated
query. We first calculate the query-level embeddings using vector extrema [15] for which we use pretrained word embeddings (trained on
the GoogleNews corpus) and for each dimension of the word vectors,
we take the most extreme value amongst all word vectors in the query
and use that value in the query-level embedding. This approach prioritizes words carrying important semantic information over common
ones [28]. Then, we compute the similarity between the generated
query and the target query vectors using the cosine similarity.
Retrieval Based Query Similarity. In the real world’s application
of the query suggestion task, the suggested query is going to be
submitted to a search engine if it was selected by the user. Utilizing
this fact, we can evaluate the quality of the suggested query in terms
of how similar the retrieved results from the search engine with
respect to this query are to the results retrieved using the target
query. To this end, we suggest three different evaluation metrics:
To calculate the first metric, we submit the target query to an
external collection of documents and retrieve the ranked list of top
documents using a retrieval function and consider this list as the
reference list. We do the same but given the generated query. Then,
we use a ranking similarity metric to calculate the agreement of these
two ranked lists (denoted as simret ). We are not directly measuring if
the suggested query is addressing the actual user information need.
But we can reasonably conclude how well we do in terms of helping
a user which lacks the ability for query reformulation by suggesting
him a query which retrieves results that are similar to those of a
well-reformulated query.
To implicitly tackle this issue of not measuring how we address
the user information need, we define the second metric. We consider the ranked list retrieved from the expanded target query using
Pseudo Relevance Feedback (PRF) as the reference list. It has been
shown that PRF usually improves the performance of retrieval [9, 18]
by decreasing the vocabulary gap of a query and relevant documents.
Hence considering its results as the reference list, we can estimate
how well the generated query performs in terms of retrieving results
from an (in average) better version of the target query, which is more
likely to address the user information need (denoted as sim+
ret ).
In order to even better evaluate how good the generated query
is addressing the actual user information need, we take all sessions
with length l > 2 from the test data. We then use the first bl/2c queries
in the session as the context for generating the next query. After this,
we retrieve the ranked lists of documents for each of the next dl/2e
queries in the session and merge them using normalized scoring.
The merged list is used as the reference ranked list and we calculate
its agreement with the retrieved results given the generated query
(denoted as sim++
ret ).
The two last proposed metrics are in fact aiming at better evaluation of generated suggestions with respect to the actual user information needs when the user judgments are not available.
4
EXPERIMENTS AND RESULTS
In this section, we first describe the dataset we used for training,
validation, and testing as well as our experimental setups. Then we
present the main results of the model followed by discussions and
analysis.
4.1
Dataset and Experimental Setups
‘ We used AOL search logs [34] as the largest publicly available
search log which is also used by similar research works for the evaluation [4, 10, 21, 38, 42]. This query set contains web queries initiated
by 657,426 unique users in the AOL search engine that were sampled
from a three-months period from March 1, 2006, to May 31, 2006. We
preprocessed the data by eliminating non-alphanumeric characters,
spelling error correction, and lowercasing. Then we segmented the
log into sessions, using a simple standard segmentation heuristic,
i.e. intervals of at least 30 minutes idle time denotes a session boundary [24]. The final dataset consists of ∼ 9 million queries in ∼ 3
million sessions. We sorted the sessions time-wise and partitioned
them into three parts: the main training set, consisting of 70% of the
sessions which is used to train our proposed model and the baseline
methods. We also created an L2R training set, which includes 20% of
sessions for training the L2R algorithm (just used in the evaluation
based on discrimination). 10% of sessions are used as a test set.
The evaluation based on discrimination is similar to previous research [36, 42]. Given the anchor query, i.e. the last query in the
context, we first select top-20 candidate queries ranked based on the
frequency of their co-appearances with the anchor query in the same
session in the main training set, as the Most Popular Suggestions
(MPS). We then extract a set of 17 features employed by Sordoni
et al. [42]. These include:
(1) Features which capture the whole session history, like the score
calculated by Query Variable Markov Model as one of the contextaware query suggestion models [19], character n-gram similarity
between the candidate query and the 10 most recent queries in
the context [31, 40], and average Levenshtein distance between
the candidate and queries in the context [25].
(2) Features which only take the most recent query into account,
like frequency of the anchor query in the main training set, the
number of times the candidate follows the anchor query in the
main training set, and Levenshtein distance between the anchor
query and the candidate [25, 33].
(3) Features which characterize the candidate independently, like
frequency of the candidate in the main training set and the length
of the candidate in terms of the number of words and characters.
When we extracted features for all candidates, we train the LambdaMART as the L2R method by labeling the target query as relevant
and all others as non-relevant. We call this trained model BaseRanker. We use the BaseRanker in addition to the score from Sordoni
et al. [42] paper as an additional feature, as one of the baselines.
Besides these baselines, we train the original seq2seq model with
attention (explained in Section 2.1) with the whole context flattened
to a sequence as the source sequence and the target query as the
target sequence. We also add the score from this model to the BaseRanker as an additional feature as one of the baselines. We also add
the score from our model (calculated using 24) to the base-ranker to
assess how it further helps to improve the quality of the ranking of
suggestions. Besides this, we want to investigate how different part
of our model including the query-aware attention and the copy mechanism (CM) individually affect the performance of query suggestion,
So we also evaluate the seq2seq model with query-aware attention
only (seq2seq + QaA), and seq2seq model with copy mechanism
(seq2seq + CM) only as additional baselines.
In the evaluation based on generation, for the retrieval based
Table 1: Performance of the different methods as discriminative models. (x ) indicates that the improvements with
respect to the method in row x is statistically significant,
at the 0.05 level using the paired two-tailed t-test with
Bonferroni correction.
#
Model
MRR
1
2
3
4
5
6
7
MPS
BaseRanker
BaseRanker + Seq2Seq
BaseRanker + HRED [42]
BaseRanker + (Seq2Seq + QaA)
BaseRanker + (Seq2Seq + CM)
BaseRanker + ACG
0.5216
0.5530(1)
0.5679(1,2)
0.5727(1,2)
0.5744(1,2)
0.5851(1,2,3,4,5)
0.5941(1,2,3,4,5,6)
metrics, we use ClueWeb09 Category B corpus with over 50 million
English documents as the external collection of documents. As the
retrieval function, we use JS-divergence retrieval model with the
Dirichlet prior smoothing and as the PRF method, we use RM3 [1].
We further use the rank-biased overlap (RBO) [47] at level 100 to
measure the agreement of two ranked lists and report the average
RBOs over all instances in the test set.
We used TensorFlow [13, 44] to implement our model. The parameters of our model are optimized employing the Adam optimizer [26]
and using the computed gradient of the loss to perform the backpropagation algorithm. With respective to the main validation set,
which is 10% of the sessions from the main training set we tuned
some hyper-parameters of the model using batched GP bandits with
an expected improvement acquisition function [11]. In our model,
the number of hidden neurons in each of encoders (forward and backward; word-level and query-level) and the decoder were selected
from [64,128,256]. The initial learning rate and the dropout parameter were selected from {10−3 ,10−5 } and {0.0,0.2,0.5}, respectively.
We considered embedding sizes {300,500} for the input embedding
of the word level encoder. The batch size in our experiments was set
to 128. We set the vocabulary size to 90k (the same for the baselines),
and the beam size for the beam search decoding was set to 4.
It is noteworthy that for all the baseline models we also tuned
the hyper-parameters if their optimized value is not reported in the
corresponding papers.
4.2
Main Results
In this section, we first report the main results of our proposed model
in the task of query suggestion compared to the baselines, both in
terms of discrimination and generation quality.
Evaluation based on Discrimination. First, we evaluate the performance of our model as a discriminative model in the setup explained in Section 3.1. Table 1 presents the performance of our model
and the baseline models. As it can be seen, by adding the score from
ACG, we can gain the highest improvement compared to the baseline
models. This is mainly because of the fact that in the scoring process
of candidates, rare terms from the true target query that occur in the
candidate query have a chance of getting a high probability in the
distribution learned by ACG if they appear in one of the previously
submitted queries.
Table 2: Performance of the different methods as generative models.
#
Method
Overlap Based
Embedding Based
PER (%)
simemb
simret
Retrieval Based
sim+
ret
sim++
ret
1
2
3
4
seq2seq
BaseRanker + seq2seq (top-1)
seqsSeq + QaA
seq2seq + CM
84.11 (+− 6.3)
72.23 (+− 8.1)
80.90 (+− 5.0)
71.16 (+− 3.5)
0.5170 (+− 0.003)
0.5019 (+− 0.006)
0.5517 (+− 0.004)
0.6119 (+− 0.003)
0.1630 (+− 0.008)
0.4375 (+− 0.009)
0.2012 (+− 0.009)
0.3563 (+− 0.009)
0.2424 (+− 0.009)
0.3751 (+− 0.008)
0.2916 (+− 0.008)
0.4173 (+− 0.009)
0.1955 (+− 0.008)
0.3916 (+− 0.008)
0.2330 (+− 0.008)
0.3950 (+− 0.008)
5
6
HRED [42]
BaseRanker + HRED [42] (top-1)
81.50 (+− 4.9)
72.36 (+− 7.3)
0.5455 (+− 0.004)
0.5200 (+− 0.004)
0.2667 (+− 0.008)
0.4504 (+− 0.009)
0.3250 (+− 0.009)
0.3812 (+− 0.009)
0.3443 (+− 0.007)
0.4091 (+− 0.007)
7
8
ACG
BaseRanker + ACG (top-1)
68.03 (+− 3.6)
70.66 (+− 7.1)
0.6473 (+− 0.004)
0.5196 (+− 0.004)
0.3612 (+− 0.008)
0.4594 (+− 0.008)
0.4366 (+− 0.009)
0.3927 (+− 0.009)
0.4315 (+− 0.008)
0.4111 (+− 0.007)
It is interesting that the improvement of just having the hierarchical setup for reading the session history, i.e. the source sequence
in both HRED and seq2seq with query-aware attention is not statistically significant with respect to the seq2seq model with a simple
attention mechanism (models 4 and 5 compared to model 3). On the
other hand, when the model is equipped with the copy mechanism,
the performance improvement by adding query-aware attention is
statistically significant (model 7 compared to model 6). This suggests
that taking query-level attention into account along with the wordlevel attention, is particularly important for the copy mechanism
which directly uses the attention weights as output probabilities, but
is less important for the decoder itself.
Evaluation based on Generation. The next set of experiments
aims at assessing the quality of the models in terms of generating the
next query according to the metrics we introduced in Section 3.2. Table 2 presents the performance of different models. We also evaluate
the top ranked query from different models used in the evaluation
based on discrimination setup, but in terms of evaluation based on
generation.
Among the metrics for evaluation based on generation, PER is a
strict metric as it is based on the exact word overlap between generated query and target query. Nonetheless, the generated next query
by ACG has the lowest error in terms of PER, even lower than the
corresponding discriminative model in which the score from ACG
is used as one of the features (model 7 compared to model 8). The
relatively low error rate of model 4 is an indicator that the success of
ACG to generate exact words in the target query is mostly reasoned
by its ability to copy terms from the context queries. It is supported
by the statistics that we extracted from our test set, where in average
38% of terms in the target queries are retained from the previous
queries in the same session.
Embedding similarity of the generated and the target queries
relaxes the hard assumption in PER by taking words semantic similarity into consideration. ACG performs better than all the baseline in
terms of generating queries that are semantically similar to the target
queries. Regarding the results of the models we report in this paper,
generative approaches lead to queries with higher simemb scores.
On the other hand, in discriminative approaches are in average more
successful in terms of PER. This is assumed to be reasoned by the
fact that in the generative models in this paper, we explicitly learn
representations for words, in a downstream task of learning query
suggestions, which helps to better capture the semantic similarity
among the words. During generation, the model tries to generate
the semantically most plausible words for the suggested query based
on the learned representations.
Regarding the retrieval based evaluation, with respect to the simret
metric, BaseRanker which uses ACG as one of its features outper++
forms all baseline models. In contrast, for both sim+
ret and simret our
ACG as a generative model is the best performing method. Since PRF
implicitly models the retrieval with soft matching, using sim+
ret we
also consider cases where the results given the generated query do
not have an immediate agreement with the results given the target
query, but the generated query is semantically related to the target
query. Also with sim++
ret , we move one step further by evaluating if
the results given the generated query are going to satisfy the user at
some point in the search session, even if it is not the case given the
immediate target query. This explains why we achieve better results
using generative approaches with respect to these two metrics.
4.3
Discussions and Analysis
In this section, we first investigate the generative ability of our model
in a multiple query suggestion setup. We then study the effect of
session length on the performance of ACG. At the end, we will assess
the robustness of ACG compared to the baseline in dealing with
noisy data.
Multiple Query Suggestion. Major search engines usually provide
more than one suggestion for the users. Here, we investigate the
quality of the top-10 queries that our model generates. To do so,
during the decoding, after generating the first suggestion, we ignore
the fact that the first suggestion was generated through the beam
search, i.e. h/qi was generated. Instead, we continue decoding until
the next suggestion is generated. We repeat this process to generate
10 suggestions which lead to a sorted list of queries based on their
likelihood of generation.
The performance of ACG considering the queries at different
ranks, i.e. position in the list of queries, with respect to the different
metrics, is illustrated in Figure 3.
As it is expected, in general, the quality of the generated queries
decreases as the rank of the query increases. The drops in terms of
++
simret , sim+
ret , and simret are less dramatic compared to simemb .
It is also really interesting that the performance of the secondly
generated query is relatively low compared to the first query in terms
of all metrics except the sim++
ret , where the score is even slightly higher
medium
large
Session length
(a) Evaluation based on Discrimination
1
2
3
4
5
6
Rank
7
8
9
10
Figure 3: Performance of the generated queries at different
ranks.
than the score for the first query. Although this boost is not statistically significant. But regarding the meaning of the sim++
ret metric
(Section 3.2), it seems the generated queries in the second and the
third rank are somehow future desirable queries which although
they do not match with the immediate user query, are related to the
ultimate user information need.
Effect of Session Length. In order to investigate the effect of session length on the performance of our model, we separate the test
set into three categories:
(1) sessions with 2 queries (short) 66.09% of the test set
(2) sessions with 3−4 queries (medium) 22.36% of the test set
(3) sessions with > 4 queries (long) 11.55% of the test set
We report the performance of our model in terms of both ability of
discrimination and generation, compared to the HRED [42], using
test sets with short, medium, and long sessions. The results are
depicted in Figure 4.
In Figure 4a, we evaluate the performance of our model and the
baseline model as an additional feature for the BaseRanker and measure the MRR of the ranked list of suggestions. ACG appears to
perform best in all test sets and its performance is not only stable
across the test sets with different session lengths but also progressive
when the length of the sessions increases. However, HRED seems
to fail for long sessions, which can be the result of topically broad
information needs or changes of the search topic. This problem is
quiet challenging, although our model seems to handle this problem
by dynamically attending to the most promising part of the context.
This is more the case for evaluation based on generation. However,
it has been shown that when the length of a session increases, the
percentage of repeating previously-used terms also increases [25],
which means that the next query is more likely to contain the terms
used before when it appears in the latter steps of a session. This is
where the ability of ACG for copying terms from the context queries
kicks in that compensates the performance loss which might occur
due to the complexity of the long sessions.
Attending the Promising Parts of the Context. Analyzing the
context of a session to understand which part is useful for query
suggestion is considered as one of the most challenging stages of this
task. This is due to the fact that noisy words are common to appear
0.3
0.4561
0.2716
0.4512
0.4282
0.3224
0.4
ACG
0.2
small
0.3
0.5
0.3353
0.6040
0.5861
0.5583
0.5
0.4
HRED
0.6
sim+
ret
0.5
0.6
ACG
0.5341
HRED
0.7
MRR
Performance
0.6
sim++
ret
0.6268
sim+ret
simret
0.5925
simemb
small
medium
large
Session length
(b) Evaluation based on Generative
Figure 4: Performance of ACG compared to HRED on
sessions with different lengths.
in a given context and this can happen in several circumstances:
Either when users switch to another search topic quickly after one
search session and the assumption of 30 minutes as the idle time
duration for detecting session boundaries fails. Or when users issue
navigational queries during a search session, like e.g. google. Or in
cases where users use non-discriminating words in their query that
are no effective indicators for the search intend.
In this situation, a session-aware query suggestion system should
be able to handle noisy parts by neglecting them. To test the ability
of our model compared to the baselines on being robust against these
noises, we manipulate the sessions in training, test, and validation
sets in three ways. We train and evaluate the query suggestion
models on this manipulated data:
• Noise term insertion: In order to assess the robustness of the models against noise terms in the session, we first select the 200 most
frequent terms excluding stopwords from the main training set.
Then, for each session in the data, we sample from this list with a
probability that is proportional to the frequency of the terms and
insert the sampled noise term at a random position of a random
query in the session.
• Noise query insertion: In order to assess the robustness of the models against noise queries, similar to the previous case, we extract a
list of 100 most frequent queries in the main training set as noise
queries. For each session, we sample from the list and insert the
noise query at a random position in the session.
• Noise session insertion: In the last case, we aim at assessing the
ability of the models in session boundary detection. To do so, for
each session, we randomly pick another session from the same
user, if there is any, and insert it at the beginning of the session.
Table 3 presents the results of seq2seq, HRED, and ACG as generative models in terms of simemb . As shown, ACG achieves a significantly better performance in all situations compared to HRED and
seq2seq. In terms of performance loss compared to the non-noisy
case (results Table 2), all models are relatively robust to noise term
insertion. However, seq2seq is not able to handle noise query and
noise session insertion. Although both HRED and ACG seem to be
still robust against noise query insertion, inserting noise sessions
considerably affects the performance of HRED, which is also in accordance with the performance drop of HRED in encoding long sessions
(Figure 4). This is while ACG controls the performance loss in this
situation by providing insight into how different parts in the given
context contribute to the next query generation. In other words,
Table 3: Performance (and performance loss) of the different
methods as generative models on noisy data, in terms of
simemb .
Method
Noise term
insertion
Noise query
insertion
Noise session
insertion
seq2seq
HRED [42]
ACG
0.4973 (-3.8%)
0.5380 (-1.4%)
0.6366 (-1.6%)
0.4419 (-14.5%)
0.5140 (-5.8%)
0.6019 (-7.0%)
0.3969 (-23.2%)
0.4505 (-17.4%)
0.5878 (-9.1%)
ACG is able to detect the boundary of the session by attending to the
promising scope of the context.
Error Analysis. As discussed and showed in the experiments, the
main superiority of the model over the HRED, i.e. the main baseline,
is its ability to copy an arbitrary part of the input sequence during
decoding. We analyzed cases where ACG fails against HRED, in
different situations and the main point of failure of ACG is where
we remove the spelling correction from the preprocessing. Looking
into the output of our model we noticed that it usually fails in cases
where the next query in the session is just a spelling correction of
the previous queries and the copy mechanism tends to repeat the
spelling error. For instance, given the query “bebefit pediatrician”,
ACG suggests “bebefit pediatric center” as the next query, while the
obvious reformulation for the given query would be the replacement
of “bebefit” with “benefit”. The easiest way to tackle this problem is
to have spell error correction as part of the preprocessing. We also
consider training the model with a spelling correction dataset that
helps to pick generating the correct version of words with errors
instead of copying them. We will leave this for future work.
5
RELATED WORK
Query Suggestion. The biggest challenge in the task of query suggestion is to understand the actual user intent for suggesting the next
query. To aim this, several studies leveraged the “wisdom of crowds”
by mining the session structure and other information in query log
data to find alternative queries for suggestions. For example, Huang
et al. [22] tried to find within session query pairs co-occurrences and
rank suggestions based on the frequency of co-occurrence. Boldi
et al. [5] took the structure of the sessions into account by building a query-flow graph to estimate how likely a user moves from
one query to the next query in the same session. Another group of
methods tries to find similar queries in log data. It is assumed that
similar queries have larger overlap between their respective clicks
and based on this assumption. E.g., Mei et al. [29] used a random walk
over a bi-partied query-document graph or Baeza-Yates et al. [2]
used k-means to find similar queries based on clicked data. He et al.
[19] proposed a session-based method based on Variable Memory
Markov model (QVMM) to build a suffix tree to model the user query
sequence for query suggestions. Santos et al. [36] and Ozertem et al.
[33] tried to first extract candidates and then employed learning-torank methods for ranking suggestions considering several features
(similar to our BaseRanker baseline). These methods perform poorly
for long-tail queries. To tackle this issue, Vahabi et al. [45] proposed
to find suggestions for long-tail queries by comparing their search
results. Another group of research tries to generate new suggestions
by leveraging search logs and external resources. E.g., Szpektor et al.
[43] used WordNet along with a template generation method, or Jain
et al. [23] used a CRF for segmenting queries followed by a machine
learning stage to filter out poor suggestions. Recently, Sordoni et al.
[42] proposed a context-aware method which uses a hierarchical
RNN to encode the session information and generate the sequence
of words as the next query. Their work is the most similar to ours.
We additional introduce the query-aware attention for capturing the
hierarchical structure of the session. We also augment our model
with copy mechanism to better model the query reformulation which
lead to a significant improvement in the performance.
Hierarchical Structure Decoding. There are several attempts to
make recurrent neural networks able to deal with structured data.
E.g., Li et al. [27] proposed a hierarchical auto-encoder to encode
and reconstruct multi-sentence paragraphs, taking both word and
sentence levels into account. Serban et al. [39] used a similar architecture to Sordoni et al. [42] paper in the context of dialogue systems
to encode the context of the dialogue at the utterance level. Chung
et al. [8] proposed a model encodes the temporal dependencies with
different timescales by updating probabilities for different units to
capture the latent hierarchical structure in the sequence. Yang et al.
[48] proposed a two-level attention mechanism which uses the word
level attention to lean the sentence level attention to improve the
performance of document classification. The query-aware attention
mechanism we propose in this paper is similar to theirs. However, we
learn the query-level and word level attentions in separate processes
and then integrate information from both during decoding.
Incorporating Copy Mechanism. The idea of incorporating a
copy mechanism originally comes from the pointer networks [46],
which is in fact a seq2seq model with attention that produces an
output sequence consisting of elements from the input sequence.
The pointer network has been used to create hybrid approaches that
mix copying (pointing) and generation during decoding in different
tasks like neural machine translation [17], language modeling [30],
and summarization [16, 32, 37]. Our approach is the close to Gu et al.
[16] and See et al. [37] works but we considered learning to copy,
to generate, and to make the decision of copying or generating as
separate tasks by using multi-objective learning.
6
CONCLUSION AND FUTURE WORKS
In this paper, we proposed a session-aware query suggestion model
by augmenting the seq2seq model with a query-aware attention
mechanism to make it able to encode the structure of the session. We
also incorporate a copy mechanism during decoding which helps to
model term retention in query reformulation. Finally, we proposed
new metrics to evaluate generative models on the task of query suggestion. Our experiments show that our proposed model boost the
performance of seq2seq models and outperforms baselines both in
terms of discrimination and generation.
For future work, we are going to extend our model to integrate
information from clicked documents as additional signals in the process of generating the next query. This information also can be used
for the evaluation based on generation paradigm. For evaluation
purposes, we can also use the content of documents retrieved using
the generated query and estimate how well the generated query
addresses the user information need in terms of the ability to provide
relevant content. Another interesting direction is to evaluate our
model in a query auto-completion task. This is possible by changing
the setting of the model to have a character-based seq2seq model.
Besides encoding the previously submitted queries in the session, we
can feed the decoder with the prefix of the new query and generate
the most probable sequence of following characters.
REFERENCES
[1] Nasreen Abdul-Jaleel, James Allan, W Bruce Croft, Fernando Diaz, Leah Larkey,
Xiaoyan Li, Mark D Smucker, and Courtney Wade. 2004. UMass at TREC 2004:
Novelty and HARD. (2004).
[2] Ricardo Baeza-Yates, Carlos Hurtado, and Marcelo Mendoza. 2004. Query
recommendation using query logs in search engines. In International Conference
on Extending Database Technology. Springer, 588–596.
[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine
translation by jointly learning to align and translate. In Proceedings of the
International Conference on Learning Representations (ICLR’15).
[4] Ziv Bar-Yossef and Naama Kraus. 2011. Context-sensitive query auto-completion.
In Proceedings of the 20th international conference on World wide web. ACM,
107–116.
[5] Paolo Boldi, Francesco Bonchi, Carlos Castillo, Debora Donato, and Sebastiano
Vigna. 2009. Query suggestions using query-flow graphs. In Proceedings of the
2009 workshop on Web Search Click Data. ACM, 56–63.
[6] Huanhuan Cao, Daxin Jiang, Jian Pei, Qi He, Zhen Liao, Enhong Chen, and
Hang Li. 2008. Context-aware query suggestion by mining click-through and
session data. In Proceedings of the 14th ACM SIGKDD international conference on
Knowledge discovery and data mining. ACM, 875–883.
[7] Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau,
Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase
representations using RNN encoder-decoder for statistical machine translation.
arXiv preprint arXiv:1406.1078 (2014).
[8] Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. 2017. Hierarchical multiscale
recurrent neural networks. In Proceedings of the International Conference on
Learning Representations (ICLR’17).
[9] Mostafa Dehghani, Hosein Azarbonyad, Jaap Kamps, Djoerd Hiemstra, and
Maarten Marx. 2016. Luhn revisited: Significant words language models. In
Proceedings of the 25th ACM International on Conference on Information and
Knowledge Management. ACM, 1301–1310.
[10] Mostafa Dehghani, Hamed Zamani, Aliaksei Severyn, Jaap Kamps, and W. Bruce
Croft. 2017. Neural Ranking Models with Weak Supervision. In Proceedings of
the 40th International ACM SIGIR Conference on Research and Development in
Information Retrieval (SIGIR ’17). 65–74.
[11] Thomas Desautels, Andreas Krause, and Joel W Burdick. 2014. Parallelizing
exploration-exploitation tradeoffs in Gaussian process bandit optimization.
Journal of Machine Learning Research 15, 1 (2014), 3873–3923.
[12] Carsten Eickhoff, Jaime Teevan, Ryen White, and Susan Dumais. 2014. Lessons
from the journey: a query log analysis of within-session learning. In Proceedings
of the 7th ACM international conference on Web search and data mining. ACM,
223–232.
[13] Martı́n Abadi et al. 2015. TensorFlow: Large-Scale Machine Learning on
Heterogeneous Systems. (2015). http://tensorflow.org/ Software available from
tensorflow.org.
[14] Marzieh Fadaee, Arianna Bisazza, and Christof Monz. 2017. Data Augmentation
for Low-Resource Neural Machine Translation. In Proceedings of the Annual
Meeting of the Association for Computational Linguistics (ACL’17).
[15] Gabriel Forgues, Joelle Pineau, Jean-Marie Larchevêque, and Réal Tremblay. 2014.
Bootstrapping dialog systems with word embeddings. In NIPS, Modern Machine
Learning and Natural Language Processing Workshop.
[16] Jiatao Gu, Zhengdong Lu, Hang Li, and Victor OK Li. 2016. Incorporating copying
mechanism in sequence-to-sequence learning. In Proceedings of the Annual
Meeting of the Association for Computational Linguistics (ACL’16).
[17] Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua
Bengio. 2016. Pointing the unknown words. In Proceedings of the Annual Meeting
of the Association for Computational Linguistics (ACL’16).
[18] Donna Harman and Chris Buckley. 2009. Overview of the reliable information
access workshop. Information Retrieval 12, 6 (2009), 615.
[19] Qi He, Daxin Jiang, Zhen Liao, Steven CH Hoi, Kuiyu Chang, Ee-Peng Lim, and
Hang Li. 2009. Web query recommendation via sequential query prediction. In
Data Engineering, 2009. ICDE’09. IEEE 25th International Conference on. 1443–1454.
[20] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural
computation 9, 8 (1997), 1735–1780.
[21] Bo-June Paul Hsu and Giuseppe Ottaviano. 2013. Space-efficient data structures
for top-k completion. In Proceedings of the 22nd international conference on World
Wide Web. ACM, 583–594.
[22] Chien-Kang Huang, Lee-Feng Chien, and Yen-Jen Oyang. 2003. Relevant
term suggestion in interactive web search based on contextual information in
query session logs. Journal of the American Society for Information Science and
Technology 54, 7 (2003), 638–649.
[23] Alpa Jain, Umut Ozertem, and Emre Velipasaoglu. 2011. Synthesizing high utility
suggestions for rare web search queries. In Proceedings of the 34th international
ACM SIGIR conference on Research and development in Information Retrieval. ACM,
805–814.
[24] J Jansen Bernard, Amanda Spink, Chris Blakely, and Sherry Koshman. 2007.
Defining a session on Web search engines. Journal of the American Society for
Information Science and Technology 58, 6 (2007), 862–871.
[25] Jyun-Yu Jiang, Yen-Yu Ke, Pao-Yu Chien, and Pu-Jen Cheng. 2014. Learning user
reformulation behavior for query auto-completion. In Proceedings of the 37th
international ACM SIGIR conference on Research & development in information
retrieval. ACM, 445–454.
[26] Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic
optimization. arXiv preprint arXiv:1412.6980 (2014).
[27] Jiwei Li, Minh-Thang Luong, and Dan Jurafsky. 2015. A hierarchical neural autoencoder for paragraphs and documents. arXiv preprint arXiv:1506.01057 (2015).
[28] Chia-Wei Liu, Ryan Lowe, Iulian V Serban, Michael Noseworthy, Laurent Charlin,
and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical
study of unsupervised evaluation metrics for dialogue response generation. (2016).
[29] Qiaozhu Mei, Dengyong Zhou, and Kenneth Church. 2008. Query suggestion
using hitting time. In Proceedings of the 17th ACM conference on Information and
knowledge management. ACM, 469–478.
[30] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017.
Pointer Sentinel Mixture Models. In Proceedings of the International Conference
on Learning Representations (ICLR’17).
[31] Bhaskar Mitra. 2015. Exploring session context using distributed representations
of queries and reformulations. In Proceedings of the 38th International ACM SIGIR
Conference on Research and Development in Information Retrieval. ACM, 3–12.
[32] Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, and others. 2016.
Abstractive text summarization using sequence-to-sequence rnns and beyond. In
Proceedings of the Annual Meeting of the Association for Computational Linguistics
(ACL’16).
[33] Umut Ozertem, Olivier Chapelle, Pinar Donmez, and Emre Velipasaoglu.
2012. Learning to suggest: a machine learning framework for ranking query
suggestions. In Proceedings of the 35th international ACM SIGIR conference on
Research and development in information retrieval. ACM, 25–34.
[34] Greg Pass, Abdur Chowdhury, and Cayley Torgeson. 2006. A picture of search..
In InfoScale, Vol. 152. 1.
[35] Eldar Sadikov, Jayant Madhavan, Lu Wang, and Alon Halevy. 2010. Clustering
query refinements by user intent. In Proceedings of the 19th international
conference on World wide web. ACM, 841–850.
[36] Rodrygo LT Santos, Craig Macdonald, and Iadh Ounis. 2013. Learning to rank
query suggestions for adhoc and diversity search. Information Retrieval 16, 4
(2013), 429–451.
[37] Abigail See, Peter J Liu, and Christopher D Manning. 2017. Get To The Point:
Summarization with Pointer-Generator Networks. In Proceedings of the Annual
Meeting of the Association for Computational Linguistics (ACL’17).
[38] Christian Sengstock and Michael Gertz. 2011. CONQUER: a system for efficient
context-aware query suggestions. In Proceedings of the 20th international
conference companion on World wide web. ACM, 265–268.
[39] Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle
Pineau. 2016. Building end-to-end dialogue systems using generative hierarchical
neural network models. In Thirtieth AAAI Conference on Artificial Intelligence.
[40] Milad Shokouhi. 2013. Learning to personalize query auto-completion. In
Proceedings of the 36th international ACM SIGIR conference on Research and
development in information retrieval. ACM, 103–112.
[41] Marc Sloan, Hui Yang, and Jun Wang. 2015. A term-based methodology for query
reformulation understanding. Information Retrieval Journal 18, 2 (2015), 145–165.
[42] Alessandro Sordoni, Yoshua Bengio, Hossein Vahabi, Christina Lioma, Jakob
Grue Simonsen, and Jian-Yun Nie. 2015. A hierarchical recurrent encoder-decoder
for generative context-aware query suggestion. In Proceedings of the 24th ACM
International on Conference on Information and Knowledge Management. 553–562.
[43] Idan Szpektor, Aristides Gionis, and Yoelle Maarek. 2011. Improving recommendation for long-tail queries via templates. In Proceedings of the 20th international
conference on World wide web. ACM, 47–56.
[44] Yuan Tang. 2016. TF.Learn: TensorFlow’s High-level Module for Distributed
Machine Learning. arXiv preprint arXiv:1612.04251 (2016).
[45] Hossein Vahabi, Margareta Ackerman, David Loker, Ricardo Baeza-Yates, and
Alejandro Lopez-Ortiz. 2013. Orthogonal query recommendation. In Proceedings
of the 7th ACM conference on Recommender systems. ACM, 33–40.
[46] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In
Advances in Neural Information Processing Systems. 2692–2700.
[47] William Webber, Alistair Moffat, and Justin Zobel. 2010. A similarity measure for indefinite rankings. ACM Transactions on Information Systems (TOIS) 28, 4 (2010), 20.
[48] Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy.
2016. Hierarchical attention networks for document classification. In Proceedings
of NAACL-HLT. 1480–1489.
| 2cs.AI
|
Understanding Neural Networks Through Deep Visualization
Jason Yosinski
Cornell University
YOSINSKI @ CS . CORNELL . EDU
arXiv:1506.06579v1 [cs.CV] 22 Jun 2015
Jeff Clune
Anh Nguyen
University of Wyoming
JEFFCLUNE @ UWYO . EDU
ANGUYEN 8@ UWYO . EDU
Thomas Fuchs
Jet Propulsion Laboratory, California Institute of Technology
FUCHS @ CALTECH . EDU
Hod Lipson
Cornell University
HOD . LIPSON @ CORNELL . EDU
Abstract
Recent years have produced great advances in
training large, deep neural networks (DNNs), including notable successes in training convolutional neural networks (convnets) to recognize
natural images. However, our understanding of
how these models work, especially what computations they perform at intermediate layers, has
lagged behind. Progress in the field will be
further accelerated by the development of better tools for visualizing and interpreting neural
nets. We introduce two such tools here. The
first is a tool that visualizes the activations produced on each layer of a trained convnet as it
processes an image or video (e.g. a live webcam stream). We have found that looking at live
activations that change in response to user input
helps build valuable intuitions about how convnets work. The second tool enables visualizing
features at each layer of a DNN via regularized
optimization in image space. Because previous
versions of this idea produced less recognizable
images, here we introduce several new regularization methods that combine to produce qualitatively clearer, more interpretable visualizations.
Both tools are open source and work on a pretrained convnet with minimal setup.
Published in the Deep Learning Workshop, 31 st International
Conference on Machine Learning, Lille, France, 2015. Copyright
2015 by the author(s).
1. Introduction
The last several years have produced tremendous progress
in training powerful, deep neural network models that are
approaching and even surpassing human abilities on a variety of challenging machine learning tasks (Taigman et al.,
2014; Schroff et al., 2015; Hannun et al., 2014). A flagship
example is training deep, convolutional neural networks
(CNNs) with supervised learning to classify natural images
(Krizhevsky et al., 2012). That area has benefitted from the
combined effects of faster computing (e.g. GPUs), better
training techniques (e.g. dropout (Hinton et al., 2012)), better activation units (e.g. rectified linear units (Glorot et al.,
2011)), and larger labeled datasets (Deng et al., 2009; Lin
et al., 2014).
While there has thus been considerable improvements in
our knowledge of how to create high-performing architectures and learning algorithms, our understanding of how
these large neural models operate has lagged behind. Neural networks have long been known as “black boxes” because it is difficult to understand exactly how any particular, trained neural network functions due to the large number of interacting, non-linear parts. Large modern neural
networks are even harder to study because of their size;
for example, understanding the widely-used AlexNet DNN
involves making sense of the values taken by the 60 million trained network parameters. Understanding what is
learned is interesting in its own right, but it is also one
key way of further improving models: the intuitions provided by understanding the current generation of models
should suggest ways to make them better. For example,
the deconvolutional technique for visualizing the features
learned by the hidden units of DNNs suggested an architectural change of smaller convolutional filters that led to
state of the art performance on the ImageNet benchmark in
2013 (Zeiler & Fergus, 2013).
gus (2013), which highlights the portions of a particular
image that are responsible for the firing of each neural unit.
We also note that tools that enable understanding will especially benefit the vast numbers of newcomers to deep
learning, who would like to take advantage of off-the-shelf
software packages — like Theano (Bergstra et al., 2010),
Pylearn2 (Goodfellow et al., 2013), Caffe (Jia et al., 2014),
and Torch (Collobert et al., 2011) — in new domains, but
who may not have any intuition for why their models work
(or do not). Experts can also benefit as they iterate ideas for
new models or when they are searching for good hyperparameters. We thus believe that both experts and newcomers
will benefit from tools that provide intuitions about the inner workings of DNNs. This paper provides two such tools,
both of which are open source so that scientists and practitioners can integrate them with their own DNNs to better
understand them.
Network-centric approaches investigate a network directly
without any data from a dataset. For example, Erhan
et al. (2009) synthesized images that cause high activations for particular units. Starting with some initial input
x = x0 , the activation ai (x) caused at some unit i by
this input is computed, and then steps are taken in input
space along the gradient ∂ai (x)/∂x to synthesize inputs
that cause higher and higher activations of unit i, eventually terminating at some x∗ which is deemed to be a preferred input stimulus for the unit in question. In the case
where the input space is an image, x∗ can be displayed
directly for interpretation. Others have followed suit, using the gradient to find images that cause higher activations
(Simonyan et al., 2013; Nguyen et al., 2014) or lower activations (Szegedy et al., 2013) for output units.
The first tool is software that interactively plots the activations produced on each layer of a trained DNN for userprovided images or video. Static images afford a slow, detailed investigation of a particular input, whereas video input highlights the DNNs changing responses to dynamic input. At present, the videos are processed live from a user’s
computer camera, which is especially helpful because users
can move different items around the field of view, occlude
and combine them, and perform other manipulations to actively learn how different features in the network respond.
These gradient-based approaches are attractive in their simplicity, but the optimization process tends to produce images that do not greatly resemble natural images. Instead, they are composed of a collection of “hacks” that
happen to cause high (or low) activations: extreme pixel
values, structured high frequency patterns, and copies of
common motifs without global structure (Simonyan et al.,
2013; Nguyen et al., 2014; Szegedy et al., 2013; Goodfellow et al., 2014). The fact that activations may be effected by such hacks is better understood thanks to several recent studies. Specifically, it has been shown that
such hacks may be applied to correctly classified images
to cause them to be misclassified even via imperceptibly
small changes (Szegedy et al., 2013), that such hacks can
be found even without the gradient information to produce
unrecognizable “fooling examples” (Nguyen et al., 2014),
and that the abundance of non-natural looking images that
cause extreme activations can be explained by the locally
linear behavior of neural nets (Goodfellow et al., 2014).
The second tool we introduce enables better visualization
of the learned features computed by individual neurons at
every layer of a DNN. Seeing what features have been
learned is important both to understand how current DNNs
work and to fuel intuitions for how to improve them.
Attempting to understand what computations are performed at each layer in DNNs is an increasingly popular direction of research. One approach is to study each layer as
a group and investigate the type of computation performed
by the set of neurons on a layer as a whole (Yosinski et al.,
2014; Mahendran & Vedaldi, 2014). This approach is informative because the neurons in a layer interact with each
other to pass information to higher layers, and thus each
neuron’s contribution to the entire function performed by
the DNN depends on that neuron’s context in the layer.
With such strong evidence that optimizing images to cause
high activations produces unrecognizable images, is there
any hope of using such methods to obtain useful visualizations? It turns out there is, if one is able to appropriately
regularize the optimization. Simonyan et al. (2013) showed
that slightly discernible images for the final layers of a convnet could be produced with L2 -regularization. Mahendran
and Vedaldi (2014) also showed the importance of incorporating natural-image priors in the optimization process
when producing images that mimic an entire-layer’s firing
pattern produced by a specific input image. We build on
these works and contribute three additional forms of regularization that, when combined, produce more recognizable, optimization-based samples than previous methods.
Because the optimization is stochastic, by starting at different random initial images, we can produce a set of opti-
Another approach is to try to interpret the function computed by each individual neuron. Past studies in this vein
roughly divide into two different camps: dataset-centric
and network-centric. The former requires both a trained
DNN and running data through that network; the latter requires only the trained network itself. One dataset-centric
approach is to display images from the training or test
set that cause high or low activations for individual units.
Another is the deconvolution method of Zeiler & Fer2
Figure 1. The bottom shows a screenshot from the interactive visualization software. The webcam input is shown, along with the whole
layer of conv5 activations. The selected channel pane shows an enlarged version of the 13x13 conv5151 channel activations. Below it,
the deconv starting at the selected channel is shown. On the right, three selections of nine images are shown: synthetic images produced
using the regularized gradient ascent methods described in Section 3, the top 9 image patches from the training set (the images from the
training set that caused the highest activations for the selected channel), and the deconv of the those top 9 images. All areas highlighted
with a green star relate to the particular selected channel, here conv5151 ; when the selection changes, these panels update. The top
depicts enlarged numerical optimization results for this and other channels. conv52 is a channel that responds most strongly to dog faces
(as evidenced by the top nine images, which are not shown due to space constraints), but it also responds to flowers on the blanket on the
bottom and half way up the right side of the image (as seen in the inset red highlight). This response to flowers can be partially seen in
the optimized images but would be missed in an analysis focusing only on the top nine images and their deconv versions, which contain
no flowers. conv5151 detects different types of faces. The top nine images are all of human faces, but here we see it responds also to the
cat’s face (and in Figure 2 a lion’s face). Finally, conv5111 activates strongly for the cat’s face, the optimized images show catlike fur
and ears, and the top nine images (not shown here) are also all of cats. For this image, the softmax output layer top two predictions are
“Egyptian Cat” and “Computer Keyboard.” All figures in this paper are best viewed digitally, in color, significantly zoomed in.
3
layer. The conv5 layer has size 256×13×13, which we depict as 256 separate 13×13 grayscale images. Each of the
256 small images contains activations in the same spatial
x-y spatial layout as the input data, and the 256 images are
simply and arbitrarily tiled into a 16×16 grid in row-major
order. Figure 2 shows a zoomed in view of one particular channel, conv5151 , that responds to human and animal
faces. All layers can be viewed in the software tool, including pooling and normalization layers. Visualizing these
layers provides intuitions about their effects and functions.
mized images whose variance provides information about
the invariances learned by the unit.
To summarize, this paper makes the following two contributions:
1. We describe and release a software tool that provides
a live, interactive visualization of every neuron in a
trained convnet as it responds to a user-provided image or video. The tool displays forward activation values, preferred stimuli via gradient ascent, top images
for each unit from the training set, deconv highlighting
(Zeiler & Fergus, 2013) of top images, and backward
diffs computed via backprop or deconv starting from
arbitrary units. The combined effect of these complementary visualizations promotes a greater understanding of what a neuron computes than any single method
on its own. We also describe a few insights we have
gained from using this tool. (Section 2).
Although this visualization is simple to implement, we find
it informative because all data flowing through the network
can be visualized. There is nothing mysterious happening
behind the scenes. Because this convnet contains only a
single path from input to output, every layer is a bottleneck
through which all information must pass en-route to a classification decision. The layer sizes are all small enough that
any one layer can easily fit on a computer screen.1 So far,
we have gleaned several surprising intuitions from using
the tool:
2. We extend past efforts to visualize preferred activation
patterns in input space by adding several new types
of regularization, which produce what we believe are
the most interpretable images for large convnets so far
(Section 3).
• One of the most interesting conclusions so far has
been that representations on some layers seem to be
surprisingly local. Instead of finding distributed representations on all layers, we see, for example, detectors
for text, flowers, fruit, and faces on conv4 and conv5.
These conclusions can be drawn either from the live
visualization or the optimized images (or, best, by using both in concert) and suggest several directions for
future research (discussed in Section 4).
Both of our tools are released as open source and are available at http://yosinski.com/deepvis. While the
tools could be adapted to integrate with any DNN software framework, they work out of the box with the popular
Caffe DNN software package (Jia et al., 2014). Users may
run visualizations with their own Caffe DNN or our pretrained DNN, which comes with pre-computed images optimized to activate each neuron in this trained network. Our
pre-trained network is nearly identical to the “AlexNet”
architecture (Krizhevsky et al., 2012), but with local reponse normalization layers after pooling layers following
(Jia et al., 2014). It was trained with the Caffe framework
on the ImageNet 2012 dataset (Deng et al., 2009).
• When using direct file input to classify photos from
Flickr or Google Images, classifications are often correct and highly confident (softmax probability for correct class near 1). On the other hand, when using input from a webcam, predictions often cannot be correct because no items from the training set are shown
in the image. The training set’s 1000 classes, though
numerous, do not cover most common household objects. Thus, when shown a typical webcam view of a
person with no ImageNet classes present, the output
has no single high probability, as is expected. Surprisingly, however, this probability vector is noisy and
varies significantly in response to tiny changes in the
input, often changing merely in response to the noise
from the webcam. We might have instead expected
unchanging and low confidence predictions for a given
scene when no object the network has been trained to
classify is present. Plotting the fully connected layers
(fc6 and fc7) also reveals a similar sensitivity to small
input changes.
2. Visualizing Live Convnet Activations
Our first visualization method is straightforward: plotting
the activation values for the neurons in each layer of a convnet in response to an image or video. In fully connected
neural networks, the order of the units is irrelevant, so plots
of these vectors are not spatially informative. However, in
convolutional networks, filters are applied in a way that respects the underlying geometry of the input; in the case
of 2D images, filters are applied in a 2D convolution over
the two spatial dimensions of the image. This convolution
produces activations on subsequent layers that are, for each
channel, also arranged spatially.
1
The layer with the most activations is conv1 which, when
tiled, is only 550x550 before adding padding.
Figure 1 shows examples of this type of plot for the conv5
4
conv5 layer, one can find many invariant detectors
for faces, shoulders, text, etc. by moving oneself
or objects in front of the camera. Even though the
1000 classes contain no explicitly labeled faces or
text, the network learns to identify these concepts simply because they represent useful partial information
for making a later classification decision. One face
detector, denoted conv5151 (channel number 151 on
conv5), is shown in Figure 2 activating for human
and lion faces and in Figure 1 activating for a cat
face. Zhou et al. (2014) recently observed a similar
effect where convnets trained only to recognize different scene types — playgrounds, restaurant patios,
living rooms, etc. — learn object detectors (e.g. for
chairs, books, and sofas) on intermediate layers.
The reader is encouraged to try this visualization tool out
for him or herself. The code, together with pre-trained
models and images synthesized by gradient ascent, can be
downloaded at http://yosinski.com/deepvis.
3. Visualizing via Regularized Optimization
The second contribution of this work is introducing several
regularization methods to bias images found via optimization toward more visually interpretable examples. While
each of these regularization methods helps on its own, in
combination they are even more effective. We found useful combinations via a random hyperparameter search, as
discussed below.
Formally, consider an image x ∈ RC×H×W , where C = 3
color channels and the height (H) and width (W ) are both
227 pixels. When this image is presented to a neural network, it causes an activation ai (x) for some unit i, where
for simplicity i is an index that runs over all units on all layers. We also define a parameterized regularization function
Rθ (x) that penalizes images in various ways.
Our network was trained on ImageNet by first subtracting the per-pixel mean of examples in ImageNet before inputting training examples to the network. Thus, the direct
input to the network, x, can be thought of as a zero-centered
input. We may pose the optimization problem as finding an
image x∗ where
Figure 2. A view of the 13×13 activations of the 151st channel on
the conv5 layer of a deep neural network trained on ImageNet, a
dataset that does not contain a face class, but does contain many
images with faces. The channel responds to human and animal
faces and is robust to changes in scale, pose, lighting, and context,
which can be discerned by a user by actively changing the scene
in front of a webcam or by loading static images (e.g. of the lions)
and seeing the corresponding response of the unit. Photo of lions
via Flickr user arnolouise, licensed under CC BY-NC-SA 2.0.
x∗ = arg max(ai (x) − Rθ (x))
(1)
x
In practice, we use a slightly different formulation. Because we search for x∗ by starting at some x0 and taking
gradient steps, we instead define the regularization via an
operator rθ (·) that maps x to a slightly more regularized
version of itself. This latter definition is strictly more expressive, allowing regularization operators rθ that are not
• Although the last three layers are sensitive to small
input changes, much of the lower layer computation
is more robust. For example, when visualizing the
5
the gradient of any Rθ . This method is easy to implement
within a gradient descent framework by simply alternating
between taking a step toward the gradient of ai (x) and taking a step in the direction given by rθ . With a gradient
descent step size of η, a single step in this process applies
the update:
∂ai
x ← rθ x + η
∂x
and instead show only the main object, letting other regions
be exactly zero if they are not needed. We implement this
bias using an rθ (x) that computes the norm of each pixel
(over red, green, and blue channels) and then sets any pixels with small norm to zero. The threshold for the norm,
θn pct , is specified as a percentile of all pixel norms in x.
Clipping pixels with small contribution: Instead of clipping pixels with small norms, we can try something slightly
smarter and clip pixels with small contributions to the activation. One way of computing a pixel’s contribution to an
activation is to measure how much the activation increases
or decreases when the pixel is set to zero; that is, to compute the contribution as |ai (x) − ai (x−j )|, where x−j is x
but with the j th pixel set to zero. This approach is straightforward but prohibitively slow, requiring a forward pass
for every pixel. Instead, we approximate this process by
linearizing ai (x) around x, in which case the contribution
of each dimension of x can be estimated as the elementwise product of x and the gradient. We then sum over
all
Pthree channels and take the absolute value, computing
| c x ◦ ∇x ai (x)|. We use the absolute value to find pixels with small contribution in either direction, positive or
negative. While we could choose to keep the pixel transitions where setting the pixel to zero would result in a large
activation increase, these shifts are already handled by gradient ascent, and here we prefer to clip only the pixels that
are deemed not to matter, not to take large gradient steps
outside the region where the linear approximation is most
valid. We define this rθ (x) as the operation that sets pixels
with contribution under the θc pct percentile to zero.
(2)
We investigated the following four regularizations. All are
designed to overcome different pathologies commonly encountered by gradient descent without regularization.
L2 decay: A common regularization, L2 decay penalizes
large values and is implemented as rθ (x) = (1−θdecay )·x.
L2 decay tends to prevent a small number of extreme pixel
values from dominating the example image. Such extreme
single-pixel values neither occur naturally with great frequency nor are useful for visualization. L2 decay was also
used by Simonyan et al. (2013).
Gaussian blur: Producing images via gradient ascent
tends to produce examples with high frequency information (see Supplementary Section S1 for a possible reason).
While these images cause high activations, they are neither
realistic nor interpretable (Nguyen et al., 2014). A useful
regularization is thus to penalize high frequency information. We implement this as a Gaussian blur step rθ (x) =
GaussianBlur(x, θb width ). Convolving with a blur kernel is more computationally expensive than the other regularization methods, so we added another hyperparameter
θb every to allow, for example, blurring every several optimization steps instead of every step. Blurring an image
multiple times with a small width Gaussian kernel is equivalent to blurring once with a larger width kernel, and the
effect will be similar even if the image changes slightly
during the optimization process. This technique thus lowers computational costs without limiting the expressiveness
of the regularization. Mahendran & Vedaldi (2014) used a
penalty with a similar effect to blurring, called total variation, in their work reconstructing images from layer codes.
If the above regularization methods are applied individually, they are somewhat effective at producing more interpretable images; Figure 3 shows the effects of each individual hyperparameter. However, preliminary experiments
uncovered that their combined effect produces better visualizations. To pick a reasonable set of hyperparameters
for all methods at once, we ran a random hyperparameter
search of 300 possible combinations and settled on four that
complement each other well. The four selected combinations are listed in Table 1 and optimized images using each
are shown for the “Gorilla” class output unit in Figure 4.
Of the four, some show high frequency information, others low frequency; some contain dense pixel data, and others contain only sparse outlines of important regions. We
found the version in the lower-left quadrant to be the best
single set of hyperparameters, but often greater intuition
can be gleaned by considering all four at once. Figure 5
shows the optimization results computed for a selection of
units on all layers. A single image for every filter of all
five convolutional layers is shown in Supplementary Figure S1. Nine images for each filter of all layers, including
each of the 1000 ImageNet output classes, can be viewed
at http://yosinski.com/deepvis.
Clipping pixels with small norm: The first two regularizations suppress high amplitude and high frequency information, so after applying both, we are left with an x∗ that contains somewhat small, somewhat smooth values. However,
x∗ will still tend to contain non-zero pixel values everywhere. Even if some pixels in x∗ show the primary object
or type of input causing the unit under consideration to activate, the gradient with respect to all other pixels in x∗ will
still generally be non-zero, so these pixels will also shift to
show some pattern as well, contributing in whatever small
way they can to ultimately raise the chosen unit’s activation. We wish to bias the search away from such behavior
6
Figure 4. Visualizations of the preferred inputs for different class units on layer fc8, the 1000-dimensional output of the network just
before the final softmax. In the lower left are 9 visualizations each (in 3×3 grids) for four different sets of regularization hyperparameters
for the Gorilla class (Table 1). For all other classes, we have selected four interpretable visualizations produced by our regularized
optimization method. We chose the four combinations of regularization hyperparameters by performing a random hyperparameter search
and selecting combinations that complement each other. For example, the lower left quadrant tends to show lower frequency patterns,
the upper right shows high frequency patterns, and the upper left shows a sparse set of important regions. Often greater intuition can
be gleaned by considering all four at once. In nearly every case, we have found that one can guess what class a neuron represents by
viewing sets of these optimized, preferred images. Best viewed electronically,
zoomed in.
7
Figure 5. Visualization of example features of eight layers of a deep, convolutional neural network. The images reflect the true sizes
of the features at different layers. In each layer, we show visualizations from 4 random gradient descent runs for each channel. While
these images are hand picked to showcase the diversity and interpretability of the visualizations, one image for each filter of all five
convolutional layers is shown in Figure S1 in supplementary information. One can recognize important features of objects at different
scales, such as edges, corners, wheels, eyes, shoulders, faces, handles, bottles, etc. The visualizations show the increase in complexity
and variation on higher layers, comprised of simpler components from lower layers. The variation of patterns increases with increasing
layer number, indicating that increasingly invariant representations are learned. In particular, the jump from Layer 5 (the last convolution
layer) to Layer 6 (the first fully-connected layer) brings about a large increase in variation. Best viewed electronically, zoomed in.
8
✓b every = 4 ✓
= 0.5
✓b width = 1.0 decay
✓n pct = 95
✓c pct = 95
✓b every = 10 ✓
= 0.0
✓b width = 0.0 decay
✓n pct = 0
✓c pct = 0
Table 1. Four hyperparameter combinations that produce different
styles of recognizable images. We identified these four after reviewing images produced by 300 randomly selected hyperparameter combinations. From top to bottom, they are the hyperparameter combinations that produced the top-left, top-right, bottomleft, and bottom-right Gorilla class visualizations, respectively, in
Figure 4. The third row hyperparameters produced most of the
visualizations for the other classes in Figure 4, and all of those in
Figure 5.
θdecay
0
0.3
0.0001
0
Figure 3. The effects of each regularization method from Section 3 when used individually. Each of the four rows shows a
linear sweep in hyperparameter space from no regularization (left)
to strong regularization (right). When applied too strongly, some
regularizations cause the optimization to fail (e.g. L2 decay, top
row) or the images to be less interpretable (small norm and small
contribution clipping, bottom two rows). For this reason, a random hyperparameter search was useful for finding joint hyperparameter settings that worked well together (see Figure 4). Best
viewed electronically, zoomed in.
θb width
0.5
0
1.0
0.5
θb every
4
0
4
4
θn pct
50
20
0
0
θc pct
0
0
0
90
ies have shown that discriminative networks can easily be
fooled or hacked by the addition of certain structured noise
in image space (Szegedy et al., 2013; Nguyen et al., 2014).
An oft-cited reason for this property is that discriminative
training leads networks to ignore non-discriminative information in their input, e.g. learning to detect jaguars by
matching the unique spots on their fur while ignoring the
fact that they have four legs. For this reason it has been
seen as a hopeless endeavor to create a generative model
in which one randomly samples an x from a broad distribution on the space of all possible images and then iteratively transforms x into a recognizable image by moving
it to a region that satisfies both a prior p(x) and posterior
p(y|x) for some class label y. Past attempts have largely
supported this view by producing unrealistic images using
this method (Nguyen et al., 2014; Simonyan et al., 2013).
4. Discussion and Conclusion
We have introduced two visual tools for aiding in the interpretation of trained neural nets. Intuition gained from these
tools may prompt ideas for improved methods and future
research. Here we discuss several such ideas.
The interactive tool reveals that representations on later
convolutional layers tend to be somewhat local, where
channels correspond to specific, natural parts (e.g. wheels,
faces) instead of being dimensions in a completely distributed code. That said, not all features correspond to
natural parts, raising the possibility of a different decomposition of the world than humans might expect. These
visualizations suggest that further study into the exact nature of learned representations — whether they are local to
a single channel or distributed across several — is likely to
be interesting (see Zhou et al. (2014) for work in this direction). The locality of the representation also suggests that
during transfer learning, when new models are trained atop
the conv4 or conv5 representations, a bias toward sparse
connectivity could be helpful because it may be necessary
to combine only a few features from these layers to create
important features at higher layers.
However, the results presented here suggest an alternate
possibility: the previously used priors may simply have
been too weak (see Section S1 for one hypothesis of why
a strong p(x) model is needed). With the careful design or
learning of a p(x) model that biases toward realism, one
may be able to harness the large number of parameters
present in a discriminately learned p(y|x) model to generate realistic images by enforcing probability under both
models simultaneously. Even with the simple, hand-coded
p(x) models we use in this paper as regularizers, complex dependencies between distant pixels already arise (cf.
the beetles with structure spanning over 100 pixels in Figure 4). This implies that the discriminative parameters also
contain significant “generative” structure from the training
dataset; that is, the parameters encode not only the jaguar’s
spots, but to some extent also its four legs. With better,
learned probabilistic models over the input and activations
of higher layers, much more structure may be apparent.
Work by Dai et al. (2015) shows some interesting results
in this direction. While the images generated in this paper are far from being photo-realistic, they do suggest that
The second tool — new regularizations that enable improved, interpretable, optimized visualizations of learned
features — will help researchers and practitioners understand, debug, and improve their models. The visualizations
also reveal a new twist in an ongoing story. Previous stud9
transferring discriminatively trained parameters to generative models — opposite the direction of the usual unsupervised pretraining approach — may be a fruitful area for
further investigation.
Jia, Yangqing, Shelhamer, Evan, Donahue, Jeff, Karayev, Sergey,
Long, Jonathan, Girshick, Ross, Guadarrama, Sergio, and Darrell, Trevor. Caffe: Convolutional architecture for fast feature
embedding. arXiv preprint arXiv:1408.5093, 2014.
Acknowledgments
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoff. Imagenet
classification with deep convolutional neural networks. In
Advances in Neural Information Processing Systems 25, pp.
1106–1114, 2012.
The authors would like to thank the NASA Space Technology Research Fellowship (JY) for funding, Wendy Shang,
Yoshua Bengio, Brian Cheung, and Andrej Karpathy for
helpful discussions, and Freckles the cat for her feline
countenance.
Lin, Tsung-Yi, Maire, Michael, Belongie, Serge, Hays, James,
Perona, Pietro, Ramanan, Deva, Dollár, Piotr, and Zitnick,
C. Lawrence. Microsoft COCO: common objects in context.
CoRR, abs/1405.0312, 2014. URL http://arxiv.org/
abs/1405.0312.
Mahendran, A. and Vedaldi, A. Understanding Deep Image Representations by Inverting Them. ArXiv e-prints, November
2014.
References
Bergstra, James, Breuleux, Olivier, Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guillaume, Turian,
Joseph, Warde-Farley, David, and Bengio, Yoshua. Theano: a
CPU and GPU math expression compiler. In Proceedings of
the Python for Scientific Computing Conference (SciPy), June
2010. Oral Presentation.
Nguyen, Anh, Yosinski, Jason, and Clune, Jeff. Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images. ArXiv e-prints, December 2014.
Schroff, F., Kalenichenko, D., and Philbin, J. FaceNet: A Unified Embedding for Face Recognition and Clustering. ArXiv
e-prints, March 2015.
Collobert, Ronan, Kavukcuoglu, Koray, and Farabet, Clément.
Torch7: A matlab-like environment for machine learning.
In BigLearn, NIPS Workshop, number EPFL-CONF-192376,
2011.
Simonyan, Karen, Vedaldi, Andrea, and Zisserman, Andrew.
Deep inside convolutional networks: Visualising image
classification models and saliency maps. arXiv preprint
arXiv:1312.6034, presented at ICLR Workshop 2014, 2013.
Dai, Jifeng, Lu, Yang, and Wu, Ying Nian. Generative modeling
of convolutional neural networks. In International Conference
on Learning Representations (ICLR), 2015.
Szegedy, Christian, Zaremba, Wojciech, Sutskever, Ilya, Bruna,
Joan, Erhan, Dumitru, Goodfellow, Ian J., and Fergus, Rob. Intriguing properties of neural networks. CoRR, abs/1312.6199,
2013.
Deng, Jia, Dong, Wei, Socher, Richard, Li, Li-Jia, Li, Kai,
and Fei-Fei, Li. Imagenet: A large-scale hierarchical image
database. In Computer Vision and Pattern Recognition, 2009.
CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.
Taigman, Yaniv, Yang, Ming, Ranzato, Marc’Aurelio, and Wolf,
Lior. Deepface: Closing the gap to human-level performance
in face verification. In Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on, pp. 1701–1708. IEEE,
2014.
Erhan, Dumitru, Bengio, Yoshua, Courville, Aaron, and Vincent, Pascal. Visualizing higher-layer features of a deep network. Technical report, Technical report, University of Montreal, 2009.
Torralba, Antonio and Oliva, Aude. Statistics of natural image
categories. Network: computation in neural systems, 14(3):
391–412, 2003.
Glorot, Xavier, Bordes, Antoine, and Bengio, Yoshua. Deep
sparse rectifier networks. In Proceedings of the 14th International Conference on Artificial Intelligence and Statistics.
JMLR W&CP Volume, volume 15, pp. 315–323, 2011.
Goodfellow, Ian J, Warde-Farley, David, Lamblin, Pascal, Dumoulin, Vincent, Mirza, Mehdi, Pascanu, Razvan, Bergstra,
James, Bastien, Frédéric, and Bengio, Yoshua. Pylearn2:
a machine learning research library.
arXiv preprint
arXiv:1308.4214, 2013.
Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. How transferable are features in deep neural networks? In Ghahramani,
Z., Welling, M., Cortes, C., Lawrence, N.D., and Weinberger,
K.Q. (eds.), Advances in Neural Information Processing Systems 27, pp. 3320–3328. Curran Associates, Inc., December
2014.
Goodfellow, Ian J, Shlens, Jonathon, and Szegedy, Christian. Explaining and Harnessing Adversarial Examples. ArXiv e-prints,
December 2014.
Zeiler, Matthew D and Fergus, Rob. Visualizing and understanding convolutional neural networks. arXiv preprint
arXiv:1311.2901, 2013.
Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G.,
Elsen, E., Prenger, R., Satheesh, S., Sengupta, S., Coates, A.,
and Ng, A. Y. Deep Speech: Scaling up end-to-end speech
recognition. ArXiv e-prints, December 2014.
Zhou, Bolei, Khosla, Aditya, Lapedriza, Àgata, Oliva, Aude, and
Torralba, Antonio. Object detectors emerge in deep scene cnns.
CoRR, abs/1412.6856, 2014. URL http://arxiv.org/
abs/1412.6856.
Hinton, Geoffrey E, Srivastava, Nitish, Krizhevsky, Alex,
Sutskever, Ilya, and Salakhutdinov, Ruslan R. Improving neural networks by preventing co-adaptation of feature detectors.
arXiv preprint arXiv:1207.0580, 2012.
10
Supplementary Information for:
Understanding Neural Networks Through Deep Visualization
Jason Yosinski
Jeff Clune
Anh Nguyen
Thomas Fuchs
Hod Lipson
S1. Why are gradient optimized images
dominated by high frequencies?
In the main text we mentioned that images produced by
gradient ascent to maximize the activations of neurons in
convolutional networks tend to be dominated by high frequency information (cf. the left column of Figure 3). One
hypothesis for why this occurs centers around the differing
statistics of the activations of channels in a convnet. The
conv1 layer consists of blobs of color and oriented Gabor
edge filters of varying frequencies. The average activation
values (after the rectifier) of the edge filters vary across
filters, with low frequency filters generally having much
higher average activation values than high frequency filters.
In one experiment we observed that the average activation
values of the five lowest frequency edge filters was 90 versus an average for the five highest frequency filters of 5.4, a
difference of a factor of 17 (manuscript in preparation)2,3 .
The activation values for blobs of color generally fall in the
middle of the range. This phenomenon likely arises for reasons related to the 1/f power spectrum of natural images in
which low spatial frequencies tend to contain higher energy
than high spatial frequencies (Torralba & Oliva, 2003).
Now consider the connections from the conv1 filters to a
single unit on conv2. In order to merge information from
both low frequency and high frequency conv1 filters, the
connection weights from high frequency conv1 units may
generally have to be larger than connections from low frequency conv1 units in order to allow both signals to affect
the conv2 unit’s activation similarly. If this is the case, then
due to the larger multipliers, the activation of this particular
conv2 unit is affected more by small changes in the activations of high frequency filters than low frequency filters.
2
Li, Yosinski, Clune, Song, Hopcroft, Lipson. 2015. How
similar are features learned by different deep neural networks? In
preparation.
3
Activation values are averaged over the ImageNet validation
set, over all spatial positions, over the channels with the five
{highest, lowest} frequencies, and over four separately trained
networks.
YOSINSKI @ CS . CORNELL . EDU
JEFFCLUNE @ UWYO . EDU
ANGUYEN 8@ UWYO . EDU
FUCHS @ CALTECH . EDU
HOD . LIPSON @ CORNELL . EDU
Seen in the other direction: when gradient information is
passed from higher layers to lower layers during backprop,
the partial derivative arriving at this conv2 unit (a scalar)
will be passed backward and multiplied by larger values
when destined for high frequency conv1 filters than low frequency filters. Thus, following the gradient in pixel space
may tend to produce an overabundance of high frequency
changes instead of low frequency changes.
The above discussion focuses on the differing statistics of
edge filters in conv1, but note that activation statistics on
subsequent layers also vary across each layer.4 This may
produce a similar (though more subtle to observe) effect in
which rare higher layer features are also overrepresented
compared to more common higher layer features.
Of course, this hypothesis is only one tentative explanation
for why high frequency information dominates the gradient. It relies on the assumption that the average activation
of a unit is a representative statistic of the whole distribution of activations for that unit. In our observation this
has been the case, with most units having similar, albeit
scaled, distributions. However, more study is needed before a definitive conclusion can be reached.
S2. Conv Layer Montages
One example optimized image using the hyperparameter
settings from the third row of Table 1 for every filter of all
five convolutional layers is shown in Figure S1.
4
We have observed that statistics vary on higher layers, but in
a different manner: most channels on these layers have similar average activations, with most of the variance across channels being
dominated by a small number of channels with unusually small
or unusually large averages (Li, Yosinski, Clune, Song, Hopcroft,
Lipson. 2015. How similar are features learned by different deep
neural networks? In preparation.)
conv5
conv1
conv2
conv3
conv4
Figure S1. One optimized, preferred image for every channel of all five convolutional layers. These images were produced with the
hyperparameter combinations from the third row of Table 1. Best viewed electronically, zoomed in.
12
| 9cs.NE
|
The Complexity of Distributed Edge Coloring
with Small Palettes∗
Yi-Jun Chang
University of Michigan
Qizheng He
IIIS, Tsinghua University
arXiv:1708.04290v1 [cs.DC] 14 Aug 2017
Seth Pettie
University of Michigan
Wenzheng Li
IIIS, Tsinghua University
Jara Uitto
ETH Zürich
Abstract
The complexity of distributed edge coloring depends heavily on the palette size as a function
of the maximum degree ∆. In this paper we explore the complexity of edge coloring in the
LOCAL model in different palette size regimes. Our results are as follows.
• We simplify the round elimination technique of Brandt et al. [9] and prove that (2∆ − 2)edge coloring requires Ω(log∆ log n) time w.h.p. and Ω(log∆ n) time deterministically, even
on trees. The simplified technique is based on two ideas: the notion of an irregular running
time (in which network components terminate the algorithm at prescribed, but irregular
times) and some general observations that transform weak lower bounds into stronger ones.
• We√give a randomized edge coloring algorithm that can use palette sizes as small as ∆ +
Õ( ∆), which is a natural barrier for randomized approaches. The running time of the
algorithm is at most O(log ∆ · TLLL ), where TLLL is the complexity of a permissive version
of the constructive Lovász local lemma.
• We develop a new distributed Lovász local lemma algorithm for tree-structured dependency
graphs, which leads to a (1 + )∆-edge coloring algorithm for trees running in O(log log n)
time. This algorithm arises from two new results: a deterministic O(log n)-time LLL algorithm for tree-structured instances, and a randomized O(log log n)-time graph shattering
method for breaking the dependency graph into independent O(log n)-size LLL instances.
• A natural approach to computing (∆ + 1)-edge colorings (Vizing’s theorem) is to extend
partial colorings by iteratively re-coloring parts of the graph, e.g., via “augmenting paths.”
We prove that this approach may be viable, but in the worst case requires recoloring
subgraphs of diameter Ω(∆ log n). This stands in contrast to distributed algorithms for
Brooks’ theorem [34], which exploit the existence of O(log∆ n)-length augmenting paths.
∗
Supported by NSF grants CCF-1514383 and CCF-1637546 and ERC Grant No. 336495 (ACDC).
1
Introduction
We study edge coloring 1 problems in the well known LOCAL2 model of distributed computation [29,
35], which for clarity we bifurcate into RandLOCAL and DetLOCAL depending on whether random
bits are available. The distributed complexity of computing a k-edge coloring depends heavily on
the value of k (relative to the maximum degree ∆) and whether vertices can generate random bits.
In Section 1.1 we review previous edge coloring algorithms in descending order by palette size (see
Table 1) and in Section 1.4 we summarize our contributions.
1.1
Edge Coloring Algorithms
Edge-coloring can be interpreted as a vertex coloring problem on the line graph L(G), in which
edges becomes vertices and two edges are adjacent if they share an endpoint; the line graph has
ˆ = 2∆ − 2. Applied to L(G), Linial’s [29] vertex coloring algorithm will compute
maximum degree ∆
2
ˆ
ˆ
ˆ
an O(∆ )-edge coloring in O(log∗ n−log∗ ∆+1)
time. Using the fastest deterministic
(∆+1)-vertex
√
√
O(
log
n)
coloring algorithms [34, 21], (2∆ − 1)-edge coloring is solved in min{2
, Õ( ∆) + log∗ n}
time. Barenboim, Elkin, and Maimon gave deterministic algorithms for (2k ∆)-edge coloring (k ≥ 2)
in Õ(k∆1/2k ) + log∗ n, by reduction to [21].
Barenboim, Elkin, Pettie, and Schneider [7] proved that O(log ∆) iterations of the natural randomized (2∆ − 1)-edge coloring algorithm effectively shatters the graph into uncolored components
of poly(log n) vertices, which can then be colored using a deterministic algorithm.3 Elkin, Pettie,
and Su [18] proved that when ∆ > (log n)1+γ , (2∆ − 1)-edge coloring can be solved in O(log∗ n)
time in RandLOCAL. Very recently Fischer, Ghaffari, and Kuhn proved that (2∆ − 1)-edge coloring can be solved in O(log7 ∆ log n) time in DetLOCAL. Together with [7] and [18], this implies
a RandLOCAL algorithm running in min{O((log log n)8 ), O(log7 ∆ log log n)}) time. By using a
slightly larger palette of (2 + )∆ colors, > 1/ log ∆, Ghaffari et al. [24] (cf. [25]) gave a faster
DetLOCAL edge coloring algorithm running in time O(−1 log2 ∆ log log ∆(log log log ∆)1.71 log n).
The “2∆ − 1” arises because it is the smallest palette size with the property that any partial
coloring can be extended to a total coloring, by the trivial greedy algorithm. Below the greedy
threshold 2∆ − 1, iterative coloring algorithms must be more careful in how they proceed. In
particular, at intermediate stages in the algorithm, edges must keep their available palettes relatively
large compared to the size of their uncolored neighborhood.
Using the Rödl nibble technique, Dubhashi, Grable, and Panconesi [16] gave a RandLOCAL
algorithm for (1 + )∆-edge coloring in O(log n) time, provided that ∆ is sufficiently large, e.g.,
even when is constant, ∆ > (log n)1+γ . Elkin, Pettie, and Su [18] gave RandLOCAL algorithms
for (1 + )∆-edge coloring that are faster when ∆ is large and work for all ∆ via a reduction to
the distributed Lovász local lemma (LLL); see Section 1.3 for a discussion of the distributed LLL.
1
A k-edge coloring is a function φ : E → {1, . . . , k} such that edges sharing an endpoint are colored differently;
“k” is called the palette size.
2
The LOCAL model has the following features. The input graph G = (V, E) is identical to the distributed network;
vertices are identified with processors and edges with bi-directional communication links; each v ∈ V initially knows
deg(v), a port-numbering of its incident edges, and global parameters such as n = |V | and ∆ = maxv deg(v);
time is divided into synchronized rounds, and in each round each processor can perform unlimited computation
and communicate an unbounded-length message to each of its neighbors, which is delivered before the next round.
Depending on the problem the vertices may carry additional input labels. The output of a LOCAL algorithm is
typically a labeling of V or E that satisfies some constraints. In the RandLOCAL model the output labeling is correct
w.h.p. (1 − 1/poly(n)). In the DetLOCAL model each vertex is assigned a unique O(log n)-bit ID; the output labeling
must always be correct.
3
The problem of coloring one of these components is a (2 deg(v) − 1)-list edge coloring problem, i.e., v’s palette
includes an arbitrary set of 2 deg(v) − 1 colors.
1
The (1 + )∆-edge coloring problem is solved in O(log∗ n ·
l
log n
∆1−o(1)
m
) time via the Chung-Pettie-
[12].4
Su LLL algorithm
The running time of the Dubhashi-Grable-Panconesi and Elkin-Pettie-Su
algorithms depend polynomially on −1 . In both algorithms it is clear that need not be constant,
but it is not self-evident how small it can be
√ made as a function of ∆. The natural limit for
randomized coloring strategies is a (∆ + O( ∆))-size palette,5 which was achieved in 1987 by
Karloff and Shmoys [27] in the context of parallel (PRAM) algorithms, but has never been achieved
in the LOCAL model.
We cannot hope to use fewer than ∆ + 1 colors on general graphs. Vizing [37] proved that ∆ + 1
suffices for any graph, and Holyer [26] proved that it is NP-hard to tell if a graph is ∆-colorable.
The
√
best sequential (∆ + 1)-edge coloring algorithms [1, 22] run in O(min{∆m log n, m n log n})√time
and are not suited for implementation in the LOCAL model. When the palette size is ∆ + o( ∆),
a natural way to solve the problem [1, 22] is to begin with any maximal partial coloring, and
then iteratively recolor portions of the graph (e.g., along “augmenting paths”) so that at least one
uncolored edge can be legally colored. This approach was successfully employed by Panconesi and
Srinivasan [33] in their distributed algorithm for Brooks’ theorem.6 They proved that for any partial
coloring, there exists an augmenting path with length O(log∆ n), and that given a (∆ + 1)-vertex
coloring, a ∆-vertex coloring could be computed in O(log2 n log∆ n) additional time.
1.2
Lower Bounds
Linial’s Ω(log∗ n) lower bound for O(1)-coloring the ring [29, 32] shows that f (∆)-edge coloring also
cannot be computed in o(log∗ n) time, for any function f .
None of the other published lower bounds apply
q directly to the edge coloring problem. Kuhn,
Moscibroda, and Wattenhofer’s Ω(min{ logloglog∆∆ , logloglogn n }) lower bounds apply to MIS and maximal
matching, but not to any vertex/edge coloring problem. Linial’s O(log∆ n) lower bound [29] (see [36,
p. 265]) on o(∆/ ln ∆)-vertex coloring trees does not imply anything for edge-coloring trees. The
lower bounds of Brandt et al. [9] (in RandLOCAL) and Chang, Kopelowitz, and Pettie [10] (in
DetLOCAL) for sinkless orientation and ∆-vertex coloring trees do not naturally generalize to edge
coloring. Indeed, Brandt et al.’s lower bound technique oscillates between sinkless orientation and
a closely related problem called sinkless coloring, whose input already consists of a ∆-edge colored
graph.
1.3
The Distributed Lovász Local Lemma
Randomized coloring algorithms are often composed of O(1)-round routines that commit to a partial
coloring, whose local probability of failure is small, as a function of ∆. Using a distributed Lovász
local lemma (LLL) algorithm, we can guarantee global success with probability 1−1/poly(n) or even
1. Table 2 summarizes distributed LLL algorithms under different symmetric criteria p · f (d) < 1,
where p is the local probability of failure and d is the maximum degree in the dependency graph.7
Chang and Pettie [11] conjectured that the RandLOCAL complexity of the LLL under some
polynomial criterion (e.g., p(ed)c < 1 for some fixed c) is O(log log n), matching the Brandt et
al. [9] lower bound. If this conjecture were true, the results of [10, Theorem 3] indicate what the
4
If ∆ is sufficiently small (∆ log n for some specific > 0), this algorithm can be sped up using a recent LLL
algorithm of Fischer and Ghaffari [19].
5
This is the threshold at which we have a constant probability of being able to color e, given a random feasible
coloring of its neighborhood.
6
Which states that any graph with ∆ ≥ 3 having no (∆ + 1)-cliques is ∆-vertex colorable.
7
In coloring algorithms it is typical to see d = poly(∆) and p = exp(−Ω(dΩ(1) )).
2
Palette Size
f (∆)
Time
Lower bound: Ω(log∗ n)
O(∆2 )
log∗ n − log∗ ∆ + 1
∆1+
C log ∆ log n
∆ log n
log4 n
C ∆
∆ log n
2k ∆
(2 + )∆
(Rand)
R
?
k∆1/2k + log n
?
?
log ∆ +
log∗ n
?
√
2O( log log n)
R?
min{(log log n)8 ,
7
log ∆ log log n}
Lower bound: Ω(log∆ log n)
[4]
k≥2
[4]
[6, 25]
> 1/ log ∆
[24]
Vertex coloring L(G)
[34]
Vertex coloring L(G)
[21]
Vertex coloring L(G)
[7]
(log n)1+o(1)
[18]
∆>
R
[18]
?
[20]
R?
[7]+[18]+[20]
R
new
Lower bound: Ω(log∆ n)
−1 log −1
new
+ log n
l
∆+1
Vertex coloring L(G)
[25]
−1 log ∆2+o(1) log n
log7 ∆ log n
√
∆ + Õ( ∆)
[29]
−3 log11 n
√
2O( log log n)
(1 + )∆
Vertex coloring L(G)
Vertex coloring L(G)
∗
√
Õ( ∆) + log∗ n
2∆ − 2
References
[29, 32]
[14]
√
2O( log n)
2∆ − 1
Notes
∆ = O(1)
m
n
(−2 log −1 + log∗ ∆) 2 ∆log
1−o(1)
l
m
n
log −1 2 ∆log
+ log∗ n
1−o(1)
m
l
n
+ (log log n)3+o(1)
log −1 2 ∆log
1−o(1)
m
l
n
log ∆ 2 ∆log
+ (log log n)3+o(1)
1−o(1)
diameter(G)
(log n)1+γ()
R
∆>
R
∆ > ∆
R
∆ > (log n)1+o(1)
new
∆ > ∆
new
R?
[16]
[18]
R?
new
?
[37]
Table 1: A history of notable edge coloring algorithms and lower bounds, in descending
order by palette size. Some (2∆ − 1)-edge coloring algorithms that follow from vertex
coloring L(G), such as [2, 28, 5, 3], have been omitted for brevity. RandLOCAL algorithms are marked with R; all others work in DetLOCAL. Those algorithms that are the
“best” in any sense are marked with a ?.
3
Symmetric
LLL Criterion
ep(d + 1) < 1
Rand/Det
Rand
Time
O(MIS · log1/ep(d+1) n)
O(WeakMIS · log1/ep(d+1) n)
Rand
also asymmetric crition
O(log d · log1/ep(d+1) n)
Rand
also asymmetric crit.
also asymmetric crition
epd2
<1
O(log1/epd2 n)
Rand
poly(d)2d
<1
O(log n/ log log n)
Rand
p(ed)λ < 1
O(n1/λ ·
√
2O( log n) )
p(ed)4λ < 1
O(d2 + (log n)1/λ · 2O(
p(ed)32
<1
√
2O( log log n)
d2 +1
<1
O(d2 + log∗ n)
p(ed)
Notes
Reference
also asymmetric crition
[31]
√
log log n) )
[23]+[12]
Any λ ≥ 1
Rand
Any λ ≥ 8
[19]
[19]
Requires d <
(log log n)1/5
Det
p · f (d) ≤ 1
p(ed)2 < 1
p(ed)λ < 1
p(ed)λ < 1
[19]
[19]
Lower Bounds (apply to tree-structured instances)
p · f (d) < 1 Ω(log∗ n)
Rand Any f
p · f (d) ≤ 1
[12]
[12]
Det
Rand
[12]
[12]
Rand
Any f (d) ≤ 2d
[9]
Det
2d
[10]
LLL for Tree-Structured Instances
O(log n)
Det
log n
O(max{logλ n, log log n })
Det λ ≥ 2
log log n
Rand λ ≥ 2(4r + 8r)
O(max{logλ log n, log log log n })
new
new
new
Ω(loglog(1/p) log n)
Ω(loglog(1/p) n)
Any f (d) ≤
∗
Table
√ 2: A survey of distributed LLL algorithms. MIS = O(min{d + log n, log d +
O(
log
log
n)
}) [5, 23] is the complexity of computing a maximal independent set in a
2
graph with maximum degree d. WeakMIS = O(log d) [23] is the task of finding an
independent set I such that the probability that v is not in/adjacent to I is 1/poly(d).
All lower bounds apply even to tree-structured instances. The lower bounds of [9,
10] (Ω(logd log n) randomized, Ω(logd n) deterministically) were for an LLL instance
satisfying p2d ≤ 1. By a change of parameters, they also imply stronger lower bounds
(substituting log(1/p) for d) under any LLL criterion p · f (d) ≤ p2d ≤ 1.
4
optimal algorithm should look like: it must combine an O(log n)-time DetLOCAL LLL algorithm
and an O(log log n)-time graph shattering routine to break the dependency graph into poly(log n)size LLL instances. Fischer and Ghaffari [19] exhibited a deterministic n1/λ+o(1) -time algorithm
for LLL criterion p(ed)λ < 1, and an O(d2 + log∗ n) routine to shatter the dependency graph into
log n-size components.
1.4
New Results
We present new upper and lower bounds on the complexity of edge coloring in the regimes between
palette size ∆ and 2∆ − 2, i.e., strictly below the “greedy” threshold 2∆ − 1.
Round Elimination. Our first result is a lower bound on (2∆−2)-edge coloring using a simplified
version of Brandt et al.’s [9] round elimination technique. Roughly speaking, their idea is to convert
any randomized t-round algorithm with local error probability p into a (t − 1)-round algorithm
with error probability ≈ p1/∆ . By iterating the procedure they obtain a 0-round algorithm with
−t
error probability ≈ p∆ . If any 0-round algorithm must have constant probability of failure, then
t = Ω(log∆ log p−1 ). By setting p = 1/poly(n) we get Ω(log∆ log n) RandLOCAL lower bounds for
some problems, e.g., sinkless orientation. We present a much simplified round elimination technique
that appears to give quantitatively worse bounds, but which can be automatically strengthened to
match those of [9]. Rather than try to shave one round off the running time of every processor,
it is significantly simpler to do it piecemeal, which leads us to the useful concept of an irregular
time profile. Suppose that the graph is initially k-edge colored, k being at least 2∆ − 1 so as not to
trivialize the problem. An algorithm has irregular time profile t = (t1 , . . . , tk ) if edges with input
color i choose their output color by examining only their ti -neighborhood. In our simplified roundelimination technique, we show that any algorithm with time profile (t, t, · · · , t, t − 1, · · · , t − 1) and
| {z } |
{z
}
i
k−i
error probability p can be transformed into one with time profile (t, t, · · · , t, t − 1, · · · , t − 1) and
{z
}
| {z } |
i−1
k−i+1
error probability O(p1/3 ), only by changing the algorithm for edges initially colored i. By iterating
this process we arrive at Ω(∆−1 log log p−1 ) lower bounds, which has a weaker dependence on ∆
than [9]. By following the proofs of Chang, Kopelowitz, and Pettie [10], any randomized lower bound
of this type implies Ω(log∆ n) lower bounds in DetLOCAL [10, Theorem 5], and hence Ω(log∆ log n)
lower bounds in RandLOCAL [10, Theorem 3].
Faster (1 + )∆-edge Coloring. The (1 + )∆-edge coloring algorithms of [16, 18] are slow (with
a polynomial dependence on −1 ) and have limits on how small can be, as a function of ∆. We
prove that the most “natural” randomized algorithm converges
exponentially faster with −1 and
√
can achieve palette sizes close
√ to the minimum of ∆ + Õ( ∆) allowed by the nibble method. In
particular, for any = Ω̃(1/ ∆), (1 + )∆-edge coloring is reducible to O(log −1 ) instances of the
Lovász local lemma with local failure probability exp(−(2 ∆)1−o(1) ), plus one instance of O(∆)-edge
coloring, which can be solved quickly using [7, 18, 24]. When 2 ∆ log n the error is 1/poly(n);
otherwise we can invoke a distributed LLL algorithm [31, 12, 19].
Upper Bounds on Trees. Our lower bounds on (2∆ − 2)-edge coloring apply even to trees. In
order to adapt our randomized (1 + )∆-edge coloring algorithms to trees, we need a special LLL
algorithm for tree structured dependency graphs. Using the framework of Fischer and Ghaffari [19],
we give a deterministic O(max{logλ n, log n/ log log n})-time LLL algorithm for such instances under
5
criterion p(ed)λ < 1, λ ≥ 2. The algorithm is based on a special network decomposition algorithm for
tree-structured graphs, in which one color class has diameter O(logλ n) while the other color classes
have diameter 0. We also present a new graph shattering routine for tree-structured LLL instances
that runs in time O(logλ log n), improving the O(d2 +log∗ n)-time shattering routine of [19] when d is
not too small. (The new graph shattering method can be viewed as an algorithm that computes the
final state of a certain contagion dynamic exponentially faster than simulating the actual contagion.)
By composing these results we obtain a randomized O(max{logλ log n, log log n/ log log log n}) LLL
algorithm for trees, which essentially matches the lower bound of [9] and the conjectured upper
bound for general instances [11, Conjecture 1]. See Table 2.
A Distributed Vizing’s Theorem? Suppose that a distributed (∆ + 1)-edge coloring algorithm
begins with a partial coloring and iteratively recolors subgraphs, always increasing the subset of
colored edges. If this algorithm works correctly given any partial coloring, we prove that it takes
Ω(∆ log n) time in any LOCAL model, and more generally, (∆ + c)-coloring takes Ω( ∆c log n) time.
This establishes a quantitative difference between the “locality” of Vizing’s theorem and Brooks’
theorem [33].
Organization of the paper. In Section 2 we give lower bounds on (2∆ − 2)-edge coloring. In
Section 3 we give a randomized (1 + )∆-edge coloring algorithm, whose detailed analysis appears
in Appendices D and E. In Section 4 we give new LLL algorithms for tree-structured dependency
graphs. We conclude in Section 5.
Appendix A contains some results on edge-coloring trees deterministically, which may be folklore.
Appendix B has some network decomposition algorithms for trees. Appendix C gives lower bounds
on a class of “recoloring” algorithms for Vizing’s theorem.
2
Lower Bound for (2∆ − 2)-Edge Coloring
The sinkless orientation problem is to orient (direct) the edges such that no vertex has out-degree
zero. Since this problem becomes harder with fewer edges, we let ∆ denote the minimum degree in
this problem, whereas in the edge-coloring problem ∆ is still the maximum degree. We first observe
that sinkless orientation on 2-colored bipartite graphs is reducible to (2∆ − 2)-edge coloring.
Theorem 1. Suppose Ae.c. is a t-round (2∆−2)-edge coloring algorithm with local failure probability
p. There is a (t + 1)-round sinkless orientation algorithm As.o. for 2-colored bipartite graphs with
minimum degree ∆ whose local failure probability is p.
Proof. Ae.c. produces a proper partial (2∆ − 2)-edge coloring φ : E → {1, . . . , 2∆ − 2, ⊥} such that
for all v ∈ V , Pr(∃(u, v) : φ(u, v) = ⊥) ≤ p, i.e., a vertex errs if not all of its edges are colored.
Suppose we are given a bipartite graph G = (V, E) with a 2-coloring V → {0, 1} and minimum
degree ∆. In the first round of As.o. , each vertex selects ∆ of its incident edges arbitrarily and
notifies the other endpoint whether it was selected. Let G0 = (V, E 0 ) be the subgraph of edges
selected by both endpoints. The algorithm As.o. runs Ae.c. on G0 for t rounds to get a partial
coloring φ : E 0 → {1, . . . , 2∆ − 2, ⊥}, then orients the edges (in the direction 0 → 1 or 0 ← 1) as
6
follows. Let {u0 , u1 } ∈ E be an edge with uj colored j ∈ {0, 1}.
if {u0 , u1 } ∈ E 0 and φ(u0 , u1 ) ∈ {1, 2, . . . , ∆ − 1, ⊥},
0→1
or if only u0 selected {u0 , u1 }
As.o. ({u0 , u1 }) =
if {u0 , u1 } ∈ E 0 and φ(u0 , u1 ) ∈ {∆, . . . , 2∆ − 2},
0←1
or if only u1 selected {u0 , u1 }
The only way a vertex v can be a sink is if (i) v has degree exactly ∆ in G0 , (ii) v is colored 1,
and (iii) each edge e incident to v has φ(e) ∈ {1, 2, . . . , ∆ − 1, ⊥}. Criterion (iii) only occurs with
probability at most p.
Thus, any lower bound for sinkless orientation on 2-colored graphs also applies to (2∆ − 2)-edge
coloring. Define T∆ to be an infinite ∆-regular tree whose vertices are properly 2-colored by {0, 1}
and whose edges are assigned a proper (2∆ − 1)-coloring uniformly at random.8
For simplicity we suppose that the edges host processors (not vertices), and that two edges can
communicate if they are adjacent in the line graph L(T∆ ). Define N t (e) to be all edges within
distance t of e in the line graph; we also use N t (e) to refer to all information (edge coloring, vertex
coloring, random bits, etc.) stored in their processors. By definition, a time-t algorithm has time
profile (t, t, t, · · · , t).
Lemma 1. (Round Elimination Lemma) Suppose As.o. is a sinkless orientation algorithm for
T∆ with error probability p and time profile (t, t, · · · , t, t − 1, · · · , t − 1), i.e., edges colored {1, . . . , i}
| {z }
i
halt after t rounds and the others after t − 1 rounds. There exists a sinkless orientation algorithm
A0s.o. with error probability 3p1/3 and time profile (t, t, · · · , t, t − 1, · · · , t − 1).
| {z }
i−1
Proof. Only edges colored i modify their algorithm; all others behave identically under A0s.o. and
As.o. . Let e0 = {u0 , u1 } be an edge colored i with uj colored j ∈ {0, 1} and let the remaining edges
incident to u0 and u1 be {e1 , . . . , e∆−1 } and {e∆ , . . . , e2∆−2 }, respectively. Consider the following
two events regarding the output of As.o. .
E0 : ∀j ∈ [1, ∆ − 1], As.o. (ej ) = 0 ← 1
I.e., u0 has outdegree 0 in G − {e0 }
E1 : ∀j ∈ [∆, 2∆ − 2], As.o. (ej ) = 0 → 1
I.e., u1 has outdegree 0 in G − {e0 }
If both events hold, then either u0 or u1 must be a sink, so
Pr(E0 ∩ E1 ) ≤ 2p
(1)
On edge e0 , A0s.o. runs for t − 1 rounds and determines whether the following events occur.
h
i
h
i
E0? : Pr(E0 | N t−1 (e0 )) ≥ p1/3 ,
E1? : Pr(E1 | N t−1 (e0 )) ≥ p1/3
8
One could generate such a coloring as follows: pick an edge and assign it a random color,
then iteratively pick
any vertex u with one incident edge colored, choose ∆ − 1 colors at random from the 2∆−2
possibilities, then assign
∆−1
them to u’s remaining uncolored edges uniformly at random. Randomized algorithms that run on T∆ know the edge
coloring and how it was generated. Thus, the probability of failure depends on the random bits generated by the
algorithm, and those used to generate the edge-coloring.
7
Note that if we inspect N t−1 (e0 ) (and condition on the information seen), the events E0 and E1
become independent since they now depend on disjoint sets of random variables.9 Thus,
Pr(E0 ∩ E1 | N t−1 (e0 )) = Pr(E0 | N t−1 (e0 )) · Pr(E1 | N t−1 (e0 ))
(2)
Since E0? , E1? are determined by N t−1 (e0 ), (2) implies that Pr(E0 ∩ E1 | E0? ∩ E1? ) ≥ p2/3 , and with (1)
we deduce that
Pr(E0? ∩ E1? ) ≤ 2p1/3
(3)
The algorithm A0s.o. orients e0 as follows.
A0s.o. (e0 )
=
0→1
0←1
if E0? holds
otherwise
The failure probability at a vertex not adjacent to any edge colored i is the same under As.o. and
A0s.o. . We calculate the failure probabilities of the remaining vertices now.
Pr(u0 is a sink) = Pr(E0? ∩ E0 )
≤ Pr(E0 | E0? ) ≤ p1/3 ,
by definition of E0?
Pr(u1 is a sink) = Pr(E0? ∩ E1 )
≤ Pr(E0? ∩ E1? ) + Pr(E1 ∩ E1? )
≤ 2p1/3 + p1/3 = 3p1/3 ,
by (3) and the definition of E1? .
Lemma 2. Any sinkless orientation algorithm for T∆ with local error probability p has time complexity Ω(∆−1 log log p−1 ).
Proof. Let As.o. be a t-round algorithm with error probability p, i.e., it has time profile (t, t, . . . , t).
Applying Lemma 1 t(2∆ − 1) times we get an algorithm A0s.o. with time profile (0, 0, . . . , 0) and
−t(2∆−1)
error probability p0 = O(p3
). We now claim that p0 must also be at least 8−∆ . Any 0round orientation algorithm can be characterized by a real vector (q1 , . . . , q2∆−1 ), where qi is the
probability that an edge colored i is oriented as 0 → 1. Without loss of generality, suppose that
q1 , . . . , q∆ ≥ 1/2. Fix any v ∈ V (T∆ ) labeled 1. The probability that v is a sink is at least the
probability that its edges are initially colored {1, . . . , ∆} and that they are all oriented away from
−1 −∆
v, hence p0 ≥ 2∆−1
·2
≥ 2−3∆ . Combining the upper and lower bounds on p0 we have
∆
−1 3
23∆ ≥ p−1
0 = Ω((p )
−t(2∆−1)
)
and taking logs twice we have
log(3∆) ≥ log log p−1 − t(2∆ − 1) log 3 − O(1)
which implies that t = Ω(∆−1 log log p−1 ).
Theorem 2. Even on 2-colored trees or 2-colored graphs of girth Ω(log∆ n), sinkless orientation and
(2∆ − 2)-edge coloring require Ω(log∆ log n) time in RandLOCAL and Ω(log∆ n) time in DetLOCAL.
9
E0 depends on
S
j∈[1,∆−1]
N t (ej )\N t−1 (e0 ) and E1 depends on
8
S
j∈[∆,2∆−2]
N t (ej )\N t−1 (e0 ).
Proof. Consider any sinkless orientation/(2∆ − 2)-edge coloring algorithm with local probability
of failure p. Lemma 2 applies to any vertex v and any radius t such that N t (v) is consistent
with a subgraph of T∆ . Thus, on degree-∆ trees or graphs of girth Ω(log∆ n) [15, 8], we get
Ω(min{∆−1 log log p−1 , log∆ n}) lower bounds. Following the same proof as [10, Theorem 5], this
implies an Ω(log∆ n) lower bound in DetLOCAL, which, according to [10, Theorem 3], implies an
Ω(log∆ log n) lower bound in RandLOCAL. In other words, the weak RandLOCAL lower bound
Ω(∆−1 log log n) implied by Lemma 2 automatically implies a stronger lower bound.
3
Randomized Edge Coloring Algorithm
Elkin, Pettie, and Su [18] showed that for any constant > 0, there is a number ∆ such that for
∆ > ∆ , ∆(1 + )-edge coloring can be solved in
O(TLLL (n, poly(∆), exp(−2 ∆/poly(log ∆))) + T ∗ (n, O(∆)))
rounds in the RandLOCAL model, where
TLLL (n, d, p) is the RandLOCAL complexity for constructive LLL with the parameters d and p on
an n-vertex dependency graph.
T ∗ (n, ∆0 ) is the RandLOCAL complexity for 5∆0 -edge coloring on an n-vertex graph of maximum
degree ∆0 .
It is unclear to what extent the algorithm of [18] still works if we allow = o(1). For instance, it is
unknown whether (∆ + ∆0.8 )-edge coloring can be solved in RandLOCAL.
The algorithm of [18] is based on the Rödl Nibble method as follows. After the ith iteration of the
algorithm, a certain invariant Hi+1 holds at each vertex w.h.p. in ∆, given that Hi holds everywhere
beforehand. To ensure that the invariant Hi+1 is met, a distributed Lovász Local Lemma algorithm
is applied in each iteration of the algorithm. Their algorithm requires O( 12 log( 1 )) iterations,10
which is inefficient if is small, e.g., 1/poly(∆). In this section, we prove the following theorem,
which improves upon the algorithm of [18].
2.5
Theorem 3. Let = ω log√∆ ∆ be a function of ∆. If ∆ > ∆ is sufficiently large there is a
RandLOCAL algorithm for (1 + )∆-edge coloring in time
O (log(1/)) · TLLL n, poly(∆), exp(−2 ∆/ log4+o(1) ∆) + T ∗ (n, O(∆)) .
Notice that exp(−2 ∆/ log4+o(1) ∆) = ω(log ∆), so we may use a distributed LLL algorithm
under any criterion p(ed)λ < 1. There is an inherent tradeoff between the palette size and the
runtime in Theorem 3. Selecting smaller allows us to use fewer colors, but it leads to a higher
p = exp(−2 ∆/ log4+o(1) ∆), which may increase the runtime of an LLL algorithm.
10
To design a ∆(1 + )-edge coloring algorithm for very small ,
√one cannot to afford to have too many iterations.
Observe that each iteration necessarily incur at least a 1 ± O(1/ ∆) factor of drift to the parameters (e.g.,
√ palette
size), since these parameters are upper bounded by O(∆). If the number of iterations is much higher than ∆, then
the effect of the drift becomes non-negligible. For instance, if the number of iterations is Ω(1/2 ) (which is the case
of [18]), then we cannot make = o(∆1/4 ).
9
Runtime of 5∆0 -edge Coloring. It is known that T ∗ (n, ∆0 ) is at most O(log ∆0 ) plus the
DetLOCAL complexity of 3∆0 -edge coloring on poly(log n)-size graphs. This is achieved by ap˜ + 1)-vertex coloring algorithm of [7] to the line graph, where ∆
˜ = 2∆0 − 2 is the
plying the (∆
maximum degree of the line graph.
For the special case of ∆0 = log1+Ω(1) n, (2∆0 − 1)-edge coloring can be solved in RandLOCAL
O(log∗ n) rounds [18]. The state-of-the-art DetLOCAL algorithm [24] for (2 + x)∆0 -edge coloring
has complexity
O(log2 ∆0 · x−1 · log log ∆0 · log1.71 log log ∆0 · log n)
for any x > 1/ log ∆0 . Thus, combining [18, 7, 24], we have
T ∗ (n, ∆0 ) = O(log3 log n · log log log n · log1.71 log log log n) = (log log n)3+o(1) .
This is achieved as follows. If ∆0 = Ω(log2 n), we run the O(log∗ n)-time RandLOCAL algorithm
of [18]. Otherwise, we run the RandLOCAL graph shattering phase of [7] (using the first 2∆0 colors)
followed by the DetLOCAL algorithm of [24] (using the remaining 3∆0 colors) on each component.
Runtime
on Trees.
Consider running our algorithm on a tree with palette size (1 + )∆, where
2.5+x
log √
∆
=Ω
, for some positive constant x. Then the LLL parameters are d = poly(∆) and
∆
p = exp(−2 ∆/ log4+o(1) ) in Theorem 3, which satisfy the criterion p(ed)λ < 1 with λ = Ω(logx ∆).
Using our randomized LLL algorithm for trees (Section 4), we have
TLLL n, poly(∆), exp(−2 ∆/ log4+o(1) ) = O max{ logloglogloglogn n , loglog ∆ log n} .
We claim that T ∗ (n, ∆0 ) = O(log∗ ∆0 + log∆0 log n) on trees. This is achieved as follows. First,
do a O(log∗ ∆0 )-time randomized procedure to partially color the graph using the first 2∆0 colors
so that the remaining uncolored components have size poly(log n). This can be done using the
algorithm of [18] without invoking distributed LLL. Then, apply our deterministic O(log∆0 ñ)-time
algorithm for ∆0 -edge coloring trees (Appendix A) to each uncolored component separately, using
a set of ∆0 fresh colors.
To sum up, the time complexity of (1 + )∆-edge coloring trees is
O log(1/) · max{ logloglogloglogn n , loglog ∆ log n} + log∗ ∆ + log∆ log n
= O log(1/) · max{ logloglogloglogn n , loglog ∆ log n} .
This matches our Ω(log∆ log n) lower bound (Section 2) when 1/, ∆ = O(1).
3.1
Main Algorithm
Our algorithm has two phases. The goal of the first phase is to color a subset of the edges using
def
the colors from C1 = {1, . . . , ∆(1 + ξ)} such that the subgraph induced by the uncolored edges has
degree less than ∆0 = 51 ( − ξ)∆ = Θ(∆). The first phase consists of O(log(1/)) executions of a
distributed Lovász Local Lemma algorithm. The second phase colors the remaining edges using the
def
colors from C2 = {∆(1 + ξ) + 1, . . . , ∆(1 + )} using the fastest available coloring algorithm, which
takes T ∗ (n, ∆0 ) time.
10
Algorithm. In what follows we focus on the first phase. We write Gi to denote the graph induced
by the set of uncolored edges at the beginning of the ith iteration. Each edge e in Gi has a palette
Ψi (e) ⊆ C1 . We write degi (v) to denote the number of edges incident to v in Gi and degc,i (v) to
denote the number of edges incident to v that have color c in their palettes. For the base case, we
set G1 = G and Ψi (e) = C1 for all edges. In the graph Gi , for each vertex v, each edge e, and each
color c, we maintain the following invariant Hi .
Invariant Hi :
(i) degi (v) ≤ di ,
(ii) degc,i (v) ≤ ti ,
and (iii) |Ψi (e)| ≥ pi .
Parameters. Given two numbers η ≥ 1 and ξ ∈ (0, ) (which are functions of ∆), we define three
sequences of numbers {di }, {ti }, and {pi } as follows.
Base case (i = 1):
def
def
t1 = ∆
d1 = ∆
def
p1 = ∆(1 + ξ)
Inductive step (i > 1):
def
di−1 = di−1 · 1 − (1 − 1/pi−1 )2(ti−1 −1)
ti−1
def
2ti−1
ti−1 = ti−1 · 1 −
1 − (1 − 1/pi−1 )2ti−1
(1 − 1/pi−1 )
pi−1
2
ti−1
def
2ti−1
pi−1 = pi−1 · 1 −
(1 − 1/pi−1 )
pi−1
def
di = (1 + δi−1 )di−1
def
ti = (1 + δi−1 )ti−1
def
pi = (1 − δi−1 )pi−1
Drifts (all i):
def
δi =
βi
η
def
βi =
pi
−1
ti
(Notice that β1 = ξ)
The choice of parameters are briefly explained as follows. Consider an ideal situation where
degi−1 (v) = di−1 , degc,i−1 (v) = ti−1 , and |Ψi−1 (e)| = pi−1 for all c, e, and v. Consider a very
simple experiment called OneShotColoring, in which each (uncolored) edge attempts to color itself
by selecting a color uniformly at random from its available palette. An edge e successfully colors
itself with probability (1 − 1/pi−1 )2(ti−1 −1) , since there are 2(ti−1 − 1) edges competing with e for
c ∈ Ψi−1 (e), and each of these 2(ti−1 − 1) edges selects c with probability 1/pi−1 . Thus, by linearity
of expectation, the expected degree of v after OneShotColoring is di−1 , and the parameter di is
simply di−1 with some slack. The parameters {ti−1 , ti , pi−1 , pi } carry analogous meanings. The
term βi represents the second-order error. We need control over {βi } since it influences the growth
of the three sequences {di }, {ti }, and {pi }.
For the base case, it is straightforward to see that we have deg1 (v) = ∆, degc,1 (v) = ∆, and
|Ψ1 (e)| = ∆(1 + ξ), and thus G1 satisfies the invariant H1 . For the inductive step, given that Hi
is met in Gi , we use a distributed LLL algorithm (based on OneShotColoring) to color a subset of
edges in Gi so that the graph induced by the uncolored edges (i.e., Gi+1 ) satisfies Hi+1 . We repeat
this procedure until the terminating condition di ≤ 15 ( − ξ)∆ is met, and then we proceed to the
second phase.
2.5
Analysis. Recall that = ω( log√∆ ∆ ). We set η to be any function of ∆ that is ω(log ∆) such that
2.5
≥ η√∆ . We set ξ = 6η
. The following lemma shows that under certain criteria, the parameters
{di }, {ti }, {pi }, and {βi } are very close to their “ideal” values. See Appendix D for proof.
11
Lemma 3. Consider an index i > 1. Suppose min{di−1 , ti−1 , pi−1 } = ω(log ∆), βi−1 = o(1/ log ∆),
and δi−1 = o(βi−1 / log ∆). Then the following four equations hold.
di = di−1 · (1 ± o(1/ log ∆))(1 − e−2 )
ti = ti−1 · (1 ± o(1/ log ∆))(1 − e−2 )2
pi = pi−1 · (1 ± o(1/ log ∆))(1 − e−2 )2
βi = βi−1 · (1 ± o(1/ log ∆))/(1 − e−2 )
Based on Lemma 3, we have the following lemma.
Lemma 4. Let i? = O(log(1/)) = O(log ∆) be the largest index such that βi? −1 ≤ 1/η. Then the
following four equations hold for any 1 < i ≤ i? .
di = (1 ± o(1/ log ∆))i−1 ∆(1 − e−2 )i−1 = (1 ± o(1))∆(1 − e−2 )i−1
ti = (1 ± o(1/ log ∆))i−1 ∆(1 − e−2 )2(i−1) = (1 ± o(1))∆(1 − e−2 )2(i−1)
pi = (1 ± o(1/ log ∆))i−1 ∆(1 − e−2 )2(i−1) = (1 ± o(1))∆(1 − e−2 )2(i−1)
βi = (1 ± o(1/ log ∆))i−1 ξ/(1 − e−2 )i−1 = (1 ± o(1))ξ/(1 − e−2 )i−1
Proof. To prove the lemma, it suffices to show that the condition of Lemma 3 is met for all indices
1 < i ≤ i? . We prove this by an induction on i. Suppose by the induction hypothesis the four
equations hold at index i − 1. We show that the condition of Lemma 3 is met for the index i,
and so the four equations also hold for index i. Due to 1/η = o(1/ log ∆), we already have βi−1 =
o(1/ log ∆) and δi−1 = o(βi−1 / log ∆). It remains to prove that min{di−1 , ti−1 , pi−1 } = ω(log ∆).
min{di−1 , ti−1 , pi−1 } ≥ (1 ± o(1))∆(1 − e−2 )2(i−1)
(Induction hypothesis for di−1 , ti−1 , pi−1 )
2
(1 ± o(1))ξ
= (1 ± o(1))∆ ·
(Induction hypothesis for βi−1 )
βi−1
≥ (1 ± o(1))ξ 2 η 2 ∆
(βi−1 ≤ 1/η)
= Ω(η 5 )
= ω(log ∆)
It remains to show that (i) the number of iterations
it takes to reach the terminating
condition is
4+o(1)
2
O(log 1/), and (ii) in each iteration, in TLLL n, poly(∆), exp(− ∆/ log
∆) time, invariant
Hi can be maintained. By Lemma 4, we have:
di? = (1 ± o(1))∆(1 − e−2 )i
? −1
(Lemma 4 for di? )
= (1 ± o(1))∆ · ξ/βi?
(Lemma 4 for βi? )
≤ (1 ± o(1))ξη∆
(βi? > 1/η)
1
For our choices of η and ξ, we have di? ≈ ξη∆ = ∆
6 . Thus, the terminating condition di ≤ 5 (−ξ)∆
1
∆
?
must be reached before the i -iteration (since 5 ( − ξ)∆ > 6 ). The number of iterations it takes
to reach the terminating condition is O(log 1/) by Lemma 4 for di .
12
For each 1 < i ≤ i? , we have:
δi2 · min{di , ti , pi } = βi2 ti /η 2
(Definition of δi )
2
= (1 ± o(1)) · ξ/(1 − e−2 )i−1 · ∆(1 − e−2 )2(i−1) /η 2 (Lemma 4 for ti , βi )
= (1 ± o(1)) · ∆(ξ/η)2
= Ω(2 ∆/η 4 )
(Definition of ξ)
= ω(log ∆).
(Definition of )
We will later see in Section 3.2 that this implies that any LLL algorithm with parameters d =
poly(∆) and p = exp(−Ω(∆2 /η 4 )) suffices to maintain the invariant in each iteration. Notice that
if we select η = log1+o(1) ∆, then p = exp(−2 ∆/ log4+o(1) ∆), as desired.
3.2
Maintenance of Invariant
In this section we show how to apply
a distributed LLL algorithm, with parameters d = poly(∆)
2
and p = exp(−Ω δi · min{di , ti , pi } , to achieve the following task: given a graph Gi meeting the
property Hi , color a subset of edges of Gi so that the graph induced by the remaining uncolored
edges satisfies the property Hi+1 . We write Ψ(e) = Ψi (e) for notational simplicity. Consider the
following modification to the underlying graph Gi :
• Each edge e discards colors from its palette to achieve uniform palette size pi .
• Each vertex v locally simulates some imaginary subtrees attached to v and obeying Hi to
achieve uniform color degree ti . That is, if a color c appears in the palette of some edge
incident to a vertex v, then c must appear in the palette of exactly ti edges incident to v.
Consider the following 1-round coloring procedure on the modified graph.
OneShotColoring.
(1) Each edge e selects a color Color? (e) ∈ Ψ(e) uniformly at random.
(2) An edge e successfully colors itself Color? (e) if no neighboring edge also selects Color? (e).
We write S(v) to denote the set of real edges incident to v, and we write Nc (v) to denote the
set of real and imaginary edges incident to v that have c in their palettes. Let S (v) (resp., Nc (v))
be the subset of S(v) (resp., Nc (v)) that are still uncolored after OneShotColoring. Let Ψ (e) be
the result of removing all colors c from Ψ(e) such that some edge incident to e successfully colors
itself by c.
The following concentration bound implies that Hi+1 holds with high probability in the graph
induced by the real uncolored edges after OneShotColoring, and thus we can apply a distributed
LLL algorithm to obtain Gi+1 that meets the invariant Hi+1 . See Appendix E for proof.
Lemma 5. Suppose that Hi holds. The following concentration bounds hold for any δ > 0.
Pr [|S (v)| > (1 + δ)di ] < exp −Ω(δ 2 di )
Pr [|Nc (v)| > (1 + δ)ti | Nc (v) 6= ∅] < exp −Ω(δ 2 ti )
Pr [|Ψ (e)| < (1 − δ)pi | e remains uncolored ] < exp −Ω(δ 2 pi )
13
We write N k (v) to denote the set of all vertices within distance k of v. It is straightforward
to see that (i) S (v) depends only on the colors selected by the edges whose endpoints both are
in N 2 (v), (ii) Nc (v) depends only on the colors selected by the edges whose endpoints are both in
N 3 (v), and (iii) Ψ (e) depends only on the colors selected by the edges whose endpoints are both
in N 2 (u) ∪ N 2 (v), where e = {u, v}.
Thus, the parameters for the distributed LLL are d = poly(∆)
and p = exp −Ω δi2 · min{di , ti , pi } , as desired.
4
Lovász Local Lemma on Trees
Lovász Local Lemma. Consider a set of independent random variables V and a set of bad
events E, where each A ∈ E depends on a subset vbl(A) ⊂ V. Define the dependency graph as
GE = (E, {(A, B) | vbl(A) ∩ vbl(B) 6= ∅)}). Symmetric versions of the Lovász local lemma are
stated in terms of d, the maximum degree in GE , and p = maxA∈E Pr[A]. A standard version of the
LLL says that if ep(d + 1) < 1 then Pr[∩A∈E A] > 0. The (constructive) LLL problem is to assign
values to all variables in V such that no event in E happens.
Tree-Structured LLL. Let T be a tree. Each vertex v holds some variables V(v) and is associated
with
S a bad event E(v) that depends only on variables within distance r/2 of v,Si.e., vbl(E(v)) =
r/2 (v) V(u). If S is a subset of the vertices, we use vbl(S) to be short for
v∈S vbl(E(v)) =
Su∈N S
V(u).
The
dependency
graph
for
this
set
E
of
bad
events
is
exactly
T r , obtained
r/2
v∈S u∈N
(v)
r
by connecting vertices at distance at most r in T , so d ≤ (∆(T )) . Tree-structured dependency
graphs arise naturally from any constant-time (r/2 time) RandLOCAL experiment that is run on a
tree topology.
4.1
Deterministic LLL Algorithms
Recall that a (λ, γ)-network decomposition is a partition of the vertex set into V1 , . . . , Vλ such
that connected components induced by Vi have diameter at most γ. Fischer and Ghaffari [19]
showed that given a (λ, γ)-decomposition of G2E , an LLL instance satisfying p(ed)λ < 1 is solvable
in O(λ(γ + 1)) time. We use a slight generalization of network decompositions. A (λ1 , γ1 , λ2 , γ2 )network decomposition is a partition of the vertices into V1 , . . . , Vλ1 , U1 , . . . , Uλ2 such that connected
components induced by Vi have diameter at most γ1 and those induced by Ui have diameter at most
γ2 .
Lemma 6 (Fischer and Ghaffari [19]). Suppose that a (λ1 , γ1 , λ2 , γ2 )-network decomposition of
G2E is given. Any LLL instance on GE satisfying p(ed)λ1 +λ2 < 1 can be solved in DetLOCAL in
O(λ1 (γ1 + 1) + λ2 (γ2 + 1)) time.
The proof of Theorem 4 uses new network decompositions for trees; see Appendix B.
Theorem 4. Any tree-structured LLL satisfying p(ed)λ < 1 with λ ≥ 2 can be solved in DetLOCAL
in O(max{logλ s, logloglogs s }) time, where s ≤ n is the size of any distance-O(1) dominating set of the
tree.
Proof. Recall that the dependency graph is T r for some tree T and constant r. In Appendix B we
show that a standard (2, O(log s))-decomposition for (T r )2 = T 2r is computable in O(log s) time,
and if λ = Ω(1) is sufficiently large, a (1, O(logλ s), O(λ2 ), 0)-decomposition for T 2r is computable
in O(logλ s) time.
When λ = O(1) is sufficiently small, we apply Lemma 6 with the first network decomposition.
Because the decomposition has two parts, this works with LLL criterion p(ed)2 < 1. When λ
14
is sufficiently large we compute a (1, O(logλ̂ s), O(λ̂2 ), 0)-decomposition in O(logλ̂ s) time, where
q
λ̂ = min{λ, logloglogs s }. We solve the LLL by applying Lemma 6, which takes time O(λ̂2 + logλ̂ s) =
O(max{logλ s, logloglogs s }). Observe that because of the λ̂2 term, we cannot benefit from LLL instances
q
with λ logloglogs s .
Combining Theorem 4 with the O(d2 + log∗ n) graph shattering routine of [19] we obtain a
O(d2 + max{logλ log n, logloglogloglogn n })-time RandLOCAL LLL algorithm for criterion p(ed)λ < 1, λ ≥ 4,
which is efficient only when d is very small. In Section 4.2 we give a new method for computing
a partial assignment to the variables that effectively shatters a large dependency graph into many
independent subproblems, each satisfying a polynomial LLL criterion w.r.t. the unassigned variables.
4.2
Shattering the Dependency Graph
In this section we prove the following efficient shattering lemma.
Lemma 7. Suppose we are given a tree-structured LLL instance T r satisfying LLL criterion p(ed)λ <
1, where λ ≥ 2(4r + 8r). There is a RandLOCAL algorithm that computes a partial assignment in
O(logλ log n) time with the following properties.
1. No bad event occurs under the assignment.
2. The connected components of T r induced by events with unassigned variables have size at most
poly(d) log n. Moreover, each such component contains a distance- 3r
2 dominating set (in the
tree T ) with size at most log n.
√
3. Conditioned on the partial assignment, the probability of any bad event is at most p0 = p
and each component satisfies LLL criterion p0 (ed)λ/2 < 1.
By applying Lemma 7 and then Theorem 4 to each component, we can now efficiently solve
tree-structured LLL instances in O(log log n) time or faster, independent of the dependency graph
degree d.
Theorem 5. Let T r be a tree-structured LLL instance satisfying criterion p(ed)λ < 1 with λ ≥
2(4r + 8r). This LLL can be solved in RandLOCAL in O(max{logλ log n, logloglogloglogn n }) time.
The statement of Lemma 7 suggests an algorithm to compute such a partial assignment φ.
First, draw a total assignment φ to V according to the distribution of the variables. If any bad
event E(v) occurs under φ, update φ by unsetting all variables in vbl(E(v)). In general, whenever
√
Pr[E(v)|φ] exceeds p, update φ by unsetting all variables in vbl(E(v)). This can be viewed as a
contagion dynamic played out on the dependency graph. Bad events that occur under the initial
total assignment are infected, and infected vertices can cause nearby neighbors to become infected.
If this contagion process were actually simulated, it would take Ω(log n) parallel steps to reach
a state. We give an algorithm that computes a stable set (satisfying the other requirements of
Lemma 7) exponentially faster, by avoiding a direct simulation.
Let u be a vertex in the unoriented tree T . Define Cu (k, [i, j]) to be the set of vertices that belong
to kth subtree of u such that the distance to u lies in the interval [i, j]. For example, Cu (k, [1, 1])
d (u) to be the number of different k s.t.
is the kth neighbor of u. For any vertex set S, define deg
S
Cu (k, [1, r]) ∩ S 6= ∅.
Choose µ ≥ 4 and λ0 ≥ 1 to be any integers such that λ ≥ 2(µr + λ0 ). The following bad events
are defined with respect to a fixed total assignment φ to the variables.
15
B(S, v) : Pr [E(v) | vbl(E(v))\ vbl(S)] ≥ (ed)−λ/2
[
B(v) :
B(S, v).
S⊂N r (v),|S|≤µr
In other words, B(S, v) is the event that, if we were to resample vbl(S) (but leave other variables
in vbl(E(v)) set according to φ), the probability that E(v) occurs is at least (ed)−λ/2 . The event
B(v) occurs if it is possible to find a subset S of the right cardinality such that B(S, v) occurs.
We can now consider the probability that these events occur, over a randomly selected total
assignment φ.
Pr[B(S, v)] ≤
φ
Prφ [E(v)]
(ed)−λ
r
0
≤
≤ (ed)−(µ +λ )
−λ/2
Prφ [E(v) | B(S, v)]
(ed)
r
and, by a union bound over the dµ choices of S,
X
0
Pr[B(v)] ≤
Pr[B(S, v)] ≤ (ed)−λ .
φ
S
φ
Intuitively B(v) is the event that E(v) is too close to happening, i.e., relatively few variables
need to be resampled to give E(v) a likely probability of happening. We imagine a contagion
process that samples a total variable assignment φ, and initially infects S consisting of all v such
d (w) > µ until S is
that B(v) occurs. An uninfected vertex w becomes infected (joins S) if deg
S
d (v) > µ” is a good proxy for the
stable. Lemma 8 proves that the criterion for infection “ deg
S
harder-to-analyze criterion “E(v) is dangerously close to happening.”
Lemma 8. Fix a total variable assignment φ, and let S be any vertex set such that (i) v ∈ S if B(v)
d (v) > µ. Then for any v, Pr[E(v) | vbl(E(v))\ vbl(S)] <
occurs under φ, and (ii) v ∈ S if deg
S
−λ/2
(ed)
.
Proof. If v ∈ S, then the probability of seeing E(v) after resampling vbl(S) is p < (ed)−λ , so assume
v 6∈ S.
To prove the lemma, it suffices to show that there exists a vertex set S 0 such that (i) S 0 ⊂ N r (v),
(ii) |S 0 | ≤ µr , and (iii) vbl(S 0 )∩vbl(E(v)) = vbl(S)∩vbl(E(v)), i.e., (iii) says that resampling vbl(S 0 )
is equivalent to resampling vbl(S), from v’s point of view. Since v 6∈ S, we know B(S 0 , v) does not
occur, and so Pr[E(v) | vbl(E(v))\ vbl(S 0 )] < (ed)−λ/2 .
Root the tree at v. We call a vertex u ∈ S “highest” if u is in N r (v) and no ancestor of u is in S.
If H is the set of highest vertices, then vbl(S) ∩ vbl(E(v)) = vbl(H) ∩ vbl(E(v)), so we only need
to bound |H| by µr . Suppose, for the sake of contradiction, that |H| ≥ µr + 1. Define the path
(v = v0 , v1 , . . . , vr ) by selecting vi be the child of vi−1 whose subtree contains the largest number of
vertices in H. We prove by induction that vi contains at least µr−i +1 H-vertices in its subtree. The
base case i = 0 holds by assumption. If vi has µ + 1 subtrees containing
H-vertices,
then vi would
be infected, so by the pigeonhole principle vi+1 must have at least (µr−i + 1)/µ = µr−(i+1) + 1
H-vertices in its subtree. Hence the subtree of vr contains µ0 + 1 = 2 H-vertices, which is a
contradiction since the only vertex eligible to be in H is vr itself.
4.3
Contagion Process
A (q0 , r, µ)-contagion process on an n-vertex tree T is defined as follows. Initially each vertex is
infected with probability q0 and these events are independent for vertices at distance greater than
16
d (v) > µ, then v becomes infected. Our
r. If S is the set of infected vertices at some time and deg
S
goal is to quickly compute a set S that is both stable and small.
Definition 1. A set S is stable if it causes no more infection and small if each connected component
induced by N r/2 (S) contains a distance- 3r
2 dominating set (in the tree T ) of size at most log n.
Lemma 9 connects the contagion problem to finding a partial assignment satisfying the criteria
of Lemma 7. Theorem 6 shows that one can efficiently compute small stable sets in contagion
processes.
Lemma 9. Suppose there is a τ -round RandLOCAL algorithm for finding a small stable set for
0
a ((ed)−λ , r, µ)-contagion process. Then there exists a (τ + O(1))-round RandLOCAL algorithm
for finding a partial assignment to a tree-structured LLL instance satisfying Lemma 7(1–3), under
criterion p(ed)λ < 1, where λ ≥ 2(µr + λ0 ).
Proof. Consider the contagion process defined by choosing a partial assignment to the variables and
initially infecting any vertex v where B(v) occurs. The lower bound on λ implies Pr[B(v)] ≤ q0 =
0
(ed)−λ and according to Lemma 8, any stable set satisfies Lemma 7(3).
Theorem 6. Consider a (q0 , r, µ)-contagion process played on an n-vertex tree T with maximum
degree ∆. In RandLOCAL, we can compute a small stable set S in O(logµ log n) time, where r is
constant, q0 ≤ e−1 d−8r , d = ∆r and µ ≥ 4.
Combining Lemma 9 and Theorem 6 together, Lemma 7 is proved as follows. Pick the largest
µ such that λ ≥ 2(µr + 8r). By Lemma 9 we only need to show a O(logλ log n)-round RandLOCAL
algorithm to find a small stable set for a ((ed)−8r , r, µ)-contagion process. Since λ ≥ 2(4r + 8r) and
µ ≥ 4, by Theorem 6, a small stable set can be computed in O(logµ log n) = O(logλ log n) rounds.
Theorem 6 is proved in the remainder of this section. The algorithm for Theorem 6 simulates
a more virulent contagion process for τ steps using threshold µ/2 rather than µ, then simulates a
reverse-contagion for τ steps, where vertices become uninfected if they were not initially infected
and they have nearby infected vertices in at most µ subtrees. We prove that when τ = O(logµ log n),
the final infected set Lτ is both stable and small. Algorithm 1 presents pseudocode of this process.
Observe that the sets generated by this algorithm have the following containment.
U0 ⊆ · · · ⊆ Uτ = L0 ⊇ · · · ⊇ Lτ .
1
2
3
4
5
6
7
8
9
U0 ← {u ∈ V | u is initially infected}
(u ∈ U0 if B(u) occurs initially);
for 1 ≤ i ≤ τ do
d
Ui ← Ui−1 ∪ {u ∈ V | deg
Ui−1 (u) > µ/2};
end
L0 ← Uτ ;
for 1 ≤ i ≤ τ do
d
Li ← Li−1 \{u ∈ Li−1 \U0 | deg
Li−1 (u) ≤ µ};
end
return Lτ ;
Algorithm 1: Find Small Stable Set
d (v) ≤ µ with high probability, and so Lτ is stable.
Lemma 10. For each vertex v ∈
/ Lτ , deg
Lτ
17
Proof. Root the tree at v. Define S(u) to be the subtree rooted at u. Define Cu0 (k, [i, j]) to be
d
Cu (k, [i, j]) ∩ S(u) and deg0 S (u) to be deg
S∩S(u) (u). Let Lτ +1 be the set of all vertices u s.t.
d
degLτ (u) > µ. In general, if {X(u)} is an ensembleSof events associated with vertices and S a
subset of vertices, we define X(S) to be the event u∈S X(u) and X to be the set of vertices
{v | X(v) occurs}. Consider the following three sequences of events {Fi }0≤i≤τ , {Hi }0≤i≤τ , and
{F̃i }0≤i≤τ .
Fi (u) : (u ∈
/ Ui ) ∧ (u ∈ Li+1 ),
H0 (u) : (u ∈ U0 ),
Hi+1 (u) : H0 (u) ∨ (deg0 Hi (u) ≥ µ/2),
F̃0 (u) : Hτ (u),
F̃i+1 (u) : deg0 F̃i (u) ≥ µ/2.
Notice that v cannot belong to both Uτ \Lτ and Lτ +1 , since otherwise v is contained in L1 , . . . , Lτ .
Therefore, Fτ (v) = (v ∈
/ Lτ ) ∧ (v ∈ Lτ +1 ), and so it suffices to bound the probability of event Fτ (v)
by 1/poly(n) to prove this lemma. We make the following two observations.
• Observation 1: (u ∈ Ui ) ⇒ Hi (u). The base case (i = 0) follows from the definition of Hi .
Assume that (u ∈ Ui−1 ) ⇒ Hi−1 (u). We have:
0
0
d
u ∈ Ui \U0 =⇒ deg
Ui−1 (u) > µ/2 =⇒ deg Ui−1 (u) ≥ µ/2 =⇒ deg Hi−1 (u) ≥ µ/2,
and this implies (u ∈ Ui ) ⇒ Hi (u).
• Observation 2: Fi (u) ⇒ F̃i (u). The base case (i = 0) follows from Observation 1: F0 (u) ⇒
(u ∈ L1 ) ⇒ (u ∈ Uτ ) ⇒ Hτ (u) ⇒ F̃0 (u). Assume that Fi−1 (u) ⇒ F̃i−1 (u). By definition,
d
d (u) > µ while u ∈
/ Ui implies deg
u ∈ Li+1 implies deg
Li
Ui−1 (u) ≤ µ/2. That is,
0
0
d
Fi (u) =⇒ deg
Fi−1 (u) > µ/2 =⇒ deg Fi−1 (u) ≥ µ/2 =⇒ deg F̃i−1 (u) ≥ µ/2,
which is exactly F̃i (u).
Since Fτ (u) ⇒ F̃τ (u), it remains to bound the probability of event F̃τ (v) by 1/poly(n). For
convenience, we write pi = maxu Pr[F̃i (u)] and qi = maxu Pr[Hi (u)]. We will show that (i)
µ τ /r
2
2
pτ ≤ (∆2(r +1) p0 )( 2 ) and (ii) p0 = qτ ≤ ∆r +2 q0 . Therefore,
pτ ≤ (∆2(r
2 +1)
µ τ /r
p0 )( 2 )
≤ (∆3r
2 +4
µ τ /r
q0 ) ( 2 )
≤ 1/poly(n),
as desired. In the remainder of the proof we derive these two inequalities.
Upper Bound of pτ . Notice that the event F̃i (u) is contained in the following event: “there exist
µ/2 many different k such that F̃i−1 (Cu0 (k, [1, r])) occurs”. A consequence of this observation is that
F̃i (Cu0 (k, [1, r])) ⇒ F̃i−1 (Cu0 (k, [2, 2r])). Similarly, we have:
F̃i−1 (Cu0 (k, [1, r])) =⇒ F̃i−2 (Cu0 (k, [2, 2r]))
=⇒ · · ·
=⇒ F̃i−r (Cu0 (k, [r, r2 ])).
Thus, the F̃i (u) is contained in the following event: “there exists µ/2 many different k such
that F̃i−r (Ck0 (r, r2 )) occurs”. Notice that the events F̃i−r (Cu0 (k, [r, r2 ])) for all k are independent, since F̃i (u) only depends on variables associated with vertices in N r (u). By a union bound,
18
2
∆
Pr[F̃i−r (Cu0 (k, [r, r2 ]))] ≤ ∆r pi−r . Taking a union bound over all µ/2
choices of indices for k, we
have
2
2
pi ≤ ∆µ/2 (∆r pi−r )µ/2 ≤ (∆r +1 pi−r )µ/2 .
Since µ/2 ≥ 2, we have
∆2(r
which implies pτ ≤ (∆2(r
2 +1)
2 +1)
pi ≤ (∆2(r
µ τ /r
p0 )( 2 )
2 +1)
pi−r )µ/2 ≤ · · · ≤ (∆2(r
2 +1)
µ i/r
p0 )( 2 )
,
.
Upper Bound of p0 . We use a similar argument to derive the bound of p0 = qτ . Notice that
Hi (u) is contained in the event (u ∈ U0 ) ∨ (deg0 Hi−1 (u) ≥ µ/2). A simple consequence of this is
that Hi (Cu0 (k, [1, r])) ⇒ H0 (Cu0 (k, [1, 1]) ∨ Hi−1 (Cu0 (k, [2, 2r])). Similarly, we have:
Hi (Cu0 (k, [1, r])) =⇒ H0 (Cu0 (k, [1, 1]) ∨ Hi−1 (Cu0 (k, [2, 2r]))
=⇒ · · ·
=⇒ H0 (Cu0 (k, [1, r − 1]) ∨ Hi−r (Cu0 (k, [r, r2 ])).
Thus, Hi (u) is contained in the union of the event H0 (N r (u)) and the following event: “there exist
µ/2 many different k such that Hi−r (Ck0 (r, r2 )) occurs”. The events Hi−r (Cu0 (k, [r, r2 ])) for all k
are independent, since F̃i (u) only depends on the variables associated with vertices in N r (u). By a
2
union bound, Pr[Hi−r (Cu0 (k, [r, r2 ]))] ≤ ∆r qi−r . By taking a union bound over N r (u) and another
∆
choices of indices for k, we have:
union bound over all µ/2
2
qi ≤ ∆r+1 q0 + ∆µ/2 (∆r qi−r )µ/2 .
We show by induction that qi ≤ ∆r
2 +2
q0 for each i. Suppose that qi−r ≤ ∆r
2
qi ≤ ∆r+1 q0 + ∆µ/2 (∆r qi−r )µ/2 ≤ ∆r
2
Since q0 ≤ e−1 ∆−8r and µ ≥ 4, we have (∆2r
(by an implicit assumption that ∆ ≥ 2).
2 +3
2 +1
q0 )µ/2 ∆r
q0 + (∆2r
2 +1
2 +3
2 +2
q0 , then
q0 )µ/2 .
q0 , and so qi ≤ 2∆r
2 +1
q0 ≤ ∆ r
2 +2
q0
Lemma 11. With high probability, each connected component in the subgraph of T induced by
N r/2 (Uτ ) contains a distance- 3r
2 dominating set (in the tree T ) of size at most log n, and so Lτ is
small.
Proof. Define H = T [r+1,3r] to be the graph obtained by connecting vertices in T at distance in
the range [r + 1, 3r]. We first show that, with high probability, H has no connected component D
such that (i) |D| ≥ log n, and (ii) at least half of vertices in D belong to U0 . The existence of such
a component implies that H contains a subtree of log n vertices with at least half of them in U0 .
There are at most 4log n different such tree topologies and each can be embedded into H in less that
n · ∆2r log n ways. Moreover, there are at most 2log n ways of choosing the positions of vertices in
log n/2
U0 on trees and the probability that such a particular tree occurs in H is at most q0
. A union
bound over all trees lets us conclude that such component exists with probability at most
log n/2
4log n · n · ∆2r log n · 2log n · q0
≤ 1/poly(n).
Let S be a connected component in T induced by N r/2 (Uτ ). We pick a distance- 3r
2 dominating
set D of S (in the tree T ) greedily, preferring vertices in U0 over U1 , and U1 over U2 , etc. Each
19
time a vertex is picked we remove from consideration all vertices in its r-neighborhood. Recall that
U0 ⊆ . . . ⊆ Uτ , and notice that D does not contain any vertex not in Uτ . The reason that D is
a distance- 3r
2 dominating set is as follows. The set D is obviously a distance-r dominating set of
Uτ ∩ S (in the tree T ), and Uτ ∩ S is a distance-r/2 dominating set of S (in the tree T ).
We write ui to denote the ith vertex added to D, and write Di = {u1 , . . . , ui }. Let mi denote the
number of connected components induced by Di in the graph T [r+1,2r] (which is the graph obtained
by connecting vertices in T at distance in the range [r + 1, 2r]). We claim that if ui ∈
/ U0 , then
mi < mi−1 . This implies that at least half of the vertices in D belong to U0 . Observe that the set
D is connected in H, and so |D| < log n with high probability.
We prove this claim in the remainder of the proof. Consider the moment some ui ∈
/ U0 is added
to D. We will show that the connected component of Di in the graph T [r+1,2r] that contains u is
formed by merging ui with at least two connected components of Di−1 in the graph T [r+1,2r] .
The algorithm (for finding a small stable set) added u to Uj because u had at least µ/2 ≥ 2
subtrees containing Ui−1 -vertices that are within distance-r to u. Let T1 and T2 be any two such
subtrees. For each k = 1, 2, let vk be a Uj−1 -vertex contained in both Tk and N r (u). Then there
must be a vertex wk ∈ N r (vk ) such that wk has been already added to D, since otherwise we should
pick vk instead of u. Observe that w1 and w2 belong to separate connected components of Di−1 in
the graph T [r+1,2r] , but w1 , w2 , and u are in the same component of Di in the graph T [r+1,2r] since
wk ∈ N r (vk ) ⊆ N 2r (u) for k = 1, 2.
5
Conclusion
Consider this strange phenomenon. The black-box transformations of [10] and [11] imply the
do not have short,
existence of efficient algorithms that are not entirely constructive, i.e., they
√
O(
log
log
n) LLL algorithm
coherent descriptions. For example, Fischer and Ghaffari’s randomized 2
1/5
(for d < (log log n) √ ) implies [10, Theorem 3] that the deterministic complexity of LLL (for
d < (log n)1/5 ) is 2O( log n) . One could then apply the transformation of [11, Theorem 5] to obtain
a deterministic O(log n)-time LLL algorithm for tree-structured dependency graphs (for d = O(1)).
Algorithms constructed in this way are neither useful nor enlightening, but they exist and suggest
that humanly comprehensible algorithms for these problems also exist. (See, e.g., our O(log n)-time
LLL algorithm for trees in Section 4 and Appendix B.)
Our randomized (1+)∆-edge coloring algorithm, together
with [19, 10], implies the existence of
√
O(
log
n)
a deterministic algorithm for ∆ < (log n) running in 2
time. This suggests the existence of
a simple, direct algorithm using network decompositions [33], but to date, network decompositions
have only been used for problems that admit “greedy” algorithms, e.g., MIS or (∆ + 1)-vertex
coloring, but not more difficult problems like k-edge coloring, k < 2∆ − 1. Finding simple, explicit,
and deterministic algorithms for (1 + )∆-coloring is a challenging open problem.
We have proved that the LLL on tree-structured dependency graphs can be solved in O(log log n)
time w.h.p. (or faster), confirming [11, Conjecture 1] for this case. It is unclear if there is any hope
of extending this type of algorithm to general LLL instances, absent a breakthrough in network
decomposition technology [19, 34].
References
[1] E. Arjomandi. An efficient algorithm for colouring the edges of a graph with ∆ + 1 colours.
INFOR: Information Systems and Operational Research, 20(2):82–101, 1982.
20
[2] B. Awerbuch, A. V. Goldberg, M. Luby, and S. A. Plotkin. Network decomposition and locality
in distributed computation. In Proceedings 30th IEEE Symposium on Foundations of Computer
Science (FOCS), pages 364–369, 1989.
[3] L. Barenboim. Deterministic (∆ + 1)-coloring in sublinear (in ∆) time in static, dynamic
and faulty networks. In Proceedings of the 2015 ACM Symposium on Principles of Distributed
Computing (PODC), pages 345–354, 2015.
[4] L. Barenboim and M. Elkin. Deterministic distributed vertex coloring in polylogarithmic time.
J. ACM, 58(5):23, 2011.
[5] L. Barenboim, M. Elkin, and F. Kuhn. Distributed (∆ + 1)-coloring in linear (in ∆) time.
SIAM J. Comput., 43(1):72–95, 2014.
[6] L. Barenboim, M. Elkin, and T. Maimon. Deterministic distributed (∆ + o(∆))-edge-coloring,
and vertex-coloring of graphs with bounded diversity. In Proceedings of the 2017 ACM Symposium on Principles of Distributed Computing (PODC), 2017.
[7] L. Barenboim, M. Elkin, S. Pettie, and J. Schneider. The locality of distributed symmetry
breaking. J. ACM, 63(3):20:1–20:45, 2016.
[8] B. Bollobás. Extremal graph theory, volume 11 of London Mathematical Society Monographs.
Academic Press Inc. [Harcourt Brace Jovanovich Publishers], London, 1978.
[9] S. Brandt, O. Fischer, J. Hirvonen, B. Keller, T. Lempiäinen, J. Rybicki, J. Suomela, and
J. Uitto. A lower bound for the distributed Lovász local lemma. In Proceedings 48th ACM
Symposium on the Theory of Computing (STOC), pages 479–488, 2016.
[10] Y.-J. Chang, T. Kopelowitz, and S. Pettie. An exponential separation between randomized
and deterministic complexity in the LOCAL model. In Proceedings 57th IEEE Symposium on
Foundations of Computer Science (FOCS), pages 615–624, 2016.
[11] Y.-J. Chang and S. Pettie. A time hierarchy theorem for the LOCAL model. In Proceedings
58th IEEE Symposium on Foundations of Computer Science (FOCS), 2017.
[12] K.-M. Chung, S. Pettie, and H.-H. Su. Distributed algorithms for the Lovász local lemma and
graph coloring. Distributed Computing, 30:261–280, 2017.
[13] R. Cole and U. Vishkin. Deterministic coin tossing with applications to optimal parallel list
ranking. Information and Control, 70(1):32–53, 1986.
[14] A. Czygrinow, M. Hanckowiak, and M. Karonski. Distributed O(∆ log n)-edge-coloring algorithm. In Proc. ESA 2001, pages 345–355, 2001.
[15] X. Dahan. Regular graphs of large girth and arbitrary degree. Combinatorica, 34(4):407–426,
2014.
[16] D. P. Dubhashi, D. A. Grable, and A. Panconesi. Near-optimal, distributed edge colouring via
the nibble method. Theor. Comput. Sci., 203(2):225–251, 1998.
[17] D. P. Dubhashi and A. Panconesi. Concentration of Measure for the Analysis of Randomized
Algorithms. Cambridge University Press, 2009.
21
[18] M. Elkin, S. Pettie, and H. H. Su. (2∆−1)-edge coloring is much easier than maximal matching
in the distributed setting. In Proceedings 26th ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 355–370, 2015.
[19] M. Fischer and M. Ghaffari. Sublogarithmic distributed algorithms for Lovász local lemma
with implications on complexity hierarchies. In Proceedings 31st International Symposium on
Distributed Computing (DISC), 2017.
[20] M. Fischer, M. Ghaffari, and F. Kuhn. Deterministic distributed edge coloring via hypergraph
maximal matching. In Proceedings 58th IEEE Symposium on Foundations of Computer Science
(FOCS), 2017.
[21] P. Fraigniaud, M. Heinrich, and A. Kosowski. Local conflict coloring. In Proceedings 57th
Annual IEEE Symposium on Foundations of Computer Science (FOCS), 2016.
[22] H. N. Gabow, T. Nishizeki, O. Kariv, D. Leven, and O. Terada. Algorithms for edge-coloring
graphs. Technical Report TRECIS-8501, Tohoku University, 1985.
[23] M. Ghaffari. An improved distributed algorithm for maximal independent set. In Proceedings
27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 270–277, 2016.
[24] M. Ghaffari, J. Hirvonen, F. Kuhn, Y. Maus, J. Suomela, and J. Uitto. Improved distributed
degree splitting and edge coloring. In Proceedings 31st International Symposium on Distributed
Computing (DISC), 2017.
[25] M. Ghaffari and H.-H. Su. Distributed degree splitting, edge coloring, and orientations. In
Proceedings 28th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 2505–
2523, 2017.
[26] I. Holyer. The NP-completeness of edge-coloring. SIAM Journal on Computing, 10(4):718–720,
1981.
[27] H. J. Karloff and D. B. Shmoys. Efficient parallel algorithms for edge coloring problems. J.
Algorithms, 8(1):39–52, 1987.
[28] F. Kuhn and R. Wattenhofer. On the complexity of distributed graph coloring. In Proceedings
25th Annual ACM Symposium on Principles of Distributed Computing (PODC), pages 7–15,
2006.
[29] N. Linial. Locality in distributed graph algorithms. SIAM J. Comput., 21(1):193–201, 1992.
[30] G. L. Miller and J. H. Reif. Parallel tree contraction–Part I: fundamentals. Advances in
Computing Research, 5:47–72, 1989.
[31] R. A. Moser and G. Tardos. A constructive proof of the general Lovász local lemma. J. ACM,
57(2), 2010.
[32] M. Naor. A lower bound on probabilistic algorithms for distributive ring coloring. SIAM
J. Discrete Mathematics, 4(3):409–412, 1991.
[33] A. Panconesi and A. Srinivasan. The local nature of ∆-coloring and its algorithmic applications.
Combinatorica, 15(2):255–280, 1995.
22
[34] A. Panconesi and A. Srinivasan. On the complexity of distributed network decomposition.
J. Algor., 20(2):356–374, 1996.
[35] D. Peleg. Distributed Computing: A Locality-Sensitive Approach. SIAM, 2000.
[36] S. Pettie and H.-H. Su. Distributed algorithms for coloring triangle-free graphs. Information
and Computation, 243:263–280, 2015.
[37] V. G. Vizing. On an estimate of the chromatic class of a p-graph. Diskret. Analiz No., 3:25–30,
1964.
23
A
Deterministic Algorithms for Edge-coloring Trees
Let T = (V, E) be a tree with n vertices and N + (v) be the inclusive neighborhood of v. We decompose T using two operations inspired by Miller and Reif [30], the second of which is parameterized
by an integer k ≥ 2.
• Rake: Remove all leaves and isolated vertices from T .
• Compress: Remove the set {v ∈ V | for every u ∈ N + (v), degT (u) ≤ k} from T .
Theorem 7. Alternately applying compress and rake logk n + 1 times removes all vertices from any
n-vertex tree.
Proof. Root T at an arbitrary vertex and let size(v) be the number of vertices in the subtree
rooted at v. We prove by induction that if size(v) ≤ k i , v will be removed after the first i + 1
rounds of compress and rake. The claim is trivially true when i = 0. Assume the claim is true
for i − 1. Let v be any vertex with size(v) ∈ (k i−1 , k i ] and define V 0 to be the set of u such that
(i) size(u) ∈ (k i−1 , k i ] and (ii) u is in the subtree rooted at v. Notice that each vertex u ∈ V 0
has degV 0 (u) ≤ k, since otherwise size(u) > k i . By the inductive hypothesis, all descendants of v
except V 0 have been removed after i rounds of compress/rake. The (i + 1)th compress will remove
any remaining vertices in V 0 − {v} (the degree of the parent of v is unbounded, so v may not be
removed), and if v still remains, the (i + 1)th rake will remove it.
Theorem 8. There is an O(log∆ n)-time DetLOCAL algorithm for ∆-edge coloring a tree, where
∆ ≥ 3.
Proof. Let β be the constant such that Linial’s algorithm [29] finds a β∆2 -edge coloring in O(log∗ n−
log∗ ∆ + 1) time. We begin by decomposing T with compress/rake steps, using parameter k =
max{2, b(∆/β)1/3 c}. Define Ti = (Vi , Ei ) to be the forest before the ith round of compress and
rake, and let Vic and Vir be those vertices removed by the ith compress and rake, respectively.
We edge-color the trees Tlogk n , . . . , T0 = T in this order. Given a coloring of Ti+1 , we need to
color the remaining uncolored edges in Ti . Let u ∈ Ti+1 be a vertex and v1 , . . . , vx ∈ Vir be the
vertices incident to u removed by a rake. At this point u is incident to at most ∆ − x colored
edges. It assigns to {u, v1 }, . . . , {u, vx } distinct available colors from its palette. We now turn to
the vertices removed by a compress operation. First suppose that ∆ is large enough such that
k = b(∆/β)1/3 c. Let φ be a βk 2 -edge coloring of the (as yet uncolored) subgraph of Ti incident to
Vic . Partition the palette {1, . . . , ∆} into βk 2 parts P1 , . . . , Pβk2 . Each part has size ∆/(βk 2 ) ≥ k.
Each v ∈ Vic colors each edge {v, u} any available color in Pφ({v,u}) . Since degTi (u) ≤ k, at most
k − 1 of its incident edges may already be colored. All calls to Linial’s βk 2 -edge coloring algorithm
can be executed in parallel, so the overall time is O(logk n + log∗ n − log∗ k) = O(log∆ n).
c
When k = 2, the subgraph induced by V0c ∪ · · · ∪ Vlog
consists of a set of paths. In O(log∗ n)
kn
time, find an initial 3-edge coloring of these paths. We now color Tlogk n , . . . , T0 in this order.
Coloring the edges removed during a rake is done as before. The set Vic removed in one compress
induces some paths, each end-edge of which may be adjacent to one (previously colored) edge in
Ti+1 . If the initial color of an end-edge conflicts with the coloring of Ti+1 , recolor it any available
color. When k = 2 this procedure takes O(log∗ n + logk n) = O(log∆ n) time.
Theorem 9. Any oriented tree T can be (∆ + 1)-edge colored in O(log∗ n) time.
Proof. Initially pick color φ0 ({u, parent(u)}) = i if ID(u) is the ith largest ID among its siblings.
−1
Observe that for any i, φ−1
0 (i) is a subgraph consisting of oriented paths, and that φ0 (∆) is at
24
most one edge, attached to the root. For each i ∈ {1, . . . , ∆ − 1}, in parallel, recolor φ−1
0 (i) using
the color-set {i, ∆, ∆ + 1} in such a way that the most ancestral edge in each path remains colored
i. The result is a legal (∆ + 1)-edge coloring. This takes O(log∗ n) time [13, 29].
Theorem 10. Any ∆-edge coloring algorithm for oriented trees takes Ω(log∆ n) time in RandLOCAL.
Proof. Let T be an oriented ∆-regular tree with height log∆ n and A be an edge-coloring algorithm
running in (log∆ n)/3 time. The color of {u, parent(u)} is uniquely determined by the colors of
the edges incident to leaf-descendants of u. Since the (log∆ n)/3-neighborhood of u and its leafdescendants generally do not intersect, u has a 1/∆ chance of guessing the correct edge color; if it
guesses incorrectly, there must be a violation somewhere in the subtree rooted at u.
B
Network Decomposition of Trees
Recall that a (λ, γ)-network decomposition is a decomposition of the vertices of a graph into λ
parts V1 , . . . , Vλ such that each connected component in the graph induced by each Vi has diameter
at most γ. We generalize this notation as follows. A (λ1 , γ1 , λ2 , γ2 )-network decomposition is a
decomposition of the vertices of a graph into λ1 + λ2 parts V1 , . . . , Vλ1 , U1 , . . . , Uλ2 such that each
connected component in the graph induced by each Vi has diameter at most γ1 and each connected
component in the graph induced by each Ui has diameter at most γ2 .
A distance-d dominating set of a graph G is a vertex set S such that for each vertex v in the graph
G, there exists u ∈ S such that dist(u, v) ≤ d. In this section we give two network decomposition
algorithms for T k where T = (V, E) is an n-vertex tree that contains a distance-d dominating set
S of size s. In our application d and k are constants. We assume all vertices agree on the numbers
(d, k, s). We do not need a specific dominating set S be given as an input.
B.1
A (2, O(k log s + d))-Network Decomposition
Theorem 11. Let T be a tree containing a distance-d dominating set of size s. There is a DetLOCAL
algorithm A that computes a (2, O(k log s + d))-network decomposition of T k in O(k log s + d +
k log∗ n) time.
In what follows we prove Theorem 11. Consider the following two tree operations. They are
similar to the ones described in [11], which are inspired by Miller and Reif [30]. The second operation
is parameterized by an integer ` ≥ 2. In our application we set ` = 3k.
• Rake: Remove all leaves and isolated vertices.
• Compress: Remove all vertices that belong to some path P such that (i) all vertices in P have
degree at most 2, and (ii) the number of vertices in P is at least `.
Let A0 be the algorithm on the tree T defined as follows: (1) do 3d + 1 rakes; (2) repeat log s times:
do 1 compress and then ` − 1 rakes.
Lemma 12. Algorithm A0 removes all vertices in T .
Proof. Let S be any size-s distance-d dominating set of T . Root T at an arbitrary vertex and
let size(v) be the number of vertices in the subtree rooted at v that belong to S. For any vertex
v ∈ V , we prove by induction that (i) if size(v) ≤ 1, then v is removed in Step (1) of A0 , and (ii) if
1 < size(v) ≤ 2i , then v is removed on or before the ith iteration of Step (2) of A0 .
25
For the case size(v) ≤ 1, the height of the subtree rooted at v is at most 3d, and so the entire
subtree (including v) must be removed after 3d + 1 rakes. For the case 2i−1 < size(v) ≤ 2i , we
assume by inductive hypothesis that all vertices u with size(u) ≤ 2i−1 have been removed before
the ith iteration of Step (2). With respect to the vertex v, define V 0 to be the set of all vertices
u such that (i) size(u) > 2i−1 , and (ii) u is in the subtree rooted at v. The set V 0 induces a path
such that v is an endpoint, since otherwise size(v) > 2 · 2i−1 = 2i . Let C be a connected component
induced by vertices in V 0 that are not removed yet. If |C| ≥ `, then all vertices in C are removed
after 1 compress. Otherwise, all vertices in C are removed after ` − 1 rakes.
Notice that a (2, ·)-network decomposition of T k is simply a partition V = V1 ∪ V2 such that for
both c = 1, 2, any two vertices u and v in two separate components of Vc must have distT (u, v) > k.
Recall that A0 performs L = (3d + 1) + ` log s rake/compress operations. We write Ui to denote
the
SL set of all vertices that are removed during the ith rake/compress operation. For any labeling
j=i Uj → {1, 2}, define the property Pi as follows.
• Each connected component induced by vertices of the same label have diameter at most
2tir + (6k − 2)tic , where tir (resp., tic ) is the total number of rakes (resp., compresses) after the
ith tree operation.
S
• Let u and v be any two vertices in L
j=i Uj with the same label c ∈ {1, 2}. If u and v belong to
S
separate components induced by the vertices in L
j=i Uj labeled c, then the distance between
SL
u and v in j=i Uj is strictly more than k.
Notice that any labeling V → {1, 2} with the property P1 is a (2, O(k log s + d))-network decomposition of T k . This is because 2t1r + (6k − 2)t1c ≤ 6kL = O(k log s + d).
We
S are now in a 0position to present the algorithm A. The algorithm A begins by computing
V = L
i=1 Ui using A . Then, for i = L down to 1, label all vertices v ∈ Ui as follows.
S
• (Case: the ith operation is rake) Let v ∈ Ui . For the case that v is a leaf of L
j=i Uj , let u be
SL
/ Ui , then v adopts the same label as u. Otherwise,
the unique neighbor of v in j=i Uj . If u ∈
SL
u ∈ Ui must also be a leaf of j=i Uj , and we label both u and v the same by any c ∈ {1, 2}.
S
For the case that v is an isolated vertex of L
j=i Uj , we label v by any c ∈ {1, 2}.
• (Case: the ith operation is compress) Let P be a path that is a connected component of Ui .
Notice that the number of vertices in P is at least ` = 3k. Compute a labeling of the vertices
in P meeting the following conditions: (i) each connected component induced by vertices of
the same
label has size within [k, 3k), (ii) if v is an endpoint of P that is adjacent to a vertex
S
u∈ L
j=i+1 Uj , then the label of v is the same as the label of u. If a (2k + 1, 2k)-ruling set of
P is given, such a labeling can be computed in O(k) time.
S
It is straightforward to prove by an induction that the above labeling for L
j=i Uj has the property
∗
Pi . The total running time of A is O(k log s + d + k log n), since the ruling set computation of
paths removed by compress operation can be computed in O(k log∗ n) time in parallel.
B.2
A (1, O(k logλ s + d), O(λ2 ), 0)-Network Decomposition
Theorem 12. Let T be a tree containing a distance-d dominating set of size s. There is a
DetLOCAL algorithm A that computes a (1, O(k logλ s + d), O(λ2 ), 0)-network decomposition of T k
in O(k logλ s+d+k log∗ s) time, where λ = Ω(k) is sufficiently large (i.e., λ ≥ ck for some universal
constant c).
26
In what follows we prove Theorem 12. Consider the following two tree operations. We denote
the underlying graph before the ith tree operation as Ti .
• Rake: Remove all leaves and isolated vertices.
• Compress: Remove all vertices v such that |NT2.5k
(v)| ≤ λ.
i
Set m = λ/(2.5k). Let A∗ be the algorithm on the tree T defined as follows: (1) do 3d + 1 rakes;
(2) repeat logm s times: do 1 compress and then 2.5k rakes.
Lemma 13. Algorithm A∗ removes all vertices in T .
Proof. Let S be any size-s distance-d dominating set of T . Root T at an arbitrary vertex and let
size(v) be the number of vertices in the subtree rooted at v that belong to S. We prove by induction
that (i) if size(v) ≤ 1, then v is removed in Step (1) of A∗ , and (ii) if 1 < size(v) ≤ mi , v is removed
within the first i iterations in Step (2) of A∗ .
For the case size(v) ≤ 1, the height of subtree rooted at v is at most 3d, and so the entire subtree
(including v) must be removed after 3d + 1 rakes. For the case mi−1 < size(v) ≤ mi , Assume the
claim is true for i − 1. Let v be any vertex with size(v) ∈ (mi−1 , mi ] and define V 0 to be the set of u
such that (i) size(u) > mi−1 and (ii) u is in the subtree rooted at v. By the inductive hypothesis, all
descendants of v except those in V 0 have been removed after i − 1 iterations of Step (2). Therefore,
the set V 0 induces a subtree rooted at v having at most m − 1 leaves, which means the degree of
each vertex in V 0 is at most m. For those vertices u ∈ V 0 whose depth is at least 2.5k from v,
|N 2.5k (u)| ≤ m(2.5k) ≤ λ, so they will be removed after one compress. Then the rest will also be
removed during the next 2.5k rakes.
Now we present our algorithm A for network decomposition. First apply A∗ to T , decomposing
it with rakes and compresses. Next, for any vertex v removed by compress, we mark all vertices in
N k/2 (v), i.e.,
M = {u | ∃v removed by compress, u ∈ N k/2 (v)}
is the set of all marked vertices. For any two vertices u, v ∈ M, add an edge between them if
d(u, v) ≤ k in T . Denote the resulting graph by G.
The (1, O(k logλ s + d), O(λ2 ), 0) network decomposition of T k is computed by assigning color
0 to all unmarked vertices, and color the remaining vertices in G with {1, . . . , O(λ2 )}. We will
show that (i) ∆(G) ≤ λ, and so the coloring can be computed using Linial’s algorithm [29] in
O(k log∗ s) time, (ii) each connected component induced by unmarked vertices (in T k ) has diameter
O(k logλ s + d).
• After deleting those vertices removed by compress from T , the diameter of each connected
component (in T ) is O(k logλ s + d), since the total number of rakes is O(k logλ s) + 3d + 1.
We show that the diameter of each connected component of unmarked vertices in T k is still
O(k logλ s + d). Consider any pair of unmarked vertices u and v. Notice that u and v must be
removed by rake. Suppose that u and v are not connected in T after deleting those vertices
removed by compress from T . Assume the first time they become disconnected is iteration i,
which is due to the removal of a vertex w in compress. Since all vertices in N k/2 (w) are marked,
any path in T connecting u and v must has a subpath consisting of at least 2(k/2) + 1 > k
marked vertices. Thus, u and v are also disconnected in T k after deleting all marked vertices.
• For any marked vertex v, we claim that |N k (v) ∩ M| ≤ λ (in T ), and so ∆(G) ≤ λ. Let u be
the first vertex marked in N k (v). The vertex u is added to M due to the removal of a vertex
27
w ∈ N k/2 (u) in a compress operation (it is possible that u = w). Suppose that w was removed
in iteration i. Then we have |NT2.5k
(w)| ≤ λ. Notice that N k (v) ∩ M ⊆ NTki (v) ⊆ NT2.5k
(w),
i
i
since d(v, w) ≤ d(v, u) + d(u, w) ≤ 1.5k, and since NTki (v) contains all possible marked vertices
within distance-k of v.
C
Lower Bounds for Augmenting Path-Type Algorithms
In this section, we show that for c ∈ [1, ∆
3 ], any algorithm for (∆ + c)-edge coloring based on
recoloring subgraphs to extend partial colorings needs Ω( ∆c log n) rounds.
Theorem 13. Let ∆ be the maximum degree and c ∈ [1, ∆
3 ]. For any n, there exist an n-vertex
graph G = (V, E) and a partial edge coloring φ : E → {1, . . . , ∆ + c, ⊥} with the following property.
For any coloring φ0 : E → {1, . . . , ∆ + c, ⊥} that colors a strict superset of the edges colored by φ,
φ and φ0 differ on a subgraph of diameter Ω( ∆c log n).
As a special case, suppose that G is a partially (∆ + c)-edge colored graph, with exactly one
uncolored edge e0 . A natural approach to color G is to find an “augmenting path” e0 e1 · · · e` , and
then recolor the path. That is, for 0 ≤ i ≤ ` − 1, let the new color of ei be the old color of ei+1 ,
and then color the last edge e` by choosing any available color (if possible). This approach leads
to a distributed algorithm for Brooks’ theorem [34]. However, Theorem 13 implies the existence of
a graph where any augmenting path has length Ω( ∆c log n), and so any “augmenting paths”-based
algorithm for constructive Vizing’s theorem must take at least Ω( ∆c log n) time, which is inefficient
for large ∆.
Construction. The construction of the partially edge-colored graph for Theorem 13 is as follows.
Without loss of generality, assume that ∆ + c is even, and let k = ∆+c
2 . Divide the color palette
{1, · · · , ∆ + c} into two sets S1 = {1, · · · , k} and S2 = {k + 1, · · · , ∆ + c}. Let k 0 = ∆ − k. Let
e0 = {u0 , v0 } be the uncolored edge. We construct a graph consists of ` layers of vertices, which is
denoted as G∗ (`, ∆, c).
Layer 0 consists of only u0 and v0 . The vertex u0 has k neighbors other than v0 , which form
layer 1. The edges between u0 and these k vertices are colored by S1 = {1, · · · , k}. Suppose that
layers 0 . . . i have been constructed. Layer i + 1 is constructed as follows.
Let ni be the current number of vertices in layer i. We divide these ni vertices into b nk0i c groups of
size k 0 , and the remaining ni mod k 0 vertices are ungrouped. Each group forms a complete bipartite
graph Kk0 ,k with k corresponding new vertices in layer i + 1. In total, b nk0i ck new vertices are added
to layer i + 1. Notice that a complete bipartite graph Kk0 ,k can be properly k-edge colored. If i is
even, we use the palette S1 to color these complete bipartite graphs; otherwise we use S2 . The ni
mod k 0 ungrouped vertices in layer i are promoted to layer i + 2. When we assigning groups to the
vertices in layer i + 2, we prioritize these vertices to ensure that each vertex is promoted at most
twice. We have the formula:
j n
k
i
ni+1 = b 0 ck + (ni−1 mod k 0 ) /k 0 · k 0 , i ≥ 2.
k
The construction of v0 ’s side is similar. The only difference is that we switch the roles of S1 and
S2 for coloring complete bipartite graphs. We call the partial edge coloring φ. See Figure 1 for a
concrete example which shows the construction of layers 0, . . . , 5. Notice that some vertices in layer
4 have been promoted to layer 6 (during the construction of layer 5).
Notice that the ratio of the number of vertices in two consecutive layers is Θ(logk/k0 n) =
∆
Θ( c log n), and so we can set ` = Θ( ∆c log n). The distance from the uncolored edge to any
28
Figure 1: An example of the construction, where ∆ = 5, c = 1, k = 3 and k 0 = 2. Edges
with palette S1 = {1, 2, 3} are colored in red, and edges with palette S2 = {4, 5, 6} are
colored in blue.
vertex in layer ` is at least 3` = Θ( ∆c log n), since any edge can skip at most two layers. To prove
Theorem 13, it suffices to prove the following lemma.
Lemma 14. It is impossible to extend the partial edge coloring φ of G∗ (`, ∆, c) to a total edge
coloring by only recoloring edges within layer 0 to layer ` − 1.
Proof. Without loss of generality, we only consider u0 ’s side. Let us refer to the edges connecting
layer i and higher layers as the edges of layer i. For an odd (resp., even) `, fix the colors of edges
of layer ` using only the color palette S1 (resp., S2 ). Consider the case ` is odd (the other case is
similar). Observe that each grouped vertex in layer ` − 1 has k neighbors in layer `. Therefore,
the edges of layer ` − 1 must have their colors picked from S2 . By an induction, edges of layer 1
must be colored using S1 . Similarly, edges of layer 1 at v0 ’s side must be colored using S2 , and so
no available color is left for e0 = {u0 , v0 }. Thus, even if all edges within layer 0 to layer ` − 1 are
allowed to be recolored, we cannot obtain a proper edge coloring of G∗ (`, ∆, c).
D
Estimates of Parameters
In this section, we prove Lemma 3. We assume min{di−1 , ti−1 , pi−1 } = ω(log ∆), βi−1 = o(1/ log ∆),
and δi−1 = o(βi−1 / log ∆). The two terms (1 − 1/pi−1 )2ti−1 and pti−1
(1 − 1/pi−1 )2ti−1 show up
i−1
in the definition of di−1 , ti−1 , and pi−1 . We begin by showing that these two terms are both
e−2 (1 + o(1/ log ∆)). We use the fact pti−1
= βi−11 +1 in the following calculation.
i−1
29
(1 − 1/pi−1 )2ti−1 = e−2ti−1 /pi−1 (1 − O(ti−1 /p2i−1 ))
= e−2 · e2(1−ti−1 /pi−1 ) (1 − O(ti−1 /p2i−1 ))
1
−2
2(1−ti−1 /pi−1 )
1−O
=e ·e
(1 + βi−1 )pi−1
= e−2 · e2(1−ti−1 /pi−1 ) (1 − o(1/ log ∆))
= e−2 · e2βi−1 /(βi−1 +1) (1 − o(1/ log ∆))
= e−2 · (1 + O(2βi−1 /(βi−1 + 1)))(1 − o(1/ log ∆))
= e−2 · (1 + o(1/ log ∆))(1 − o(1/ log ∆))
= e−2 (1 + o(1/ log ∆)).
ti−1
ti−1
(1 − 1/pi−1 )2ti−1 = e−2 ·
· (1 + (1/ log ∆))
pi−1
pi−1
= e−2 (1 + o(1/ log ∆))/(1 + βi−1 )
(∗)
by (*)
= e−2 (1 + o(1/ log ∆))/(1 + o(1/ log ∆))
= e−2 (1 ± o(1/ log ∆)).
We are in a position to derive the first three equations in Lemma 3 (i.e., estimates of di , ti , and
pi ). Recall that δi−1 = o(1/ log2 ∆) and 1/pi−1 = o(1/ log ∆).
di = di−1 · (1 + δi−1 ) 1 − (1 − 1/pi−1 )2(ti−1 −1)
= di−1 · (1 + o(1/ log2 ∆)) 1 − e−2 (1 + o(1/ log ∆))/(1 − 1/pi−1 )2
= di−1 · (1 + o(1/ log2 ∆)) 1 − e−2 (1 + o(1/ log ∆))
= di−1 · (1 ± o(1/ log ∆))(1 − e−2 ).
ti−1
2ti−1
1 − (1 − 1/pi−1 )2ti−1
(1 − 1/pi−1 )
ti = ti−1 · (1 + δi−1 ) 1 −
pi−1
2
2
= ti−1 · (1 + o(1/ log ∆)) 1 − e−2 (1 ± o(1/ log ∆))
= ti−1 · (1 ± o(1/ log ∆))(1 − e−2 )2 .
2
ti−1
pi = pi−1 · (1 − δi−1 ) 1 −
(1 − 1/pi−1 )2ti−1
pi−1
2
= pi−1 · (1 − o(1/ log2 ∆)) 1 − e−2 (1 ± o(1/ log ∆))
= pi−1 · (1 ± o(1/ log ∆))(1 − e−2 )2 .
30
Finally, we derive the last equation in Lemma 3 (estimate of βi ).
pi
−1
ti
(1 − δi−1 )pi−1
=
−1
(1 + δi−1 )ti−1
βi =
ti−1
2t
pi−1 1 − pi−1 (1 − 1/pi−1 ) i−1
= (1 − O(δi−1 )) ·
·
−1
ti−1
1 − (1 − 1/pi−1 )2ti−1
pi−1
2ti−1
ti−1 − (1 − 1/pi−1 )
−1
= (1 − O(δi−1 )) ·
1 − (1 − 1/pi−1 )2ti−1
pi−1
pi−1
2ti−1
−
1
+
O(δ
)
−
+
(1
−
1/p
)
i−1
i−1
ti−1
ti−1
=
2t
i−1
1 − (1 − 1/pi−1 )
pi−1
pi−1
2ti−1
−
1
+
O(δ
)
−
+
(1
−
1/p
)
i−1
i−1
ti−1
ti−1
=
1 − e−2 (1 + o(1/ log ∆))
βi−1 − O(δi−1 )
=
(1 − e−2 )(1 − o(1/ log ∆))
βi−1 (1 − o(1/ log2 ∆))
=
(1 − e−2 )(1 − o(1/ log ∆))
= βi−1 · (1 ± o(1/ log ∆))/(1 − e−2 ).
E
Definition of pi−1 and ti−1
by (*)
−
pi−1
+ (1 − 1/pi−1 )2ti−1 = −Θ(1)
ti−1
δi−1 = o(1/ log2 ∆)
Concentration Bounds
In this section we prove Lemma 5. For notational simplicity, we ignore all subscripts i. We assume
the entire graph has uniform color degree t and uniform palette size p. Recall (from Lemma 3)
that we assume t = Θ(p). Each vertex v is associated with a subset S(v) of (real) edges incident
to v such that |S(v)| ≤ d. Throughout this section, we no longer make a distinction between real
and imaginary edges and vertices; both of them are considered in the calculation. We write Xi to
denote (X1 , . . . , Xi ). We use the following tools to derive the concentration bounds. Theorem 14 is
from Dubhashi-Panconesi [17]. Lemma 15 follows from straightforward calculation.
Theorem 14 ([17, Equation (8.5)]). Let X1 , . . . , Xn be an arbitrary set of random variables. Let
def
f (X1 , . . . , Xn ) be such that E[f ] is finite. We write Di = E[f |Xi ] − E[f |Xi−1 ]. Suppose that there
exist M and σi2 for 1 ≤ i ≤ n meeting the following conditions.
• For any assignment of random variables Xi−1 , Var[Di |Xi−1 ] ≤ σi2 .
• For 1 ≤ i ≤ n, and for any assignment of random variables Xi , |Di | ≤ M .
2
s
Then Pr[f > E[f ] + s] ≤ exp − 2 Pn σ2 +M s/3 .
( i=1 i
)
Lemma 15. Let X be a random variable such that (i) E[X] = 0, (ii) Pr[X = a] = α and Pr[X =
b] = 1 − α, and (iii) |a − b| ≤ k. Then we have the following.
• Var[X] ≤ α(1 − α)k 2 ≤ αk 2 .
• |b| ≤ αk.
31
• |a| ≤ (1 − α)k ≤ k.
The main challenge of proving Lemma 5 is to handle a large number of correlated random
variables. This means that Chernoff bound is not applicable. The advantage of using Theorem 14
is that it allows us to bound the “influence” of each variable individually in terms of variance.
Notice that there are situations where maxXi Di2 is much larger than maxXi−1 Var[Di |Xi−1 ], where
Di = E[f |Xi ] − E[f |Xi−1 ] is defined in Theorem 14. In these situations, Theorem 14 yields a better
concentration bound than the one from Azuma’s inequality. Throughout this section, we use the
following notation. For each edge e and each color c, define ze,c as the indicator random variable
that e successfully colors itself c, thus ze,c = 0 if c ∈
/ Ψ(e).
E.1
Concentration of Vertex Degree
Let v • be a vertex. We claim that E[|S (v • )|] ≤ d . An edge e successfully colors itself with
probability (1 − 1/p)2(t−1) , since there are 2(t − 1) edges competing with e for Color? (e), and each
of these 2(t − 1) edges selects Color? (e) with probability 1/p. Thus, by linearity of expectation,
E[|S (v • )|] = (1 − (1 − 1/p)2(t−1) )|S(v • )| ≤ (1 − (1 − 1/p)2(t−1) )d = d .
def
def
def
For brevity, we write S = S(v • ), S = S (v • ),and z = |S| − |S |. The goal of this section is
2
to show that Pr[z < E[z] − s] < exp −Ω(s
the desired concentration bound
/|S|) , which implies
•
2
Pr [|S (v )| > (1 + δ)d ] < exp −Ω(δ d) , by setting s = δd .
def P
def P
Notations. We write ze =
c∈Ψ(e) ze,c and zc =
e∈S ze,c . In other words, ze is the indicator
random variable that e successfully colors itself; zc is the indicator random variable that some edge
in S successfully colors itself by c.
Let S 0 denote the set of edges such that e0 ∈ S 0 if there exists e = {v • , u} ∈ S such that (i)
Ψ(e) ∩ Ψ(e0 ) 6= ∅, and (ii) e0 is incident to e on u. For each edge e0 ∈ S 0 and for each each color
c ∈ Ψ(e0 ), we define R(e0 , c) as the subset of S such thatP
e ∈ R(e0 , c) if (i) e is incident to e0 , and
0
0
0
(ii) c ∈ Ψ(e). We write w(e , c) = |R(e , c)| and w(e ) = c∈Ψ(e0 ) w(e0 , c). Observe that the value
w(e0 , c) can exceed 2 when e0 ∈
/ S is an imaginary
edge incident to v • . Intuitively, w(e0 ) measures
P
?
the influence of Color (e) on z. Notice that e0 ∈S 0 w(e0 ) ≤ |S|pt.
We consider the sequence of random variables (X1 , . . . , X|S|+|S 0 | ), where the initial |S 0 | variables
are the colors selected by the edges in S 0 , in arbitrary order, and the remaining |S| variables are the
colors selected by the edges in S, in arbitrary order. We let z = f (X1 , . . . , X|S|+|S 0 | ) in Theorem 14.
To prove the
concentration bound, it suffices to show that we can set M = O(1) and σi2
Pndesired
2
to achieve i=1 σi = O(|S|). In what follows, we analyze the effect of exposing the value of the
random variable Xi , given that all random variables in Xi−1 have been fixed.
Exposing an Edge in S 0 . Consider the case where Xi = Color? (e? ) is the color selected by an
edge e? ∈ S 0 . Our goal is to show that
PVar[Di |X2i−1 ] = O(w(e)/(pt)) and |Di | = O(1). Hence we
set σi2 = O(w(e)/(pt)), which implies 1≤i≤|S
P 0 | σi = O(|S|).
By the linearity of expectation,
S Di = c (E[zc |Xi ] − E[zc |Xi−1 ]), where the summation ranges
over all colors c that appear in e∈S Ψ(e). We write Di,c = E[zc |Xi ] − E[zc |Xi−1 ], and make the
following observations:
• Di,c 6= 0 only if c ∈ Ψ(e? ). For each c ∈ Ψ(e? ), Di,c depends only on whether e? selects the
color c, which occurs with probability 1/p. In particular, Di,c < 0 only if e? selects c, and
Di,c > 0 only if e? does not select c. Thus, Cov[Di,c , Di,c0 |Xi−1 ] ≤ 0 for all color pairs {c, c0 }.
32
• For each e ∈ S, both E[ze,c |Xi ] and E[ze,c |Xi−1 ] are within [0, 1/p], since ze,c = 1 only if e
selects c, which occurs with probability 1/p. Thus, maxXi Di,c − minXi Di,c ≤ w(e? , c)/p.
By Lemma 15 (with k ≤ w(e? , c)/p and α = 1/p), we have Var[Di,c |Xi−1 ] ≤ (1/p)(w(e? , c)/p)2 . We
bound the variance Var[Di |Xi−1 ] as follows.
X
X
Var[Di |Xi−1 ] =
Var[Di,c |Xi−1 ] +
Cov[Di,c , Di,c0 |Xi−1 ]
c,c0
c
=
X
O((w(e? , c)/p)2 /p)
Cov[Di,c , Di,c0 |Xi−1 ] ≤ 0
c
=
X
O(w(e? , c)/p2 )
w(e? , c) < t = Θ(p)
c
= O(w(e? )/p2 )
= O(w(e? )/(pt)).
We bound |Di | as follows. Consider c ∈ Ψ(e? ). Recall that we already have the bound |Di,c | ≤
w(e? , c)/p ≤ (t − 1)/p. If c is not selected by e? , which occurs with probability 1 − 1/p, we have
a tighter bound |Di,c | ≤ w(e? , c)/p2 ≤ (t − 1)/p2 by Lemma 15 with k ≤ w(e? , c)/p and α = 1/p.
Therefore,
X
t−1
t−1
|Di | ≤
|Di,c | ≤ 1 ·
+ (p − 1) · 2 = O(1).
p
p
c
Exposing an Edge in S. Consider the case wherePXi = Color? (e? ) is the color selected by the
edge e? ∈ S. Suppose that Xi = c? . Recall Di = c Di,c . It is straightforward to see that (i)
|Di,c | ≤ 1 if c = c? , (ii) |Di,c | ≤ 1/p if c ∈ Ψ(e? ) − {c? }, and (iii)
Thus,
P |Di,c | = 0 otherwise.
2
2
|Di | = O(1), and Var[Di |Xi−1 ] = O(1). We set σi = O(1), and so |S 0 |<i≤|S|+|S 0 | σi = O(|S|).
E.2
Concentration of Palette Size
Let e• = {u, v} be an edge, and let c• = Color? (e• ) be the color selected by e• . We do not consider
c• as a random variable (i.e., we expose the color selected by e• first). Let E be the event that e•
does not successfully color itself. Since e• remains uncolored with at least a constant probability,
we are allowed to ignore the condition “e• remains uncolored” in Lemma 5 in the subsequent calculation. To prove the desired concentration bound regarding palette size, it suffices to show that (i)
| E[|Ψ (e• )|] − p | = O(1), and (ii) Pr[|Ψ (e• )| < (1 − δ) E[|Ψ (e• )|]] < exp(−Ω(δ 2 E[|Ψ (e• )|])).
Notations. We write Su (resp., Sv ) to denote the set of edges e incident to e• on u (resp., v)
such that Ψ(e) ∩ Ψ(e• ) − {c• } =
6 ∅. We write S 0 to denote the set of edges such that e0 ∈ S 0 if there
exists e ∈ Su ∪ Sv meeting the following conditions: (i) e0 is incident to e, (ii) e0 ∈
/ Su ∪ Sv ∪ {e• },
0
•
•
•
and (iii) Ψ(e) ∩ Ψ(e ) ∩ Ψ(e ) − {c } =
6 ∅. Notice that Ψ (e ) is determined by the colors selected
0
by the edges in Su ∪ Sv ∪ S . We have |Su | ≤ (p − 1)(t − 1) < pt, |Sv | ≤ (p − 1)(t − 1) < pt, and
|S 0 | ≤ 2(p − 1)(t − 1)2 < 2pt2 .
Expected Value. In what follows, consider a color c ∈ Ψ(e• ) − {c• }.
• Let e ∈ Su ∪ Sv such that c ∈ Ψ(e). We have E[ze,c ] = p1 (1 − p1 )2t−3 . Notice that e• selects
c• 6= c, so there are 2t − 3 (instead of 2t − 2) edges competing with e for the color c.
33
def
• Let e0 = {u, x} ∈ Su and e00 = {v, y} ∈ Sv such that c ∈ Ψ(e0 ) ∩ Ψ(e00 ). We define ze0 ,e00 ,c =
0 00
ze0 ,c · ze00 ,c . If x = y, then ze0 ,e00 ,c = 0. Otherwise, E[ze0 ,e00 ,c ] = p12 (1 − p1 )4t−6−b(e ,e ) , where
b(e0 , e00 ) ≤ 3 is the number of edges e such that (i) e 6= e• , and (ii) e is incident to both e0 and
e00 .
Let zc be the indicator random variable that some edge incident to e• successfully colors itself by
c, that is,
X
X
def
ze0 ,e00 ,c .
ze,c −
zc =
e0 ,e00 : e0 ∈Su , e00 ∈Sv , c∈Ψ(e0 )∩Ψ(e00 )
e : e∈Su ∪Sv , c∈Ψ(e)
The number of edges e ∈ Su ∪ Sv such that c ∈ Ψ(e) is exactly 2t − 2. The number of pairs
(e0 = {u, x} ∈ Su , e00 = {v, y} ∈ Sv ) such that c ∈ Ψ(e0 )∩Ψ(e00 ) and x 6= y is at least (t−1)2 −(t−1)
and at most (t − 1)2 . By linearity of expectation (recall t = Θ(p)),
E[zc ] =
def
Define z =
P
c∈Ψ(e• )−{c• } zc .
2t
t2
(1 − 1/p)2t − 2 (1 − 1/p)4t ± O(1/p).
p
p
Then, we have:
E[|Ψ (e• )|] = |Ψ(e• )| − E[z]
2t
= p · 1 − (1 − 1/p)2t +
p
2t
= p · 1 − (1 − 1/p)2t +
p
= p ± O(1).
|Ψ (e• )| = |Ψ(e• )| − z
t2
(1 − 1/p)4t ± O(1/p)
p2
t2
4t
(1 − 1/p)
± O(1)
p2
Definition of p
Hence | E[|Ψ (e• )|] − p | = O(1).
Concentration Bound. Consider the sequence of random variables (X1 , . . . , X|Su |+|Sv |+|S 0 | ),
where the initial |S 0 | variables are the colors selected by the edges in S 0 , in arbitrary order, and the
remaining |Su | + |Sv | variables are the colors selected by the edges in Su ∪ Sv , in arbitrary order. Let
z = f (X1 , . . . , X|Su |+|Sv |+|S 0 | ) in Theorem 14. To prove the desired concentration bound, it suffices
P|Su |+|Sv |+|S 0 | 2
to show that we can set M = O(1) and σi2 such that i=1
σi = O(p).
Exposing an Edge in S 0 . Consider the case where Xi = Color? (e? ) is the color selected by
the edge e? ∈ S 0 . The goal is to show that |Di | = O(1/t). This implies (i) |Di | = O(1) and
(ii) Var[Di |Xi−1 ] = O(1/t2 ), and so we may set σi2 = O(1/t2 ). Since |S 0 | = O(pt2 ), we have
P|S 0 | 2
i=1 σi = O(p).
Let R denote the set of edges in Su ∪ Sv that are incident to e? . Notice that 1 ≤ |R| ≤ 2. We
define:
X
X
def
zc(i) =
ze0 ,c −
ze0 ,e00 ,c .
e0 : e0 ∈R, c∈Ψ(e0 )
e0 ,e00 : e0 ∈Su , e00 ∈Sv , c∈Ψ(e0 )∩Ψ(e00 ), {e,e00 }∩R6=∅
(i)
Intuitively, zc is the result of subtracting all terms from the definition of zc not involving edges
(i)
(i)
in R. We now argue that E[zc |Xi ] − E[zc |Xi−1 ] = E[zc |Xi ] − E[zc |Xi−1 ]. This is due to the
two observations: (i) If e ∈
/ R, then E[ze,c |Xi ] = E[ze,c |Xi−1 ]. (ii) If {e0 , e00 } ∩ R = ∅, then
E[ze0 ,e00 ,c |Xi ] = E[ze0 ,e00 ,c |Xi−1 ].
34
Consider a color c ∈ Ψ(e? ) ∩ Ψ(e• ) − {c• }. The probability that some edge in R selects c is
(i)
(i)
at most 2/p, so the conditional expectations E[zc |Xi ] and E[zc |Xi−1 ] must be within [0, 2/p].
(i)
(i)
Therefore, | E[zc |Xi ] − E[zc |Xi−1 ]| ≤ 2/p. For the case of c 6= Xi , which occurs with probability
(i)
(i)
1 − 1/p, we have a tighter bound | E[zc |Xi ] − E[zc |Xi−1 ]| ≤ 2/p2 by Lemma 15 with k ≤ 2/p and
α = 1/p. Thus, we can bound |Di | as follows.
X
| E[zc |Xi ] − E[zc |Xi−1 ]|
|Di | ≤
c∈Ψ(e• )−{c• }
X
=
| E[zc(i) |Xi ] − E[zc(i) |Xi−1 ]|
c∈Ψ(e? )∩Ψ(e• )−{c• }
≤ (2/p) + (2/p2 )(|Ψ(e? ) ∩ Ψ(e• ) − {c• }| − 1)
= O(1/p) = O(1/t).
Exposing an Edge in Su ∪ Sv . Consider the case where Xi = Color? (e? ) is the color selected
def
by the edge e? ∈ Su ∪ Sv . We define w(e? ) = |Ψ(e? ) ∩ Ψ(e• ) − {c• }|. The goal is to show that (i)
|Di | = O(1) and (ii) Var[Di |Xi−1 ] = O(w(e? )/p). By setting σi2 = O(w(e? )/p), we achieve
|S 0 |+|Su |+|Sv |
X
i=|S 0 |+1
σi2 =
X
O(w(e)/p) = O(pt/p) = O(t) = O(p).
e∈Su ∪Sv
P
By the linearity of expectation, Di = c∈Ψ(e? )∩Ψ(e• )−{c• } Di,c , where Di,c = E[zc |Xi ]−E[zc |Xi−1 ].
Since both E[zc |Xi ] and E[zc |Xi−1 ] are within [0, 1], we have |Di,c | ≤ 1. We have a tighter bound
|Di,c | ≤ 1/p in the event that Color? (e? ) 6= c (by Lemma 15 with k ≤ 1 and α = 1/p). Thus,
|Di | ≤ 1 + (w(e? ) − 1)/p = O(1).
In order to prove that Var[Di |Xi−1 ] = O(w(e? )/p), we need the following two observations.
• Consider a color c ∈ Ψ(e? )∩Ψ(e• )−{c• }. Recall that |Di,c | ≤ 1/p for the case c is not selected
by e? , which occurs with probability 1 − 1/p. Thus, E[Di,c · Di,c |Xi−1 ] ≤ (1/p) · 1 + (1 − 1/p) ·
1/p2 = O(1/p).
• Consider two distinct colors c and c0 in Ψ(e? ) ∩ Ψ(e• ) − {c• }. If e? selects c or c0 (which occurs
with probability 2/p), Di,c · Di,c0 ≤ 1 · (1/p). Otherwise Di,c · Di,c0 ≤ (1/p) · (1/p). Therefore,
E[Di,c · Di,c0 |Xi−1 ] ≤ (2/p) · 1/p + (1 − 2/p) · 1/p2 = O(1/p2 ).
We now bound Var[Di |Xi−1 ] as follows.
X
Var[Di |Xi−1 ] ≤
X
E[Di,c · Di,c0 |Xi−1 ]
c∈Ψ(e? )∩Ψ(e• )−{c• } c0 ∈Ψ(e? )∩Ψ(e• )−{c• }
≤ w(e? ) · O(1/p) + w(e? )(w(e? ) − 1) · O(1/p2 )
= O(w(e? )/p).
E.3
Concentration of Color Degree
Let v • be a vertex and let c• be a color that appears in Ψ(e) for some e incident to v. For convenience,
def
we write R = Nc• (v • ). Define R as the subset of R such that e = {v • , u} ∈ R if (i) e is not
successfully colored by a color in Ψ(e) − {c• }, and (ii) no edge incident to e on u successfully colors
def
itself c• . We write z = |R \ R |. Let E 0 be the event that Nc• (v • ) 6= ∅. Observe that if E 0 occurs,
35
then no edge incident to v • successfully colors itself c• . Thus, conditioning on E 0 happening, R \ R
equals Nc• (v • ).
Our goal is to show that (i) Pr[z < E[z] − s] < exp(−Ω(s2 /t)), and (ii) E[|R |] = |R| − E[z] =
t ±O(1). Since E 0 occurs with constant probability, the above (i) and (ii) together imply the desired
concentration bound Pr[|Nc• (v • )| > (1 + δ)t | E 0 ] < exp(−Ω(δ 2 t)), by setting s = δt .
Expected Value. With respect to an edge e = {v • , u} ∈ R, we define the following notations
based on the definition of R .
• Define zea as the indicator random variable that some edge incident to e on u successfully
colors itself c• . We have E[zea ] = (t − 1) · p1 (1 − p1 )2t−2 = pt (1 − p1 )2t ± O(1/p).
• Define zeb as the indicator random variable that e is successfully colored by a color in Ψ(e) −
{c• }. We have E[zeb ] = (p − 1) · p1 (1 − p1 )2t−2 = (1 − p1 )2t ± O(1/p).
def
Let zea,b = zea · zeb . Notice that zea and zeb are nearly independent but not independent. Let
P
def
ze = zea + zeb − zea,b , and thus z = |R \ R | = e∈R ze . We calculate E[zea,b ] as follows. Let e0 be
any edge incident to e such that c• ∈ Ψ(e0 ), and let c be any color in Ψ(e) − {c• }. With respect to
(e, e0 , c), we define the following two sets:
• Sa is the set of all edges e00 such that e00 ∈ Sa if (i) e00 6= e, e0 , (ii) e00 is incident to e0 , and (iii)
c• ∈ Ψ(e00 ). Intuitively, Sa is the set of all edges other than e that contend with e0 for the
color c• . Notice that |Sa | = 2t − 3 (since Ψ(e) must contain c• ).
• Sb is the set of all edges e00 such that e00 ∈ Sb if (i) e00 6= e, e0 , (ii) e00 is incident to e, and (iii)
c ∈ Ψ(e00 ). Intuitively, Sb is the set of all edges other than e0 that contend with e for the color
c. Notice that 2t − 3 ≤ |Sb | ≤ 2t − 2 (since Ψ(e0 ) may or may not contain c).
Let x(c, e0 ) denote the probability that (i) e0 successfully colors itself c• and (ii) e successfully colors
itself c. In view of the definition of Sa and Sb , we have:
x(c, e0 ) =
1
p2
Y
(1 − 1/p)
e00 ∈Sa \Sb
Y
(1 − 1/p)
Y
(1 − 2/p)
e00 ∈Sa ∩Sb
e00 ∈Sb \Sa
1
(1 − 1/p)|Sa \Sb | (1 − 1/p)|Sb \Sa | (1 − 2/p)|Sa ∩Sb |
p2
1
|Sa ∩ Sb |
|Sa \Sb |
|Sb \Sa |
2|Sa ∩Sb |
= 2 (1 − 1/p)
(1 − 1/p)
(1 − 1/p)
1−O
p
p2
=
(Notice that |Sa ∩ Sb | < t = Θ(p).)
1
(1 − 1/p)|Sa |+|Sb | (1 − O(1/p))
p2
1
= 2 (1 − 1/p)4t−O(1) (1 − O(1/p))
p
1
= 2 (1 − 1/p)4t − O(1/p3 ).
p
=
36
We now calculate E[zea,b ] and show that E[|R |] = |R| − E[z] = t ± O(1).
X
E[zea,b ] =
x(c, e0 )
(c,e0 ) : e0 incident to e, c• ∈ Ψ(e0 ), c ∈ Ψ(e) − {c• }
= (t − 1)(p − 1) ·
=
1
(1 − 1/p)4t − O(1/p3 )
p2
t
(1 − 1/p)4t − O(1/p).
p
|R | = |R| − E[z]
X
=t−
E[zea ] + E[zeb ] − E[zea,b ]
e∈R
=t· 1−
=t· 1−
t
(1 − 1/p)2t − (1 − 1/p)2t +
p
t
(1 − 1/p)2t − (1 − 1/p)2t +
p
= t − O(1).
t
(1 − 1/p)4t − O(1/p)
p
t
4t
(1 − 1/p)
− O(1)
p
Definition of t
Concentration Bound. Based on the definition of zea and zeb , we define the following sets.
• Let R1 be the set of all edges e such that (i) e ∈
/ R, (ii) c• ∈ Ψ(e), and (iii) e is incident
to some edge in R. Similarly, let R2 be the set of all edges e such that (i) e ∈
/ R ∪ R1 , (ii)
c• ∈ Ψ(e), and (iii) e is incident to some edge in R1 . Notice that the value zea , for any e ∈ R,
is determined by the information regarding which edges in R ∪ R1 ∪ R2 select c• . We write
α = |R ∪ R1 ∪ R2 |.
• Let R0 be the set of all edges e0 such that (i) e0 ∈
/ R and (ii) there exists e ∈ R such that
0
b
Ψ(e) ∩ Ψ(e ) 6= ∅. Notice that the the value ze , for any e ∈ R, is determined by the colors
selected by the edges in R ∪ R0 . We write β = |R ∪ R0 |.
For each e ∈ R, zea is simply the summation of ze0 ,c• over all edges e0 ∈ R1 incident to e. For
each e00 ∈ R2 , we write w(e00 ) to denote the
of edges in R1 incident to e00 . Intuitively, w(e00 )
P number
? 00
a
measures the influence of Color (e ) on e∈R ze .
We consider the sequence of random variables (X1 , . . . , Xα+β ), where the initial α random
variables reveal which edges in R ∪ R1 ∪ R2 select the color c• according to the ordering R2 , R1 , R,
and the remaining β random variables reveal the colors selected by the edges in R ∪ R0 according to
the ordering R0 , R. We let z = f (X1 , . . . , Xα+β ) in Theorem 14. To prove the desired concentration
P
2
bound, it suffices to show that we can set M = O(1) and σi2 such that α+β
i=1 σi = O(t). In what
follows, we analyze the effect of exposing the value of Xi , given variables Xi−1 have been fixed.
Revealing whether c• is Selected by an Edge in R ∪ R1 ∪ R2 . Consider the case where Xi
reveals whether c• is selected by an edge in e? ∈ R ∪ R1 ∪ R2 (i.e., 1 ≤ i ≤ α). There are at most
two distinct outcomes of Di |Xi−1 , in which one occurs with probability 1/p. Thus, we have:
2
Var[Di |Xi−1 ] ≤ max Di |Xi−1 − min Di |Xi−1 /p = O(max |Di |2 /p).
Xi
Thus, to achieve
Pα
2
i=1 σi
Xi
= O(t) and M = O(1), we only need:
37
Xi
• For e? ∈ R2 , |Di | = O(w(e? )/p). Hence we can set σi2 = O(w(e? )/p2 ), as Var[Di |Xi−1 ] =
O((w(e? )/p)2 /p) = O(w(e? )/p2 ).
• For e? ∈ R ∪ R1 , |Di | = O(1). Hence we can set σi2 = O(1/p), as Var[Di |Xi−1 ] = O(1/p).
P
P
Notice that e∈R2 w(e) < t3 , |R1 | < t2 , and |R| = t. Thus, αi=1 σi2 = O(t). With respect to the
edge e? ∈ R ∪ R1 ∪ R2 , we make the following definitions.
def
def
Y a = {e0 ∈ R1 : e0 = e? or e0 is incident to e? } Dia =
def
def
Y b = {e ∈ R : e = e? or e is incident to e? }
Dib =
P
e0 ∈Y a
P
e∈Y b
E[ze0 ,c• |Xi ] + E[ze0 ,c• |Xi−1 ]
| E[zeb |Xi ] − E[zeb |Xi−1 ]|
Intuitively, Y a and Y b are the subsets of R1 and R that are “relevant” to Di in the following sense:
E[ze00 ,c• |Xi ] = E[ze00 ,c• |Xi−1 ] for all e00 ∈ R1 \ Y a ; and E[zeb0 |Xi ] = E[zeb0 |Xi−1 ] for all e0 ∈ R \ Y b .
Our plan of bounding |Di | is as follows. First we show that |Di | ≤ 4Dia + Dib in Claim 1, and then
we bound Dia and Dib separately in Claim 2 and Claim 3. The three claims together establish a
desired bound on |Di |.
Claim 1. |Di | ≤ 4Dia + Dib .
Proof. We define the following notations.
def
P1 = {(e, e0 ) : e ∈ R \ Y b , e0 ∈ Y a , e is incident to e0 }
def
P2 = {(e, e0 ) : e ∈ Y b , e0 ∈ R1 \ Y a , e is incident to e0 }
def
P3 = {(e, e0 ) : e ∈ Y b , e0 ∈ Y a , e is incident to e0 }
P
def
Qj = − (e,e0 )∈Pj E[ze0 ,c• · zeb |Xi ] − E[ze0 ,c• · zeb |Xi−1 ]
def P
Fj = e∈R E[zej |Xi ] − E[zej |Xi−1 ]
(for each j = 1, 2, 3)
(for each j = a, b)
The definitions of P1 , P2 , and P3 depend on Y a and Y b , which depend on the edge e? . For instance,
if e? ∈ R, then Y b = R, which implies that P1 = ∅. Recall that the edge e? can be any edge in
R ∪ R1 ∪ R2 , and the proof of this claim applies to all choices of e? ∈ R ∪ R1 ∪ R2 .
Notice that for any pair (e ∈ R, e0 ∈ R1 ) such that e is incident to e0 but (e, e0 ) ∈
/ P1 ∪ P2 ∪ P3 ,
we must have E[ze0 ,c• · zeb |Xi ] = E[ze0 ,c• · zeb |Xi−1 ] due to the definition of Y a and Y b . We rewrite
the term Di as follows.
Di = E[z|Xi ] − E[z|Xi−1 ]
X
=
(E[ze |Xi ] − E[ze |Xi−1 ])
e∈R
=
X
(E[zea |Xi ] − E[zea |Xi−1 ]) + E[zeb |Xi ] − E[zeb |Xi−1 ] − E[zea · zeb |Xi ] − E[zea · zeb |Xi−1 ]
e∈R
(zea is the summation of ze0 ,c• over all edges e0 ∈ R1 incident to e.)
X
= Fa + Fb −
E[ze0 ,c• · zeb |Xi ] − E[ze0 ,c• · zeb |Xi−1 ]
(e,e0 ) : e∈R, e0 ∈R1 , e0 incident to e
(Any pair (e, e0 ) ∈
/ P1 ∪ P2 ∪ P3 contributes 0 in the summation.)
= Fa + Fb + Q1 + Q2 + Q3 .
38
To prove this claim it suffices to show that (i)|Fa + Q1 | ≤ 2Dia , (ii) |Fb + Q2 | ≤ Dib , and |Q3 | ≤ 2Dia .
(Expand Fa using the fact that zea is the summation of ze0 ,c• over all edges e0 ∈ R1 incident to e.)
X
|Fa + Q1 | ≤ Q1 +
E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
(e,e0 ) : e∈R, e0 ∈R1 , e0 incident to e
(Any pair (e, e0 ) ∈
/ P1 ∪ P3 contributes 0 in the summation.)
X
≤ Q1 +
E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
(e,e0 )∈P1 ∪P3
X
≤
E[ze0 ,c• (1 − zeb )|Xi ] − E[ze0 ,c• (1 − zeb )|Xi−1 ]
(e,e0 )∈P1
+
X
E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
(e,e0 )∈P3
/ R \ Y b .)
(E[zeb |Xi−1 ] = E[zeb |Xi ] when e ∈
X
≤
(1 − E[zeb |Xi−1 ]) E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
(e,e0 )∈P1
+
X
E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
(e,e0 )∈P
3
(0 ≤ 1 − E[zeb |Xi−1 ] ≤ 1)
X
≤
E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
(e,e0 )∈P1 ∪P3
(Any edge e0 ∈ R1 is incident to at most 2 edges in R.)
X
≤2
E[ze0 ,c• |Xi ] − E[ze0 ,c• |Xi−1 ]
≤
e0 ∈Y a
2Dia .
0
a
For each e ∈ Y b , we
P write B(e) to denote the
P set of all edges e ∈ R1 \ Y that• are incident to e.
Notice that 0 ≤ E[ e0 ∈B(e) ze0 ,c• |Xi−1 ] = E[ e0 ∈B(e0 ) ze0 ,c• |Xi ] ≤ 1, since e = {v , u} and all edges
in B(e) share the vertex u.
|Fb + Q2 | ≤ Q2 +
X
E[zeb |Xi ] − E[zeb |Xi−1 ]
e∈Y b
(By definition of B(e) and Q2 .)
X
X
X
≤
E zeb 1 −
ze0 ,c• Xi − E zeb 1 −
ze0 ,c• Xi−1
e∈Y b
e0 ∈B(e)
e0 ∈B(e)
39
(For e0 ∈ R1 \ Y a , we have E ze0 ,c• |Xi = E ze0 ,c• |Xi−1 .)
X
X
1 − E
ze0 ,c• Xi−1 E[zeb |Xi ] − E[zeb |Xi−1 ]
≤
e0 ∈B(e)
e∈Y b
≤
=
|Q3 | ≤
X
E[zeb |Xi ] − E[zeb |Xi−1 ]
e∈Y b
Dib .
X
E[z
e0 ,c•
·
zeb |Xi ]
+ E[z
e0 ,c•
·
zeb |Xi−1 ]
(e,e0 )∈P3
≤
X
E[ze0 ,c• |Xi ] + E[ze0 ,c• |Xi−1 ]
(e,e0 )∈P3
(Any edge e0 ∈ R1 is incident to at most 2 edges in R.)
X
≤2
E[ze0 ,c• |Xi ] + E[ze0 ,c• |Xi−1 ]
≤
e0 ∈Y a
2Dia .
Claim 2. If e? ∈ R2 , then Dia = O(w(e? )/p). If e? ∈ R ∪ R1 , then Dia = O(1).
Proof. We first consider the case that e? ∈ R2 . In this case |Y a | = w(e? ). Recall that Y a ⊆ R1 ,
and so all e ∈ Y a have not decided whether to select c• by the time Xi is revealed. Therefore,
both E[ze,c• |Xi ] and E[ze,c• |Xi−1 ] are within the range [0, 1/p], and so Dia = O(w(e? )/p). Next,
consider the case that e? ∈ R ∪ R1 . In this case all edges in Y a must have a common vertex
?
a
a
with
P two edges in Y can successfully color themselves by c. Hence Di ≤
P e , and so at most
e∈Y a E[ze,c• |Xi−1 ] ≤ 2 + 2 = 4.
e∈Y a E[ze,c• |Xi ] +
Claim 3. If e? ∈ R1 ∪ R2 , then Dib = O(1/p). If e? ∈ R, then Dib = O(1).
P
Proof. Recall that zeb = c∈Ψ(e)−{c• } ze,c for any edge e ∈ Y b , and so
Dib ≤
X
e∈Y
b
X
| E[ze,c |Xi ] − E[ze,c |Xi−1 ]|.
c∈Ψ(e)−{c• }
We first show that | E[ze,c |Xi ] − E[ze,c |Xi−1 ]| = O(1/p2 ) if e? 6= e. We write k1 (resp., k2 ) to denote
the number of edges incident to e that have decided to select c• (resp., have decided to not select
c• ) by the time Xi is revealed.
(e has decided to select c• )
0
1
E[ze,c |Xi−1 ] = p−1
· (1 − 1/p)2t−1−k1 −k2 · (1 − 1/(p − 1))k2 (e has decided to not select c• )
1
2t−1−k1 −k2 · (1 − 1/(p − 1))k2
(e has not made any decision)
p · (1 − 1/p)
In any case, E[ze,c |Xi−1 ] = O(1/p). There are two possibilities of E[ze,c |Xi ] based on Xi , i.e.,
whether e? selects c• .
(
E[ze,c |Xi−1 ]/(1 − 1/p)
(e? selects c• )
E[ze,c |Xi ] =
E[ze,c |Xi−1 ] · (1 − 1/(p − 1))/(1 − 1/p) (e? does not select c• )
40
In any case, | E[ze,c |Xi ] − E[ze,c |Xi−1 ]| = O(1/p2 ). We are now in a position to bound Dib . For the
case that e? ∈ R1 ∪R2 , we have |Y b | ≤ 2 and e? ∈
/ Y b , and so Dib ≤ 2·(p−1)·O(1/p2 ) = O(1/p). For
?
b
the case that e ∈ R, we have |Y | = |R| = t and e? ∈ Y b , and so Dib ≤ 1+(t−1)·(p−1)·O(1/p2 ) =
O(1).
Revealing the Color Selected by an Edge in R ∪ R0 . Observe that zea , for all e ∈ R, are
determined by {Xj : j ∈ [α]}. If zea = 1, then ze = 1 regardless of the value of zeb ; if zea = 0, then
ze = zeb . For those edges e ∈ R such that ze is not determined by {Xj : j ∈ [α]}, the random
number ze = zeb behaves the same as that in the analysis of concentration of vertex degree, so the
same analysis for the concentration of the vertex degree can be applied here (think of S = R and
S 0 = R0 ).
P
In more detail, for each edge e0 ∈ R0 , we define w0 (e0 ) as e∈R, e0 incident to e |Ψ(e0 ) ∩ Ψ(e) − {c• }|.
NoticePthat w(e0 ) is identical to the definition of w0 (e0 ) except that the color c• is ruled out, and we
have e0 ∈R0 w0 (e0 ) ≤ |R|(p − 1)(t − 1) < pt2 .
Now consider the color Xi = Color? (e? ) selected by the edge e? ∈ R ∪ R0 . The same analysis for
the concentration of the vertex degree implies the following.
• If e? ∈ R0 , then |Di | = O(1) and Var[Di |Xi−1 ] = O(w0 (e? )/(pt)). Hence we set σi2 =
O(w0 (e? )/(pt)).
• If e? ∈ R, then |Di | = O(1) and Var[Di |Xi−1 ] = O(1). Hence we set σi2 = O(1).
P
2
Thus, α+β
j=α+1 σi = O(t), as desired.
41
| 8cs.DS
|
Realizing Half-Diminished Reality from Video
Stream of Manipulating Objects
Hayato Okumoto, Mitsuo Yoshida and Kyoji Umemura
arXiv:1709.08340v1 [cs.CV] 25 Sep 2017
Department of Computer Science and Engineering
Toyohashi University of Technology
Toyohashi, Aichi, Japan
[email protected], [email protected], [email protected]
Abstract—When we watch a video, in which human hands
manipulate objects, these hands may obscure some parts of
those objects. We are willing to make clear how the objects are
manipulated by making the image of hands semi-transparent,
and showing the complete images of the hands and the object.
By carefully choosing a Half-Diminished Reality method, this
paper proposes a method that can process the video in real time
and verifies that the proposed method works well.
Keywords—Video Enhancement; Virtual Reality; Diminished
Reality
I. I NTRODUCTION
When we watch how objects are manipulated using surveillance cameras, hands usually conceal or obscure some parts of
those manipulated objects. Realizing Half-diminished reality
in this paper means to generate a video where both the hands
and all parts of the objects are shown at the same time.
An example of this situation is a broadcast of a Japanese
chess match, where players manipulate pieces and that an
entire understanding of where those pieces are situated is
always important in order to understand the current move.
In spite of this importance, at the timing of when the move
is decided, the corresponding piece is hidden by the player’s
hand. This situation always happens for a video showing
when objects are manipulated by hand. This problem, in the
case of a chess game, is caused by a situation where the
hands (foreground image) conceal both the piece and the
board (background image). One notable characteristic of the
situation is the existing interaction between the foreground
image and the background image. For example, the piece in
the hand belongs to the foreground image where it belonged
to background image previously but will soon belong to the
background image once again. The timing of this transition is
also important information of the interaction.
In this paper, we will propose an image processing method
that satisfies the following conditions. (A) The background
image has no concealed areas. (B) Both the background image
and foreground image are integrated into a single image. (C)
The movements of the background image and foreground
image are synchronized. (D) The processing should be in real
time so that it can be used for live broadcast. (E) The input
of processing is a single video from a fixed camera.
978–1–5090–1636–5/16/$31.00 c 2016 IEEE
This kind of image processing is known as Diminished
Reality (DR), which is a part of the Virtual Reality (VR)
research field. DR is to show us a virtual space, where
some objects in the real world have either vanished or are
transparent. Though many researches in DR [1], [2], [3] make
objects completely transparent, some researches [4], [5] focus
on semi-transparency, which is called Half-Diminished Reality
(H-DR). DR focuses on obtaining a clear and sound image of
the background; whereas H-DR regards both the foreground
and the background images as important. Buchmann et al. [4]
proposed to show a H-DR image through a head mounted
display to improve the efficiency of manual operation, and
discussed the level of transparency and the efficiency of the
operation.
He and Zhang [6] proposed a notable DR method which
satisfies condition D and E. He and Zhang obtained the
complete contents of a whiteboard (Background image) in a
video conference system by using a single camera. It obtained
the background image by substituting the background image
with only the stable part of the current image. Since hands
(foreground image) are not usually stable, only the written
contents on the whiteboard appear in the output image. If
we can obtain a complete background image, we can obtain
the H-DR image that satisfies conditions A and B (but not
C) by superimposing the background image and the current
image. Since it requires considerable time to judge whether a
part is stable or not, the update of the background image is
delayed from the change of the current image. As a result, a
direct application of He and Zhang’s proposal cannot satisfy
condition C. Condition C is very challenging since at the
timing of transition of the object image from the foreground
image to the background image, the hands are usually holding
the objects and a true background image may not be available
at this time.
In this paper, we focus our attention to the application of
broadcasting. With broadcasting, though it requires processing
the image at the same rate of the current image, it is acceptable
if there is a delay of several seconds from input to output. This
presents an idea where we can adjust the timing of the current
image so that it can be synchronized with the background
image.
The contribution of these works is as follows. (1) We
define the situation of H-DR where interactions between the
foreground and background image exist. (2) By distinguishing
the delay from real time, we show that a time shift approach is
usable for real time and a synchronized setting. (3) We explain
a valid situation where the proposal is effective.
II. R ELATED W ORK
DR processing usually consists of two steps. The first step
is to distinguish the foreground image from the background
image. The second step is to complement that part of the
background image where the foreground object conceals or
obscures the background image.
A. Detecting Foreground Image
One of the well-used methods to detect the foreground
image is background subtraction. The simplest method of
background subtraction method is as follows. First, we capture
the background image without a foreground object. Then, we
compare the color and brightness of the current image and the
captured background. If the difference is greater than some set
threshold, then we can regard the area as a part of foreground
image. Wren et al. [7] enhance this method. Instead of using
one static background image, it uses the average of several
images as the background. This method obtains the foreground
object in real time, but the result is particularly sensitive
to shadows and changes in lighting. Instead, using a single
background image, the statistical model for a background
image is commonly used [8]. Some works use GaussianMixture [9], [10], [11], and other works use a non-parametric
estimation using kernel [12], [13], [14]. There is also a work
that uses Eigen vectors from a series of images to model the
background [15]. However, all of these works have a problem
when the foreground and background are interacting. There is
another work that does not construct the color model, but uses
an RGB-Depth camera in order to detect the foreground [5].
This approach requires a special camera, and is not suitable
for the application of Japanese chess where the foreground
object and the background object may be the same position.
He and Zhang [6] use the difference between the current
image and the previous image but not the fixed background
images. For a certain time, if there is no significant difference,
the part of the image is regarded as stable. If the image is
changing during a certain period, the corresponding part of
the image is treated as a foreground region. The key concept
of this approach is that the foreground image is unstable.
Please note that an unstable region may not be a foreground
region, but that a stable region is always a background region.
In this approach, the background region, which is uncovered
by a foreground object, is also treated the same way as
the moving foreground region. In this sense, this approach
does not correctly detect a particular foreground region, but
successfully detects the region that contains all of foreground
regions.
B. Complementing Background Image
One of the natural ways to complement a background image
is to use multiple cameras with different angles towards an
object. Using the transformation of an image, it recovers the
hidden part from other videos taken from different angles [16],
[17], [3]. Apparently though, it has a problem when hands
(foreground objects) hold the objects (background objects),
since there is no angle at which we can see the objects.
Another approach is to complement the hidden background
in the same context as a neighboring background region. This
approach assumes that the background is a wall or some
other uniform texture. However, this is not a case where the
background is a collection of objects, for example, a board
and pieces of Japanese chess.
He and Zhang [6] use a time shift approach. Since only a
single camera is used in this approach, the information complementing the background image comes from the previous
images. In this approach, the region that is regarded as the
foreground part is not updated. As a result, the past images
are used to complement the region that is currently concealed
by a foreground object. The background image is updated by
the current image in only the part where the image is regarded
as stable. A nice feature of He and Zhang’s approach is the
ability to update the background. Please note that the change of
background is inevitably delayed in order to judge its stability.
Since they aimed to recover the information from a white
board, in other words, they aimed at DR only, they did not
need to consider the interaction between the foreground image
and the background image. Therefore, this delay was not an
issue for them.
III. P ROPOSED M ETHOD
We aimed at an enhanced visual image from a surveillance
camera showing some hands manipulating some objects. In
this case, we are willing to see the complete shape of objects,
and also see how hands manipulate objects. An object belongs
to the foreground image when it is held by hands. The same
object belongs to a background image when it is not held by
hands. We are willing to show the information about how the
object is placed at the end time of the manipulation.
First, we generate a DR image, where an image of foreground objects that obscure a background image is diminished.
Then, we superimpose the current video image onto the
background image. Then we will get a H-DR image, where
the foreground objects are half-transparent. In this image, we
can see both the hands and the objects. Since some objects
may move between the foreground image and the background
image, it is a requirement that both images are synchronized.
Our approach is to adjust the foreground image to obtain
the final result. This approach will work if the person who
is manipulating the objects and the person who is watching
the video are independent, which is a typical case for a
surveillance camera.
When the manipulation finishes, the object is held by hand.
Therefore, a time shift approach would be a reasonable choice
Color
Value
Input from Camera
Output of Background
Color
Value
d = Background Estimation Time
Time
Time
t
Fig. 1. A foreground object moves over the background. Since the color
value keeps changing during object movement. If we suppress updating of
the image, the foreground object is erased.
t+d
Fig. 2. An object in the background becomes a part of the foreground, and
then becomes a part of background again. The background image will reflect
the change after a certain duration passes.
to complement the image of the object. We choose the He and
Zhang’s approach [6] to realize DR.
A. Implementation of Diminished Reality
We use the temporal change of color values in pixels. As
is shown in Fig. 1, when a moving foreground object covers
the background image, the color value of the corresponding
image is not stable within a certain time. In this case, we do
not update the pixel by the current image. As is shown in the
red line in Fig. 1. The foreground object is erased from output.
As is shown in Fig. 2, when an object in the background moves
to another location, the first and last color values are different.
While the object is moving, the color value is changing. As
before, the background image is not updated during movement
and the color value before the movement is used for output.
Suppose that the movement has finished at the time, t. The
input color value becomes stable. After some duration, the
system estimates that the movement is finished, and updates
the output as the input value. This estimation can be done
simultaneously for the block of pixels. We choose this block
as 16 × 16 pixels square. We define the length of estimating
duration as background estimation time, d. Please note that
movement in the current image finishes at the time t, while
the output is updated at the time t + d. We need to adjust d as
the method of manipulation. For the case of Japanese chess, we
choose d to be 3 seconds. This d should be adjusted according
to the movement of the foreground. If the foreground is an
industial robot, the d could be less than 1 second.
B. Implementation of Half-Diminished Reality
Though we can see both hands and objects by superimposing the background image and the current image, there is a
problem as is shown Fig. 3. Fig. 3 shows the current image
and the background image that are superimposed at the time t.
Since the update of the background image is delayed, the hand
and the piece are not synchronized. Accordingly, the output is
confusing.
It is possible to delay the current image by time d, using
a ring frame buffer. Then we can get a delayed image as is
Fig. 3. The result of a superimposed image of the current image and the
background image at the end time of a movement. The hand and the piece
are not synchronized.
shown as the blue line in Fig. 4. Fig. 5 shows the delayed
image and the background image that are superimposed at the
time t + d. The hand and the piece are synchronized. We can
see the character on the piece at the time t + d, which may
be informative for a video audience.
IV. V ERIFICATION OF THE R EQUIREMENT
We will verify that the proposed method satisfies the conditions that we set at the beginning. Fig. 8 shows the output
sequence of our system.
Fig. 8 (1) shows the initial situation where the hand is not
in the figure. Fig. 8 (2) shows the situation when the hand
covers the board and starts the movement. Fig. 8 (3) shows
the situation in the middle of the movement, where the piece
is shown in its initial position. Fig. 8 (4) shows the situation
at the end of movement. In this figure, the piece in its initial
position is about to disappear, and the piece at its next position
is appearing. Fig. 8 (5) shows the situation where the hand is
not in the figure after movement. Since we can always see
the character on the piece, condition A is satisfied. Since we
Color
Value
Output of Foreground
Output of Background
Time
+d
Fig. 4. Relationship between the delayed image and the background image.
At the end of the movement in the delayed image, the background image
changes.
Fig. 6. Both current image and output (in the display) is shown at the time
when the hand starts moving the piece. The hand is not yet appearing in the
display.
Fig. 5. The result of a superimposed image of the delayed image and the
background image at the time of reflecting change. The hand and piece are
synchronized.
Fig. 7. After 3 seconds from Fig. 6. The hand appears in the display. It
shows that the output is delayed but keeping up with the input.
can also see the hands and piece by the superimposed image,
condition B is satisfied. As in Fig. 8 (4), the movement of the
hand and piece is synchronized and condition C is satisfied.
Fig. 6 and Fig. 7 show both the input and output of our
system. Fig. 6 shows the situation where the hand is about to
move a piece. Fig. 7 is 3 seconds later, where a hand appears
in the display. The input is captured by a web camera. We use
a 3.3GHz Intel Core i3-3220 (2 cores CPU). The resolution
is HD(1280 by 720 pixels). The output goes to network using
Motion JPEG. The load is 26% at the frame rate 9.5 fps, using
a single thread, and the load is 47% at a frame rate of 13 fps
by multithread implementation. This implies that conditions D
and F are satisfied.
V. D ISCUSSIONS
By this proposed method, the output is delayed from the
input by several seconds. This may limit the situation where
the proposed method may be useful. We assume that the person
who manipulates the object is different from the person who
watches the output. If this assumption is not applicable, the
delay might be a problem. Nevertheless, we can assume this
in many cases especially when using a surveillance camera.
The other assumption that we have here is that the background object is always stable and the foreground object is
not stable longer than duration d. There may be some cases
where an object in the background moves spontaneously.
Nevertheless, we can assume this for the Japanese chess
situation. It is also the case for many assembling video.
VI. C ONCLUSION
In this paper, we have shown the implementation of HDR, where there are interactions between a background and a
foreground. We propose to use movement to realize DR, and
then to adjust the current image by superimposing onto video.
It may be valuable that we are able to see the complete object
image at the end time of its manipulation. We have chosen
the broadcast of a Japanese chess match as an example. The
same technique will be applicable in many situations where a
surveillance camera is used.
Fig. 8. A sequence of output video is shown. We can always see the characters on the pieces, and the movement of hands. The pieces change and the hand
movements are synchronized.
ACKNOWLEDGMENT
A part of this research was supported by JSPS KAKENHI
Grant Number 26330396.
R EFERENCES
[1] J. Herling and W. Broll, “PixMix: A real-time approach to high-quality
Diminished Reality,” in 1th IEEE International Symposium on Mixed
and Augmented Reality 2012, Science and Technology Papers, 2012,
pp. 141–150.
[2] V. Lepetit and M.-O. Berger, “A semi-automatic method for resolving
occlusion in augmented reality,” in Proc. of IEEE Conference on
Computer Vision and Pattern Recognition, vol. 2, 2000, pp. 225–230.
[3] S. Zokai, J. Esteve, Y. Genc, and N. Navab, “Multiview paraperspective
projection model for diminished reality,” in Proc. of 2nd IEEE and ACM
International Symposium on Mixed and Augmented Reality, 2003, pp.
217–226.
[4] V. Buchmann, T. Nilsen, and M. Billinghurst, “Interaction With Partially
Transparent Hands And Objects,” in Sixth Australasian User Interface
Conference, vol. 40, 2005, pp. 17–20.
[5] K. Sugimoto, H. Fujii, A. Yamashita, and H. Asama, “Half-diminished
reality image using three RGB-D sensors for remote control robots,”
in 2014 IEEE International Symposium on Safety, Security, and Rescue
Robotics, 2014, pp. 1–6.
[6] L. W. He and Z. Zhang, “Real-time whiteboard capture and processing using a video camera for remote collaboration,” IEEE Trans. on
Multimedia, vol. 9, no. 1, pp. 198–206, 2007.
[7] C. Wren, A. Azarbayejani, T. Darrell, and A. Pentland, “Pfinder: realtime tracking of the human body,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 19, no. 7, pp. 780–785, 1997.
[8] M. Piccardi, “Background subtraction techniques: a review,” in 2004
IEEE International Conference on Systems, Man and Cybernetics, vol. 4,
2004, pp. 3099–3104.
[9] T. S. F. Haines and T. Xiang, “Background Subtraction with Dirichlet
Process Mixture Models,” IEEE Trans. on Pattern Analysis and Machine
Intelligence, vol. 36, pp. 670–683, 2014.
[10] C. Stauffer and W. Grimson, “Adaptive background mixture models for
real-time tracking,” in Proc. of 1999 IEEE Computer Society Conference
on Computer Vision and Pattern Recognition, vol. 2, 1999, pp. 246–252.
[11] Z. Zivkovic, “Improved adaptive Gaussian mixture model for background subtraction,” in Proc. of the 17th International Conference on
Pattern Recognition, 2004., vol. 2, 2004, pp. 28–31.
[12] A. Elgammal, D. Harwood, and L. Davis, “Non-parametric model for
background subtraction,” Computer Vision, vol. 1843, pp. 751–767,
2000.
[13] B. Han, D. Comaniciu, and L. Davis, “Sequential kernel density
approximation through mode propagation: applications to background
modeling,” in Proc. of Asian Conference on Computer Vision, vol. 4,
2004, pp. 818–823.
[14] Y. Liu, H. Yao, W. Gao, X. Chen, and D. Zhao, “Nonparametric
background generation,” Journal of Visual Communication and Image
Representation, vol. 18, no. 3, pp. 253–263, 2007.
[15] N. M. Oliver, B. Rosario, and A. P. Pentland, “A Bayesian Computer
Vision System for Modelling Human Interactive,” IEEE Trans. on
Pattern Analysis and Machine Intelligence, vol. 22, pp. 831–843, 2000.
[16] A. Enomoto and H. Saito, “Diminished Reality using Multiple Handheld
Cameras,” in Proc. of Asian Conference on Computer Vision, vol. 7,
2007.
[17] T. Hashimoto, Y. Uematsu, and H. Saito, “Generation of see-through
baseball movie from multi-camera views,” in 2010 IEEE International
Workshop on Multimedia Signal Processing, 2010, pp. 432–437.
| 1cs.CV
|
Learning crystal plasticity using digital image correlation: Examples from
discrete dislocation dynamics
Stefanos Papanikolaoua,b , Michail Tzimasa , Hengxu Songa,b , Andrew C.E. Reidc , Stephen A. Langerd
a The
West Virginia University, Department of Mechanical & Aerospace Engineering, Morgantown, WV, 26505
Johns Hopking University,Department of Mechanical Engineering,3400 N Charles St, Baltimore, Maryland, 21218
c Material Measurement Laboratory, National Institute of Standards and Technology, Gaithersburg, MD 20899
d Information Technology Laboratory, National Institute of Standards and Technology, Gaithersburg, MD 20899
arXiv:1709.08225v1 [cond-mat.mtrl-sci] 24 Sep 2017
b The
Abstract
Digital image correlation (DIC) is a well-established, non-invasive technique for tracking and quantifying the
deformation of mechanical samples under strain. While it provides an obvious way to observe incremental and
aggregate displacement information, it seems likely that DIC data sets, which after all reflect the spatially-resolved
response of a microstructure to loads, contain much richer information than has generally been extracted from
them. In this paper, we demonstrate a machine-learning approach to quantifying the prior deformation history of
a crystalline sample based on its response to a subsequent DIC test. This prior deformation history is encoded
in the microstructure through the inhomogeneity of the dislocation microstructure, and in the spatial correlations
of the dislocation patterns, which mediate the system’s response to the DIC test load. Our domain consists of
deformed crystalline thin films generated by a discrete dislocation plasticity simulation. We explore the range of
applicability of machine learning (ML) for typical experimental protocols, and as a function of possible size effects
and stochasticity. Plasticity size effects may directly influence the data, rendering unsupervised techniques unable
to distinguish different plasticity regimes.
Keywords: digital image correlation (DIC), dislocation dynamics, plastic deformation, principal component
analysis (PCA), clustering, dislocation patterns
1. Introduction
The prior deformation history and/or material processing of mechanical components may dramatically influence their mechanical properties without nominally altering their chemical composition (Frazier, 2014). Given a
mechanical component design, it is mandatory to identify whether the mechanical properties of the component are
in accord with its nominal properties. In crystals, a major component of property sensitivity is plastic deformation.
How can we inexpensively and swiftly distinguish whether a crystalline sample is pristine or if it has been heavily
deformed in the past? Scanning electron microscopy (SEM), X-ray powder diffraction, and Transmission Electron
Microscopy (TEM), are insightful but relatively cumbersome ways of answering this question. Moreover, they
may be consistently successful for elasticity mapping or for molecular microstructures with the same crystal type
but different interatomic potentials (Kalidindi, 2015) but it is not clear what to expect when plastic inhomogeneity
and strain localization are present, as in ductile metals and alloys. In contrast, nanoindentation or digital image
correlation (DIC) (Anuta, 1970; Keating et al., 1975) are both inexpensive, swift and applicable to a vast range of
material classes. In this paper, we focus on DIC: Is it possible to devise a technique for generic quantification of
prior plastic deformation in crystals by using DIC? What would be the basic requirements of such an endeavour?
For crystal plasticity, the principal signature of plastic deformation has been strain localization, typically in the
form of shear bands (Bigoni and Hueckel, 1991; Asaro and Lubarda, 2006). However, this is only a fragmentary
and incomplete measurement, and for a complete characterization of plasticity, full strain correlation signatures
should be taken into account. Moreover, computing the full strain correlation signatures becomes mandatory at
Preprint submitted to Journal of the Mechanics and Physics of Solids
September 26, 2017
the microscale, where size effects and stochastic deformation can be a defining factor for plastic deformation. A
particularly popular method for tracking local strain fields is digital image correlation (DIC) (Anuta, 1970; Keating
et al., 1975). DIC is crucially dependent on optical resolution standards as well as frame rate, rendering difficult the
direct observation of strain localization in crystal plasticity. Experimental studies have shown that high-resolution
atomic force microscopy (AFM) has reached values less than a nm, to calculate step heights of a grain boundary
in the microstructure (Chen and Boehlert, 2013). In contrast, studies on fatigue crack growth have shown that
obtaining multiple optical microscope images (like DIC does) of cracks is possible and inexpensive, with resolution
of 10 nm and more (Chen et al., 2011; Z. Chen, 2012; Chen et al., 2012; Carroll et al., 2013; Khademi et al., 2015,
2016; Khademi Bieler et al., 2016; Khademi et al., 2017), through coupling with transmission electron microscopy
(TEM) or electron backscatter diffraction (EBSD) (Chen et al., 2012, 2011). In this paper, we emulate DIC using
discrete dislocation dynamics (DDD) simulations, in order to quantify and classify the statistical correlations of
various plastic deformation stages.
Strain correlation patterns in a microstructure may be investigated through the material knowledge system
(MKS) framework (Fast, 2011; Fast and Kalidindi, 2011), which can efficiently quantify generic correlations. MKS
uses statistical approaches to create processing-structure-property relationships and has the potential to bridge
multiple length scales using localization and homogenization linkages. MKS is made more practical and flexible
by the Python Materials Knowledge System (PyMKS) (Wheeler et al., 2014) framework, which is an objectoriented set of tools and examples, written in Python, that provide high-level access to the MKS frameworks
for rapid creation and analysis of structure property-processing relationships. We will be using the MKS and
PyMKS frameworks along with statistical tools from scikit-learn (Pedregosa et al., 2011) for efficiently classifying
correlations.
Mitchell defines machine learning (ML) as a natural outgrowth of the intersection of Computer Science and
Statistics (Mitchell, 2006). ML has rapidly evolved in the last decade to include almost all scientific communities (Russell and Norvig, 1995; Friedman, 1998; Werning et al., 2010; James et al., 2013). From the most simple
Gaussian representation of a data set (Rasmussen and Nickisch, 2010) to complex data mining applications and
evaluation, ML has made major inroads within materials science and holds considerable promise for materials
research and discovery. Some examples of successful applications of ML within materials research include accelerated and accurate predictions (using past historical data) of phase diagrams, crystal structures, and material
properties (Mueller et al., 2016).
In our work, we use ML for quantifying plastic deformation in crystals. A simple example of crystal plasticity arises during uniaxial compression of single crystals. In this context, the simplest realistic case is a thin
film, which can be modeled by simple 2D-edge discrete dislocation dynamics (2D-DDD) simulations (Van der
Giessen and Needleman, 1995). Such 2D simulations have been able to explain experimentally observed aspects
of size dependent plasticity in thin films (Nicola et al., 2006; Deshpande et al., 2012; Vlassak et al., 2014; Deshpande et al., 2013; Shishvan and Van der Giessen, 2010). The benefit of 2D-DDD for thin films is that it allows
for efficient numerical evaluation of inter-dislocation correlations. Nevertheless, the methods developed here are
directly applicable to more complex numerical approaches such as 3D-DDD (Wang et al., 2006) or continuous
crystal plasticity models (Asaro and Lubarda, 2006), or even actual DIC experiments. Furthermore, thin films
have demonstrated a rich phenomenology, since they display strong size effects and stochastic plastic fluctuations
as the sample thickness decreases (Papanikolaou et al., 2017; Van der Giessen et al., 2006). In this paper, we
explore whether such size effects and stochastic fluctuations influence the applicability of ML methods. In this
paper, we build an approach for explicit statistical analysis of plastically deformed microstructures by mimicking
the DIC protocol (Fig. 1). We simulate thin film compression with 2D discrete dislocation dynamics and then we
use correlation algorithms as a statistical approach, analogous to detecting digital image correlations in thin films,
as a way to infer the existing deformation history from total strain profiles. Such strain profiles may originate in
non-invasive mechanical testing (such as DIC or Digital Strain Imaging), and we assume they do. Nevertheless, the
proposed method is applicable to any strain profile, however it is obtained. ML protocols help build our statistics
model for recognition of differently deformed microstructures. Principal component analysis (PCA) is used along
with 2-point correlation statistics (Niezqoda et al., 2008) to help us recognize and find possible correlations of slip
2
bands formed due to stress, and then a clustering algorithm is used to characterize these possible correlations in an
efficient way.
The remainder of this paper is organized as follows: In section 2 we describe our 2D-DDD model for obtaining
mock data samples and we also describe our approach for quantifying strain correlation patterns. In section 3
we discuss the ML approaches we use for processing our data samples, and present our results based on various
parameters. In section 4 we present a summary of our work and the conclusions drawn.
Figure 1: Schematic of the DIC protocol in modeling and sequence of sample loading: A schematic of the sequence of events when
loading, unloading and reloading a sample is shown, with the corresponding stress and strain graphs vs the time step of simulation. A sample
is obtained from 2D-DDD simulation and it’s loaded to a specific strain value – Prior loaded state (found at specific time steps). Then, the
sample is unloaded to zero stress and the remaining plastic strain can be calculated – Prior unloaded states. Finally, the sample is reloaded to a
DIC-testing strain – Test reloaded state. The DIC process begins when the sample is at the prior unloaded state. Even though a sample has been
plastically deformed (at the prior loaded state), the samples obtained for DIC experiments can be polished, thus the surface of a sample is not
able to provide information about deformation (the prior unloaded state sample can be seen in the figure as having smooth surface). Randomly
placed DIC-tracking nanoparticles that are detected optically and contribute to correlation statistics are applied to the sample. Then, as the
sample is reloaded, the permament deformation can be seen via DIC, since there are changes in the distances between tracked nanoparticles.
3
Figure 2: DIC for various loading histories: A material has an assumed prior deformation history (prior loaded states). We unload the sample
and obtain the prior unloaded states. How does the result of DIC, which characterizes the test reloaded states, reflect the prior history?
2. The model
2.1. Discrete dislocation dynamics
The geometry of the model problem is shown in Fig. 3 for single slip system (a) and double slip systems (b).
In this work, our primary focus will be on the double slip system samples, and we will compare the performance
with single slip in Section 3.3. Samples are modeled (Papanikolaou et al., 2017) by a rectangular profile of width
w and aspect ratio α (α = h/w).
Figure 3: The 2D discrete dislocation plasticity model of uniaxial compression of thin films: Slip planes (lines) span the sample, equally
spaced at d = 10b, but planes close to corners are deactivated to maintain a smooth loading boundary. Surface and bulk dislocation sources
are present (disks) and forest obstacles are spread homogeneously across the active slip planes. Initially the sample is free of stress and mobile
dislocations (a) Single slip system (b) Double slip system
Plastic flow occurs by the nucleation and glide of edge dislocations, on single or double slip systems. We
study sample widths ranging in powers of 2 from w = 0.125 (or w0 ) to 2 µm with α = h/w = 4. The top and
bottom edges (x = 0, w) are traction free, allowing dislocations to exit the sample. Loading is taken to be ideally
strain-controlled, by prescribing the y-displacement at the lateral edges (y = 0, h). The applied strain rate (for both
loading and unloading regimes), ḣ/h = 104 s−1 , is held constant across all our simulations, similar to experimental
practice. Plastic deformation of the crystalline samples is described using the discrete dislocation framework for
small strains (Van der Giessen and Needleman, 1995). Each dislocation is treated as a singularity in a linear elastic
background solid with Young’s modulus E and Poisson ratio ν, whose analytic solution is known at any position.
4
Available slip planes are separated by 10 Burgers vectors and are oriented at ±30○ from the loading direction
(Fig. 3). At the beginning of the calculation, the crystal is stress free and there are no mobile dislocations. Bulk
sources are randomly distributed over slip planes, and their strength is selected randomly from a Gaussian distribution with mean value τ̄nuc = 50 MPa and 10 % standard deviation. We only consider glide of dislocations,
neglecting the possibility of climb. The motion of dislocations is determined by the component of the PeachKoehler force in the slip direction. Each sample contains a random distribution of forest dislocation obstacles and
surface and bulk dislocation sources. Once nucleated, dislocations can either exit the sample through the tractionfree sides, annihilate with a dislocation of opposite sign when their mutual distance is less than 6b, or become
pinned at an obstacle. Our simple obstacle model is that a dislocation stays pinned until its Peach-Koehler force
exceeds the obstacle-dependent value τobs b. The strength of the obstacles τobs is taken to be 300 MPa with 20
% standard deviation. Our simulations are carried out for material parameters that are reminiscent of aluminum:
E = 70 GPa, ν = 0.33. We consider 50 random realizations for each parameter case. If dislocations approach the
physical boundary of the sample then a geometric step is created on the surface along the slip direction (see Fig. 4).
The simulation is carried out incrementally, using a time step that is a factor 20 smaller than the nucleation
time tnuc = 10 ns. At the beginning of every time increment, nucleation, annihilation, pinning at and release
from obstacle sites are evaluated. After updating the dislocation structure, the new stress field in the sample is
determined, using the finite element method to solve for the image fields (Van der Giessen and Needleman, 1995).
Figure 4: Strain profiles captured from the 2D-DDD simulation at the prior loaded and test reloaded states – w = 2 µm – Double slip
system: A sample is loaded to 10 % strain and the resultant strain profile is shown. Then, the sample is unloaded to zero stress and is reloaded
to DIC-testing strain of 0.1 %. (a) Strain profile for the prior loaded sample at 10 % strain. Plastic steps are allowed to emerge on the film
surface (Papanikolaou et al., 2017). The strain map is unitless. (b) A strain profile at the test-reloaded state is shown, after subtracting the
residual plastic deformation at the prior-unload state. We consider such strain profiles as analogous to typical DIC experimental strain profiles.
For higher strain localization we have higher values in the colormap (light blue to red), while for little to no strain localization we have smaller
values in the colormap (darker blue). Since we subtract the remaining plastic strain at the prior unloaded states, all the previous localizations
have been removed from the strain profile, thus most of the strain profile has a dark blue color (0 values in colormap). The strain map is unitless.
The DIC test that we wish to imitate would measure the strain field in the sample after it has been strained
and relaxed, as described above, and then subjected to a subsequent “DIC-testing” strain. (Since we can measure
the strain field directly in the simulations, there is no need to simulate the DIC tracking particles.) We consider a
testing reload regime that is governed mainly by the degree of invasiveness we introduce to the data set. All tests
have been carried out for prior loaded states in three different regions (0.1 %, 1 %, 10 %) of total strain. These
regions represent samples under different degrees of plastic deformation. Fig. 2 shows a schematic of the way
we create our data set: A sample is loaded to 3 different strains (prior loaded states). From each state we unload
to obtain the prior unloaded states. In the prior unloaded states, the samples are stress free, but there is some
remaining strain due to plasticity. We then reload the samples to a specific DIC-testing strain. In particular, the
strain difference between the test reloaded states and prior unloaded states is constant in all samples.
Samples of different widths (w) undergo the same unload-reload protocol to create our data set. We have
the option to select at which strain the unload process begins, as well as the DIC-testing strain level we want to
introduce. We perform tests at two values of the reloading strain; the small-reload data set has a DIC-testing strain
5
of 0.1 %, and the large-reload data set has strain of 1.0 % : that is the difference in strain between prior unloaded
states and test reloaded states is 0.1 or 1 %. Fig. 5 shows various characteristic stress-strain curves like that shown
schematically in Fig. 1 (a). The prior loaded state points, as well as the prior unloaded and test reloaded points are
shown. Note that there is less noise in the data for larger widths w (Papanikolaou et al., 2017). Fig. 5 (c) shows a
stress strain curve obtained through reloading to larger DIC-testing strain, 1 %. The main difference is at the reload
points, which show further deformation of the sample, in contrast to samples reloaded to smaller strain (0.1 %).
Figure 5: Examples of stress strain curves obtained from simulations at various widths and DIC-testing strains: Samples are loaded to 3
different deformations levels (prior loaded states –diamonds). The sample is then unloaded to prior unloaded state (squares). The predeformed
sampled is then reloaded to a DIC-testing strain (0.1 % for “small−reload” and 1 % for “large−reload” data sets –disks). Each color represents
a different loading-unloading-reloading region of the sample. (a) w=0.125 µm, DIC-testing strain = 0.1 %. The prior loaded states are found
at 0.1, 1 and 10 % strain. The prior unloaded states are found at 0.01, 0.59 and 9.61 % strain. Test reloaded states are at 0.11, 0.69 and 9.71
% strain. (b) w=2.0 µm, DIC-testing strain = 0.1 %. The prior loaded states are found at 0.1, 1 and 10 % strain. The prior unloaded states are
found at 0.01, 0.76 and 9.73 % strain. Test reloaded states are at 0.11, 0.86 and 9.83 % strain. (c) w=2 µm, DIC-testing strain = 1 %. The prior
loaded states are found at 0.1, 1 and 10 % strain. The prior unloaded states are found at 0.01, 0.82 and 9.69 % strain. Test reloaded states are
at 1.01, 1.82 and 10.69 % strain, respectively.
2.2. Emulation of the digital image correlation (DIC) technique for plastically deforming crystals
Digital image correlation (DIC) is an optical method that employs tracking and image registration techniques.
DIC techniques have been increasing in popularity, including micro- and nano-scale mechanical testing applications, due to its relative ease of implementation and use. Advances in computer technology and digital cameras as
well as white-light optics have been instrumental in the development of the technique (Anuta, 1970; Keating et al.,
1975). DIC compares a series of images of a sample at different stages of deformation, tracks pixels movement in
a representative volume and calculates displacement and strain with a correlation algorithm. The tracking is done
by applying speckled registration dots to the surface of samples. The dots move along as the sample is deformed.
Each dot moves both absolutely and relatively to other nearby dots, making possible a reliable identification of
the same registration dots in different images, even though the dots have moved. This identification can be used
to measure deformation, displacement, strain, and optical flow and it is widely applied in many areas of science
and engineering. Through this process we can easily obtain strain fields that may help characterizing a particular
elastoplastic behavior (Anuta, 1970; Keating et al., 1975; McCormick and Lord, 2012; Roux et al., 2009). DIC
only measures relative displacements, it cannot by itself distinguish elastic from plastic strain – for that, one has to
unload the sample and see how much of the strain is recovered.
Our data set consists of samples of various widths under different states of deformation, obtained by simulating
thin film uniaxial compression with our 2D-DDD algorithm. Fig. 6 shows samples of different widths w under
deformation. In particular, (a) and (c) show samples at 0.1 % total DIC-testing strain while (b) and (d) show
samples at 1 % total DIC-testing strain. These figures are produced through removing the remaining plastic strain
(at the prior unloaded states of each sample) from the test reloaded states of each sample. In DIC these figures
would be the final images produced after reloading the sample.
Clearly it is not straightforward to characterize the plastic behavior of the samples without prior knowledge (i.e.
the degree of plasticity incurred from the prior loaded and test reloaded states). For example, in Fig. 6, without
prior knowledge we would not know that the samples in (a),(b) are loaded to 10 % strain while in (c),(d) the
6
samples are loaded to 1 % strain. The figures appear to be quite different, and the similarity of their histories is not
recognizable by eye. Furthermore, (a) and (c) appear to be similar and without prior knowledge, we wouldn’t know
that the samples were loaded to different strains (the same is true for (b) and (d)). The figures show similarities
that only ML’s trained eye is able to detect. Indeed, in later sections we will show how ML algorithms can show
that the differences in figures are quantifiable and fundamentally different. With the help of ML we are able to find
the initial deformation history of various samples, as long as the DIC-testing strain does not overwrite it.
Figure 6: Variety of strain profiles in 2D-DDD simulations for smaller and larger DIC-testing strain: A sample is loaded to a high
deformation of strain (which could be either 1 % or 10 %) and then unloaded. This predeformed sample is then reloaded to a DIC-testing strain.
In (a),(b) the samples are loaded to 10 % strain, while in (c),(d) the samples are loaded to 1 % strain (a) Small DIC-testing strain (0.1 %). w=2
µm. Double slip system simulation. (b) Large DIC-testing strain (1 %). w=1 µm. Single slip system simulation (c) Small DIC-testing strain
(0.1 %). w=1 µm. Single slip system simulation. (d) Large DIC-testing strain (1 %). w=2 µm. Double slip system simulation. For description
of color map see Fig. 4 (b).
2.3. Microstructural Discretization
In this section, we manipulate our microstructural data to make it suitable as input to the subsequent machinelearning steps, closely following the scheme of the Materials Knowledge System (MKS) (Fast and Kalidindi,
2011)
In the general MKS scheme, one selects one or more spatially-varying quantities which characterize the microstructure. The space of all possible values of these quantities is called the local state space, H, and a point in
this space is denoted h. Some care with the vocabulary is required, since physically speaking, these quantities are
simply the values of fields of interest to us, and may or may not correspond to thermodynamic state variables.
In particular, in the current study, we take as our quantity of interest the determinant of deviatoric total strain
(other invariants work equally well) φ ≡ 21 (ε2xx +ε2yy )−ε xx εyy +2ε2xy , where the tensor ε is the total strain, including the
plastic part, which is of course not a thermodynamic state variable. The DDD simulations are naturally discretized,
so we simply use the DDD grid to index points in space. (See Appendix A– Fig. A.24 for additional details.)
In the MKS method, one further considers a “microstructure function”, defined on the product space of the
microstructure state variables H, and physical space x, m(h, x). In general use, this function may be thought of a
probability density on these spaces. In our case, where we have a succession of particular microstructural instances,
the microstructure function corresponding to each instance is a delta-function in h at each point in space.
In order to obtain data suitable for constructing two-point correlations, it is necessary to bin the state variables.
We make use of the PyMKS software (Wheeler et al., 2014) which offers tools to accomplish this. The most basic
7
h-axis discretization scheme is the so-called “primitive basis” scheme, in which one selects some number N of
evenly-spaced levels, h0 , h1 , . . . hN , and, at a point in space where the state variable has value h, selects amplitudes
ωi for these levels such that ∑i ωi hi = h, with the additional restriction that only the hi ’s which are directly below
and directly above the local value h are nonzero, and ∑i ωi = 1. The entire system is thus described by a set of
values {ωi } in each spatial bin x, from which we can compute auto- and cross-correlation functions (ωi , ωi ) and
(ωi , ω j ). In our simulations we discretize the state space into 3 different bins, corresponding to 3 local states ω1 ,
ω2 , and ω3 at low, intermediate, and high local strains.
Figure 7: w = 2 µm – 2D projection of PCA results for thin films – Double slip system: ω1 ,ω1 auto-correlation. (a) Projection of data set
on first two principal components. Red blobs denote samples with 0.1 % strain prior loaded state, blue triangles samples with 1 % strain prior
loaded state and green squares denote samples with 10 % strain prior loaded state, respectively. (b) First principal component of PCA, shown
in sample coordinates (Fig. 3). (c) Second principal component of PCA, shown in sample coordinates (Fig. 3). The colormaps are unitless.
Figure 8: Efficiency of ML compared to visual inspection of results – w = 2 µm – Doubleslip system: ω1 ,ω1 auto-correlation. The colors
follow the definition of Fig. 7. (a) Strain profiles obtained through our model. The top figure corresponds to strain profiles for a sample initially
loaded to 0.1 % strain. The middle figure is a sample strain profile obtained from initial loading of 1 % strain. Finally, the bottom figure is
a sample strain profile through initial loading at 10 % strain. All strain profiles have been obtained at the test reloaded state by subtracting
the remaining plastic strain at the prior unloaded state. (b) PCA map for samples that have similar strain profiles as (a). 3 distinct clusters
are formed , even though samples that were initially loaded to 1 and 10 % strain have somewhat similar strain profiles (in some areas). The
projection is upon the first and second principal components. For description of color map see Fig. 4 (b).
8
Figure 9: w = 0.5 µm – 2D projection of PCA results for thin films – Double slip system: ω1 ,ω1 auto-correlations. The colors follow the
definition of Fig. 7. (a) Projection of data set on first two principal components with a clustering algorithm applied to the data set, demonstrating
a failure in clustering the various deformation levels. (b) Projection of data set on first two principal components without a clustering algorithm
applied to the data set, justifying (a). (c) First principal component of PCA, shown in sample coordinates (Fig. 3). (d) Second principal
component of PCA, shown in sample coordinates (Fig. 3). For description of colormaps, see Fig. 7.
3. Pre-processing, clustering and classification for strain profiles of crystalline thin films
The classification of deformation histories through investigations of strain profiles is not always simple. Certainly, strain localization and shear banding can be strong signatures of local plastic deformation, however with
our small DIC-testing strains, localization is not easy to observe. Instead, signatures of plastic deformation are
concentrated in collective features of the 2-point correlation function (see Eq. B.6). In order to identify and classify these collective features, a statistical approach needs to be implemented in a multitude of training samples. In
the following, we use Principal Component Analysis (PCA) as a step in classifying deformation histories. (There
are many other machine learning methods that we could have used instead, but a survey of possible methods lies
beyond the scope of this manuscript.) We also discuss the applicability of a classification algorithm to the PCA
results.
3.1. Principal Component Analysis
After obtaining a set of samples with different plastic deformation histories and calculating statistical correlation functions of each sample, we must show that differences between the correlations can be used to distinguish
samples with different “prior” plastic strain levels. A handy tool to apply is principal component analysis (PCA).
PCA is a statistical approach that finds an ordered set of orthogonal basis vectors that efficiently describes
the variance in a data set (Jon Shlens, 2003). These vectors are called principal components. The number of
principal components is necessarily less than the possible number of observations and the total number of original
variables. The basis is defined in such a way that the first principal component has the largest possible variance (that
is, accounts for as much of the variability in the data as possible), and the following components in turn have the
highest variance possible, while still being orthogonal to all the preceding components. For this reason, it is critical
that all samples in the data set were acquired under identical controlled conditions with a clear understanding of
9
the origin of the variability. PCA is commonly used as one step in a series of analyses; one can use it to reduce the
number of variables, especially when there are too many predictors relative to the number of observations.
PCA is accomplished by the application of Singular Value Decomposition (SVD) (Press, 2007). Given an data
matrix, D, whose columns each contain the components of a single data point, D may be decomposed to a diagonal
matrix of singular values, S, and left/right singular vector matrices V and U, with D = VT SU. The columns of V
and U are the left and right singular vectors of D. The V vectors that correspond to the largest singular values
capture the most characteristic features. After calculating the V vectors, we project the data samples onto the
subspace defined by them. In decreasing order of their corresponding singular values, we denote the first three
basis vectors (principle components) as PC1, PC2, and PC3.
In Machine Learning, the PCA technique is used to project data vectors which lie in a high-dimensional space
to a smaller dimensional space whose basis vectors are selected to capture most of the variation in the original
data. In the textbook scheme, one typically has an n × m array, with n indexing the samples, and m being the
dimensionality of the feature space the vectors lie in. In our case, we wish to do PCA on our correlation data,
which we can treat as a vector by listing the data values from the grid in raster order.
For any given sample, it is possible to compute the two-point correlations out to any distance that can be
supported by the data, but in order to meaningfully compare correlations across samples of different sizes, it is
necessary to have uniform input data to our eventual machine-learning pipeline. For this reason, we truncate the
two-point correlation vector on to a square grid that we believe is large enough to be a useful statistical characterization of the spatial arrangement of the strain field, but small enough to give reasonable execution speeds in our
numerical analysis operations.
Having projected the correlation data to the PCA-derived subspace, we can address our main objective, to
find the prior deformation history of the samples from the strain correlation measurements. We generate data on
samples of different widths w (0.125, 0.25, 0.5, 1.0, 2.0 m) for both single and double slip systems, reloaded to two
values of the DIC-testing strain which we call “large-reload” (1 %) and “small-reload” (0.1 %). The aspect ratio
is fixed (h/w = 4), and the number of pixels in each of the DDD simulations is fixed to 80 (in both dimensions).
We calculate the two-point correlation matrix for all samples of each width and restrict the data to the 40 by 40
region. Continuing, we adjust the truncated correlation matrix by subtracting the average. The end result is used as
an input for the start of our ML method. The information is fed to the PCA algorithm in order to change from the
high dimensional space of the correlation matrix, to a more controlled subspace of the principal components. In
our calculations the first three components of PCA describe more than 98 % of the variance in the data sets, so we
restrict the computation to those three components. Next, we project the correlation matrix onto the new reduced
subspace. Visually, at this point, we are able to examine the possible clusters that have formed due to the different
PCA vectors that have been produced. Furthermore, we are able to identify the samples that belong to each cluster,
which we use as an input for examining the accuracy of classification algorithm. The set of points belonging to
the 3D subspace of PCA (PC1,PC2,PC3) are used as an input to the classification algorithm to examine whether
or not the visually identified clusters are replicated by the classification algorithm.
10
Figure 10: w = 1 µm, The choice of the correlation domain and how it impacts the PCA maps – Double slip system : ω1 ,ω1 autocorrelation. For description of colormaps and colors of PCA maps, see Fig. 7. Projection of data set on first two principal components. (a)
40 by 40 domain of correlation matrix. Highly smooth in the center and towards the boundaries of the domain. (b) 100 by 100 domain of the
correlation matrix. A highly focused area near the center of the domain is shown, where the phenomena are focused. The smoothness present
in (a) is slowly removed from this domain. (c) 400 by 400 domain of correlation matrix. Along the y-axis we have hit the actual boundaries
of the sample. We have rich phenomenology present towards the center of the correlation matrix and at the boundaries. (d) PCA maps for 40
by 40 domain. (e) PCA map for 100 by 100 domain. The variance of the data has changed and the projections have shifted. The information
provided by (b) doesn’t change the cluster formations, but introduces unnecessary information that has shifted the results along the PC1 and
PC2 axes. (f). PCA map for 400 by 400 domain. The variance of the data has changed even more compared to (e). The distances between
the blue and green clusters have increased an order of magnitude compared to (e) and 2 orders of magnitude compared to (d). The information
provided by (c) doesn’t affect the clusters that are formed from our algorithm.
3.2. Clustering and classification
We use the Continuous k-Nearest Neighbors (CkNN) algorithm (Berry and Sauer, 2017) to classify samples
after running PCA on the data set. CkNN searches for samples that are close to each other (k-nearest neighbors)
and arranges them into clusters. The algorithm is an unsupervised method that detects natural clusters within
a data set, and our interest in it is the degree to which the natural clusters correspond to the prior deformation.
After different clusters have been identified, the algorithm classifies the samples based on the cluster to which they
belong. The input to this algorithm must consist of a set of points, which in our case is the projections of the
correlation matrix on the 3 principal components. As an output, the algorithm produces the classified samples,
based on the cluster to which they belong.
The size of the data set, as in most classification algorithms, imposes a limitation on the algorithm. For double
slip system simulations our data set consists of 150 samples for each w while for single slip system simulations our
data set consists of 27 samples for each w. This introduces a limitation on classification, especially for single slip
systems. The algorithm groups data samples with similar PCA vectors into one cluster. We find that the algorithm
works better for larger data sets. The method is successful if the samples with different prior loading are grouped
into different clusters. Note that the clustering is done in three dimensions using all three PC components, but most
of our plots are two dimensional, which can sometimes hide the degree of clustering.
In Fig. 11 we show a 3D PCA map for material samples of w = 1 µm. It is obvious that the clustering isn’t
affected by Mthe 3rd dimension, and in this case the information provided by PC3 is irrelevant to our results.
11
The results shown on this paper, except for results shown in section 3.5, are extracted by applying the PCA and
the CkNN algorithm to the whole data set. The same PCA and CkNN steps are applied to all the simulations. The
remainder of this paper discusses how well the clustering algorithm works in various situations.
Figure 11: w = 1 µm, 3D projection of PCA results for thin films – Double slip system : ω1 ,ω1 auto-correlation. The colors follow the
definition of Fig. 7. 3 different clusters are shown like in Fig. C.25. Introducing the 3rd component into the PCA map, doesn’t affect the results.
Figure 12: w = 2 µm, – Strain profiles and 2D projection of PCA results for thin films – Single slip system : ω1 ,ω1 auto-correlation. The
colors follow the definition of Fig. 7. Projection of data set on first two principal components. (a) Different strain profiles are seen for a single
slip system of w = 2 µm. These strain profiles are obtained through subtracting the remaining plastic strain at the prior unloaded state, from
the test reloaded state. The top profile is for sample initially loaded to 0.1 % strain, then unloaded to zero stress and reloaded to DIC-testing
strain of 0.1 %. The middle profile is for sample initially loaded to 1 % strain and then unloaded and reloaded. The bottom figure is for sample
initially loaded to 10 % strain and then underwent the unload-reload process. (b) PCA projection of of our results for samples of w = 2 µm.
The similarity between strain profiles at 1 and 10 % doesn’t affect the formation of separate clusters for samples that were initially loaded at
these strains. For description of color map see Fig. 4 (b).
12
Figure 13: w = 2 µm – 2D projection of PCA results for thin films – Comparison between single and double- slip system : ω1 ,ω1
auto-correlation. The colors follow the definition of Fig. 7. Projection of data set on first two principal components. (a) Single slip system
projection (b) Double slip system projection.
Figure 14: w = 2 µm – Comparison of principal components among double and single slip systems: Components shown correspond to
the analysis of Fig. 13. (a) First principal component of double slip system. (b) First principal component of single slip system. (c) Second
principal component of double slip system. (d) Second principal component of single slip system. For description of colormaps, see Fig. 7.
13
3.3. Distinguishing plasticity regimes for small DIC-testing strain (0.1 %)
We ran a multitude of tests for different w. For large w (> 0.5µm), our algorithm correctly clusters and classifies
data into 3 different groups, one for each of the prior strain values, which was the main objective of our work. Fig. 7
(a) shows that clustering is easily observed for w = 2 µm, where 3 distinct clusters appear in the PCA of the mω1
autocorrelation. It is clear that there is enough cluster separation to reliably classify plastically deformed metals
into heavily deformed and less deformed categories. For these larger sized systems the CkNN algorithm has 100
% accuracy, while for smaller sized systems with w ≤ 0.5 µm the clustering algorithm fails to cluster data points
according to their deformation state. That is evident in Fig. 9 (b), where one can see what a correct clustering and
classification would look like for specimens of w = 0.5 µm. In Fig. 9 (a) one can observe the results after the CkNN
algorithm is applied to the data set. Other figures in the Appendix C.1 show how specimens of various sizes are
classified using the CkNN algorithm. The plastic noise fluctuations in the system, as well as the finite size of the
system, interferes with the classification of smaller sized data samples, while for larger w the samples are classified
correctly.
Fig. 7 (b) and (c) show the representation of the first 2 principal components, shown in their natural sample
coordinates (i.e., the PCA vectors have been converted back to the 2D grid representation of a correlation function)
with units denoting number of elements (in regular DIC, it would correspond to number of pixels in each direction).
If two correlation functions were randomly chosen from the data set, the difference between them would most likely
look like 7 (b) (with some scaling) mixed with a smaller amount of 7 (c). Note that the first principal component is
roughly isotropic, while the second is strongly anisotropic. Figs. 7, 9 show the progression of PCA’s effectiveness
as a clustering technique as sample width decreases. We can observe that the first PCA component at larger w
is relatively isotropic. While in Fig. 7 (b) we notice a concrete isotropy of the first principal component of the
analysis, it gradually becomes anisotropic as the sample width decreases (Fig. 9 (b)). This change is correlated
with the onset of stochastic fluctuations at small scales and mechanical annealing (Shan et al., 2008) that promotes
concrete slip bands even at small DIC-testing strains. While both principal components for w = 2 µm (Fig. 7 (b,c))
are smooth, they gradually become less structured as w decreases (Fig. 9 (b,c)), naturally an effect of stochastic
fluctuations at small length scales. For w = 2 µm there is a distinct difference between the first and second principal
components, related to a spatial symmetry breaking. This distinction disappears as w decreases. For smaller w, due
to the emerging crystal plasticity size effects (Papanikolaou et al., 2017), the data set is not as distinguishable as
we would have wanted with our clustering technique, because of the noise associated with strengthening. (Fig. 5
(a)). Examples may be seen in Fig. 9 and in Appendix C.1. These figures show the behavior of our classifier and
how they compare to the actual initial deformation of our samples.
One deficiency of our procedure emerged as we examined the results: as w decreases, the distance between the
PCA-transformed samples also decreases. It is known that classification algorithms have an inherent limitation:
when the distance between points in one cluster is similar to the distance separating two clusters, then the algorithm
has difficulty distinguishing the clusters. In particular, Fig. 7 shows that the cluster distances in the PC1 direction
are of order of magnitude 10−2 to 10−1 . For w ≤ 0.5 µm the cluster between PCA-transformed samples is on the
order of 10−3 to 10−2 , similar to the distance between the samples itself, and the data samples cannot be classified
correctly. For smaller sized systems, it is evident that samples with 0.1 % and 1 % prior loaded strain (red circles
and blue triangles, respectively) are so close to each other that the classifier regards them as belonging to the same
cluster.
3.4. Distinguishing plasticity regimes for single slip samples
As mentioned in Sec. 2.1, we model single and double slip systems. So far, we have shown how emergent shear
bands can be observed in our simulations for both of these systems (Fig. 6), as well as PCA results for double slip
(Figs 7, 9). PCA results for single slip are consistent with double slip, as shown in Fig. 12. Specifically, Fig. 12 (a)
shows results of single slip system simulations for w = 2 µm and (b) for w = 1 µm. The clustering properties for
these larger sized systems are similar to the properties observed for similar systems for double- slip simulations.
Fig. 13 displays a direct comparison between the results of single slip and double slip systems with w = 2 µm. The
PCA results contain distinctly separated clusters.
14
Fig. 14 compares the principal components of the structure factors for single and double slip systems.
Figure 15: w = 2 µm – 2D projection of PCA results for thin films – Double slip system – Validation: ω1 ,ω1 auto-correlation. Red blobs
denote training samples with 0.1 % strain prior loaded state, blue triangles training samples with 1 % strain prior loaded state and green squares
denote training samples with 10 % strain prior loaded state, respectively. Red stars depict testing samples of 0.1 % prior loaded state, blue
stars testing samples of 1 % prior loaded state and green stars testing samples of 10 % prior loaded state. Validated-split data set. Projection
on first two principal components.
3.5. Validation and accuracy of the algorithm
A ML algorithm, in order to be considered successful, should be validated with “unknown” data sets (testing
data) which have the same features as the data set the algorithm was designed for (training data). In many cases,
testing data sets are hard to find, so the whole data set is split into two parts (not necessarily a half and half split),
and the ML algorithm can be trained on part of the data set and its effectiveness tested on the rest. The results
shown in the rest of this paper are an application of PCA transformation and the CkNN algorithm on the whole
data set (for a given w).
For validation purposes, we “trained” the algorithm by computing the PCA transformation from a randomly
chosen half of the w = 2 µm samples and applying the CkNN algorithm. Then we applied the PCA transformation
to the remaining half of the samples and examined whether or not they were projected into the correct clusters.
The results are shown in Fig. 15. It is evident that the testing data perfectly matches the training set. Similarly
“training” the algorithm to samples of various sizes (i.e., half of the samples instead of all the samples), follows
the results of section 3.3. For samples with w ≥ 1 µm the “testing” data set is projected to the 3 classified clusters
that have formed. In contrast, for smaller sized systems, the training data set is misclassified (as happens when
examining the whole data set) and the testing data set falls within the misclassified results.
We can quantify the degradation of the clustering process using some of the tools provided in the scikit-learn
metrics module (Pedregosa et al., 2011). In particular, we examine the accuracy score of the algorithm, as well
as the Fβ score. Accuracy is the ratio of the number of classified samples by our algorithm (prediction) over the
number of the true classification of our samples. We apply the CkNN algorithm and generate clusters. Because
we know the prior strain for each sample, we can immediately check whether the clusters correspond to the strain
levels. Perfect clustering is when each cluster contains only samples with identical prior strains. The results are
summarized in Fig. 16. For w ≥ 1 µm the accuracy score is 1 as seen in Fig. 16 (a); that is, all the samples are
correctly classified. For smaller samples w ≤ 0.5 µm (or w/w0 ≤ 22 as in the figure), we have a 0.33 accuracy score,
because only the samples of one cluster are correctly classified. The CkNN algorithm predicts that most samples
belong to that cluster, but this doesn’t affect the accuracy score, since the samples that belong to that cluster are
correctly classified.
To quantify the performance of the classification process, we use the Fβ score (David Powers, 2011; Yates and
Ribeiro-Neto, 2011),
15
Fβ = (1 + β2 ) ⋅
(β2
p⋅r
⋅ p) + r
(1)
where precision p is the number of correctly classified samples in the cluster divided by the number of all
classified samples in the same cluster, and recall r is the number of correctly classified samples in the cluster
divided by the number of samples that should have been returned (classified). The β number changes the weight
of recall vs precision. For β > 1 then recall is weighted more than precision, while for β < 1 precision is weighted
more than recall. For β = 1 then we have the F1 -score, with precision and recall having the same weight in the
equation. In our work, we examine the F1 , F2 and F0.5 -score. In Fig. 16 (b,c,d) we see the respective results for
these scores. For samples with w ≥ 1 µm (or w/w0 ≥ 23 we have value of 1 on all scores and all clusters, while for
smaller w we observe that the line with the squares, which corresponds to samples with 10 % initial compressive
loading returns non-zero values, varying as the β value changes. For the samples belonging to that cluster we don’t
get the highest possible result, because the number of correctly classified samples is smaller than the number of
samples in the cluster (i.e., the precision is small). The line with the circles, which corresponds to samples with 0.1
% initial strain loading, has value 0 for w ≤ 0.5 µm because no samples have been classified as belonging to that
cluster. The last line, with the triangles corresponding to samples with 1 % initial loading has non-trivial values
because in some cases there are some samples that are classified correctly (the recall and precision are very small).
In summation: For the “square” cluster we have low precision but high recall,since we classify the samples that
actually belong that the cluster correctly, but we also classify samples from other clusters; for the “triangle cluster”
we have low recall and low precision, since we classify a small number of samples into that cluster.
16
Figure 16: Accuracy scores of classified samples - 0.1 % DIC-testing strain:. ω1 ,ω1 auto-correlations. The x-axis of each graph is the base
2 logarithm of the various sample widths we examine. (a) Accuracy score for the samples. Maximum value 1 means that all the samples have
been correctly classified. (b) F1 -score of our 3 clusters that are formed. The line with the squares represents the cluster with samples of 10
% initial strain loading, while the line with the triangles is for the cluster with samples initially loaded to 1 % strain. Finally, the line with the
circles is for the cluster with samples initially loaded to 0.1 % strain. For smaller sized systems we have observed that most of the samples are
classified as belonging in the “square” cluster, hence the scored value for that cluster only. Since the algorithm correctly classifies the samples
that were initially loaded to 10 % strain, but also classifies more samples as belonging to that cluster, then the score does not have the maximum
value of 1 but lower. (c) F2 -score of our 3 cluster that have formed. The definition of the colored lines follows (b). Since for F2-score we
have increased weight of the recall, the 0.7 maximum value is expected for the “square” cluster. (d) F0.5 -score of our 3 clusters. The colors
definitions follow (b). Since we have reduced weight of the precision, for lower sample widths it is expected to have lower score than F1 for
the “square” cluster.
17
Figure 17: Large-reload vs. small-reload DIC-testing – Example of PCA projection results for thin films of w = 0.5 µm:. ω1 ,ω1 autocorrelations. The colors follow the definition of Fig. 7. (a) Sample with prior loaded state at 10 % strain is unloaded to zero stress and then
reloaded to small DIC-testing strain (0.1 %). The strain profile shown is captured after removing the prior unloaded state strain profile. (b)
Small DIC-test reloaded state strain (0.1 %), without a clustering algorithm applied to data set. Projection on two principal components. Actual
representation of the data set, with some mixing of the samples. The clusters have shifted closer to one another but not indistinguishable. (c)
Sample with prior loaded state at 10 % strain is unloaded to zero stress and then reloaded to large DIC-testing strain (1 %). The strain profile
shown is captured after removing the prior unloaded state strain profile. The difference in strain profiles (a) and (c) is clearly shown. (d)
Large DIC-test reloaded state (1 %) without a clustering algorithm applied to the data set. Actual representation of the data set. For the higher
DIC-testing strain of 1 %, we can see that there is much more mixing of the samples. Reloading to higher strain values, adds plastic memory
to the samples, rendering our process inapplicable for these cases. For description of color map see Fig. 4 (b).
18
Figure 18: First and second principal component of PCA application on thin films of w = 0.5 µm shown in sample coordinates: (a) First
principal component - Small DIC-test reloaded state strain (0.1 %) (b) First principal component- Large DIC-test reloaded state strain (1 %)
(c) Second principal component - Small DIC-test reloaded state strain (0.1 %) (d) Second principal component- Large DIC-test reloaded state
strain (1 %). For description of colormaps, see Fig. 7.
3.6. Distinguishing plasticity regimes for large DIC-testing (1 %) total strain
The construction of the “large-reload” data set was intended to show the delicate handling needed for us to
get the cluster separation we require. In particular, when reloading to 1 % DIC-testing strain our method does not
produce the desired separation. As the DIC-testing strain increases so do strain localization features and shear band
sizes. With a shear band spanning the whole specimen, we expect that the statistical correlations differ significantly
from the statistical correlations of the “small-reload” data set. That is due to the overall effect of localization, from
a structural correlation viewpoint: as load is applied to the sample and before strain localization emerges in the
form of shear bands, crystal plasticity must be homogeneous and relatively isotropic. This is mainly due to the
dislocation sources and obstacles that are placed in an unbiased manner (randomly) across the slip planes in our
simulations. However, after localization emerges, correlations are dominated by shear banding and become highly
anisotropic. Our methods pick up that transition. Indeed, even in the case of lower reload strain, while some
separation is present, for the prior loaded state of 0.1 % strain and 1 % strain the distance between clusters is small,
while in smaller sized systems (Fig. 9) samples are unclassifiable. Coupled with the invasive nature of the total
strain imposed on the DIC-test reloaded state, the inseparability of the clusters makes the samples’ prior loaded
state history indistinguishable.
Fig. 17 compares the small and largereload DIC-testing regimes. Fig. 17 (b) and (d) show the results of PCA
with a clustering algorithm applied to the data set. From Fig. 17 (d) we can see that higher DIC-testing strain
renders samples indistinguishable in PCA coordinates. The separation that was present for the low DIC-testing
strain (0.1 %) is missing for higher values. Figs. 17 (a,c) show the strain profiles captured when the sample is
19
reloaded to low (a) DIC-testing strains and high (c). A more comprehensive comparison for these regimes can be
found in Appendix C.3. It is obvious that for higher DIC-testing strain there is much more mixing of the samples,
thus the classification algorithm fails. Fig. 18 shows another difference between the two DIC testing regines. For
small reload strain the first principal component (a) appears relatively isotropic, while it becomes highly anisotropic
for larger reload strain (c). This observation extends to other components (e.g., 2nd (b,d)) and is correlated to the
emergent anisotropy of strain localization.
3.7. Dependence of unsupervised learning capacity on pre-processing aspects.
As discussed in Sec. 2.3, the discretization scheme defines the form and dimensions of the correlation functions
to which we apply a PCA transformation. We can choose to examine correlations between different local states
ω. We can categorize samples based on their deformation history either for ω1 ,ω1 auto-correlations or ω2 ,ω2 autocorrelations. We find that cross-correlations aren’t helpful for classifying samples according to their deformation
levels. Figs. 19 shows results obtained from various correlation functions: in general, as w decreases, we observe
that distances between each cluster are also decreasing. In particular, Fig. 19 (b) shows that the distances in each
cluster are measured in an order of magnitude 10−4 to 10−3 while in Fig. 19 (a),(c) the order of magnitude is 10−2
to 10−1 . This difference in Fig. 19 (a),(c) is enough for the clustering algorithm to find the different deformation
levels and classify with 100 % accuracy our data set.
Figure 19: Auto-correlations vs. cross-correlations for pre-processing – Example of PCA projection maps for w = 2 µm: The colors
follow the definition of Fig. 7. w = 2 µm. (a) ω1 ,ω1 auto-correlations (b) ω1 ,ω2 cross-correlations (c) ω2 ,ω2 auto-correlations.
Another choice we can make is the quantity that characterizes the microstructure. Until now, we considered an
isotropic measure of the total deformation strain in the sample. Our classification scheme produces similar results
if we use the more common 2nd invariant of the strain deformation tensor, J2 = εik εki . Fig. 20 shows the results for
different microstructural measure calculations. For larger systems (w = 1,w = 2) the only notable difference is the
overall variance of the data in PCA coordinates.
20
Figure 20: Effect of strain invariant type for pre-processing – Examples of PCA projection maps: The colors follow the definition of
Fig. 7. (a) ω1 ,ω1 auto-correlations. w = 2 µm. Isotropic measurements. (b) ω1 ,ω1 auto-correlations. w = 2 µm. J2 invariant. (c) ω2 ,ω2
auto-correlations. w = 1 µm. Isotropic measurements. (d) ω2 ,ω2 auto-correlations. w = 1 µm. J2 invariant.
Finally, we may use the plastic strain as the microstructural deformation state variable (following Kalidindi -it
is not actually a thermodynamic state variable), which would effectively correspond to applying DIC on the test
unloaded state dislocation ensembles. The test unloaded state would be the state that would emerge after unloading
the sample on the test reloaded state. Fig. 21 shows that classification still works and there is an observable
difference of the data variance in PCA coordinates. For more figures on the differences in pre-processing aspects
see Appendix C.4.
Figure 21: Total vs. Residual/Plastic strain for pre-processing: Examples of PCA projection maps: The colors follow the definition of
Fig. 7. (a) Plastic strain. ω1 ,ω1 auto-correlations. w = 2 µm (b) Total strain. ω1 ,ω1 auto-correlations. w = 2 µm
21
3.8. Independence from the choice of discretization schemes and dimension reduction methods
We find that our methods are not sensitive to reasonable variations of the microstructural binning of the local
strain variable. As a test, we discretize the microstructure into L = 2, 3, 4 and 5 parts. We are able to distinguish
the initial deformation history of all the samples when calculating the ω1 ,ω1 auto-correlations and the ω2 ,ω2 autocorrelations. These results are independent of the discretization scheme (i.e. the number of local states used).
Fig. 22 shows the results for data samples of w = 1 µm, as the number of local states L increases. Clustering and
classification is possible, and the clustering algorithm has 100 % accuracy independently of the number of local
states, but the overall noise of the data increases with the number of local states. The noise is due to the use of
a fixed number of DDD simulations for each prior strain level. The signal strength in each correlation function
increases with system size and the number of dislocations, but decreases as the data is distributed into more bins L.
This effect is more pronounced for cross-correlations because they decrease for short distances and our correlation
function range is truncated. Hence we do not obtain classifiable results for any cross correlations.
Figure 22: Effect of discretization schemes on pre-processing: Examples of PCA projection maps for w = 1 µm: The colors follow the
definition of Fig. 7. ω1 ,ω1 auto-correlations. (a) 2 local states (b) 3 local states (c) 4 local states (d) 5 local states
22
While PCA is one of the most common and useful tools for dimensionality reduction, some data sets could be
so large that it is impractical. With that in mind, we compared our PCA results with other common algorithms, such
as Incremental Principal Component Analysis (IPCA) and the Truncated Singular Value Decomposition (TSVD).
IPCA uses a different form of processing a data set that allows for partial computations which in most cases match
the results of PCA. Incremental PCA stores estimates of component variances and updates the variance ratio of
a component incrementally. It is faster and uses memory more efficiently than PCA. TSVD on the other hand,
implements a variant of singular value decomposition (SVD) that only computes the largest singular values. Given
that PCA works on the basis of the singular value decomposition, we expect little to no difference with this method.
No significant differences are seen when applying some of these variations of PCA to our data sets. The results
are shown on Fig. 23. Note that no additional parameters, other than the initialization of the different methods,
have been modified; in particular, the same clustering algorithm is used as with the PCA methods. The TSVD
results do not display any differences from regular PCA, besides slight changes in data variance and data cluster
positions. The IPCA results, on the other hand, are mirrored from the PCA results in both the PC1 and PC2 axes
(negative values). If we calculated the absolute values we would see just minor differences in data variance and
cluster positions as in the TSVD results.
Figure 23: Comparison of different dimensionality reduction methods for w = 2 µm: The colors follow the definition of Fig. 7. ω1 , ω1
auto- correlations. (a) PCA (b) TSVD (c) IPCA
4. Summary
We examined the applicability of ML algorithms to practical and relatively inexpensive experimental methods
for the detection of mechanical deformation and crystal plasticity, such as DIC. The principal question was whether
it is possible to detect the degree of prior plastic deformation of thin films, especially when they display significant
plasticity size effects. Our conclusion is that ML algorithms can achieve our objective with varying levels of success in various situations. Through mimicking the DIC protocol in two dimensional discrete dislocation plasticity
simulations, we identified realistic cases (single and double slip thin films with widths larger than 1 µm) where
data clustering and classification is possible, based on the degree of prior plastic deformation. However, when size
effects come into play, we found that such clustering and classification becomes gradually more difficult, since the
intrinsic, plasticity induced crackling noise causes large variance in smaller systems. Our simulations uncovered a
crucial parameter for the applicability of our methods, namely the DIC-testing total strain during reloading. While
for a small-reload level of 0.1 % (half of the commonly defined engineering yield stress, found at at engineering
strain 0.2 %), our methods are highly successful, they are clearly not successful one order of magnitude higher, at
1 %. In summary, we demonstrated that the detection of prior plastic deformation is possible through the use of
principal component analysis and the CkNN algorithm, for films wider than 0.5 µm and DIC-testing total strains
smaller than 0.5 %. Our future goals include quantitatively predicting the prior strain (and its uncertainty) from
the test strain, and also applications to actual DIC data from experiments.
The main difference between our simulation and actual DIC data sets is the resolution. Since we obtain direct
strain measurements we assume we have infinite resolution for examining our data set, in contrast to DIC, which
23
has limited resolution. DIC is limited in that regard in a 3-fold way: the average observable distance in experiments,
the smallest distance between DIC markers and the resolution of the machinery that is used.
5. Acknowledgements
We would like to thank D. Wheeler and S. Kalinindi for illuminating discussions. We would like to thank
Marilyn Y. Vasquez Landrove for sharing the CkNN clustering code (Berry and Sauer, 2017), and Erik Van Der
Giessen for insightful comments on our work. We also acknowledge funding from Department of Commerce under
award No. 1007294R (SP).
24
References
W. E. Frazier, Metal additive manufacturing: a review. Journal of Materials Engineering and Performance, 2014, 23(6), 1917-1928.
D. Bigoni and T. Hueckel, Uniqueness and localization –I. Associative and non-associative elastoplasticity. Int. J. Solids and Structures, 1991,
28 (2), 197-213.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas,
A. Passos, D. Cournapeau, M. Brucher, M. Perrot and E. Duchesnay, Scikit-learn: Machine Learning in Python, J. Mach. Learn. Res., 2011,
12, 2825-2830.
C.E. Rasmussen, H. Nickisch, Gaussian Processes for Machine Learming (GMPL) Toolbox, J. Mach. Learn. Res., 2010, 11, 3011-3015.
R. Asaro and V. Lubarda, Mechanics of solids and materials, Cambridge University Press, 2006.
Z. Chen and C.J. Boehlert, Evaluating the Plastic Anisotropy of AZ31 Using Microscopy Techniques, 2013, J. Mater., 65 (9), 1237-1244, doi:
10.1007/s11837-013-0672-6.
Z. Chen, A. Shyam, J. Huang, R.F. Decker, S.E. LeBeau and C.J. Boehlert, The Small Fatigue Crack Growth Behavior of an AM60 Magnesium
Alloy, Metall. Mater. Trans. A, 2013, 44 (2), 1045-1058, doi: 10.1007/s11661-012-1449-1.
J. Shlens, A Tutorial on Principal Component Analysis Derivation, Discussion and Singular Value Decomposition, 2003
Z. Chen, J. Huang, R.F. Decker, S.E. LeBeau, L.R. Walker, O.B. Cavin, T.R. Watkins and C.J. Boehlert, The Effect of Thermomechanical
Processing on the Tensile, Fatigue, and Creep Behavior of Magnesium Alloy AM60, Metall. Mater. Trans. A, 2011, 42 (5), 1386-1399.
V. Khademi, M. Ikeda, C.J. Boehlert, The Effect of Temperature on the Tensile Deformation Behavior of a Beta Titanium Allow,
an In-Situ SEM Study, Proceedings of the 13th World Conference on Titanium, John Wiley & Sons, Inc., 2015, 1109-1116, doi:
10.1002/9781119296126.ch188
V. Khademi, C.J. Boehlert, R.T. Bieler, M. Ikeda, S. Daly, Z. Chen, A correlation between Digital Image Correlation and Grain Misorientation
Distribution Mapping to capture the localized plastic deformation in a polycrystalline titanium alloy, TMS Annual Meeting & Exhibition,
San Diego, CA, USA, 2017.
V. Khademi, R.T. Bieler, M. Ikeda, C.J. Boehlert, Influence of Grain Size and Crystallographic Orientation on Localized Plastic Strain Distribution in Polycrystalline Beta Titanium Alloys, Materials Science & Technology 16, Salt Lake City, UT, USA, 2016
V. Khademi, R.T. Bieler, M. Ikeda, C.J. Boehlert, Quantifying the Local and Global Misorientation Distributions as a Function of Crystallographic Orientation and Level of Plastic Strain in Polycrystalline Materials by Utilizing EBSD Mapping, TMS 16, Nashville, TN, USA,
2016
Z. Chen, A Study of the Effects of Processing and Alloying on the Microstructure and Deformation Behavior of Wrought Magnesium Alloys,
Thesis (Ph.D., Materials Science Engineering) - Michigan State University, 2012.
Z.W. Shan, R.K. Mishra, S.S Asif, O.L. Warren and A.M. Minor, Mechanical annealing and source-limited deformation in submicrometrediameter Ni crystals. Nature materials, 2008, 7 (2), 115-119.
J. D. Carroll, W. Abuzaid, J. Lambros and H. Sehitoglu, High resolution digital image correlation measurements of strain accumulation in
fatigue crack growth, Int. J. Fatigue, 2013, 57, 140-150.
S.R. Kalidindi,J.A. Gomberg, Z.T. Trautt and C.A. Becker, Application of data science tools to quantify and distinguish between structures and
models in molecular dynamics datasets, Nanotechnology, 2015, 26 (34), 344006, doi: 10.1088/0957-4484/26/34/344006
S.R. Kalidindi, Computationally-Efficient Fully-Coupled Multi-Scale Modeling of Materials Phenomena Using Calibrated Localization Linkages, ISRN Materials Science, 2012, Article ID 305692.
T.N. Fast and S.R. Kalidindi, Formulation and Calibration of Higher-Order Elastic Localization Relationships Using the MKS Approach, Acta
Materialia, 2011, 59 (11), 4595-4605.
T.N. Fast, Developing higher-order materials knowledge systems, Thesis (PhD, Materials engineering) Drexel University, 2011, doi:1860/4057.
D. Wheeler , D. Brough , T. Fast, S.R. Kalidindi and A. Reid, PyMKS: Materials Knowledge System in Python, 2014.
D.C. Mattis and R. Swendsen, Statistical Mechanics Made Simple, 2nd Edition, World Scientific Publishing Co. Pte. Ltd., 2008
S.R. Niezgoda, D.T. Fullwood and S.R. Kalidindi, Delineation of the Space of 2-Point Correlations in a Composite Material System, Acta
Materialia, 2008, 56 (18), 52855292.
S.R. Niezgoda, D.M. Turner, D.T. Fullwood and S.R. Kalidindi, Optimized Structure Based Representative Volume Element Sets Reflecting
the Ensemble-Averaged 2-Point Statistics, Acta Materialia, 2010, 58 (13), 44324445.
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach (2nd ed.) 1995, Prentice Hall. ISBN 978-01379039
M.N. Wernick, Y. Yang, J.G. Brankov, G. Yourganov and S.C. Strother, Machine Learning in Medical Imaging, IEEE Signal Processing Mag.,
2010, 27 (4), pp. 2538.
G. James, D. Witten, T. Hastie and R. Tibshirani, An Introduction to Statistical Learning, 2013, New York: Springer
J.H. Friedman, Data Mining and Statistics: What’s the connection?, 29th Symposium on the interface, 1998.
Z.Q. Wang, N.M. Ghoniem, S.Swaminarayan and R. LeSar, A parallel algorithm for 3D dislocation dynamics, J. Comput. Phys., 2006, 219 (2),
608621.
R. Asaro and V. Lubarda, Mechanics of Solids and Materials, Cambridge: Cambridge University Press, 2006.
C. Ayas, V.S. Deshpande, and M. Geers, Tensile response of passivated films with climb-assisted dislocation glide, J. Mech. Phys. Solids, 2012,
60, 1626-1643.
K. Danas and V.S. Deshpande, Plane-strain discrete dislocation plasticity with climb-assisted glide motion of dislocations, Model. Simul.
Mater. Sci. Eng., 2013, 21(4).
K.M Davoudi, L. Nicola and J.J. Vlassak, Bauschinger effect in thin metal films: Discrete dislocation dynamics study, J. Applied Phys., 2014,
115 (1).
L. Nicola, Y.Xiang, J.J. Vlassak, E. Van der Giessen and A. Needleman, Plastic deformation of freestanding thin films: experiments and
modeling, J. Mech. Phys. Solids, 2006, 54 (10), 2089-2110.
25
S.S. Shishvan, S. Mohammadi, M. Rahimian and E. Van der Giessen, Plane-strain discrete dislocation plasticity incorporating anisotropic
elasticity, Int. J. Solids and Structures, 2011, 48 (2), 374-387.
D.S. Balint,V.S Deshpande, A. Needleman and E. Van der Giessen, Size effects in uniaxial deformation of single and polycrystals: a discrete
dislocation plasticity analysis, Model. Simul. Mater. Sci. Eng., 2006, 14, 409-422.
D.T. Fullwood, S.R. Kalidindi, and B.L. Adams, Second - Order Microstructure Sensitive Design Using 2-Point Spatial Correlations, Chapter
12 in Electron Backscatter Diffraction in Materials Science, 2nd Edition , Eds. A. Schwartz, M. Kumar, B. Adams, D. Field, Springer, NY,
2009.
Mitchell T., The discipline of machine learning, Carnegie Mellon University, 2006, (Technical Report CMUML-06-108).
S. Papanikolaou, H. Song and E. Van der Giessen, Obstacles and sources in dislocation dynamics: Strengthening and statistics of abrupt plastic
events in nanopillar compression, J. Mech. Phys. Solids, 2017, 102, 17-29.
J.P. Hirth and J. Lothe, Theory of dislocations, 1982, 2nd ed. (John Wiley and Sons, New York).
E. Van der Giessen and A. Needleman, Discrete dislocation plasticity: a simple planar model, Model. Simul. Mater. Sci. Eng., 1995, 3 (5), 689.
T. Mueller, A.G. Kusne, and R. Ramprasad, Machine Learning in Materials Science, in Reviews in Computational Chemistry, Volume 29 (eds
A. L. Parrill and K. B. Lipkowitz), John Wiley & Sons, Inc, Hoboken, NJ.
P.E Anuta, Spatial registration of multispectral and multitemporal digital imagery using fast Fourier transform techniques, IEEE Trans. Geosci.
Electron., 1970, GE-8, 353368.
T.J. Keating , P.R. Wolf and F.L. Scarpace , An Improved Method of Digital Image Correlation, Photogrammetric Engineering and Remote
Sensing, 1975, 41(8), 9931002.
N. McCormick and J.D. Lord, Digital Image Correlation, Materials Today, 2012, 13(12), 52-54.
S. Roux, J. Rthor and F. Hild, Digital Image Correlation and Fracture: An Advanced Technique for Estimating Stress, J. Phys. D: Applied
Physics, 2009, 42(21).
T. Berry and T. Sauer, Consistent Manifold Representation for Topological Data Analysis, 2017.
P.V. Makarov, S. Schmauder, O.I. Cherepanov, I.Yu. Smolin, V.A. Romanova, R.R. Balokhonov, D.Yu. Saraev, E. Soppa, P. Kizler, G. Fischer,
S. Hu, M. Ludwig, Simulation of elastic-plastic deformation and fracture of materials at micro-, meso- and macrolevels, Theoretical and
Applied Fracture Mechanics, 2001, 37(1-3), 183-244.
R. Baeza-Yates and B. Ribeiro-Neto, 2011, Modern Information Retrieval, Addison Wesley, pp. 327-328.
D.M.W. Powers, Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness & Correlation, J. Mach. Learn. Tech.,
2011, 2(1), 37 63.
W.H. Press, Numerical recipes 3rd edition: The art of scientific computing, Cambridge university press, 2007.
26
Appendix A. Discrete Dislocation Dynamics
In our DDD simulation, plastic flow occurs by the nucleation and glide of edge dislocations, on single or double
slip systems. With the typical Burgers vector of FCC crystals being b = 0.25 nm, we study sample widths ranging in
powers of 2 from w = 0.125µm to 2µm with α = h/w = 4–32. The lateral edges (x = 0, w) are traction free, allowing
dislocations to exit the sample. Loading is taken to be ideally strain-controlled, by prescribing the y-displacement at
the top and bottom edges (y = 0, h). The applied strain rate (for both loading and unloading regimes), ḣ/h = 104 s−1 ,
is held constant across all our simulations, similar to experimental practice. Plastic deformation of the crystalline
samples is described using the discrete dislocation framework for small strains (Van der Giessen and Needleman,
1995), where the determination of the state in the material employs superposition. As each dislocation is treated
as a singularity in a linear elastic background solid with Young’s modulus E and Poisson ratio ν, whose analytic
solution is known at any position, and because the sample does not extend to infinity: the displacement, strain and
stress fields need to be corrected by smooth image fields (denoted by ˆ below) to satisfy boundary conditions at the
edges. Hence, the displacements ui , strains εi j , and stresses σi j are written as
ui = ũi + ûi , εi j = ε̃i j + ε̂i j , σi j = σ̃i j + σ̂i j ,
(A.1)
where the (˜) field is the sum of the fields of all N dislocations in their current positions, i.e.
N
(J)
N
(J)
N
(J)
ũi = ∑ ũi , ε̃i j = ∑ ε̃i j , σ̃i j = ∑ σ̃i j .
J=1
J=1
(A.2)
J=1
Image fields are obtained by solving a linear elastic boundary value problem using finite elements with the boundary conditions changing as the dislocation structure evolves under the application of mechanical load. At the
beginning of the calculation, the crystal is stress free and there are no mobile dislocations. This corresponds to
a well-annealed sample, yet with pinned dislocation segments left that can act either as dislocation sources or as
obstacles. Dislocations are generated from sources when the resolved shear stress τ at the source location is sufficiently high (τnuc ) for a sufficiently long time (tnuc ). We consider bulk dislocation sources and obstacles (Van der
Giessen and Needleman, 1995). A dislocation configuration of one of the simulations, at 10% total strain, is shown
in Fig. A.24(a), together with the xx−component of the total strain (b) and the shear stress (c).
−2
The bulk sources are randomly distributed over slip planes at a density ρbulk
nuc = 60 µm , while their strength
is selected randomly from a Gaussian distribution with mean value τ̄nuc = 50 MPa and 10% standard deviation.
Bulk sources are designed to mimic the Frank-Read mechanism in two dimensions (Hirth and Lothe, 1982), such
that they generate a dipole of dislocations at distance Lnuc , when activated. The initial distance between the two
dislocations in the dipole is
Lnuc =
b
E
,
4π(1 − ν2 ) τnuc
(A.3)
at which the shear stress of one dislocation acting on the other is balanced by the local shear stress. Surface
dislocation sources are successively placed at opposite ends of slip planes, which corresponds to a surface density
sur
of around ρnuc
= 175/ µm. Once a single dislocation is generated from a surface source, it is put at 10b from the
free surface. Under this circumstance, our surface nucleated dislocation has an effective nucleation strength of 312
MPa (Papanikolaou et al., 2017). We only consider glide of dislocations, neglecting the possibility of climb. The
evolution of dislocations is determined by the component of the Peach-Koehler force in the slip direction. For the
Ith dislocation, this is given by
f (I) = n(I) ⋅ (σ̂ + ∑ σ̃(J) ) ⋅ b(I) ,
(A.4)
J≠I
where n(I) is the slip plane normal and b(I) is the Burgers vector of dislocation I. This force will cause the
dislocation I to glide, following over-damped dynamics, with velocity
27
Figure A.24: Boundary plastic steps and stress/strain profiles in dislocation configurations: (a) A dislocation snapshot at 10 % strain for
a sample of width w=2 and double slip systems. A high density of dislocation is observed, close to 1015 /m2 . T’s and ’s have been used to
describe the dislocations that create positive stress (T’s) and negative stress (’s) on the different slip systems. (b) The xx -strain profile that
corresponds to the dislocation configuration of (a) is shown in the displaced coordinates. The boundary is filled with rough plastic steps. The
dashed rectangular box denotes the region for which we consider spatial strain correlations, that are then used for ML purposes. The strain map
is unitless. (c) The shear stress τ xy of the configuration in (a,b) is shown. The stress map is in MPa.
v(I) =
f (I)
,
B
(A.5)
where B is the drag coefficient. In this paper, its value is taken as B = 10−4 Pa s, which is representative for
aluminum. Each sample contains a random distribution of forest dislocation obstacles and surface dislocation
sources, as well as a random distribution of bulk dislocation sources. Once nucleated, dislocations can either
exit the sample through the traction-free sides, annihilate with a dislocation of opposite sign when their mutual
distance is less than 6b, or become pinned at an obstacle. Point obstacles are included to account for the effect
of blocked slip caused by precipitates and forest dislocations on out-of-plane slip systems that are not explicitly
described. They are randomly distributed over slip planes with a constant density that corresponds on average, to
one source, either surface or bulk, for every 8 randomly-distributed obstacles. In this way the densities of sources
and obstacles remains the same as the sample dimensions change, but there is a statistical preference towards
always accompanying sources with obstacles in order to avoid statistical outlier behaviors. We model obstacles in
a simple way where a dislocation stays stays pinned until its Peach-Koehler force exceeds the obstacle-dependent
value τobs b. The strength of the obstacles τobs is taken to be 300 MPa with 20 % standard deviation. Our simulations
are carried out for material parameters that are reminiscent of aluminum: E = 70 GPa,ν = 0.33. We consider 50
random realizations for each parameter case.
28
The simulation is carried out in an incremental manner, using a time step that is a factor 20 smaller than the
nucleation time tnuc = 10 ns. At the beginning of every time increment, nucleation, annihilation, pinning at and
release from obstacle sites are evaluated. After updating the dislocation structure, the new stress field in the sample
is determined, using the finite element method to solve for the image fields (Van der Giessen and Needleman,
1995).
Appendix B. Microstructural discretization and N-point correlation statistics
To avoid confusion, we will be referring to the strain fields as the “microstructure”, following Kalidindi’s terminology (Kalidindi, 2012). However, it is clear that the complete microstructure should also include the material
and all of its properties that go into determining the strain field. The discretization of our microstructure is the separation of the local strain information into various stages (low,intermediate and high local strains). The continuous
local state variable h, the local state space H and the microstructure function m(h, x) are used to represent a single
microstructure in a digital format. The local state space H can be thought of as the complete collection of all the
necessary state variables that are needed to uniquely define the material structure at a given location. The local
state variable h is one point in the local state space, or one configuration of state variables. In our simulations,
we consider the isotropic strain measure φ which takes the role of the continuous local state variable φ ≡ h. The
microstructure µ(x) can be described by a distribution m(h, x), the microstructure function, for a local state h at a
position x:
µ(x) = ∫ hm(h, x)dh .
(B.1)
H
The domain of a microstructure is discretized in both real and state space. Binning in real space follows
naturally from the DDD simulation method, dividing the total volume into S rectangular bins, neglecting bins that
are too close to the boundary (see Appendix A– Fig. A.24). The local state space H can be binned by expanding
m in a set of basis functions. We use the primitive basis function from PyMKS, Λl (h), which has a triangular form
Λl (h) = max(1 − ∣
h − hl
∣ , 0)
δh
(B.2)
where δh is the bin width. For linear binning of H into L − 1 bins, δh = H/(L − 1), with any bin defined by the
edges l ∈ [1, L] with values hl−1 , hl . The discretized microstructure function is then
m[l, s] =
1
∫ ∫ Λl (h)m(h, x)dxdh
∆x H s
(B.3)
in spatial bin s and state bin l. In the above expressions, variables in round brackets are continuous and variables
in square brackets are discrete.
A microstructure function discretized with this basis is subject to the constraint
L
∑ m[l, s] = 1
(B.4)
l=1
which is equivalent to saying that every location (spatial cell) is filled with some configuration of local states.
N-point spatial correlations provide a way to quantify material structure, using statistics. 1-point statistics
is based on the probability that a specified local state will be found in any randomly selected spatial bin in a
microstructure (Niezqoda et al., 2008; Fullwood et al., 2009; Niezqoda et al., 2010). 1-point statistics computes
the volume fractions of the local states in the microstructure.
f [l] =
1
∑ m[s, l]
S s
29
(B.5)
In this equation, f [l] is the probability of finding the local state l in any randomly selected spatial bin in the
microstructure, m[s, l] is the microstructure function (the digital representation of the microstructure), S is the total
number of spatial cells in the microstructure and s is a specific spatial cell.
2-point spatial correlations (also known as 2-point statistics) contain information about the fractions of local
states as well as the first order information on how the different local states are distributed in the microstructure.
2-point statistics is based upon the probability of having a vector placed randomly in the microstructure with one
end of the vector be on one specified local state and the other end on another specified local state. This vector could
have any length or orientation that the discrete microstructure allows. The equation for 2-point statistics reads:
f [r∣l, l′ ] =
1
′
∑ m[s, l]m[s + r, l ]
S s
(B.6)
In this equation f [r∣l, l′ ] is the conditional probability of finding the local states l and l′ at a distance and
orientation away from each other defined by the vector r. When the 2 local states are the same l = l′ , the correlation
is called an auto-correlation. If the 2 local states are not the same, it is a cross-correlation. Our algorithm combines
2-point spatial correlations and the primitive basis function. Given our microstructure and through simple one line
commands of PyMKS we find the correlations we want in the spatial bins. The relevant information of auto and
cross-correlations are stored into matrices for further processing. The discretization of the microstructure into the
arbitrary chosen local states allows for multiple cross and auto-correlations.
In our case we discretize our microstructure in 3 different bins, with 3 possible local states, at low, intermediate
and high local strains. These local states create the following possible correlations: let ω1 be local state 1, ω2
local state 2 and ω3 local state 3. We have (ω1 ,ω1 ) auto-correlations, (ω2 ,ω2 ) and (ω3 ,ω3 ) auto-correlations.
Furthermore, it is also possible to observe (ω1 ,ω2 ), (ω1 ,ω3 ) and (ω2 ,ω3 ) cross-correlations.
30
Appendix C. Results
Appendix C.1. Distinguishing plasticity regimes for small DIC-testing strain (0.1 %)
This section of the Appendix contains results that correspond to Section 3.3 of the main text. Figs. C.25 and
C.26 show results for samples with w = 1, w = 0.25 and w = 0.125 µm. In Fig. C.26 (b) and (d) we show how
a correct classification of the samples should look like, in contrast to (a) and (c), which show the results of our
classification scheme.
Figure C.25: w = 1 µm – 2D projection of PCA results for thin films – Double slip system: ω1 ,ω1 auto-correlation. The colors follow
the definition of Fig. 7. (a) Projection of data set on first two principal components. (b) First principal component of PCA, shown in sample
coordinates (Fig. 3). (c) Second principal component of PCA, shown in sample coordinates (Fig. 3). For description of colormaps, see Fig. 7.
Figure C.26: w = 0.25 µm, w = 0.125 µm – 2D projection of PCA results for thin films – Double slip system : (ω1 ,ω1 ) auto-correlations.
The colors follow the definition of Fig. 7. (a) Projection of data set on first two principal components with a clustering algorithm applied to
the data set. w = 0.25 µm. (b) Projection of data set on first two principal components without a clustering algorithm applied to the data
set. w = 0.25 µm. Actual representation of initial deformations. (c) Projection of data set on first two principal components with a clustering
algorithm applied to the data set. w = 0.125 µm. (d) Projection of data set on first two principal components without a clustering algorithm
applied to the data set. Actual representation of initial deformations. w = 0.125 µm.
31
Appendix C.2. Distinguishing plasticity regimes for single slip samples
This section contains results from simulations of single slip systems. In particular, we show the results for
w = 0.25 and w = 1 µm to emphasize the likeness between results of single slip and double slip systems.
Figure C.27: w = 0.25 µm – 2D projection of PCA results for thin films – Single slip system: ω1 ,ω1 auto-correlation. The colors follow the
definition of Fig. 7. (a) Projection of data set on first two principal components with a clustering applied to the data set, demonstrating failure
in clustering. (b) Projection of data set on first two principal components without a clustering algorithm applied to the data set, justifying the
failure of clustering in (a).
Figure C.28: w = 1 µm – 2D projection of PCA results for thin films – Single slip system: ω1 ,ω1 auto-correlation. The colors follow the
definition of Fig. 7. Projection of data set on first two principal components with a clustering applied to the data set, demonstrating failure in
clustering.
32
Appendix C.3. Distinguishing plasticity regimes for large DIC-testing strain (1 %)
This section contains a direct connection between small and large DIC-testing strain when reloading the samples. In Figs. C.29 (a) we show the results for sample of w = 0.5 µm when we use our classifier. In (b) we show
how a correct classification should look like for the same samples, when reloading to 0.1 % DIC-testing strain. In
contrast, Figs. C.29 (c) and (d) show samples when reloading to 1 % DIC-testing strain. While the sample distinction is applied with (c) showing classified data and (d) showing how a correct classification should be, we can also
observe that due to the high reload strain value, we added plastic memory to the samples. Thus, we completely
changed the PCA projections of our samples.
Figure C.29: Large-reload vs. small-reload DIC-testing – Example of PCA projection results for thin films of w = 0.25 µm:. ω1 ,ω1
auto-correlations. The colors follow the definition of Fig. 7. (a) Small DIC-test reloaded state strain (0.1 %), with clustering algorithm applied
to data set (b) Small DIC-test reloaded state strain (0.1 %), without a clustering algorithm applied to data set. Actual representation of data
set. (c) Large DIC-test reloaded state strain (1 %) with a clustering algorithm applied to the data set. (d) Large DIC-test reloaded state (1 %)
without a clustering algorithm applied to the data set. Actual representation of the data set.
33
Appendix C.4. Dependence of unsupervised learning capacity on pre-processing aspects.
The final section of this appendix, shows more results for different preprocessing aspects. It corresponds to
Section 3.8 of the main text. Fig. C.30 shows the classified results different types of auto-correlations, for samples
of w=1 and w=2 µm. Fig. C.31 shows results when using the plastic strain as the microstructural state variable
(see Section 3.8) and applying the correlation statistics through the φ isotropic measurement for the local strain or
the J2 invariant.
Figure C.30: Auto-correlations vs. cross-correlations for pre-processing – More examples of PCA projection maps for w = 1, 2 µm: The
colors follow the definition of Fig. 7. (a) ω1 ,ω1 auto-correlations. w = 2 µm (b) ω2 ,ω2 auto-correlations. w = 2 µm (c) ω1 ,ω1 auto-correlations.
w = 1 µm (d) ω2 ,ω2 auto-correlations. w = 1 µm
34
Figure C.31: Combined plastic strain calculation and 2nd strain-tensor invariant – Examples of PCA projection maps: The colors follow
the definition of Fig. 7. (a) ω1 ,ω1 auto-correlations. w=2 µm. Isotropic measurement. (b) ω1 ,ω1 auto-correlations. w = 2 µm. J2 invariant. (c)
ω1 ,ω1 auto-correlations. w = 1 µm. Isotropic measurement. (d) ω1 ,ω1 auto-correlations. w = 1 µm. J2 invariant.
35
| 5cs.CE
|
Art of singular vectors and universal adversarial perturbations
Ivan Oseledets
Skolkovo Institute of Science and Technology
Institute of Numerical Mathematics of the
Russian Academy of Sciences
119333, Gubkina St. 8
Moscow, Russia
arXiv:1709.03582v2 [cs.CV] 20 Nov 2017
Valentin Khrulkov
Skolkovo Institute of Science and Technology
143025, Nobel St. 3, Skolkovo Innovation Center
Moscow, Russia
[email protected]
[email protected]
Abstract
tion has been studied in [10, 8]. This also motivated the
research on defenses against various kinds of attack strategies [16, 5].
In the recent work Moosavi et al. [14] have shown that
there exist universal adversarial perturbations — imageagnostic perturbations that cause most natural images to be
misclassified. They were constructed by iterating over a
dataset and recomputing the ”worst” direction in the space
of images by solving an optimization problem related to geometry of the decision boundary. Universal adversarial perturbations exhibit many interesting properties such as their
universality across networks, which means that a perturbation constructed using one DNN will perform relatively well
for other DNNs.
We present a new algorithm for constructing universal
perturbations based on solving simple optimization problems which correspond to finding the so-called (p, q)singular vector of the Jacobian matrices of feature maps of
a DNN. Our idea as based on the observation that since the
norm of adversarial perturbations is typically very small,
perturbations in the non-linear maps computed by the DNN
can be reasonably well approximated by the Jacobian matrix. The (p, q)-singular vector of a matrix A is defined as
the solution of the following optimization problem
Vulnerability of Deep Neural Networks (DNNs) to adversarial attacks has been attracting a lot of attention in
recent studies. It has been shown that for many state of the
art DNNs performing image classification there exist universal adversarial perturbations — image-agnostic perturbations mere addition of which to natural images with high
probability leads to their misclassification. In this work we
propose a new algorithm for constructing such universal
perturbations. Our approach is based on computing the socalled (p, q)-singular vectors of the Jacobian matrices of
hidden layers of a network. Resulting perturbations present
interesting visual patterns, and by using only 64 images we
were able to construct universal perturbations with more
than 60 % fooling rate on the dataset consisting of 50000
images. We also investigate a correlation between the maximal singular value of the Jacobian matrix and the fooling
rate of the corresponding singular vector, and show that the
constructed perturbations generalize across networks.
1. Introduction
Deep Neural Networks (DNNs) with great success have
been applied to many practical problems in computer vision
[11, 20, 9] and in audio and text processing [7, 13, 4]. However, it was discovered that many state-of-the-art DNNs
are vulnerable to adversarial attacks [6, 14, 21], based on
adding a perturbation of a small magnitude to the image.
Such perturbations are carefully constructed in order to
lead to misclassification of the perturbed image and moreover may attempt to force a specific predicted class (targeted attacks), as opposed to just any class different from
the ground truth (untargeted attacks). Potential undesirable
usage of adversarial perturbations in practical applications
such as autonomous driving systems and malware detec-
kAvkq → max,
kvkp = 1,
(1)
and if we desire kvkp = L instead, it is sufficient to multiply
the solution of (1) by L. Universal adversarial perturbations
are typically generated with a bound in the ∞-norm, which
motivates the usage of such general construction. To obtain
the (p, q)-singular vectors we use a modification of the standard power method, which is adapted to arbitrary p-norms.
The main contributions of our paper are
• We propose an algorithm for generating universal adversarial perturbation, using the generalized power
1
method for computing the (p, q)-singular vectors of the
Jacobian matrices of the feature maps.
and due to homogeneity of the problem defined by eq. (5)
it is sufficient to solve it for kεkp = 1. The solution of
(5) is defined up to multiplication by −1 and is called the
(p, q)-singular vector of Ji (x). Its computation in a general
case is the well-known problem [2, 3]. In several cases e.g.
p = ∞, q = ∞ and p = 1, q = 1, algorithms for finding
the exact solution of the problem (5) are known [22], and
are based on finding the element of maximal absolute value
in each row (column) of a matrix. However, this approach
requires iterating over all elements of the matrix A and thus
has complexity O(nm) for the matrix of size n × m. Typical size of such a matrix appearing in our setting, e.g. taking VGG-19 network, output of the first pooling layer, and
batch size of 64 (usage of a batch of images is explained
further in the text), would be 9633792 × 802816, which requires roughly 30 TB of memory to store and makes these
algorithms completely impractical. In order to avoid these
problems, we switch to iterative methods. Instead of evaluating and storing the full matrix A we use only the matvec
function of A, which is the function that given an input vector v computes an ordinary product Av without forming the
full matrix A, and typically has O(n) complexity. In many
applications that deal with extremely large matrices using
matvec functions is essentially mandatory.
For computing the (p, q)-singular vectors there exists a
well-known Power Method algorithm originally developed
by Boyd [3], which we explain in the next section. We also
present a modification of this method in order to construct
universal adversarial perturbations.
• Our method is able to produce relatively good universal adversarial examples from a relatively small number of images from a dataset.
• We investigate a correlation between the largest (p, q)singular value and the fooling rate of the generated
adversarial examples; this suggests that this singular
value can be used as a quantitative measure of the robustness of a given neural network and can be in principle incorporated as the regularizer for the DNNs.
• We analyze various properties of the computed adversarial perturbations such as generalization across networks and dependence of the fooling rate on the number of images used for construction of the perturbation.
2. Problem statement
Suppose that we have a standard feed-forward DNN
which takes a vector x as the input, and outputs a vector
of probabilities p(x) for the class labels. Our goal given parameters q ≥ 1 and L > 0 is to produce a vector ε such
that
arg max p(x) 6= arg max p(x + ε),
kεkq = L,
(2)
for as many x in a dataset as possible. Efficiency of a given
universal adversarial perturbation ε for the dataset X of the
size N is called the fooling rate and is defined as
|{x ∈ X : arg max p(x) 6= arg max p(x + ε)}|
.
N
3. Generalized power method
(3)
Suppose that for some linear map A we are given the
matvec functions of A and A> . Given parameter r ≥ 1 we
also define a function
Let us denote the outputs of the i-th hidden layer of the
network by fi (x). Then for a small vector ε we have
ψr (x) = sign x|x|r−1 ,
fi (x + ε) − fi (x) ≈ Ji (x)ε,
where
Ji (x) =
∂fi
∂x
which applies to vectors element-wise. As usual for r ≥ 1
we also define r0 such that 1r + r10 = 1. Then, given some
initial condition x, one can apply the following algorithm 1
to obtain a solution of (5). In the case p = q = 2 it becomes
the familiar power method for obtaining the largest eigenvalue and the corresponding eigenvector, applied to the matrix A> A.
The discussion so far applies to finding an adversarial perturbation for an instance x. To produce universal adversarial perturbation we would like to maximize the left-hand
size of (5) uniformly across all the images in the dataset X.
For this we introduce a new optimization problem
X
kJi (xj )εkqq → max, kεkp = L.
(7)
,
x
is the Jacobian matrix of fi . Thus, for any q-norm
kfi (x + ε) − fi (x)kq ≈ kJi (x)εkq ,
(4)
We can conclude that for perturbations which are small in
magnitude in order to sufficiently perturb the output of a
hidden layer, it is sufficient to maximize right-hand side of
the eq. (4). It seems reasonable to suggest that while propagating further in the network it will dramatically change the
predicted label of x.
Thus to construct an adversarial perturbation for an individual image x we need to solve
kJi (x)εkq → max,
kεkp = L,
(6)
xj ∈X
A solution of the problem defined by eq. (7) uniformly perturbs the output of the i-th layer of the DNN, and thus can
(5)
2
Algorithm 2 Stochastic Power Method for generating universal adversarial perturbations
Algorithm 1 Power method for generating the (p, q)singular vectors of a linear map A
1:
2:
3:
4:
5:
6:
7:
8:
Inputs: initial condition x, the matvec functions of A
and A>
x
. (p, q)-singular vector
x ← kxk
p
s ← kAxkq
. (p, q)-singular value
while not converged do
Sx ← ψp0 (A> ψq (Ax))
Sx
x ← kSxk
p
s ← kAxkq
return x, s
1:
2:
3:
4:
5:
6:
power method to compute ”more general” singular vectors.
However in our experiments we discovered that it almost
does not affect the fooling rate of the generated universal
perturbation.
serve as the universal adversarial perturbation due to the
reasons discussed in the introduction. Note that the problem given in eq. (7) is exactly equivalent to
kJi εkq → max,
Inputs: a batch of images Xb = {x1 , x2 , . . . xb }, fi (x)
- fixed hidden layer of the DNN
for xj ∈ X do
Construct the matvec functions of Ji (xj ) and
Ji> (xj )
Construct the matvec functions of Ji (Xb ) and Ji> (Xb )
defined in eq. (9)
Run algorithm 1 with desired p and q
return ε
. the universal perturbation
kεkp = L,
5. Efficient implementation of the matvec functions
where Ji is the matrix obtained via stacking Ji (xj ) vertically for each xj ∈ X. To make this optimization problem
tractable, we apply the same procedure to some randomly
chosen subset of images (batch) Xb ⊂ X, obtaining
X
kJi (xj )εkqq → max, kεkp = L,
(8)
Matrices involved in algorithm 2 for typical DNNs are
too large to be formed explicitly. However, using automatic
differentiation available in most deep learning packages it is
possible to construct matvec functions which then are evaluated in a fraction of a second. To compute the matvecs we
follow the well-known approach based on Pearlmutter’s Roperator [17], which could be briefly explained as follows.
Suppose that we are given an operation gradx [f ](x0 ) which
computes the gradient of a scalar function f (x) with respect
to the vector variable x at the point x0 . Let fi (x) be some
fixed layer of the DNN, such that x ∈ Rn and fi (x) ∈ Rm ,
thus Ji (x) ∈ Rm×n and for vectors v1 ∈ Rn , v2 ∈ Rm
we would like to compute Ji (x)v1 , Ji> (x)v2 at some fixed
point x. These steps are presented in algorithm 3.
For a given batch of images this algorithm is run only once.
xj ∈Xb
and hypothesize that the obtained solution will be a good
approximate to the exact solution of (7). We present this
approach in more detail in the next section.
4. Stochastic power method
Let us choose a fixed batch of images Xb =
{x1 , x2 . . . xb } from the dataset and fix a hidden layer of the
DNN, defining the map fi (x). Denote sizes of x, fi (x) by n
and m correspondingly. Then, using the notation from section 2 we can compute Ji (xj ) ∈ Rm×n for each xj ∈ Xb .
Let us now stack these Jacobian matrices vertically obtaining the matrix Ji (Xb ) of size bm × n:
Ji (x1 )
Ji (x2 )
Ji (Xb ) =
(9)
... .
Ji (xb )
Algorithm 3 Constructing the matvec functions of the Jacobian matrix of a hidden layer of a DNN
1:
2:
3:
Note that to compute the matvec functions of Ji (Xb ) and
Ji> (Xb ) it suffices to be able to compute the individual
matvec functions of Ji (x) and Ji> (x). We will present an
algorithm for that in the next section and for now let us assume that these matvec functions are given. We can now
apply algorithm 1 to the matrix Ji (Xb ) obtaining Stochastic Power Method (SPM). Note that in algorithm 2 we could
in principle change the batch Xb between iterations of the
4:
5:
Inputs :v1 ∈ Rn , v2 ∈ Rm - vectors to compute the
matvec functions of, fi (x) - fixed hidden layer of the
DNN
Ji> (x)v2 ← gradv [hv, fi (x)i](v2 )
g(v2 ) ← hJi> (x)v2 , v1 i
Ji (x)v1 ← gradv2 [g](0m )
return Ji (x)v1 , Ji> (x)v2
Let us summarize our approach for generating universal
perturbations. Suppose that we have some dataset of natural images X and a fixed deep neural network trained to perform image classification. At first we choose a fixed random
3
block2_conv1
0
block2_conv2
0
block2_pool
0
25
25
25
50
50
50
50
75
75
75
75
100
100
100
100
125
125
125
125
150
150
150
150
175
175
175
175
200
200
200
0
25
50
75
100
125
150
175
200
0
25
50
75
100
125
150
175
200
block3_conv1
0
25
200
0
25
50
75
100
125
150
175
200
0
25
50
75
100
125
150
175
200
150
175
200
150
175
200
(a) VGG-16
block2_conv1
0
block2_conv2
0
block2_pool
0
25
25
25
50
50
50
50
75
75
75
75
100
100
100
100
125
125
125
125
150
150
150
150
175
175
175
175
200
200
200
0
25
50
75
100
125
150
175
200
0
25
50
75
100
125
150
175
200
block3_conv1
0
25
200
0
25
50
75
100
125
150
175
200
0
25
50
75
100
125
(b) VGG-19
conv1
0
res2a_branch2a
0
bn2a_branch2a
0
25
25
25
50
50
50
50
75
75
75
75
100
100
100
100
125
125
125
125
150
150
150
150
175
175
175
175
200
200
200
0
25
50
75
100
125
150
175
200
0
25
50
75
100
125
150
175
200
activation_476
0
25
200
0
25
50
75
100
125
150
175
200
0
25
50
75
100
125
(c) ResNet50
Figure 1: Universal adversarial perturbations constructed using various layers of various DNNs.
batch of images Xb from X and specify a hidden layer of
the DNN. Then using algorithm 3 we construct the matvec
functions of the matrix defined by eq. (9). Finally we run
algorithm 2 to obtain the perturbation and then rescale it if
necessary.
sible interpretation of these patterns can be given if we note
that extremely similar images were computed in [15] in relation to feature visualization. Namely, for various layers
in GoogLeNet [20] the images which activate a particular neuron were computed. In particular, visualization of
the layer conv2d0, which corresponds to edge detection,
looks surprisingly similar to several of our adversarial perturbations. Informally speaking, this might indicate that
adversarial perturbations constructed as the (p, q)-singular
vectors attack a network by ruining a certain level of image
understanding, where in particular first layers correspond to
edge detection. This is partly supported by the fact that the
approach used for feature visualization in [15] is based on
computing the Jacobian matrix of a hidden layer and maximizing the response of a fixed neuron, which is in spirit
related to our method.
To measure how strongly the (p, q)-singular vector disturbs
the output of the hidden layer based on which it was constructed, we evaluate the corresponding singular value. We
have computed it for all the layers of VGG-16, VGG-19 and
ResNet50. Results are given in fig. 2. Note that in general
singular values of the layers of ResNet50 are much smaller
in magnitude than those of the VGG nets, which is further
shown to roughly correspond to the obtained fooling rates.
Convergence of algorithm 2 is analyzed in fig. 3. We ob-
6. Experiments
In this section we analyze various adversarial perturbations constructed as discussed in section 5. For testing
purposes we use the ILSVRC 2012 validation dataset [18]
(50000 images).
6.1. Adversarial perturbations
In our experiments we chose p = ∞, q = 10 and
computed the (p, q) - singular vectors for various layers
of VGG-16 and VGG-19 [19] and ResNet50 [9]. q = 10
was chosen to smoothen optimization problem and effectively serves as the replacement for q = ∞, for which the
highest fooling rates were reported in [14]. We also investigate other values of q in section 6.3. Batch size in algorithm 2 was chosen to be 64 and we used the same 64 images to construct all the adversarial perturbations. Some
of the computed singular vectors are presented in figs. 1a
to 1c. We observe that computed singular vectors look visually appealing and present interesting visual patterns. Pos4
12000
300
6000
10000
250
Singular value
5000
8000
200
6000
150
2000
4000
100
1000
2000
50
4000
3000
0
2.5
5.0
7.5
10.0
12.5
15.0
0
0.0
17.5
Layer number
0
2.5
5.0
7.5
10.0
12.5
15.0
17.5
20.0
0
25
50
Layer number
(a) VGG-16
75
100
125
150
175
Layer number
(b) VGG-19
(c) ResNet50
Figure 2: (p, q)-singular values for all the layers of various DNNs. Values p = ∞, q = 10 were used.
quired roughly 3000 images to achieve (see section 6.4 for
further comparison). Examples of images after addition of
the adversarial perturbation with the highest fooling rate for
VGG-19 are given in fig. 6, and their predicted classes for
various adversarial attacks (for each network we choose the
adversarial perturbation with the highest fooling rate) are
reported in tables 5 to 7. We note that the top-1 class probability for images after the adversarial attack is relatively low
in most cases, which might indicate that images are moved
away from the decision boundary. We test this behavior by
computing the top-5 probabilities for several values of the
∞-norm of the adversarial perturbation. Results are given
in fig. 4. We see that top-1 probability decreases significantly and becomes roughly equal to the top-2 probability.
Similar behavior was noticed in some of the cases when the
adversarial example failed to fool the DNN — top-1 probability still has decreased significantly. It is also interesting
to note that such adversarial attack indeed introduces many
new edges in the image, which supports the claim made in
the previous section.
serve that a relatively low number of iterations is required
to achieve good accuracy. In particular if each evaluation of
the matvec functions takes O(n) operations, the total complexity is O(dn) for d iterations, which for d as small as
60 is a big improvement compared to O(n2 ) of the exact
algorithm.
100
10−1
Relative error
10−2
10−3
10−4
VGG-19 block2 pool
VGG-19 block3 conv1
ResNet50 conv1
ResNet50 activation 103
ResNet50 activation 109
10−5
10−6
10−7
0
10
20
30
40
50
60
Iteration number
Figure 3: Convergence of the (∞, 10)-singular value. Relative error of the singular value w.r.t iteration number is
shown.
Top-1
Top-2
Top-3
Top-4
Top-5
0.6
p(x)
6.2. Fooling rate and singular values
As a next experiment we computed and compared the
fooling rate of the perturbation by various computed singular vectors. We choose 1 kεk∞ = 10 – recall that this can be
achieved just by multiplying the computed singular vector
by the factor of 10. Results are given in tables 1 to 3.
We see that using only 64 images allowed us to achieve
more than 40% fooling rate for all the investigated networks
on the dataset containing 50000 images of 1000 different
classes. This means that by analyzing less than 0.15% of
the dataset it is possible to design strong universal adversarial attacks generalizing to many unseen classes and images. Similar fooling rates reported in [14, Figure 6] re-
0.4
0.2
0.0
0
5
10
15
20
25
kεk∞
Figure 4: Top-5 probabilities predicted by the VGG-19 w.r.t
∞-norm of the universal adversarial perturbation. Tests
were run for image 3 from fig. 6. Universal adversarial perturbation with highest fooling rate in table 2 was chosen.
As a next experiment we investigate the dependence of the
achieved fooling rate on the batch size used in algorithm 2.
Some of the results are given in fig. 5. Surprisingly, in-
1 Pixels
in the images from the dataset are normalized to be in [0, 255]
range, so by choosing kεk∞ = 10 we make the adversarial perturbations
quasi-imperceptible to human eye
5
Layer name
block2 pool
block3 conv1
block3 conv2
block3 conv3
1165.74
0.52
2200.08
0.39
3146.66
0.50
6282.64
0.50
Singular value
Fooling rate
Table 1: Fooling rates for VGG-16
Layer name
block2 pool
block3 conv1
block3 conv2
block3 conv3
784.82
0.60
1274.99
0.33
1600.77
0.50
3063.72
0.52
Singular value
Fooling rate
Table 2: Fooling rates for VGG-19
Layer name
conv1
res3c branch2a
bn5a branch2c
activation 8
Singular value
Fooling rate
59.69
0.44
19.21
0.35
138.81
0.34
15.55
0.34
Table 3: Fooling rates for ResNet50
some subset I of the layers (preferably closer to the input)
of the DNN and include the term
X
kJi (X)kqq ,
Fooling rate
0.60
0.59
i∈I
0.58
in the regularizer, where X indicates the current learning
batch. We plan to analyze this approach in future work.
Finally, we investigate if our adversarial perturbations generalize across different networks. For each DNN we have
chosen the adversarial perturbation with the highest fooling
rate from tables 1 to 3 and tested it against other networks.
Results are given in table 4. We see that these adversarial
perturbations are indeed doubly universal, reasonably well
generalizing to other architectures. Surprisingly, in some
cases the fooling rate of the adversarial perturbation constructed using other network was higher than that of the
network’s own adversarial perturbation. This universality
might be explained by the fact that if Deep Neural Networks
independently of specifics of their architecture indeed learn
to detect low-level patterns such as edges, then adding an
edge-like noise has a high chance to ruin the prediction. It
is interesting to note that the adversarial perturbation obtained using block2 pool layer of VGG-19 is the most
efficient one, in correspondence with its interesting edgelike structure.
0.57
0.56
20
30
40
50
60
70
80
90
100
Batch size
Figure 5: Dependence of the fooling rate on the batch size.
block2 pool layer in VGG-19 was used for the experiment.
creasing the batch size does not significantly affect the fooling rate and by using as few as 16 images it is possible
to construct the adversarial perturbations with 56% fooling
rate. This suggests that the singular vector constructed using Stochastic Power Method reasonably well approximates
solution of the general optimization problem (7).
It appears that higher singular value of the layer does not
necessarily indicate higher fooling rate of the corresponding singular vector. However, as shown on fig. 2 the singular
values of various layers VGG-19 are in general larger than
those of VGG-16, and of VGG-16 are in general larger than
the singular values of ResNet50, which is roughly in correspondence between the maximal fooling rates we obtained
for these networks. Moreover, layers closer to the input of
the DNN seem to produce better adversarial perturbations,
than those closer to the end.
Based on this observation we hypothesize that to defend the
DNN against this kind of adversarial attack one can choose
6.3. Dependence of the perturbation on q
In the analysis so far we have chosen q = 10 as an approximate to q = ∞. However, any value of q can be used
for constructing the adversarial perturbations and in this
subsection we investigate how the choice of q affects the
fooling rate and the generated perturbations (while keeping p = ∞). Perturbations computed for several different
values of q are presented in fig. 7, and the corresponding
6
Figure 6: Examples of images misclassified after the adversarial attack (the attack based on block2 pool layer of VGG-19
is shown). Predicted classes are given in tables 5 to 7.
Figure 7: Adversarial perturbations constructed for various values of q. Presented images correspond to values q uniformly
increasing from 1.0 to 5.0. block2 pool layer of VGG-19 was used.
VGG-19
ResNet50
0.650
0.52
0.48
0.41
0.60
0.60
0.47
0.39
0.38
0.44
0.625
Fooling rate
VGG-16
VGG-19
ResNet50
VGG-16
Table 4: Generalization of the adversarial perturbations
across networks. Columns indicate the DNN for which
the adversarial perturbation was computed, rows indicate
on which network it was tested. Adversarial perturbations
with highest fooling rates in tables 1 to 3 were chosen.
0.600
0.575
0.550
0.525
0.500
2
4
6
8
10
q
Figure 8: Dependence of the fooling rate on the value of q.
As before, p = ∞ and norm of the adversarial perturbation
was set to be 10.
fooling rates are reported in fig. 8. We observe that bigger
values of q produce more clear edge-like patterns, which is
reflected in the increase of the fooling rate. However, the
maximal fooling rate seems to be achieved at q ≈ 5, probably because it is ’smoother’ substitute for q = ∞ than
q = 10, which might be important in such large scale problems.
constructed using these two methods for various batch sizes.
Results are presented in fig. 9. Note that our method indeed captures the universal attack vector relatively fast, and
the fooling rate stabilizes on roughly 63%, while the fooling rate of the perturbation constructed by the UAP method
starts low and then gradually increases as more images are
added. Running time of our algorithm depends on which
hidden layer we use. As an example for block2 pool
layer of VGG-19 the running time per iteration of the power
method for a batch of one image was roughly 0.06 seconds
(one NVIDIA Tesla K80 GPU was used and the algorithm
was implemented using Tensorflow [1] and numpy libraries). Since the running time per iteration linearly de-
6.4. Comparison of the algorithms
In this subsection we perform a comparison of the algorithm presented in Moosavi et al., which we refer to as UAP,
and our method. For the former we use the Python implementation https://github.com/LTS4/universal/. Since one of
the main features of our method is an extremely low number of images used for constructing the perturbation, we decided to compare the fooling rates of universal perturbations
7
p(x)
p(x + ε)
image 1
image 2
image 3
image 4
mashed potato 53.9%
head cabbage 29.4%
pole 37.6%
rubber eraser 39.3%
fountain 55.0%
carousel 61.4%
goblet 12.1%
bucket 36.6%
Table 5: VGG-16
p(x)
p(x + ε)
image 1
image 2
image 3
image 4
mashed potato 68.3%
flatworm 26.5%
flagpole 37.4%
letter opener 19.9%
fountain 74.5%
pillow 20.6%
coffee mug 23.1%
candle 40.3%
Table 6: VGG-19
p(x)
p(x + ε)
image 1
image 2
image 3
image 4
mashed potato 94.2%
stole 21.7%
totem pole 43.1%
fountain pen 27.6%
flagpole 35.3%
monitor 9.76%
chocolate sauce 22.0%
goblet 40.3%
Table 7: ResNet50
7. Related work
pends on the batch size b, the total running time could be
estimated as 0.06bd seconds for d iterations. By fixing
b = 32 and d = 30 we obtain that the total running time
to generate the universal perturbation with approximately
60% fooling rate on the whole dataset is roughly 1 minute
(we did not include the time required to compute the symbolic Jacobian matvecs since it is performed only once, and
is also required in the implementation of UAP, though different layer is used). In our hardware setup the running
time of the UAP algorithm with batch size 128 was approximately 10 minutes, and the fooling rate of roughly 20% was
achieved. According to [14, Figure 6] approximately 3000
images will be required to obtain the fooling rates of order
60%.
Many different methods [6, 14, 10, 21, 12] have been
proposed to perform adversarial attacks on Deep Neural
Networks in the white box setting where the DNN is fully
available to the attacker. Two works are especially relevant for the present paper. Goodfellow et al. [6] propose
the fast gradient sign method, which is based on computing the gradient of the loss function l(x) at some image x
and taking its sign as the adversarial perturbation. This approach allows one to construct rather efficient adversarial
perturbations for individual images and can be seen as a
particular case of our method. Indeed if we take the batch
size to be equal to 1 and the loss function l(x) as the hidden layer, then sign gradx l(x) is exactly the solution of the
problem (5) with p = 1 and L = 1 (since l(x) is just a
number this problem does not depend on q). Second work
is Moosavi et al. [14] where the universal adversarial perturbations have been proposed. It is based on a sequential
solution of nonlinear optimization problems followed by a
projection onto p = ∞ (p = 2) sphere, which iteratively
computes the ’worst’ possible direction towards the decision boundary. Optimization problems proposed in the current work are simpler in nature and well-studied, and due
to their homogeneous property the adversarial perturbation
with an arbitrary norm is obtained by simply rescaling the
once computed perturbation, in contrast with the algorithm
in [14].
0.6
Fooling rate
0.5
0.4
0.3
0.2
SingularFool
UAP
0.1
0.0
20
40
60
80
100
120
Batch size
Figure 9: Dependence of the fooling rate on the number
of images used for constructing the universal perturbation.
SingularFool denotes the method proposed in the current
paper, UAP denotes the algorithm presented in [14]. q = 5
and block2 pool layer of VGG-19 were used.
8. Conclusion
In this work we explored a new algorithm for generating
universal adversarial perturbations and analyzed their main
properties, such as generalization across networks, dependence of the fooling rate on various hyperparameters and
8
having certain visual properties. We have showed that by
using only 64 images a single perturbation fooling the network in roughly 60% cases can be constructed, while the
previous known approach required several thousand of images to obtain such fooling rates. In a future work we plan to
address the relation between feature visualization [15] and
adversarial perturbations, as well as analyzing the defense
approach discussed in section 6.2.
[11] Y. LeCun, Y. Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory
and neural networks, 3361(10):1995, 1995. 1
[12] Y. Liu, X. Chen, C. Liu, and D. Song. Delving into transferable adversarial examples and black-box attacks. arXiv
preprint arXiv:1611.02770, 2016. 8
[13] T. Mikolov, S. Kombrink, L. Burget, J. Černockỳ, and
S. Khudanpur. Extensions of recurrent neural network language model. In Acoustics, Speech and Signal Processing
(ICASSP), 2011 IEEE International Conference on, pages
5528–5531. IEEE, 2011. 1
[14] S.-M. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and
P. Frossard. Universal adversarial perturbations. arXiv
preprint arXiv:1610.08401, 2016. 1, 4, 5, 8
[15] C. Olah, A. Mordvintsev, and L. Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/featurevisualization. 4, 9
[16] N. Papernot, P. McDaniel, X. Wu, S. Jha, and A. Swami.
Distillation as a defense to adversarial perturbations against
deep neural networks. In Security and Privacy (SP), 2016
IEEE Symposium on, pages 582–597. IEEE, 2016. 1
[17] B. A. Pearlmutter. Fast exact multiplication by the hessian.
Neural computation, 6(1):147–160, 1994. 3
[18] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
et al. Imagenet large scale visual recognition challenge.
International Journal of Computer Vision, 115(3):211–252,
2015. 4
[19] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014. 4
[20] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In Proceedings of the
IEEE conference on computer vision and pattern recognition, pages 1–9, 2015. 1, 4
[21] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan,
I. Goodfellow, and R. Fergus. Intriguing properties of neural
networks. arXiv preprint arXiv:1312.6199, 2013. 1, 8
[22] L. N. Trefethen and D. Bau III. Numerical linear algebra,
volume 50. Siam, 1997. 2
Acknowledgements
This study was supported by the Ministry of Education and Science of the Russian Federation (grant
14.756.31.0001), by RFBR grants 16-31-60095-mola-dk, 16-31-00372-mola and by Skoltech NGP program.
References
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen,
C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia,
R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané,
R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster,
J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker,
V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 7
[2] A. Bhaskara and A. Vijayaraghavan. Approximating matrix
p-norms. In Proceedings of the twenty-second annual ACMSIAM symposium on Discrete Algorithms, pages 497–511.
SIAM, 2011. 2
[3] D. W. Boyd. The power method for lp norms. Linear Algebra
and its Applications, 9:95–101, 1974. 2
[4] F. A. Gers, J. Schmidhuber, and F. Cummins. Learning to
forget: Continual prediction with LSTM. 1999. 1
[5] I. Goodfellow, N. Papernot, and P. McDaniel. cleverhans v0.
1: an adversarial machine learning library. arXiv preprint
arXiv:1610.00768, 2016. 1
[6] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. arXiv preprint
arXiv:1412.6572, 2014. 1, 8
[7] A. Graves, A.-r. Mohamed, and G. Hinton. Speech recognition with deep recurrent neural networks. In Acoustics,
speech and signal processing (icassp), 2013 ieee international conference on, pages 6645–6649. IEEE, 2013. 1
[8] K. Grosse, N. Papernot, P. Manoharan, M. Backes, and
P. McDaniel. Adversarial perturbations against deep neural networks for malware classification. arXiv preprint
arXiv:1606.04435, 2016. 1
[9] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
770–778, 2016. 1, 4
[10] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533,
2016. 1, 8
9
| 2cs.AI
|
Weighted Online Problems with Advice
⋆
Joan Boyar, Lene M. Favrholdt, Christian Kudahl, and Jesper W. Mikkelsen
arXiv:1606.05210v2 [cs.DS] 14 Aug 2017
Department of Mathematics and Computer Science, University of Southern Denmark
Abstract. Recently, the first online complexity class, AOC, was introduced. The class consists of many online problems where each request
must be either accepted or rejected, and the aim is to either minimize or
maximize the number of accepted requests, while maintaining a feasible
solution. All AOC-complete problems (including Independent Set, Vertex
Cover, Dominating Set, and Set Cover) have essentially the same advice
complexity. In this paper, we study weighted versions of problems in
AOC, i.e., each request comes with a weight and the aim is to either minimize or maximize the total weight of the accepted requests. In contrast
to the unweighted versions, we show that there is a significant difference
in the advice complexity of complete minimization and maximization
problems. We also show that our algorithmic techniques for dealing with
weighted requests can be extended to work for non-complete AOC problems such as Matching in the edge arrival model (giving better results
than what follow from the general AOC results) and even non-AOC problems such as scheduling.
1
Introduction
An online problem is an optimization problem for which the input is divided into
small pieces, usually called requests, arriving sequentially. An online algorithm
must serve each request, irrevocably, without any knowledge of possible future
requests. The quality of online algorithms is traditionally measured using the
competitive ratio [12, 16], which is essentially the worst case ratio of the online
performance to the performance of an optimal offline algorithm, i.e., an algorithm
that knows the whole input sequence from the beginning and has unlimited
computational power.
For some online problems such as Independent Set or Vertex Cover, the best
possible competitive ratio is linear in the sequence length. This gives rise to the
question of what would happen, if the algorithm knew something about future
requests. Semi-online settings, where it is assumed that the algorithm has some
specific knowledge such as the value of an optimal solution, have been studied
(see [6] for many relevant references). The extra knowledge may also be more
problem specific such as an access graph for paging [4, 8].
⋆
This work was partially supported by the Villum Foundation, grant VKR023219,
and the Danish Council for Independent Research, Natural Sciences, grant DFF1323-00247.
In contrast to problem specific approaches, advice complexity [3, 9, 11] is a
quantitative and standardized way of relaxing the online constraint. The main
idea of advice complexity is to provide an online algorithm, Alg, with some
partial knowledge of the future in the form of advice bits provided by a trusted
oracle which has unlimited computational power and knows the entire request
sequence. Informally, the advice complexity of an algorithm is a function of input
sequence length, and for a given n, it is the maximum number of advice bits read
for input sequences of length n. The advice complexity of a problem is a function
of input sequence length and competitive ratio, and for a given competitive ratio
c, it is the best possible advice complexity of any c-competitive algorithm for
the problem. Advice complexity is formally defined in Section 2.
Upper bounds on the advice complexity for a problem can sometimes lead
to (or come from) semi-online algorithms, and lower bounds can show that such
algorithms do not exist. Since its introduction, advice complexity has been a
very active area of research. Lower and upper bounds on the advice complexity
have been obtained for a large number of online problems; a recent list can be
found in [17]. For a survey on advice complexity, see [6].
Recently in [7], the first complexity class for online problems, AOC, was
introduced. The class consists of online problems that can be described in the
following way: The input is a sequence of requests and each request must either
be accepted or rejected. The set of accepted requests is called the solution. For
each request sequence, there is at least one feasible solution. The class contains
minimization as well as maximization problems. For a minimization problem,
the goal is to accept as few requests as possible, while maintaining a feasible
solution, and for maximization problems, the aim is to accept as many requests
as possible. For minimization problems, any super set of a feasible solution is
also a solution, and for maximization problems, any subset of a feasible solution
is also a feasible solution. The AOC-complete problems are the hardest problems
in the class in terms of their advice complexity. The class AOC is formally defined
in Section 3.
In this paper, we consider a generalization of the problems in the class AOC
in which each request comes with a weight. The goal is now to either minimize or
maximize the total weight of the accepted requests. We separately consider the
classes of maximization and minimization problems. For AOC-complete maximization problems, we get advice complexity results quite similar to those for
the unweighted versions of the problems. On the other hand, for AOC-complete
minimization problems, the results are a lot more negative: using less than one
advice bit per request leads to unbounded competitive ratios, so this gives a
complexity class containing harder problems than AOC. This is in contrast to unweighted AOC-complete problems, where minimization and maximization problems are equally hard in terms of advice complexity. Recently, differences between (unweighted) AOC minimization and maximization problems were found
with respect to online bounded analysis [5] and min- and max-induced subgraph
problems [13].
2
Our upper bound techniques are also useful for non-complete AOC problems
such as Matching in the edge arrival model, as well as non-AOC problems such
as Scheduling.
Previous results. For any AOC-complete problem, Θ(n/c) advice bits are necessary and sufficient to obtain a competitive ratio of c. More specifically, for
competitive ratio c, the advice complexity is B(n, c) ± O(log n), where
(c − 1)c−1
B(n, c) = log 1 +
n,
(1)
cc
and an/c ≤ B(n, c) ≤ n/c, a = 1/(e ln(2)) ≈ 0.53. This is an upper bound on
the advice complexity of all problems in AOC. In [7], a list of problems including
Independent Set, Vertex Cover, Dominating Set, and Set Cover were proven
AOC-complete.
The paper [1] studies a semi-online version of scheduling where it is allowed to keep several parallel schedules and choose the best schedule in the
end. The scheduling problem considered is makespan minimization on m identical machines. Using (1/ε)O(log(1/ε)) parallel schedules, a (4/3 + ε)-competitive
algorithm is obtained. Moreover, a (1 + ε)-competitive algorithm which uses
(m/ε)O(log(1/ε)/ε) parallel schedules is given along with an almost matching lower
bound. Note that keeping s different schedules until the end corresponds to working with s different online algorithms. Thus, this particular semi-online model
easily translates to the advice model, the advice being which of the s algorithms
to run. In this way, the results of [1] correspond to a (4/3 + ε)-competitive
algorithm using O(log2 (1/ε)) advice bits and a (1 + ε)-competitive algorithm using O(log(m/ε) · log(1/ε)/ε) advice bits. In particular, note that this algorithm
uses constant advice in the size of the input and only logarithmic advice in the
number of machines.
In [15], scheduling on identical machines with a more general type of objective
function (including makespan, minimizing the ℓp -norm, and machine covering)
was studied. The paper considers the advice-with-request model where a fixed
number of advice bits are provided along with each request. The main result
is a (1 + ε)-competitive algorithm that uses O((1/ε) · log(1/ε)) advice bits per
request, totaling O((n/ε) · log(1/ε)) bits of advice for the entire sequence.
Our results. We prove that adding arbitrary weights, AOC-complete minimization problems become a lot harder than AOC-complete maximization problems:
– For AOC-complete maximization problems, the weighted version is not significantly harder than the unweighted version: For any maximization problem
in AOC (this includes, e.g., Independent Set), the c-competitive algorithm
given in [7] for the unweighted version of the problem can be converted
into a (1 + ε)c-competitive algorithm for the weighted version using only
O((log2 n)/ε) additional advice bits. Thus, a (1 + ε)c-competitive algorithm
using at most B(n, c) + O((log2 n)/ε) bits of advice is obtained. For the
3
weighted version of non-complete AOC maximization problems, a better advice complexity than B(n, c) may be obtained: For any c-competitive algorithm for an AOC maximization problem, P, using b advice bits can be
converted into a O(c · log n)-competitive algorithm for the weighted version
of P using b + O(log n) advice bits. For Weighted Matching in the edge arrival model, this implies a O(log n)-competitive algorithm reading O(log n)
bits of advice. We show that this is best possible in the following sense:
For a set of weighted AOC problems including Matching, Independent Set
and Clique, no algorithm reading o(log n) bits of advice can have a competitive ratio bounded by any function of n. Furthermore, any O(1)-competitive
algorithm for Matching must read Ω(n) advice bits.
– For all minimization problems known to be AOC-complete (this includes,
e.g., Vertex Cover, Dominating Set, and Set Cover), n − O(log n) bits of
advice are required to obtain a competitive ratio bounded by a function of
n. This should be contrasted with the fact that n bits of advice trivially
yields a strictly 1-competitive algorithm.
If the largest weight wmax cannot be arbitrarily larger than the smallest
weight wmin , the c-competitive algorithm given in [7] for the unweighted version can be converted into a c(1 + ε)-competitive algorithm for the weighted
versions using B(n, c) + O(log2 n + log(log(wmax /wmin )/ε)) advice bits in
total.
Our main upper bound technique is a simple exponential classification scheme
that can be used to sparsify the set of possible weights. This technique can
also be used for problems outside of AOC. For example, for scheduling on related machines, we show that for many important objective functions (including makespan minimization and minimizing the ℓp -norm), there exist (1 + ε)competitive algorithms reading O((log2 n)/ε) bits of advice. For scheduling on
m unrelated machines where m is constant, we get a similar result, but with
O((log n)m+1 /εm ) advice bits. Finally, for unrelated machines, where the goal is
to maximize an objective function, we show that under some mild assumptions
on the objective function (satisfied, for example, for machine covering), there is
a (1 + ε)-competitive algorithm reading O((log n)m+1 /εm ) bits of advice.
For scheduling on related and unrelated machines, our results are the first
non-trivial upper bounds on the advice complexity. For the case of makespan
minimization on identical machines, the algorithm of [1] is strictly better than
ours. However, for minimizing the ℓp -norm or maximizing the minimum load on
identical machines, we exponentially improve the previous best upper bound [15]
(which was linear in n).
2
Preliminaries
Throughout the paper, we let n denote the number of requests in the input.
We let R+ denote the set containing 0 and all positive real numbers. We let log
denote the binary logarithm log2 . For k ≥ 1, [k] = {1, 2, . . . , k}. For any bit
4
string y, let |y|0 and |y|1 denote the number of zeros and the number of ones,
respectively, in y. We write x ⊑ y if for all indices, i, xi = 1 ⇒ yi = 1.
2.1
Advice complexity and competitive analysis
In this paper, we use the “advice-on-tape” model [3]. Before the first request
arrives, the oracle, which knows the entire request sequence, prepares an advice
tape, an infinite binary string. The algorithm Alg may, at any point, read some
bits from the advice tape. The advice complexity of Alg is the maximum number
of bits read by Alg for any input sequence of at most a given length. Opt is an
optimal offline algorithm.
Advice complexity is combined with competitive analysis to determine how
many bits of advice are necessary and sufficient to achieve a given competitive
ratio.
Definition 1 (Competitive analysis [12, 16] and advice complexity [3]).
The input to an online problem, P, is a request sequence σ = hr1 , . . . , rn i. An
online algorithm with advice, Alg, computes the output y = hy1 , . . . , yn i, where
yi is computed from ϕ, r1 , . . . , ri , where ϕ is the content of the advice tape. Each
possible output for P is associated with a cost/profit. For a request sequence σ,
Alg(σ) (Opt(σ)) denotes the cost/profit of the output computed by Alg (Opt)
when serving σ.
If P is a minimization (maximization) problem, then Alg is c(n)-competitive
if there exists a constant, α, such that, for all n ∈ N, Alg(σ) ≤ c(n)·Opt(σ)+α,
(Opt(σ) ≤ c(n) · Alg(σ) + α), for all request sequences, σ, of length at most
n. If the relevant inequality holds with α = 0, we say that Alg is strictly c(n)competitive.
The advice complexity, b(n), of an algorithm, Alg, is the largest number of
bits of ϕ read by Alg over all possible request sequences of length at most n. The
advice complexity of a problem, P, is a function, f (n, c), c ≥ 1, such that the
smallest possible advice complexity of a strictly c-competitive online algorithm
for P is f (n, c).
We only consider deterministic online algorithms (with advice). Note that
both b(n) and c(n) in the above definition may depend on n, but, for ease
of notation, we often write b and c instead of b(n) and c(n). Also, with this
definition, c ≥ 1, for both minimization and maximization problems.
2.2
Complexity classes
In this paper, we consider the complexity class AOC from [7].
Definition 2 (AOC [7]). A problem, P, is in AOC (Asymmetric Online Covering) if it can be defined as follows: The input to an instance of P consists of a
sequence of n requests, σ = hr1 , . . . , rn i, and possibly one final dummy request.
An algorithm for P computes a binary output string, y = y1 . . . yn ∈ {0, 1}n,
where yi = f (r1 , . . . , ri ) for some function f .
5
For minimization (maximization) problems, the score function, s, maps a
pair, (σ, y), of input and output to a cost (profit) in N ∪ {∞} (N ∪ {−∞}).
For an input, σ, and an output, y, y is feasible if s(σ, y) ∈ N. Otherwise, y is
infeasible. There must exist at least one feasible output. Let Smin (σ) (Smax (σ))
be the set of those outputs that minimize (maximize) s for a given input σ.
If P is a minimization problem, then for every input, σ, the following must
hold:
1. For a feasible output, y, s(σ, y) = |y|1 .
2. An output, y, is feasible if there exists a y ′ ∈ Smin (σ) such that y ′ ⊑ y.
If there is no such y ′ , the output may or may not be feasible.
If P is a maximization problem, then for every input, σ, the following must
hold:
1. For a feasible output, y, s(σ, y) = |y|0 .
2. An output, y, is feasible if there exists a y ′ ∈ Smax (σ) such that y ′ ⊑ y.
If there is no such y ′ , the output may or may not be feasible.
Recall that no problem in AOC requires more than B(n, c) + O(log n) bits
of advice (see Eq. (1) for the definition of B(n, c)). This result is based on a
covering design technique, where the advice indicates a superset of the output
bits that are 1 in an optimal solution.
The problems in AOC requiring the most advice are AOC-complete [7]:
Definition 3 (AOC-complete [7]). A problem P ∈ AOC is AOC-complete if
for all c > 1, any c-competitive algorithm for P must read at least B(n, c) −
O(log n) bits of advice.
In [7], an abstract guessing game, minASGk (Minimum Asymmetric String
Guessing with Known History), was introduced and shown to be AOC-complete.
The minASGk-problem itself is very artificial, but it is well-suited as the starting
point of reductions. All minimization problems known to be AOC-complete have
been shown to be so via reductions from minASGk.
The input for minASGk is a secret string x = x1 x2 . . . xn ∈ {0, 1}n given
in n rounds. In round i ∈ [n], the online algorithm must answer yi ∈ {0, 1}.
Immediately after answering, the correct answer xi for round i is revealed to the
algorithm. If the algorithm answers yi = 1, it incurs a cost of 1. If the algorithm
answers yi = 0, then it incurs no cost if xi = 0, but if xi = 1, then the output
of the algorithm is declared to be infeasible (and the algorithm incurs a cost of
∞). The objective is to minimize the total cost incurred. Note that the optimal
solution has cost |x|1 . See the appendix for a formal definition of minASGk and
for definitions of other AOC-complete problems.
The problem minASGk is based on the binary string guessing problem [2,11].
Binary string guessing is similar to asymmetric string guessing, except that any
wrong guess (0 instead of 1 or 1 instead of 0) gives a cost of 1.
In Theorem 1, we show a very strong lower bound for a weighted version of
minASGk. In Theorem 2, via reductions, we show that this lower bound implies
6
similar strong lower bounds for the weighted version of other AOC-complete
minimization problems.
Definition 4 (Weighted AOC). Let P be a problem in AOC. We define the
weighted version of P, denoted Pw , as follows: A Pw -input σ = h{r1 , w1 },
{r2 , w2 }, . . . , {rn , wn }i consists of n P-requests, r1 , ..., rn , each of which has a
weight wi ∈ R+ . The P-request ri and its weight wi are revealed simultaneously.
An output y = y1 . . . yn ∈ {0, 1}n is feasible for the input σ if and only if y is
feasible for the P-input hr1 , . . . , rn i. The cost (profit) of an infeasible solution is
∞ (−∞).
If P is a minimization problem, then the cost of a feasible Pw -output y for
an input σ is
n
X
wi yi
s(σ, y) =
i=1
If P is a maximization problem, then the profit of a feasible Pw -output y for
an input σ is
n
X
wi (1 − yi )
s(σ, y) =
i=1
3
Weighted Versions of AOC-Complete Minimization
Problems
In the weighted version of minASGk, minASGkw, each request is a weight for
the current request and the value 0 or 1 of the previous request. Producing a
feasible solution requires accepting (answering 1 to) all requests with value 1,
and the cost of a feasible solution is the sum of all weights for requests which
are accepted.
We start with a negative result for minASGkw and then use it to obtain
similar results for the weighted online version of Vertex Cover, Set Cover, Dominating Set, and Cycle Finding.
Theorem 1. For minASGkw , no algorithm using less than n bits of advice is
f (n)-competitive, for any function f .
Proof. Let Alg be any algorithm for minASGkw reading at most n − 1 bits of
advice. We show how an adversary can construct input sequences where the cost
of Alg is arbitrarily larger than that of Opt. We only consider sequences with at
least one 1. It is easy to see that for the unweighted version of the binary string
guessing problem, n bits of advice are necessary in order to guess correctly each
time: If there are fewer than n bits, there are only 2n−1 possible advice strings,
so, even if we only consider the 2n − 1 possible inputs with at least one 1, there
are at least two different request strings, x and y, which get the same advice
string. Alg will make an error on one of the strings when guessing the first bit
where x and y differ, since up until that point Alg has the same information
about both strings.
7
We describe a way to assign weights to the requests in minASGkw such that
if Alg makes a single mistake (either guessing 0 when the correct answer is 1 or
vice versa), its performance ratio is unbounded. We use a large number a > 1,
which we allow to depend on n. All weights are from the interval [1, a] (note that
they are not necessarily integers). We let x = x1 , . . . , xn be the input string and
set w1 = a1/2 . For i > 1, wi is given by:
(
−i
wi−1 · a(−2 ) , if xi−1 = 0
wi =
−i
wi−1 · a(2 ) ,
if xi−1 = 1
Since the weights are only a function of previous requests, they do not reveal
any information to Alg about future requests.
Observation 1 For each i, the following hold:
−n
(a) If xi = 0, then wj ≤ wi · a(−2 ) for all j > i.
−n
(b) If xi = 1, then wj ≥ wi · a(2 ) for all j > i.
We argue for each set of inequalities in the observation: P
j
−(i+1)
−(i+1)
)
)
(a): If xi = 0, for each j > i, wj = wi · a(−2
· a k=i+2 (±2
, where
the plus or minus depends on whether xk = 0 or xk = 1. The value wj is largest
−j
−n
if all of the xk values are 1, in which case wj = wi · a(−2 ) ≤ wi · a(−2 ) .
(b): The argument of xi = 1 is similar, changing minus to plus and vice versa.
We claim that if Alg makes a single mistake, its performance ratio is not
bounded by any function of n. Indeed, if Alg guesses 0 for a request, but the
correct answer is 1, the solution is infeasible and Alg gets a cost of ∞.
We now consider the case where Alg guesses 1 for a request j, but the
correct answer is 0. This request gives a contribution of wj = ab , for some
0 < b < 1, to the cost of the solution produced by Alg. Define j ′ such that
wj ′ = max{wi | xi = 1}. Since Opt only answers 1 if xi = 1, this is the largest
contribution to the cost of Opt from a single request.
−n
−n
If j ′ > j, Observation 1(a) gives that wj ′ ≤ wj · a(−2 ) = ab · a(−2 ) =
−n
−n
ab−2 . The cost of Opt is at most n · wj ′ ≤ n · ab−2 . Thus,
−n
Alg(x)
a2
ab
≥
.
−n =
b−2
Opt(x)
n
n·a
Since a can be arbitrarily large (recall that it can be a function of n), we see
that no algorithm can be f (n)-competitive for any specific function f .
−n
If j ′ < j, Observation 1(b) gives us that wj ≥ wj ′ · a(2 ) . Using wj = ab , we
−n
get ab−2 ≥ wj ′ . We can repeat the argument from the case where j ′ > j to
see that no algorithm can be f (n)-competitive for any specific function f .
⊓
⊔
In order to show that similar lower bounds apply to all minimization problems
known to be complete for AOC, we define a simple type of advice preserving
reduction for online problems. These are much less general than those defined
8
by Sprock in his PhD dissertation [18], mainly because we do not allow the
amount of advice needed to change by a multiplicative factor.
Let OptP (σ) denote the value of the optimal solution for request sequence
σ for problem P, and let |σ| denote the number of requests in σ.
Definition 5. Let P1 and P2 be two online minimization problems, and let I1
be the set of request sequences for P1 and I2 be the set of request sequences for
P2 . For a given function g : N → R+ , we say that there is a length preserving
g-reduction from P1 to P2 , if there is a transformation function f : I1 → I2
such that
– for all σ ∈ I1 , |σ| = |f (σ)|, and
– for every algorithm Alg2 for P2 , there is an algorithm Alg1 for P1 such
that for all σ1 ∈ I1 , the following holds:
If Alg2 produces a feasible solution for σ2 = f (σ1 ) with advice φ(σ2 ), then
Alg1 , using at most |φ(σ2 )| + g(|σ2 |) advice bits, produces a feasible solution
for σ1 such that
• Alg1 (σ1 ) ≤ Alg2 (σ2 ) + OptP1 (σ1 ) and OptP1 (σ1 ) ≥ OptP2 (σ2 ), or
• Alg1 (σ1 ) = OptP1 (σ1 )
Note that the transformation function f is length-preserving in that the
lengths of the request sequences for the two problems are identical. This avoids
the potential problem that the advice for the two problems could be functions
of two different sequence lengths. The amount of advice for the problem being
reduced to is allowed to be an additive function, g(n), longer than for the original
problem, because this seems to be necessary for some of the reductions showing
that problems are AOC-complete. Since the reductions are only used here to
show that no algorithm is F (n)-competitive for any function F , the increase in
the performance ratio that occurs with these reductions is insignificant.
The following lemma shows how length-preserving reductions can be used.
Lemma 1. Let P1 and P2 be online minimization problems. Suppose that at
least b1 (n, c) advice bits are required to be (c + 1)-competitive for P1 and suppose
there is a length preserving g(n)-reduction from P1 to P2 . Then, at least b1 (n, c)−
g(n) advice bits are needed for an algorithm for P2 to be c-competitive.
Proof. Let f be the transformation function associated with g. Suppose for the
sake of contradiction that there is a (strictly) c-competitive algorithm Alg2 for
P2 with advice complexity b2 (n, c) < b1 (n, c)−g(n). Then there exists a constant
α such that for any request sequence σ1 ∈ I1 , either Alg1 (σ1 ) = OptP1 (σ1 ) or
Alg1 (σ1 ) ≤ Alg2 (σ2 ) + OptP1 (σ1 )
≤ c · OptP2 (σ2 ) + α + OptP1 (σ1 )
≤ (c + 1) · OptP1 (σ1 ) + α,
where σ2 = f (σ1 ). Thus, Alg1 is (strictly) (c + 1)-competitive, with less than
b1 (n, c) bits of advice, a contradiction.
⊓
⊔
9
All known AOC-complete problems were proven complete using length-preserving
reductions from minASGk, so the following holds for the weighted versions of
all such problems:
Theorem 2. For the weighted online versions of Vertex Cover, Cycle Finding,
Dominating Set, Set Cover, an algorithm reading less than n − O(log n) bits of
advice cannot be f (n)-competitive for any function f .
Proof. The reductions in [7] showing that these problems are AOC-complete are
length preserving O(log n)-reductions from minASGk, and hence, the theorem
follows from Lemma 1. For Vertex Cover, the following O(log n)-reduction can
be used (the other three reductions are given in the Appendix B):
Each input σ = hx1 , x2 , . . . , xn i to the problem minASGk, is transformed to
f (σ) = hv1 , v2 , . . . , vn i, where V = {v1 , v2 , . . . , vn } is the vertex set of a graph
with edge set E = {(vi , vj ) : xi = 1 and i < j}. Let V1 = {vi ∈ V : x1 = 1}. Note
that V1 \ {vn } is a minimum vertex cover of the graph and that no algorithm
can reject more than one vertex from V1 , since V1 induces a clique.
The advice used by the minASGk algorithm Alg1 consists of the advice
used by the Vertex Cover algorithm Alg2 and O(log n) bits that are either all
0 or give (an encoding of) an index to a position i in the input sequence, such
that vi ∈ V1 and Alg2 rejects vi .
Let VAlg2 ⊆ V be the vertex cover constructed by Alg2 and let XAlg1 be the
set of requests on which Alg1 returns a 1. Then either XAlg1 = VAlg1 or XAlg1 =
VAlg2 ∪ {vi }, where {vi } = V1 \ VAlg1 . Thus, Alg1 (σ) ≤ Alg2 (f (σ)) + Opt(σ),
since wi ≤ Opt(σ).
⊓
⊔
4
Exponential Sparsification
Assume that we are faced with an online problem for which we know how to
obtain a reasonable competitive ratio, possibly using advice, in the unweighted
version (or when there are only few possible different weights). We use exponential sparsification, a simple technique which can be of help when designing
algorithms with advice for weighted online problems by reducing the number of
different possible weights the algorithm has to handle. The first step is to partition the set of possible weights into intervals of exponentially increasing length,
i.e., for some small ε, 0 < ε < 1,
R+ =
∞
[
k=−∞
(1 + ε)k , (1 + ε)k+1 .
How to proceed depends on the problem at hand. We now informally explain
the meta-algorithm
that we repeatedly use in this paper. Note that if w1 , w2 ∈
(1 + ε)k , (1 + ε)k+1 and w1 ≤ w2 , then w1 ≤ w2 ≤ (1 + ε)w1 . For many online
problems, this means that an algorithm can treat all requests whose weights
belong to this interval as if they all had weight (1 + ε)k+1 with only a small loss
in competitiveness.
10
Consider now a set of weights and let wmax denote
the largest weight in
kmax
, (1 + ε)kmax +1 .
the set. Let kmax be the integer for which
w
∈
(1
+
ε)
max
We say that a request with weight w ∈ (1 + ε)k , (1 + ε)k+1 is unimportant if
k < kmax − ⌈log1+ε (n2 )⌉. Furthermore, we will often categorize the request as
important if kmax −⌈log1+ε (n2 )⌉ ≤ k < kmax +1 and as huge if k ≥ kmax +1. Each
2
unimportant request has weight w ≤ (1 + ε)k+1 ≤ (1 + ε)kmax −⌈log1+ε (n )⌉−1+1 ≤
wmax /n2 , so the total sum of the unimportant weights is O(wmax /n). For many
weighted online problems, this means that an algorithm can easily serve the
requests with unimportant weights, as follows. In maximization problems, this
is done by rejecting them. In minimization problems, it is done by accepting
them. Thus, exponential sparsification (when applicable) essentially reduces the
problem of computing a good approximate solution for a problem with n distinct
weights to that of computing a good approximate solution with only O(log1+ε n)
distinct weights.
For a concrete problem, several modifications of this meta-algorithm might
be necessary. Often, the most tricky part is how the algorithm can learn kmax
without using too much advice. One approach that we often use is the following:
The oracle encodes the index i of the first request whose weight is close enough
to (1 + ε)kmax that the algorithm only needs a little bit of advice to deduce kmax
from the weight of this request. If it is somehow possible for the algorithm to
serve all requests prior to i reasonably well, then this approach works well.
Our main application of exponential sparsification is to weighted AOC problems. We begin by considering maximization problems. Note that no assumptions
are made about the weights of Pw in Theorem 3.
Theorem 3. If P ∈ AOC is a maximization problem, then for any c > 1 and
0 < ε ≤ 1, Pw has a strictly (1 + ε)c-competitive algorithm using B(n, c) +
O(ε−1 log2 n) advice bits.
Proof. Fix ε > 0. Let σ = h{r1 , w1 }, . . . , {rn , wn }i be the input and let x =
x1 . . . xn ∈ {0, 1}n specify an optimal solution for σ, with zeros indicating membership in the optimal solution. Throughout most of this proof, we assume that
n is sufficiently large. The necessary conditions are discussed at the end of the
proof, along with how to handle small n.
Define s = 1 + ε/2. Let VOpt = {i : xi = 0}. Note that VOpt contains exactly
those rounds in which Opt answers 0 and thus accepts. Furthermore, for k ∈ Z,
k
let V k = {i : sk ≤ w(i) < sk+1 } and let VOpt
= VOpt ∩V k . Finally, let imax ∈ VOpt
be such that w(imax ) ≥ w(i) for every i ∈ VOpt .
m
The oracle computes the unique m ∈ Z such that imax ∈ VOpt
. We say that
2
m−⌈logs (n2 )⌉
a request ri is unimportant if w(i) < s
, important if sm−⌈logs (n )⌉ ≤
w(i) < sm+1 , and huge if w(i) ≥ sm+1 . The oracle computes the index i′ of the
′
first important request in the input sequence. Assume that i′ ∈ V m . The oracle
writes the length n of the input onto the advice tape using a self-delimiting
encoding1 , and then writes the index i′ and the integer m − m′ (which is at most
1
For example, ⌈log n⌉ could be written in unary (⌈log n⌉ ones, followed by a zero)
before writing n itself in binary.
11
⌈logs (n2 )⌉) onto the tape, using a total of O(log n) bits. This advice allows the
algorithm to learn m as soon as the first important request arrives. From there
on, the algorithm will know if a request is important, unimportant, or huge.
Whenever an unimportant or a huge request arrives, the algorithm answers 1
(rejects the request). We now describe how the algorithm and oracle work for
the important requests.
For each 0 ≤ j ≤ ⌈logs (n2 )⌉, let nm−j = V m−j . For the requests (whose
indices are) in V m−j , we use the covering design based c-competitive algorithm
for unweighted AOC-problems. This requires B(nm−j , c) + O(log nm−j ) bits of
advice. Since B(n, c) is linear in n, this means that we use a total of
⌈logs (n2 )⌉
X
b=
j=0
B(nm−j , c) + O(log nm−j ) ≤ B(n, c) + O(logs n · log n)
bits of advice. Note that logs (n) ≤ 2ε−1 log n for ε/2 ≤ 1, giving the bound on
the advice in the statement of the theorem.
We now prove that the algorithm achieves the desired competitiveness. We
can ignore the huge requests, since neither Alg nor Opt accepts any of them.
k
Let VAlg be those rounds in which Alg answers 0 and let VAlg
= VAlg ∩ V k .
2
We consider the important requests first. Fix 0 ≤ j ≤ ⌈logs (n )⌉. Let nOpt
m−j =
m−j
m−j
VOpt
, i.e., nOpt
which are also in the
m−j is the number of requests in V
m−j
optimal solution VOpt . By construction, we have nOpt
m−j ≤ c VAlg . Since the
largest possible weight of a request in V m−j is at most s times larger than the
m−j
) ≤
smallest possible weight of a request in V m−j , this implies that w(VOpt
m−j
s · c · w(VAlg ). Thus, we get that
⌈logs (n2 )⌉
⌈logs (n2 )⌉
X
m−j
w(VOpt
)
≤
m−j
s · c · w(VAlg
) = s · c · Alg(σ).
(2)
j=0
j=0
s
X
We now consider the unimportant requests. If ri is unimportant, then w(i) ≤
′
≤ sm /n2 ≤ w(ximax )/n2 ≤ Opt(σ)/n2 . This implies that
m′ −⌈logs (n2 )⌉
∞
X
m−j
w(VOpt
)≤n
j=⌈logs (n2 )⌉+1
Opt(σ)
Opt(σ)
=
.
2
n
n
We conclude that
⌈logs (n2 )⌉
Opt(σ) = w(VOpt ) =
X
m−j
w(VOpt
)+
j=0
∞
X
j=⌈logs (n2 )⌉+1
By Eq. (3),
⌈logs (n2 )⌉
X
1
m−j
1−
w(VOpt
),
Opt(σ) ≤
n
j=0
12
m−j
w(VOpt
).
(3)
n
so by Eq. (2), Opt(σ) ≤ n−1
· s · c · Alg(σ).
n
Note that for n ≥ n0 = 2+2ε
ε , ( n−1 )(1 + ε/2) ≤ (1 + ε). For inputs of length
less than n0 , the oracle writes an optimal solution onto the advice tape, using at
most n0 bits. Since n0 ≤ 4ε , b ∈ O(ε−1 log2 n) as required. For inputs of length
at least n0 , we use the algorithm described above. Thus, for every input σ, it
holds that Opt(σ) ≤ (1 + ε)cAlg(σ). Since ε was arbitrary, this proves the
theorem.
⊓
⊔
It may be surprising that adding weights to AOC-complete maximization problems has almost no effect, while adding weights to AOC-complete minimization
problems drastically changes the advice complexity. In particular, one might
wonder why the technique used in Theorem 3 does not work for minimization
problems. The key difference lies in the beginning of the sequence. Let wmax be
the largest weight of a request accepted by Opt.
For maximization problems, the algorithm can safely reject all requests before
the first important one. For minimization problems, this approach does not work,
since the algorithm must accept a superset of what Opt accepts in order to
ensure that its output is feasible. Thus, rejecting an unimportant request that
Opt accepts may result in an infeasible solution. This essentially means that the
algorithm is forced into accepting all requests before the first important request
arrives. Accepting all unimportant requests is no problem, since they will not
contribute significantly to the total cost. However, accepting even a single huge
request can give an unbounded contribution to the algorithm’s cost. As shown
in Theorem 1, it is not possible in general for the algorithm to tell if a request
in the beginning of the sequence is unimportant or huge without using a lot of
advice.
However, if the ratio of the largest to the smallest weight is not too large,
exponential sparsification is also useful for minimization problems in AOC. Essentially, when this ratio is bounded, it is possible for the algorithm to learn a
good approximation of wmax when the first request arrives. This is formalized in
Theorem 4, the proof of which is very similar to the proof of Theorem 3.
Theorem 4. If P ∈ AOC is a minimization problem and 0 < ε ≤ 1, then Pw
with all weights in [wmin , wmax ] has a (1 + ε)c-competitive algorithm with advice
complexity at most
wmax
2
−1
−1
B(n, c) + O ε log n + log ε log
.
wmin
Proof. Fix ε > 0. Let σ = h{r1 , w1 }, . . . , {rn , wn }i be the input and let x =
x1 . . . xn ∈ {0, 1}n specify an optimal solution for σ, with ones indicating membership in the optimal solution. Define s = 1 + ε/2. Let VOpt = {i : xi = 1}.
Note that VOpt contains exactly those rounds in which Opt answers 1 and thus
accepts. Furthermore, for k ∈ Z, let V k = {i : sk ≤ w(i) < sk+1 } and let
k
VOpt
= VOpt ∩ V k . Finally, let imax ∈ VOpt be such that w(imax ) ≥ w(i) for
every i ∈ VOpt .
13
m
The oracle computes the unique m ∈ Z such that imax ∈ VOpt
. We say that
2
m−⌈logs (n2 )⌉
a request ri is unimportant if w(i) < s
, important if sm−⌈logs (n )⌉ ≤
m+1
m+1
w(i) < s
, and huge if w(i) ≥ s
. The oracle also computes the unique
′
′
m ∈ Z such that sm ≤ w1 < sm +1 and writes the values n and m − m′ on the
tape in a self-delimiting encoding.
The number of advice bits needed to write m − m′ is O(log(m − m′ )).
log (m − m′ ) ≤ log (logs w(imax ) − logs w1 ) + 1
wmax
≤ log logs
+1
wmin
wmax
≤ log 2ε−1 log
+ 1, since logs n ≤ 2ε−1 log n, for ε/2 ≤ 1
wmin
Note that since the length of m−m′ is not known, we need to use a self-delimiting
max
encoding, which means that we use O(log n + log(ε−1 log w
wmin )) advice bits at
the beginning.
This advice allows the algorithm to learn m as soon as the first request
arrives. From there on, the algorithm will know if a request is important, unimportant, or huge. Whenever a huge request arrives, the algorithm answers 0
(rejects the request). When an unimportant request arrives, the algorithm answers 1 (accepts the request). We now describe how the algorithm and oracle
work for the important requests.
For the important requests (whose indices are) in V m−j , we use the covering design based c-competitive algorithm for unweighted AOC-problems. This
is similar to what we do in the proof of Theorem 3. The same calculations
yield an upper bound on this advice of B(n, c) + O(logs n · log n). Note that
logs (n) ≤ 2ε−1 log n for ε/2 ≤ 1, giving the bound on the advice in the statement of the theorem.
First, we note that the solution produced is valid, since it is a superset of the
solution of Opt.
We now argue that the cost of the solution is at most (1 + ε)c times the cost
of Opt. Following the proof of Theorem 3 and switching the roles of Opt and
Alg, we have by construction that the cost of the important requests for the
algorithm is at most sc times larger than the cost for Opt on the important
requests. For the huge requests, both this algorithm and Opt incur a cost of
zero.
We now consider the unimportant requests. If ri is unimportant, then
2
w(i) < sm−⌈logs (n
)⌉
≤ sm /n2 ≤ w(ximax )/n2 ≤ Opt(σ)/n2 .
This implies that
∞
X
w(V m−j ) ≤ n
j=⌈logs (n2 )⌉+1
Opt(σ)
Opt(σ)
=
.
n2
n
(4)
Thus, even if the algorithm accepts all unimportant requests and Opt accepts
. In total, the algorithm gets
none of them, it only accepts an additional Opt(σ)
n
14
a cost of at most (1 + n1 )(1 + ε/2)cOpt(σ). For n ≥ n0 = 2+ε
ε , this is at
most (1 + ε)cOpt(σ). For inputs of length less than n0 , the oracle will write an
optimal solution onto the advice tape, using at most n0 bits. Since n0 ≤ 3ε , b ∈
O(ε−1 log2 n) as required. For inputs of length at least n0 , we use the algorithm
described above. Thus, for every input σ, it holds that Opt(σ) ≤ (1+ε)cAlg(σ).
⊓
⊔
5
Matching and Other Non-Complete AOC Problems
We first provide a general theorem that works for all maximization problems in
AOC, giving better results in some cases than that in Theorem 3.
Theorem 5. Let P ∈ AOC be a maximization problem. If there exists a ccompetitive P-algorithm reading b bits of advice, then there exists a O(c · log n)competitive Pw -algorithm reading O(b + log n) bits of advice.
Proof. Use exponential sparsification on the weights with an arbitrary ε, say
ε = 1/2, and let s = 1 + ε. For a given request sequence, σ, let wmax be the
maximum weight that OptPw accepts. The oracle computes the unique m ∈ Z
such that wmax ∈ [sm , sm+1 ). The important requests are those with weight w,
2
where sm−⌈logs (n )⌉ ≤ w < sm+1 .
We consider only the ⌈logs (n2 )⌉ + 1 important intervals, i.e., the intervals
i i+1
[s , s ), m − ⌈logs (n2 )⌉ ≤ i ≤ m, and index them by i. Let k be the index of
the interval of weights contributing the most weight to OptPw (σ). The advice
is a self-delimiting encoding of the index, j, of the first request with weight
w ∈ [sk , sk+1 ), plus the advice used by the given c-competitive P-algorithm.
This requires at most b + O(log(n)) bits of advice.
The algorithm rejects all requests before the jth. From the jth request, the
algorithm calculates the index k. The algorithm accepts those requests which
would be accepted by the P-algorithm when presented with the subsequence
of σ consisting of the requests with weights in [sk , sk+1 ). Since, by exponential
sparsification, OptPw accepts total weight at most n1 OptPw (σ) from requests
with unimportant weights, and it accepts at least as much from interval k as
from any of the other ⌈logs (n2 )⌉ + 1 intervals considered, OptPw accepts weight
OptPw (σ)
at least (1 − n1 ) ⌈log (n
2 )⌉+1 from interval k. The algorithm, Alg, described here
s
accepts at least 1c as many requests as OptPw does in this interval, and each
1
of the requests it accepts is at least a fraction
s as large
as the largest weight
in this interval. Thus, c(1 + ε)Alg(σ) ≥
O(c log n)-competitive.
1
1− n
⌈logs (n2 )⌉+1
OptPw (σ), so Alg is
⊓
⊔
In the online matching problem, edges arrive one by one. Each request contains the names of the edge’s two endpoints (the set of endpoints is not known
from the beginning, but revealed gradually as the edges arrive). The algorithm
must irrevocably accept or reject them as they arrive, and the goal is to maximize
the number of edges accepted. The natural greedy algorithm for this problem is
15
well known to be 2-competitive. In terms of advice, the problem is known to be
in AOC, but is not AOC-complete [7]. We remark that a version of unweighted
online matching with vertex arrivals (incomparable to our weighted matching
with edge arrivals) has been studied with advice in [10].
Corollary 1. There exists a O(log n)-competitive algorithm for Weighted Matching reading O(log n) bits of advice.
Proof. The result follows from Theorem 5 since there exists a 2-competitive
algorithm without advice for (unweighted) Matching.
⊓
⊔
5.1
Lower bounds
First, we present a result which holds for the weighted versions of many maximization problems in AOC. It also holds for the weighted versions of AOC-complete
minimization problems, but Theorem 2 gives a much stronger result.
Theorem 6. For the weighted online versions of Independent Set, Clique, Disjoint Path Allocation, and Matching, an algorithm reading o(log n) bits of advice
cannot be f (n)-competitive for any function f .
To prove Theorem 6, we start by proving the following lemma from which
the theorem easily follows.
Lemma 2. Let P ∈ AOC and suppose there exists a family (σn )n∈N of P-inputs
with the following properties:
1. σn = hr1 , r2 , . . . , rn i consists of n requests.
2. σn+1 is obtained by adding a single request to the end of σn .
3. If P is a maximization problem, the feasible solutions are those in which at
most one request is accepted.
If P is a minimization problem, the feasible solutions are those in which at
least one request is accepted.
Then, no algorithm for the weighted problem Pw reading o(log n) bits of advice
can be f (n)-competitive for any function f .
Proof. Let Alg be a Pw -algorithm reading at most b = o(log n) bits of advice.
Let f (n) > 0 be an arbitrary non-decreasing function of n. We will show that
for all sufficiently large n, there exists an input of length n such that the profit
obtained by Opt is at least f (n) times as large as the profit obtained by Alg.
Since f (n) was arbitrary, it follows that Alg is not f (n)-competitive for any
function f .
Since b = o(log n), there exists an N ∈ Z such that for any n ≥ N , Alg
reads less than log(n) − 1 bits of advice on inputs of length at most n. Fix an
n ≥ N . For 1 ≤ i ≤ n, define the Pw -input σ
bi = h{r1 , f (n)}, {r2 , f (n)2 }, . . . ,
i
{ri , f (n) }i. Consider the set of inputs {b
σ1 , . . . , σ
bn }. For every 1 ≤ i ≤ n, the
number of advice bits read by Alg on the input σ
bi is at most log(n) − 1 (since
16
the length of the input σ
bi is i ≤ n). Thus, by the pigeonhole principle, there must
exist two integers n1 , n2 with n1 < n2 such that Alg reads the same advice on
bn1 , then it achieves a profit of 0
bn2 . If Alg rejects all requests in σ
σ
bn1 and σ
bn1 , then it
while Opt obtains a profit of f (n)n1 . If Alg accepts a request in σ
bn1 and
obtains a profit of at most f (n)n1 . Since Alg reads the same advice on σ
σ
bn2 and since the two inputs are indistinguishable for the first n1 requests, this
bn2 . But
means that Alg also obtains a profit of at most f (n)n1 on the input σ
σn2 ) ≥ f (n)n2 −n1 ≥ f (n).
σn2 )/Alg(b
Opt(b
σn2 ) = f (n)n2 , and hence Opt(b
For minimization problems, we can use the same arguments and the input
sequence σ
bi = h{r1 , f (n)−1 }, {r2 , f (n)−2 }, . . . , {ri , f (n)−n }i.
⊓
⊔
Proof (Proof of Theorem 6). For Independent Set, we can use the above lemma
with a family of cliques (Kn )n∈N , and for Clique, we can use a family of independent sets. For Matching, we can use a family of stars (K1,n )n∈N . For Disjoint Path
Allocation, we use a path P2n = hv1 , v2 , . . . , v2n i and ri = hvi , vi+1 , . . . vi+n i. ⊓
⊔
Returning to the example of Weighted Matching, we now know that O(log n)
bits suffice to be O(log n)-competitive, and that no algorithm can be f (n)competitive for any function f with o(log n) bits of advice. In order to prove
that a linear number of advice bits is necessary to achieve constant competitiveness for Weighted Matching, we use a direct product theorem from [14]. This
uses the concept defined in [14] of a problem being Σ-repeatable. Informally,
this means that it is always possible to combine r (sufficiently profitable) input
sequences I1 , I2 , . . . , Ir into a single input g(I1 , I2 , . . . , Ir ) such that serving this
single input gives profit close to that of serving each of the Ii independently and
adding the profits.
Definition 6. Let P be an online maximization problem and I be the set of
possible input sequences. Assume that for every input in I, there are only a
finite number of valid outputs. Let I ∗ be the set of concatenations of sequences
(rounds) from I. P is Σ-repeatable with parameters (k1 , k2 , k3 ) if there exists a
function g : I ∗ → I satisfying the following:
– For every σ ∗ ∈ I ∗ with r rounds, |g(σ ∗ )| ≤ |σ ∗ | + k1 r
– For every deterministic algorithm Alg for P , there is a deterministic algorithm Alg∗ for sequences from I ∗ , such that for every σ ∗ ∈ I ∗ with r
rounds, Alg∗ (σ ∗ ) ≥ Alg(g(σ ∗ ) − k2 r,
– Let Opt∗ denote an optimal algorithm for sequences from I ∗ . For every
σ ∗ ∈ I ∗ with r rounds, Opt∗ (σ ∗ ) ≤ Opt(g(σ ∗ )) + k3 r.
Theorem 7. An O(1)-competitive algorithm for Weighted Matching must read
Ω(n) bits of advice.
Proof. We prove the lower bound using a direct product theorem [14]. According
to [14], it suffices to show that: (i) Weighted Matching is Σ-repeatable, and
(ii) for every c, there exists a probability distribution pc with finite support
such that for every deterministic algorithm Det without advice, it holds that
17
Epc [Opt(σ)] ≥ c · Epc [Det(σ)]. Also, there must be a finite upper bound on the
profit an algorithm can obtain on an input in the support of pc .
It is trivial to see that Weighted Matching is Σ-repeatable. Fix c ≥ 1 and let
k = 2c− 1. We define the probability distribution pc by specifying a probabilistic
adversary: The input graph will be a star K1,m consisting of m edges for some
1 ≤ m ≤ k. In round i, the adversary reveals the edge ei = (v, vi ) where vi is
a new vertex and v is the center vertex of the star. The edge ei has weight 2i .
If i < k, then with probability 1/2 the adversary will proceed to round i + 1,
and with probability 1/2 the input sequence will end. If the adversary reaches
round k, it will always stop after revealing the edge ek of round k. Note that the
support of pc and the largest profit an algorithm can obtain on any input in the
support of pc are both finite.
Let X be the random variable which denotes the number of edges revealed
by the adversary. Note that Pr(X = j) = 2−j if 1 ≤ j < k. Consequently,
Pr(X = k) = 1 − Pr(X < k) = 1 −
k−1
X
2−i = 2−(k−1) .
(5)
i=1
Let Det be a deterministic algorithm without advice. We may assume that Det
decides in advance on some 1 ≤ j ≤ k and accepts the edge ej (the only other
possible deterministic strategy it to never accept an edge, but this is always
strictly worse than following any of the k strategies that accepts an edge). If
X < j, then the profit obtained by Det is zero. If X ≥ j, then Det obtains a
profit of 2j . It follows that
E[Det(σ)] = Pr(X ≥ j)2j = (1 − Pr(X < j))2j = 2−(j−1) 2j = 2.
The optimal algorithm Opt always accepts the last edge of the input. Thus, if
X = j, then the profit of Opt is 2j . It follows that
E[Opt(σ)] =
k
X
Pr(X = j)2j =
j=1
Thus, we conclude that E[Opt(σ)] ≥
k−1
X
j=1
k+1
2
2−j 2j + 2−(k−1) 2k = k + 1.
E[Det(σ)] = c E[Det(σ)].
⊓
⊔
In particular, we cannot achieve constant competitiveness using O(log n) bits
of advice for Weighted Matching. We leave it as an open problem to close the
gap between ω(1) and O(log n) on the competitiveness of Weighted Matching
algorithms with advice complexity O(log n).
6
Scheduling with Sublinear Advice
For the scheduling problems studied, the requests are jobs, each characterized
by its size. Each job must be assigned to one of m available machines. If the
machines are identical, the load of a job on any machine is simply its size. If
18
the machines are related, each machine has a speed, and the load of a job, J,
assigned to a machine with speed s is the size of J divided by s. If the machines
are unrelated, each job arrives with a vector specifying its load on each machine.
Consider a sequence σ = hr1 , . . . , rn i of n jobs that arrive online. Each job
ri ∈ σ has an associated weight-function wi : [m] → R+ . Upon arrival, a job
must irrevocably be assigned to
Pone of the m machines. The load Lj of a machine
j ∈ [m] is defined as Lj =
i∈Mj wi (j) where Mj is the set of (indices of)
jobs scheduled on machine j. The total load of a schedule for σ is the vector
L = (L1 , . . . , Lm ). We say that (L1 , . . . , Lm ) ≤ (L′1 , . . . , L′m ) if and only if
Li ≤ L′i for 1 ≤ i ≤ m. A scheduling problem of the above type is specified by an
objective function f : Rm
+ → R+ and by specifying if the goal is to minimize or
maximize f (L) = f (L1 , . . . , Lm ) ∈ R+ . We assume that f is non-decreasing, i.e.,
f (L) ≤ f (L′ ) for all L ≤ L′ . Some of the classical choices of objective function
include:
– Minimizing the ℓp -norm fp (L) = fp (L1 , . . . , Lm ) = k(L1 , . . . , Lm )kp for some
1/p
P
p
L
1 ≤ p ≤ ∞. That is, for 1 ≤ p < ∞, the goal is to minimize
j∈[m] j
and for p = ∞, the goal is to minimize the makespan maxj∈[m] Lj .
– Maximizing the minimum load f (L) = minj∈[m] Lj . This is also known as
machine covering. Note that this objective function is not a norm2 , but it
does satisfy that f (αL) = αf (L) for every α ≥ 0 and L ∈ Rm
+.
We begin with a result for unrelated machines.
Theorem 8. Let P be a scheduling problem on m unrelated machines where the
goal is to minimize an objective function f . Assume that f is a non-decreasing
norm. Then, for 0 < ε ≤ 1, there exists a (1+ε)-competitive P-algorithm reading
O ( 4ε log(n)+2)m log(n) bits of advice. In particular, if m = O(1) and ε = Ω(1),
then there exists a (1 + ε)-competitive algorithm reading O(polylog(n)) bits of
advice.
Proof. Since the objective function f is a norm on Rm , we will denote it by k · k.
Let 1j be the jth unit vector (the vector with a 1 in the jth coordinate and 0
elsewhere).
Fix an input sequence σ. The oracle starts by computing an arbitrary optimal
schedule for σ. Throughout most of this proof, we assume that n is sufficiently
large. The necessary conditions are discussed at the end of the proof, along with
how to handle small n.
Let LOpt be the load-vector of this schedule. Thus, Opt(σ) = kLOpt k. Let
s = 1 + ε/2 and let k be the unique integer such that sk ≤ kLOpt k < sk+1 . A
job ri ∈ σ is said to be unimportant if there exists a machine j ∈ [m] such that
2
kwi (j)1j k < sk−⌈logs (n )⌉ . A job which is not unimportant is important.
The oracle uses O(log n) bits to encode n using a self-delimiting encoding.
It then writes the index i′ of the first important job ri′ onto the advice tape
(or indicates that σ contains no important jobs) using ⌈log(n + 1)⌉ bits. Let
2
f is a norm if f (αv) = |α|f (v), f (u + v) ≤ f (u) + f (v), and f (v) = 0 ⇒ v = 0.
19
j ′ be the machine minimizing kwi′ (j ′ )1j ′ k, where ties are broken arbitrarily.
The oracle also writes ∆i′ = k − ki′ , where ki′ is the unique integer such that
ski′ ≤ kwi′ (j ′ )1j ′ k < ski′ +1 onto the advice tape using O(log logs n) bits.
Scheduling unimportant jobs. If a job ri ∈ σ is unimportant, then the algorithm schedules the job on the machine j minimizing kwi (j)1j k where ties are
broken arbitrarily. We now explain how the algorithm knows if a job is unimportant or not. If ri ∈ σ is a job that arrives before the first important job,
i.e., if i < i′ , then ri is unimportant by definition. When job ri′ arrives, the
algorithm can deduce k since it knows ∆i′ from the advice and since it can compute minj kwi′ (j)1j k without help. Knowing k (and the number of jobs n), the
algorithm is able to tell if a job is unimportant or not.
Scheduling important jobs. We now describe how the algorithm schedules the
important jobs. To this end, we define the type of an important job. For an
important job ri , let ∆i (1), . . . , ∆i (m) be defined as follows: For 1 ≤ j ≤ m,
if there exists an integer ki (j) ≤ k such that ski (j) ≤ kwi (j)1j k < ski (j)+1 ,
then ∆i (j) = k − ki (j) (since ri is important, ∆i (j) ≤ ⌈logs (n2 )⌉). If no such
integer exists, then it must be the case that kwi (j)1j k ≥ sk+1 > kLOpt k. In this
case, we let ∆i (j) = ⊥ be a dummy symbol. The type of ri is the vector ∆i =
(∆i (1), . . . , ∆i (m)). Note that there are only (⌈logs (n2 )⌉ + 2)m different types.
For each possible type ∆ = (∆(1), . . . , ∆(m)), the oracle writes the number,
a∆ , of jobs of type ∆ onto the advice tape. This requires at most (⌈logs (n2 )⌉ +
2)m ⌈log(n + 1)⌉ bits of advice.
Note that since k · k is a norm, if ri ∈ σ is of type ∆i = (∆i (1), . . . , ∆i (m)),
then sk−∆i (j) k1j k−1 ≤ wi (j) ≤ sk−∆i (j)+1 k1j k−1 if ∆i (j) 6= ⊥ and wi (j) >
kLOpt kk1j k−1 if ∆i (j) = ⊥. The algorithm computes an optimal schedule Sbimp
for the input σ
b which for each possible type ∆ contains a∆ jobs with weightfunction w
b∆ where w
b∆ (j) = sk−∆(j)+1 k1j k−1 if k(j) 6= ⊥ and w
b∆ (j) = ∞
otherwise. This choice of weight-function ensures that if ri ∈ σ is a job of type
∆i , then for each j with ∆i (j) 6= ⊥,
wi (j) < w
b∆i (j) ≤ s · wi (j).
(6)
When an important job of σ arrives, the algorithm computes the type of the job.
Based solely on this type, the algorithm schedules the important jobs in σ by
following the schedule Sbimp for σ
b . Let Limp be the load-vector of the important
jobs of σ scheduled by Alg. Note that by Eq. (6), the weight-function of an
important job of σ is strictly smaller (for all machines) than the weight-function
of the corresponding job of σ
b. Thus, since f is non-decreasing kLimp k is bounded
from above by the cost of the schedule Sbimp for σ
b.
Putting it all together. The optimal schedule for σ computed by the oracle
b be the load-vector of this schedule. By Eq. (6),
induces a schedule of σ
b. Let L
b
we get that kLk ≤ skLOpt k. Thus, the cost of Sbimp (which was an optimal
b ≤ skLOpt k.
scheduling of σ
b) is at most kLk
Let Lunimp be the load-vector of the unimportant jobs scheduled by Alg.
Furthermore, let Mj be the set of indices of the unimportant jobs scheduled by
20
Alg on machine j. By subadditivity,
kLunimpk =
m X
X
wi (j)1j ≤
j=1 i∈Mj
≤n
kLOpt k
kLOpt k
≤
.
n2
n
m X
X
j=1 i∈Mj
kwi (j)1j k <
m X
X
2
sk−⌈logs (n
)⌉
j=1 i∈Mj
We are finally able to bound the cost of the entire schedule created by Alg:
Alg(σ) = kLimp + Lunimp k ≤ kLimp k + kLunimpk ≤ (s + 1/n)kLOpt k
Recall that s = 1 + ε/2. Thus, if n ≥ 2/ε, then Alg(σ) ≤ (1 + ε)Opt(σ). For
inputs of length less than 2/ε, the oracle can simply encode the optimal solution
using at most 2ε ⌈log m⌉ bits of advice. The total amount of advice used by our
algorithm is at most
(⌈logs (n2 )⌉ + 2)m ⌈log(n + 1)⌉ + O(log n + log logs n) = O (4ε−1 log(n) + 2)m log(n) .
⊓
⊔
For the following discussion, assume that ε = Θ(1). We remark that the
(1 + ε)-competitive algorithm in Theorem 8 is only of interest if the number of
machines m is small compared to the number of jobs n. As already noted, the
most interesting aspect of Theorem 8 is that our algorithm uses only polylog(n)
bits of advice if m is a constant. More generally, if m = o(log n/ log log n),
then our algorithm will use o(n) bits of advice. On the other hand, if m =
Θ(log n), then our algorithm uses Ω(log(n)log(n) ) bits of advice, which is worse
than the trivial 1-competitive algorithm which uses n⌈log m⌉ = O(n log log n)
bits of advice when m = Θ(log n).
The advice complexity of the algorithm in Theorem 8 depends on the number
of machines m because we want the result to hold even when the machines are
unrelated. We now show that when restricting to related machines, we can obtain
a (1 + ε)-competitive algorithm using O(ε−1 log2 n) bits of advice, independent
of the number of machines. The proof resembles that of Theorem 8. The main
difference is that we are able to reduce the number of types to O(log2 n).
Theorem 9. Let P be a scheduling problem on m related machines where the
goal is to minimize an objective function f . Assume that f is a non-decreasing
norm. Then, for 0 < ε ≤ 1, there
exists a (1 + ε)-competitive P-algorithm with
advice complexity O ε−1 log2 n .
Proof. Since the objective function f is a norm on Rm , we will denote it by k · k.
Fix an input sequence σ. The oracle starts by computing an arbitrary optimal
schedule for σ. Let s = 1 + ε/2. The oracle uses O(log n) bits to encode n using
a self-delimiting encoding.
Let C1 , . . . , Cm be the speeds of the m machines. Assume without loss of generality that k1j k/Cj attains its minimum value when j = 1. Define B = k11 k/C1 .
21
Let LOpt be the load-vector of the fixed optimal schedule. Thus, Opt(σ) =
kLOpt k. Let k be the unique integer such that sk ≤ kLOpt k < sk+1 . A job ri ∈ σ
2
is said to be unimportant if its weight, wi , satisfies wi B < sk−⌈logs (n )⌉ . A job
which is not unimportant is important. Note that wi B is always bounded from
above by kLOpt k since ri must be placed on some machine. The oracle writes
the index i′ of the first important job ri′ onto the advice tape (or indicates that
σ contains no important jobs) using ⌈log n⌉ + 1 bits. The oracle also writes the
′
′
unique integer k ′ such that sk−k ≤ wi′ B < sk−k +1 onto the advice tape, using
O(log logs (n)) bits.
We now explain how the algorithm knows if a job is unimportant or not. If
ri ∈ σ is a job that arrives before the first important job, i.e., if i < i′ , then ri
is unimportant by definition. When job ri′ arrives, the algorithm can deduce k
since it knows k ′ from the advice and since it can compute wi′ B without help.
Knowing k (and the number of jobs n), the algorithm is able to tell if a job is
unimportant or not.
If a job ri ∈ σ is unimportant, then the algorithm schedules the job on
machine 1.
Scheduling important jobs. We now describe how the algorithm schedules the
important jobs. To this end, we define the type of an important job. The type
of an important job ri is the non-negative integer ti such that sk−ti ≤ wi B <
sk−ti +1 . Note that there are only ⌈logs (n2 )⌉+ 1 different types. For each possible
type 0 ≤ t ≤ ⌈logs (n2 )⌉, the oracle writes the number of jobs at of that type
onto the advice tape. This requires at most O(log s (n2 ) log(n)) bits of advice.
Note that since k · k is a norm, if ri ∈ σ is of type ti , then sk−ti B −1 ≤ wi ≤
sk−ti +1 B −1 . The algorithm computes an optimal schedule Sbimp for the input σ
b
which for each possible type 0 ≤ t ≤ ⌈logs (n2 )⌉ contains at jobs with weight
w
bt = sk−ki +1 B −1 . This choice of weight ensures that if ri ∈ σ is a job of type
ti , then,
wi < w
bti ≤ s · wi .
(7)
When an important job of σ arrives, the algorithm computes the type of the job.
Based solely on this type, the algorithm schedules the important jobs in σ by
following the schedule Sbimp for σ
b . Let Limp be the load-vector of the important
jobs of σ scheduled by Alg. Note that by Eq. (7), the weight of an important
job of σ is strictly smaller than the weight of the corresponding job of σ
b. Thus,
kLimp k is bounded from above by the cost of the schedule Sbimp for σ
b.
Putting it all together. The fixed optimal schedule for σ induces a scheduling
b be the load-vector of this schedule. By Eq. (7), we get that L
b ≤
of σ
b . Let L
b
sLOpt . Thus, the cost of Simp (which was an optimal scheduling of σ
b) is at most
b ≤ skLOpt k.
kLk
Let Wu be the total weight of unimportant jobs scheduled on machine 1 by
Alg. We have that
2
k(Wu /C1 )11 k = Wu B ≤ nsk−⌈logs (n
22
)⌉
≤n
kLOpt k
kLOpt k
≤
.
n2
n
We are finally able to bound the cost of the entire schedule Simp ∪Sunimp created
by Alg:
Alg(σ) = kLimp + (Wu /C1 )11 k ≤ kLimp k + k(Wu /C1 )11 k ≤ (s + 1/n)kLOpt k
Recall that s = 1 + ε/2. Thus, if n > 2/ε, then Alg(σ) ≤ (1 + ε)Opt(σ). For
inputs of length less than 2/ε, the oracle can simply encode the optimal solution
using at most 2ε ⌈log m⌉ bits of advice. The total amount of advice used by our
algorithm is O(ε−1 log2 n).
⊓
⊔
We now consider scheduling problems where the goal is to maximize an objective function f . Recall that we assume that the objective function is nondecreasing. The most notable example is when f is the minimum load. In the
following theorem, we show how to schedule almost optimally on unrelated machines with only a rather weak constraint on f (weaker than f being a norm).
Theorem 10. Let P be a scheduling problem on m unrelated machines where the
goal is to maximize an objective function f . Assume that f is non-decreasing,
that f (αL) ≤ αf (L) for every α ≥ 0, and L ∈ Rm
+ . Then, for every 0 <
ε ≤ 1, there exists a (1 + ε)-competitive P-algorithm with advice complexity
O(( 4ε log(n) + 2)m m2 log n). In particular, if m = O(1) and ε = Ω(1), the advice
complexity is O(polylog(n)).
Proof. Fix an input sequence σ and an arbitrary optimal schedule. Let s =
1+ε/2. The oracle uses O(log n) bits to encode n using a self-delimiting encoding.
For 1 ≤ j ≤ m, let Lj be the load on machine j in the optimal schedule.
Furthermore, let kj be the unique integer such that skj ≤ Lj < skj +1 . We say
2
that a job ri is unimportant to machine j if wi (j) < skj −⌈logs (n )⌉ , important to
2
machine j if skj −⌈logs (n )⌉ ≤ wi (j) < skj +1 and huge to machine j if wi (j) ≥
skj +1 . Note that if ri is huge to machine j, Opt does not schedule ri on machine
j. A job which is important to at least one machine is called important. All
other jobs are called unimportant. Note that, by definition, any unimportant job
is unimportant (and not huge) to the machine where it is scheduled by Opt.
We number the machines such that the first job which is important to machine
j arrives no later than the first job which is important to machine j ′ for every
j < j ′ . This numbering is written to the advice tape, using O(m log m) advice
bits.
The algorithm works in m + 1 phases (some of which might be empty). Phase
0 begins when the first request arrives. For 1 ≤ j ≤ m, phase j − 1 ends and
phase j begins when the first important job for machine j arrives. Note that
the same job could be the first important job for more than one machine. Phase
m ends with the last request of σ. For each phase, j, the oracle writes the
index, i, of the request starting the phase and the unique integer ∆i (j) such
that skj −∆i (j) ≤ wi (j) < skj −∆i (j)+1 .
The unimportant jobs are scheduled arbitrarily by our algorithm (it will
become clear from the analysis of the algorithm that any choice will do).
23
We now describe how the algorithm schedules the important jobs in phase
j for 1 ≤ j ≤ m. By definition, at any point in phase j, we have received an
important job for machines 1, 2, . . . , j and no important job for machine j + 1
has yet arrived.
The type of a job ri in phase j is a vector ∆i = (∆i (1), . . . , ∆i (j)) where
′
′
∆i (j ′ ) is the interval of ri on machine j ′ (so skj −∆i (j ) ≤ wi (j ′ ) < skj −∆i (j )+1 )
or ⊥ if the job is not important to machine j ′ . Note that there are ⌈(2 +
logs (n2 ))j ⌉ possible job types in phase j. The oracle considers how the jobs
in phase j are scheduled in the fixed optimal schedule. For each job type ∆
and each machine 1 ≤ j ′ ≤ j, the oracle encodes the number of jobs of that
type which are scheduled on machine j ′ during phase j. This can be done using O(⌈2 + logs (n2 )⌉m m log n) bits of advice for a single phase, and O(⌈2 +
logs (n2 )⌉m m2 log n) bits of advice for all m phases.
Equipped with the advice described above, the algorithm simply schedules
the important jobs in the current phase based on their types. This ensures that,
for each machine j, the total load of important jobs that Opt schedules on
machine j is at most s times as large as the total load of important jobs scheduled
by Alg on machine j (since if ri and ri′ are important to machine j and of the
same type, then wi (j) < s · wi′ (j)).
In order to finish the proof, we need to show that the contribution of unimportant jobs to Opt(σ) is negligible (recall that all jobs are either important
or unimportant). To this end, let Lunimp
(resp. Limp
j
j ) be the load on machine
j of the unimportant (resp. important) jobs scheduled on that machine in the
+ Limp
optimal schedule. Note that Lj = Lunimp
j . By the definition of an unimj
portant job (and since there trivially can be no more than n unimportant jobs),
we find that for every 1 ≤ j ≤ m,
Lj
.
n
n
Thus, Lj = Lunimp
+ Limp
≤ Lj /n + Limp
from which Lj ≤ n−1
· Limp
follows,
j
j
j
j
assuming that n > 1. Since this holds for all machines, and since as previously
imp
argued Limp
Opt ≤ s · LAlg ≤ s · LAlg , we get that
2
< n · s−⌈logs (n
Lunimp
j
)⌉
· Lj ≤
n
n
Limp
LAlg .
Opt ≤ s
n−1
n−1
By assumption, the objective function f satisfies f (αL) ≤ αf (L) and is nondecreasing. Thus, we conclude that
n
n
n
LAlg ≤ s
f (LAlg ) = s
Alg(σ).
Opt(σ) = f (LOpt ) ≤ f s
n−1
n−1
n−1
LOpt ≤
For n ≥ 2 + 2ε , this gives a ratio of at most 1 + ε.
⊓
⊔
References
1. Susanne Albers and Matthias Hellwig. Online makespan minimization with parallel
schedules. In SWAT, volume 8503 of LNCS, pages 13–25, 2014.
24
2. Hans-Joachim Böckenhauer, Juraj Hromkovič, Dennis Komm, Sacha Krug, Jasmin
Smula, and Andreas Sprock. The string guessing problem as a method to prove
lower bounds on the advice complexity. Theor. Comput. Sci., 554:95–108, 2014.
3. Hans-Joachim Böckenhauer, Dennis Komm, Rastislav Královič, Richard Královič,
and Tobias Mömke. On the advice complexity of online problems. In ISAAC,
volume 5878 of LNCS, pages 331–340, 2009.
4. Allan Borodin, Sandy Irani, Prabhakar Raghavan, and Baruc h Schieber. Competitive paging with locality of reference. Journal of Computer and System Sciences,
50(2):244–258, 1995.
5. Joan Boyar, Leah Epstein, Lene M. Favrholdt, Kim S. Larsen, and Asaf Levin.
Online bounded analysis. In CSR, volume 9691 of LNCS, pages 131–145, 2016.
6. Joan Boyar, Lene M Favrholdt, Christian Kudahl, Kim S Larsen, and Jesper W
Mikkelsen. Online algorithms with advice: A survey. ACM Computing Surveys
(CSUR), 50(2):19, 2017.
7. Joan Boyar, Lene M. Favrholdt, Christian Kudahl, and Jesper W. Mikkelsen. Advice complexity for a class of online problems. In STACS, volume 30 of LIPIcs,
pages 116–129, 2015. Full paper to appear in Theory of Computing Systems.
8. Marek Chrobak and John Noga. LRU is better than FIFO. Algorithmica,
23(2):180–185, 1999.
9. Stefan Dobrev, Rastislav Královič, and Dana Pardubská. Measuring the problemrelevant information in input. RAIRO - Theor. Inf. Appl., 43(3):585–613, 2009.
10. Christoph Dürr, Christian Konrad, and Marc P. Renault. On the power of advice
and randomization for online bipartite matching. In ESA, pages 37:1–37:16, 2016.
11. Yuval Emek, Pierre Fraigniaud, Amos Korman, and Adi Rosén. Online computation with advice. Theor. Comput. Sci., 412(24):2642–2656, 2011.
12. Anna R. Karlin, Mark S. Manasse, Larry Rudolph, and Daniel D. Sleator. Competitive snoopy caching. Algorithmica, 3:77–119, 1988.
13. Dennis Komm, Rastislav Královič, Richard Královič, and Christian Kudahl. Advice complexity of the online induced subgraph problem. In MFCS, volume 58 of
LIPIcs, pages 59:1–59:13, 2016.
14. Jesper W. Mikkelsen. Randomization can be as helpful as a glimpse of the future
in online computation. In ICALP, volume 55 of LIPIcs, pages 39:1–39:14, 2016.
15. Marc P. Renault, Adi Rosén, and Rob van Stee. Online algorithms with advice for
bin packing and scheduling problems. Theor. Comput. Sci., 600:155–170, 2015.
16. Daniel D. Sleator and Robert E. Tarjan. Amortized efficiency of list update and
paging rules. Commun. ACM, 28(2):202–208, 1985.
17. Jasmin Smula. Information Content of Online Problems: Advice versus Determinism and Randomization. PhD thesis, ETH, Zürich, 2015.
18. Andreas Sprock. Analysis of hard problems in reoptimization and online computation. PhD thesis, ETH, Zürich, 2013.
25
Appendix
A
AOC-Complete Problems
For completeness, we state the full definition of minASGk from [7]:
Definition 7 ( [7]). The minimum asymmetric string guessing problem with
known history, minASGk, has input h?, x1 , . . . , xn i, where x = x1 . . . xn ∈
{0, 1}n, for some n ∈ N. For 1 ≤ i ≤ n, round i proceeds as follows:
1. If i > 1, the algorithm learns the correct answer, xi−1 , to the request in the
previous round.
2. The algorithm answers yi = f (x1 , . . . , xi−1 ) ∈ {0, 1}, where f is a function
defined by the algorithm.
The output y = y1 . . . yn computed by the algorithm is feasible, if x ⊑ y. Otherwise, y is infeasible. The cost of a feasible output is |y|1 , and the cost of an
infeasible output is ∞.
In addition to minASGk, the class of AOC-complete problems also contains
many graph problems. The following four graph problems are studied in the
vertex-arrival model, so the requests are vertices, each presented together with
its edges to previous vertices. The first three problems are minimization problems
and the last one is a maximization problem. In Vertex Cover, an algorithm must
accept a set of vertices which constitute a vertex cover, so for every edge in
the requested graph, at least one of its endpoints is accepted. For Dominating
Set, the accepted vertices must constitute a dominating set, so every vertex in
the requested graph must be accepted, or one its neighbors must be accepted.
In Cycle Finding, an algorithm must accept a set of vertices inducing a cyclic
graph. For Independent Set, the accepted vertices must form an independent set,
i.e., no two accepted vertices share an edge.
For Disjoint Path Allocation a path P is given, and the requests are subpaths
of P . The aim is to accept as many edge disjoint paths as possible.
For Set Cover, the requests are finite subsets from a known universe, and the
union of the accepted subsets must be the entire universe. The aim is to accept
as few subsets as possible.
B
Reductions for Theorem 2
In the proof of Theorem 2, a reduction sketch was given for the weighted online
version of Vertex Cover. Here we include sketches for the reductions for the
weighted versions of Cycle Finding, Dominating Set and Set Cover.
Cycle Finding Each input σ = hx1 , x2 , . . . , xn i to the problem minASGk, is
transformed to f (σ) = hv1 , v2 , . . . , vn i, where V = {v1 , v2 , . . . , vn } is the vertex
set of a graph with edge set E = {(vj , vi ) : f ′ (xi ) = j} ∪ {(vmin , vmax )}, where
f ′ (xi ) is the largest j < i such that xj = 1, max is the largest i such that xi = 1,
26
and min is the smallest i such that xi = 1. If |σ|1 > 2, the vertices corresponding
to 1s form the only cycle in the graph.
The advice used by the minASGk algorithm Alg1 consists of the advice
used by the Cycle Finding algorithm Alg2 in combination with 1 bit indicating
whether or not |σ|1 ≤ 2 and in this case (an encoding of) one or two indices
of 1s in the input sequence. If |σ|1 > 2, then Alg2 accepts some vertices, and
Alg1 returns a 1 for the xi corresponding to each of those vertices.
If Alg1 returns a non-optimal feasible set, Alg2 does too, and the sets have
the same weights, so Alg1 (σ) ≤ Alg2 (f (σ)) + Opt(σ). In this case, the weights
of the optimal solutions for σ and f (σ) are both the sum of the weights of the
elements corresponding to 1s in σ, so f is a length preserving O(f (n))-reduction.
Dominating Set Each input σ = hx1 , x2 , . . . , xn i to the problem minASGk, is
transformed to f (σ) = hv1 , v2 , . . . , vn i, where V = {v1 , v2 , . . . , vn } is the vertex
set of a graph with edge set E = {(vi , vmax )}., where max is the largest i such
that xi = 1.
The advice used by the minASGk algorithm Alg1 consists of the advice
used by the Dominating Set algorithm Alg2 in combination with 1 bit indicating
whether or not |σ|1 = 0. If |σ|1 ≥ 1, then there is another bit of advice indicating
whether or not Alg2 accepted vmax . If Alg2 did not accept vmax , the advice also
contains an index of a vertex corresponding to a 0 in σ which was accepted, plus
the index of vmax .
If Alg1 ’s solution is feasible, but not optimal, then |σ|1 > 0 and Alg2
accepts some vertices, and Alg1 returns a 1 for the xi corresponding to each
of those vertices (though, in the case where vmax was rejected, it answers 1 for
xmax and answers 0 for the earlier request indicated by the advice).
If |σ|1 > 0 a minimum weight dominating set for f (σ) consists of exactly
those vertices corresponding to 1s in σ, so the weights of the optimal solutions
for σ and f (σ) are both the sum of the weights of the elements corresponding to
1s in σ, unless Alg2 did not accept vmax . However, the weight of xmax ≤ Opt(σ),
so Alg1 (σ) ≤ Alg2 (f (σ)) + Opt(σ). Thus, f is a length preserving O(log(n)reduction.
Set Cover This reduction is very similar to that for Dominating Set. Each
input σ = hx1 , x2 , . . . , xn i to the problem minASGk, max is the largest i such
that xi = 1. In the set cover instance, the universe is {1, . . . , n}, and f (σ) is a
set of n requests, where request i is {i}, unless i = max, in which case, the set
consists of max and all of the j where xj = 0.
As with the reduction to Dominating Set, this is a length preserving O(log(n)reduction.
27
| 8cs.DS
|
arXiv:1709.09398v1 [math.GR] 27 Sep 2017
ON LINEARITY OF AUTOMORPHISM GROUPS OF
ALGEBRAS AND GROUPS
VITALIĬ ROMAN’KOV
Abstract. The main purpose of this paper is to describe some published results and outline corresponding approaches which when applied to automorphism groups of algebras or groups establish that these
groups are linear or non-linear.
1. On linearity of automorphism groups of groups
There are some useful tests for linearity, such as
(1) A linear group has ascending chain condition on centralizers.
(2) (Malcev) Finitely generated linear groups are residually finite.
(3) (Malcev) A solvable linear group is nilpotent-by-(abelian-by-finite).
(4) (Tits) A linear group either contains a free group of rank two, or is
solvable-by-(locally finite).
1.1. Non-linear automorphism groups of finitely generated solvable groups. In [1], Tits showed that a finitely generated matrix group
either contains a solvable normal subgroup of a finite index (i.e., is almost
solvable) or else contains a noncyclic free subgroup. In [2], Bachmuth and
Mochizuki conjectured that Tits’ alternative is satisfied in every finitely
generated group of automorphisms of a finitely generated solvable group.
They point out that their conjecture holds for solvable abelian-by-nilpotent
groups and in some other cases. It turned out that this conjecture breaks
down, in general. It has been independently shown by Hartley [3] and the
author [4] with absolutely different approaches. We will give these results
and corresponding approaches starting with [4].
Theorem 1.1. (Romankov [4]). The direct wreath product of the group of
IA-automorphisms of an arbitrary finitely generated solvable group with an
infinite cyclic group is embeddable into the group of automorphisms of some
finitely generated solvable group.
Recall that the IA-automorphisms of a group G are those automorphisms
which are identical modulo the commutator group G′ .
Corollary 1.2. If A is an arbitrary finitely generated almost solvable group,
then the wreath product B = AwrC∞ (C∞ denotes the infinite cyclic group)
is embeddable into the group of automorphisms of some finitely generated
solvable group.
1
2
VITALIĬ ROMAN’KOV
The deduced corollary yields a negative answer to the Bachmuth and
Mochizuki’s question since, under its conditions, the Tits’ alternative is
satisfied in the group G if and only if A is solvable. Note that A can be
chosen as arbitrary finite group.
In [2], Bachmuth and Mochizuki also conjectured that in every finitely
generated group of automorphisms of a finitely generated solvable group,
there is a subnormal series of a finite length whose factors are either abelian
or matrix groups over a commutative Noetherian ring.
Corollary 1.3. There exists a finitely generated group of automorphisms
of a finitely generated solvable group which does not have the indicated
property.
The infinite countable direct power G∞ of a simple finite non-abelian
group G is not embeddable into a group with the indicated subnormal series. It is easy to prove this by induction on the length of the series using
the remark of Merzlyakov that the direct product of an infinite number of
non-abelian groups cannot be represented by matrices (see [5], Example 3).
The proofs for fields and for commutative Noetherian rings are identical.
This power G∞ is the basic subgroup of GwrC∞ . By Corollary 1.3, the
latter group is embeddable in the group of automorphisms of some finitely
generated solvable group.
Recall, that a group is called perfect if it equals its derived subgroup. In
[3], a group G is called perfectly distributed, if every subgroup of a finite index
of G contains a non-trivial finitely generated perfect subgroup. Clearly no
perfectly distributed group can be soluble-by-finite.
Theorem 1.4. (Hartley [3]). There exists a finitely generated solvable group
G of derived length three whose automorphism group contains subgroups K ≤
L such that L is finitely generated, L/K is infinite cyclic, K is perfectly
distributed and locally finite.
Moreover, it can be even arranged that K is locally a direct power of any
finite non-abelian simple group. Clearly L is not solvable-by-finite, nor does
it contain a non-abelian free subgroup.
The group G of Theorem 1.4 is actually an extension of a locally finite
group by an infinite cyclic group. But a somewhat more complicated version
of the construction for Theorem 1.4 gives
Theorem 1.5. 3 (Hartley [3]). There exists a finitely generated solvable
group G of derived length four whose automorphism group contains a torsionfree subgroup L having a normal subgroup K such that L is finitely generated,
L/K is infinite cyclic, K is perfectly distributed, every finitely generated subgroup of K is abelian-by-finite.
1.2. On linearity of automorphism groups of finitely generated free
groups. Let Fn be an absolutely free group with basis Xn = {x1 , ..., xn }.
Theorem 1.6. (a) (Krammer [6]) Aut F2 is linear.
ON LINEARITY OF AUTOMORPHISM GROUPS OF ALGEBRAS AND GROUPS
3
(b) (Formanek and Procesi [7]). Aut Fn is not linear for n ≥ 3.
The proof of (b) in [7] uses the representation theory of algebraic groups
to show that a kind of diophantine equation between the irreducible representations of a group G is impossible unless G is abelian-by-finite. This
leads to statement, which says that the HNN-extension
(1)
H(G) =< G × G, t|t(g, g)t1 = (1, g) for all g ∈ G >
cannot be a linear group if G is not nilpotent-by-abelian-by-finite. The
statement (b) then is proved by showing that for n3, the automorphism
group of a free group Fn contains H(F2 ).
Refer as a unitriangular automorphism of Fn with respect to Xn to every
automorphism ϕ determined by a mapping of the form
(2)
ϕ : x1 7→ x1 , xi 7→ vi xi for i = 2, , n,
where vi = vi (x1 , ..., xi−1 is an arbitrary element of Fi−1 = gp(x1 , ..., xi−1 ).
Every collection (v2 , ..., vn ) ∈ F1 × ... × Fn−1 determines an automorphism
ϕ of Fn . The unitriangular automorphisms of Fn constitute a subgroup
UTAutFn , which we call the group of unitriangular automorphisms of Fn .
To simplify expressions, we denote it by Un . It is easy to see that, up to
isomorphism, Un is independent of the choice of Xn .
Theorem 1.7. (Romankov [8]). The group Un of unitriangular automorphisms of the free group Fn of rank is linear if and only if n ≤ 3.
As U1 is trivial and U2 is infinite cyclic, these groups are obviously linear.
In [8], it is shown that U3 is isomorphic to a subgroup of AutF2 . Then by
Krammer’s theorem [6] U3 is linear.
We need two statements about any group G that follow.
Theorem 1.8. (a) (Formanek, Procesi [7]). Let ρ be a linear representation
of H(G). Then the image of G × {1} has a subgroup of finite index with
nilpotent derived subgroup, i.e., is nilpotent-by-abelian-by-finite.
(b) (Brendle, Hamidi-Tehrani [9]). Let N be a normal subgroup of H(G)
such that the image of G × {1} in H(G)/N is not nilpotent-by-abelian-byfinite. Then H(G)/N is not linear.
In [8], a subgroup G of U4 has been constructed that is isomorphic to
a quotient H(F2 )/N such that the image of G × {1} in H(F2 )/N is not
nilpotent-by-abelian-by-finite.
1.3. On linearity of automorphism groups of finitely generated relatively free groups. Given an arbitrary variety C of groups, denote by Gn
the free group in C with a fixed basis Yn = {y1 , ..., yn }. Refer as a unitriangular automorphism of Gn with respect to Yn to every automorphism φ
determined by a mapping of the form
(3)
φ : y1 7→ y1 , yi 7→ ui yi for i = 2, , n,
4
VITALIĬ ROMAN’KOV
where ui = ui (y1 , ..., yi−1 ) is an arbitrary element of Gi−1 = gp(y1 , ..., yi−1 ).
Every collection (u2 , ..., un ) ∈ G1 × ... × G1 × ... × Gn−1 determines an automorphism φ of Gn . The unitriangular automorphisms of Gn constitute
a subgroup UTAut Gn , which we call the group of unitriangular automorphisms of Gn . To simplify expressions, we denote it by Vn . It is easy to see
that, up to isomorphism, Vn is independent of the choice of Yn .
Theorem 1.9. (Auslander and Baumslag [10], Merzlyakov [5]). The automorphism group of every finitely generated nilpotent-by-finite group is linear.
Moreover, the holomorph of a group of this type (hence, its automorphism
group as well) admits a faithful matrix representation over the ring Z of
integers.
This implies in particular that the automorphism groups of a finite rank
relatively free group Gn in nilpotent-by-finite varieties C of groups admit
faithful matrix representations.
Theorem 1.10. (Olshanskii [11]). If a relatively free group Gn is neither
free nor nilpotent-by-finite then AutGn is not linear.
Observe also that Aut G∞ for every nontrivial variety C includes a subgroup consisting of all permutations of an infinite set Y∞ of free generators,
which is isomorphic to the symmetric group on the infinite set. It is well
known that the latter is not linear. Olshanskiis proof of Theorem 1.10 uses
the following reasoning.
For every group G there is a homomorphism σ : G → Aut G associating
to each h ∈ G the inner automorphism σ(h) : f 7→ hf h−1 . The kernel of
σ is the center C(G) of G and the image of σ is the group Inn G of inner
automorphisms of G. The latter is normal in Aut G, which means that we
may regard the quotient G/C(G) ≃ InnG as a normal subgroup of Aut G.
It is shown in [11], that in the case of neither free nor nilpotent-by-finite
relatively free group Gn there exists an automorphism θ of Gn such that
the extension P of Gn /C(Gn ) by means of θ is not linear. Furthermore,
θ was chosen so that the eigenvalues of the induced linear transformation
(Gn )ab = Gn /G′n of abelianization, which in this case, on assuming that Aut
Gn is a linear group, is a rank n free abelian group, were not roots of unity.
The automorphism θ is certainly not unitriangular since all eigenvalues of
each unitriangular automorphism are equal to 1. The group Inn Gn /C(Gn )
also fails to consist of unitriangular automorphisms. Thus, the result of
Olshanskii and his method of proof give no information on the possible
linearity of the group Vn = UTAutGn .
Theorem 1.11. (Erofeev and Romankov [12]).
(a) Let Gn be a rank n ≥ 2 relatively free group in an arbitrary variety
C of groups. The following hold: V1 is trivial, while V2 is the cyclic group
of the order equal to the exponent of C. These groups admit faithful matrix
representations. For n ≥ 3, if Gn1 is nilpotent group then so is Vn . If Gn1
is nilpotent-by-finite then Vn is linear over Z.
ON LINEARITY OF AUTOMORPHISM GROUPS OF ALGEBRAS AND GROUPS
5
(b) Let Gn be a rank n ≥ 3 relatively free group in an arbitrary nontrivial
variety C of groups distinct from the variety of all groups. If Gn1 is not
a nilpotent-by-finite group then the group Vn = UTAut Vn of unitriangular
automorphisms of Gn admits no faithful matrix representation over any field.
Thus, the claims of Theorem 1.11 yield exhaustive information on the
linearity of groups of unitriangular automorphisms of relatively free groups
of finite rank in the proper varieties of groups. Namely, Vn admits a faithful
matrix representation over some field if and only if Gn1 is a nilpotent-byfinite group.
1.4. On linearity of automorphism groups determined by test elements. Let G be a group. An element g ∈ G is called a test element (for
recognizing automorphisms) if, whenever ϕ(g) = g for an endomorphism ϕ
of G, it follows that ϕ is an automorphism. This definition was introduced
by V. Shpilrain in [13]. Let T (G) denote the set of all test elements of G.
We say that a subgroup H = H(g) of AutG is determined by the test
element g ∈ t(G) if H = {ϕ ∈ AutG : ϕ(g) = g}. In fact, H(g) is a stabilizer
of g in AutG. A question arises: what subgroups of AutG are determined
by test elements?
If we define a relation ≤ on the set T (G) by g ≤ f ⇔ H(g) ⊆ H(f ), a
natural question arises about structure of T (G) with the relation ≤ .
More generally, we say that a subset T ⊂ G is a test set if, whenever
ϕ(g) = g for every g ∈ T for an endomorphism ϕ of G, it follows that ϕ is
an automorphism. The test rank t(G) of G is defined as the minimal size
of a test subset of G. Similarly, we introduce the relation ≤ on the set of all
test subsets of G.
Now we just indicate some results about linearity of subgroups determined
by test elements. One of these results is implied by the following theorem.
Theorem 1.12. (Timoshenko [15]). Let Mn be the free metabelian group
of rank n ≥ 2. Then
(a) T (M2 ) = M2′ \ {1};
(b) t(Mn ) = n − 1;
(c) a set of elements {g1 , ..., gn−1 } is a (minimal) test set of Mn if and
only if, when {g1 , ..., gn−1 } ⊆ Mn′ , they are linear independent over the group
ring ZG/G′ and every endomorphism φ fixing each of these elements induces
an automorphism of G/G′ .
Corollary 1.13. Every automorphism φ ∈ AutMn determined by a minimal test set {g1 , ..., gn−1 } mentioned fixes every element u ∈ Mn′ , thus by
Shmel’kin’s theorem [14] is inner corresponding to some element h ∈ Mn′ .
It follows, that every subgroup H = H({g1 , ..., gn−1 }) determined by a
minimal test set {g1 , ..., gn−1 } coincides with the subgroup InnMn′ (Mn ) of
InnMn corresponding to conjugations by elements of Mn′ . Hence, H ≃ Mn′
and clearly it is a linear abelian group.
6
VITALIĬ ROMAN’KOV
It was shown in [18], that in the case of the free solvable group S2,3 of
(2)
rank 2 and class 3 there is a test element g that lies in the second term S2,3
′
(see
of S2,3 . Since this term has dimension 1 over the group ring ZS2,3 /S2,3
Remark below for explanation), it follows that the subgroup H = H(g) of
(2)
AutS2,3 is isomorphic to S2,3 , hence is linear. By [19], a similar assertions
are true for any free solvable group S2,n of rank 2.
Remark 1.14. Let Sn,d be a free solvable group of rank n ≥ 2 and class
(d−1)
d ≥ 2. Then the last non-trivial member Sr,d can be treated as a module
(d−1)
via conjugation. Moreover, this module
over the group ring ZSr,d /Sr,d
can be naturally embedded in the linear space over the division ring (sfield)
(d−1)
T containing ZSr,d /Sr,d The dimension of this linear space is equal n − 1.
(see for instance [16]).
The following statement generalize one of the statements of Theorem 1.12.
Theorem 1.15. (Timoshenko [17]). Let Sn,d be a free solvable group of
rank n ≥ 2 and class d ≥ 2. Then
t(Sn,d ) = n − 1.
More detailed information about minimal test sets of free solvable groups
can be can be obtained by the following theorem.
Theorem 1.16. (Gupta, Timoshenko [19]). Let Fn be a free group of rank
n ≥ 2. Let R be a non-trivial normal subgroup of that lies in Fn′ . Suppose
that the group ring ZFn /R is a domain that satisfies the Ore condition. Then
for any test set {g1 , ..., gn−1 } of Fn /R′ all elements gi belong to R/R′ and
are linear independent over ZFn /R (in other words, over the division ring
of fractions of ZFn /R that exists by our assumptions).
It follows in particular that every minimal test set T = {g1 , ..., gn−1 } of
a free solvable group Sn,d of rank n and class d lies in the last non-trivial
(d−1)
term Sn,d of the derived series of Sn,d , and the subgroup H of AutSn,d
(d−1)
determined by T coincides with Sn,d
and thus is linear.
Problem 1.17. Let Fn be a free group of rank n ≥ 3, and g ∈ Fn be a test
element. Is the corresponding subgroup H(g) linear?
Since AutF2 is linear a similar assertion for F2 is obviously true. The case
F3 is looking as most interesting.
If G∞ is a relatively free group of the countable infinite rank with the
basis Y∞ = {y1 , y2 , ..., yi , ...}, the automorphism group AutG∞ contains the
(non-linear) subgroup S∞ of all permutations of the countable infinite set
Y∞ . Hence AutG∞ is non-linear. Some good properties of Aut G∞ for the
case of arbitrary metabelian or nilpotent-by-abelian variety can be found in
[20].
ON LINEARITY OF AUTOMORPHISM GROUPS OF ALGEBRAS AND GROUPS
7
2. On linearity of automorphism groups of relatively free
algebras
Let F be a field. We restrict exposition to considering the following classical algebras over F (the subscript n ≥ 2 stands for the rank, i.e., the number
of free generating elements): the free Lie algebra Ln , the free associative algebra An , the absolutely free algebra Fn , and the algebra Pn of polynomials.
It is irrelevant whether we consider these algebras with or without identity.
The automorphism group and the tame automorphism group of an algebra
Cn are denoted by Aut Cn and TAutCn , respectively. By definition, the
subgroup TAut Cn is generated in Aut Cn by all elementary automorphisms
and all non degenerate linear changes of generators. Respectively, TAutCn
is called the subgroup of tame automorphisms. By definition, the elementary
automorphisms are of the form
(4)
φ : xi 7→ αxi + f (x1 , ..., xi−1 , xi+1 , ..., xn ), xj 7→ xj for ij,
where i, j = 1, ..., n, α is a nonzero element of F, and f (x1 , ..., xi−1 , , ..., xi+1 , , xn )
is an element of the subalgebra generated by the generators mentioned. It
is well known that if C2 coincides with P2 or A2 then Aut C2 = TAut C2
for the case of an arbitrary field F (see [21] - [25]). If F is a field of characteristic zero, then Aut P3 6= TAutP3 by [26]-[27], and Aut A3 6= TAut A3
by [28]-[29].
An automorphism of an algebra Cn is said to be unitriangular if it has
the form
(5)
ϕ : xi 7→ xi + fi (x1 , ..., xi−1 ),
where i = 1, ..., n, and fi (x1 , ..., xi1 ) is an element of the subalgebra generated by the generators mentioned. Let TUn denote a subgroup of of TAut
Cn which is generated by elementary unitriangular automorphisms of the
form
(6)
φ : xi 7→ xi + f (x1 , ..., xi−1 ), xj 7→ xj for i 6= j,
where i, j = 1, ..., n and f (x1 , ..., xi1 ) is an element of the subalgebra generated by the generators mentioned. We assume that x1 is kept fixed under
any unitriangular automorphism. Formally, if algebra contains constants,
this element may have an image of the form x1 + α, where α ∈ F. Under
such definition, the group of unitriangular automorphisms differs inessentially from the group defined in the way indicated above, which the former
contains as a subgroup. The following result will also be valid.
Theorem 2.1. (Romankov, Chirkov, Shevelin [30]). The group of tame
automorphisms of the free Lie algebra Ln (the free associative algebra An ,
the absolutely free algebra Fn , the algebra Pn of polynomials) of rank n ≥ 4
over a field F of characteristic zero admits no faithful representation by
matrices over any field.
8
VITALIĬ ROMAN’KOV
More exactly, it was established in all these cases that the group of
tame automorphisms contains a solvable subgroup of unitriangular automorphisms TUn in which the commutant of every subgroup of finite index
is not nilpotent. By a theorem of A. I. Malcev, this is impossible in matrix
groups.
Similarly to TUn , we define a group Tn of triangular automorphisms
which is generated by elementary triangular automorphisms of the form
(7)
φ : xi 7→ αxi + f (x1 , ..., xi−1 ), xj 7→ xj for i 6= j,
where i, j = 1, ..., n, and α is a nonzero element of F. The group Tn consists
of all automorphisms of the form
(8)
φ : xi 7→ αi xi + fi (x1 , ..., xi−1 ),
where i = 1, ..., n, αi is a nontrivial element of F, and fi (x1 , ..., xi−1 ) are
elements of the subalgebras generated by the generators specified.
Theorem 2.2. (a) (Sosnovskii [31]). For n ≥ 3 the group AutPn over a
field F of characteristic zero is not linear.
(b) (Bardakov, Neschadim, Sosnovskii [32]). For n ≥ 3 the group AutAn
over a field F of characteristic zero is not linear.
Theorem 2.3. (Romankov [33]). For algebras Pn , An , and Fn every finitely
generated subgroup G of a group of triangular automorhisms admits a faithful representation by triangular matrices over F. Consequently, the group
G is soluble. At the same time, every finitely generated subgroup H of a
group of unitriangular automorphisms admits a faithful representation by
unitriangular matrices over F. Hence the group H is nilpotent.
A shorter version of this paper was published in [34].
References
[1] Tits J. Free subgroups in linear groups // J. Algebra. 1972. Vol. 20, No. 2. P. 250-270.
[2] Bachmuth S., Mochizuki H.Y. Automorphisms of solvable groups // Bull. Am. Math.
Soc. 1975. Vol. 8, No. 2. P. 420-422.
[3] Hartley B. A conjecture of Bachmuth and Mochizuki on automorphisms of soluble
groups // Can. J. Math. 1976. V. 28, No. 6. P. 1302-1310.
[4] Romankov V.A. Embedding some interplacings in groups of automorphisms of finitely
generated solvable groups // Algebra and Logic. 1976. Vol. 15. P. 187-192.
[5] Merzlyakov Yu.I. Matrix representation of automorphisms, extensions, and solvable
groups // Algebra and Logic. 1968. Vol. 7, No. 3. P. 169-192.
[6] Krammer D.The hypercenter of linear groups // Invent. Math. 2000. Vol. 142, No. 3.
P. 451-586.
[7] Formanek E., Procesi C. The automorphism groups of a free group is not linear //
J. Algebra. 1992. Vol.149, No. 2. P. 494499.
[8] Romankov V. A. The linearity problem for the unitriangular automorphism groups of
free groups // Journal of Siberian Federal University. Mathematics & Physics. 2013.
Vol. 6, No. 4. P. 516520.
[9] Brendle T. E., Hamidi-Tehrani H. On the linearity problem for mapping class groups
// Algebraic and Geometric Topology. 2001. Vol. 1. P. 445-468.
ON LINEARITY OF AUTOMORPHISM GROUPS OF ALGEBRAS AND GROUPS
9
[10] Auslander L., Baumslag G. Automorphism groups of finitely generated nilpotent
groups // Bull. Amer. Math. Soc. 1967. Vol. 73. P. 716717.
[11] Olshanskii A. Yu. Linear automorphism groups of relatively free groups // Turkish
J. Math. 2007. Vol. 31. P. 105111.
[12] Erofeev S.Yu., Romankov V.A. On the groups of unitriangular automorphisms of
relatively free groups // Siberian Math. J. 2012. Vol. 53, No. 5. P. 792-799.
[13] Shpilrain V. Recognizing automorphisms of the free groups// Arch. Math. 1994. Vol.
62. P. 385-392.
[14] Shmel’kin A.L. Two notes on free solvable groups // Algebra i Logika. 1967. Vol. 6,
No. 2. P. 95109. (Russian).
[15] Timoshenko E.I. Test elements and test rank of a free metabelian group// Siberian
Math. J. 2000. Vol. 41, No. 6. P. 1200-1204.
[16] Roman’kov V.A. Normal automorphisms of discrete groups// Siberian Math. J. 1983.
Vol. 24, No. 4. P. 138-149.
[17] Timoshenko E.I. Computing test rank for a free solvable group// Algebra and Logic.
2006. Vol. 45, No. 4. P. 254-264.
[18] Roman’kov V.A. Test elements for free solvable groups of rank 2// Algebra and Logic.
2001. Vol. 40, No. 2. 106-111.
[19] Gupta C.K., Timoshenko E.I. Test rank for some free polynilpotent groups// Algebra
and Logic. 2003. Vol. 42, No. 1 P. 20-28.
[20] Bryant R.M. and Roman’kov V.A. Automorphism groups of relatively free groups//
Proc. Camb. Philos. Soc. Vol. 127, No. 3. 411-424.
[21] Czerniakiewicz A.J. Automorphisms of a free associative algebra of rank 2. I // Trans.
Amer. Math. Soc. 1971. Vol. 160. P. 393-401.
[22] Czerniakiewicz A.J. Automorphisms of a free associative algebra of rank 2. II //
Trans. Amer. Math. Soc. 1972. Vol. 171. P. 309-315.
[23] Jung H.W. Uber ganze birationale Transformationen der Ebene // J. Reine Angew.
Math. 1942. Vol. 184. P. 161-174.
[24] Kulk W. van der. On polynomial rings in two variables // Nieuw Arch. Wiskd. 1953.
III. Ser., No. 1. P. 33-41.
[25] Macar-Limanov L.G. Automorphisms of free algebra with two generators// Funct.
analysis and applications. 1970. Vol. 4. P. 107-108. (Russian).
[26] Umirbaev U.U., Shestakov I.P. Subalgebras and automorphisms of polynomial rings//
Dokl. RAS. 2002. Vol. 386, No. 6. P. 745-748.
[27] Shestakov I.P., Umirbaev U.U. The tame and the wild automorphisms of polynomial
rings in three variables // J. Amer. Math. Soc. 2004. Vol. 17, No. 1. P. 197-227.
[28] Umirbaev U.U. Tame and wild automorphisms of polynomial algebras and free associative algebras // Preprint of Max Plank Inst. fur Math. 2004. No. 108, MPIM,
Bonn.
[29] Umirbaev U.U. The Anick automorphism of free associative algebras // J. Reine
Angew. Math. 2007. Vol. 605. P. 165-178.
[30] Roman kov V. A., Chirkov I. V., Shevelin M. A. Nonlinearity of the automorphism
groups of some free algebras // Siberian Math. J. 2004. Vol. 45, No. 5. P. 974977.
[31] Sosnovskii Yu. V. The hypercentral structure of the group of unitriangular automorphisms of a polynomial algebra // Siberian Math. J. 2007. Vol. 48, No. 3. P. 555558.
[32] Bardakov V.G., Neshadim M.V., Sosnovsky Yu.V. Groups of triangular automorphisms of a free associative algebra and a polynomial algebra // J. Algebra. 2012.
Vol. 362. P. 201-220.
[33] Romankov V. A. The local structure of groups of triangular automorphisms of relatively free algebras // Algebra and Logic. 2012. Vol. 56, No. 5. P. 425-434.
[34] Roman’kov V.A. A short survey on linearity of automorphism groups of algebras and
groups// Herald of Omsk University. 2016. 1. P. 22-25.
10
VITALIĬ ROMAN’KOV
Institute of Mathematics and Information Technologies, Dostoevsky Omsk
State University
E-mail address: [email protected]
| 4math.GR
|
arXiv:1706.08213v1 [math.GR] 26 Jun 2017
Idempotent ordered semigroup
K. Hansda
Department of Mathematics, Visva-Bharati University,
Santiniketan, Bolpur - 731235, West Bengal, India
[email protected]
June 27, 2017
Abstract
An element e of an ordered semigroup (S, ·, ≤) is called an ordered idempotent if e ≤ e2 . We
call an ordered semigroup S idempotent ordered semigroup if every element of S is an ordered
idempotent. Every idempotent semigroup is a complete semilattice of rectangular idempotent
semigroups and in this way we arrive to many other important classes of idempotent ordered
semigroups.
1
Introduction
Idempotents play an important role in different major subclasses of a regular semigroup. A regular
semigroup S is called orthodox if the set of all idempotents E(S) forms a subsemigroup, and S is a
band if S = E(S).
T. Saito studied systematically the influence of order on idempotent semigroup [8]. In [1],
Bhuniya and Hansda introduced the notion of ordered idempotents and studied different classes of
regular ordered semigroups, such as, completely regular, Clifford and left Clifford ordered semigroups
by their ordered idempotents. The purpose of this paper to study ordered semigroups in which
every element is an ordered idempotent. Complete semilattice decomposition of these semigroups
automatically suggests the looks of rectangular idempotent semigroups and in this way we arrive to
many other important classes of idempotent ordered semigroups.
1
The presentation of the article is as follows. This section is followed by preliminaries. In
Section 3, is devoted to the idempotent ordered semigroups and characterizations of different type of
idempotent ordered semigroups.
2
Preliminaries
An ordered semigroup is a partially ordered set (S, ≤), and at the same time a semigroup (S, ·) such
that for all a, b, c ∈ S; a ≤ b implies ca ≤ cb and ac ≤ bc. It is denoted by (S, ·, ≤). Throughout this
article, unless stated otherwise, S stands for an ordered semigroup. For every subset H ⊆ S, denote
(H] = {t ∈ S : t ≤ h, for some h ∈ H}. Let I be a nonempty subset of an ordered semigroup S. I
is a left (right) ideal of S, if SI ⊆ I (IS ⊆ I) and (I] = I. I is an ideal of S if I is both a left and
a right ideal of S. An (left, right) ideal I of S is proper if I 6= S. S is left (right) simple if it does
not contain proper left (right) ideals. S is called t-simple if it is both left and right simple and S is
called simple if it does not contain any proper ideals.
Kehayopulu [2] defined Green’s relations on a regular ordered semigroup S as follows:
aLb if (S 1 a] = (S 1 b], aRb if (aS 1 ] = (bS 1 ], aJ b if (S 1 aS 1 ] = (S 1 bS 1 ], and H = L ∩ R.
These four relations L, R, J and H are equivalence relations.
An equivalence relation ρ on S is called left (right) congruence if for every a, b, c ∈ S; a ρ b implies
ca ρ cb (ac ρ bc). By a congruence we mean both left and right congruence. A congruence ρ is called
a semilattice congruence on S if for all a, b ∈ S, a ρ a2 and ab ρ ba. By a complete semilattice
congruence we mean a semilattice congruence σ on S such that for a, b ∈ S, a ≤ b implies that aσab.
Equivalently: There exists a semilattice Y and a family of subsemigroups {S}α∈Y of type τ of S such
that:
1. Sα ∩ Sβ = φ for any α, β ∈ Y with α 6= β,
S
2. S = α ∈ Y Sα ,
3. Sα Sβ ⊆ Sα β for any α, β ∈ Y,
4. Sβ ∩ (Sα ] 6= φ implies β α, where is the order of the semilattice Y defined by
:= {(α, β) | α = α β (β α)} [7].
An ordered semigroup S is called left regular if for every a ∈ S, a ∈ (Sa2 ]. An element e
of (S, ·, ≤) is called an ordered idempotent [1] if e ≤ e2 . As an immediate example of idempotent
ordered semigroups, we can consider (N, ·, ≤) which is an idempotent ordered semigroup but (N, ·)
2
is not an idempotent semigroup. An ordered semigroup S is called H−commutative if for every
a, b ∈ S, ab ∈ (bSa].
If F is a semigroup, then the set Pf (F ) of all finite subsets of F is a semilattice ordered semigroup
with respect to the product ′ ·′ and partial order relation ′ ≤′ given by: for A, B ∈ Pf (F ),
A · B = {ab | a ∈ A, b ∈ B} and A ≤ B if and only if A ⊆ B.
3
Idempotent ordered semigroups
We have discussed the elementary properties of ordered idempotents. In this section we characterize
ordered semigroups of which every element is an ordered idempotent. Here we show that these
ordered semigroups are analogous to bands.
We now make a natural analogy between band and idempotent ordered semigroup.
Theorem 3.1. Let B be a semigroup. Then Pf (B) is idempotent ordered semigroup if and only if B
is a band.
Proof. Let B be a band and U ∈ Pf (B). Choose x ∈ U. Then x2 ∈ U 2 implies x ∈ U 2 . Then
U ⊆ U 2 . So Pf (B) is idempotent ordered semigroup.
Conversely, assume that B be a semigroup such that Pf (B) is an idempotent ordered semigroup.
Take y ∈ B. Then Y = {y} ∈ Pf (B). Thus Y ⊆ Y 2 , which implies y = y 2. Hence B is a band.
Proposition 3.2. Let B be a band, S be an idempotent ordered semigroup and f : B −→ S be a
semigroup homomorphism. Then there is an ordered semigroup homomorphism φ : Pf (B) −→ S
such that the following diagram is commutative:
f
B
l
S
φ
Pf (B)
where l : B −→ Pf (B) is given by l(x) = {x}.
Proof. Define φ : Pf (F ) −→ S by: for A ∈ Pf (F ), φ(A) = ∨a∈A f (a). Then for every A, B ∈
Pf (F ), φ(AB) = ∨a∈A,b∈B f (ab) = ∨a∈A,b∈B f (a)f (b) = (∨a∈A f (a))(∨b∈B f (b)) = φ(A)φ(B), and
if A ≤ B, then φ(A) = ∨a∈A f (a) ≤ ∨b∈B f (b) = φ(B) shows that φ is an ordered semigroup
homomorphism. Also φ ◦ l = f .
3
Lemma 3.3. In an idempotent ordered semigroup S, am ≤ an f or every a ∈ S and m, n ∈
N with m ≤ n.
Every idempotent ordered semigroup S is completely regular and hence J is the least complete
semilattice congruence on S, by Lemma ??. In an idempotent ordered semigroup S, the Green’s
relation J can equivalently be expressed as: for a, b ∈ S,
aJ b if there are x, y, u, v ∈ S such that a ≤ axbya and b ≤ buavb.
Now we characterize the J −class in an idempotent ordered semigroup.
Definition 3.4. An idempotent ordered semigroup S is called rectangular if for all a, b ∈ S, there
are x, y ∈ S such that a ≤ axbya.
Example 3.5. (N, ·, ≤) is a rectangular idempotent ordered semigroup,whereas if we define a o b =
min{a, b} f or all a, b ∈ N then (N, ◦, ≤) is an idempotent ordered semigroup but not rectangular.
Also we have the following equivalent conditions.
Lemma 3.6. Let S be an idempotent ordered semigroup. Then the following conditions are equivalent:
1. S is rectangular;
2. for all a, b ∈ S, there is x ∈ S such that a ≤ axbxa;
3. for all a, b, c ∈ S there is x ∈ S such that ac ≤ axbxc.
Proof. (1) ⇒ (3): Let a, b, c ∈ S. Then there are x, y ∈ S such that a ≤ axbya. This implies that ac ≤
axbyac ≤ ax(bya)(bya)c ≤ (axbyab)(axbyab)yac ≤ a(axbyabya)b(axbyabya)c ≤ atbtc, where t =
axbyabya ∈ S.
(3) ⇒ (2): Let a, b ∈ S. Then there is x ∈ S such that a2 ≤ axbxa. Then a ≤ a2 implies that
a ≤ axbxa.
(2) ⇒ (1): This follows directly.
As we can expect, we show that the equivalence classes in an idempotent ordered semigroup S
determined by J are rectangular.
Theorem 3.7. Every idempotent ordered semigroup is a complete semilattice of rectangular idempotent ordered semigroups.
4
Proof. Let S be an idempotent ordered semigroup. Then J is the least complete semilattice congruence on S. Now consider a J -class (c)J for some c ∈ S. Since J is a complete semilattice congruence,
(c)J is a subsemigroup of S. Let a, b ∈ (c)J . Then there is x ∈ S such that a ≤ axbxa, which implies
that a ≤ a(axb)b(bxa)a, that is, a ≤ aubva where u = axb and v = bxa. Also the completeness of
J implies that (a)J = (a2 xbxa)J = (axb)J = (bxa)J , and u, v ∈ (c)J . Thus (c)J is a rectangular
idempotent ordered semigroup.
Definition 3.8. An idempotent ordered semigroup S is called left zero if for every a, b ∈ S, there
exists x ∈ S such that a ≤ axb.
An idempotent ordered semigroup S is called right zero if for every a, b ∈ S, there is x ∈ S such
that a ≤ bxa.
Now it is evident that every left zero and right zero idempotent ordered semigroup is rectangular.
Proposition 3.9. An idempotent ordered semigroup is left zero if and only if it is left simple.
Proof. First suppose that S is a left zero idempotent ordered semigroup and a ∈ S. Then for any
b ∈ S, there is x ∈ S such that b ≤ bxa, which shows that b ∈ (Sa]. Thus S = (Sa] and hence S is
left simple.
Conversely, assume that S is left simple. So for every a, b ∈ S, there is s ∈ S such that a ≤ sb.
Then a ≤ a2 gives that a2 ≤ asb. Thus S is a left zero idempotent ordered semigroup.
Lemma 3.10. In an idempotent ordered semigroup S, the following conditions are equivalent:
1. For all a, b ∈ S, there is x ∈ S such that ab ≤ abxba;
2. For all a, b ∈ S, there is x ∈ S such that ab ≤ axbxa;
3. For all a, b ∈ S, there is x, y ∈ S such that ab ≤ axbya.
Proof. (1) ⇒ (3): This follows directly.
(3) ⇒ (2): This is similar to the Lemma 3.6.
(2) ⇒ (1): Let a, b ∈ S. Then there is x ∈ S such that bab ≤ baxbxba. Now since ab ≤
ababab, we have ab ≤ ab(abaxbx)ba.
We now introduce the notion of left regularity in an idempotent ordered semigroup.
Definition 3.11. An idempotent ordered semigroup S is called left regular if for every a, b ∈ S there
is x ∈ S such that ab ≤ axbxa.
5
Theorem 3.12. An idempotent ordered semigroup S is left regular if and only if L = J is the least
complete semilattice congruence on S.
Proof. First we assume that S is left regular. Let a, b ∈ S be such that aJ b. Then there are
u, v, x, y ∈ S such that a ≤ ubv and b ≤ xay. Since S is left regular, there are s, t ∈ S such
that bv ≤ bsvsb and ay ≤ atyta. Then a ≤ ubsvsb and b ≤ xatyta; which shows that aLb. Thus
J ⊆ L. Again L ⊆ J on every ordered semigroup and hence L = J . Since every idempotent ordered
semigroup is completely regular, it follows that L is the least complete semilattice congruence on S,
by Lemma ??
Conversely, let L is the least complete semilattice congruence on S. Consider a, b ∈ S. Then
abLba implies that ab ≤ xba for some x ∈ S. This implies that
ab ≤ abab ≤ abxba.
Thus S is a left regular idempotent ordered semigroup, by Lemma 3.10.
In the following result a left regular idempotent ordered semigroup has been characterized by
left zero idempotent ordered semigroup.
Theorem 3.13. Let S be an idempotent ordered semigroup. Then the following conditions are
equivalent:
1. S is left regular;
2. S is a complete semilattice of left zero idempotent ordered semigroups ;
3. S is a semilattice of left zero idempotent ordered semigroups.
Proof. (1) ⇒ (2): In view of Theorem 3.12, it is sufficient to show that each L-class is a left zero
idempotent ordered semigroup. Let L be an L-class and a, b ∈ L. Then L is a subsemigroup,
since L is a semilattice congruence. Since aLb there is x ∈ S such that a ≤ xb. This implies that
a ≤ a3 ≤ a2 xb ≤ a2 xb2 ≤ aub, where u = axb.
By the completeness of L, a ≤ xb implies that (a)L = (axb)L , and hence u ∈ L. Thus S is left
zero idempotent ordered semigroup.
(2) ⇒ (3): Trivial.
(3) ⇒ (1): Let ρ be a semilattice congruence on S such that each ρ-class is a left zero idempotent
ordered semigroup. Consider a, b ∈ S. Then ab, ba ∈ (ab)ρ shows that there is s ∈ (ab)ρ such that
ab ≤ absba ≤ absbsba ≤ a(bsb)b(bsb)a. Hence S is left regular.
6
The following theorem gives several alternative characterizations of an H−commutative idempotent ordered semigroup.
Lemma 3.14. Let S be an idempotent ordered semigroup. Then the following conditions are equivalent:
1. S is H-commutative;
2. for all a, b ∈ S, ab ∈ (baS] ∩ (Sba];
3. S is a complete semilattice of t-simple idempotent ordered semigroups;
4. S is a semilattice of t-simple idempotent ordered semigroups.
Proof. (1) ⇒ (2): Consider a, b ∈ S. Since S is H− commutative, there is
u ∈ S such that ab ≤ bua. Also for u, a ∈ S, ua ≤ asu for some s ∈ S. Thus ab ≤ basu, which shows
that ab ∈ (baS]. Similarly ab ∈ (Sba]. Hence ab ∈ (baS] ∩ (Sba].
(2) ⇒ (3): Suppose that J be an J -class in S and a, b ∈ J. Since J is rectangular there is
x ∈ J such that a ≤ axbxa. Also by the given condition (2) there is u ∈ J such that bxa ≤ xaub.
So a ≤ ax2 aub ≤ vb where v = ax2 au. Since J is a complete semilattice congruence on S, (a)J =
(a2 x2 aub)J = (ax2 au)J = (v)J . So v ∈ J. This shows that J is left simple. Similarly it can be
shown that J is also right simple. Thus S is a complete semilattice of t-simple idempotent ordered
semigroups.
(3) ⇒ (4): This follows trivially.
(4) ⇒ (1): Let S be the semilattice Y of t-simple idempotent ordered semigroups {Sα }α∈Y and ρ
be the corresponding semilattice congruence on S. Then there are α, β ∈ Y such that a ∈ Sα and b ∈
Sβ . Then ba, ab ∈ Sαβ . Since Sαβ is t-simple, ab ≤ xba for some x ∈ Sαβ . Now for x, ba ∈ Sαβ there
is y ∈ Sαβ such that x ≤ bay. This finally gives ab ≤ bta, where t = ayb.
Definition 3.15. An idempotent ordered semigroup (S, ., ≤) is called weakly commutative if for any
a, b ∈ S there exists u ∈ S such that ab ≤ bua.
Theorem 3.16. For an idempotent ordered semigroup S, the followings are equivalent:
1. S is weakly commutative;
2. for any a, b ∈ S, ab ∈ (baS] ∩ (sba];
3. S is complete semilattice of left and right simple idempotent ordered semigroups .
Proof. (1) ⇒ (2): Let a, b ∈ S. Then ∃u ∈ S s.t ab ≤ bua for u, a ∈ S, there exists z ∈
S such that ua ≤ azu. Thus ab ≤ bua ≤ baza f or za ∈ S. So ab ≤ (baS]. Similarly ab ∈ (Sba].
Hence ab ∈ (baS] ∩ (sba].
7
(2) ⇒ (3): Since S is an idempotent ordered semigroup, we have by theorem[] ρ is complete
semilattice congruence. We now have to show that, for each z ∈ S, J = (z)ρ is left and right simple.
For this let us choose a, b ∈ J. Then there exists x, y ∈ S such that a ≤ axbya. So from the given
condition bya ∈ (syab] and therefore there is s1 ∈ S such that bya ≤ s1 yab. Therefore a ≤ axs1 yab.
Now since ρ is complete semilattice congruence on S, we have (a)ρ = (a2 xs1 yab)ρ = (axs1 yab)ρ =
(axs1 ya)ρ. Thus a ≤ ub, where u = axs1 ya ∈ J. Hence J is left simple and similarly it is right
simple.
(3) ⇒ (1): Let S is complete semilattice Y of left and right simple idempotent ordered semigroups {Sα }α∈Y . Thus S = {Sα }α∈Y . Take a, b ∈ S. Then there are α, β ∈ Y such that a ∈
Sα and b ∈ Sβ . Thus ab ∈ Sαβ . So ab, ba ≤ Sαβ . Then there are u, v ∈ Sαβ such that ab ≤
uba and ab ≤ bav implies ab ≤ ab2 ≤ bta, where t = avub. Hence S is weakly commutative. This
completes the proof.
Definition 3.17. An idempotent ordered semigroup (S, ·, ≤) is called normal if for any a, b, c ∈
S, ∃ x ∈ S such that abca ≤ acxba.
Theorem 3.18. For an idempotent ordered semigroup S, the followings are equivalent:
1. S is normal;
2. aSb is weakly commutative, for any a, b ∈ S;
3. aSa is weakly commutative, for any a ∈ S.
Proof. (1) ⇒ (2): Consider axb, ayb ∈ aSb f or x, y ∈ S. As S is normal, ∃u, v ∈ S such that
(axb)(ayb) ≤ (axb)(ayb)(axb)(ayb) ≤ aybuxba2 xbayb, f or xba, yb ∈ S ≤ (ayb)uxb(bay)v(a2 x)b, f or a2 x, bay ∈
S ≤ (ayb)(uxb2 ayva)(axb). This implies (axb)(ayb) ≤ (ayb)t(axb) ≤ (ayb)(ayb)t(axb)(axb), t =
uxb2 ayva and thus (axb)(ayb) ≤ aybsaxb, where s = aybtaxb ∈ aSb. Thus aSb is weakly commutative.
(2) ⇒ (3): This is obvious by taking b = a.
(3) ⇒ (1): Let a, b, c ∈ S. Then abca, aca ∈ aSa. Since aSa is weakly commutative. Then there
is s ∈ aSa such that (abca)aca ≤ acasaabca. Now for aba, abca ∈ aSa, there is t ∈ aSa such that
abaabca ≤ abcataba. Thus abca ≤ (abca)(abca) ≤ abca2 ca2 bca ≤ abca2 ca2 ba2 bca = (abcaaca)(abaabca) ≤
(acasa2 ca)(abcataba) ≤ acuba; where u = asa2 bca2 bcata ∈ S. Hence S is normal.
Definition 3.19. An idempotent ordered semigroup (S, ·, ≤) is called left normal (right normal) if
for any a, b, c ∈ S, there exists x ∈ S such that abc ≤ acxb (abc ≤ bxac).
8
Theorem 3.20. Let S be a left normal idempotent ordered semigroup, then
1. L is the least complete semilattice congruence on S;
2. S is a complete semilattice of LZidempotent ordered semigroups.
Proof. (1): Let a, b ∈ S such that aρb. Then there are x, y, u, v ∈ S such that
a ≤ a(xbya), b ≤ b(uavb).
(1)
Since S is left normal, we have for x, b, ya ∈ S, xbya ≤ xyatb for some t ∈ S. Similarly there is
s ∈ S such that uavb ≤ uvbsa. So from 1, a ≤ (axyat)b and b ≤ (buvbs)a. Hence aLb. Thus ρ ⊆ L.
Again, let a, b ∈ S such that aLb. Thus there re u, v ∈ S such that a ≤ ub and b ≤ va. By
lemma[], we have a ≤ a3 = aaa ≤ auba ≤ aubba for some u, b ∈ S. Therefore aρb. Thus L ⊆ ρ.
Thus L = ρ.
(2): Here we are only to proof that each L-class is a left zero. For this let L-class (x)L =
L, (say) for some x ∈ S. Clearly (x)L is a subsemigroup of S. Take a, b ∈ L. Then y, z ∈
S such that a ≤ yb, b ≤ za. Since S is left normal, there is t ∈ S such that a ≤ yb ≤ (yb)b ≤ yzab.
This implies a ≤ a2 ≤ a(ayzb)b. Thus (a)L = (a2 yzb)L = (ayzb)L . Therefore L is left zero.
Hence S is a complete semilattice of left zero idempotent ordered semigroups.
Theorem 3.21. Let S be a idempotent ordered semigroup, then S is normal if and only if L is right
normal band congruence and R is left normal band congruence.
Proof. First we shall see that L is left congruence on S. For this let us take a, b ∈ S such that aLb and c ∈
S. Then there is x, y ∈ S such that a ≤ xb, b ≤ ya. Now as S is normal idempotent ordered semigroup, ca ≤ cxb ≤ cxbcxb ≤ cxbx(s1 )cb for some s1 ∈ S. Thus a ≤ s2 cb where s2 = cxbxs1 ∈ S.
Again cb ≤ s4 ca where s4 = cyays3 ∈ S. So caLcb. It finally shows that L is congruence on S.
Similarly it can be shown that R is congruence on S.
Next consider that a, b, c ∈ S are arbitrary. Then since S is normal idempotent ordered semigroup,
abc ≤ abcabc ≤ abcbt1 ac ≤ acb(t1 t2 bac) for some acbt1 t2 ∈ S. Also bac ≤ bacbac ≤ bacat3 bc ≤
(bct3 t4 abc) for some bct3 t4 ∈ S. So abcLbac. Similarly abcRacb. This two relations respectively
shows that L is right normal band congruence and R is left normal band congruence.
Conversely, suppose that L is right normal band congruence and R is left normal band congruence. Consider a, b, c ∈ S. Then abcRacb and bcaLcba. Then ∃x1 , x2 ∈ S such that
abc ≤ (acb)x1 and bca ≤ x2 cba.
9
Now then abc ≤ (abc)bca ≤ (acb)x1 bca ≤ ac(bx1 x2 c)ba f or some bx1 x2 c ∈ S. Hence S is a Nidempotent ordered semigroup.
References
[1] A.K.Bhuniya and K. Hansda, Complete semilattice of ordered semigroups, arXiv:1701.01282v1.
[2] N. Kehayopulu, Note on Green’s relation in ordered semigroup , Math. Japonica, 36(1991),
211-214.
[3] N. Kehayopulu, On completely regular poe-semigroups, Math. Japonica 37(1992), 123-130.
[4] N. Kehayopulu, On regular duo ordered semigroups, Math. Japonica, 37(1992), 535-540.
[5] N. Kehayopulu, On intra-regular ordered semigroups, Semigroup Forum, 46(1993), 271-278.
[6] N. Kehayopulu, On completely regular ordered semigroups, Scinetiae Mathematicae 1(1)(1998),
27-32.
[7] N. Kehayopulu and M. Tsingelis, Semilattices of Archimedean ordered semigroups, Algera Colloquium 15:3(2008), 527-540.
[8] T. Saito, Ordered idempotent semigroups, J. Math. Soc. Japan 14(2)(1962), 150-169.
10
| 4math.GR
|
arXiv:1603.03987v1 [cs.SC] 13 Mar 2016
Binomial Difference Ideals
Xiao-Shan Gao, Zhang Huang, Chun-Ming Yuan
KLMM, Academy of Mathematics and Systems Science
The Chinese Academy of Sciences, Beijing 100190, China
Abstract
In this paper, binomial difference ideals are studied. Three canonical representations
for Laurent binomial difference ideals are given in terms of the reduced Gröbner basis
of Z[x]-lattices, regular and coherent difference ascending chains, and partial characters
over Z[x]-lattices, respectively. Criteria for a Laurent binomial difference ideal to be
reflexive, prime, well-mixed, and perfect are given in terms of their support lattices.
The reflexive, well-mixed, and perfect closures of a Laurent binomial difference ideal
are shown to be binomial. Most of the properties of Laurent binomial difference ideals
are extended to the case of difference binomial ideals. Finally, algorithms are given to
check whether a given Laurent binomial difference ideal I is reflexive, prime, well-mixed,
or perfect, and in the negative case, to compute the reflexive, well-mixed, and perfect
closures of I. An algorithm is given to decompose a finitely generated perfect binomial
difference ideal as the intersection of reflexive prime binomial difference ideals.
Keywords. Laurent binomial difference ideal, binomial difference ideal, Z[x]-lattice,
difference characteristic set, Gröbner basis of Z[x]-module, generalized Hermite normal
form.
1
Introduction
A polynomial ideal is called binomial if it is generated by polynomials with at most two
terms. Binomial ideals were first studied by Eisendbud and Sturmfels [7], which were further
studied in [1, 6, 19, 16, 23] and were applied in algebraic statistics [22], chemical reactions
[20], and error-correcting codes [25].
In this paper, we initiate the study of binomial difference ideals and hope that they will
play similar roles in difference algebraic geometry to their algebraic counterparts. Difference
algebra and difference algebraic geometry were founded by Ritt [24] and Cohn [4], who
aimed to study algebraic difference equations in the way polynomial equations were studied
in commutative algebra and algebraic geometry [4, 13, 17, 26].
We now describe the main results of this paper. In Section 3, we prove basic properties
of Z[x]-lattices. By a Z[x]-lattice, we mean a Z[x]-module in Z[x]n . Z[x]-lattices play the
same role as Z-lattices do in the study of binomial ideals. Here, x is used to denote the
difference operator σ. For instance, a3 σ(a)2 is denoted as a2x+3 . Since Z[x] is not a PID,
the Hermite normal form for a matrix with entries in Z[x] does not exist. In this section,
we introduce the concept of generalized Hermite normal form and show that a matrix is a
1
generalized Hermite normal form if and only if its columns form a reduced Gröbner basis for
a Z[x]-lattice.
In Section 4, we give three canonical representations for Laurent binomial difference
ideals in terms of reduced Gröbner bases of Z[x]-lattices, difference characteristic sets, and
partial characters. Gröbner bases play an important role in the study of binomial ideals [7].
In general, a binomial difference ideal is not finitely generated and does not have a finite
Gröbner basis. Instead, the theory of characteristic set for difference polynomial systems [9]
is used for similar purposes. It is shown that any Laurent binomial difference ideal can be
written as [A], where A is a regular and coherent difference ascending chain consisting of
binomial difference polynomials.
Let I be a proper Laurent binomial difference ideal and L = {f ∈ Z[x]n | Yf − cf ∈ I}
the support lattice of I, which is a Z[x]-lattice. In Section 5, we give criteria for a Laurent
binomial difference ideal to be prime, reflexive, well-mixed, and perfect in terms of its support
lattice. The criterion for prime ideals is similar to the algebraic case, but the criteria for
reflexive, well-mixed, and perfect difference ideals are unique to difference algebra and are
first proposed in this paper. Furthermore, it is shown that the reflexive, well-mixed, and
perfect closures of a Laurent binomial difference ideal I with support lattice L are still
binomial, whose support lattices are the x-, M -, and the P -saturation of L, respectively. It
is further shown that any perfect Laurent binomial difference ideal I can be written as the
intersection of Laurent reflexive prime binomial difference ideals whose support lattices are
the x-Z-saturation of the support lattice of I.
In Section 6, binomial difference ideals are studied. It is shown that a large portion of the
properties for binomial ideals proved in [7] can be easily extended to the difference case. We
also identify a class of normal binomial difference ideals which are in a one to one correspondence with Laurent binomial difference ideals. With the help of this correspondence, most
properties proved for Laurent binomial difference ideals are extended to the non-Laurent
case.
In Section 7, algorithms are given to check whether a Z[x]-lattice L is Z-, x-, M-, or Psaturated, or equivalently, whether a Laurent binomial difference ideal I is prime, reflexive,
well-mixed, or perfect. If the answer is negative, we can also compute the Z-, x-, M-,
or P-saturation of L and the reflexive, well-mixed, or perfect closures of I. Based on these
algorithms, we give an algorithm to decompose a finitely generated perfect binomial difference
ideal as the intersection of reflexive prime binomial difference ideals. This algorithm is
stronger than the general decomposition algorithm in that for general difference polynomials,
it is still open on how to decompose a finitely generated perfect difference ideal as the
intersection of reflexive prime difference ideals [9].
A distinctive feature of the algorithms presented in this paper is that problems about
difference binomial polynomial ideals are reduced to problems about Z[x]-lattices which are
pure algebraic and have simpler structures.
2
2
Preliminaries about difference algebra
In this section, some basic notations about difference algebra will be given. For more details
about difference algebra, please refer to [4, 9, 13, 17, 26].
2.1
Difference polynomial and Laurent difference polynomial
An ordinary difference field, or simply a σ-field, is a field F with a third unitary operation
σ satisfying: for any a, b ∈ F, σ(a + b) = σ(a) + σ(b), σ(ab) = σ(a)σ(b), and σ(a) = 0 if
and only if a = 0. We call σ the transforming operator of F. If a ∈ F, σ(a) is called the
transform of a and is denoted by a(1) . For n ∈ Z>0 , σ n (a) = σ n−1 (σ(a)) is called the n-th
transform of a and denoted by a(n) , with the usual assumption a(0) = a. If σ −1 (a) is defined
for each a ∈ F, F is called inversive. Every σ-field has an inversive closure [4]. A typical
example of inversive σ-field is Q(λ) with σ(f (λ)) = f (λ + 1).
In this paper, F is assumed to be inversive and of characteristic zero. Furthermore, we
use σ- as the abbreviation for difference or transformally.
We
P introduce the following useful notation. Let x be an algebraic indeterminate and
p = si=0 ci xi ∈ Z[x]. For a in any σ-over field of F, denote
s
Y
(σ i a)ci .
a =
p
i=0
For instance, a
x2 −1
= a(2) /a. It is easy to check that for p, q ∈ Z[x], we have
ap+q = ap aq , apq = (ap )q , (ab)p = ap bp .
By a[n] we mean the set {a, a(1) , . . . , a(n) }. If S is a set of elements, we denote S [n] = ∪a∈S a[n] .
Let S be a subset of a σ-field G which contains F. We will denote respectively by F[S],
F(S), F{S}, and FhSi the smallest subring, the smallest subfield, the smallest σ-subring,
and the smallest σ-subfield of G containing F and S. If we denote Θ(S) = {σ k a|k ≥ 0, a ∈ S},
then we have F{S} = F[Θ(S)] and FhSi = F(Θ(S)).
Now suppose Y = {y1 , . . . , yn } is a set of σ-indeterminates over F. The elements of
(k)
F{Y} = F[yj : j = 1, . . . , n; k ∈ N] are called σ-polynomials over F in Y, and F{Y} itself
is called the σ-polynomial ring over F in Y. A σ-polynomial ideal, or simply a σ-ideal, I
in F{Y} is an ordinary algebraic ideal which is closed under transforming, i.e. σ(I) ⊂ I.
If I also has the property that a(1) ∈ I implies that a ∈ I, it is called a reflexive σ-ideal.
A prime σ-ideal is a σ-ideal which is prime as an ordinary algebraic polynomial ideal. For
convenience, a prime σ-ideal is assumed not to be the unit ideal in this paper. A σ-ideal I
is called well-mixed if f g ∈ I implies f gx ∈ I for f, g ∈ F{Y}. A σ-ideal I is called perfect
if for any a ∈ N[x] \ {0} and p ∈ F{Y}, pa ∈ I implies p ∈ I. If S is a subset of F{Y}, we
use (S), [S], hSi, and {S} to denote the algebraic ideal, the σ-ideal, the well-mixed σ-ideal,
and the perfect σ-ideal generated by S.
An n-tuple over F is an n-tuple of the form η = (η1 , . . . , ηn ) where the ηi are selected
from a σ-overfield of F. For a σ-polynomial f ∈ F{Y}, η is called a σ-zero of f if when
(j)
(j)
substituting yi by ηi in f , the result is 0.
3
Q
For f = (f1 , . . . , fn )τ ∈ Z[x]n , we define Yf = ni=1 yifi . Yf is called a Laurent σmonomial in Y and f is called its support. A nonzero vector f = (f1 , . . . , fn )τ ∈ Z[x]n is said
to be normal if the leading coefficient of fs is positive, where s is the largest subscript such
that fs 6= 0.
A Laurent σ-polynomial over F in Y is an F-linear combination of Laurent σ-monomials in
Y. Clearly, the set of all Laurent σ-polynomials form a commutative σ-ring under the obvious
sum, product, and the usual transforming operator σ, where all Laurent σ-monomials are
invertible. We denote the σ-ring of Laurent σ-polynomials with coefficients in F by F{Y± }.
Let p be a Laurent σ-polynomial in F{Y± }. An n-tuple (a1 , . . . , an ) over F with each ai 6= 0
is called a nonzero σ-solution of p if p(a1 , . . . , an ) = 0.
2.2
Characteristic set for a difference polynomial system
Let f be a σ-polynomial in F{Y}. The order of f w.r.t. yi is defined to be the greatest
(k)
number k such that yi appears effectively in f , denoted by ord(f, yi ). If yi does not appear
in f , then we set ord(f, yi ) = −∞. The order of f is defined to be maxi ord(f, yi ), that is,
ord(f ) = maxi ord(f, yi ).
The elimination ranking R on Θ(Y) = {σ k yi |1 ≤ i ≤ n, k ≥ 0} is used in this paper:
l
i > σ yj if and only if i > j or i = j and k > l, which is a total order over Θ(Y). By
convention, 1 < θyj for all θyj ∈ Θ(Y).
σk y
(k)
Let f be a σ-polynomial in F{Y}. The greatest yj w.r.t. R which appears effectively
in f is called the leader of f , denoted by ld(f ) and correspondingly yj is called the leading
variable of f , denoted by lvar(f ) = yj . The leading coefficient of f as a univariate polynomial
in ld(f ) is called the initial of f and is denoted by If .
Let p and q be two σ-polynomials in F{Y}. q is said to be of higher rank than p if
(k)
(k)
(k)
ld(q) > ld(p) or ld(q) = ld(p) = yj and deg(q, yj ) > deg(p, yj ).
(k+l)
(k)
Suppose ld(p) = yj . q is said to be reduced w.r.t. p if deg(q, yj
all l ∈ N.
(k)
) < deg(p, yj ) for
A finite sequence of nonzero σ-polynomials A = A1 , . . . , Am is said to be a difference
ascending chain, or simply a σ-chain, if m = 1 and A1 6= 0 or m > 1, Aj > Ai and Aj is
reduced w.r.t. Ai for 1 ≤ i < j ≤ m.
A σ-chain A can be written as the following form
A : A11 , . . . , A1k1 , . . . , Ap1 , . . . , Apkp
(1)
where lvar(Aij ) = yci for j = 1, . . . , ki and ord(Aij , yci ) < ord(Ail , yci ) for j < l. The
following are two σ-chains
A1 = y1x − 1, y12 y22 − 1, y2x − 1
A2 = y12 − 1, y1x − y1 , y22 − 1, y2x + y2
(2)
Let A = A1 , A2 , . . . , At be a σ-chain with Ii as the initial of Ai , and f any σ-polynomial.
Then there exists an algorithm, which reduces f w.r.t. A to a polynomial r that is reduced
4
w.r.t. A and satisfies the relation
t
Y
i=1
Iei i · f ≡ r, mod [A],
(3)
where the ei ∈ N[x] and r = prem(f, A) is called the σ-remainder of f w.r.t. A [9].
A σ-chain C contained in a σ-polynomial set S is said to be a characteristic set of S,
if S does not contain any nonzero element reduced w.r.t. C. Any σ-polynomial set has a
characteristic set. A characteristic set C of a σ-ideal J reduces to zero all elements of J .
(o )
Let A : A1 , . . . , At be a σ-chain, Ii = I(Ai ), yli i = ld(Ai ). A is called regular if for any
j
j
j ∈ N, Iix is invertible w.r.t A [9] in the sense that [A1 , . . . , Ai−1 , Iix ] contains a nonzero
(o +k)
σ-polynomial involving no yli i , k = 0, 1, . . .. To introduce the concept of coherent σchain, we need to define the ∆-polynomial first. If Ai and Aj have distinct leading variables,
we define ∆(Ai , Aj ) = 0. If Ai and Aj (i < j) have the same leading variable yl , then
oi = ord(Ai , yl ) < oj = ord(Aj , yl ). Define
∆(Ai , Aj ) = prem((Ai )x
oj −oi
, Aj ).
(4)
Then A is called coherent if prem(∆(Ai , Aj ), A) = 0 for all i < j [9].
Let A be a σ-chain. Denote IA to be the minimal multiplicative set containing the initials
of elements of A and their transforms. The saturation ideal of A is defined to be
sat(A) = [A] : IA = {p ∈ F{Y} : ∃h ∈ IA , s.t. hp ∈ [A]}.
The following result is needed in this paper.
Theorem 2.1 [9, Theorem 3.3] A σ-chain A is a characteristic set of sat(A) if and only if
A is regular and coherent.
3
Z[x]-lattice
In this section, we prove basic properties of Z[x]-lattices, which will play the role of lattices
in the study of binomial ideals.
For brevity, a Z[x]-module in Z[x]n is called a Z[x]-lattice. Since Z[x] is a Noetherian
ring, any Z[x]-lattice L has a finite set of generators f = {f1 , . . . , fs } ⊂ Z[x]n :
L = SpanZ[x] {f1 , . . . , fs } , (f1 , . . . , fs ).
A matrix representation of f or L is
M = [f1 , . . . , fs ]n×s ,
with fi to be the i-th column of M . We also denote L = (M ). The rank of a Z[x]-lattice L
is defined to be the rank of any matrix representation of L, which is clearly well defined.
5
We list some basic concepts and properties of Gröbner bases of modules. For details,
please refer to [5].
Denote ǫi to be the i-th standard basis vector (0, . . . , 0, 1, 0, . . . , 0)τ ∈ Z[x]n , where 1
lies in the i-th row of ǫi . A monomial m in Z[x]n is an element of the form axk ǫi ∈ Z[x]n ,
where a ∈ Z and k ∈ N. The following monomial order > of Z[x]n will be used in this paper:
axα ǫi > bxβ ǫj if i > j, or i = j and α > β, or i = j, α = β, and |a| > |b|.
With the above order, P
any f ∈ Z[x]n can be written in a unique way as a linear combination of monomials, f = li=1 hi , where hi 6= 0 and h1 > h2 > · · · > hl . The leading term
of f is defined to be LT(f ) = h1 . For any G ⊂ Z[x]n , we denote by LT(G) the set of leading
terms of G.
The order > can be extended to elements of Z[x]n as follows: for f , g ∈ Z[x]n , f < g if
and only if LT(f ) < LT(g).
Let G ⊂ Z[x]n and f ∈ Z[x]n . We say that f is G-reduced with respect to G if any
monomial of f is not a multiple of LT(g) by an element in Z[x] for any g ∈ G.
Definition 3.1 A finite set f = {f1 , . . . , fs } ⊂ Z[x]n is called a Gröbner basis for the Z[x]lattice L generated by f if for any g ∈ L, there exists an i, such that LT(g)|LT(fi ). A
Gröbner basis f is called reduced if for any f ∈ f , f is G-reduced with respect to f \ {f }. In
this paper, it is always assumed that f1 < f2 < · · · < fs .
Let f be a Gröbner basis. Then any f ∈ Z[x]n can be reduced to a unique normal form
by f , denoted by grem(f , f ), which is G-reduced with respect to f .
Definition 3.2 Let f , g ∈ Z[x]n , LT(f ) = axk ei , LT(g) = bxs ej , s ≤ k. Then the Spolynomial of f and g is defined as follows: if i 6= j then S(f , g) = 0; otherwise S(f , g) =
a
if b | a;
f − b xk−s g,
b
k−s g,
(5)
f
−
x
if a | b;
a
uf + vxk−s g, if a ∤ b and b ∤ a, where gcd(a, b) = ua + vb.
The following basic property for Gröbner basis is obviously true for Z[x]-lattices and a
polynomial-time algorithm to compute Göbner bases for Z[x]-lattices is given in [15].
Theorem 3.3 (Buchberger’s Criterion) The following statements are equivalent.
1) f = {f1 , . . . , fs } ⊂ Z[x]n is a Gröbner basis.
2) grem(S(fi , fj ), G) = 0 for all i, j.
3) f ∈ (f ) if and only if grem(f , f ) = 0.
We will study the structure of a Gröbner basis for a Z[x]-lattice by introducing the
concept of generalized Hermite normal form. First, we consider the case of n = 1.
Lemma 3.4 Let B = {b1 , . . . , bk } be a reduced Gröbner basis of a Z[x]-module in Z[x],
b1 < · · · < bk , and LT(bi ) = ci xdi ∈ N[x]. Then
6
1) 0 ≤ d1 < d2 < · · · < dk .
2) ck | · · · |c2 |c1 and ci 6= ci+1 for 1 ≤ i ≤ k − 1.
3) ccki |bi for 1 ≤ i < k. If e
b1 is the primitive part of b1 , then e
b1 |bi for 1 < i ≤ k.
4) The S-polynomial S(bi , bj ) can be reduced to zero by B for any i, j.
Proof: 1) and 4) are consequences of Theorem 3.3. To prove 2), assume that there exists an
l such that cl−1 | · · · |c2 |c1 but cl 6 |cl−1 . Let r = gcd(cl , cl−1 ) = p1 cl + p2 cl−1 , where p1 , p2 ∈ Z.
Then |r| < |cl−1 | and |r| < |cl |. Since cl−1 | · · · |c2 |c1 , we have |r| < |ci |, i = 1, . . . , l. Let
g = p1 bl + p2 xdl −dl−1 bl−1 . Then LT(g) = rxdl which is reduced w.r.t. B and g ∈ (B),
contradicting to the definition of Gröbner bases.
We prove 3) by induction on k. When k = 2, let b1 = c1 xd1 + s11 xd1 −1 + · · · + s1d1
and b2 = c2 xd2 + s21 xd2 −1 + · · · + s2d2 . Then, c2 |c1 and d1 < d2 . Let c1 = c2 t, we need
to show t|b1 . Since the S-polynomial S(b1 , b2 ) = tb2 − xd2 −d1 b1 can be reduced to zero by
b1 , we have tb2 − xd2 −d1 b1 = u(x)b1 , where u(x) ∈ Z[x] and deg(u(x)) < d2 − d1 . Then,
tb2 = (xd2 −d1 + u(x))b1 , and t|b1 follows since xd2 −d1 + u(x) is a primitive polynomial in
Z[x]. The claim is true. Assume that for k = l − 1, the claim is true, then eb1 |bi for
1 ≤ i ≤ l − 1. We will prove the claim for k = l. Since S(b1 , bl ) = cc1l bl − xdl −d1 b1
Pl−1
can be reduced to zero by B. We have cc1l bl − xdl −d1 b1 =
i=1 fi bi with fi ∈ Z[x] and
Pl−1
c1
d
−d
1
l
deg(fi bi ) ≤ dl − 1. Then, bl = x
b1 +
fi bi . By induction, e
b1 is a factor of the
i=1
cl
right hand side ofPthe above equation. Thus e
b1 |bl . Let bi = si b′1 for 1 ≤ i ≤ l, we have
l−1
c1
d
−d
1s +
l
1
i=1 fi si where deg(si ) = di − d1 and s1 ∈ Z. Since deg(fi si ) ≤ dl − d1 − 1,
cl sl = x
c
c1
c1
we have cl |s1 and cl |b1 . For any 1 ≤ i < j < l, assume ccil |bi . We will show that cjl |bj . Since
P
cj−1
c
j−1
dj −dj−1 b
S(bj−1 , bj ) = j−1
j−1 =
i=1 fi bi , we have cl is a factor of the right hand
c j bj − x
c
cj−1
cj
side of the above equation, for cj−1 |cj−2 | · · · |c1 . Then, j−1
cl | cj bj and cl |bj . The claim is
proved.
Example 3.5 Here are three Gröbner bases in Z[x]: {2, x}, {12, 6x + 6, 3x2 + 3x, x3 + x2 },
{9x + 3, 3x2 + 4x + 1}.
To give the structure of a reduced Gröbner basis similar to that in Example 3.5, we
introduce the concept of generalized Hermite normal form. Let
c
1,1
...
cr1 ,1
0
...
C =
0
...
0
...
0
...
...
...
...
...
...
...
...
...
...
c1,l1
...
cr1 ,l1
0
...
0
...
0
...
0
c1,l1 +1
...
cr1 ,l1 +1
cr1 +1,1
...
cr2 ,1
...
0
...
0
...
...
...
...
...
...
...
...
...
...
...
...
...
cr1 +1,l2
...
cr2 ,l2
...
0
...
0
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
0
...
0
...
...
...
...
...
...
...
crt−1 +1,1
...
crt ,1
whose elements are in Z[x]. It is clear that m = rt and s =
be the column of the matrix C whose last nonzero element is
Pt
...
...
...
...
...
...
...
...
...
...
i=1 li .
...
...
...
...
...
...
...
cr
t−1 +1,lt
...
crt ,lt
(6)
m×s
We denote by cri ,j to
cri ,j = ci,j,0 xdij + · · · + ci,j,dij .
Then the leading monomial of cri ,j is cri ,j,0 xdri ,j ǫri . It is clear that rk(L) = t.
7
(7)
Definition 3.6 The matrix C in (6) is called a generalized Hermite normal form if it satisfies
the following conditions:
1) 0 ≤ dri ,1 < dri ,2 < · · · < dri ,li for any i.
2) cri ,li ,0 | · · · |cri ,2,0 |cri ,1,0 .
c
,0
3) S(cri ,j1 , cri ,j2 ) = xdri ,j2 −dri ,j1 cri ,j1 − crri ,j,j1 ,0
cri ,j2 can be reduced to zero by the column
i 2
vectors of the matrix for any 1 ≤ i ≤ t, 1 ≤ j1 < j2 ≤ li .
4) cri ,j is G-reduced w.r.t. the column vectors of the matrix other than cri ,j , for any 1 ≤
i ≤ t, 1 ≤ j ≤ li .
It is clear that {cri ,1 , . . . , cri ,li } is a reduced Gröbner basis in Z[x]. Then, as a consequence
of Theorem 3.3 and Lemma 3.4, we have
Theorem 3.7 f = {f1 , . . . , fs } ⊂ Z[x]n is a reduced Gröbner basis such that f1 < f2 < · · · <
fs if and only if [f1 , . . . , fs ] is a generalized Hermite normal form.
Example 3.8 The following matrices are generalized Hermite normal forms
x 2 0
2 x−1 0 0
M1 =
, M2 =
0 2 x
0 0
2 x−1
whose columns constitute the reduced Gröbner bases of the Z[x]-lattices.
basis. Let S(fi , fj ) = mij fi − mji fj be the
Let f = {f1 , . . . , fs } be a reduced Gröbner P
S-polynomial of fi , fj and grem(S(fi , fj ), f ) =
k ck fk be the normal representation of in
e i , fj )
terms of the Gröbner basis f . Then the syzygy polynomial S(f
X
e i , fj ) = mij ǫi − mji ǫj −
S(f
ck ǫk ,
k
is an element in Z[x]s , where ǫk is the k-th standard basis vector of Z[x]s . Define an order
in Z[x]s as follows: axα ǫi ≺ bxβ ǫj if LT(axα fi ) > LT(bxβ fj ) in Z[x]n . By Schreyer’s
Theorem [5, p. 212], we have
Theorem 3.9 Let F = [f1 , . . . , fs ]n×s ∈ Z[x]n×s be a generalized Hermite normal form.
e i , fj ) form a Gröbner basis of the Z[x]-lattice ker(F ) =
Then the syzygy polynomials S(f
{X ∈ Z[x]s |F X = 0} under the newly defined order ≺.
Let C be defined in (6) and k ∈ N. Introduce the following notations:
i −1
{cri ,k , xcri ,k , . . . , xdeg(cri ,k+1 )−deg(crt ,k )−1 cri ,k },
C− = ∪ti=1 ∪lk=1
k
C + = ∪ti=1 ∪∞
k=0 {x cri ,li }.
C∞ = C− ∪ C
+
8
(8)
6 3x 0 3 2x
Example 3.10 Let C =
0 0 6 3x x3 + x
6 0 3 3x
Then C− =
and
0 6 3x 3x2
C∞ =
6 3x 3x2 3x3 · · ·
0 0 0
0
···
.
0 3 3x 2x
2x2
···
2
3
6 3x 3x x + x x4 + x2 · · ·
.
We need the following properties about C∞ . By saying the infinite set C∞ is linear
independent over Z, we mean any finite subset of C∞ is linear independent over Z. Otherwise,
C∞ is said to be linear dependent.
Lemma 3.11 The columns of C∞ in (8) are linear independent over Z.
Proof: Suppose C is given in (6). The leading term of c ∈ C∞ is LT(c) = axl ǫri for i = 1, . . . , t
and l ∈ N. Furthermore, for two different c1 and c2 in CS such that LT(c1 ) = axl1 ǫri
and LT(c2 ) = bxl2 ǫri , we have l1 6= l2 . Then LT(C∞ ) = {aili xli ǫri | i = 1, . . . , t; li =
di1 , di1 + 1, . . . ; aili ∈ Z} are linear independent over Z, where di1 is from (7). Then C∞ are
also linear independent over Z.
Lemma 3.12 Let C be a generalized Hermite normal form. Then any g ∈ (C) can be written
uniquely as a linear combination of finitely many elements of C∞ over Z.
Proof: g ∈ (C) can be written as a linear combination of elements of C∞ over Z by the
procedure to compute grem(g, C) = 0 [5]. The uniqueness is a consequence of Lemma 3.11.
4
Canonical Representations for Laurent binomial σ-ideal
In this section, we will give three canonical representations for a proper Laurent binomial
ideal.
4.1
Laurent binomial σ-ideal
In this section, several basic properties of Laurent binomial σ-ideals will be proved.
By a Laurent σ-binomial in Y, we mean a σ-polynomial with two terms, that is, aYg +bYh
where a, b ∈ F ∗ = F \ {0} and g, h ∈ Z[x]n . A Laurent σ-binomial of the following form is
said to be in normal form
p = Yf − cf
where cf ∈ F ∗ = F \ {0} and f ∈ Z[x]n is normal. The vector f is called the support of p.
−f
−f − c−1 which is called the inverse of p. It
For p = Yf − cf , we denote pb = −c−1
f
f Y p = Y
is clear that any Laurent σ-binomial f can be written uniquely as f = aM (Yf − cf ) where
a ∈ F ∗ , M is a Laurent σ-monomial, and Yf − cf is in normal form. Since aM is a unit
9
in F{Y± }, we can use the normal σ-binomial Yf − cf to represent f , and when we say a
Laurent σ-binomial we always use its normal representation.
A Laurent σ-ideal is called binomial if it is generated by Laurent σ-binomials.
Lemma 4.1 Let Yfi − ci , i = 1, . . . , s be contained
Q in a Laurent binomial σ-ideal I and
f = a1 f1 + · · · + as fs , where ai ∈ Z[x]. Then Yf − si=1 cai i is in I.
Proof: It suffices to show that if p1 = Yf1 − c1 ∈ I and p2 = Yf2 − c2 ∈ I, then Ynf1 − cn1 ∈ I
xf1 − σ(c ) ∈ I, and Yf1 +f2 − c c ∈ I, which are indeed true
for n ∈ N, Y−f1 − c−1
1
1 2
1 ∈ I, Y
−1 −f1
f
n
n
nf
n
1
1
(Yf1 − c1 ) ∈ I,
since Y − c1 = (Y ) − c1 contains p1 as a factor, Y−f1 − c−1
1 = −c1 Y
Yxf1 − σ(c1 ) = σ(Yf1 − c1 ) ∈ I, and Yf1 +f2 − c1 c2 = Yf1 (Yf2 − c2 ) + c2 (Yf1 − c1 ) ∈ I.
As a direct consequence, we have
Proposition 4.2 Let I be a proper Laurent binomial σ-ideal and
L(I) := {f ∈ Z[x]n | ∃cf ∈ F ∗ s.t. Yf − cf ∈ I}.
(9)
Then L(I) is a Z[x]-lattice, which is called the support lattice of I. Furthermore, Let
L(I) = (f1 , . . . , fs ). Then I = [Yf1 − cf1 , . . . , Yfs − cfs ]. That is, a Laurent binomial σ-ideal
is finitely generated and [f1 , . . . , fs ] is called a matrix representation for I
Proof: Let I1 = [Yf1 − cf1 , . . . , Yfs − cfs ]. It suffices to show I ⊂ I1 . Since I is Laurent
binomial, it has a set of generators of the form fh = Yh − ch . Then h ∈ L(I) = (f1 , . . . , fs ).
By Lemma 4.1, there exists a e
ch ∈ F such that feh = Yh −e
ch ∈ I1 . Then fh −feh = e
ch −ch ∈ I.
e
Since I is proper, we have fh − fh = 0 or fh ∈ I1 and hence I ⊂ I1 .
Similarly, we can prove
Corollary 4.3 Let I = [Yf1 − c1 , . . . , Yfs − cs ] be a proper Laurent
Ps binomial σ-ideal and
let h1 , . . . , hr be another set of generators
of
(f
,
.
.
.
,
f
),
and
h
=
1
s Q
i
k=1 ai,k fk , i = 1, . . . , r,
Qs
a1,i
ar,i
s
h
h
r
1
where ai,k ∈ Z[x]. Then I = [Y − i=1 ci , . . . , Y − i=1 ci ].
We now show to check whether a Laurent binomial σ-ideal is proper.
Proposition 4.4 Let I = [Yf1 − c1 , . . . , Yfs − cs ] be a Laurent binomial σ-ideal and M =
[f1 , . . . , fs ] ∈ Z[x]n×s . Let ker(M ) = {h ∈ Z[x]s |Q
M h = 0} be generated by u1 , . . . , ut , where
u
ui = (ui,1 , . . . , uis ). Then I =
6 [1] if and only if si=1 ci l,i = 1 for l = 1, . . . , t.
Q
u
Proof: “⇒” Let fi = Yfi − ci . Suppose c = si=1 ci l,i 6= 1 for some l. Replacing
ci by
Qs
ul,i
fi − f in the above equation and noting that u ∈ ker(M ), we have c =
Y
c
=
i
l
i=1 i
Qs
Qs
M
·u
f
u
i
l
l,i
+ g = 1 + g where g ∈ I. Then 0 6= c − 1 ∈ I and I = [1], a
= i=1 Y
i=1 (y − fi )
contradiction.
“⇐” Suppose the contrary. Then there exist gi ∈ F{Y± } such that
g1 f1 + · · · + gs fs = 1.
10
(10)
(k)
(j)
Let l be the maximal c such that yc occurs in some fi , o the largest j such that yl occurs
o
(o)
in some fk , and d = maxsk=1 deg(fk , yl ). Let fk = Yfk − ck = Ik yldx − ck . Since (10) is an
o
j
identity about the algebraic variables yix , we can set yldx = ck /Ik in (10) to obtain a new
identity. In the new identity, fk becomes zero and the left hand side of (10) has at most s − 1
summands. We will show that this procedure can be continued for the new identity. Then
the left hand side of (10) will eventually becomes zero, and a contradiction is obtained and
the lemma is proved.
o
If ord(fi , yl ) < o or ord(fi , yl ) = o and deg(fi , ylx ) < d for some i, then fi is not
o
changed in the above procedure. Let us assume that for some v, deg(fv , ylx ) = d and
o
fv = Yfv − cv = Iv yldx − cv . Then after the substitution, fv = ck Iv /Ik − cv = ck fev where
fev = Iv /Ik − cv /ck . We claim that either fev = 0 or Iv /Ik is a proper monomial, and as a
consequence, the above substitution can continue. To prove the claim, it suffices to show
that if Iv = Ik then cv = ck . If Iv = Ik , then fv = fk , that is fv − fk = 0 is a syzygy among
fi and let ǫvk be the corresponding syzygy vector. Then ǫvk ∈ ker(M ) can be written as a
ǫvk can be written as
linear combination Q
of u1 , . . . , us . Let c = (c1 , . . . , cs )τ . Then cv c−1
k =c
u
s
l,i
−1
a product of cul = i=1 ci = 1, and thus cv ck = 1.
4.2
Characteristic set of Laurent binomial σ-ideal
We show how to modify the characteristic set method presented in section 2.2 in the case of
Laurent binomial σ-ideals. First, assume that all Laurent σ-binomials are in normal form,
which makes the concepts of order and leading variables unique.
Second, when defining the concepts of rank and q to be reduced w.r.t. p, we need to re(o)
(o)
(k+l)
place deg(p, yj ) by |deg(p, yj )|. Precisely, q is said to be reduced w.r.t. p if |deg(q, yj
)| <
(k)
(k)
|deg(p, yj )| for all l ∈ N, where ld(p) = yj . For instance, y1−2x y2 − 1 is not reduced w.r.t.
y12 − 1. With these changes, the concepts of σ-chain and characteristic set can be defined
in the Laurent σ-binomial case. For instance, the σ-chain A2 in (2) becomes the following
Laurent normal form:
Ae2 = y12 − 1, y1−1 y1x − 1, y22 − 1, y2−1 y2x − 1
(11)
Third, the σ-remainder for two Laurent σ-binomials need to be modified as follows. We
first consider how to compute prem(f, g) in the simple case: o = ord(f, yl ) = ord(g, yl ),
(o)
(o)
where yl = lvar(g). Let g = Ig (yl )d − cg , where d = deg(g, yl ) and Ig is the initial of
(o)
g. As mentioned above, g is in normal form, that is d > 0. Let df = deg(f, yl ) and
(o)
f = If (yl )df − cf . We consider two cases.
In the first case, let us assume df ≥ 0. If df < dg , then set r = prem1 (f, g) to be f .
Otherwise, perform the following basic step
r := prem1 (f, g) = (f − g
If (o)
cf
If (o) df −dg
)/cg = (yl )df −dg − .
(y )
Ig l
Ig
cg
(12)
Let hr , hf , fg be the supports of r, f, g, respectively. Then
hr = hf − hg .
11
(13)
Set f = r and repeat the procedure prem1 for f and g. Since df decreases strictly after each
iteration, the procedure will end and return prem(f, g) = r which satisfies
r =
If (o)
cf
f
− hg = k (yl )df −kdg − k
ckg
Ig
cg
hr = hf − khg
(14)
(15)
d
where k = ⌊ dfg ⌋ and h ∈ F{Y± }.
In the second case, we assume df < 0. The σ-remainder can be computed similar to
(o)
the first case. Instead of g, we consider b
g = (Ig )−1 (yl )−dg − c−1
g . If |df | < dg , then set
r = prem1 (f, g) to be f . Otherwise, perform the following basic step
(o)
(o)
r := prem1 (f, g) = cg (f − gbIg If (yl )df +dg ) = If Ig (ylg )df +dg − cf cg .
In this case, equation (13) becomes hr = hf + hg . To compute prem(f, g), repeat the above
basic step for f = r until |df | < dg .
For two general σ-binomials f and g, prem(f, g) is defined as follows: if f is reduced
w.r.t g, set prem(f, g) = f . Otherwise, let yl = lvar(g), of = ord(f, yl ), and og = ord(g, yl ).
Define
prem(f, g) = prem(. . . , prem(prem(f, g (of −og ) ), g (of −og −1) ), . . . , g).
Let A : A1 , . . . , As be a Laurent binomial σ-chain and f a σ-binomial. Then define
prem(f, A) = prem(. . . , prem(prem(f, As ), As−1 ), . . . , A1 ).
In summary, we have
Lemma 4.5 Let A = A1 , . . . , As be a Laurent binomial σ-chain, f a σ-binomial, and r =
prem(f, A). Then r is reduced w.r.t. A and satisfies
cf ≡ r, mod [A],
(16)
where c ∈ F ∗ . Furthermore, let the supports of r and f be hr and hf , respectively. Then
hf − hr is in the Z[x]-lattice generated by the supports of Ai .
Similar to section 2.2, the concepts of coherent and regular σ-chains can be extended to
the Laurent case. Since any σ-monomial is a unit in F{Y± }, the concept of regular σ-chain
need to be strengthened as follows. A σ-chain A is called Laurent regular if A is regular and
any σ-monomial is invertible w.r.t A. Then, following [9], Theorem 2.1 can be extended to
the following Laurent version straightforwardly.
Theorem 4.6 A Laurent σ-chain A is a characteristic set of sat(A) if and only if A is
coherent and Laurent regular.
For Laurent binomial σ-chains, we have
Lemma 4.7 Any Laurent binomial σ-chain A is Laurent regular.
12
Proof: Since the initials of A are σ-monomials, it suffices to show that any σ-monomial is
invertible w.r.t A. By [9, p.248], a σ-monomial M is invertible w.r.t A is M is invertible
w.r.t an extension AM of A when both M and AM are treated as algebraic polynomials in
j
yix , where AM is an algebraic Laurent binomial chain. By [10, p.1150], M is invertible w.r.t
AM if the successive Sylvester resultant Resl(M, AM ) of M and AM is nonzero. Since A
o
is Laurent binomial, B ∈ AM has the form B = I(ykx )m + U , where I is the initial of B
o
o
and U a σ-monomial does not contain ykx . Let N = J(ykx )n be any σ-monomial with J as
o
the initial. Then the Sylvester resultant of M and B w.r.t ykx is J m U n which is a nonzero
σ-monomial. As a consequence, Resl(M, AM ) is also a nonzero σ-monomial and hence A is
Laurent regular.
We now give the first canonical representation for Laurent binomial σ-ideals.
Theorem 4.8 I is a proper Laurent binomial σ-ideal if and only there exists a Laurent
coherent σ-chain A such that I = sat(A) = [A].
Proof: Let I 6= [1] and A the characteristic set of I. Then [A] ⊂ I ⊂ sat(A). From (16),
we have sat(A) ⊂ [A] and then I = sat(A) = [A]. By Theorem 4.6, A is coherent. To prove
the other side of the theorem, let A be a Laurent coherent σ-chain. By Lemma 4.7, A is
also Laurent regular. By Theorem 4.6, A is a characteristic set of I = sat(A). Then I is
proper.
Corollary 4.9 Let I be a Laurent reflexive prime binomial σ-ideal in F{Y± }. Then dim(I) =
n − rk(L(I)).
Proof: By Theorem 4.8, I = [A], where A : Yc1 − c1 , . . . Ycs − cs . Let C = [c1 , . . . , cs ] is
the matrix representation for I and in the form of (6). Since I is reflexive and prime, by
Theorem 4.3 of [9], dim(I) = n − t = n − rk(L(I)).
Corollary 4.10 A Laurent binomial σ-ideal is radical.
Proof: By Theorem 4.8, I = [A], where A : Yh1 − c1 , . . . , Yhr − cr is the characteristic set
(o )
of I. Let Ai = Yhi − ci and yli i = ld(Ai ). A is also saturated in the sense that its separant
∂Ai
±
(oi ) are σ-monomials and hence units in F{Y }. Then similar to the differential case [2],
∂yl
i
it can be shown that sat(A) = [A] is a radical σ-ideal.
Let f1 < f2 < · · · < fs be elements in Z[x]n , ci ∈ F ∗ , and
f
= {f1 , . . . , fs } ⊂ Z[x]n
(17)
±
fi
Af = {A1 , . . . , Aa } ⊂ F{Y } with Ai = Y − ci , i = 1, . . . , s
In the rest of this section, we will establish a connection between f and Af . From Definition
3.6, we have
Lemma 4.11 For i < j, Aj is reduced w.r.t. Ai if and only if fj is G-reduced w.r.t. fi .
13
Lemma 4.12 For f and Af in (17) and a Laurent σ-binomial f = Yf − c, if prem(f, Af ) =
Yg − cg , then g = grem(f , f ).
(o )
Proof: Let us first consider prem1 in (12) for f and Ai = Yfi − ci = Ii (yli i )di − ci , where
ld(Ai ) = yli and Ii is the initial of Ai . From (13), the support of r = prem1 (f, Ai ) is f − fi . It
is clear that LT(fi ) = di xoi ǫli . Let fi = di xoi ǫli + f i . Similarly, write f = df xoi ǫli + m where
df xoi ǫli is the leading term of f w.r.t. ǫli and df ≥ di ≥ 0. Then a basic step to compute
grem(f , fi ) is to compute grem1 (f , fi ) = f − fi = (df − di )xoi ǫli + f − f i , which is the support
of prem1 (f, Ai ).
Using the basic step grem1 to compute grem(f , f ), we have a sequence of elements in
Z[x]n : g0 = f , g1 , . . . , gt = grem(f , f ). Correspondingly, using the basic step prem1 to
compute prem(f , Af ), we have a sequence of σ-binomials f0 = f, f1 , . . . , ft = prem(f, Af )
such that the support of fi is gi for i = 1, . . . , t.
Lemma 4.13 If f in (17) is a reduced Gröbner basis and [Af ] 6= [1], then Af is a coherent
σ-chain.
Proof: By Lemma 4.11, Af is a σ-chain. Let Ai = Yfi − ci and Aj = Yfj − cj (i < j) have
d xo j
oi
the same leading variable yl , and Ai = Ii yldi x − ci , Aj = Ij yl j − cj . From Definition 3.6,
we have oi < oj and dj |di . Let di = tdj where t ∈ N. According to (14), we have
xoj −oi
∆(Ai , Aj ) = prem((Ai )
oj −oi
(Ii )x
, Aj ) =
Ijt
oj −oi
(ci )x
−
ctj
.
(18)
di
dj fj .
dj xoj ǫl ,
Then the support of ∆(Ai , Aj ) is xoj −oi fi −
Since LT(Ai ) = di xoi ǫl and LT(Aj ) =
we have N = lcm(di xoi , dj xoj ) = di xoj .
According to Definition 3.2, the S-vector of fi and fj is
S(fi , fj ) = xoj −oi fi −
di
fj .
dj
Since f is a Gröbner basis, we have g = grem(S(fi , fj ), f ) = 0. Since the support of ∆(Ai , Aj )
is S(fi , fj ), by Lemma 4.12, R = prem(∆(Ai , Aj ), Af ) = Yg − c = 1 − c for some c ∈ F. Since
[Af ] is proper and R = 1 − c ∈ [Af ], we have R = 0 and hence Af is coherent.
We now give the main result of this section.
Theorem 4.14 For f and Af defined in (17), Af is a coherent σ-chain if and only if f is a
reduced Gröbner basis and [Af ] 6= [1].
Proof: Lemma 4.13 proves one side of the theorem. For the other direction, let Af be a coherent σ-chain. By Lemma 4.11, fi is G-reduced to fj for i 6= j. By Theorem 4.8, [Af ] is proper.
Use the notations introduced in the proof of Lemma 4.13. Since S(fi , fj ) is the support of
∆(Ai , Aj ), by Lemma 4.12, fij = grem(S(fi , fj ), f ) is the support of prem(∆(Ai , Aj ), Af ).
Since Af is coherent, prem(∆(Ai , Aj ), Af ) = Yfij − c = 0 for any i and j, and this is possible
only when fij = grem(S(fi , fj ), f ) = 0 and c = 1 due to the fact [Af ] 6= [1]. Hence f is a
reduced Gröbner basis.
14
4.3
Partial character and Laurent binomial σ-ideal
In this section, we will show that proper Laurent binomial σ-ideals can be described uniquely
with their partial characters.
Definition 4.15 A partial character ρ on Z[x]n is a homomorphism from a Z[x]-lattice Lρ
to the multiplicative group F ∗ satisfying ρ(xf ) = (ρ(f ))x = σ(ρ(f )) for f ∈ Lρ .
Let ρ be a partial character over Z[x]n and Lρ = (f1 , . . . , fs ), where f = {f1 , . . . , fs } is a
reduced Gröbner basis. Define
I(ρ) := [Yf − ρ(f ) | f ∈ Lρ ].
f1
fs
A(ρ) := Y − ρ(f1 ), . . . , y − ρ(fs ).
(19)
(20)
The Laurent binomial σ-ideal I(ρ) has the following properties.
Lemma 4.16 I(ρ) = [A(ρ)] 6= [1] and A(ρ) is a characteristic set of I(ρ).
Proof: By Lemma 4.1 and the property of partial character, I(ρ) = [A(ρ)].
P By Proposition
4.4, in orderQto prove I(ρ) 6= [1], it suffices
to show
P
Q that for any syzygy i ai fi = 0 among
fi , we have i ρ(fi )ai = 1. Indeed, ρ( i ai fi ) = i ρ(fi )ai = 1, since ρ is a homomorphism
from the Z[x]-module Lρ to F ∗ . Since f is a reduced Gröber basis, by Theorem 4.14, A is a
characteristic set of I(ρ).
Lemma 4.17 A Laurent σ-binomial Yf − cf is in I(ρ) if and only if f ∈ Lρ and cf = ρ(f ).
Proof: By Lemma 4.16, A(ρ) is a characteristic set of I(ρ). Since f = y f − cf is a σ-binomial
in I(ρ), we have r = prem(f, A) = 0. By Lemma 4.5, f is in the Z[x]-module Lρ . The other
side is obviously true and the lemma is proved.
We now show that all Laurent binomial σ-ideals are defined by partial characters.
Theorem 4.18 The map ρ ⇒ I(ρ) gives a one to one correspondence between the set of
proper Laurent binomial σ-ideals and partial characters on Z[X]n .
Proof: By Lemma 4.16, a partial character defined a proper Laurent binomial σ-ideal. For
the other side, let I ⊆ F{Y± } be a proper Laurent binomial σ-ideal. I is generated by its
members of the form y f − cf for f ∈ Z[x]n and cf ∈ F ∗ . Let Lρ = L(I) which is defined
in (9) and ρ(f ) = cf . Since I is proper, cf is uniquely determined by f . By Lemma 4.1
and Proposition 4.2, ρ is a partial character which is uniquely determined by I. It is clear
I(ρ) = I. To show the correspondence is one to one, it suffices to show ρ(I(ρ)) = ρ which
is a consequence of Lemma 4.17. The theorem is proved.
As a summary of this section, we have the following canonical representations for a proper
Laurent binomial σ-ideal, which follows directly from Theorems 4.8, 4.14, 4.18.
Theorem 4.19 I is a proper Laurent binomial σ-ideal if and only if
15
(1) I = [A], where A is a coherent Laurent binomial σ-chain.
(2) I = [A], where A = Yf1 − c1 , . . . , Yfs − cs , fi ∈ Z[x]n , ci ∈ F ∗ , f = {f1 , . . . , fs } is a
reduced Gröbner basis of a Z[x]-lattice, and [A] 6= [1].
(3) I = I(ρ) = [A], where ρ is a partial character on Z[x]n and A = A(ρ).
Furthermore, A is a characteristic set of I and (f ) is the support lattice of I.
5
Criteria for prime, reflexive, and perfect Laurent binomial
σ-ideals
In this section, we give criteria for a Laurent binomial σ-ideal to be prime, reflexive, wellmixed, and perfect in terms of its support lattice.
5.1
Reflexive and prime Laurent binomial σ-ideals
In this section, we first give criteria for reflexive and prime Laurent binomial σ-ideals and
then give a decomposition theorem for perfect Laurent binomial σ-ideals.
For the σ-indeterminates Y = {y1 , . . . , yn } and t ∈ N, we will treat the elements of Y[t]
as algebraic indeterminates, and F[Y[±t] ] is the Laurent polynomial ring in Y[t] . Let I be a
Laurent binomial σ-ideal in F{Y± }. Then it is easy to check that
It = I ∩ F[Y[±t] ]
is a Laurent binomial ideal in F[Y[±t] ].
Denote Z[x]t to be the set of elements in Z[x] with degree ≤ t. Then Z[x]nt is the Zmodule generated by xi ǫl for i = 0, . . . , t, l = 1, . . . , n. It is clear that Z[x]nt is isomorphic to
Zn(t+1) as Z-modules by mapping xi ǫl to the ((l − 1)(t + 1) + i + 1)-th standard basis vector
in Zn(t+1) . Hence, we treat them as the same in this section. Let L be a Z[x]-lattice and
t ∈ N. Then
Lt = L ∩ Z[x]nt = L ∩ Zn(t+1)
is a Z-module in Zn(t+1) . Similarly, it can be shown that when restricted to Z[x]nt , a partial
character ρ on Z[x]n becomes a partial character ρt on Zn(t+1) .
Lemma 5.1 With the notations introduced above, we have It = I ∩ F[Y[±t] ] = I(ρt ).
Proof: It suffices to show that the support lattice of It is Lρt = Lt . By Lemma 4.17,
Yf − cm ∈ It if and only if f ∈ L ∩ Z[x]nt , or equivalently, maxm∈f deg(m, x) ≤ t, which is
equivalent to f ∈ Lt .
Definition 5.2 Let L be a Z[x]-module in Z[x]n .
• L is called Z-saturated if, for any a ∈ Z and f ∈ Z[x]n , af ∈ L implies f ∈ L.
16
• L is called x-saturated if, for any f ∈ Z[x]n , xf ∈ L implies f ∈ L.
• L is called saturated if it is both Z- and x- saturated.
Theorem 5.3 Let ρ be a partial character over Z[x]n . If F is algebraically closed and
inversive, then
(a) Lρ is Z-saturated if and only if I(ρ) is prime;
(b) Lρ is x-saturated if and only if I(ρ) is reflexive;
(c) Lρ is saturated if and only if I(ρ) is reflexive prime.
Proof: It is clear that (c) comes from (a) and (b). Let I = I(ρ) and L = Lρ .
(a): I is a Laurent prime σ-ideal if and only if It is a Laurent prime ideal for all t. From
Lemma 5.1, the support of It is Lt . Then by [7, Thm 2.1], It is a Laurent prime ideal if and
only if Lt is a Z-saturated Z-module. Furthermore, a Z[x]-lattice L is Z-saturated if and
only if Lt is a Z-saturated Z-module for all t. Thus, (a) is valid.
(b): Suppose I is reflexive. For xf ∈ L, by Lemma 4.17, there is a Yxf − c ∈ I. Since F
is reflexive, c = dx for d ∈ F. Then σ(Yf − d) ∈ I and hence Yf − d ∈ I since I is reflexive.
By Lemma 4.17 again, f ∈ L and L is x-saturated. To prove the other direction, assume L
is x-saturated. For f x ∈ I, we have an expression
fx =
s
X
i=1
fi (Yfi − ci )
(21)
where Yfi −ci ∈ I and fi ∈ F{Y± }. Let d = maxsi=1 deg(Yfi −ci , y1 ) and assume Yf1 = M1 y1d .
(j)
Replace y1d by c1 /M1 in (21). Since (21) is an identity for the variables yi , this replacement
is meaningful and we obtain a new identity. Yf1 − c1 becomes zero after the replacement.
Due to the way to chose d, if another summand, say Yf2 − c2 , is affected by the replacement,
then Yf2 = M2 y1d . After the replacement, Yf2 − c2 becomes c1 (M2 /M1 − c2 /c1 ) which is also
in I by Lemma 4.17. In summary, after the replacement, the right hand side of (21) has
less than s summands and the left hand side of (21) does not changed. Repeat the above
procedure, we will eventually obtain a new indenity
fx =
s̄
X
i=1
f¯i (Yxgi − c̄i )
(22)
where Yxgi − c̄i ∈ I and f¯i ∈ F{Y± }. We may assume that any yi does not appear in f¯i .
Otherwise, by setting yi to be 1, the left hand side of (22) is not changes and a new identity
is obtained. Since F is inversive, c̄i = exi and f¯i = gix for ei ∈ F and gi ∈ F{Y± }. By
Lemma 4.17, Yxgi − exi ∈ I implies xgi ∈ L. Since L is x-saturated,
Ps̄ xgi ∈ gL implies gi ∈ L
g
i
and hence P
Y − ei ∈ I by Lemma 4.17 again. From (22), σ(f − i=1 gi (Y i − ei )) = 0 and
hence f = s̄i=1 gi (Ygi − ei ) ∈ I. (b) is proved.
17
Definition 5.4 Let L ⊂ Z[x]n be a Z[x]-lattice. The Z-saturation of L is satZ (L) = {f ∈
Z[x]n | ∃a ∈ Z s.t. af ∈ L}. The x-saturation of L is satx (L) = {f ∈ Z[x]n | xf ∈ L}. The
saturation of L is sat(L) = {f ∈ Z[x]n | ∃a ∈ Z, ∃k ∈ N s.t. axk f ∈ L}.
It is clear that the Z-saturation (x-saturation) of L is Z-saturated (x-saturated) and
sat(L) = satZ (satx (L)) = satx (satZ (L)).
Theorem 5.5 Let I be a Laurent binomial σ-ideal and L the support lattice of I. If F is
inversive, then the reflexive closure of I is also a Laurent binomial σ-ideal whose support
lattice is the x-saturation of L.
Proof: Let Ix be the reflexive closure of I and Lx = satx (L). Suppose I = [f1 , . . . , fr ], where
fi = Yfi − ci . Then L = (f1 , . . . , fr ). If L is x-saturated, by Theorem 5.3, I is reflexive.
Otherwise, there exist k1 ∈ N, bi ∈ Z[x], and h1 ∈ Z[x]n such that h1 6∈ L and
k1
x h1 =
r
X
i=1
bi fi ∈ L.
(23)
Q
k
a) ∈ F.
By Lemma 4.1, Yx 1 h1 −e
a is in I, where e
a = ri=1 cbi i . Since F is inversive, a = σ −k1 (e
k
h
h
h
1
1
1
1
Then, σ (Y − a) ∈ I, and hence Y − a ∈ Ix . Let I1 = [f1 , . . . , fr , Y − a]. It
is clear that L1 = (f1 , . . . , fr , h1 ) is the support lattice of I1 . Then I
I1 ⊂ Ix and
L
L1 ⊂ Lx . Repeating the above procedure for I1 and L1 , we obtain I2 and L2 =
(f1 , . . . , frP
, h1 , h2 ) such that h2 6∈ L1 and xk2 h2 ∈ L1 . We claim thatPL2 ⊂ Lx . Indeed, let
r
r
k
k1 +k2 h = xk1 (xk2 h ) = xk1
k1
2
x hP
2 =
0 h1 . Then by (23), x
2
2
i=1 ei fi +e
i=1 ei fi +e0 (x h1 ) =
P
r
r
xk1 i=1 ei fi + e0 i=1 bi fi ∈ L and the claim is proved. As a consequence, I2 ⊂ Ix .
Continuing the process, we have I
I1
· · · It ⊂ Ix and L L1
· · · Lt ⊂ Lx
such that Li is the support lattice of Ii . The process will terminate, since Z[x]n is Northerian.
The final Z[x]-lattice Lt is x-saturated and hence It is reflexive by Theorem 5.3. Since Lx is
the smallest x-saturated Z[x]-lattice containing L and L ⊂ Lt ⊂ Lx , we have Lt = Lx and
It = Ix .
Corollary 5.6 Let L ⊂ Z[x]n be a Z[x]-lattice. Then rk(L) = rk(satx (L)) and rk(L) =
rk(satZ (L)).
Proof: From the proof of Theorem 5.5, satx (L) = (L, h1 , . . . , ht ) and for each hi , there is
a positive integer ni such that xni hi ∈ L. Let A be a representation matrix of L. Then
a representation matrix B of Lx can be obtained by adding to A a finite number of new
columns which are linear combinations of columns of A divided by some xd . Therefore,
rk(A) = rk(B). We can prove rk(L) = rk(satZ (L)) similarly.
We now give a decomposition theorem for perfect σ-ideals.
Theorem 5.7 Let I be a Laurent binomial σ-ideal, L the support lattice of I, and LS the
saturation of L. If F is algebraically closed and inversive, then {I} is either [1] or can be
written as the intersection of Laurent reflexive prime binomial σ-ideals whose support lattice
is LS .
18
Proof: Let Ix be the reflexive closure of I and Lx = satx (L). By Theorem 5.5, Lx is the
support lattice of Ix . Suppose Ix = [f1 , . . . , fr ], fi = Yfi − ci , i = 1, . . . , r, and Lx =
(f1 , . . . , fr ). If Lx is Z-saturated, then by Theorem 5.3, Ix is reflexive prime. Otherwise,
there exist k1 ∈ N, ai ∈ Z[x], and h1 ∈ Z[x]n such that h1 6∈ Lx and
By Lemma 4.1, Yk1 h1
k1 h1 = a1 f1 + · · · + ar fr ∈ Lx .
Q
−e
a ∈ I, where e
a = ri=1 cai i . Since F is algebraically closed,
Yk1 h1 − e
a=
k1
Y
l=1
(24)
(Yh1 − e
al ) ∈ I x
where e
al , l = 1, . . . , k1 are the k1 roots of e
a. By the difference Nullstellensatz [4, p.87], we
have the following decomposition
{I} = ∩kl11=1 {I1l1 }
where I1l = [f1 , . . . , fr , Yh1 − e
al ]. Check whether I1l1 = [1] with Proposition 4.4 and discard
those trivial ones. Then the support lattice for any of I1l is L1 = (f1 , . . . , fr , h1 ). Similar to
the proof of Theorem 5.5, we can show that Ix I1l and Lx L1 ⊂ LS .
Repeating the process, we have Ix I1l1
· · · Itlt for li = 1, . . . , ki and Lx
L2 · · · Lt ⊂ LS such that Li is the support lattice of Iili for li = 1, . . . , ki and
L1
{I} = ∩klii=1 {Iili }, i = 1, . . . , t.
The process will terminate, since Z[x]n is Northerian. Since LS is the smallest Z-saturated
Z[x]-lattice containing Lx and Lx ⊂ Lt ⊂ LS , we have Lt = satZ (Lx ) = satZ (satx (L)) = LS .
Then Itlt is reflexive prime and the theorem is proved.
Since the reflexive prime components of I have the same support lattice, by Corollary
4.9, they also have the same dimension.
Corollary 5.8 Any Laurent binomial σ-ideal I is dimensionally unmixed.
5.2
Well-mixed and perfect Laurent binomial σ-ideals
In this section, we give criteria for a Laurent binomial σ-ideal to be well-mixed and perfect
in terms of its support lattice and show that the well-mixed and perfect closures of a Laurent
binomial σ-ideal are still binomial.
For S ⊂ F{Y± }, let S ′ = {f gx |f g ∈ S}. We define inductively: S0 = S, Sn = [Sn−1 ]′ , n =
1, 2, . . .. The union of the Sn is clearly a well-mixed σ-ideal and is contained in every wellmixed σ-ideal containing S. Hence this union is hSi. If I ⊂ F{Y± } is a Laurent σ-ideal, then
hIi ia called the well-mixed closure of I. We first prove some basic properties of well-mixed
σ-ideals. Note that these properties are also valid in F{Y}.
Lemma 5.9 Let I1 , . . . , Is be prime σ-ideals. Then I = ∩si=1 Ii is a well-mixed σ-ideal.
19
Proof: It is obvious.
Lemma 5.10 Let S1 , S2 be two subsets of F{Y± } which satisfy a ∈ Si implies σ(a) ∈ Si , i =
1, 2. Then [S1 ]n [S2 ]n ⊂ [S1 S2 ]n .
Proof: Let s ∈ [S1 ]1 and t ∈ [S2 ]1 . Then s = f1 g1x and t = f2 g2x where f1 g1 ∈ [S1 ], f2 g2 ∈ [S2 ].
Then, f1 g1 f2 g2 ∈ [S1 S2 ], and st = f1 f2 (g1 g2 )x ∈ [S1 S2 ]1 . Hence, [S1 ]1 [S2 ]1 ⊂ [S1 S2 ]1 . By
induction, [S1 ]n [S2 ]n ⊂ [S1 S2 ]n .
±
Lemma 5.11
a ∈ Si implies
2 be two subsets of F{Y } which
p satisfy p
p σ(a) ∈ Si , i =
p Let S1 , Sp
1, 2. Then [S1 S2 ]n = [S1 ]n ∩ [S2 ]n for n ≥ 1, and hS1 i ∩ hS2 i = hS1 S2 i.
Proof: The last statement is an immediate consequence of the first one. Since [Sp
1 S2 ] ⊂ [Si ],
we
have
[S
S
]
⊂
[S
]
for
i
=
1,
2,
and
[S
S
]
⊂
[S
]
∩
[S
]
follows.
Hence,
[S1 S2 ]n ⊂
1 2 n
i n
1 2 n
1 n
2 n
p
2
2
[S1 ]n ∩ [Sp
2 ]n . Let a ∈ [Sp
1 ]n ∩ [S2 ]n we have
pa ∈ [S1 ]n [S2 ]n . By Lemma 5.10, a ∈ [S1 S2 ]n .
Hence a ∈ [S1 S2 ]n , and [S1 ]n ∩ [S2 ]n ⊂ [S1 S2 ]n follows.
, . .p
. , Im be Laurent σ-ideals.
Lemma
p 5.12 Let I1m
I
i
=
∩
Then h∩m
i=1 hIi i.
i=1 i
p Qm
p Qm
√
m I . Then
Proof:
Let
I
=
∩
I
=
[
I
].
By
Lemma
5.11,
we
have
h i=1 Ii i =
i
i
i=1
i=1
qQ
p
p
p
p Qm
n−1
m
hIn i = . . . = ∩i=1 hIi i. Now we show that hIi = h i=1 Ii i. Since
i=1 hIi i ∩
p
p
p
p
p Qm
Qm
h
I
i
⊂
hIi.
By
Lemma
5.11,
hIi
=
hIi
∩
·
·
·
∩
hIi =
I
⊂
I,
we
have
i
i
i=1
i=1
p
p
p
p
pQ
p Qm
m
hI m i ⊂ h m
hIi
=
h
hIi
=
∩
hI
i.
I
i,
and
hence
I
i.
Then,
i
i=1
i=1 i
i=1 i
Now, we prove a basic property for a σ-field F.
√
2πi
Lemma 5.13 Let ζm = e m be the primitive m-th root of unity, where i = −1 and m ∈
Z≥2 . If F is algebraically closed, then there exists an om ∈ [0, m−1] such that gcd(om , m) = 1
om . Furthermore, the perfect σ-ideal {y m − 1} in F{y} is
and σ(ζm ) = ζm
{y m − 1} = [y m − 1, y x − y om ]
(25)
where y is a σ-indeterminate.
Q
j
Proof: Since F is algebraically closed, ζm is in F. From y m −1 = m−1
j=0 (y −ζm ) = 0, we have
Q
j
σ(y)m − 1 = m−1
j=0 (σ(y) − ζm ) = 0. Then, there exists an om such that 0 ≤ om ≤ m − 1 and
om . Suppose gcd(o , m) = d > 1 and let o = dk, m = ds, where s ∈ [1, m − 1].
σ(ζm ) = ζm
m
m
s
om s = ζ dks = ζ km = 1, which implies ζ s = 1, a contradiction.
Then σ(ζm ) = ζm
m
m
m
j
By the difference Nullstellensatz [4, p.87], we have {y m − 1} = ∩m−1
j=0 [y − ζm ]. In order
j
m
x
om ]. Since y x − y om =
to show (25), it suffices to show ∩m−1
j=0 [y − ζm ] = [y − 1, y − y
j x
jom
j
j x
xj
) + ζm
] for any 0 ≤ j ≤ m − 1,
(y − ζm
) + ζm
− y om = (y − ζm
− y om ∈ [y − ζm
j
j
m−1
m
x
x
o
m
∈ ∩j=0 [y − ζm ] and hence [y − 1, y − y om ] ⊂ ∩m−1
we have y − y
j=0 [y − ζm ]. Let
j
x
om ∈ [y − ζ j ], for j = 0, . . . , m − 1, from f ∈ [y − ζ j ], we have
f ∈ ∩m−1
m
m
j=0 [y − ζm ]. Since y − y
20
Q
P
k
j
j
f = gj (y − ζm
) + k hjk (y x − y om )x , where gj , hjk ∈ Q{y}. Then f m = m−1
j=0 (gj (y − ζm ) +
Q
P
m−1
m
x
om
m
x
om
x
om kx
j=0 gj (y −1)+p, where p ∈ [y −y ]. Hence, f ∈ [y −1, y −y ]
k hjk (y −y ) ) =
j
m
x
om ]. The lemma is proved.
and ∩m−1
j=0 [y − ζm ] ⊂ [y − 1, y − y
The number om introduced in Lemma 5.13 depends on F only and is called the m-th
transforming degree of unity. In the following corollaries, F is assumed to be algebraically
closed and hence om is fixed for any m ∈ N. From the proof of Lemma 5.13, we have
j
Corollary 5.14 y x − y om ∈ ∩m−1
j=0 [y − ζm ].
Corollary 5.15 For n, m, k in N, if n = km then on = om mod m.
on . From, σ(ζ k ) = σ(ζ ) = ζ om , we
Proof: By definition, ζnk = ζm . Then, σ(ζnk ) = ζnkon = ζm
m
m
n
on = ζ om . Then o = o mod m.
have ζm
n
m
m
Lemma 5.16 hy m − 1i = {y m − 1} = [y m − 1, y x − y om ].
Q
j
Proof: By Lemma 5.13, it suffices to show y x − y om ∈ hy m − 1i. Since y m − 1 = m−1
j=0 (y − ζm )
Q
j
i )x = (y x − ζ om i ), we have f = (y x − ζ om i )
m
and (y − ζm
i
m
m
0≤j≤m−1,j6=i (y − ζm ) ∈ hy − 1i for
x
o
m
i = 0, . . . , m − 1. We will show that y − y ∈ (f0 , . . . , fm−1 ). To show this, we need the
i
P
1 Pm−1 ζm
1
formula ym1−1 = m−1
i=0 y−ζ i from [11, p. 494]. We have
i=0 m(ζ i )m−1 (y−ζ i ) = m
m
m
m−1
1 X i
ζm f i =
m
i=0
=
m
m−1
1 X i ym − 1 x
om i
ζm
(y − ζm
)
i
m
y − ζm
1
m
i=0
m−1
X
i
ζm
i=0
= yx −
1
m
m−1
ym − 1 x
1 X i y m − 1 om i
ζm
y
−
ζ
i
i m
y − ζm
m
y − ζm
i=0
m−1
X
i=0
y m − 1 (om +1)i
ζ
.
i m
y − ζm
Pm−1
j
y m −1 (om +1)i
1 (om +1)j
1 (om +1)j y m −1
ζm
ζm
= m
. Then, g(ζm
) = m
j |y=ζ j
i ζm
y−ζm
m
y−ζm
Q
j
j om
1 (om +1)j j(m−1)
1 om j
i
i
ζm
0≤i≤m−1,i6=j (ζm − ζm ) = m ζm
1≤i≤m−1 (ζm − 1) = m ζm m = (ζm ) . Since
j
j om
deg(g(y)) ≤ m − 1 and g(ζm
) = (ζm
) for j = 0, . . . , m − 1, we have g(y) = y om . Hence
y x − y om ∈ (f0 , . . . , fm−1 ) ⊂ hy m − 1i.
Let g(y) =
Q
1
m
i=0
Corollary 5.17 For m ∈ N, a ∈ F ∗ , and f ∈ Z[x]n , we have Y(x−om )f −ax−om ∈ hYmf −am i.
f
Proof: Let z = Ya and I = [Ymf − am ]. Then z m − 1 ∈ I. By Lemma 5.16, z x−om − 1 ∈
f
hz m − 1i ⊂ hIi. Then ( Ya )x−om − 1 ∈ hIi or Y(x−om )f − ax−om ∈ hIi.
The following example shows that the generators of a well-mixed or perfect ideal depend
on the difference field F.
21
√
√
Example
5.18 Let F = Q( −3) and p = y13 − 1. Following
Lemma
5.13, if σ( −3) =
√
√
√
−3, we have o3 = 1 and hpi = {p} = [p, y1x − y1 ]. If σ( −3) = − −3, we have o3 = 2 and
hpi = {p} = [p, y1x − y12 ].
Motivated by Corollary 5.17, we have the following definition.
Definition 5.19 If a Z[x]-lattice L satisfies
mf ∈ L ⇒ (x − om )f ∈ L
(26)
where m ∈ N, f ∈ Z[x]n , and om is defined in Lemma 5.13, then it is called M-saturated.
For any Z[x]-lattice L, the smallest M-saturated Z[x]-lattice containing L is called the Msaturation of L and is denoted by satM (L).
The following result gives an effective version for condition (26).
Lemma 5.20 A Z[x]n -lattice L is M-saturated if and only if the following condition is true:
Let L1 = satZ (L) = (g1 , . . . , gs ) such that mi gi ∈ L for mi ∈ N. Then (x − omi )gi ∈ L.
Proof: We need only P
to show (x − omi )gi ∈ L implies (26). For any mf ∈ L, we have
f ∈ L1 and hence f = ri=1 qi gi , where qi ∈ Z[x]. Let t = lcm(m, m1 , . . P
. , ms ). By Corollary
r
5.15,
we
have
o
=
o
+
c
m
,
where
c
∈
Z.
Then
(x
−
o
)f
=
t
m
i
i
i
t
i=1 qi (x − ot )gi =
Pr
Pir
)g
−
q
(x
−
o
q
c
m
g
∈
L.
By
Corollary
5.15,
o
=
o
+
cm, where c ∈ Z.
mi i
t
m
i=1 i
i=1 i i i i
Then (x − om )f = (x − ot )f + cmf ∈ L.
We now give a criterion for a Laurent binomial σ-ideal to be well-mixed.
Theorem 5.21 Let ρ be a partial character and F an algebraically closed and inversive σfield. If I(ρ) is well-mixed, then Lρ is M-saturated. Conversely, if Lρ is M-saturated, then
either hI(ρ)i = [1] or I(ρ) is well-mixed.
Proof: Suppose that I(ρ) is well-mixed. If there exists an m ∈ N such that mf ∈ Lρ , then by
Lemma 4.17, there exists a c ∈ F ∗ such that Ymf − c ∈ I(ρ). Since F is algebraically closed,
there exists an a ∈ F ∗ such that c = am . Then, Ymf − am ∈ I(ρ). Since I(ρ) is well-mixed,
by Corollary 5.17, Y(x−om )f − ax−om ∈ I(ρ), and by Lemma 4.17 again, (x − om )f ∈ Lρ
follows and Lρ is M-saturated.
Conversely, let Lρ be M-saturated. If Lρ is Z-saturated, then by Theorem 5.3, I(ρ)
is prime and hence well-mixed by Lemma 5.9. Otherwise, there exists an m1 ∈ N, and
f ∈ Z[x]n such that f 6∈ Lρ and m1 f ∈ Lρ . By Lemma 4.17, there exists an a ∈ F ∗ such that
Ym1 f − am1 ∈ I(ρ). We claim that either hI(ρ)i = [1] or
1 −1
I(ρ) = ∩m
l1 =0 Il1
2πi
(27)
l1 ] and ζ
m
where Il1 = [I(ρ), Yf − aζm
m1 = e 1 . By (26), (x − om1 )f ∈ Lρ . By Lemma
1
4.17, there exists a b ∈ F ∗ such that Y(x−om1 )f − b ∈ I(ρ). Since Ym1 f − am1 ∈ I(ρ), by
l1 ] for any l . Then b − ax−om1 =
Corollary 5.14, we have Y(x−om1 )f − ax−om1 ∈ [Yf − aζm
1
1
22
Y(x−om1 )f − ax−om1 − (Y(x−om1 )f − b) ∈ Il1 for any l1 . If b 6= ax−om1 , Il1 = [1] for all l1 ,
1 −1
and hence 1 ∈ ∩m
l1 =0 Il1 ⊂ hI(ρ)i by Lemma 5.16 and hI(ρ)i = [1] follows. Now suppose
1 −1
b = ax−om1 or ax = baom1 . To prove (27), it suffices to show ∩m
l1 =0 Il1 ⊂ I(ρ). Let
P
s
j
f
l1
1 −1
f ∈ ∩m
j=0 pj σ (Y − aζm1 ), where fl1 ∈ I(ρ). By
l1 =0 Il1 . From f ∈ Il1 , we have f = fl1 +
om1
Lemma 5.13, σ(ζm1 ) = ζm1 . We thus have
l1
l1
σ(Yf − aζm
) = Yxf − bYom1 f + bYom1 f − σ(aζm
)
1
1
l1 om1
= Yom1 f (Y(x−om1 )f − b) + b(Yom1 f − aom1 ζm1
l1 om1
) + (baom1 − σ(a))ζm1
.
l1 ) =
Since Y(x−om1 )f − b ∈ I(ρ) and baom1 − σ(a) = baom1 − ax = 0, we have σ(Yf − aζm
1
l1 ), where g
gl1 + ql1 (Yf − aζm
∈
I(ρ).
Using
the
above
equation
repeatedly,
we
have
l
1
1
Qm1 −1
f − aζ l1 )) =
l1 ), where h ∈ I(ρ). Then, f m1 =
(h
+
p
(Y
f = hl1 + pl1 (Yf − aζm
l1
l1
l1
m1
l1 =0
1
Qm1 −1
Q 1 −1
f − aζ l1 ) = s + (Ym1 f − am1 )
p
∈
I(ρ),
where
s
is
in
I(ρ).
By
p
(Y
s+ m
l1
l1
m1
l1 =0
l1 =0
Corollary 4.10, we have f ∈ I(ρ). The claim is proved.
The support lattice for any of [Il1 ] is L1 = (Lρ , f ). Similar to the proof of Theorem 5.5,
we can show that I(ρ) Il1 and Lρ L1 . If L1 is not Z-saturated, there exists a k > 1 and
g ∈ Z[x]n such that g 6∈ L1 and kg ∈ L1 . Let m2 = km1 . We have m2 g = km1 g ∈ Lρ and
there exists a c ∈ F ∗ such that Ym2 g − cm2 ∈ I(ρ). Hence, (x − om2 )g ∈ Lρ ⊂ L1 and there
l2 ],
exists a d ∈ F ∗ , such that Y(x−om2 )g −d ∈ I(ρ). Let L2 = (L1 , g) and Il1 ,l2 = [Il1 , Yg −cζm
2
l2 = 0, . . . , m2 − 1. Then L1
L2 and L2 is the support lattice for all Il1 ,l2 provided
Il1 ,l2 6= [1]. Similar to the above, it can be shown that d − cx−om2 ∈ Il1 ,l2 for any l1 , l2 .
If d − cx−om2 6= 0, then Il1 ,l2 = [1] for any l1 , l2 and hIl1 i = [1] by Lemma 5.16. Since
1 −1
Laurent binomial σ-ideals are radical, hI(ρ)i = ∩m
l1 =0 hIl1 i = [1] by Lemma 5.12 and (27). If
2 −1
d−cx−om2 = 0, it can be similarly proved that Il1 = ∩m
l2 =0 Il1 ,l2 for any l1 . As a consequence,
m1 −1 m2 −1
1 −1
we have either hI(ρ)i = [1] or I(ρ) = ∩m
l1 =0 Il1 = ∩l1 =0 ∩l2 =0 Il1 ,l2 .
Repeating the process, we have either hI(ρ)i = [1] or
m1 −1
mt −1
1 −1
I(ρ) = ∩m
l1 =0 Il1 = · · · = ∩l1 =0 · · · ∩lt =0 Il1 ,...,lt
where Lρ L1 · · · Lt ⊂ satZ (Lρ ). Since Z[x]n is Notherian, the procedure will terminate
and Lt is Z-saturated. Since each Il1 ,...,lt is either [1] or a prime σ-ideal, and hence either
hI(ρ)i = [1] or I(ρ) is well-mixed by Lemma 5.9.
The following example shows that hI(ρ)i = [1] can indeed happen in Theorem 5.21.
Example 5.22 Let I = [A], where A = {y12 + 1, y1x − y1 , y22 + 1, y2x + y2 } is a σ-chain. The
support lattice of I is M-saturated. We have y22 − y12 = y22 + 1 − (y12 + 1) ∈ I. Then by
Corollary 5.17, y1 y2x − y1x y2 ∈ hIi. From y1x − y1 , y2x + y2 ∈ I, we have y1 y2 ∈ hIi and hence
1 ∈ hIi. This also shows that a binomial σ-ideal is generally not well-mixed.
Theorem 5.23 Let F be an algebraically closed and inversive σ-field and I = I(ρ) a Laurent
binomial σ-ideal. Then the well-mixed closure of I is either [1] or a Laurent binomial σ-ideal
whose support lattice is satM (Lρ ).
23
Proof: Suppose that hI(ρ)i 6= [1]. If L is not M-saturated, then there exists an m ∈ N and
f ∈ Z[x]n such that f 6∈ L, mf ∈ L, and (x − om )f 6∈ L. By Lemma 4.17, there exists a
c ∈ F ∗ such that Ymf − cm ∈ I(ρ). Let I1 = [I, Y(x−om )f − cx−om ] and L1 = (L, (x − om )f ).
I1 ⊂ hIi
By Corollary 5.17, Y(x−om )f − cx−om ∈ hI(ρ)i. Let LM = satM (L). Then I
and L L1 ⊂ LM . Repeat the procedure to construct Ii and Li for i = 2, . . . , t such that
I
I1
···
It ⊂ {I} and L
L1
···
Lt ⊂ LM . Since Z[x]n is Notherian, the
procedure will terminate at, say t. Then Lt = LM is M-saturated. By Lemma 5.27, Lt is
also x-saturated. By Theorem 5.21, It ⊂ hIi is well-mixed and hence It = hIi.
By the proof of Theorem 5.23, we have
Corollary 5.24 A Z[x]-lattice and its M-saturation have the same rank.
Example 5.25 Let p = y22 − y12 . Following the proof of Theorem 5.23, it can be shown that
hpi = {p} = [y1−2 y22 − 1, y11−x y2x−1 − 1] = [y22 − y12 , y1 y2x − y1x y2 ].
In the rest of this section, we prove similar results for the perfect closure of Laurent
binomial σ-ideals. We first give a definition.
Definition 5.26 If a Z[x]-lattice is both x-saturated and M-saturated, then it is called Psaturated. For any Z[x]-lattice L, the smallest P-saturated Z[x]-lattice containing L is called
the P-saturation of L and is denoted by satP (L).
Lemma 5.27 For any Z[x]-lattice L, satP (L) = satx (satM (L)) = satM (satx (L)).
Proof: Let L1 = satx (satM (L)) and L2 = satM (satx (L)). It suffices to show L1 = L2 . We
claim that L1 is P-saturated. Let mf ∈ L1 for m ∈ N. Then mxa f ∈ satM (L) for some
a ∈ N, which implies (x − om )xa f ∈ L ⊂ satx (satM (L)) = L1 . Since L1 is x-saturated,
(x − om )f ∈ L1 and the claim is proved. Since L ⊂ satM (L), satx (L) ⊂ satx (satM (L)) = L1 .
From the claim, L1 is P-saturated and hence L2 ⊂ satM (L1 ) = L1 .
For the other direction, we claim that L2 is x-saturated. Let xf ∈ satM (satx (L)) ⊂
satZ (satx (L)). Then there exists an m ∈ N, such that mf ∈ satx (L) which implies (x −
om )f ∈ satM (satx (L)) and hence om f = xf − (x − om )f ∈ satM (satx (L)) follows. By Lemma
5.13, gcd(om , m) = 1. Then f ∈ satM (satx (L)), and the claim is true. Since satM (L) ⊂
satM (satx (L)) = L2 = satx (satM (satx (L))), we have L1 ⊂ L2 .
It is easy to check that a σ-ideal I is perfect if and only if I is reflexive, radical, and
well-mixed. Since a Laurent binomial σ-ideal I is always radical, I is perfect if and only if I
is reflexive and well-mixed. From this observation, we can deduce the following result about
perfect Laurent binomial σ-ideal ideals.
Theorem 5.28 Let ρ be a partial character and F an algebraically closed and inversive σfield. If I(ρ) is perfect, then Lρ is P-saturated. Conversely, if Lρ is P-saturated, then either
{I(ρ)} = [1] or I(ρ) is perfect. Furthermore, the perfect closure of I(ρ) is either [1] or a
Laurent binomial σ-ideal whose support lattice is satP (Lρ ).
24
Proof: If I(ρ) is perfect, then it is well-mixed and reflexive. By Theorems 5.21 and Theorem 5.3, Lρ is M-saturated and x-saturated, and hence P-saturated. Conversely, if Lρ is
P-saturated, it is M-saturated and x-saturated. By Theorem 5.21, either hI(ρ)i = [1] or I(ρ)
is well-mixed. If hI(ρ)i = [1], {I(ρ)} = [1]. Otherwise, by Theorem 5.3, I(ρ) is reflexive.
By Corollary 4.10, I(ρ) is radical. Then I(ρ) is perfect.
By Lemma 5.27, Lp = satP (Lρ ) = satM (satx (Lρ )). Then the perfect closure of I(ρ) is
the well-mixed closure of the reflexive closure of I(ρ), and then is either [1] or a Laurent
binomial σ-ideal whose support lattice is LP by Theorems 5.5 and 5.23.
6
Binomial σ-ideal
6.1
Basic properties of binomial σ-ideal
In this section, it is shown that certain results from [7] can be extended to the difference
case using the theory of infinite Gröbner basis.
A σ-binomial in Y is a σ-polynomial with at most two terms, that is, aYa + bYb where
a, b ∈ F and a, b ∈ N[x]n . For f ∈ Z[x]n , let f + , f − ∈ Nn [x] denote the positive part and
the negative part of f such that f = f + − f − . Consider a σ-binomial f = aYa + bYb , where
a, b ∈ F ∗ . Without loss of generality, assume a > b according to the order defined in Section
3. Then f has the following canonical representation
−
f = aYa + bYb = aYg (Yf + − cYf )
(28)
n
+ ∈ N[x]. The normal
where c = −b
a , f = a − b ∈ Z[x] is a normal vector, and g = a − f
−
vector f is called the support of f . Note that gcd(Yf + , Yf ) = 1.
A σ-ideal in F{Y} is called binomial if it is generated by, possibly infinitely many, σbinomials.
In this section, F{Y} is considered as a polynomial ring in infinitely many algebraic
j
variables Θ(Y) = {yix , i = 1, . . . , n; j ≥ 0} and denoted by S = F[Θ(Y)]. A theory of
Gröbner basis in the case of infinitely many variables is developed in [14] and will be used
j
in this section. For any m ∈ N, denote Θhmi (Y) = {yix , i = 1, . . . , n; j = 0, 1, . . . , m} and
S hmi = F[Θhmi (Y)] is a polynomial ring in finitely many variables.
A monomial order in S is called compatible with the difference structure, if yix
for k1 < k2 . Only compatible monomial orders are considered in this section.
k1
< yix
k2
Let I be a σ-ideal in F{Y}. Then I is an algebraic ideal in S. By [14], we have
Lemma 6.1 Let I be a binomial σ-ideal in F{Y}. Then for a compatible monomial order,
the reduced Gröbner basis G of I exists and satisfies
hmi
G = ∪∞
m=0 G
(29)
where Ghmi = G ∩ S hmi is the reduced Gröbner basis of I hmi = I ∩ S hmi in S hmi .
Contrary to the Laurent case, a binomial σ-ideal may be infinitely generated, as shown
by the following example.
25
i
j
j
i
Example 6.2 Let I = [y1x y2x − y1x y2x : 0 ≤ i < j ∈ N]. It is clear that I does not have a
finite set of generators and hence a finite Gröbner basis. The Gröbner basis of
m
m
I hmi = I ∩ Q[y1 , y2 ; y1x , y2x ; . . . ; y1x , y2x ]
i
j
i
i
is {y1x y2x − y1x y2x : 0 ≤ i < j ≤ m} with a monomial order satisfying y1 < y2 < y1x < y2x <
i
j
j
i
m
m
· · · < y1x < y2x . Then {y1x y2x − y1x y2x : 0 ≤ i < j ∈ N} is an infinite reduced Gröbner basis
m
m
for I in the sense of [14] when y1x and y2x are treated as independent algebraic variables.
Remark 6.3 The above concept of Gröbner basis does not consider the difference structure.
The concept may be refined by introducing the reduced σ-Gröbner basis [12]. A σ-monomial
M1 is called reduced w.r.t. another σ-monomial M2 if there do not exist a σ-monomial M0
k
and a k ∈ N such that M1 = M0 M2x . Then the reduced σ-Gröbner basis of I in Example
i
i
6.2 is {y1 y2x − y1x y2 : i ∈ Z≥1 } which is still infinite. Since the purpose of Gröbner basis
in this paper is theoretic and not computational, we will use the version of infinite Gröbner
basis in the sense of [14].
With Lemma 6.1, a large portion of the properties for algebraic binomial ideals proved
by Eisenbud and Sturmfels in [7] can be extended to the difference case. The proofs follow
the same pattern: to prove a property for I, we first show that the property is valid for I
if and only if it is valid for all I hmi , and then the corresponding statement from [7] will be
used to show that the property is indeed valid for I hmi . We will illustrate the procedure in
the following corollary. For other results, we omit the proofs.
Corollary 6.4 Let I ⊂ F{Y} be a binomial σ-ideal. Then the Gröbner basis G of I consists
of σ-binomials and the normal form of any σ-term modulo G is again a σ-term.
Proof: By a σ-term, we mean the multiplication of an element from F ∗ and a σ-monomial.
By (29), it suffices to show that corollary is valid for all Ghmi , that is, the Gröbner basis
Ghmi of I hmi consists of binomials and the normal form of any term modulo Ghmi is again
a term. Since Ghmi is the Gröbner basis of I hmi = I ∩ S hmi and I hmi is a binomial ideal in
a polynomial ring with finitely many variables, the corollary follows from Proposition 1.1 in
[7].
Corollary 6.5 A σ-ideal I is binomial if and only if the reduced Gröbner basis for I consists
of σ-binomials.
Corollary 6.6 If I is a binomial σ-ideal, then the elimination ideal I ∩ F{y1 , y2 , . . . , yr } is
binomial for every r ≤ n.
The following lemma can be proved similar to its algebraic counterpart.
Lemma 6.7 If I and J are binomial σ-ideals in F{Y} then we have I ∩ J = [tI + (1 −
t)J ] ∩ F{Y} where t is a new σ-indeterminate.
The intersection of binomial σ-ideals is not binomial in general, but from Lemma 6.7 and
[7] we have
26
′
Corollary 6.8 If I and I are binomial σ-ideals and J1 , . . . , Js are σ-ideals generated by
′
σ-monomials, then [I + I ] ∩ [I + J1 ] ∩ . . . ∩ [I + Js ] is binomial.
Corollary 6.9 Let I be a binomial σ-ideal and let J1 , . . . , Js be monomial σ-ideals.
(a) The intersection [I + J1 ] ∩ · · · ∩ [I + Js ] is generated by σ-monomials modulo I.
(b) Any σ-monomial in the sum I + J1 + · · · + Js lies in one of the σ-ideals I + Ji .
Corollary 6.10 If I is a binomial σ-ideal, then for any σ-monomial M , the σ-ideal quotients [I : M ] and [I : M ∞ ] are binomial.
Corollary 6.11 Let I be a binomial σ-ideal and J a monomial σ-ideal. If f ∈ I + J and
g is the sum of those terms of f that are not individually contained in I + J , then g ∈ J .
From [7, Theorem 3.1], we have
Theorem 6.12 If I is a binomial σ-ideal, then the radical of I is binomial.
Finally, we consider the reflexive closure of binomial σ-ideals.
Lemma 6.13 A binomial σ-ideal I is reflexive if and only if bx ∈ I ⇒ b ∈ I for any
σ-binomial b ∈ F{Y}.
Proof: It suffices to prove one side of the statement, that is, if bx ∈ I ⇒ b ∈ I for any
σ-binomial b then I is reflexive. Let p be a σ-polynomial such that px ∈ I. Then, there
exists an m ∈ N such that px ∈ I hmi = I ∩ S hmi . Let G be the (finite) reduced Gröbner basis
j
of I hmi in S hmi under the variable order yix < yk for any i, k, j > 0. By Proposition 1.1 in
[7], G consists of binomials.
px can be reduced to zero by G. Due to the chosen variable
P
order, we have px = i exi gix , where exi ∈ S hmiPand gix is a σ-binomial in S hmi . Since gix are
σ-binomials in I, we have gi ∈ I. Then, p = i ei gi ∈ I and I is reflexive.
Theorem 6.14 If I is a binomial σ-ideal, then the reflexive closure of I is binomial.
k
Proof: Let I1 be the σ-ideal generated by the σ-binomials p such that px ∈ I for a k ∈ N.
We claim that I1 is the reflexive closure of I and it suffices to show that I1 is reflexive. Let
s+1
s
∈ I. Thus p ∈ I1
p be a σ-binomial such that px ∈ I1 . Then for some s ∈ N, (px )x = px
and I1 is reflexive by Lemma 6.13.
6.2
Normal binomial σ-ideal
In this section, most of the results about Laurent binomial σ-ideals proved in Sections 4 and
5 will be extended to normal binomial σ-ideals.
j
Let m be the multiplicative set generated by yix for i = 1, . . . , n, j ∈ N. A σ-ideal I is
called normal if for M ∈ m and p ∈ F{Y}, M p ∈ I implies p ∈ I. For any σ-ideal I,
I : m = {f ∈ F{Y} | ∃M ∈ m s.t. M f ∈ I}
27
is a normal σ-ideal. For any σ-ideal I in F{Y}, it is easy to check that
F{Y± }I ∩ F{Y} = I : m.
(30)
We first prove a property for general normal σ-ideals.
Lemma 6.15 A normal σ-ideal I in F{Y} is reflexive (radical, well-mixed, perfect, prime)
if and only if F{Y± }I is reflexive (radical, well-mixed, perfect, prime) in F{Y± }.
Proof: Let I = F{Y± }I be a Laurent σ-ideal. Since I is normal, from (30) we have
I ∩ F{Y} = I. If I is reflexive, it is clear that I is reflexive. For the other direction, if
f x ∈ I, then by clearing denominators of f x , there exists a σ-monomial M x in Y such that
M x f x ∈ I ∩ F{Y} = I. Since I is reflexive, M f ∈ I and hence f ∈ I, that is, I is reflexive.
The results about radical and perfect σ-ideals can be proved similarly.
We now show that I is prime if and only if I is prime. If I is prime, it is clear that I is
also prime. For the other side, let f g ∈ I. Then there exist σ-monomials N1 , N2 such that
N1 f ∈ F{Y}, N2 g ∈ F{Y}, and hence N1 f N2 g ∈ I. Since I is prime, N1 f or N2 g is in I
that is f or g is in I. The result about well-mixed σ-ideals can be proved similarly.
Given a partial character ρ on Z[x]n , we define the following binomial σ-ideal in F{Y}
+
−
I + (ρ) = [Yf − ρ(f )Yf | f ∈ Lρ ].
(31)
We will show that any normal binomial σ-ideal can be written as the form (31).
Lemma 6.16 Let ρ be a partial character on Z[x]n and I(ρ) defined in (19). Then I + (ρ) =
I(ρ) ∩ F{Y}. As a consequence, I + (ρ) is proper and normal.
P
Proof: It is clear that I + (ρ) ⊂ I(ρ) ∩ F{Y}. If f ∈ I(ρ) ∩ F{Y}, then f = si=1 fi Mi (Yfi −
ρ(fi )) where fi ∈ F, fi ∈ Lρ , and Mi are Laurent σ-monomials in Y. There exists a σmonomial M in Y such that
Mf =
s
X
i=1
+
−
fi Ni (Yfi − ρ(fi )Yfi ) ∈ I + (ρ),
(32)
where Ni is a σ-monomial in Y. We will prove f ∈ I + (ρ) from the above equation. Without
o
loss of generality, we may assume that M = ycx for some c and o ∈ N. Note that (32) is
o
k
an algebraic identity in yix , i = 1, . . . , n, k ∈ N. If Ni contains ycx as a factor, we move
−
+
o
Fi = fi Ni (Yfi − ρ(fi )Yfi ) to the left hand side of (32) and let f1 = f − Fi /ycx . Then
o
f ∈ I + (ρ) if and only if f1 ∈ I + (ρ). Repeat the above procedure until no Ni contains ycx
as a factor.
+
−
o
If s = 0 in (32), then f = 0 and the lemma is proved. Since gcd(Yfi , Yfi ) = 1, ycx
+
−
+
cannot be a factor of both Yfi and Yfi . Let Yfi be the largest σ-monomial in (32) not
−
o
containing ycx under a given σ-monomial total order . If Yfi is the largest σ-monomial
o
in (32) not containing ycx , the proving process is similar. There must exists another σ−
+
−
+
binomial fj Nj (Yfj − ρ(fj )Yfj ) such that Ni Yfi = Nj Yfj . Let Ni = Ypi , Nj = Ypj . Then
28
+
Y fi
+pi
−
= Y fj
+pj
+
−
+
and fi+ + pi = fj− + pj . We have p = fi Ni (Yfi − ρ(fi )Yfi ) + fj Nj (Yfj −
+
−
−
fj+ +pj
fi
− ρ(fi )ρ(fj )Yfi +pi ) + (fj − ρ(ffij ) )Nj (Yfj − ρ(fj )Yfj ). Since f
ρ(fj ) (Y
+
−
fj+ + pj − (fi− + pi ) = fi+ − fi− + fj+ − fj− = fi + fj ∈ Lρ , we have Yfj +pj − ρ(fi )ρ(fj )Yfi +pi
−
+
N (Yf − ρ(f )Yf ) ∈ I + (ρ), where N is a σ-monomial. As a consequence, p ∈ I + (ρ).
−
ρ(fj )Yfj ) =
=
=
If
−
+
o
N contains ycx , move the term ρ(ffij ) N (Yf − ρ(f )Yf ) to the left hand side of (32) as we
did in the first phase of the proof. After the above procedure, equation (32) is still valid.
o
Furthermore, the number of σ-binomials in (32) does not increase, no Ni contains ycx , and
−
+
o
the largest σ-monomial Yfi or Yfi not containing ycx becomes smaller. The above procedure
o
will stop after a finite number of steps, which means s = 0 in (32) and hence ycx f = 0 which
means the original f is in I + (ρ). Then I + (ρ) = I(ρ) ∩ F{Y}.
I + (ρ) = I(ρ) ∩ F{Y} is proper. For otherwise I(ρ) = [1], contradicting to Lemma 4.16.
Note that I + (ρ)F{Y± } = I(ρ). Then I + (ρ) = I(ρ) ∩ F{Y} = I + (ρ)F{Y± } ∩ F{Y} =
I + (ρ) : m, and I + (ρ) is normal.
+
−
Lemma 6.17 Let ρ be a partial character over Z[x]n . Then Yf − cYf ∈ I + (ρ) if and only
if f ∈ Lρ and c = ρ(f ).
+
−
Proof: By Lemma 6.16, Yf − cYf ∈ I + (ρ) if and only if Yf − c ∈ I(ρ) which is equivalent
to f ∈ Lρ and c = ρ(f ) by Lemma 4.17.
Lemma 6.18 If I is a normal binomial σ-ideal, then there exists a unique partial character
−
+
ρ on Z[x]n such that I = I + (ρ) and Lρ = {f ∈ Z[x]n | Yf − ρ(f )Yf ∈ I} which is called
the support lattice of I.
Proof: We have I ·F{Y± }∩F{Y} = I : m. By Theorem 4.18, there exists a partial character
ρ such that I · F{Y± } = I(ρ). Then by Lemma 6.16, I = (I : m) = I · F{Y± } ∩ F{Y} =
−
+
I(ρ) ∩ F{Y} = I + (ρ). By Lemma 6.17, we have Lρ = {f ∈ Z[x]n | Yf − ρ(f )Yf ∈ I =
I + (ρ)}. The uniqueness of ρ comes from the fact that Lρ is uniquely determined by I.
By Lemmas 6.16 and 6.18, we have
Theorem 6.19 The map I(ρ) ⇒ I + (ρ) gives a one to one correspondence between Laurent
binomial σ-ideals in F{Y± } and normal binomial σ-ideals in F{Y}.
Due to Lemma 6.16 and Theorem 6.19, most properties of Laurent binomial σ-ideals can
be extended to normal binomial σ-ideals. As a consequence of Corollary 4.10, Lemma 6.15,
and Lemma 6.16, we have
Corollary 6.20 A normal binomial σ-ideal is radical.
As a consequence of Theorem 5.5, Lemma 6.15, and Theorem 6.19.
Corollary 6.21 If F is inversive, then the reflexive closure of I + (ρ) is also a normal binomial σ-ideal whose support lattice is the x-saturation of Lρ .
29
Corollary 6.22 If F is algebraically closed and inversive, then
(a) Lρ is Z-saturated if and only if I + (ρ) is prime;
(b) Lρ is x-saturated if and only if I + (ρ) is reflexive;
(c) Lρ is saturated if and only if I + (ρ) is reflexive prime.
Proof: It is easy to show that I(ρ) = I + (ρ)F{Y± }. Then the corollary is a consequence of
Theorem 5.3, Lemma 6.15, and Lemma 6.16.
For properties related with perfect σ-ideals, it becomes more complicated. Direct extension of Theorems 5.7, and 5.28 to the normal binomial case is not correct as shown by the
following example.
Example 6.23 Let I = [y1x −y1 , y22 − y12 , y2x + y2 ] which
is a normal binomial σ-ideal whose
x − 1 −2
0
representation matrix is L =
. Since o2 = 1, L is P -saturated. Also,
0
2 x−1
x − 1 −1
Ls = sat(L) =
. We have {I} = {I, y2 − y1 } ∩ {I, y2 + y1 } = [y1 , y2 ]. Then
0
1
{I} =
6 [1] and I is not perfect and hence Theorems 5.28 are not correct. Theorem 5.7 is also
not correct, since the supporting lattice of the prime component of I is not Ls . This example
also shows that the perfect closure of a normal binomial σ-ideal is not necessarily normal.
It can be seen that the problem is due to the occurrence of σ-monomials. For any partial
character ρ, it can be shown that
{I + (ρ)} : m = {I(ρ)} ∩ F{Y}.
(33)
We thus have the following modifications for Theorems 5.28 and 5.7.
Corollary 6.24 Let F be an inversive and algebraically closed σ-field. If I + (ρ) is perfect,
then Lρ is P -saturated. Conversely, if Lρ is P -saturated, then either {I} : m = [1] or I is
perfect. For any ρ, either {I + (ρ)} : m = [1] or {I + (ρ)} : m is a binomial σ-ideal whose
support lattice is the P -saturation of Lρ .
Proof: If I is perfect, by Lemma 6.15, I(ρ) = IF{Y± } is also perfect. By Theorem 5.28,
Lρ is P -saturated. If Lρ is P -saturated and x-saturated, by Theorem 5.28, either I(ρ) = [1]
or I(ρ) is perfect. If I(ρ) = [1], by (33), {I} : m = [1]. If I(ρ) is perfect, by Lemma 6.15,
I = I + (ρ) is also perfect.
Similar results hold for normal well-mixed σ-ideals.
In the rest of this section, we give decomposition theorems for perfect binomial σ-ideals.
We first consider normal binomial σ-ideals. By Corollary 6.20 and Example 5.22, a normal
binomial σ-ideal is radical but may not be perfect.
Theorem 6.25 Let I = I + (ρ) be a normal binomial σ-ideal and F an inversive and algebraically closed σ-field. Then {I} : m is either [1] or can be written as the intersection of
reflexive prime binomial σ-ideals whose support lattice is the saturation lattice of Lρ .
30
T
Proof: By Theorem 5.7, either {I(ρ)} = [1] or {I(ρ)} = si=1 I(ρi ), where I(ρi ) are reflexive
prime σ-ideals whose support lattices are sat(LT
6.16, either {I + (ρ)} :
ρ ). By (33) and Lemma
T
s
s
m = [1] or {I + (ρ)} : m = {I(ρ)} ∩ F{Y} = ( i=1 I(ρi )) ∩ F{Y} = i=1 (I(ρi ) ∩ F{Y}) =
T
s
+
+
i=1 I (ρi ). By Corollary 6.22, I (ρi ) is reflexive and prime whose support lattices are the
saturation of Lρ .
Now, consider general binomial σ-ideals.
Lemma 6.26 I ⊂ F{Y} is a reflexive prime binomial σ-ideal if and only if I = [yi1 , . . . , yis ]
+I1 , where {yi1 , . . . , yis } = Y ∩ I, {z1 , . . . , zt } = Y\I, and I1 is a normal binomial reflexive
prime σ-ideal in F{z1 , . . . , zt }.
j
Proof: If I is reflexive and prime, then (yix )d ∈ I if and only if yi ∈ I. Let I1 = I ∩
F{z1 , . . . , zt }. Then I = [yi1 , . . . , yis ] + I1 . I1 is clearly reflexive and prime. We still
need to show that I1 is normal. Let N f ∈ I1 for a σ-monomial N in {z1 , . . . , zt } and
f ∈ F{z1 , . . . , zt }. N cannot be in I1 . Otherwise, some zi is in I1 since I1 is reflexive and
prime, which contradicts to {z1 , . . . , zt } = Y\I. Therefore, f ∈ I1 and I1 is normal. The
other direction is trivial.
The σ-ideal I in Lemma 6.26 is said to be quasi-normal. The following result can be
proved similarly to Theorem 5.7.
Theorem 6.27 Let I be a binomial σ-ideal. If F is algebraically closed and inversive,
then the perfect σ-ideal {I} is either [1] or the intersection of quasi-normal reflexive prime
binomial σ-ideals.
Proof: We prove the theorem by induction on n. Let I1 = {I} : m. Then {I} = I1 ∩
∩ni=1 {I, yi }. It is easy to check I1 = {I : m} : m. Since I : m is normal, by Theorem 6.25,
I1 is either [1] or intersection of normal reflexive prime σ-ideals. If n = 1, then {I, yi } must
be either [y1 ] or [1]. Then the theorem is proved for n = 1. Suppose the theorem is valid for
n = 1, . . . , k − 1. Still use {I} = I1 ∩ ∩ni=1 {I, yi }. Let Ii be the σ-ideal obtained by setting
yi to 0 in I. By the induction hypothesis, Ii can be written as intersection of quasi-normal
reflexive prime σ-ideals in F{Y \ {yi }}. So the theorem is also valid for {I, yi } = {Ii , yi }.
The theorem is proved.
6.3
Characteristic set for normal binomial σ-ideal
The theory of characteristic set given in Section 4.2 will be extended to the normal σ-binomial
case.
Let ρ be a partial character over Z[x]n , Lρ = (f1 , . . . , fs ) where f = {f1 , . . . , fs } is a
reduced Gröbner basis, and
+
−
+
−
A+ (ρ) : Yf1 − ρ(f1 )Yf1 , . . . , Yfs − ρ(fs )Yfs .
We have the following canonical representation for normal binomial σ-ideals.
31
(34)
Theorem 6.28 Use the notations in (34). Then I + (ρ) = sat(A+ (ρ)). Furthermore, A+ (ρ)
is a regular and coherent σ-chain and hence is a characteristic set of I + (ρ).
Proof: Let I1 = [A+ (ρ)] : m. We claim I1 = sat(A+ (ρ)). It is clear that sat(A+ (ρ)) ⊂
[A+ (ρ)] : m = I1 . For the other direction, let p ∈ I1 and p1 = prem(p, A+ (ρ)) which
is reduced w.r.t. A+ (ρ). By (3), p1 ∈ I1 . As a consequence, p1 ∈ [A(ρ)] as Laurent σpolynomials in F{Y± }. By Lemma 4.16, A(ρ) is a characteristic set of [A(ρ)]. Since p1 is
reduced w.r.t. A+ (ρ), it is also reduced w.r.t. A(ρ). Then p1 = 0 and hence the claim is
proved.
We now prove I + (ρ) = sat(A+ (ρ)). By the above claim, Lemma 4.16, and Lemma 6.16,
sat(A+ (ρ)) = [A+ (ρ)] : m = [A+ (ρ)]F{Y± } ∩ F{Y} = [A(ρ)] ∩ F{Y} = I(ρ) ∩ F{Y} =
I + (ρ).
It remains to prove that A+ (ρ) is a characteristic set of I1 = [A+ (ρ)] : m. By definition,
it suffices to show that if p ∈ I1 is reduced w.r.t. A+ (ρ) then p = 0. Let Ai = Yfi − ρ(fi )
fi+ − ρ(f )Yfi− . Since p ∈ I , there exist a σ-monomial M and f
and A+
1
i,j ∈ F{Y}
i = Y
P
P i
j
+ xj
±
x
such that M p = i,j fi,j (Ai ) . Then in F{Y }, we have p = i,j gi,j Ai ∈ [A(ρ)], where
gi,j ∈ F{Y± }. Since p is reduced w.r.t. A+ (ρ), it is also reduced w.r.t. A(ρ). By Lemma
4.16, A(ρ) is a characteristic set of [A(ρ)] and hence p = 0. The claim is proved.
Since I1 = sat(A+ (ρ)), A+ (ρ) is also a characteristic set of sat(A+ (ρ)). By Theorem 2.1,
A+ (ρ) is regular and coherent.
Example 6.29 Let L = ([1 − x, x − 1]τ ) be a Z[x]-module and ρ the trivial partial character
on L, that is, ρ(f ) = 1 for f ∈L. By Theorem 6.28, I + (ρ) = sat[y1 y2x − y1x y2 ] ⊆ Q{y1 , y2 }.
j
i
By Theorem 6.16, I + (ρ) is a reflexive prime σ-ideal. We can show that I + (ρ) = [y1x y2x −
i
j
y1x y2x | 0 ≤ i < j ≤ m], which is an infinitely generated σ-ideal.
As a consequence of Theorem 6.19, Theorem 6.28, and Lemma 4.16, we have
Corollary 6.30 Let A(ρ) and A+ (ρ) be defined in (20) and (34), respectively. Then
([A(ρ)]F{Y± }) ∩ F{Y} = sat(A+ (ρ)).
As a consequence of Theorem 5.3, Corollary 6.22, and Theorem 6.28.
Corollary 6.31 [A(ρ)] is a reflexive (prime) σ-ideal in F{Y± } if and only if sat(A+ (ρ)) is
a reflexive (prime) σ-ideal in F{Y}.
We now prove the converse of Theorem 6.28. Let fi ∈ Z[x]n and ci ∈ F ∗ , i = 1, . . . , s.
Consider the following σ-chains
A : Yf1 − c1 , . . . , Yfs − cs
A+ : Y
f1+
f1−
− c1 Y , . . . , Y
+
Yfs
(35)
− cs Y
fs−
in F{Y± } and F{Y}, respectively. Notice that, when talking about A+ (or A), all operations
are performed in F{Y} (or F{Y± }). Since fi are assumed to be normal, A+ is a σ-chain if
and only if A is a Laurent σ-chain.
32
Lemma 6.32 Use the notations in (35). Let p = aYa + bYb = aN (Yf − c) ∈ F{Y}, where
a, b ∈ N[x]n , f ∈ Z[x]n , N is a σ-monomial, c ∈ F ∗ . If A+ is coherent and regular, then
prem(p, A+ ) = 0 implies prem(Yf − c, A) = 0.
Proof: Since prem(p, A+ ) = 0, there exists a σ-monomial M1 such that M1 p ∈ [A+ ]. Let
p1 = Yf − c. Since r1 = prem(p1 , A) = Yg − cg , by Lemma 4.5, there exists a c1 ∈ F ∗ such
that r1 − c1 p1 ∈ [A]. Then, there exists a σ-monomial M2 such that M2 N r1 , M2 N p1 ∈ F{Y}
and M2 N (r1 − c1 p1 ) ∈ [A+ ] and hence M2 M1 N (r1 − c1 p1 ) = M2 M1 N r1 − ca1 M2 M1 p ∈ [A+ ].
Let M = M1 M2 N . From M1 p ∈ [A+ ], we have M r1 ∈ [A+ ] ⊂ sat(A+ ).
+
−
d xo
Suppose Ai = Yfi − cYfi = Ii+ ycii i − cIi− , where yci is the leading variable of Ai . A
j
o +k
variable like ycxi i for k ∈ N is called a main variable of A+ . A variable yix is called a
parameter of A+ is it is not a main variable. If M contains a main variable of A+ as a factor.
o +k
Then let z = ycxi i be the largest one appearing in M under the variable ordering induced
by the lexicographical of the index (ci , oi +k). Let s = deg(M, z) and M1 = M/(z s ). We may
assume that di is a factor of s. Otherwise, let s1 = ⌊ dsi ⌋, s0 = s − s1 di , and M = M z di −s0 =
M1 z di (s1 +1) . We still have M r1 ∈ sat(A+ ). We may use Ai = 0 to eliminate z from M as
k
k
k
k
d xo
follows: M1 z s−di (cIi− )x r1 = M1 z s−di (Ii+ ycii i − Ai )x r1 = M (Ii+ )x r1 − M1 z s−di (Ai )x r1 ∈
k
sat(A+ ). Note that deg(M1 z s−di (cIi− )x , z) = s − di . Repeat the above procedure, we may
find a σ-monomial N such that N r1 ∈ sat(A+ ), N does not contain z as a factor, and any
j
variable yix in M is smaller than z in the given variable ordering. Repeat the procedure, we
may finally obtain a σ-monomial L such that L does not contain main variables of A+ as
factors and Lr1 ∈ sat(A+ ). Since L contains only parameters of A+ and r1 is reduced w.r.t.
A+ , Lr1 is also reduced w.r.t. A+ . Since A+ is regular and coherent, by Lemma 3.4, it is
the characteristic set of sat(A+ ). Therefore, Lr1 = 0, and r1 = 0.
The following example shows that if prem(p, A+ ) 6= 0 then the relation between prem(p,
and prem(Yf − c, A) may be complicated, where p = aYa + bYb = aN (Yf − c).
A+ )
+
2
Example 6.33 Let p = y2 (y2 − 1), A1 = y1−1 y22 − 1, and A+
1 = y2 − y1 . Then prem(p, A1 ) =
±
y1 − y2 in F{y2 }. But in F{y2 }, p is represented as pe = y2 − 1 and prem(e
p, A1 ) = y2 − 1.
Lemma 6.34 Use the notations in (35). A is a regular and coherent σ-chain in F{Y± } if
and only if A+ is a regular and coherent σ-chain in F{Y}.
Proof: If A is regular and coherent, by Theorem 4.18 and Corollary 4.19, there exists a
partial character ρ over Z[x]n such that Lρ = (f1 , . . . , fs ), ρ(fi ) = ci , and I(ρ) = [A]. By
Theorem 6.28, A+ = A+ (ρ) is regular and coherent.
Assume that A+ is regular and coherent. We first show that [A] 6= [1] in F{Y± }.
It suffices to show that sat(A+ ) does not contain a σ-monomial. Suppose the contrary,
there is a σ-monomial M ∈ sat(A+ ). Since A+ is a regular and coherent chain, we have
prem(M, A+ ) = 0. Now consider the procedure of prem, it can be shown that the pseudoremainder of a nonzero σ-monomial w.r.t. a binomial σ-chain is still a nonzero σ-monomial,
a contradiction.
Note that A is always regular since σ-monomials are invertible in F{Y± }. Then, it
suffices to prove that A is coherent.
33
+
−
fi − c Yfi . Assume A+ and A+ (i < j) have the same
Let Ai = Yfi − ci and A+
i
i
j
i = Y
d xo j
oi
−
+ di x
+ j
leading variable yl , and A+
− ci Ii− , A+
− cj Ij− , where Ii− = Yfi . From
i = Ii y l
j = Ij y l
Definition 3.6, we have oi < oj and di |dj . Let di = tdj where t ∈ N. From (4),
+ x
+
∆(A+
i , Aj ) = prem((Ai )
oj −oi
x
t − t
, A+
j ) = cj (Ij ) (Ii )
oj −oi
− (Ij+ )t (ci Ii+ )x
+
oj −oi
.
−
+
t
h − c Yh ), where
Comparing to (18), if ∆(Ai , Aj ) = Yh − cf , then ∆(A+
f
i , Aj ) = cj M (Y
+
+
+
+
M is a σ-monomial. Since A is coherent, prem(∆(Ai , Aj ), A ) = 0. By Lemma 6.32,
prem(∆(Ai , Aj ), A) = 0 which implies that A is coherent.
We now prove the converse of Theorem 6.28.
Theorem 6.35 Use the notations in (35). If A+ is a regular and coherent σ-chain, then
there is a partial character ρ over Z[x]n such that Lρ = (f1 , . . . , fs ), ρ(fi ) = ci , I(ρ) = [A],
and I + (ρ) = sat(A+ )
Proof: By Lemma 6.34, A is regular and coherent. By Theorem 4.14, f is a reduced Gröbner
basis for a Z[x]-lattice and [A] ⊂ F{Y± } is proper. By Theorem 4.18 and Corollary 4.19,
there exists a partial character ρ such that Lρ = (f1 , . . . , fs ), ρ(fi ) = ci , and I(ρ) = [A]. By
Theorem 6.28, I + (ρ) = sat(A+ (ρ)) = sat(A+ ).
As a consequence of Theorem 6.35 and Lemma 6.16, we have the following canonical
representation for a normal binomial σ-ideal.
Corollary 6.36 I is a normal binomial σ-ideal if and only if I = sat(A+ ), where A+ is a
regular and coherent chain given in (35).
6.4
Perfect closure of binomial σ-ideal and binomial σ-variety
In this section, we will show that the perfect closure of a binomial σ-ideal is also binomial.
We will also give a geometric description of the zero set of a binomial σ-ideal. For the perfect
closure of a binomial σ-ideal, we have
Theorem 6.37 Let F be an algebraically closed and inversive σ-field. Then the perfect
closure of a binomial σ-ideal I is binomial.
We first remark that it is not known wether the well-mixed closure of a binomial σ-ideal
is still binomial. Before proving Theorem 6.37, we first prove several lemmas. In the rest of
this section, we assume that I ⊆ S = F{Y} and m the set of σ-monomials in S.
Lemma 6.38 If I is a binomial σ-ideal, then {I} : m is either [1] or a binomial σ-ideal.
Proof: It is easy to check {I}F{Y± } = {IF{Y± }}. By (30), {I} : m = {I}F{Y± }∩F{Y} =
{IF{Y± }} ∩ F{Y}. Now the lemma follows from Theorem 5.28.
34
Lemma 6.39 If I is a σ-ideal in F{Y}, then
{I} = {I} : m ∩ {I + y1 } ∩ · · · ∩ {I + yn }
(36)
Proof: The right hand side of (36) clearly contains {I}. It suffices to show that every
reflexive prime P containing I contains one of the σ-ideals on the right-hand side of (36).
If {I} : m ⊆ P , we are done. Otherwise, there exists an element f ∈ ({I} : m) \ P which
implies that there exists a σ-monomial M such that M f ∈ {I} ⊆ P . This implies yi ∈ P
for some i. Thus, P contains {I + yi } as required.
Lemma 6.40 Let I be a binomial σ-ideal in S = F{Y} and S ′ = F{y1 , . . . , yn−1 }. If
I ′ = I ∩ S ′ , then [I + yn ] is the sum of [I ′ S + yn ] and a monomial σ-ideal in S ′ .
k
Proof: Every σ-binomial involving ynx is either contained in [yn ] or is congruent modulo [yn ]
to a σ-monomial in S ′ . Thus, all generators of I which are not in I ′ may be replaced by
σ-monomials in S ′ when forming a generating set for [I + yn ].
Lemma 6.41 Let I be a perfect binomial σ-ideal in S = F{Y}. If M is a σ-monomial
σ-ideal, then {I + M} = [I + M1 ] for some monomial σ-ideal M1 .
Proof: If 1 ∈ M, then the
T lemma is obviously valid. Otherwise, [I + M] : m = [1]. Lemma
6.39 yields {I + M} = i=n
i=1 {I + M + yi }. By Corollary 6.9, we need only to show that
{I + M + yi } is the sum of I and a monomial σ-ideal. For simplicity, let i = n and write
S ′ = F{y1 , y2 , . . . , yn−1 }. Since I is perfect, the σ-ideal I ′ = I ∩ S ′ is perfect as well. By
Lemma 6.40, [I + M + yn ] = [I ′ S + M′ S + yn ] where M′ is a monomial σ-ideal in S ′ . By
induction on n, the perfect closure of I ′ + M′ in S ′ has the form I ′ + M′1 , where M′1 is a
monomial σ-ideal of S ′ . Putting this together, we have
{I + M + yn } = {I ′ S + M′ S + yn } = [I ′ S + M′1 S + yn ]
⊆ [I + M′1 S + yn ] ⊆ {I + M + yn }.
So {I + M + yn } = [I + M′1 S + yn ] is I plus a monomial σ-ideal, as required.
Proof of Theorem 6.37: We will prove the theorem by induction on n. By Lemma 6.38,
I1 = {I} : m is binomial. For n = 1, by Lemma 6.39, {I} = I1 ∩ {I + y1 }. If {I + y1 } = 1
then {I} = I1 is binomial. Otherwise {I + y} = [y] and hence I ⊂ [y]. Since I ⊂ I1 ,
{I} = I1 ∩ [y] = [I + I1 ] ∩ [I + y] is binomial by Lemma 6.8. Suppose the lemma is valid
for n − 1 variables and let I be a binomial σ-ideal in S = F{Y}. Let Ij := I ∩ Sj , where
Sj = F{y1 , . . . , yj−1 , yj+1 , . . . , yn }. By the induction hypothesis, we may assume that the
perfect closure of each Ij is binomial. Adding these binomial σ-ideals to I, we may assume
that each Ij is perfect begin with. By Lemma 6.38, I1 = {I} : m is binomial. Then
there exists a binomial σ-ideal I ′ , say I ′ = I1 , such that I1 = [I + I ′ ]. By Lemma 6.40,
[I + yj ] = [Ij S + Jj S + yj ], where Jj is a monomial σ-ideal in Sj . Since Ij is perfect, the σideal Ij S is perfect, so we can apply Lemma 6.41 with M = [Jj S +yi] to see that there exists
a monomial σ-ideal Mj in S such that {I +yj } = {I
Tj S +Jj S +yj } = [Ij S +Mj ] = [I +Mj ].
By Lemma 6.39 and Corollary 6.8, {I} = [I + I ′ ] ∩nj=1 [I + Mj ] is binomial.
35
Example 6.42 Let p = y22 − y12 . Following the proof of Theorem 6.37, {p} = ({p} : m) ∩
[y1 , y2 ]. By Example 5.25 and Corollary 6.30, I1 = {p} : m = sat[y22 − y12 , y1 y2x − y1x y2 ] =
j
j
i
i
[y1 y2x − y1x y2 , y21+x − y11+x | i, j ∈ N]. Thus, {p} = I1 ∩ [y1 , y2 ] = I1 .
In the rest of this section, we give a geometric description of the zero set of a binomial
σ-ideal, which is a generalization of Theorem 4.1 in [7] to the difference case. The basic idea
of the proof also follows [7], except we need to consider the distinction between the perfect
σ-ideals and radical ideals.
We decompose the affine n-space (A)n into the union of 2n σ-coordinate flats:
(A∗ )Ω := {(a1 , a2 , . . . , an ) | ai 6= 0, i ∈ Ω; ai = 0, i ∈
/ Ω}
where Ω runs over all subsets of {1, 2, . . . , n}. The Cohn closure of (A∗ )Ω in (A)n is defined
by the σ-ideal
M (Ω) := [yi |i ∈
/ Ω] ⊂ F{Y}.
The σ-coordinate ring of (A∗ )Ω is the Laurent polynomial σ-ring F{Ω± } := F{yi , yi−1 , i ∈ Ω}.
′
We can define a coordinate projection (A∗ )Ω −→ (A∗ )Ω whenever Ω ⊆ Ω′ ⊆ {1, 2, . . . , n} by
setting all those coordinates not in Ω to zero.
If X is any σ-variety of (A)n and I = I(X) ⊆ F{Y}, then the Cohn closure of the
intersection of X with (A∗ )Ω corresponds to the σ-ideal
IΩ := [I + M (Ω)] : mΩ ⊂ F{Y}
Q
m (x)
where mΩ = { i∈Ω yi i |mi (x) ∈ N[x]}. Since I is perfect, by the difference Nullstellsatz
[4, p.87]
\
I = {IΩ }.
Ω
If I is binomial, then by Corollary 6.10 the σ-ideal IΩ is also binomial.
Lemma 6.43 Let R := F{z1 , z1−1 , . . . , zt , zt−1 } ⊂ R′ := F{z1 , z1−1 , . . . , zt , zt−1 , y1 , . . . , ys } be
a Laurent polynomial σ-ring and a polynomial σ-ring over it. If B ⊂ R′ is a binomial σideal and M ⊂ R′ is a monomial σ-ideal such that [B + M ] is a proper σ-ideal in R′ , then
[B + M ] ∩ R = B ∩ R.
Proof: This is a σ-version of [7, Lemma 4.2], which can be proved similarly.
We can make a classification of all binomial σ-varieties X by intersecting X with (A∗ )Ω ,
since by Theorem 6.37, the perfect closure of a binomial σ-ideal is still binomial.
Theorem 6.44 Let F be any algebraically closed and inversive σ-field. A σ-variety X ⊂ An
is generated by σ-binomials if and only if the following three conditions hold.
(1) For each (A∗ )Ω , the σ-variety X ∩ (A∗ )Ω is generated by σ-binomials.
(2) The family of sets U = {Ω ⊆ {1, 2, . . . , n}|X ∩ (A∗ )Ω 6= ∅} is closed under taking
intersections.
36
(3) If Ω1 , Ω2 ∈ U and Ω1 ⊂ Ω2 , then the coordinate projection (A∗ )Ω2 −→ (A∗ )Ω1 maps
X ∩ (A∗ )Ω2 onto a subset of X ∩ (A∗ )Ω1 .
The above theorem can be reduced to the following algebraic version.
Theorem 6.45 Let F be any algebraically closed and inversive σ-field. A perfect σ-ideal
I ⊂ F{Y} is binomial if and only if the following three conditions hold.
(1) For each Ω ⊆ {1, . . . , n}, IΩ is binomial.
(2) U = {Ω ⊆ {1, 2, . . . , n} | {IΩ } =
6 [1]} is closed under taking intersections.
(3) If Ω1 , Ω2 ∈ U and Ω1 ⊂ Ω2 , then IΩ1 ∩ F{Ω1 } ⊂ IΩ2 , where F{Ω1 } = F{yi | yi ∈ Ω1 }.
Proof: Suppose I is a perfect σ-ideal in F{Y}. Since I is binomial, by Lemma 6.10 IΩ is
also binomial and (1) is proved. To prove (2) by contradiction, assume that for Ω1 , Ω2 ∈ U ,
{IΩ1 } =
6 [1], {IΩ2 } =
6 [1], {IΩ1Q
∩Ω2 } = [1]. We consider two cases. If IΩ1 ∩Ω2 = [1], then for
some m(x) ∈ N[x] we have ( i∈Ω1 ∩Ω2 yi )m(x) ∈ [I + M (Ω1 ) + M (Ω2 )]. By Corollary 6.9,
Q
( i∈Ω1 ∩Ω2 yi )m(x) is either in [I + M (Ω1 )] or [I + M (Ω2 )], so IΩ1 or IΩ2 is [1]. For the second
case, we have IΩ1 ∩Ω2 6= [1] and {IΩ1 ∩Ω2 } = [1]. Then there exist a finite number of proper
σ-binomials B1 , . . . , Bs and σ-monomials m1 , . . . , ms in F{Ω1 ∩ Ω2 } such that mi Bi ∈ I
and {B1 , . . . , Bs , yi , i ∈
/ Ω1 ∩ Ω2 } = [1]. We thus have {B1 , . . . , Bs } = [1]. Since mi Bi ∈
I ∩ F{Ω1 ∩ Ω2 }, we have Bi ∈ IΩ1 and Bi ∈ IΩ2 and thus {IΩ1 } = {IΩ2 } = [1]. To prove
(3), given Ω1 , Ω2 ∈ U and Ω1 ⊂ Ω2 , we have IΩ2 = [IΩ2 : mΩ1 ]. Set R′ = F{Ω±
/ 1 },
1 }{{yi }i∈Ω
then
′
[I + M (Ω1 )]R′ ∩ F{Ω±
1 } ⊆ I Ω2 R .
′
Since Ω1 ∈ U , the σ-ideal [I +M (Ω1 )]R is proper. By Lemma 6.43, we have [I +M (Ω1 )]R′ ∩
±
±
′
′
F{Ω±
1 } = IR ∩ F{Ω1 } ⊂ IΩ2 R ∩ F{Ω1 }. So IΩ1 ∩ F{Ω1 } ⊂ IΩ2 .
To prove the other driection, let I be a perfect σ-ideal satisfying the three conditions.
By the difference Nullstellensatz, I = ∩Ω∈U {IΩ }. By condition (2), U is a partially ordered
set under the inclusion for subsets of {1, . . . , n}. For each Ω ∈ U , we set J (Ω) = [IΩ ∩
F{Ω}]F{Y} with the properties that if Ω1 ⊂ Ω2 , {J (Ω1 )} ⊂ {J (Ω2 )}. Note that [MΩ1 ∩Ω2 ] ⊂
[MΩ1 + MΩ2 ]. Then we have
I = ∩Ω∈U {IΩ } = ∩Ω∈U {J (Ω) + M (Ω)}.
Now we will prove that
∩Ω∈U {J (Ω) + M (Ω)} = {∩Ω∈U M (Ω) +
X
Ω∈U
{J (Ω) ∩ (∩Ωη +Ω M (Ωη ))}}.
(37)
If Ω2 ⊇ Ω1 , then {J (Ω2 ) + M (Ω2 )} ⊇ {J (Ω2 )} ⊇ {J (Ω1 )} ⊇ {J (Ω1 ) ∩ ∩Ωη +Ω1 M (Ωη )}.
If Ω2 + Ω1 , we have {J (Ω2 ) + M (Ω2 )} ⊇ M {Ω2 } ⊇ {J (Ω1 ) ∩ ∩Ωη +Ω1 M (Ωη )}. So the
left hand side contains the right hand side P
of (37). For the other direction, consider a
reflexive prime σ-ideal P ⊇ [∩Ω∈U M (Ω) + Ω∈U {J (Ω) ∩ ∩Ωη +Ω M (Ωη )}] and set V =
{Ω ∈ U |M (Ω) ⊂ P }. Then V is a finite partially order set and nonempty since P ⊇
37
∩Ω∈U M (Ω) and {MΩ1 ∩Ω2 } ⊂ {MΩ1 + MΩ2 }. Let Ω0 be the smallest element of V such that
P ⊇ MΩ0 . At the same time, P ⊃ J (Ω0 ) ∩ ∩Ωη +Ω0 M (Ωη ), then P ⊇ J (Ω0 ). Therefore,
P ⊇ J (Ω0 ) +P
M (Ω0 ) and P contains the left hand side of (37) and (37) is proved. Since
∩Ω∈U M (Ω) + Ω∈U {J (Ω) ∩ ∩Ωη +Ω M (Ωη )} is binomial, the theorem follows from (37).
7
Algorithms
In this section, we give algorithms for most of the results in the preceding sections. In
particular, we give an algorithm to decompose a finitely generated perfect binomial σ-ideal
as the intersection of reflexive and prime binomial σ-ideals. The following basic algorithms
will be used.
• Algorithm GHNF. Let f be a finite set of Z[x]n . The algorithm computes the
generalized Hermite normal form of [f ], or equivalently, the reduced Gröbner basis of
the Z[x]-module (f ) [5, p. 197]. A polynomial-time algorithm is given in [15].
• Algorithm GKER. For a matrix M ∈ Z[x]n×s , compute a set of generators of the
Z[x]-lattice: kerZ[x] (M ) = {X ∈ Z[x]s | M X = 0}. This can be done by combining
Algorithm GHNF and Theorem 3.9.
Let D be Z, Q[x], Zp [x], or Q[x]/(q(x)), where q(x) is an irreducible polynomial in Q[x].
Then D is either a PID or a field. The following algorithms will be used.
• Algorithm HNF. For M ∈ Dn×s , compute the Hermite normal form of M [3, p.68].
• Algorithm KER. For a matrix M ∈ Dn×s , compute a basis for the D-module:
kerD (M ) = {X ∈ Ds | M X = 0} [3, p.74].
7.1
x-saturation of Z[x]-lattice
In this section, we give algorithms to check whether a Z[x]-lattice L is x-saturated and in
the negative case to compute the x-saturation of L.
n
Let f1 , . . . , fs ∈ Z[x]
Ps and L = (f1 , . . . , fs ). If L is not x-saturated, then there exist
e =
gi ∈P
Z[x] such that i=1 gi fi = xh and h 6∈ L. Setting gi (x) = gi (0) + xe
gi (x) and h
s
h − i=1 e
gi (x)fi , we have
s
X
e
gi (0)fi = xh
(38)
i=1
e∈
where h
6 L. Setting x = 0 in the above equation, we have
s
X
gi (0)fi (0) = 0,
i=1
38
that is, G = (g1 (0), . . . , gs (0))τ is in the kernel of the matrix F = [f1 (0), . . . , fs (0)] ∈ Zn×s ,
e This
which can be computed efficiently [3, page 74]. From G and (38), we may compute h.
observation leads to the following algorithm.
Algorithm 1 — XFactor([f1 , . . . , fs ])
Input: A generalized Hermite normal form [f1 , . . . , fs ] ⊂ Z[x]n×s .
Output:∅, if the Z[x]-lattice L = (f1 , . . . , fs ) is x-saturated; otherwise, a finite
set {(hi , ei )P
| i = 1, . . . , r} such that ei = (ei1 , . . . , eis )τ ∈ Zs , hi 6∈ L,
and xhi = sl=1 eil fl ∈ L, i = 1, . . . , r.
Set F = [f1 (0), . . . , fs (0)] ∈ Zn×s .
Compute a basis E ⊂ Zs of the Z-module kerZ (F ) with Algorithm KER.
Set H = ∅.
While E 6= ∅
4.1. Let e = (e1 , . . . , es )τ ∈ E and E = E \ {e}.
4.2. Let h = (e1 f1 + · · · + es fs )/x.
4.3. If grem(h, {f1 , . . . , fs }) 6= 0, then add (h, e) to H.
5. Return H.
1.
2.
3.
4.
We now give the algorithm to compute the x-saturation of a Z[x]-lattice.
Algorithm 2 — SatX(f1 , . . . , fs )
Input: A finite set f = {f1 , . . . , fs } ⊂ Z[x]n .
Output:A set of generators of satx (f1 , . . . , fs ) .
1. Compute the generalized Hermite normal form g of f with Algorithm GHNF.
2. Set H = XFactor(g).
3. If H = ∅, then output g; otherwise set f = Col(g) ∪ {h | (h, f ) ∈ H} and goto step 1.
Note. Col(g) is the set of columns of g.
Example 7.1 Let C be the following generalized Hermite normal form.
−x + 2 1 1
C = [f1 , f2 , f3 ] = 3x + 2 1 2x + 1 .
0
2x x2
2 1 1
In XFactor(C), the kernel of the following matrix [f1 (0), f2 (0), f3 (0)] = 2 1 1 is
0 0 0
τ
τ
generated by e1 = (0, −1, 1) and e2 = (1, −2, 0) . In step 4.2 of XFactor, we have
h = (−f2 + f3 )/x = (0, 2, x − 2)τ . One can check that (0, 2, x − 2)τ 6∈ (C). In SatX,
39
computing the generalized Hermite normal form of
−x + 2 1
C1 = 3x + 2 −3
0
4
C ∪ {(0, 2, x − 2)τ }, we have
0
.
2
x−2
XFactor(C1 ) returns ∅. So, (C1 ) is the x-saturation of (C).
Proposition 7.2 Algorithms SatX and XFactor are correct.
Proof: From the output of Algorithm XFactor, in step 3 of SatX, we have (g) ( (g ∪
{h | (h, f ) ∈ H}) ⊆ satx (f ). Since Z[x]n is a Noetherian Z[x]-module, SatX will terminate
and return the x-saturation of (f ). So, it suffices to show the correctness of Algorithm
XFactor.
We first explain step 4.2 of Algorithm XFactor. Since e ∈ kerZ (F ), h(0) = [f1 (0), . . . ,
fs (0)]e = [0, . . . , 0]τ . Therefore, x is a factor of e1 f1 + · · · + es fs and thus h = (e1 f1 + · · · +
es fs )/x ∈ Z[x]n .
To prove the correctness of Algorithm XFactor, it suffices to show that L = satx (L) if
and only if for each e ∈ E, e1 f1 + · · · + es fs = xh implies h ∈ L.
Let E = {e1 , . . . , ek } where ei ∈ Zs . If L = satx (L), then it is clear that (f1 , . . . , fs )ei =
xhi implies hi ∈ L. To prove the other direction,
Ps let [f1 , . . . , fs ]ei = xhi for 1 ≤ i ≤
k, where hi ∈ L. Let xf ∈ L. P
Then xf =
i=1 ci (x)fi , where ci (x) ∈ Z[x]. If for
s
each i, x|ci (x), then we have P
f =
(c
(x)/x)f
i ∈ L,
i=1 i
Psand the lemma is proved. Oths
erwise, set x = 0 in xf =
i=1 ci (0)fi (0) = 0. Hence Q =
i=1 ci (x)fi , we obtain
τ ∈ ker (F ) and hence there exist a ∈ Z, i = 1, . . . , k such that Q =
[c
(0),
.
.
.
,
c
(0)]
1
s
Z
i
Pk
Pk
Pk
e
e
i=1 ai ei . Then, [f1 , . . . , fs ]Q =
i=1 ai [f1 , . . . , fs ]ei =
i=1 ai xhi = xh, where h =
Pk
i=1 ai hi ∈ L. Then,
xf
Ps
Ps
Ps
(x)fi
=
i=1 ci (x)fi = P
i=1 ci (0)fi +
i=1 xciP
e + x s ci (x)fi ,
= [f1 , . . . , fs ]Q + x si=1 ci (x)fi = xh
i=1
e + Ps ci (x)fi ∈ L and the lemma is
where ci (x) = (ci (x) − ci (0))/x ∈ Z[x]. Hence, f = h
i=1
proved.
7.2
Z-saturation of Z[x]-lattice
The key idea to compute satZ (L) for a Z[x]-lattice L ∈ Z[x]n is as follows. Let f =
{f1 , . . . , fs }. Then (f ) is not Z-saturated
P if and only if a linear combination of fi contains a
nontrivial prime
P factor in Z, that is, i gi fi = pf , where p is a prime number and f 6∈ (f ).
Furthermore, i gi fi = pf with gi 6= 0 mod p is valid if and only if f1 , . . . , fs are linear dependent over Zp [x]. The fact that Zp [x] is a PID allows us to compute such linear relations using
methods of Hermite normal forms [3]. The following algorithm is based on this observation.
40
Algorithm 3 — ZFactor
Input: A generalized Hermite normal form C = [c1 , . . . , cs ] ⊂ Z[x]n of form (6).
Output:∅, if L = (C) is Z-saturated; otherwise, a finite set {(hi , ki , ei ) | i =
1, . . . , r}, such thatPhi ∈ Z[x]n , ki ∈ N, ei = (ei1 , . . . , eis )τ ∈ Z[x]s ,
hi 6∈ L and ki hi = sl=1 eil cl ∈ L for i = 1, . . . , r.
1. Read theQnumbers t, ri , li , cri ,1,0 , i = 1, . . . , t from (6).
2. Set q = ti=1 cri ,1,0 ∈ N.
3. For any prime factor p of q do
3.1. Set F = [cr1 ,l1 , cr2 ,l2 , . . . , crt ,lt ] ∈ Zp [x]n×t .
3.2. Compute a basis G ⊂ Zp [x]s of the Zp [x]-module kerZp [x] (F )
with Algorithm KER.
P
3.3. If G 6= ∅, for each g = [g1 , . . . , gt ]τ ∈ G, let ti=1 gi cri ,li = ph in Z[x]n .
Return the setP
of (h, p, e), where e = [e1 , . . . , es ]τ ∈ Z[x]s such that
esk = gk , sk = ki=1 li , k = 1, . . . , t and ej = 0 for other j.
3.4. Compute the Hermite normal form B = {b1 , . . . , bt }
of {cr1 ,l1 , . . . , crt ,lt } in Zp [x]n with Algorithm HNF.
P
3.5. Let C− = {f1 , . . . , fl } be given in (8) and e
fi = grem(fi , B) = fi + tk=1 ai,k bk ,
in Zp [x]n , where ai,k ∈ Zp [x].
P
3.6. If e
fi = 0 for some i, then fi + tk=1 ai,k bk = phi in Z[x]n .
Return the set of (hiP
, p, ei ) where ei is a vector in Z[x]s such that
(c1 , . . . , cs )ei = fi + tk=1 ai,k bk = phi .
3.7. Set E = [e
f1 , . . . , e
fl ] ∈ Zp [x]n×l .
3.8. Compute a basis D of {X ∈ Zlp | EX = 0} as a vector space over Zp .
P
3.9. If D 6= ∅, for each b = [b1 , . . . , bl ]τ ∈ D, li=1 bie
fi = ph in Z[x]n .
s
Return the set of
Pl(h, p,ee) where e is a vector in Z[x] such that
(c1 , . . . , cs )e = i=1 bi fi = ph.
4. Return ∅.
Remark 7.3 In steps 3.6 and 3.9, we need to compute ei or e. Since B = {b1 , . . . , bt } is
the Hermite normal form of c = {cr1 ,l1 , . . . , crt ,lt } in Zp [x]n , there exists an invertible matrix
Mt×t such that
P [b1 , . . . , bt ] = [cr1 ,l1 , . . . , crt ,lt ]M . In Step 3.6, ei can be obtained from the
relation fi + tk=1 ai,k bk = phi and the relation [b1 , . . . , bt ] = [cr1 ,l1 , . . . , crt ,lt ]M . Step 3.9
can be treated similarly.
Remark 7.4 In step 3.8, we need to compute a basis for the vector space {X ∈ Zlp | EX = 0}
over Zp . We will show how to do this. A matrix F ∈ Zp [x]m×s is said to be in standard form
if F has the structure in (6) and deg(cri ,k1 , x) < deg(cri ,k2 , x) for i = 1, . . . , t and k1 < k2 .
The matrix E ∈ Zp [x]n×l can be transformed into standard form using the following
operations: (1) exchange two columns and (2) add the multiplication of a column by an
element from Zp to another column. Equivalently, there exists an inversive matrix U ∈ Zl×l
p
41
such that E · U = S is in standard form. Suppose that the first k columns of S are zero
vectors. Then the first k columns of U constitute a basis for ker(E). This can be proved
similarly to that of the algorithm to compute a basis for the kernel of a matrix over a PID
[3, page 74].
We now give the algorithm to compute the Z-saturation.
Algorithm 4 — SatZ(f0 , . . . , fs )
Input: A set of vectors f = {f0 , . . . , fs } ⊂ Z[x]n .
Output:A reduced Gröbner basis g such that (g) = satZ (f ).
1. Compute generalized Hermite normal form g of f .
2. Set S =ZFactor(g).
3. If S = ∅, return g; otherwise set f = Col(g) ∪ {h | (h, k, f ) ∈ S} and goto step 1.
Example 7.5 Let C be the following generalized Hermite normal form:
2
x + 2x − 2 x + 2 1
C=
.
0
4
2x
Then, t = 2, r1 = 1, l1 = 1, r2 = 2, l2 = 2, q = 4, c1,1 = [x2 + 2x − 2, 0]τ , c2,1 = [x + 2, 4]τ ,
c2,2 = [1, 2x]τ . Apply algorithm ZFactor to C. We have p = 2. In steps 3.1 and 3.2,
x2 1
F =
and ker(F ) is generated by G = {[−1, x2 ]τ }. In step 3.3, x2 c22 − c11 =
0 0
2(1 − x, x3 )τ and return (1 − x, x3 )τ .
In Algorithm SatZ, (1 − x, x3 )τ is added into C and
2
x + 2x − 2 x + 2 1 1 − x
C1 =
,
0
4
2x x3
which is also a generalized Hermite normal form.
Applying Algorithm
2 ZFactor
to C1 . We have p = 2 and t = 2. In steps 3.1-3.3, G = ∅.
x 1−x
x+2 1 x
In step 3.4, B =
. In step 3.5, C− =
and e
fi 6= 0 for all i.
3
0 x
4
2x 2x2
x 1 x
. In Step 3.8, D = {b}, where b = [1, 0, −1]τ . In Step 3.9,
In step 3.7, E =
0 0 0
(x + 2, 4)τ − (x, 2x2 )τ = 2(1, 2 − x2 )τ . Add (1, 2 − x2 )τ into C1 and compute the generalized
Hermite normal form, we have
2
x + 2x − 2 x + 2 1 −1
C2 =
.
0
4
2x x2 − 2
Apply Algorithm ZFactor again, it is shown that C2 is Z-saturated.
We will prove the correctness of the algorithm. We denote by satp (L) the set {f ∈
Z[x]n | pf ∈ L} where p ∈ Z is a prime number. An infinite set S is said to be linear
P independent over a ring R if any finite set of S is linear independent over R, that is ki=1 ai gi = 0
for ai ∈ R and gi ∈ S implies ai = 0, i = 1, . . . , k.
42
Lemma 7.6 Let C be a generalized Hermite normal form and L = (C). Then satp (L) = L
if and only if C∞ is linear independent over Zp , where C∞ is defined in (8).
Proof: “ ⇒ ” Assume the contrary, that is, C∞ = {h1 , h2 , . . .} P
defined in (8) is linear
dependent P
over Zp . Then there exist ai ∈ Zp not all zero, such that ri=1 ai hi = 0 in Zp [x]n
and hence ri=1 ai hi = pg in Z[x]n . By Lemma 3.11, hi are linear independent over Zp and
hence g 6=P0. Since satp (L) P
= L, we have g ∈ L. By Lemma 3.12, there exist bi ∈ Z such
that g = ri=1 bi hi . Hence ri=1 (ai − pbi )hi = 0 in Z[x]n . By Lemma 3.11, ai = pbi and
hence ai = 0 in Zp [x], a contradiction.
“ ⇐ ” Assume the contrary,
that is, there exists a g ∈ Z[x]n , such that g 6∈ L and pg ∈ L.
Pr
By Lemma
pg = i=1 ai hi , where ai ∈ Z. p cannot be a factor of all
Prai . Otherwise,
Pr a3.12,
i
h
∈
L.
Then
some
of
a
is
not
zero
in
Z
,
which
means
g =
i
p
i=1 ai hi = 0 is
i=1 p i
nontrivial linear relation among Ci over Zp , a contradiction.
From the “ ⇒ ” part of the above proof, we have
P
ai hi = 0 be a nontrivial linear relation among hi in Zp [x]n , where
Corollary 7.7 Let ri=1P
n
ai ∈ Zp . Then, in Z[x] , ri=1 ai hi = ph and h 6∈ (C).
Lemma 7.8 Let C = [c1 , . . . , cs ] be a generalized Hermite normal form and L = (C). Then
satp (L) = L if and only if C∞ is linear independent over Zp for the prime factors of q defined
in step 2 of Algorithm ZFactor.
Proof: By Definition 3.6, the leading monomial of xk cQri ,j ∈ C∞ is of the form cri ,j,0 xk+dri ,j ǫri
and cri ,li ,0 | . . . |cri ,2,0 |cri ,1,0 . If p is coprime with ti=1 cri ,1,0 , then cri ,j,0 6= 0 mod p for
1 ≤ j ≤ li . Therefore, the leading monomials of the elements of C∞ are linear independent
over Zp , and
Qt hence C∞ is linear independent over Zp . Therefore, it suffices to consider prime
factors of i=1 cri ,1,0 .
To check whether C∞ is linear independent over Zp , we first consider a subset of C∞ in
the following lemma.
Lemma 7.9 Let C be the generalized Hermite normal form given in (6). Then C + defined in
(8) is linear independent over Zp if and only if {cr1 ,l1 , cr2 ,l2 , . . . , crt ,lt } are linear independent
over Zp [x].
P P
P
ai,j xj cri ,li =
Proof: This is obvious since
i pi cri ,li , where ai,j ∈ Z and pi =
i
j
P
j
j ai,j x .
Lemma 7.10 Let B be a Hermite normal form in Zp [x]n and g = {g1 , . . . , gr } ⊂ Zp [x]n .
Then g ∪ B∞ is linear dependent over Zp if and only if
ei = grem(gi , B) = 0 in Zp [x]n for some i, or
• either g
• the residue set {grem(gi , B) | i = 1, . . . , r} are linear dependent over Zp .
43
Proof: We may assume that grem(gi , B) = 0 does not happen, since it gives a nontrivial
ei = gi mod B∞ . g ∪ B∞ is linear dependent over
linear relation of g ∪ B∞ . By Lemma 3.12, g
P
Zp if and only if there existPai ∈ Zp not all zero such that i ai gi = 0 mod B∞ over Zp ,
e
which is valid if and only if i ai gei =
P
P0 mod B∞ . Since gi are G-reduced with respect to B,
e
e
e
a
g
=
0
mod
B
if
and
only
if
a
g
=
0,
that
is
g
∞
i are linear dependent over Zp .
i i i
i i i
Proposition 7.11 Algorithm SatZ is correct.
Proof: Since Z[x]n is Notherian, the algorithm terminates and it suffices to show that
Algorithm ZFactor is correct. Let C = [c1 , . . . , cs ]. By Lemma 7.6, to check whether
satZ (c1 , . . . , cs ) is Z-saturated, it suffices to check for any prime p, C∞ is linearQindependent
on Zp . Furthermore, by Lemma 7.8, it suffices to consider prime factors of ti=1 cri ,1,0 in
step 3. This explain why only prime factors of q are considered in Step 3.
In steps 3.1 and 3.2, we check whether C + in (8) is linear independent over Zp . By Lemma
7.9, we need only to consider whether C1 = {cr1 ,l1 , cr2 ,l2 , . . . , crt ,lt } is linear independent over
Zp [x]. It is clear that C1 is linear independent over Zp [x] if and only if G = ∅, where G is
given in step 3.2.
τ
Pt In step 3.3, C1 is linear dependent
Pt over Zp . If G 6= ∅ for any g n= [g1 , . . . , gt ] ∈ G,
i=1 gi cri ,li = 0 in Zp [x]. Hence
i=1 gi cri ,li = ph where h ∈ Z[x] . By Corollary 7.7,
h 6∈ L. The correctness of Algorithm ZFactor is proved in this case.
In steps 3.4 - 3.10, we handle the case where C + is linear independent over Zp . In step
3.4, we compute the Hermite normal form of C1 in Zp [x]n , which is possible because Zp [x]n
is a PID [3]. Furthermore, we have [3]
[cr1 ,l1 , . . . , crt ,lt ]N = [b1 , . . . , bt ]
where {b1 , . . . , bt } is a Hermite normal form and N is an inversive matrix in Zp [x]t×t . Then
C∞ = C− ∪ C + is linear independent over Zp if and only if
j
j
Ce = C− ∪ B∞ = C− ∪ ∪∞
j=0 {x b1 , . . . , x bt } is linear independent over Zp .
(39)
By Lemma 7.10, property (39) is valid if and only if grem(c, B) 6= 0 for all c ∈ C− and the
residue set Ce− is linear independent over Zp , which are considered in step 3.7 and steps 3.83.10, respectively. Then we either prove L is Z-saturated or find a nontrivial linear relation
for elements in C∞ over Zp . By Corollary 7.7, such a relation leads to an h ∈ satZ (L) \ L.
The correctness of the algorithm is proved.
As a direct consequence of Lemma 5.20 and Algorithm ZFactor, we have the algorithm
to compute the M -saturation.
44
Algorithm 5 — SatM(f0 , . . . , fs )
Input: A set of vectors f = {f0 , . . . , fs } ⊂ Z[x]n .
Output:A generalized Hermite normal form g such that satM (f ) = (g).
1. Using Algorithm ZFactor, we can compute mi ∈ N and gi ∈ Z[x]n , i = 1, . . . , s
such that satZ (f ) = (g1 , . . . , gs ) and mi gi ∈ (f ).
2. Let S = ∅ and for i = 1, . . . , s, if mi 6= 1 then S = S ∪ {(x − omi )gi }.
3. Compute the generalized Hermite normal form g of f ∪ S and return g.
Notices that if mi = 1 then omi = 0 and gi ∈ (f ). The numbers mi need not to be
unique for the following reasons. Suppose mi = ni k and ni gi ∈ (f ). Then by Corollary 5.15,
omi = oni + cni and hence (x − oni )gi = (x − omi )gi + cni gi ∈ (f ). That is, we can replace
mi by its factor ni .
Proposition 7.12 Algorithm SatM is correct.
Proof: Let L1 = (f ) and L2 = (f , (x − om1 )g1 , . . . , (x − oms )gs ). We claim that satZ (L1 ) =
satZ (L2 ). Since L1 ⊂ L2 , satZ (L1 ) ⊂ satZ (L2 ). Since satZ (L1 ) = (g1 , . . . , gs ), we have L2 ⊂
satZ (L1 ) and hence satZ (L2 ) ⊂ satZ (L1 ). The claim is proved. Then satZ (L2 ) = (g1 , . . . , gs )
and mi gi ∈ L1 ⊂ L2 . Since (x − omi )gi ∈ L2 , i = 1, . . . , s, L2 is M -saturated by Lemma
5.20.
7.3
Algorithms for Laurent binomial and binomial σ-ideals
In this section, we will present several algorithms for Laurent binomial and binomial σ-ideals,
and in particular a decomposition algorithm for binomial σ-ideals. We first give an algorithm
to compute the characteristic set for a Laurent binomial σ-ideal.
Algorithm 6 — CharSet
Input: F : a finite set of Laurent σ-binomials in F{Y± }.
Output:∅, if [F ] = [1]; otherwise, a regular and coherent Laurent binomial σchain A such that [A] = [F ] and A is a characteristic set of the σ-ideal
[F ].
1.
2.
3.
4.
5.
6.
Let F = {Yf1 − c1 , . . . , Yfr − cr } and f = {f1 , . . . , fr }.
Compute a set of generators H ⊂ Z[x]r of kerZ[x] ([f1 , . . . , fr ]) with Algorithm GKER.
Q
If for any h = (h1 , . . . , hr )τ ∈ H, ri=1 chi i 6= 1, then return ∅.
Compute the reduced GröbnerPbasis g of f with Algorithm GHNF.
Let g = {g1 , . . . gs } and gi = rk=1 ai,k fk , where ai,k ∈ Q
Z[x], i = 1, . . . , s.
a
g
i
Return A = {g1 , . . . , gs }, where gi = Y − di and di = rk=1 cki,k , i = 1, . . . , s.
Proposition 7.13 Algorithm CharSet is correct.
45
Proof: Steps 1-3 uses Proposition 4.4 to check whether [F ] = [1]. Note that (f ) and (g)
are the support lattices of the binomial σ-ideals [F ] and [G], respectively. By Corollary 4.3,
[F ] = [G]. By Theorem 4.14, A is a regular and coherent σ-chain and hence a characteristic
set of [A].
We now show how to compute the reflexive closure for a Laurent binomial σ-ideal.
Algorithm 7 — Reflexive
Input: P : a finite set of Laurent σ-binomials in F{Y± }, where F is inversive.
Output:A: a regular and coherent Laurent binomial σ-chain such that [A] is the
reflexive closure of [P ].
1. Let G = CharSet(P ). If G = ∅, return 1.
2. Let G = {g1 , . . . , gs }, gi = Ygi − di , and g = [g1 , . . . , gs ] ∈ Z[x]n×s .
3. H =XFactor(g).
4. If H = ∅, return G.
5. Let H = {(hi , ei ) | i = 1, . .Q
. , r} and ei = (ei1 , . . . , eis )τ .
e
6. Let P := G ∪ {Yhi − σ −1 ( sj=1 dj i,j ), i = 1, . . . , r}, and go to step 1.
Proposition 7.14 Algorithm Reflexive is correct.
Proof: The algorithm basically follows the proof of Theorem 5.5. By CharSet, [P ] = [G]
and G is a regular and coherent σ-chain. In step 4, if H = ∅, then (g) is x-saturated, and by
Theorem 5.3, [G] is reflexive and the theorem is proved. Otherwise, we execute steps 5 and
Q
b
6. Let I1 = [P ], L1 = (f ), I2 = [G ∪ {Yhi − σ −ki ( sj=1 dji,j ), i = 1, . . . , r}], and L2 = L(I2 ).
Then, we have I1 I2 ⊂ Ix and L1 L2 ⊂ Lx , where Lx = satx (L1 ) and Ix is the reflexive
closure of I1 . Similar to the proof of Theorem 5.5, the algorithm will terminate and output
the reflexive closure of [P ].
Remark 7.15 Similar to Algorithm Reflexive, we can give algorithms to check whether
a Laurent binomial σ-ideal I is well-mixed, perfect, or prime, and in the negative case to
compute the well-mixed or perfect closures of I. The details are omitted.
We give a decomposition algorithm for perfect Laurent binomial σ-ideals.
46
Algorithm 8 — DecLaurent
Input: P : a finite set of Laurent σ-binomials in F{Y± }, where F is inversive
and algebraically closed.
Output:∅, if {P } = [1] or regular and coherent binomial σ-chains C1 , . . . , Ct such
that [Ci ] are Laurent reflexive prime σ-ideals and {P } = ∩ti=1 [Ci ] is a
minimal decomposition.
1. Let F =Reflexive(P ). If F = 1, return ∅.
2. Set R = ∅ and F = {F }.
3. While F 6= ∅.
3.1. Let F = {Yf1 − c1 , . . . , Yfr − cr } ∈ F, F = F \ {F }.
3.2. Let G = CharSet(F ). If G = ∅, goto step 3.
3.3. Let G = {g1 , . . . , gs }, gi = Ygi − di , and g = [g1 , . . . , gs ] ∈ Z[x]n×s .
3.4. H =ZFACTOR(g).
3.5. If H = ∅, add G to R.
τ
3.6. Let H = {(hi , ki , ei ) | i = 1, . . . , r} and ei = (ei1 , . . . , eQ
is ) .
e
s
3.7. For i = 1, . . . , t, let ri,1 , . . . , ri,ki be the ki -th roots of j=1 dj i,j .
3.8. For l1 = 1, . . . , k1 ,. . . ,lt = 1, . . . , kt , add G ∪ {Yh1 − r1,l1 , . . . , Yht − rt,lt } to F.
4. Return R.
k
k
Example 7.16 Let P = {g1 , g2 , g3 }, where g1 = y1−2 y1x −1, g2 = y2−2 y2x −1, g3 = y1 y2−x y32 −
1, and k ≥ 2. [P ] is already reflexive, so step 1 does nothing. P is already a regular and
coherent chain, so G = P . Let g1 = [xk − 2, 0, 0]τ , g2 = [0, xk − 2, 0]τ , g3 = [1, −x, 2]τ be
the supports of g1 , g2 , and g3 , respectively. Then [g1 , g2 , g3 ] is already a generalized Hermite
normal form.
In Steps 3.4-3.6, H = {(h1 , k1 , e1 )}, where h1 = [1, −x, xk ]τ , k1 = 2, e1 = [−1, x, xk ]τ .
k
k
In Step 3.7, r1,1 = 1 and r1,2 = −1. In Step 3.8, P1 = y1 y2−x y3x − 1 and P2 = y1 y2−x y3x + 1
are added to G to obtain C1 = {g1 , g2 , g3 , P1 } and C2 = {g1 , g2 , g3 , P2 }. Both [C1 ] and [C2 ] are
reflexive and prime, and are returned.
To see why the algorithm is correct, from e1 = [−1, x, xk ], we have (g1 +1)−1 (g2 +1)x (g3 +
k
k
= y12 y2−2x y32x = 1 mod[P ]. Hence, P1 P2 = y12 y2−2x y32x − 1 ∈ [P ].
k
1)x
Proposition 7.17 Algorithm DecLaurent is correct.
Proof: The algorithm basically follows the proof of Theorem 5.7. The proof is similar to that
of Theorem 7.14. By the proof of Theorem 5.7, we obtain a minimal decomposition.
In the rest of this section, we give a decomposition algorithm for binomial σ-ideals. Before
giving the main algorithm, we give a sub-algorithm DecMono which treats the σ-monomials.
Basically, it gives the following decomposition
n
Y
V( yifi ) = V(y1 ) ∪ V(y2 /y1 ) ∪ · · · ∪ V(yn /{y1 , . . . , yn−1 })
i=1
where 0 6= fi ∈ N[x] and V(yc /S) is the set of zeros of yc = 0 not vanishing any of the
variables in S. The correctness of the algorithm comes directly from the above formula.
47
Algorithm 9 — DecMono
Input: (Y0 , B, Y1 ): Y0 , Y1 are disjoint subsets of Y and B a finite set of σbinomials or σ-monomials in F{Y}.
Output:(Y0i , Bi , Y1i ): Y0i , Y1i are disjoint subsets of Y, Bi consists of proper
σ-binomials, and V(Y0 ∪ B/Y1 ) = ∪ri=1 V(Y0i ∪ Bi /Y1i ).
1. Set R = ∅ and F = {(Y0 , B, Y1 )}.
2. While F 6= ∅.
2.1. Let C = (Y0 , B, Y1 ) ∈ F, F = F \ {C}.
k
2.2. For all yc ∈ Y0 , let B1 = Byc =0 (replace ycx by 0) and delete 0 from B1 .
2.3. If B1 contains no σ-monomials, add (Y0 , B1 , Y1 ) to R and goto step 2.
Q
2.4. Let M = ki=1 ycfii ∈ B1 , where 0 6= fi ∈ N[x]. B1 = B1 \ {M }.
2.5. Let Y2 := {yc1 , . . . , yck } \ Y1 . If Y2 = ∅, go to step 2; else let Y2 = {yt1 , . . . , yts }.
2.6. For i = 1, . . . , s, add (Y0 ∪ {yti }, B1 , Y1 ∪ {yt1 , . . . , yti−1 }) to F.
3. Return R.
We now give the main algorithm. The algorithm basically follows the proof of Theorem
6.27.
main modification is that instead of the perfect ideal decomposition {F } = ({F } :
T The
n
m) ∩i=1 {F, yi }, we use the following zero decomposition
[
V(F ) = V({F } : m) ∪ni=1 V(F ∪ {yi }/{y1 , . . . , yi−1 }).
The purpose of using the later decomposition is that many redundant components can be
easily removed by the following criterion V(F/D) = ∅ if F ∩ D 6= ∅, which is done in step
2.5 of Algorithm DecMono.
Algorithm 10 — DecBinomial
Input: F : a finite set of σ-binomials in F{Y}.
Output:∅, if {F } = [1] or (C1 , Y1 ), . . . , (Cr , Yr ), where Yi ⊂ Y and Ci are regular
and coherent σ-chains containing σ-binomials or variables in Y \ Yi such
that sat(Ci ) are reflexive prime σ-ideals and {F } = ∩ri=1 sat(Ci ).
1. Set R = ∅ and F =DecMono(∅, F, ∅).
2. While F 6= ∅.
2.1. Let C = (Z0 , B, Z1 ) ∈ F, F = F \ {C}.
2.2. If B = ∅, add (Z0 , Z1 ) to R.
2.3. Let E = DecLaurent(B) in F{Z± }, where Z = Y \ Z0 and m = |Z|.
2.4. If E = ∅ goto step2.
2.5. Let E = {E1 , . . . , El } and El = {Zfl,1 − cl,1 , . . . , Zfl,sl − cl,sl }, where fl,k ∈ Z[x]m .
+
−
f+
f−
2.7. Add ({Z0 , Zfl,1 − cl,1 Zfl,1 , . . . , Z l,sl − cl,sl Z l,sl }, Z1 ) to R, l = 1, . . . , k.
2.8. Let Z = {yc1 , . . . , ycs }. For i = 1, . . . , s, do
F = F∪ DecMono(Z0 ∪ {yci }, B, Z1 ∪ {yc1 , . . . , yci−1 }).
3. Return R.
48
k
k
Example 7.18 Let A = {A1 , A2 , A3 }, where A1 = y1x − y12 , A2 = y2x − y22 , A3 = y1 y32 − y2x ,
and k ≥ 2. In Step 1, we have F = {(∅, A, ∅)}. From Example 7.16, in Step 2.3, we have
E = {C1 , C2 }, where C1 and C2 are given in Example 7.16. In Step 2.7, (E1 , ∅) and (E2 , ∅) are
k
added to R, where E1 = {A1 , A2 , A3 , Q1 }, E2 = {A1 , A2 , A3 , Q2 }, and Q1 = y1 y3x −y2x , Q2 =
k
y1 y3x +y2x . In Step 2.8, Z = {y1 , y2 , y3 } and ({y1 , y2 }, ∅, ∅), ({y2 , y3 ), {A1 }, {y1 }) are added to
F. Finally, we have the following decomposition {A} = [y1 , y2 ]∩[A1 , y2 , y3 ]∩sat(C1 )∩sat(C2 ),
where all components are reflexive and prime.
Remark 7.19 Using the algorithm in [9], the following decomposition is obtained: {A} =
[y1 , y2 ] ∩ sat(A). From Example 7.16, sat(A) is not prime. Then, Algorithm DecBinomial
is stronger than the general decomposition algorithm given in [9].
Proposition 7.20 Algorithm DecBinomial is correct.
Proof: In step 1, σ-monomials in F are treated. In step 2, we will treat the components of
F one by one. In step 2.1, the component (Z0 , B, Z1 ) is taken from F. In step 2.3, {B} is
decomposed as {B} = ∩kl=1 [El ] in F{Z± }, where El are regular and coherent σ-chains and
[El ] are reflexive prime ideals. By (30) and Corollary 6.30, we have
{B} : m = {B} ∩ F{Z} = ∩kl=1 ([El ]F{Z± }) ∩ F{Y} = ∩kl=1 sat(El+ ),
+
−
f+
(40)
f−
where El+ = {Zfl,1 − cl,1 Zfl,1 , . . . , Z l,sl − cl,sl Z l,sl }, l = 1, . . . , k. Since El is regular and
coherent, by Lemma 6.34, El+ is also regular and coherent. Since [El ] is reflexive and prime,
by Corollary 6.31, sat(El+ ) is also reflexive and prime. Note that E = ∅ in step 2.4 if and
only if [B] contains a σ-monomial.
Since B ⊂ F{Z}, we have the following decomposition
V(Z0 ∪ B/Z1 ) = V(Z0 ∪ ({B} : m)/Z1) ∪si=1 V(Z0 ∪ B ∪ {yci }/{yc1 , . . . , yci−1 } ∪ Z1 ),
where V(Z0 ∪ B ∪ {yci }/{yc1 , . . . , yci−1 } ∪ Z1 ) is further simplified with Algorithm DecMono
in step 2.8. From (40),
V(Z0 ∪ {B} : m/Z1 ) = ∪kl=1 V(sat(Z0 , El+ )/Z1 ) = ∪kl=1 V([Z0 , sat(El+ )]/Z1 ),
where {Z0 , El+ } is a regular and coherent σ-chain since El+ does not contain variables in Z0 .
The above formula explains why ({Z0 , El+ }, Z1 ) is added to R in steps 2.5-2.7.
Let the algorithm returns R = {(Ci , Yi ); i = 1, . . . , m}. From the above proof, we have
V(F ) = ∪kl=1 V(sat(Ci )/Yi ). Since Yi ∩ Ci = ∅ and sat(Ci ) is a reflexive and prime σ-ideal,
the Cohn closure of V(sat(Ci )/Yi ) is V(sat(Ci )) and hence
V(F ) = ∪kl=1 V(sat(Ci )/Yi ) = ∪kl=1 V(sat(Ci )).
By the difference Hilbert Nullstellensatz, {F } = ∩kl=1 {sat(Ci )} = ∩kl=1 sat(Ci ). The algorithm
terminates, since after each execution of step 2, in the new components (Y0l , Bl , Y1l ) added
to F in step 2.8, Bl contains at least one less variables than B.
49
8
Conclusion
In this paper, we initiate the study of binomial σ-ideals. Two basic tools used to study
binomial σ-ideals are the Z[x]-lattice and the characteristic set instead of the Z-lattice and
the Gröbner basis used in the algebraic case [7].
For Laurent binomial σ-ideals, two main results are proved. Canonical representations
for proper Laurent binomial σ-ideals are given in terms of Gröbner basis of Z[x]-lattices,
regular and coherent σ-chains in F{Y± }, and partial characters over Z[x]n , respectively. It
is also shown that a Laurent binomial σ-ideal is radical and dimensionally un-mixed. We also
give criteria for a Laurent binomial σ-ideal to be reflexive, well-mixed, perfect, and prime in
terms of its support lattice. It is shown that the reflexive, well-mixed, and perfect closures
of a Laurent binomial σ-ideal I is still binomial whose support lattices are the x-, M-, and
P-saturation of the support lattice of I .
For binomial σ-ideals, we show that certain properties of algebraic binomial ideals given
in [7] can be extended to the difference case using the theory of Gröbner basis in the case of
infinitely many variables. It is shown that most properties of Laurent binomial σ-ideal can
be extended to the normal binomial σ-ideals.
Algorithms are given for the key results of the paper. We give algorithms to check whether
a given Laurent binomial difference ideal I is reflexive, prime, well-mixed, or perfect, and
in the negative case, to compute the reflexive, well-mixed, and perfect closures of I. An
algorithm is given to decompose a finitely generated perfect binomial difference ideal as the
intersection of reflexive prime binomial difference ideals.
Finally, we make a remark about differential binomial ideals. The study of binomial
differential ideals is more difficult, because the differentiation of a binomial is generally
not a binomial. Differential toric varieties were defined in [18] and were used to connect
the differential Chow form [8] and differential sparse resultant [18]. But, contrary to the
difference case, the defining ideal for a differential toric variety is generally not binomial.
References
[1] M. Barile, M. Morales, A. Thoma. Set-theoretic Complete Intersections on Binomials,
Proc. AMS, 130(7), 1893-1903, 2001.
[2] D. Bouziane, A. Kandri Rody, H. Maârouf. Unmixed-dimensional Decomposition of a
Finitely Generated Perfect Differential Ideal, Journal of Symbolic Computation, 31(6),
631-649, 2001.
[3] H. Cohen. A Course in Computational Algebraic Number Theory, Springer-Verlag,
Berlin, 1993.
[4] R. M. Cohn. Difference Algebra, Interscience Publishers, New York, 1965.
[5] D. Cox, J. Little, D. O’Shea. Using Algebraic Geometry, Springer-Verlag, New York,
1998.
50
[6] A. Dickenstein, L. F. Matusevich, E. Miller. Combinatorics of Binomial Primary Decomposition, Mathematische Zeitschrift, 264(4), 745-763, 2010.
[7] D. Eisenbud and B. Sturmfels. Binomial Ideals, Duke Math. J., 84(1), 1-45, 1996.
[8] X. S. Gao, W. Li, C. M. Yuan. Intersection Theory in Differential Algebraic Geometry:
Generic Intersections and the Differential Chow Form, Trans. of Amer. Math. Soc.,
365(9), 4575-4632, 2013.
[9] X. S. Gao, Y. Luo, C. M. Yuan. A Characteristic Set Method for Ordinary Difference
Polynomial Systems, Journal of Symbolic Computation, 44(3), 242-260, 2009.
[10] X.S. Gao, J. Van der Hoeven, C.M. Yuan, G.L. Zhang, Characteristic Set Method for
Differential-difference Polynomial Systems, Journal of Symbolic Computation, 44(9),
1137-1163, 2009.
[11] K. O. Geddes, S. R. Czapor, G. Labahn. Algorithms for Computer Algebra, Kluwer
Academic Publishers, Boston, MA, 1992.
[12] V. P. Gerdt and D. Robertz. Computation of Difference Gröbner Bases, Computer
Science Journal of Moldova, 20(2), 203-226, 2012.
[13] E. Hrushovski. The Elementary Theory of the Frobenius Automorphisms, Available
from http://www.ma.huji.ac.il/˜ ehud/, July, 2012.
[14] K. I. Iima and Y. Yoshino. Gröbner Bases for the Polynomial Ring with Infinite Variables
and Their Applications, Communications in Algebra, 37:10, 3424-3437, 2009.
[15] R. J. Jing, C. M. Yuan, X. S. Gao. A Polynomial-time Algorithm to Compute Generalized Hermite Normal Form of Matrices over Z[x], arXiv:1601.01067, 2016.
[16] U. Koppenhagen and E. W. Mayr. An Optimal Algorithm for Constructing the Reduced
Gröbner Basis of Binomial Ideals, Proc. ISSAC’96, ACM Press, 55-62, 1996.
[17] A. Levin. Difference Algebra, Springer-Verlag, New Work, 2008.
[18] W. Li, C. M. Yuan, X. S. Gao. Sparse Differential Resultant for Laurent Differential
Polynomials, Foundations of Computational Mathematics, 15(2), 451-517, 2015.
[19] I. O. Mart´inez de Castilla, R. P. Sánchez. Cellular Binomial Ideals. Primary Decomposition of Binomial Ideals, Journal of Symbolic Computation, 30(4), 383-400, 2000.
[20] M. P. Millán, A. Dickenstein, A. Shiu, C. Conradi. Chemical Reaction Systems with
Toric Steady States, Bull. Math. Biol., 74, 1027-1065, 2012.
[21] E. Miller and B. Sturmfels. Combinatorial Commutative Algebra, Springer-Verlag, New
York, 2005.
[22] L. Pachter and B. Sturmfels (eds). Algebraic Statistics for Computational Biology, Cambridge University Press, 2005.
51
[23] I. Peeva and B. Sturmfels. Syzygies of Codimension 2 Lattice Ideals, Mathematische
Zeitschrift, 229(1), 163-194, 1998.
[24] J. F. Ritt and J. L. Doob. Systems of Algebraic Difference Equations, American Journal
of Mathematics, 55(1), 505-514, 1933.
[25] M. Saleemi and K. H. Zimmermann. Linear Codes as Binomial Ideals, International
Journal of Pure and Applied Mathematics, 61(2), 147-156, 2010.
[26] M. Wibmer. Algebraic Difference Equations, Preprint, 2013.
52
| 0math.AC
|
arXiv:1612.00277v1 [cs.PL] 1 Dec 2016
NSAD 2016
Sparsity Preserving Algorithms for Octagons
Jacques-Henri Jourdan
MPI-SWS, Inria Paris
Abstract
Known algorithms for manipulating octagons do not preserve their sparsity, leading typically to quadratic or
cubic time and space complexities even if no relation among variables is known when they are all bounded.
In this paper, we present new algorithms, which use and return octagons represented as weakly closed
difference bound matrices, preserve the sparsity of their input and have better performance in the case their
inputs are sparse. We prove that these algorithms are as precise as the known ones.
1
Introduction
In order to capture numerical properties of programs, static analyzers use numerical
abstract domains. The choice of a numerical abstract domain in a static analyzer is
a compromise between precision, the ability of capturing complex numerical properties, and performance. Non-relational abstract domains, such as intervals [6],
are very efficient but relatively imprecise: they cannot represent relations between
program variables. On the other hand, in order to capture numerical relations between program variables, one can express them as linear inequalities. This class of
relational numerical abstract domain is composed of linear abstract domains. A linear abstract domain corresponds to a different precision vs. performance trade-off:
they range from the less precise, efficient ones such as zones [13], pentagons [12] or
octagons [13,14] to the more precise, costly ones, such as subpolyhedra [11], octahedra [5], two variables per inequalities [16], zonotopes [15] or general polyhedra [8].
In particular, the Octagon abstract domain [13,14] accurately represents many of
the variable relationships appearing in a program, while being still reasonably fast
(all the operations have quadratic or cubic complexity on the number of variables).
It is very popular in the static analysis community, which explains why algorithmic
improvements [3,1,17] and precision improving variants [4] are regularly published.
As reported by the designers of Astrée [7], its quadratic or cubic performances
still make it unusable as-is with a reasonable number of variables. Indeed, the
1
This work was supported by Agence Nationale de la Recherche, grant ANR-11-INSE-003.
This paper is electronically published in
Electronic Notes in Theoretical Computer Science
URL: www.elsevier.com/locate/entcs
Jourdan
data structures typically used to represent octagonal abstract values, i.e., strongly
closed difference bound matrices, have a quadratic size in the number of variables for
which an upper or lower bound is known. A common solution is the use of variable
packing [13, §8.4.2], where the Octagon abstract domain is only used on small packs
of variables. The downside of packing is that no relation is stored between variables
that are not in the same pack. A variant of packing has been introduced to mitigate
the imprecision [2], but loss in precision can still occur.
The problem of the performance of octagons has already been studied: in particular, Singh et al. [17] proposed an implementation of the Octagon abstract domain
optimized in the case its representation is sparse. But they do not address the fact
that it is dense as soon as interval bounds are known for many variables, and we
anticipate that, for this reason, the sparsity is very low in their implementation.
Instead, in this paper, we propose to use new algorithms for the Octagon abstract
domain: these algorithms work on a sparse representation for octagons, so that the
cost of the analysis of two independent sets of variables is the sum of the costs
of the analyses of the two sets of variables, taken independently. Our algorithms
have the same precision as the traditional ones. Our main idea is the following:
in order to ensure an optimal precision of all the operations, the data structures
representing octagons, difference bound matrices, are usually kept strongly closed:
that is, algorithms make sure that any returned difference bound matrix is a best
abstraction. However, most often, strongly closed difference bound matrices are
dense because of the necessary strengthening step. In this paper, we propose to
weaken the maintained invariant on difference bound matrices and to keep them
weakly closed hence skipping the strengthening step. Weakly closed difference bound
matrices are not necessarily dense, so that we can use sparse data structures to
represent them. We prove that some algorithms can be kept unchanged to work on
weakly closed difference bound matrices without losing any precision and give new
algorithms for the other operations.
We begin by preliminary definitions in §2. In §3, we describe and prove the
soundness and relative precision of our new algorithms. We conclude in §4.
2
Definitions
Let V+ be a finite set of variables. We call a regular environment a function from V+
to R. A regular environment represents the numerical state of a program. The role
of the Octagon abstract domain is to approximate sets of regular environments ρ. To
that end, the abstract domain of octagons stores a set of inequalities of the following
form:
±ρ(u) ± ρ(v) ≤ Cstuv
u, v ∈ V+
(1)
This corresponds to giving bounds to sums and differences of values of ρ. Moreover, if we use twice the same variable with the same sign, we see that, using such
constraints, we can express interval constraints over values of an environment [13].
In order to handle in a unified way all the different combinations of signs in
these constraints, we introduce the set V± of signed variables. Signed variables are
of two kinds: they are either usual variables from V+ , called positive variables in the
context of signed variables, or their opposites form, negative variables. We equip
2
Jourdan
V± with an involutive operator, associating to each signed variable v its opposite v,
such that v is positive if and only if v is negative.
Regular environments are canonically extended to signed variables by taking
ρ(u) = −ρ(u). More generally, we define irregular environments as functions σ from
V± to R, and consider the set of regular environments as a subset of the set of
irregular environments. Regular environments are exactly irregular environments ρ
that satisfy the property ∀v, ρ(v) = −ρ(v).
Using this new formalism, octagonal constrains of the form (1) can be seen as
upper bounds on differences of values of ρ, a regular environment:
ρ(u) − ρ(v) ≤ Cstuv
u, v ∈ V±
(2)
This has two benefits: first, all the different kinds of constraints allowed by (1) get
factored out as one simpler form. Second, we can see these constraints as constraints
on irregular environments, and further constrain them as being regular: we see that
the study of the Octagon abstract domain starts by the study of a simpler abstract
domain, where only differences of variables are bounded. The set of constraints,
called potential constraints, of such an abstract domain is well studied in the linear optimization literature, because it corresponds to the well-known shortest path
problem in a weighted directed graph.
Such a set of constraints is represented as difference bound matrices: a difference
bound matrix, or DBM, is a matrix (Buv )(u,v)∈V±2 of elements of R ∪ {+∞}. The
meaning of these constraints is given by two concretization functions γpot and γoct ,
that associate to a DBM the set of irregular or regular environments, respectively,
satisfying all the constraints:
γpot (B) = {σ : V± → R | ∀uv ∈ V± , σ(u) − σ(v) ≤ Buv }
γoct (B) = {ρ ∈ γpot (B) | ∀u ∈ V± , ρ(u) = −ρ(u)}
(3)
(4)
Example 2.1 Consider V+ = {x; y; z} a set of three (positive) variables. The set
of signed variables is V± = {x; x; y; y; z; z}. Let A be the DBM such that Axx = 1,
Ayy = 3, Ayz = 1 and Auv = +∞ for all the other entries. The set γoct (A) contains
all the environments ρ : V± → R such that:
•
∀u ∈ V± , ρ(u) = −ρ(u)
•
ρ(x) ≤ 1/2, −ρ(y) ≤ 3/2 and ρ(y) + ρ(z) ≤ 1
This concretization is assimilated to the set of environments ρ : V+ → R over positive
variables such that ρ(x) ≤ 1/2, −ρ(y) ≤ 3/2 and ρ(y) + ρ(z) ≤ 1.
We denote as ♯ ≤ the natural order relation over DBMs, defined as follows:
A ♯ ≤ B ⇔ ∀uv ∈ V± , Auv ≤ Buv
(5)
The following easy lemma states that this order relation makes γpot and γoct increasing, which makes ♯ ≤ a good candidate for a comparison operator of the Octagon
abstract domain 2 :
2 As we see in §3.1, this is not the order relation we use as a comparison operator in our implementation
of the Octagon abstract domain.
3
Jourdan
Lemma 2.2 Let A and B be two DBMs such that A ♯ ≤ B. Then, we have:
γpot (A) ⊆ γpot (B)
γoct (A) ⊆ γoct (B)
For any non-empty set S of irregular environments, there exists a minimal (in the
sense of ♯ ≤) DBM that approximates it. That is, there exists a minimal DBM α(S)
such that S ⊆ γpot (α(S)). This property follows immediately from the definition
of α:
α(S)uv = sup{σ(u) − σ(v)}
(6)
σ∈S
This function α is called the abstraction function. We can easily see that α is an
increasing function. Moreover, α does not only return best abstractions for γpot , but
also for γoct : if the set S contains only regular environments, we can see that α(S)
is also the minimal DBM such that S ⊆ γoct (α(S)). In fact, it is easy to see that
♯ ≤ defines a complete lattice over DBMs extended with a bottom element, and that
the pairs (α, γpot ) and (α, γoct ) form Galois connections.
2.1
Closure and Strong Closure
Many DBMs have the same concretization. This is a problem, because the abstract
environments that we manipulate are therefore not necessarily the most precise ones,
and this can lead to imprecision. Thus, usually, an implementation of the Octagon
abstract domain maintains the invariant that it only manipulates “canonical” forms
of DBMs, such that B = α(γoct (B)). Such “canonical” DBMs are always the best
possible representative over all the DBMs with the same concretization.
An important fact is that we can characterize best abstractions using the values
they contain, and that we have algorithms to compute them. We expose these characterizations, together with these algorithms. Moreover, we give a weaker closedness
condition over DBMs, that does not ensure canonicity, but that allows better algorithms without loss of precision.
2.1.1 Best abstractions for γpot
A first step is to remark that canonical DBMs always have null diagonal values.
Moreover, canonical DBMs should always verify the triangular inequality. We call
such DBMs closed DBMs:
Definition 2.3 [Closed DBM] A closed DBM is a DBM B verifying the two following properties:
•
∀v ∈ V± , Bvv = 0
•
∀uvw ∈ V± , Buw ≤ Buv + Bvw
Closed DBMs are exactly best abstractions for γpot [13, Theorem 3.3.6]. Hence,
closed DBMs always have non-empty concretizations. We do not detail here the
algorithm used to detect the emptiness of the concretization of a DBM and to
compute closures: instead, we refer the interested reader to previous work [13,1].
Example 2.4 The closure α(γpot (A)) of the DBM A as defined in Example 2.1
contains the following additional finite entries:
4
Jourdan
•
∀u ∈ V± , α(γpot (A))uu = 0
•
α(γpot (A))y z = 4 (corresponding to the constraint ρ(z) − ρ(y) ≤ 4).
2.1.2 Best abstractions for γoct
We now refine the notion of closure to canonical forms for γoct . It is easy to see
that, for any non-empty set S of regular environments, α(S)uv = α(S)v u . Thus,
canonical DBMs for γoct will verify the coherence property:
Definition 2.5 [Coherent DBM] A DBM B is coherent when:
∀uv ∈ V± , Buv = Bv u
Moreover, matrix elements of the form Buu (for u ∈ V± ) impose interval constraints on values of ρ. These interval constraints can be combined to entail constraints on any difference of values of ρ. For this reason, canonical forms for γoct
will verify the following strong closedness property:
Definition 2.6 [Strongly closed DBM] A DBM B is strongly closed when it is closed
and coherent and:
Buu + Bvv
∀uv ∈ V± , Buv ≤
2
This condition is necessary and sufficient: strong closedness characterizes canonical DBMs for γoct .
Theorem 2.7 Let B be a DBM. The two following properties are equivalent:
(i) B is strongly closed
(ii) γoct (B) 6= ∅ and B = α(γoct (B))
Proof. See, e.g., [13, Theorems 4.3.2 and 4.3.3].
✷
Usually [1], to compute strong closure, one first ensures that the given matrix
is coherent, then computes a closure (i.e., a canonical representative in the sense of
γpot ), and, finally, performing a so-called strengthening step 3 :
Definition 2.8 [Strengthening] Let B be a DBM. The strengthening of B, noted
♯S(B) is defined by:
Buu + Bvv
♯
S(B)uv = min
; Buv
2
The following theorem states the correctness of the strong closure algorithm
sketched above, consisting in computing a closure followed by a strengthening:
Theorem 2.9 Let B be a coherent DBM with γoct (B) 6= ∅. Then:
α(γoct (B)) = ♯S(α(γpot (B)))
In particular, if B is coherent and closed, then ♯S(B) is strongly closed.
Proof. See, e.g., [10, Theorem 8.2.7].
3
✷
This is actually an improvement of the method described initially by Miné [13].
5
Jourdan
Example 2.10 In order to consider the strong closure of the DBM A as defined in
Example 2.1, we first need to make it coherent: let à be the DBM containing the
same entries as A, except that Ãzy = 1.
The closure of à contains the following addional finite entries:
•
∀u ∈ V± , α(γpot (Ã))uu = 0
•
α(γpot (Ã))zy = α(γpot (A))y z = 4 (corresponding to the constraint ρ(z)−ρ(y) ≤ 4)
•
α(γpot (Ã))zz = 5 (corresponding to the constraint ρ(z) ≤ 5/2).
The strong closure α(γoct (Ã)) is then obtained by strengthening α(γpot (Ã)). The
strengthening operation creates the following new entries:
•
α(γoct (Ã))xy = α(γoct (Ã))y x = 2
•
α(γoct (Ã))xz = α(γoct (Ã))zx = 3.
2.2
Weak Closedness
Usually, the implementations of the Octagon abstract domain maintain all DBMs
strongly closed, so that maximal information is known when performing an abstract
operation. However, this breaks sparsity: indeed, matrix elements of the form Buu
are non-relational interval bounds on the variables: as we expect many variables
to be bounded, the strengthening step gives finite bounds for many DBM cells,
and a strengthened DBM loses most of the sparsity. In general, a DBM has a
quadratic size in the number of variables, and therefore this loss of sparsity is costly.
Previous attempts at improving performances using sparsity [17] did not make this
observation. We believe that, when using these implementations, DBMs quickly
become dense, hence reducing the efficiency of sparse algorithms.
In our algorithms, we propose to skip the strengthening step: instead of maintaining the invariant that all the manipulated DBMs are strongly closed, we maintain
the invariant that they are weakly closed:
Definition 2.11 [Weakly closed DBM] Let B be a DBM. We say that B is weakly
closed when any of the two following equivalent statements hold:
(i) B has a null diagonal and ♯S(B) is strongly closed;
(ii) B has a null diagonal, ♯S(B) is coherent, and:
∀uvw, ♯S(B)uw ≤ Buv + Bvw
Proof. The proof of equivalence of the definitions is in [10, Definition 8.2.5].
(7)
✷
In order to make sure we do not lose precision, we will prove for each of those
operators that it computes abstract values with the same concretization as with
the usual algorithms. Equivalently, we prove that the strengthening of the abstract
values computed by our operators are equal to the abstract values computed by the
usual operators on the strengthened parameters.
A weakly closed DBM is neither necessarily strongly closed nor closed. However,
a closed and coherent DBM is always weakly closed: this helps us easily building
weakly closed DBMs from arbitrary sets of octagonal constraints.
6
Jourdan
Example 2.12 Continuing on the definitions of Example 2.10, α(γpot (Ã)) is closed
and coherent hence weakly closed. This DBM contains no entry relating the variable
x and the other variables. This is an improvement in sparsity compared to the strong
closure α(γoct (Ã)). To the best of our knowledge, this opportunity is not leveraged
by previously known algorithms, such as [17].
This notion of weak closedness has been introduced by Bagnara et al. [1, Appendix A] as an intermediate notion for proving the correctness of the tight closure
algorithm (see §3.5). To the best of our knowledge, the use of weak closedness as an
invariant for manipulating sparse DBMs is an original result of our work.
3
Operations on Difference Bound Matrices
The abstract domain of octagons defines several operations manipulating difference
bound matrices. They include lattice operations, like comparison and join, and
abstract transfer functions, which model state change in the program.
In this section, we recall the standard definition of these operations, and give
the new sparsity-preserving definition on weakly closed DBMs. All these algorithms
preserve the sparsity and weak closedness of DBMs and can be proved to be as
precise as the standard ones. More precisely, we claim that they always return
DBMs whose strengthening equals the DBMs that would have been returned by
the traditional algorithms. The implementation of the widening operation, detailed
in [10, Section 8.2.7], is more complex and omitted by lack of space.
3.1
Comparison
In order to use octagons in a static analyzer, we need to define a comparison operator,
taking two DBMs and returning a Boolean. If this Boolean is true, then we have
the guarantee that the concretization of the first operand is included in that of the
second operand.
A good candidate is ♯ ≤, the natural order relation between DBMs. Its soundness
is guaranteed by the monotonicity of γoct . In usual implementations of the Octagon
abstract domain, DBMs are kept strongly closed, hence this operator is actually as
precise as possible: it returns true if and only if the concretizations are included.
However, in the setting of weakly closed DBMs, this property does not hold. In
order not to lose precision while still using sparse DBM, we need another comparison
operator that strengthens the bounds of the left operand when they do not entail
the right operand:
Definition 3.1 [Weakly closed comparison] Let A and B be DBMs. The weakly
closed comparison of A and B, noted A ♯ ≤weak B is defined by:
A ♯ ≤weak B ≡
^
Auv ≤ Buv ∨
u,v∈V±
Buv <+∞
Auu + Avv
≤ Buv
2
That is, for every finite bound on B, we first check whether it is directly entailed
by the corresponding bound in A, and then try to entail it using non-relational
7
Jourdan
bounds. The following theorem states that it implements the comparison on concretizations, hence we can use it in a sparse context without losing precision:
Theorem 3.2 Let A a weakly closed DBM and B any DBM. The two following
statements are equivalent:
(i) γoct (A) ⊆ γoct (B)
(ii) A ♯ ≤weak B
Proof. See, e.g., [10, Theorem 8.2.9].
3.2
✷
Forgetting Variables
An important operation provided by the Octagon abstract domain is forget. When
given a DBM and a variable v, it returns another DBM where all the information
on v has been forgotten. Its concrete and abstract definitions are given by:
Definition 3.3 [Concrete forgetting] Let x ∈ V+ and S be a set of regular environments. We define:
x
Foct
(S) = {σ + [x ⇒ r; x ⇒ −r] | σ ∈ S, r ∈ R}
Definition 3.4 [Abstract forgetting]
x (B) the DBM such that:
(i) Let x ∈ V± and B be a DBM. We define ♯Fpot
♯ x
Fpot (B)uv
if u = v = x
0
= +∞ otherwise if u = x or v = x
Buv otherwise
(ii) Let x ∈ V+ and B a DBM. We define:
♯ x
Foct (B)
x ♯ x
= ♯Fpot
( Fpot (B))
It is a known result from the Octagon literature [13, Theorems 3.6.1 and 4.4.2]
x is sound when applied to any DBM. Moreover, when applied to any
that ♯Foct
strongly closed DBM, it is exact and returns a strongly closed DBM. To these
x as-is
properties, we add similar properties for weak closedness, that let us use ♯Foct
for weakly closed DBMs without loss of precision:
Theorem 3.5 Let B be a weakly closed DBM and x ∈ V+ . We have:
x (B)) = ♯F x (♯S(B))
(i) ♯S(♯Foct
oct
x (γ (B)) = γ (♯F x (B))
(ii) Foct
oct
oct
oct
x (B) is weakly closed
(iii) ♯Foct
Proof. See [10, Theorem 8.2.11].
3.3
✷
Join
The usual join operator on DBMs is the least upper bound operator for ♯ ≤:
8
Jourdan
Definition 3.6 [DBM least upper bound] Let A and B be two DBMs. The least
upper bound ♯ ∪ on DBMs is defined by:
∀uv, (A ♯ ∪ B)uv = max{Auv ; Buv }
The order relation ♯ ≤ and the operator ♯ ∪ clearly form an upper semi-lattice,
thus usual properties on Galois connections hold, providing the usual results on the
soundness and precision of this operator: ♯ ∪ is sound, and, if given strongly closed
DBMs, it returns the best strongly closed DBM approximating the concrete union.
For weakly closed DBMs, even though ♯ ∪ is sound, it would possibly lose precision
when applied to non-strongly closed DBMs. For example, the weakly closed DBM
A represents the two following inequalities on positive variables x and y:
x+x≤1
y+y ≤0
The weakly closed DBM B, in turn, represents the two following inequalities:
x+x≤0
y+y ≤1
The inequality x + y ≤ 1/2 is not present in A nor in B, even though it is in ♯S(A)
and in ♯S(B). As a result, A ♯ ∪ B contains the inequalities x + x ≤ 1 and y + y ≤ 1,
but does not entail x + y ≤ 1/2, which is entailed however by ♯S(A) ♯ ∪ ♯S(B).
The rationale behind this example is that a join can create some amount of
relationality that was not present in one or both operands. Our operator has to
reflect this fact. Care should be taken, however, not to break the sparsity of the
operands by introducing spurious finite values in the matrix. Our join for weakly
closed DBMs is defined as follows:
Definition 3.7 [Weakly closed join for octagons] Let A and B be two weakly closed
1/2
1/2
vv
vv
and Auv = Auu +A
. The weakly
DBMs. We take, for u, v ∈ V± , Buv = Buu +B
2
2
♯
closed join ∪weak is defined in two steps:
0
(i) We first define A ♯ ∪weak B. Let u, v ∈ V± . We define:
0
(A ♯ ∪weak B)uv
Auv
B
uv
=
1/2
max{Auv ; Buv }
(B ♯ ∪0
weak A)uv
if Auv = Buv
1/2
if Auv < Buv ≤ Buv
1/2
if Auv < Buv ∧ Buv < Buv
if Auv > Buv
(ii) Let u, v ∈ V± . We define:
)
(
0
(An♯ ∪weak B)uv o
if Auu < Buu ∧ Avv > Bvv
min
1/2
1/2
♯
(A ∪weak B)uv =
max Auv ; Buv
or Auu > Buu ∧ Avv < Bvv
(A ♯ ∪0
otherwise
weak B)uv
The first step can be computed by iterating over all the matrix elements that
are different in A and B. This first step thus preserves the sparsity, and consumes
9
Jourdan
computing time only for variables that are different in both branches. The second
step can be computed efficiently by first collecting in a list all the variables u for
which Auu < Buu and, in another list, all those for which Buu < Auu . By iterating
over the two lists, we can efficiently modify only the cells meeting the given condition.
It should be noted that we break in the second step only the sparsity that needs to
be broken, as the modified cells correspond to the cases where the join create new
relational information (as in the example above).
The following theorem states that this modified join operator can be used on
weakly closed DBMs without losing precision or soundness:
Theorem 3.8 Let A and B be two weakly closed DBMs. We have:
(i) ♯S(A ♯ ∪weak B) = ♯S(A) ♯ ∪ ♯S(B)
(ii) γoct (A ♯ ∪weak B) = γoct (α(γoct (A) ∪ γoct (B)))
(iii) A ♯ ∪weak B is weakly closed
Proof. See [10, Theorem 8.2.13].
3.4
✷
Assuming Constraints
An important operation for abstract domains is the assume primitive, which refines
the internal state of an abstract domain using a new assumption over the set of
approximated environments. In this section, we only consider the cases where this
operation is exact, i.e., it does not lead to any approximation. These cases amount
to assuming that ρ(x) − ρ(y) ≤ C, for C ∈ R and x and y two variables. In order
to deal with arbitrary linear inequalities or even arbitrary arithmetical constraints,
it is necessary to write some supporting module for the Octagon domain that will
translate arbitrary constraints into exact ones. Such a support module is out of
the scope of this paper: we refer the reader to [13] for more detail. Moreover, note
that the combination of assume together with the forget let us emulate variable
assignment 4 , hence we do not detail variable assignment in this paper.
We give the assume primitive in two versions: one adapted to γpot , and one
adapted to γoct . We first give the concrete semantics of this operation, which is the
same for irregular and regular environments:
Definition 3.9 [Assuming constraints in the concrete] Let C ∈ R, x, y ∈ V± and S
be a set of irregular environments. We define:
Ax−y≤C (S) = {σ ∈ S | σ(x) − σ(y) ≤ C}
It is easy to see that we can reflect exactly this operation in DBMs. Indeed, it
suffices to change the cell corresponding to the new constraint, if the old value is
larger than the new one. However, this does not maintain any kind of closedness,
whether it be the normal closure, the strong closure or the weak closedness. As a
result, it is necessary to run a closure algorithm when inserting the new constraint.
These algorithms are costly (i.e., cubic complexity), and do not leverage the fact
that the input matrix is already almost closed. For this reason, incremental closure
4
An efficient implementation would however use a specific, optimized implementation for assignments.
10
Jourdan
algorithms have been developed, with quadratic complexity. We give here a slightly
different presentation of these algorithms as the one originally given by Miné [13]:
Definition 3.10 [Assuming constraints in the abstract] Let C ∈ R, B be a DBM
and x, y ∈ V± .
(i) We define ♯Ax−y≤C
(B) the DBM such that, for u, v ∈ V± :
pot
♯ x−y≤C
Apot
(B)uv
= min{Buv ; Bux + C + Byv }
♯ x−y≤C (B) as:
(ii) If x, y ∈ V+ , we define ♯Ax−y≤C
weak (B) and Aoct
♯ x−y≤C
Aweak (B)
♯ x−y≤C
Aoct
(B)
y−x≤C ♯ x−y≤C
= ♯Apot
(Apot
(B))
= ♯S(♯Ax−y≤C
weak (B))
is sound and exact when
It is well-known [10, Theorem 8.2.14] that ♯Ax−y≤C
oct
applied to a DBM with a null diagonal. When applied to a strongly closed DBM B
with 0 ≤ C + Byx , the result is strongly closed. Therefore, an implementation of the
assume primitive in the strongly closed setting first checks whether 0 ≤ C + Byx . If
so, it returns ♯Ax−y≤C
; otherwise it returns ⊥.
oct
In particular, when applied to weakly closed DBMs, ♯Ax−y≤C
is sound and exact,
oct
since weakly closed DBMs have null diagonals. However, because this operator
uses ♯S, it breaks sparsity. The advantage of using weakly closed DBMs is that, in
the setting of weakly closed DBMs, ♯S is no longer needed: ♯Ax−y≤C
can be used
weak
as-is, provided the implementation additionally checks that 0 ≤ 2C + Byy + Bxx .
The following theorem summarizes this result, and justifies the use of this transfer
function in the context of sparse DBMs without loss of precision:
Theorem 3.11 Let C ∈ R, B a weakly closed DBM and x, y ∈ V+ . We have:
♯ x−y≤C (♯S(B))
(i) If 0 ≤ 2C + Byy + Bxx , then ♯S(♯Ax−y≤C
weak (B)) = Aoct
x−y≤C (γ (B))
(ii) γoct (♯Ax−y≤C
oct
weak (B)) = A
(iii) If B
(i)
(ii)
(iii)
(iv)
is weakly closed, the following statements are equivalent:
Ax−y≤C (γoct (B)) 6= ∅
0 ≤ ♯Ax−y≤C
weak (B)xx
0 ≤ C + Byx and 0 ≤ 2C + Byy + Bxx
♯Ax−y≤C (B) is weakly closed.
weak
Proof. See, e.g., [10, Theorem 8.2.15].
3.5
✷
Tightening
Miné [13] and Bagnara et al. [1] study the case of the Octagon abstract domain when
the considered environments take only values in Z: in contrast with the previous
sections, in this case, the strongly closed DBMs are not all canonical, so that modified
algorithms need to be used. We explain here that the use of the weakly closed setting
is compatible with the integer case. To this end, we define a different concretization
Z , that concretizes to integer environments:
function, γoct
11
Jourdan
Definition 3.12 [Integer concretization of octagons] Let B be a DBM. We define:
Z
γoct
(B) = {ρ ∈ γoct (B) | ∀u ∈ V+ , ρ(u) ∈ Z}
If we consider only integer environments, best abstractions have a slightly
stronger characterization. Such DBM are said tightly closed. We also define the
notion of weakly tightly closed DBMs, which is the analog of weakly closed DBMs
for the integer case:
Definition 3.13 [Tight closure] Let B be a DBM. B is tightly closed (respectively
weakly tightly closed) when:
•
B is strongly closed (respectively weakly closed)
•
∀uv ∈ V± , Buv ∈ Z
•
∀u ∈ V± ,
Buu
2
∈Z
Tightly closed DBMs are exactly best abstractions for integer environments [10,
Theorem 8.2.17]. Bagnara et al. [1, §6] give efficient algorithms for computing the
tight closure of a DBM. It consists in using a tightening operation before strengthening. The tightening operation is defined by:
Definition 3.14 [Tightening] Let B a DBM with elements in Z. We define ♯T (B)
be the DBM with elements in Z such that, for u, v ∈ V± :
♯
T (B)uv =
(
Buv − 1 if u = v and Buv is odd
Buv
otherwise
The following theorem gives the essential property of the tightening operation:
Theorem 3.15 Let B a weakly closed DBM with elements in Z. We suppose that
∀u ∈ V± , 0 ≤ ♯T (B)uu + ♯T (B)uu . Then ♯T (B) is weakly tightly closed.
Proof. See, e.g., [10, Theorem 8.2.18].
✷
This theorem has two consequences. First, as already explained by Bagnara
et al. [1, §6], it gives an efficient algorithm to compute tight closure: one would
compute the closure of the input matrix, then tighten it and finally strengthen it.
Second, our sparse algorithms need only small adjustments when used with integer
environments: instead of maintaining the DBMs weakly closed, we just have to make
them weakly tightly closed by tightening them after each operation.
Note, however, that tightening does not address the case of mixed environments,
where some variables are known to have integer values, and some others can have
an arbitrary real values. To the best of our knowledge, there is no known efficient
closure algorithm supporting this use case, even in the dense setting.
4
Conclusion
In this paper, we presented new algorithms for the Octagon abstract domain, which
preserve the sparsity of the representation of octagons. These algorithms are as
12
Jourdan
precise as the usual ones, and rely on a weaker invariant over difference bound
matrices, called weak closedness. We have shown that these algorithms can be used
in the context of rational or real environments as well as in the context of integer
environments.
We implemented and formally verified in Coq these algorithms in the context of
the Verasco static analyzer [9,10,18]. The use of these new algorithms improved the
performances of the Octagon abstract domain by at least one order of magnitude.
There are still possible improvements to these algorithms: in particular, we think
that it could be profitable to sparsify difference bound matrices as much as possible
after each abstract operation, while still maintaining them weakly closed. Indeed,
abstract operations may infer bounds in difference bound matrices that can actually
be deduced from non-relational bounds, therefore missing opportunity of sparsity.
We think the reduction algorithm presented by Bagnara et al. [1] can be adapted
to compute reduced difference bound matrices using only weakly closed difference
bound matrices. This would lead to a simpler widening algorithm based on a semantic definition as described by Bagnara et al. [1, §4.2]. We believe the implementation
of these new algorithms in state-of-the-art static analyzers, by using, for example,
the framework developed by Singh et al. [17] would lead to a significant performance
improvement.
References
[1] Bagnara, R., P. M. Hill and E. Zaffanella, Weakly-relational shapes for numeric abstractions: Improved
algorithms and proofs of correctness, Formal Methods in System Design 35 (2009), pp. 279–323.
[2] Bouaziz, M., TreeKs: A functor to make numerical abstract domains scalable, in: NSAD, ENTCS 287
(2012), pp. 41–52.
[3] Chawdhary, A., E. Robbins and A. King, Simple and efficient algorithms for octagons, in: APLAS,
LNCS 8858 (2014), pp. 296–313.
[4] Chen, L., J. Liu, A. Miné, D. Kapur and J. Wang, An abstract domain to infer octagonal constraints
with absolute value, in: SAS, LNCS 8723 (2014), pp. 101–117.
[5] Clarisó, R. and J. Cortadella, The octahedron abstract domain, in: SAS, LNCS 3148 (2004), pp. 312–
327.
[6] Cousot, P. and R. Cousot, Static determination of dynamic properties of programs, in: Proceedings of
the Second International Symposium on Programming (1976), pp. 106–130.
[7] Cousot, P., R. Cousot, J. Feret, L. Mauborgne, A. Miné and X. Rival, Why does astrée scale up?,
Formal Methods in System Design 35 (2009), pp. 229–264.
[8] Cousot, P. and N. Halbwachs, Automatic discovery of linear restraints among variables of a program,
in: POPL (1978), pp. 84–96.
[9] Jourdan, J., V. Laporte, S. Blazy, X. Leroy and D. Pichardie, A formally-verified C static analyzer, in:
POPL, ACM, 2015, pp. 247–259.
[10] Jourdan, J.-H., “Verasco: a Formally Verified C Static Analyzer,” Ph.D. thesis, Université Paris Diderot
(2016).
[11] Laviron, V. and F. Logozzo, Subpolyhedra: A (more) scalable approach to infer linear inequalities, in:
VMCAI, LNCS 5403 (2009), pp. 229–244.
[12] Logozzo, F. and M. Fähndrich, Pentagons: a weakly relational abstract domain for the efficient
validation of array accesses, in: SAC, ACM, 2008, pp. 184–188.
[13] Miné, A., “Weakly relational numerical abstract domains,” Ph.D. thesis, École Polytechnique (2004).
13
Jourdan
[14] Miné, A., The octagon abstract domain, HOSC 19 (2006), pp. 31–100.
[15] Putot, S. and E. Goubault, Static analysis of numerical algorithms, in: SAS, LNCS 4134 (2006), pp.
18–34.
[16] Simon, A. and A. King, The two variable per inequality abstract domain, HOSC 23 (2010), pp. 87–143.
[17] Singh, G., M. Püschel and M. Vechev, Making numerical program analysis fast, in: PLDI, ACM, 2015,
pp. 303–313.
[18] The Verasco formally verified C static analyzer, http://compcert.inria.fr/verasco/ .
14
| 6cs.PL
|
ON WEAK FRAÏSSÉ LIMITS
arXiv:1711.09295v1 [math.LO] 25 Nov 2017
ZAKHAR KABLUCHKO AND KATRIN TENT
Abstract. Using the natural action of S∞ we show that a countable hereditary class
C of finitely generated structures has the joint embedding property (JEP) and the weak
amalgamation property (WAP) if and only if there is a structure M whose isomorphism
type is comeager in the space of all countable, infinitely generated structures with age in C.
In this case, M is the weak Fraı̈ssé limit of C.
This applies in particular to countable structures with generic automorphisms and recovers a result by Kechris and Rosendal [Proc. Lond. Math. Soc., 2007 ].
1. Introduction
After writing the note [2] showing that the class of universal-homogenous structures is
comeager in some appropriate space of structures, we noticed that the methods from the
paper of Kechris and Rosendal [3] easily yield a stronger and more general result applying
to weak Fraı̈ssé classes of finitely generated structures.
Thus, the aim of the present note is to prove that for an arbitrary hereditary class C
of finitely generated (not necessarily finite) structures, the space of all countable, infinitely
generated structures with age in C contains a comeager subset of structures isomorphic to
some countable structure M if and only if C has the weak Fraı̈ssé property and M is the
weak Fraı̈ssé limit of C. As a corollary, this proves the existence and uniqueness of the weak
Fraı̈ssé limit of a weak Fraı̈ssé class. We also explain how this in turn reproves the result by
Kechris and Rosendal [3] on comeager conjugacy classes of automorphisms of such structures
in the slightly more general setting of finitely generated structures.
Similar results for classes of finite structures were obtained by Kruckman in his PhD
thesis [4] using Banach–Mazur games on posets.
2. Background and the main result
2.1. Weak Fras̈sé limits. Let L be a countable language. For background on model theory
we refer to [6]. Here and in what follows, the word ‘structure’ always means ‘L-structure’.
Definition 2.1. Let C be a countable class of (isomorphism types of) finitely generated
structures and assume that C is hereditary, that is closed under taking finitely generated
substructures. We say that C is a weak Fraı̈ssé class if C satisfies
2010 Mathematics Subject Classification. Primary: 03C15; Secondary: 54E52, 60B99.
Key words and phrases. Fraı̈ssé class, universality, homogeneity, comeager set, Baire property, Polish
group actions.
1
2
ZAKHAR KABLUCHKO AND KATRIN TENT
(WAP) (Weak Amalgamation Property) For any A ∈ C there is some B ∈ C and embedding
e : A −→ B such that for all B1 , B2 ∈ C and embeddings hi of B into Bi , i = 1, 2, there
is some C ∈ C and embeddings gi of Bi into C, i = 1, 2, such that g1 ◦h1 ◦e = g2 ◦h2 ◦e.
(JEP) (Joint Embedding Property) For any A, B ∈ C there is some D ∈ C containing A and
B as a substructure.
The class C is called a Fraı̈ssé class if in condition (WAP) we can choose B = A and
e = id. The notion of Weak Amalgamation was first studied by Ivanov in [1].
For a structure M, the age of M is the class of isomorphism types of finitely generated
substructures of M.
Definition 2.2. Let C be a countable hereditary class of finitely generated structures and
let M be a countable structure.
(1) (C-Universality) We say that M is C-universal if age(M) = C.
(2) (Weak C-Homogeneity) We say that M is weakly C-homogeneous if for any structure
A ∈ C there is some B ∈ C and embedding e : A −→ B such that for any substructures B1 , B2 ⊂ M with (A1 , B1 ), (A2 , B2 ) isomorphic to (e(A), B) any isomorphism
ϕ : (A1 , B1 ) → (A1 , B2 ) extends to an automorphism f of M with ϕ|A1 = f |A1 .
(3) (Weak C-Saturation) For any structure A ∈ C there is some B ∈ C and embedding
e : A −→ B such that for any D ∈ C containing B and any embedding h : B −→ M
there is an embedding of D into M extending h ◦ e on A.
The structure M is called C-homogeneous (C-saturated, respectively) if we can choose
A = B and e = id.
A countable structure M is called a (weak) Fraı̈ssé limit for C if it is C-universal and
(weakly) C-homogeneous.
We will show that a countable hereditary class C of finitely generated structures has a
weak Fraı̈ssé limit if and only if C satisfies (JEP) and (WAP).
2.2. The space of structures and the action of S∞ . Let C be a countable hereditary
class of finitely generated structures and let S be the set of all structures M defined on the
universe ω (or any other fixed countable universe) with the following two properties:
(i) the age of M is contained in C and
(ii) M is not finitely generated.
Metric on S: We endow S with the following metric: For two different structures M, N ∈ S
we put d(M, N) = 2−k where k ∈ {0, 1, . . .} is minimal such that the substructures of M
and N generated by the elements 0, 1, . . . , k are different, i.e. either these substructures
have different universes as subsets of ω or their universes agree, but some constant, relation
or function differs on the universe. Note that with this metric, S becomes a complete
(ultra)metric space. Consequently, the Baire category theorem holds in S.
Topology on S: For a finitely generated structure B whose isomorphism type is contained
in C and such that dom B ⊂ ω, let OB be the set of all structures A ∈ S whose restriction to
dom B coincides with B. It is easy to see that these sets (after discarding the empty ones)
3
form a basis of some topology T on S, see [2], and that T is the topology induced by the
metric d on S.
Proposition 2.3. The space (S, T ) is separable if and only if every structure from C that
can be embedded into a countable, infinitely generated structure, is finite.
Proof. Let (S, T ) be separable with a countable dense set {X1 , X2 , . . .}. Assume, by contraposition, that there is an infinite structure A0 ∈ C that can be embedded into some countable
infinitely generated structure. Taking some infinite set U ⊂ ω as a universe, we can construct
a structure A(U) isomorphic to A0 . The open sets of the form OA(U ) are non-empty, hence
any such set contains some point of the form Xi . Since there are uncountably many possible
U’s, there is at least one Xi contained in uncountably many sets of the form OA(U ) . Hence,
Xi has uncountably many finitely generated substructures with pairwise different universes.
But this is a contradiction. Conversely, if all finitely generated substructures of any M ∈ S
are finite, then there are only countably many non-empty basic open sets of the form OB .
By choosing a point in any such set, we obtain a countable dense subset of S, thus proving
separability.
Action of S∞ on S: Consider S∞ , the permutation group on ω, as a Polish group with a
neighbourhood basis of the identity consisting of pointwise stabilizers of finite sets. There is
a natural continuous action of G = S∞ on S just given by the action g ∈ G on the domain
ω of a structure and transferring the structure to the image. Under this action, the orbit of
a structure M ∈ S is the class of structures isomorphic to M and the stabilizer GM consists
of all automorphisms of M.
Proposition 2.4. (cp. [3], Prop. 3.2) Let G be a closed subgroup of S∞ acting continuously
on a topological space X and let x ∈ X. Then the following are equivalent:
(1) The orbit of x under G is non-meager.
(2) For any open subgroup V of G the orbit of x under V is somewhere dense.
Proof. ¬(2) ⇒ ¬(1): Suppose that the orbit of x under some open subgroup V ≤ G is
nowhere dense. Since V has countable index in G and the orbit of x under each coset gV is
also nowhere dense, the orbit of x under G is meager.
¬(1) ⇒ ¬(2): Conversely, assume that the orbit of x under G is meager which means that
it is contained in the countable union of closed nowhere dense sets Fn , n ∈ ω. Then the sets
Kn = {g ∈ G : g(x) ∈ Fn } are closed in G and their union is all of G. At least one of the Kn
has non-empty interior, so contains a translate of some open subgroup V . Thus the orbit of
x under (all translates of) V is nowhere dense.
2.3. Main result. Recall that a subset of a topological space is called comeager if its complement is meager or, equivalently, if the set can be represented as a countable intersection
of subsets with dense interior.
Theorem 2.5. Let C be a countable hereditary class of finitely generated structures and let
(S, T ) be the topological space of all countable structures on ω with age contained in C and
which are not themselves finitely generated. The topology T is given by the metric d as above.
Then the following are equivalent:
4
ZAKHAR KABLUCHKO AND KATRIN TENT
(1)
(2)
(3)
(4)
(5)
(6)
The class C has a weak Fraı̈ssé limit M.
There is a C-universal and weakly C-saturated structure M.
The S∞ -orbit of some structure M is comeager in S.
For some structure M the set of structures isomorphic to M is comeager in S.
The class C satisfies (JEP) and (WAP).
There is a C-universal weakly C-saturated structure M such that (*) holds:
(*) any finitely generated substructure A of M is contained in a finitely generated
substructure B ∈ M (not necessarily the one given by weak C-saturation) such
that any D ∈ C containing a copy of B embeds over A into M.
It follows in particular, that a weak Fraı̈ssé limit, if it exists, is unique up to isomorphism.
Proof. (1) ⇒ (2): Let M ∈ S be a weak Fraı̈ssé limit of C. We have to show that M is
weakly C-saturated, Let A ⊆ B ∈ C be as given by weak C-homogeneity. Assume that
(A0 , B0 ) is an isomorphic copy of (A, B) in M. Let D ∈ C contain an isomorphic copy of B.
By C-universality, M contains a substructure D0 isomorphic to D. By C-homogeneity, any
isomorphism between the copy of A inside D0 and A0 can be extended to an automorphism
of M, yielding an embedding of D over A0 into M as required.
(2) ⇒ (3): Let M ∈ S be C-universal and weakly C-saturated. Then age(M) = C
and thus the S∞ -orbit of M is dense. It now suffices to verify that the S∞ -orbit of M is
non-meager: since the action of S∞ on the metric space S is continuous, it then follows
from [5], Thm. 4.4 that the orbit is Gδ and hence comeager (given that it is dense). To
see that the S∞ -orbit of M is non-meager, it suffices by Proposition 2.4 to see that for any
open subgroup V ≤ S∞ the V -orbit of M is somewhere dense. We may assume that there
is a finite set A0 such that V = GA0 is the pointwise stabilizer of A0 . Let A ⊂ M be the
substructure generated by A0 and let B ∈ C be such that B contains an isomorphic copy of
A and such that any extension of B can be embedded into M by weak C-saturation. Now
since B ∈ age(M) we may assume that B ⊂ M. Let A′ denote the isomorphic copy of A
inside B with A′0 denoting the finite set corresponding to A0 . Then V ′ = GA′0 is conjugate
to GA0 in S∞ and it clearly suffices to prove the claim for V ′ . We claim that the V ′ -orbit of
M is dense in OB : By C-saturation of M, any extension D of B can be embedded over A′
into M. This says that for some g ∈ V ′ we have g(M) ∈ OD , which is enough.
(3) ⇔ (4): clear.
(3) ⇒ (5): Let M ∈ S be an element whose orbit under G = S∞ is comeager. Since
S is a Baire space, a comeager set is dense and thus we see that age(M) = C and hence
(JEP) holds. It remains to prove that C satisfies (WAP). Let A ∈ C. Then M contains a
substructure isomorphic to A and we may assume that A ⊂ M. Let A0 ⊂ A be a finite
set generating A. Then the pointwise stabilizer of A0 in G is an open subgroup GA0 and
by Proposition 2.4 the orbit of M under GA0 is dense in some basic open set OB for some
finitely generated structure B. Thus, M ′ := f (M) ∈ OB for some f ∈ GA0 . Let A′ = f (A)
be the structure (with domain contained in ω) obtained by transferring A by f . Then,
M ′ ∈ OB ∩ OA′ meaning that B and A′ are substructures of M ′ . Let D be the substructure
of M ′ generated by A′ and B. Then, OD ⊂ OA′ ∩ OB . To prove (WAP), let E, F ∈ C be
structures containing D (and assume that the domains of E and F are embedded into ω).
Since the orbit of M (and hence M ′ ) under GA0 is dense in OD , there exist g, h ∈ GA0 such
5
that g(M ′ ) ∈ OE and h(M ′ ) ∈ OF . Observe that both g and h equal to the identity map on
A0 , hence g|A′ = h|A′ = id. Hence the substructure of M ′ generated by g −1 (E)∪h−1 (F ) ⊂ M ′
is the required weak amalgam over A′ .
(5) ⇒ (6): Assume that C has (JEP) and (WAP). We will construct a C-universal and
weakly C-saturated structure satisfying (*). Choose an enumeration (Ci )i∈ω of all isomorphism types in C. We construct M as the union of an ascending chain
M0 ⊂ M1 ⊂ · · · ⊂ M
of elements of C. Suppose that Mi is already constructed. If i = 2n is even, let Ai+1 be
the structure embedding both Mi and Cn given by (JEP) and let Mi+1 ∈ C be the structure
given by (WAP) for Ai+1 .
We use the odd steps to ensure that the limit M is weakly C-saturated. For i = 2n + 1,
let A ∈ C and let B ∈ C be as guaranteed by (WAP) where we assume that A ⊂ B. Let
D ∈ C and assume we are given two embeddings f0 : B → Mi and f1 : B → D. Then we let
Mi+1 ∈ C be the structure guaranteed to exist by (WAP).
Assume now that we have A ⊂ B, D ∈ C and embeddings f0 : B → D and f1 : B → M.
Since B is finitely generated, the image of f0 will be contained in some Mj . Thus, in order
to guarantee the weak C-saturation of M, we have to ensure during the construction of the
the Mi , that eventually, for some odd i ≥ j, the embeddings f0 : B → Mi and f1 : B → D
were used in the construction of Mi+1 . This can be done since for each j there are – up
to isomorphism – at most countably many possibilities. Thus there exists an embedding
g1 : D → Mi+1 with f0 |A = g1 ◦ f1 |A .
By the even stages, we have C ⊂ age(M). Conversely, the finitely-generated substructures of M are the substructures of the Mi . Since the Mi belong to C, their finitely-generated
substructures also belong to C. Hence age(M) = C as required. Furthermore, since any
finitely generated substructure A is contained in some Mi , the structure M2i satisfies (*)
for A.
(6) ⇒ (1): Let M ∈ S be C-universal and weakly C-saturated such that (*) holds.
We have to show that M is weakly C-homogeneous. Let A0 ⊆ A1 ∈ C be as given by
weak C-saturation. Assume that (B0 , B1 ) and (C0 , C1) are isomorphic copies of (A0 , A1 )
in M and that ϕ0 : (B0 , B1 ) → (C0 , C1) is an isomorphism. We want to extend ϕ0 |B0
to an automorphism of M by a back-and-forth construction. Suppose we have already
constructed ϕ2i−1 : B2i−1 → C2i−1 with ϕ2i−1 |B0 = ϕ0 such that C2i−1 is an extension of
C2i−2 satisfying (*). Let B2i be the substructure given by (*) containing the substructure
of M generated by B2i−1 ∪ {i}. Then there is a copy C2i such that there is an isomorphism
ϕ2i : (B2i−1 , B2i ) → (C2i−1 , C2i ) with ϕ2i |B0 = ϕ0 . Now let C2i+1 be the structure given by
(*) containing the substructure of M generated by C2i ∪ {i} and choose an isomorphic copy
B2i+1 of C2i +1 and an isomorphism ϕ2i+1 : (B2i , B2i+1 ) → (C2i , C2i +1 ) extending ϕ2i |B2i−1 .
Thus again ϕ2i+1 |B0 = ϕ0 and we continue.
2.4. Finding automorphisms with comeager conjugacy class. The proof of Theorem 2.5 was motivated by the methods from [3]. We now show how this in turn implies their
result on comeager conjugacy classes.
6
ZAKHAR KABLUCHKO AND KATRIN TENT
For a Fraı̈ssé class C of finitely generated L-structures, let Cp denote the class of all
systems of the form S = hA, ψ : B → Ci where A, B, C ∈ C, B, C ⊂ A and ψ is an
isomorphism of B and C. An embedding of one system S = hA, ψ : B → Ci into another
system T = hD, ϕ : E → F i is an embedding f : A → D such that f embeds B into E, C
into F and such that f ◦ ψ ⊂ ϕ ◦ f . It thus makes sense to talk about (JEP) and (WAP)
and hence about Weak Fraı̈ssé limits for Cp .
We now consider comeagerness in the space Sp of structures (M, h) where M ∈ S and
h ∈ Aut(M). Then, for any finitely generated substructure A of M and isomorphic finitely
generated substructures B ⊂ A, C = h(B), the system S = hA, h|B : B → Ci is in Cp . As
before, the space Sp is a complete metric space. The group Aut(M) is endowed with the
usual topology of pointwise convergence.
Remark 2.6. We could instead consider an expanded language Lp = L∪{f } containing a new
symbol f for a partial function. Then Cp denotes the class of all structures A ∈ C expanded
by an interpretation of the symbol f as an isomorphism f : B → C where A, B, C ∈ C and
B, C ⊂ A. By embedding an Lp -structure A into an Lp -structure D, we may enlarge the
domain of the function f inside A (contrary to the usual conventions).
Note that we do not assume the Fraı̈ssé class C to consist of finite structures.
Theorem 2.7. Let C be a Fraı̈ssé class of finitely generated structures with Fraı̈ssé limit M.
Then the following are equivalent:
(1) Cp has (JEP) and (WAP).
(2) The weak Fraı̈ssé limit of Cp exists in Sp and has comeager isomorphism type in Sp .
(3) There is f ∈ Aut(M) with comeager conjugacy class in Aut(M).
If any of the above conditions holds, the weak Fraı̈ssé limit of Cp has the form (M, f ), where
f ∈ Aut(M) has a comeager conjugacy class.
Proof. (1) ⇒ (2): The proof of Theorem 2.5 (5) ⇒ (6) shows how to construct the weak
Fraı̈ssé limit (N, g). It is left to show that g is an automorphism of N. For this it suffices to
verify that g is defined on all of N and is surjective. Suppose not, so there is some a ∈ N
where g is not defined. Since N is the weak Fraı̈ssé limit of Cp it is the countable union of
systems Si = hNi , ϕi : Bi → Ci i for i < ω. Let i be minimal such that g is not defined on
all of Ni . By weak Cp -saturation, there is some system T = hD, ϕ : E → F i extending Si
such that any extension of T can be embedded over Si into N. Since C is a Fraı̈ssé class
with Fraı̈ssé limit M, there is an embedding of D into M and by ultrahomogeneity of M, ϕ
extends to an automorphism of ϕ̂ of M. Then ϕ̂ is defined on all of Ni ⊂ D. Let G be the
substructure of M generated by D and ϕ̂(D). The system hG, ϕ̂|D : D → ϕ̂(D)i is in Cp and
extends T , thus embeds over Ni into N, showing that g is defined on all of N. A similar
argument shows that g is surjective.
The proof of Theorem 2.5 (2) ⇒ (3) shows that the isomorphism type of (N, g) in Sp is
comeager.
(2) ⇒ (3): Suppose the class Cp has a weak Fraı̈ssé limit (N, g). To see that N is
isomorphic to M it suffices to check that N is weakly C-saturated. Then both M and N have
comeager isomorphism classes and thus must be isomorphic. Clearly we have age(N) = C.
7
To see that N is weakly C-saturated, let A ⊂ N be a finitely generated substructure, so
A ∈ C. Consider the system hA, id : E → Ei where E is the substructure generated by the
emptyset. Note that any partial isomorphism extens id : E → E. By (WAP) for Cp there is
some system T = hB, ϕ : C → Di such that any system containing T embeds into N. Now
let F ∈ C be a finitely generated structure containing the structure B. Then the system
hF, ϕ : C → Di is in Cp and hence embeds into N over A. This shows that we can choose B
to witness the weak C-saturation of N.
It is left to prove that the conjugacy class of g is comeager in Aut(N). To see this note
that the orbit of (N, g) under S∞ is comeager in the class Sp . Since g ∈ Aut(N), the orbit
of S∞ under the stabilizer of N is exactly the orbit of g under conjugation in Aut(N). Since
Aut(N) is a Polish group, the claim follows from the Kuratowski-Ulam theorem.
(3) ⇒ (1): The same proof as Theorem 2.5 (5) ⇒ (6) and (6) ⇒ (1) shows the required.
References
[1] A.A. Ivanov. Generic expansions of ω-categorical structures and semantics of generalized
quantifiers. J. Symb. Log., 64(2):775–789, 1999. doi: 10.2307/2586500.
[2] Z. Kabluchko and K. Tent. Universal-homogeneous structures are generic. Preprint, 2017.
Available at http://arxiv.org/abs/1710.06137.
[3] A. S. Kechris and C. Rosendal. Turbulence, amalgamation, and generic automorphisms
of homogeneous structures. Proc. Lond. Math. Soc. (3), 94(2):302–350, 2007. ISSN
0024-6115. URL https://doi.org/10.1112/plms/pdl007.
[4] A. Kruckman. Infinitary limits of finite structures. PhD Thesis, University of California,
Berkeley, 2016. Available at http://pages.iu.edu/ akruckma/thesis.pdf.
[5] R. L. Sami. Polish group actions and the Vaught conjecture. Trans. Amer. Math. Soc.,
341(1):335–353, 1994. ISSN 0002-9947. URL https://doi.org/10.2307/2154625.
[6] K. Tent and M. Ziegler.
A course in model theory, volume 40 of Lecture
Notes in Logic.
Association for Symbolic Logic, La Jolla, CA; Cambridge
University Press, Cambridge, 2012.
doi: 10.1017/CBO9781139015417.
URL
http://dx.doi.org/10.1017/CBO9781139015417.
Zakhar Kabluchko: Institut für Mathematische Stochastik, Westfälische WilhelmsUniversität Münster, Orléans-Ring 10, 48149 Münster, Germany
E-mail address: [email protected]
Katrin Tent: Mathematisches Institut und Institut für Mathematische Logik und Grundlagenforschung, Westfälische Wilhelms-Universität Münster, Einsteinstr. 62, 48149 Münster,
Germany
E-mail address: [email protected]
| 4math.GR
|
The complement of proper power graphs of finite groups
T. Anitha∗, R. Rajkumar†
arXiv:1601.03683v2 [math.GR] 17 Dec 2017
Department of Mathematics, The Gandhigram Rural Institute – Deemed to be University,
Gandhigram – 624 302, Tamil Nadu, India.
Andrei Gagarin‡
School of Mathematics, Cardiff University,
Cardiff, CF24 4AG, Wales, UK
Abstract
For a finite group G, the proper power graph P∗ (G) of G is the graph whose vertices are
non-trivial elements of G and two vertices u and v are adjacent if and only if u 6= v and um = v
or v m = u for some positive integer m. In this paper, we consider the complement of P∗ (G),
denoted by P∗ (G). We classify all finite groups whose complement of proper power graphs is
complete, bipartite, a path, a cycle, a star, claw-free, triangle-free, disconnected, planar, outerplanar, toroidal, or projective. Among the other results, we also determine the diameter and
girth of the complement of proper power graphs of finite groups.
Keywords: Complement of power graph, finite groups, diameter, girth, bipartite graph,
planar graph, toroidal graph, projective-planar graph.
2010 Mathematics Subject Classification: 05C25, 05C10.
1
Introduction
The investigation of properties of a given algebraic structure can be made by associating it with a
suitable graph, and then by analyzing the graph properties using methods of graph theory. This
∗
e-mail: [email protected]
e-mail: [email protected]
‡
e-mail: [email protected]
†
1
2
approach have been used in a great amount of literature, for example, see [1], [3], [8]. Moreover,
there are several recent papers dealing with embeddability of graphs associated with algebraic
structures on topological surfaces. For instance, see [2], [12], [21], [22], [23]. Kelarev and Quinn
[15] introduced and studied the directed power graph of a semigroup. The directed power graph
of a semigroup S is a digraph having the vertex set S, and, for u, v ∈ S, there is an arc from u
to v if and only if u 6= v and v = um for some positive integer m. Later, Chakrabarty et al. [13]
defined the undirected power graph P(G) of a group G as an undirected graph whose vertex set is
G, and two vertices u and v are adjacent if and only if u 6= v and um = v or v m = u for some
positive integer m. Recently, several interesting results have been obtained for these graphs. For
instance, see [16], [26]. Mirzargar et al. [18] investigated planarity of the undirected power graph of
finite groups, and Xuanlong Ma and Kaishun Wang [29] classified all finite groups whose undirected
power graphs can be embedded on the torus.
Further, in [19], Moghaddamfar et at. considered the graph P∗ (G), which is obtained by
removing the identity element from the undirected power graph P(G) of a given group G. This
graph is called the undirected proper power graph of G. They have studied several properties of
these graphs, including the classification of finite groups whose undirected proper power graphs
are one of strongly regular, bipartite, planar, or Eulerian. Later, in [7], Doostabedi and Farroki
have investigated various kinds of planarity, toroidality, and projective-planarity of these graphs.
An interested reader may refer to the survey [14] for further results and open problems related to
the power graph of groups and semigroups. In this paper, we consider only the undirected graphs,
and, for simplicity, use the term ‘power graph’ to refer to the undirected power graph.
In this paper, we study the properties of complement of the proper power graph of a group.
For a given group G, the complement of the proper power graph of G, denoted by P∗ (G), is a graph
whose vertex set is the set of all nontrivial elements of G, and two vertices u and v are adjacent if
and only if u 6= v, and um 6= v and v n 6= u for any positive integers m, n; in other words u and v
are adjacent if and only if u 6= v, u ∈
/ hvi and v ∈
/ hui.
The rest of the paper is organized as follows. In Section 2, we provide some preliminaries and
notations. In Section 3, we classify all finite groups whose complement of proper power graph
is complete, bipartite, C3 -free, K1,3 -free, disconnected, or having isolated vertices. Moreover, in
this section, we determine the girth and diameter of the complement of proper power graphs of
finite groups. In Section 4, we classify all finite groups whose complement of proper power graphs
is planar, toroidal, or projective-planar. As a consequence, we classify the finite groups whose
3
complement of proper power graphs is a path, a star, a cycle, outer-planar, or having no subgraphs
K1,4 or K2,3 .
2
Preliminaries and notations
In this section, we remind some concepts, notation, and results in graph theory and group theory.
We follow the terminology and notation of [11, 27] for graphs and [25] for groups. A graph G is
said to be complete if there is an edge between every pair of its distinct vertices. G is said to be
k-partite if the vertex set of G can be partitioned to k subsets, called parts of G, such that no
two vertices in the same subset of the partition are adjacent. A complete k-partite graph, denoted
by Kn1 ,n2 ,...,nk , is a k-partite graph having its parts sizes n1 , n2 , . . . , nk such that every vertex in
each part is adjacent to all the vertices in the other parts of Kn1 ,n2 ,...,nk . For simplicity, we denote
the complete k-partite graph Kn,n,··· ,n by K(k, n). The graph K1,n is called a star. Pn and Cn
respectively denote the path and cycle on n vertices. We denote the degree of a vertex v in G by
degG (v). G is said to be H-free, if it has no induced subgraph isomorphic to H.
G is said to be connected if there exists a path between any two distinct vertices in the graph;
otherwise G is said to be disconnected. The distance between two vertices u and v of a graph,
denoted by d(u, v), is the length of a shortest path between u and v in the graph, if such a path
exists, and ∞ otherwise. The diameter of a connected graph G is the maximum distance between
any two vertices in the graph, and is denoted by diam(G). The number of edges in a path or a
cycle, is called its length. The girth of G is the minimum of the lengths of all cycles in G, and
is denoted by gr(G). If G is acyclic, that is, if G has no cycles, then we write gr(G) = ∞. The
complement G of G is a graph, which has the vertices of G as its vertex set, and two vertices in
G are adjacent if and only if they are not adjacent in G. Given two simple graphs, G1 = (V1 , E1 )
and G2 = (V2 , E2 ), their union, denoted by G1 ∪ G2 , is a graph with the vertex set V1 ∪ V2 and the
edge set E1 ∪ E2 . Their join, denoted by G1 + G2 , is a graph having G1 ∪ G2 together with all the
edges joining points of V1 to points of V2 .
A graph is said to be embeddable on a topological surface if it can be drawn on the surface in
such a way that no two edges cross. The orientable genus or genus of a graph G, denoted by γ(G),
is the smallest non-negative integer n such that G can be embedded on the sphere with n handles.
G is said to be planar or toroidal, respectively, when γ(G) is either 0 or 1. A planar graph G is
said to be outer-planar if it can be drawn in the plane with all its vertices lying on the same face.
4
A crosscap is a circle (on a surface) such that all its pairs of opposite points are identified,
and the interior of this circle is removed. The nonorientable genus of G, denoted by γ(G), is the
smallest integer k such that G can be embedded on the sphere with k crosscaps. G is said to be
projective or projective-planar if γ(G) = 1. Clearly, if G0 is a subgraph of G, then γ(G0 ) ≤ γ(G)
and γ(G0 ) ≤ γ(G).
Let G be a group. The order of an element x in G is denoted by o(x). For a positive integer
n, ϕ(n) denotes the Euler’s totient function of n. For any integer n ≥ 3, the dihedral group of
order 2n is given by D2n = a, b|an = b2 = e, ab = ba−1 . For any integer n ≥ 2, the quarternion
group of order 4n is given by Q4n = a, b|a2n = b4 = 1, b2 = an , ab = ba−1 . For any α ≥ 3 and a
prime p, the modular group of order pα is given by Mpα = ha, b|ap
α−1
= bp = 1, bab−1 = ap
α−2
+ 1i.
Throughout this paper, p, q denotes distinct prime numbers.
The following results are used in the subsequent sections.
Theorem 2.1. ([27, Theorem 6.6]) A graph G is planar if and only if G contains no subgraphs
homeomorphic to K5 or K3,3 .
Theorem 2.2. ([27, Theorems 6.37, 6.38, 11.19, 11.23])
(1) γ(Kn ) =
l
(n−3)(n−4)
12
l
(2) γ(Km,n ) =
(3) γ(Kn ) =
l
l
,n≥3
(m−2)(n−2)
4
(n−3)(n−4)
6
(4) γ(Km,n ) =
m
m
m
, n, m ≥ 2
, n ≥ 3, n 6= 7; γ(Kn ) = 3 if n = 7.
(m−2)(n−2)
2
m
, n, m ≥ 2.
As a consequence, γ(Kn ) > 1 for n ≥ 8, γ(Kn ) > 1 for n ≥ 7, γ(Km,n ) > 1 if either m ≥ 4, n ≥ 5
or m ≥ 3, n ≥ 7, and γ(Km,n ) > 1 if either m ≥ 3, n ≥ 5 or m = n = 4.
Theorem 2.3.
(i) ([17, p.129]) The number of non-cyclic subgroup of order pα in any non-cyclic
group of order pm is of the form 1 + kp whenever 1 < α < m and p > 2.
(ii) ([25, Proposition 1.3]) If G is a p-group of order pn , and it has a unique subgroup of order
pm , 1 < m ≤ n, then G is cyclic or m = 1 and p = 2, G ∼
= Q2α .
(iii) ([5, Theorem IV, p.129]) If G is a p-group of order pn , then the number of subgroups of order
ps , 1 ≤ s ≤ n is congruent to 1 (mod p).
5
3
Some results on the complement of proper power graphs of
groups
Theorem 3.1. Let G be a finite group. Then P∗ (G) is complete if and only if G ∼
= Zm
2 , m ≥ 1.
Proof. Assume that P∗ (G) is complete. Then every element of G is of order 2. For, if G contains
an element x of order p (6= 2), then there exist a non-trivial element y in hxi. Then x is not
adjacent to y in P∗ (G), which is a contradiction to the hypothesis. Therefore, G is a 2-group with
exponent 2. Since any group with exponent 2 must be abelian, so G ∼
= Zm
2 , m ≥ 1. Conversely, if
∗
G∼
= Zm
2 , m ≥ 1, then every element of G is of order 2, so it follows that P (G) is complete.
Theorem 3.2. Let G be a finite group. Then P∗ (G) is K1,3 -free if and only if G is isomorphic to
one of the following:
(i) Zpn , Z6 , S3 , Zn2 , Q8 , where n ≥ 1;
(ii) 3-group with exponent 3;
(iii) non-nilpotent group of order 2n .3 or 2.3m , where n, m > 1 with all non-trivial elements are of
order 2 or 3.
Proof. Let |G| has k distinct prime divisors.
Case 1. If k ≥ 3, then G contains at least one subgroup of order p ≥ 5, let it be H. Since H is a
subgroup of prime order, so the non-trivial elements in H are not adjacent to each other in P∗ (G).
The elements in G of order q (q 6= p), are not a power of any of the elements of H, and vise versa.
So P∗ (G) contains K1,3 as an induced subgraph.
Case 2. If k = 2, then |G| = pn q m where n, m ≥ 1. If at least one of p or q ≥ 5, then P∗ (G)
contains K1,3 as an induced subgraph, as by the argument used in Case 1. So we now assume that
both p, q < 5; without loss of generality, we can take p = 2, q = 3.
If m = n = 1, then G ∼
= Z6 or S3 . It is easy to see that P∗ (Z6 ) ∼
= K1,2 ∪ K2 , which is K1,3 -free.
P∗ (S3 ) is shown in Figure 1, which is K1,3 -free.
Now we assume that either n or m > 1. Suppose G contains an element x whose order is not a
prime, then o(x) = 2l (l > 1), 3s (s > 1) or 2l 3s (0 < l ≤ n, 0 < s ≤ m).
Subcase 2a. If o(x) = 2l , l > 1, then any three non-trivial elements of hxi are power of each other.
These three elements together with the element of order 3 forms K1,3 as an induced subgraph of
P∗ (G).
6
(1 2)
b
b
(1 2 3)
b
(2 3)
b
b
(1 3)
(1 3 2)
Figure 1: P∗ (S3 ).
Subcase 2b. If o(x) = 3s , s > 1, then as in Subcase 2a, we can show that P∗ (G) contains K1,3 as
an induced subgraph.
Subcase 2c. If o(x) = 2l 3s , where 0 < l ≤ n, 0 < s ≤ m, then hxi contains an element of order 6,
let it be y. Let X1 and X2 be subsets of hyi, where X1 contains two elements of order 6, and two
elements of order 3; and X2 contains one element of order 2, and two elements of order 6. Since
either n or m > 1, so G contains a subgroup of order ps , where p = 2 or 3 and s > 1, let it be H.
Suppose that H is cyclic. Then G contains an element of order ps , where p = 2 or 3 and s > 1. By
Subcases 2a and 2b, P∗ (G) contains K1,3 as a subgraph. If H is non-cyclic, then H contains more
than two cyclic subgroups of order p. Hence G contains an element of order p, which is not in hyi,
let it be z. Then the elements in X1 and z induces K1,3 as a subgraph of P∗ (G). Thus, it remains
to consider the case when all the non-trivial elements of G are of order either 2 or 3. If we assume
that G is such a group, then by [6], G must be non-nilpotent of order either 2n .3 or 2.3m , n, m > 1.
Moreover, the degree of each vertex in P∗ (G) is either |G| − 2 or |G| − 3, and so P∗ (G) is K1,3 -free.
Case 3. If k = 1, then |G| = pn , n ≥ 1.
Subcase 3a. If G is cyclic, then obviously
P∗ (Zpn ) ∼
= K pn −1 ,
(3.1)
which is K1,3 -free.
Subcase 3b. Assume that G is non-cyclic.
Subcase 3b(i). Let p = 2. If n = 2, then G ∼
= Z2 × Z2 , and so P∗ (G) ∼
= K3 , which is K1,3 -free.
Now we assume that n > 2. If G ∼
= Zn2 , then by Theorem 3.1, P∗ (G) ∼
= K2n −1 , which is K1,3 -free.
If G ∼
= Q8 , then P∗ (G) is as shown in Figure 2, which is K1,3 -free. If G ∼
= Q2n , n ≥ 4, then G
contains a cyclic subgroup of order 8, let it be H. Here H contains a unique subgroup of order
7
b
b
b
b
a3
a3 b
a
b
b
a2
b
b
a2 b
ab
Figure 2: P∗ (Q8 ).
4. But G contains at least two cyclic subgroup of order 4, so as by the argument used in Case 1,
P∗ (G) contains K1,3 as an induced subgraph.
Next, we assume that G Zn2 and Q2n . Then G contains an element of order 22 , let it be x
and so hxi contains a unique element of order 2. By Theorem 2.3 (ii) and (iii), G contains at least
three elements of order 2. Therefore, an element of order 2, which is not in hxi together with the
non-trivial elements in hxi forms K1,3 as an induced subgraph of P∗ (G).
Subcase 3b(ii). Let p 6= 2. Then by Theorem 2.3(i), G has a subgroup H ∼
= Zp × Zp . Then
H contains p + 1 subgroups of order p. Also any two of these subgroups have trivial intersection.
Hence each non-trivial element in any of these cyclic subgroups is not a power of any non-trivial
element in another cyclic subgroups of H. Hence,
P∗ (Zp × Zp ) ∼
= K(p + 1, p − 1).
(3.2)
If p ≥ 5, then by (3.2), P∗ (G) contains K1,3 as an induced subgraph. Now assume that p = 3.
If n = 2, then by (3.2), P∗ (G) ∼
= K(4, 2). Therefore P∗ (G) is K1,3 -free. Suppose that n > 2. If
G contains at least one element of order 32 , let it be x. Then hxi contains a unique subgroup of
order 3. By Theorem 2.3(ii) and(iii), G contains at least four subgroups of order 3. Then as in the
argument used in Case 1, the element y ∈
/ hxi of order 3 together with the non-trivial elements in
hxi forms K1,3 as an induced subgraph of P∗ (G). If all the elements in G are of order 3, that is, G
is a 3-group with exponent 3. Then degree of each vertex of P∗ (G) is 3n − 3, so P∗ (G) is K1,3 -free.
The proof follows by combining together all the above cases.
Theorem 3.3. Let G be a finite group. Then the following are equivalent:
(1) P∗ (G) is isomorphic to either Zpn or Zpqm , n, m ≥ 1;
8
(2) P∗ (G) is C3 -free;
(3) P∗ (G) is bipartite.
Proof. First we prove that (1) ⇔ (2):
Let |G| has k distinct prime divisors. Now we divide the proof into the following cases.
Case 1. If k = 1, then |G| = pn . Suppose G is cyclic, then by (3.1), P∗ (G) is totally disconnected.
Now, we assume that G is non-cyclic. If p > 2, then by Theorem 2.3(i), G contains a subgroup
isomorphic to Zp ×Zp . Then by (3.2), P∗ (G) contains C3 as a subgraph. Now, let p = 2. If G Qn ,
then by Theorem 2.3(ii) and (iii), G contains at least 3 elements of order 2, and so P∗ (G) contains
C3 as a subgraph. If G ∼
= Qn , then G contains at least three cyclic subgroups of order 4, and so
P∗ (G) contains C3 as a subgraph.
Case 2. If k = 2, then |G| = pn q m . Suppose G is cyclic, then the elements of order pn , q m and
pr q s , where 0 < r < n, 0 < s < m are not powers of one another. So they form C3 as a subgraph
of P∗ (G). Suppose that either n < 2 or m < 2. Without loss of generality, we assume that n = 1.
Then every element of order p is adjacent to the elements of order q s , 0 < s ≤ m; the elements of
order pq s , 0 < s ≤ m, are adjacent to the elements of order q t , t > s. So P∗ (G) does not contains
C3 as a subgraph. Suppose G is a non-cyclic abelian, then G contains a subgroup isomorphic to
either Zp × Zp or Zq × Zq , and so by (3.2), P∗ (G) contains C3 as a subgraph. Suppose that G is
non-abelian. If n = m = 1 and q > p, then G ∼
= Zq o Zp and it contains q Sylow p-subgroups, and
a unique Sylow q-subgroup. So
P∗ (Zq o Zp ) ∼
= K(q, p − 1) + K q−1 ,
(3.3)
which contains C3 as a subgraph. If either n > 1 or m > 1, then G contains a subgroups of order
pn and q m , let them be H and K respectively. If either H or K is non-cyclic, then by Case 1,
P∗ (G) contains C3 . If H and K are cyclic, then G contains elements of order pr , 0 < r ≤ n and
q s , 0 < s ≤ m. Let z be an element in G, which is not in H and K. If the order of z is pr , where
r ≤ n, then z together with an element of order pn and the element of order q forms C3 in P∗ (G).
Similarly, if the order of z is q s , where s ≤ m, then P∗ (G) contains C3 . If order of z is pr q s , then G
contains an element of order pq. This element together with an element of order pn and q m forms
C3 in P∗ (G).
Case 3. If k ≥ 3, then G contains at least three elements of distinct prime orders, and so they
forms C3 in P∗ (G).
9
Next, we show that (1) ⇒ (3): If G ∼
= Zpn , then by (3.2), P∗ (G) is bipartite. If G ∼
= Zpqm ,
m ≥ 1, then P∗ (G) is bipartite with bipartition X and Y , where X contains the elements of order
q s , 0 < s ≤ m, and Y contains the elements of order p and the elements of order pq s , 0 < s ≤ m.
Proof of (3) ⇒ (2) is obvious.
Combining all the above cases, we get the result.
Theorem 3.4. Let G be a finite group. Then P∗ (G) is disconnected if and only if G ∼
= Zn or Q2α .
In this case, the number of components of P∗ (Zn ) is n − 1, if n = pα ; ϕ(n) + 1, otherwise. The
number of components of P∗ (Q2α ) is 2.
Proof. First we assume that G ∼
= Zn . Then all the elements of G are powers of the generators
of G. So the generators of G are isolated vertices in P∗ (G). If n = pα , then by (3.1), number of
components of P∗ (G) is n − 1. Assume that n = pn1 1 pn2 2 . . . pnk k , where pi ’s are distinct primes and
ni ≥ 1 for all i and k > 1. Let x and y be non-generators of G such that they are non-adjacent in
P∗ (G). Then hxi ⊆ hyi or hyi ⊆ hxi. Without loss of generality, we assume that hyi ⊆ hxi. Since x
is a non-generators of G, so pni i - o(x) for some i. Hence x adjacent to the element of order pni i in
P∗ (G), say z. Then y is also adjacent to z, so x − z − y is a x − y path in P∗ (G). Thus, P∗ (G) has
ϕ(n) + 1 components.
If G ∼
= Q2α , then G contains one cyclic subgroup of order 2α−1 , say H and the remaining
elements are of order 4. Since the element of order 2, say x is a power of all the elements of G. So
this element is an isolated vertex in P∗ (G). Hence all the elements in G of order 4, which is not
in H are adjacent to all the non-trivial elements of H, except x. So the number of components of
P∗ (G) is 2.
Now we assume that G Zn and Q2α . We have to show that P∗ (G) is connected. Let
|G| = pn1 1 pn2 2 · · · pnk k , where pi ’s are distinct primes. We need to consider the following two cases.
i
Case 1. Let k ≥ 2. For each i = 1, 2, . . . , k, Let Xi = {x ∈ G | o(x) = pm
i , 0 < mi ≤ ni }. Then
each element x ∈ Xi is adjacent to all the elements in Xj , i 6= j. So the subgraph induced by the
k
k
[
[
elements of
Xi is connected. Now let x ∈ G with x ∈
/
Xi . Since G is non-cyclic, so pni i - o(x)
i=1
for some i. Let H be Sylow pi -subgroup of G.
i=1
Subcase 1a. If H is cyclic, then G contains an element of order pni i , say z. Then z is not a power
S
of x. So z ∈ ki=1 Xi , and is adjacent to x in P∗ (G).
Subcase 1b. Let H be non-cyclic. If H is non-quarternion, then by Theorem 2.3 (ii) and (iii), G
contains more than two cyclic subgroups of order p. So x is adjacent to the elements of order p,
10
which are not in hxi. If H is quarternion, then H contains more than two cyclic subgroup of order
4, so x is adjacent to the elements of order 4, which are not in hxi. In both the cases there exist
k
[
z∈
Xi , which adjacent to x in P∗ (G).
i=1
Case 2. Let k = 1. Since G is non-cyclic and non-quaternion, so by Theorem 2.3(ii) and (iii), G
contains more than two cyclic subgroups of order p1 , let them be Hi := hzi i, i = 1, 2, . . . , r, for
some r ≥ 3. Then each non-trivial element in Hi is not a power of any non-trivial element in Hj ,
k
[
i 6= j. Hence the subgraph induced by an elements in
Hi is connected. Now, let x ∈ G, with
i=1
x∈
/
k
[
Hi . Then hxi contains exactly one zi , so any zj , j 6= i is not a power of x; since o(x) > o(zj ),
i=1
so x is also not a power of zj . Thus x is adjacent to zj ∈
k
[
Hi .
i=1
From the above arguments, it follows that P∗ (G) is connected. This completes the proof.
From the proof of the previous theorem, we deduce the following result. Note that this result
follows directly from [19, Lemma 8]. Here we obtain this as a consequence of the previous theorem.
Corollary 3.1. Let G be a finite group. Then P∗ (G) contains isolated vertices if and only if G ∼
= Zn
or Q2m . Moreover, the number of isolated vertices in P∗ (Zn ) is n − 1, if n = pα ; ϕ(n), otherwise.
The number of isolated vertices in P∗ (Q2m ) is 1.
Theorem 3.5. Let G be a finite group. Then diam(P∗ (G)) is ∞, if
G ∼
= Zn or Qn ; 1, if
G∼
= Zm
2 , m ≥ 1; 2, otherwise.
Proof. The possibilities of G with diam(P∗ (G)) is either ∞ or 1 follows from Theorems 3.4 and 3.1,
respectively. Now we assume that G Zn , Qn or Zm
2 , m ≥ 1. Let G has k distinct prime factors.
Case 1. Let k = 1. Since G Zn and Q2α , so by parts (ii), (iii) of Theorem 4.1, G contains at
least three subgroups of prime order, let them be hxi i , i = 1, 2, . . . , r, where r ≥ 3. Now, let x be
a non-trivial element in G. Then hxi contains exactly one hxi i, for some i. It follows that, every
xj (j 6= i) is not a power of x, and vice versa, so x is adjacent to all xj (j 6= i). Now let u, v be
non-trivial elements in G. Then hui and hvi contains a subgroup of prime order, let them be hxr i
and hxs i, respectively. If hxr i =
6 hxs i, then u and v are not a powers of each other. So u and v are
adjacent in P∗ (G). If hxr i = hxs i, then there exist xl (l 6= i), which is adjacent to both u and v in
P∗ (G). So u − xl − v is a u − v path in P∗ (G).
Case 2. Let k ≥ 2. Let |G| = pn1 1 pn2 2 . . . pnk k , where pi ’s are distinct primes and ni ≥ 1 for all i. Let
x and y be not adjacent vertices in P∗ (G). Then hxi ⊆ hyi or hyi ⊆ hxi. Without loss of generality,
11
we assume that hyi ⊆ hxi. Since G is non-cyclic, so pni i - o(x) for some i. Then by Subcases 1a and
1b in proof of Theorem 3.4, x is adjacent to the element of order pni i or pi , let that element be z.
Then y is also adjacent to z, so x − z − y is x − y path in P∗ (G)
Theorem 3.6. If G is a finite group, then gr(P∗ (G)) is ∞, if G ∼
= Zpn or Z2p ; 4, if G ∼
= Zpqm ,
where q m 6= 2; 3, otherwise.
Proof. If G Zpn , and Zpqm , n, m ≥ 1, then by Theorem 3.3, G contains C3 as a subgraph. If
G∼
= Zpn , then by (3.1), P∗ (Zpn ) is acyclic. Now let G ∼
= Zpqm . If m ≥ 1, then by Theorem 3.3,
P∗ (Zpqm ) is bipartite; If m > 1, then G contains C4 as a subgraph of P∗ (G); If m = 1, then the
non-trivial elements of G are of orders one of p, q, pq. The elements of order pq are generators of G,
and hence they are isolated vertices in P∗ (G). Also the elements of order p and q are not a power
of one another. Therefore,
P∗ (Zpq ) ∼
= Kp−1,q−1 ∪ K (p−1)(q−1) ,
(3.4)
which is acyclic, when p = 2, and it contains C4 , when p > 2. The proof follows from these
facts.
4
Embedding of the complement of proper power graphs of groups
on topological surfaces
The main results we prove in this section are the following:
Theorem 4.1. Let G be a finite group and p be prime. Then
(1) P∗ (G) is planar if and only if G is one of Zpα , Z12 , Z2p , Z3p , Z2 × Z2 , Q8 , S3 ;
(2) P∗ (G) is toroidal if and only if G is one of Z18 , Z20 , Z28 , Z3 × Z3 , Z2 × Z2 × Z2 , Z4 × Z2 , D8 ;
(3) P∗ (G) is projective if and only if G is one of Z20 , Z4 × Z2 , D8 .
As a consequence of this result, we deduce the following:
Corollary 4.1. Let G be a finite group. Then
(1) P∗ (G) is neither a path nor a star;
(2) P∗ (G) is Cn if and only if n = 3 and G ∼
= Z2 × Z2 .
12
(3) P∗ (G) does not contain K1,4 as a subgraph if and only if G is either Zpn , Z6 or Z2 × Z2 ;
(4) The following are equivalent:
(a) G is one of Zpn , Z2 × Z2 or Z2p ;
(b) P∗ (G) is outerplanar;
(c) P∗ (G) does not contain K2,3 as a subgraph.
First, we begin with the following result.
Proposition 4.1. If G is a finite group whose order has more than two distinct prime divisors,
then γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Proof. Let |G| = pn1 1 pn2 2 · · · pnk k , where pi ’s are distinct primes, ni ≥ 1 and k ≥ 3. We divide the
proof in to the following cases:
Case 1. If k = 3, then without loss of generality, we assume that p1 < p2 < p3 . Let us consider
the following subcases.
Subcase 1a. If p1 > 2, then G contains at least two elements of order p1 , at least four elements of
order p2 , and at least six elements of order p3 . Then the elements of order pα1 1 (0 < α1 ≤ n1 ) and
pα2 2 (0 < α2 ≤ n2 ) are adjacent to the elements of order pα3 3 (0 < α3 ≤ n3 ) in P∗ (G). Hence P∗ (G)
contains K5,6 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 1b. Let p1 = 2. If p2 > 3, then G contains at least four elements of order p2 , and at
least six elements of order p3 . Then the elements of order 2 and p2 are adjacent to the elements of
order p3 . Thus P∗ (G) contains K5,4 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If p2 = 3, and either n2 ≥ 2 or n3 ≥ 2, then P∗ (G) contains K3,7 as a subgraph, so γ(P∗ (G)) > 1
and γ(P∗ (G)) > 1. Now, we assume that n2 = n3 = 1. Suppose for some i (i = 1, 2, 3), the Sylow
pi -subgroup is not unique. If i = 1, then G contains at least three elements of order 2. Then the
elements of order 2 and 3 are adjacent to the elements of order p3 . Hence P∗ (G) contains K5,4 as
a subgraph. If i = 2 or 3, then G contains at least 8 elements of order pi , so P∗ (G) contains K3,7
as a subgroup. Suppose for each i, Sylow pi - subgroup of G is unique, then G ∼
= P × Z3 .p3 , where
P is the Sylow 2-subgroup of order 2n1 . If n1 = 1, then G ∼
= Z6 .p3 . In this case, P∗ (G) contains
K5,4 as a subgraph. If n1 ≥ 1, then P∗ (G) contains K3,7 as a subgraph. In both the cases, we have
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Case 2. Let k ≥ 4. Let pi , pj , pr > 2, for some i, j, r. Then the elements of order pi and pj are
13
adjacent to the elements of order pr in P∗ (G). Thus P∗ (G) contains K3,7 as a subgraph, and so
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Proof follows by combining the above cases together.
Proposition 4.1 reveals that, to prove the main result, it is enough to deal with the groups
whose order has at most two distinct prime divisors. For this purpose, first we consider the finite
cyclic groups, then we deal with the finite non-cyclic groups.
Proposition 4.2. Let G be a finite cyclic group and p be a prime. Then
(1) P∗ (G) is planar if and only if G is one of Zpα , Z12 , Z2p , Z3p ;
(2) P∗ (G) is toroidal if and only if G is one of Z18 , Z20 , Z28 ;
(3) P∗ (G) is projective if and only if G ∼
= Z20 .
Proof. Let |G| has k distinct prime divisors. Now we divide the proof into the following cases.
Case 1. If k = 1, then by (3.1), P∗ (G) is planar.
Case 2. Let k = 2. Let H and K be subgroups of Zpn qm of order pn and q m , respectively, where
n, m ≥ 1. The order of each non-trivial element in H is relatively prime to the non-trivial elements
in K. So no element in H is not a power of any element in K and vice versa. Therefore, P∗ (G)
contains Kpn −1,qm −1 as a subgraph. If n, m ≥ 2, then P∗ (G) contains K3,7 as a subgraph, so
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. Now, we assume that either n = 1 or m = 1. Without loss of
generality, we assume that m = 1. Then |G| = pn q, n ≥ 1. We need to consider the following
subcases:
Subcase 2a. If n = 1, then G ∼
= Zpq . By (3.4), γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1 if p, q ≥ 5;
otherwise, P∗ (G) is planar.
Subcase 2b. Let n ≥ 2.
Subcase 2b(i). Let p = 2. If n > 2, then G contains four elements of order 8, which are not a
power of any elements of orders q, 2q, 4q and vice versa. Hence P∗ (G) contains K4,5 as a subgraph.
Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If n = 2, then in G, the elements of order 2, 4 are not a power of elements of order q and vice
versa. Also the element of order 2 is a power of the elements of order 2q; the elements of order 2q
and 4 are not a power of each other. It follows that if q = 3, then P∗ (G) is planar, and a plane
embedding of P∗ (G) is shown in Figure 3.
14
9
b
b
1
b
b
b
7
b
4
6
5
b
2
10
b
b
b
3
11
b
8
Figure 3: A plane embedding of P∗ (Z12 ).
4
10b
b
5
b
13
11
b
b
18
b
7
b
b
b
b
3
6
b
b
12
b
16
14
b
16
10
b
b
b
19
12
8
b
1
b
15
b
b
b
2
9
b
b
b
10
4
8
b
17
b
10
Figure 4: A toroidal embedding of P∗ (Z20 ).
.
If q = 5, then γ(P∗ (G)) = 1, γ(P∗ (G)) = 1, and a toroidal and a projective embeddings of P∗ (G)
are shown in Figures 4 and 5, respectively. If q = 7, then P∗ (G) contains K3,6 as a subgraph, so
γ(P∗ (G)) > 1, γ(P∗ (G)) = 1 and a toroidal embedding of P∗ (G) is shown in Figure 6. If q > 7,
then P∗ (G) contains K3,7 as a subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2b(ii). Let p = 3. If n > 2, then G contains eighteen elements of order 27, which are
not a power of any elements of orders q, 3q, 9q, and vice versa. Hence P∗ (G) contains K4,5 as
a subgraph. Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If n = 2, then in G, the elements of
order 3, 9 are not a power of elements of order q, and vice versa. Also the element of order 3 is a
power of the elements of order 3q; the elements of order 3q and 9 are not a power of each other. It
follows that if q = 2, then P∗ (G) contains K3,6 as a subgraph, so γ(P∗ (G)) > 1; but γ(P∗ (G)) = 1,
and a toroidal embedding of P∗ (G) is shown in Figure 7. If q ≥ 5, then P∗ (G) contains K5,4 as a
subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
15
b
7
4
b
19
b
3
b
16
b
b
b
b
b
b
b
2
1
9
b
8
6
b
14
11
b
18
12
b
b
15
17
10
b
5
b
b
13
Figure 5: A projective embedding of P∗ (Z20 ).
.
Subcase 2b(iii): If p ≥ 5, then the elements in G of order q and pq are not a power of the elements
of order p2 , and vise versa. Note that G contains at least one element of orders q, at least four
elements of order pq, and at least twenty elements of order p2 . It follows that P∗ (G) contains K3,7
as a subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Case 3. Let k ≥ 3. Then by Proposition 4.1, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Combining all the above cases together, the proof follows.
Proposition 4.3. Let G be a finite non-cyclic group of order pα , where p is a prime and α ≥ 2.
Then
(1) P∗ (G) is planar if and only if G is either Z2 × Z2 or Q8 ;
(2) P∗ (G) is toroidal if and only if G is one of Z3 × Z3 , Z2 × Z2 × Z2 , Z8 × Z2 or D8 ;
(3) P∗ (G) is projective if and only if G is one of Z4 × Z2 or D8 .
Proof. We divide the proof into several cases.
Case 1. Let α = 2. Then G ∼
= Zp × Zp . If p ≥ 5, then by (3.2), P∗ (G) contains two copies of K3,3 ,
and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If p = 3, then by (3.2), P∗ (G) ∼
= K2,2,2,2 . Here P∗ (G)
contains K4,4 as a subgraph, so γ(P∗ (G)) > 1; but γ(P∗ (G)) = 1. A toroidal embedding of P∗ (G)
is shown in Figure 8. If p = 2, then
P∗ (G) ∼
= C3 ,
(4.1)
16
14
4
b
b
1
7b
b
12
b
25
5
b
23
b
b
16
b
b
b
19
18
b
b
22
b
b
26
11
24
b
12
b
16
9
b
b
27
b
b
14
b
13
15
2
b
b
b
20
17
b
10
b
b
6
b
b
14
8
b
3
21
b
b
b
4
8
14
Figure 6: A toroidal embedding of P∗ (Z28 ).
9
2
b
b
8
10
b
b
b
7
6
11 b
b
b
9b
b
1
b
b
4
b
3
16
14
15
13
b
b
b
b
12
5
b
17
b
8
10
b
b
b
b
b
9
2
4
9
Figure 7: A toroidal embedding of P∗ (Z18 ).
which is planar.
Case 2. Let α = 3.
Subcase 2a. Assume that p ≥ 3. Then up to isomorphism the only non-cyclic groups of order p3
are Zp × Zp × Zp , Zp2 × Zp , (Zp × Zp ) o Zp and Mp3 .
Subcase 2a(i). If G ∼
= Zp × Zp × Zp , then G contains p2 + 1 subgroups of order p. This implies
that Kp2 +1 is subgraph of P∗ (G). Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2a(ii). If P∗ (G) ∼
= Zp2 × Zp , then G contains p + 1 subgroups of order p, let them be Hi ,
i = 1, 2, . . . , p + 1. Also G contains p cyclic subgroups of order p2 , let them be Ni , i = 1, 2, . . . , p.
Moreover, all these subgroups contains the unique subgroups of order p, without loss of generality,
let it be H1 . Then P∗ (G) contains Kp(p−1),p(p2 −1) as a subgraph with the bipartition X and Y , where
X contains all the elements of order p in Hi , i = 2, 3, . . . , p + 1; Y contains all the non-identity
elements in Ni , i = 1, 2, . . . , p. This implies that P∗ (G) contains K4,5 , and so γ(P∗ (G)) > 1,
17
(1,0)
b
(1,2)
b
(2,1)
(2,2)
b
b
b
b
(1,0)
b
(2,0)
b
(1,2)
(0,1)
(0,2) b
b
(1,1)
b
b
(1,0)
(2,1)
b
(2,2)
b
(1,1)
b
(1,0)
Figure 8: A toroidal embedding of P∗ (Z3 × Z3 ).
.
γ(P∗ (G)) > 1.
Subcase 2a(iii). If G ∼
= (Zp × Zp ) o Zp , then G contains p2 subgroups of order p. Then the
subgraph of P∗ (G) induced by the set having one element of order p from each of these subgroups
forms Kp2 , so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2a(iv). If G ∼
= Mp3 , then the subgroup lattice of Mp3 is isomorphic to the subgroup
lattice of Zp2 × Zp , so by the above argument, we have γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2b. If p = 2, then upto isomorphism the only non-cyclic subgroup of order 8 are
Z2 × Z2 × Z2 , Z4 × Z2 , D8 and Q8 .
Subcase 2b(i). If G ∼
= Z2 × Z2 × Z2 , then the order of each element of G is 2. It follows that
P∗ (G) ∼
= K7 , so γ(P∗ (G)) > 1 and γ(P∗ (G)) = 1.
Subcase 2b(ii). If G ∼
= Z4 × Z2 , then G contains the elements (1, 0), (3, 0), (1, 1), (3, 1) of order 4,
and the elements (2, 0), (0, 1), (2, 1) of order 2. Here (2, 0) is a power of each of (1, 0), (3, 0), (1, 1), (3, 1);
(3, 0) is a power of (1, 0); (3, 1) is a power of (1, 1). Also no two remaining elements of G are
power of one another. Hence P∗ (G) contains K3,3 with bipartition X := {(1, 0), (3, 0), (2, 1)} and
Y := {(1, 1), (3, 1), (0, 1)}, so P∗ (G) is non-planar. Also γ(P∗ (G)) = 1, γ(P∗ (G)) = 1; a toroidal
and a projective embedding of P∗ (G) is shown in Figure 9 and 10, respectively.
Subcase 2b(iii). If G ∼
= D8 , then b, ab, a2 b, a3 b, a2 are the elements of order 2, and a, a3 are the
elements of order 4; these are the only elements of D8 . Here hai = a3 and it contains a2 , so a, a2 ,
a3 are not adjacent to each other. Also any two remaining elements of G are not a power of one
another. Thus a2 , b, ab, a2 b, a3 b forms K5 as a subgraph of P∗ (G), so P∗ (G) non-planar. Further,
γ(P∗ (G)) = 1, γ(P∗ (G)) = 1; a toroidal and projective embedding of P∗ (G) is shown in Figure 11
18
(1,0)
(0,1)
b
b
b
(3,1)
(1,1)
(1,0)
(2,1)
b
b
b
b
b
(3,0)
(1,1)
b
b
b
(1,0)
(0,1)
(2,0)
b
b
(1,0)
(2,1)
Figure 9: A toroidal embedding of P∗ (Z4 × Z2 ).
.
(0,1)
b
b
(1,2)
(2,0)
b
(1,0)
b
(3,1)
b
b
(1,1)
b
(3,0)
Figure 10: A projective embedding of P∗ (Z4 × Z2 ).
and 12.
Subcase 2b(iv). If G ∼
= Q8 , then by Figure 2, P∗ (G) is planar.
Case 3. Let α ≥ 4.
Subcase 3a. Let p = 2. If α = 4, then up to isomorphism there are four non-cyclic abelian groups
of order 24 , and nine non-abelian groups of order 24 . In the following, first we deal with these
non-cyclic abelian groups:
If G ∼
= Z4 × Z4 , then G contains six cyclic subgroups, say Hi , i = 1, 2, . . . , 6 of order 4. Hence
P∗ (G) contains K6,6 as a subgraph with the bipartition X, Y , where X contains all the elements
of order four in H1 , H2 and H3 ; Y contains all the elements of order four in H4 , H5 and H6 . Thus
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
19
a2
b
b
b
a2 b
a3 b
a2
ab
b
b
b
b
b
a
a3b
b
b
a2
b
a2 b
a3 b
b
b
b
b
ab
a2
Figure 11: A toroidal embedding of P∗ (D8 ).
b
b
b
ab
a
b
b
a2 b
b
a3
b
a3 b
b
a2
Figure 12: A projective embedding of P∗ (D8 ).
If G ∼
= Z8 × Z2 , then G contains two cyclic subgroups of order 8, let them be H1 , H2 ; two cyclic
subgroups of order 4, let them be N1 , N2 ; three elements of order 2, let them be xi , i = 1, 2, 3.
Here H1 , H2 contains a cyclic subgroup of order 4, and an element of order 2 in common, without
loss of generality, let them be N1 and x1 respectively. So the elements of order 4 in N2 is not a
power of any non-trivial elements in H1 , H2 . Also x2 , x3 are not elements of H1 , H2 , so they are
not a power of any non-trivial elements in H1 , H2 and vice versa. Hence P∗ (G) contains K7,4 as a
subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Z4 × Z2 × Z2 , then G contains a subgroup isomorphic to Z2 × Z2 × Z2 . So G contains
seven elements of order 2; (2, 0, 0) is one among these elements, and is a power of each of the
elements (1, 0, 0), (1, 1, 0), (1, 0, 1), (1, 1, 1) of order 4. But the remaining six elements of order 2
are not power of these four elements of order 4. Hence P∗ (G) contains K5,6 as a subgraph with
20
bipartition X and Y , where X = {(1, 0, 0), (1, 1, 0), (1, 0, 1), (1, 1, 1), (2, 0, 0)} and Y contains the
remaining six elements of order 2. Thus γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Z2 ×Z2 ×Z2 ×Z2 , then all the non-trivial elements of G are of order 2 and so P∗ (G) ∼
= K11 .
Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Next, we investigate the nine non-abelian groups of order 24 .
If G ∼
= (Z4 × Z2 ) o Z2 , then G contains four cyclic subgroups of order 4, let them be H1 , H2 , H3 ,
H4 . Among these H1 , H2 contains a unique element of order 2 in common, and H3 , H4 contains
a unique elements of order 2 in common. But G contains exactly seven elements of order 2. So
the remaining five elements in G of order 2 are not a power of any non-trivial elements in Hi ,
i = 1, 2, 3, 4. Hence P∗ (G) contains K5,8 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Z4 o Z4 , then G contains six cyclic subgroups of order 4. Let them be Hi , i = 1, 2, . . . , 6.
It follows that P∗ (G) contains K6,4 with bipartition X and Y , where X contains all the elements
of order four in H1 , H2 , H3 ; Y contains all the elements of order four in H4 , H5 . Therefore
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Z8 o5 Z2 , then G contains two cyclic subgroups of order 8, let them be H1 , H2 , these two
cyclic subgroups contains a unique element of order 2 in common. But G contains three elements
of order 2, so the remaining two elements of order 2 are not a power of non-trivial elements of H1 ,
H2 . It follows that P∗ (G) contains K6,4 as a subgraph and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= D16 , then G contains nine elements of order 2, so G contains K9 as a subgraph. Therefore,
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Z8 o3 Z2 , then G contains three subgroups of order 4, let them be H1 , H2 , H3 . These
subgroups contains a unique element of order 2 in common. But G contains five elements of order
2. So the remaining four elements of order 2 are not a power of any non-trivial elements in Hi ,
i = 1, 2, 3. It follows that G contains K6,4 as a subgraph and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Q16 , then G contains five cyclic subgroup of order 4, let them be Hi , i = 1, . . . , 5. Then
P∗ (G) contains K6,4 as a subgraph with bipartition X and Y , where X contains elements of order 4
in H1 , H2 , H3 , and Y contains elements of order 4 in H4 , H5 . So γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= D8 ×Z2 , then G contains eleven elements of order 2. Hence they forms K11 as a subgraph
of P∗ (G) . It follows that γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= Q8 × Z2 , then G contains six cyclic subgroups of order 4, let them be Hi , i = 1, . . . , 6.
Hence P∗ (G) contains K6,6 as a subgraph with bipartition X and Y , where X contains elements of
order 4 in H1 , H2 and H3 , Y contains elements of order four in H4 , H5 and H6 . Thus, γ(P∗ (G)) > 1
21
and γ(P∗ (G)) > 1.
If G ∼
= Q8 o Z2 , then G contains seven elements of order 2 and four cyclic subgroups of order 4.
Each of these cyclic subgroups contains exactly one element of order 2 in common, let it be x. Then
P∗ (G) contains K3,8 as a subgraph with bipartition X and Y , where X contains all the elements
of order 2 in G except x, and Y contains all the elements of order 4 in G. So γ(P∗ (G)) > 1 and
γ(P∗ (G)) > 1.
Assume that α ≥ 5. Then G must contain a non-cyclic subgroup of order 2α−1 . For suppose
all the subgroup of order 2α−1 are cyclic, let H, K be two subgroups among these. Since H
is a subgroup of prime index, so H is normal in G. It follows that HK is a subgroup of G. If
|H ∩K| < 2α−2 , then |HK| > |G|, which is not possible. So |H ∩K| must be 2α−2 . It follows that H
and K contains a common subgroup of order 2α−2 . Hence G has a unique subgroup of order 2α−2 .
Then by Theorem 2.3(ii), G must be cyclic, which is a contradiction to our hypothesis. Let this
subgroup of G of order 2α−1 be H. Then by previous argument, γ(P∗ (H)) > 1 and γ(P∗ (H)) > 1,
and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 3b. Let p ≥ 3. Then by Theorem 2.3(i), G contains a non-cyclic subgroup H of order
pα−1 . So by Case 2, γ(P∗ (H)) > 1 and γ(P∗ (H)) > 1, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Proof follows by combining all the cases together.
Next, we consider the groups whose order has exactly two distinct prime factors. First, we
prove the following lemma, which is used later in the proof of Proposition 4.4. We consider the
0
u,u
graph K3,3,3
obtained from the complete tripartite graph K3,3,3 by adding two new vertices u and
u0 and six edges uvi , u0 vi , i = 1, 2, 3, where each of the vertices v1 , v2 , v3 comes from a different
part of K3,3,3 , i.e. vertices v1 , v2 , v3 induce a triangle in K3,3,3 . Notice that the tripartite graph
K3,3,3 can also be considered as the complement of three vertex-disjoint triangles, which is denoted
0
u,u
by 3K3 in [10]. We need to determine whether K3,3,3
is toroidal or not.
0
u,u
Lemma 4.1. Graph K3,3,3
is non-toroidal.
Proof. We use an approach similar to the proofs of Lemma 4.4 and Proposition 4.6 in [9]. First,
we show that the graph K3,3,3 = 3K3 has a unique embedding on the torus, which is a triangulation of the torus. This is shown computationally by using an exhaustive search method in [10]
(see also some corrected computational results for [10] at http://www.combinatorialmath.ca/
G&G/TorusMaps.html). Then, we are going to show that it is not possible to extend this unique
0
u,u
embedding of K3,3,3 to an embedding of K3,3,3
on the torus.
22
From the Euler’s formula for the torus, we have n − m + f = 0, where n, m, and f are,
respectively, the numbers of vertices, edges, and faces of a 2-cell embedding on the torus (e.g., see
[10]). Since K3,3,3 has n = 9 vertices and m = 27 edges, an embedding of K3,3,3 on the torus must
have 18 faces, i.e. f = 18. Since each edge of an embedding appears either exactly once on the
boundaries of two separate faces or two times on the boundary of the same face, and each face of
an embedding is bounded by at least three edges, we have 3f ≤ 2m. The equality 3f = 2m is
possible only when each face is a triangle, which is the case of K3,3,3 on the torus. Therefore, an
embedding of K3,3,3 on the torus must be a triangulation.
On the other hand, there are exactly two different embeddings of K3,3 on the torus (e.g., see
Figure 8 in [10]), which are to be extended to an embedding of K3,3,3 by adding into the 2-cell faces
three new vertices adjacent to all six vertices of the original K3,3 . One of these embeddings of K3,3
contains faces with only four vertices on their boundary: clearly, it is not possible to triangulate
such a 4-vertex face by adding a new vertex adjacent to all six vertices of K3,3 . However, the other
embedding of K3,3 on the torus has each of its three faces containing all six vertices of K3,3 on
the face boundary. Adding a new vertex into each 6-vertex face and making it adjacent to all the
vertices on the face boundary provides a triangulation of the torus by K3,3,3 . By this construction
and symmetries of the embedding of K3,3 with the hexagonal faces, the embedding of K3,3,3 on the
torus is unique.
Figure 13: The unique embedding of K3,3,3 on the torus.
Now, each face of the toroidal embedding of K3,3,3 contains exactly three vertices, one from
each part of K3,3,3 , no two faces have the same three vertices on the face boundary, and there are
23
0
u,u
no vertices repeated on any face boundary (see Figure 13). Each of the vertices u or u0 of K3,3,3
separately can be added with its three incident edges into any of the 18 faces of this embedding of
K3,3,3 without edge crossings. However, u and u0 must be adjacent to the same three vertices on
the boundary of a triangular face. There are no two different faces with the same three vertices
on the boundary, and there are no faces with vertices repeated on the boundary. Therefore, after
adding one of u or u0 into a face of the embedding of K3,3,3 and connecting it with edges to all
three vertices on the face boundary, there is no face containing the same three vertices to add the
other vertex without edge crossings. Thus, it is not possible to extend this unique embedding of
0
0
u,u
u,u
K3,3,3 to an embedding of K3,3,3
on the torus, and K3,3,3
is non-toroidal.
Proposition 4.4. If G is a non-cyclic group of order pn q m , where p, q are distinct primes and
n, m ≥ 1. Then
(1) P∗ (G) is planar if and only if G ∼
= S3 ;
(2) γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1, if G S3 .
Proof. In the proof of Proposition 4.2, we have noticed that, to prove this result, it is enough to
consider the finite non-cyclic group G of order pn q, n ≥ 1.
Case 1. Let n = 1. Without loss of generality, we assume that p < q, then G ∼
= Zq o Zp . If
(p, q) = (2, 3), then G ∼
= S3 . By Figure 1, P∗ (S3 ) is planar. If (p, q) 6= (2, 3), then by (3.3), P∗ (G)
contains K5,4 as a subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Case 2. Let n = 2.
Subcase 2a. If G is abelian, then G ∼
= Zp × Zpq .
First we assume that p = 2. If q = 3, the structure of P∗ (G) is shown in Figure 14. Since
0
u,u
this graph is isomorphic to the graph K3,3,3
described in Lemma 4.1, it follows that P∗ (G) is non-
toroidal. Further, in Figure 14, we notice that K3,6 is a subgraph of P∗ (G) and so γ(P∗ (G)) > 1. If
q = 5, then G contains three cyclic subgroups of order 10. This implies that P∗ (G) contains K4,8 ,
and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If q ≥ 7, then G contains at least two cyclic subgroups of
order q. It follows that P∗ (G) contains K6,6 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Next, we assume that p = 3. Then G contains Z3 × Z3 as a subgraph. So G contains four cyclic
subgroups of order 3, let them be H1 , H2 , H3 , H4 . Let for each i = 1, 2, 3, 4, hi , h0i be the elements
of Hi of order 3. Also G contains an element of order q, say x. Then P∗ (G) contains K5,4 as a
subgraph with bipartition X := {h1 , h2 , h01 , h02 , x} and Y := {h3 , h4 , h03 , h04 }.
24
(0,4)
(0,2)
b
b
(1,3)
b
(0,3)
b
b
(0,5) b
(0,1)
(1,0)
b
b
(1,1)
b
(1,2)
b
(1,4)
b
(1,5)
Figure 14: The structure of P∗ (Z2 × Z6 ).
If p ≥ 5, then G contains Zp × Zp as a subgraph, so by (3.2), γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2b. Let G be non-abelian.
Subcase 2b(i). Let p = 2.
If q = 3, then G ∼
= Z3 o Z4 , D12 or A4 . If G ∼
= Z3 o Z4 , then G contains three cyclic subgroups,
say H1 , H2 , H3 of order 4; unique cyclic subgroup K of order 6, and unique element of order 2. It
follows that P∗ (G) contains K5,4 as a subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If G ∼
= D12 ,
then G contains seven elements of order 2. These elements together with the element of order 3
forms K8 as a subgraph of P∗ (G), so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If G ∼
= A4 , then G contains
eight elements of order 3, and three elements of order 2. It follows that P∗ (G) contains K3,8 as a
subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If q = 5, then G ∼
= Z5 o Z4 , Z5 o (Z2 × Z2 ) or D20 . If G ∼
= Z5 o Z4 , then G contains five elements
of order 2, and four elements of order 5. This implies that P∗ (G) contains K5,4 as a subgraph, so
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If G ∼
= Z5 o (Z2 × Z2 ), then G contains four cyclic subgroups of
order 4, and a unique subgroup of order 5. Therefore, P∗ (G) contains K4,8 , so γ(P∗ (G)) > 1 and
γ(P∗ (G)) > 1. If G ∼
= D20 , then G contains eleven elements of order 2, and so P∗ (G) contains K11
as a subgraph. Hence γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If q = 7, then G ∼
= Z7 o Z4 or D28 . If G ∼
= Z7 o Z4 , then G contains seven cyclic subgroups
of order 4, and six elements of order 7. It follows that P∗ (G) contains K5,4 as a subgraph. Thus
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If G ∼
= D28 , then G contains fourteen elements of order 2 and so
P∗ (G) contains K5,4 as a subgraph. Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If q > 7, then G contains at least two subgroups of order 4 and q, let them be H1 , H2 respectively.
25
Also each non-trivial element of H1 is adjacent to all the non-trivial elements of H2 . It follows that
P∗ (G) contains K3,8 as a subgraph, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2b(ii). Let p = 3.
(i) If q = 2, then G ∼
= S3 × Z3 , D18 or Z3 × Z3 o Z2 .
If G ∼
= S3 × Z3 , then G contains a subgroup of order 9, let it be H, and three elements of order
2. These three elements of order 2 are adjacent to all the non-trivial elements of H. This implies
that P∗ (G) contains K3,8 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G ∼
= D18 or Z3 × Z3 o Z2 , then in either case P∗ (G) contains nine elements of order 2. This
implies that P∗ (G) contains K9 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
(ii) If q 6= 2, then G contains a subgroup of each of order 9 and q, let them be H1 , H2 ,
respectively. Also every non-trivial element of H1 are adjacent to all the non-trivial elements of
H2 . It follows that P∗ (G) contains K8,4 as a subgraph and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 2b(iii). Let p ≥ 5.
Then G contains a subgroup of order p2 , let it be H. If H is non-cyclic, then by (3.2),
γ(P∗ (H)) > 1 and γ(P∗ (H)) > 1, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. Suppose H is cyclic,
then every element of H of order p2 are not a power of any element which not in H and vice versa.
Thus P∗ (G) contains K3,7 as a subgraph, and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Case 3. Let n ≥ 3.
Subcase 3a. If G is abelian, then G contains a subgroup isomorphic to Zpq × Zp × Zp or Zpq × Zp2 .
If G contains a subgroup isomorphic to Zpq × Zp × Zp , then G contains p2 + 1 cyclic subgroups
of order p, let them be Hi , i = 1, 2, . . . , p2 + 1. Also G contains a cyclic subgroup of order pq, let
it be K. Clearly K contains a unique subgroup of order p, so without loss of generality, let it be
H1 . Then Hi (i 6= 1) are not subgroups of K. So each non-trivial element of K is adjacent to
all the non-trivial elements in Hi (i 6= 1). It follows that P∗ (G) contains K3,7 as a subgraph, so
γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If G contains a subgroup isomorphic to Zpq × Zp2 , then G contains p + 1 cyclic subgroups of
order p, let them be Hi , i = 1, 2, . . . , p + 1, and two cyclic subgroups of order p2 , let them be N1 ,
N2 . These two subgroups contains a unique subgroup of order p in common, so without loss of
generality, let it be H1 . Then each non-trivial element of N1 , N2 are adjacent to all the non-trivial
elements in Hi (i 6= 1). Also G contains a subgroup of order pq. It follows that P∗ (G) contains K3,7
as a subgraph with bipartition X and Y , where X contains elements of order p2 in N1 and N2 ; Y
contains elements of order p in Hi (i 6= 1), the elements of order q, and pq in G. So γ(P∗ (G)) > 1
26
and γ(P∗ (G)) > 1.
Subcase 3b. Let G be non-abelian.
Subcase 3b(i). Let p = 2.
(i) If q = 3 and n = 3, then G contains a subgroup of order 8, let it be H. But the only groups
of order 8 are Z2 × Z2 × Z2 , Z4 × Z2 , D8 , Q8 and Z8 .
If H ∼
= Z2 × Z2 × Z2 , then H contains seven elements of order 2. These elements together with
the element of order 3 forms K8 as a subgraph of P∗ (G), so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If H ∼
= Z4 × Z2 , then H contains two cyclic subgroups of order 4, let them be H1 , H2 , and
three elements of order 2, let them be x1 , x2 , x3 . Here H1 , H2 contains an element of order 2 in
common, so without loss of generality, let it be x1 . Then x2 , x3 are adjacent to all the non-trivial
elements of H1 and H2 . Hence P∗ (G) contains K5,4 as a subgraph with the bipartition X, Y , where
X contains all the non-trivial elements in H1 and H2 ; Y contains x2 , x3 , and the elements of order
3 in G. Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If H ∼
= D8 , then H contains five elements of order 2, let them be xi , i = 1, 2, . . . , 5. Also H
contains a cyclic subgroup of order 4, let it be H1 , which contains only one element of order 2, let
it be x1 . So xi , i 6= 1 is adjacent to all the non-trivial elements in H1 . Hence P∗ (G) contains K5,4
as a subgraph with the bipartition X, Y , where X contains all the non-trivial elements in H1 , and
the elements of order 3 in G; Y contains only xi , i 6= 1. Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If H ∼
= Q8 , then G ∼
= Q8 × Z3 or Z3 o Q8 . If G ∼
= Q8 × Z3 , then G contains three cyclic
subgroups of order 12, let them be H1 , H2 , H3 . H contains three cyclic subgroup of order 4, let
them be H1 , H2 and H3 . Then P∗ (G) contains K5,4 as a subgraph with the bipartition X, Y ,
where X contains all the elements of order twelve in H1 , H2 ; Y contains all the elements of order
twelve in H3 . therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If G ∼
= Z3 o Q8 , then G contains seven
elements of order 4, let them be Hi , i = 1, 2, . . . , 7. So P∗ (G) contains K5,4 as a subgraph with the
bipartition X, Y , where X contains all the elements of order four in H1 , H2 , H3 ; Y contains all the
elements of order four in H4 , H5 . So, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. Moreover, these subgroups
contains an element of order 2 in common, let it be x. Then P∗ (G) contains K5,4 as a subgraph
with the bipartition X, Y , where X contains all non-trivial elements in H1 , H2 ; Y contains all the
elements of order 3 in G, and elements of order 4 in H3 . So γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If H ∼
= Z8 , then the elements not in H are adjacent to all the non-trivial elements in H of order
8. Hence P∗ (G) contains K7,3 as a subgraph. Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
If n > 3, then G contains a subgroup H of order 2n . If H is cyclic, then the element not in
27
H are adjacent to the elements in H of order 2n , n > 3. It follows that P∗ (G) contains K7,3 as a
subgraph. Therefore, γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If H is non-cyclic, then by Proposition 4.3,
γ(P∗ (H)) > 1 and γ(P∗ (H)) > 1, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
(ii) If q ≥ 5, then P∗ (G) contains subgroups of order 2n−1 and q, let them be H1 and H2
respectively. Then each element in H1 is not a power of any element in H2 , and vice versa. This
implies that P∗ (G) contains K7,3 as a subgraph and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Subcase 3b(ii). If p ≥ 3, then G contains a subgroup of order pn , let it be H. If H is non-cyclic,
then by Theorem 4.3, γ(P∗ (H)) > 1 and γ(P∗ (H)) > 1, so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1. If H
is cyclic, then the elements not in H are adjacent to all the elements in H of order pn . It follows
that P∗ (G) contains K7,3 as a subgraph and so γ(P∗ (G)) > 1 and γ(P∗ (G)) > 1.
Proof of Theorem 4.1 follows by combining all the propositions proved so far in this section.
Proof of Corollary 4.1. Note that, if P∗ (G) is one of star, path, Cn , outerplanar, and not containing
K1,4 or K2,3 , then P∗ (G) must be planar. So to classify the finite groups whose complement of
proper power graphs is one of these, it is enough to consider the finite group whose complement of
proper power graph is planar. It is easy to check each of such possibilities among the list of groups
given in Theorem 4.1(1), and their corresponding complement of proper power graph structure
given in (3.1), (3.4), (3.2), (4.1) and Figures 3, 2, 1. This completes the proof.
Acknowledgment
This research work of the first author is supported by Ministry of Social Justice & Empowerment
and Ministry of Tribal Affairs, India in the form of Rajiv Gandhi National Fellowship.
References
[1] A. Abdollahi, S. Akbary and H. R. Maimani, Non-commuting graph of a group, J. Algebra 28
(2006), 468–492.
[2] M. Afkhami, M. Farrokhi D. G and K. Khashyarmanesh, Planar, toroidal, and projective
commuting and non-commuting graphs, Comm. Algebra 43 (2015), 2964–2970.
[3] M. Behboodi, Zero divisor graphs for modules over commutative rings, J. Commut. Algebra 4
(2012), 175–197.
28
[4] D. Bubboloni, Mohammad A. Iranmanesh and S. M. Shaker, Quotient graph of power graph,
(2016). Preprint available at arXiv:1502.02966v3 [math.CO].
[5] W. Burnside, Theory of Groups of Finite Order, Dover Publications, Cambridge, 1955.
[6] Cheng Kai Nah, M. Deaconescu, Lang Mong Lung and Shi Wujie, Corrigendum and addendum
to ”Classification of finite groups with all elements of prime order”, Proc. Amer.Math.Soc. 117
(1993), 1205–1207.
[7] A. Doostabadi and M. Farroki D. G, Embedding of (proper) power graph of finite groups,
Available at arXiv:1402.1322v1 [math.GR] (2014).
[8] Eleonora Crestani and Andrea Lucchini, The generating graph of finite soluble groups, Israel
J. Math. 198 (2013), 63–74.
[9] A. Gagarin, W. Kocay, Embedding graphs containing K5 -subdivisions, Ars Combin. 64 (2002),
pp. 33–49.
[10] A. Gagarin, W. Kocay, D. Neilson, Embeddings of small graphs on the torus, Cubo Mat. Educ.
5 (2003), no. 2, pp. 351–371.
[11] F. Harary, Graph Theory, Addison-Wesley, Philippines, 1969.
[12] Hung-Jen Chiang-Hsieh, Classification of rings with projective zero-divisor graphs, J. Algebra
319 (2008) 2789–2802.
[13] Ivy Chakrabarty, Shamik Ghosh and M.K. Sen, Undirected power graph of semigroup, Semigroup Forum 78 (2009), 410–426.
[14] Jemal Abawajy, Andrei Kelarev and Morshed Chowdhury, Power graph: A survey, Electron.
J. Graph Theory Appl. 2 (2013), 125–147.
[15] A. V. Kelarev and S. J. Quinn, Directed graph and combinatorial properties of semigroups, J.
Algebra 251 (2002), 16–26.
[16] Min Feng, Xuanlong Ma and Kaishun Wang, The structure and metric dimension of the power
graph of a finite group, European J. Combin. 43 (2015), 82–97.
[17] G. A. Miller, H .F. Blichfeldt, L. E. Dickson, Theory and application of finite groups, John
wiley and sons, inc, London, 1916.
29
[18] M. Mirzargar, A. R. Ashrafi and M. J. Nadjafi-Arani, On the power graph of a finite group,
Filomat 26 (2012), 1201–1208.
[19] A. R. Moghaddamfar, S. Rahbariyan, and W. J. Shi, Certain properties of the power graph
associated with a finite group, J. Algebra Appl. 13 (2014), Article No. 1450040 (18 pages).
[20] E. Neufeld and W. Myrvold, Practical toroidality testing in: 8th Annual ACM-SIAM symposium on discrete algorithms, (SODA) 1997, pp. 574–580.
[21] R. Rajkumar and P. Devi, Planarity of permutability graph of subgroups of groups, J. Algebra
Appl. 13 (2014) Article No. 1350112 (15 pages).
[22] R. Rajkumar and P. Devi and Andrei Gagarin, Classification of finite groups with toroidal or
projective-planar permutability graphs, Comm. Algebra, 44(9) (2016), 3705-3726.
[23] R. Rajkumar and P. Devi, Toroidality and projective-planarity of intersection graphs of subgroups of finite groups, Available at arXiv:1505.08094v2 [math.GR] (2015).
[24] Rulin Shen and Enshi, Intersection graph of subgroup of finite group, Czechoslovak Math. J
60 (135) (2010), 945–950.
[25] W. R. Scott, Group Theory, Dover Publications, New York, 1964.
[26] Sriparna Chattopadhyay and Pratima Panigrahi, Some relations between power graphs and
Cayley graphs, J. Egyptian Math. Soc. 23 (2015), 457–462.
[27] A. T. White, Graphs, Groups and Surfaces, North-Holland Mathematics Studies, No.8. American Elsevier Publishing Co. Inc., New York, 1973.
[28] J. S. Williams, Prime graph components of finite groups, J. Algebra 69 (1981), 487–513.
[29] Xuanlong Ma and Kaishun Wang, Power graphs of genus one, Available at arXiv:1509.012104v1
[math.GR] (2015).
| 4math.GR
|
ON THE REALIZATION OF THE GELFAND CHARACTER
OF A FINITE GROUP AS A TWISTED TRACE
arXiv:1303.4800v4 [math.GR] 1 Mar 2018
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
Abstract. We show that the Gelfand character χG of a finite group G
(i.e. the sum of all irreducible complex characters of G ) may be realized
as a “ twisted trace” g 7→ T r(ρg ◦ T ) for a suitable involutive linear
automorphism T of L2 (G), where ρ is the right regular representation
of G and that under certain hypotheses T (f ) = f ◦ L (f ∈ L2 (G)),
where L is an involutive antiautomorphism of G. Moreover the natural
representation τ of G associated to the L-twisted conjugacy action of
G in the fixed point set of L turns out to be a Gelfand Model in some
cases. We show here that a sufficient condition for τ not being a Gelfand
Model is the existence of elements of order 2 in the centre of G.
Keywords: Gelfand character; Gelfand model; twisted trace; finite group
representation theory; involutions.
MSC2010: 20C15, 20C
1. Introduction
The realization of the Gelfand character χG of a finite group G, i.e. the
sum of all ordinary irreducible characters of G, is an old problem [2, 7, 8, 9].
One possible approach to this problem is to try to obtain χG by twisting
the trace of some very natural representation (V, π) of G, like its regular
representation, by a suitable linear automorphism T of its underlying space
V, so as to obtain χG (g) = T r(πg ◦ T ) for all g ∈ G. Recall that twisted
traces appear in many contexts in mathematics [1, 2, 5].
A classical result in character theory of finite groups [4] is that the cen2
tral function θ1 : G → C, defined by θ1 (g)
P= |{h ∈ G : h = g}| is a generalized character that satisfies θ1 (g) =
π∈Ĝ ν(π)χπ (g) where ν(π) =
1 P
2
g∈G χπ (g ) is the Frobenius- Schur number of the character χπ of the
|G|
irreducible representation π of G. If (π, V ) is an irreducible complex representations of a finite group G and π is self-contragradient then there exists
a non degenerate bilinear form B on V, unique up to scalar multiple, such
that B(v, w) = ǫ( π)B(w, v) where ǫ( π) = ±1. Frobenius and Schur [3] proved
that ν(π) = ǫ(π) and if ν(π) = 1 then π(G) is conjugated to a subgroup of
the orthogonal group O(n) and if ν(π) = −1 then π(G) is conjugate to a
subgroup of the symplectic group Sp(n) and n is even. Clearly if ν(π) = 1
for all irreducible representations of G, then θ1 is the Gel’fand character χG
The authors have been partially supported by Fondecyt Grant 1140510.
1
2
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
of G. Some groups for which this is the case are Sn , D2n and O(n, q), q
odd [4].
Furthermore Gow [4] proved that for G = GL(n, q), the central function
θ2 : G → C, defined by θ2 (g) = |{h ∈ G : ht h−1 = g}| affords χG .
On the other hand Kawanaka and Matsuyama
[6] defined the twisted
1 P
χ(gτ
(g))
where τ is an invoFrobenius Schur indicator ντ (χ) = |G|
g∈G
lutive automorphism of G. They proved that if the character χ is afforded
by a matrix representation R of G such that R ◦ τ = R then ντ (χ) = 1. If
this is not the case but nevertheless χ ◦ τ = χ, then ντ (χ)
P = −1. Finally if
χ ◦ τ 6= χ, then ντ (χ) = 0. Furthermore they get that χ∈Ĝ ντ (χ)χ(g) =
|{h ∈ G : τ (h)2 = g}|.
Later Bump and Ginzburg [2] considered an automorphism τ of G such
that τ r = 1, the norm map N : G → G defined by N (g) = gτ (g)τ 2 (g) · · · τ r−1 (g)
and the number M (g) of solutions of the equation N (x) = g with x ∈ G.
2πi
They proved that M (g) lies in the Ze r - algebra generated by the irreducible
P
1 P
characters and M (g) = χ∈Ĝ ǫ(χ)χ(g), where ǫτ (χ) = |G|
g∈G χ(N (g)).
If r = 2 and τ is trivial then ǫτ (χ) = ν(χ)
P and M (g) = θ1 (g); if τ is not
trivial then ǫτ (χ) = ντ (χ), and M (g) = χ∈Gb ντ (χ)χ(g). If G = GL(n, q)
and τ is the transpose inverse then M (g) = θ2 (g).
We prove that the Gelfand character χG of a finite group G may be always
realized as a twisted trace as χG (g) = T r(ρg ◦ T ), g ∈ G, where T is an
involutive automorphism of L2 (G) and ρ is the right regular representation.
Moreover we prove that if G admits an involutive anti-automorphism L such
that L(g) is conjugated to g, for all g ∈ G then, putting L∗ (f ) = f ◦ L, we
get that the central function t(g) = T r(ρg ◦ L∗ ) of g ∈ G equals
P the number
−1
of solutions to the equation h L(h) = g, and is given by π∈Gb ǫπ χπ (g),
where
ǫπ = νL (χπ ) =
1 X
χπ (L(g)g −1 ) = ±1.
|G|
g∈G
Furthermore if the number of fixed points of L is equal to the sum of
the dimensions of all irreducible representations of G, then under the above
assumptions the central function t is the Gelfand character of G. So we can
take the involutive automorphism T of L2 (G) above to be L∗ .
Let G be a group for which there exists an anti-automorphism L satisfying
all the conditions above. Then we can consider the G - set X = {h ∈ G :
L(h) = h} where the action of G in X is given by g · h = g −1 hL(g)−1 .
Let τ be the associated natural representation of G in L2 (X). Although
L2 (X) has the right dimension to be a Gelfand Model of G, we prove - using
our twisted trace description of χG - that this cannot be the case if G admits
non trivial central involutions.
ON THE REALIZATION OF THE GELFAND CHARACTER OF A FINITE GROUP AS A TWISTED TRACE3
2. The Gelfand character χG as a Twisted Trace
Let G be a finite group and let (L2 (G), ρ) and (L2 (G), σ) be the right and
left regular representation of G respectively; let (Uk , πk ) (1 ≤ k ≤ r) denote
all the irreducible unitary representations of G and let Iπk (1 ≤ k ≤ r) be
the isotypic component of type πk of ρ. Let Uk be an orthonormal basis of Uk
and (ekij (g))1≤i,j≤nk the matrix of the operator πk (g) (g ∈ G) with respect
to the basis Uk of Uk , where nk denotes the dimension of Uk . Moreover χk
denotes the character afforded by πk .
The matrix coeficients ekij (1 ≤ i, j ≤ nk , 1 ≤ k ≤ r) provide then an
orthonormal basis B for the Hilbert space L2 (G), and they satisfy the relations:
ekij (g−1 ) = ekji (g)
(1)
and
ekij (gh)
=
nk
X
ekil (g)eklj (h)
l=1
Proposition 1. Let (V, π 1 ) and (V, π 2 ) be two isomorphic representations of
a finite group G in the same vector space V such that πh1 ◦πg2 = πg2 ◦πh1 for all
g, h ∈ G and let S be an involutive automorphism of V that intertwines the
representations π 1 and π 2 of G. Then the function T r(π?1 ◦ S) = T r(π?2 ◦ S)
defined on G with values in C is a central function on G and so it is a linear
complex combination of irreducible characters of G.
Proof. For g, h in G, we have
T r(πg1−1 hg ◦ S) = T r(πg1 ◦ S ◦ πg1−1 h ) = T r(S ◦ πg2 ◦ πg1−1 h )
= T r(S ◦ πg1−1 h ◦ πg2 ) = T r(πg2 ◦ S ◦ πg1−1 h ) = T r(S ◦ πg1 ◦ πg1−1 h )
= T r(S ◦ πh1 ) = T r(πh1 ◦ S) = T r(S ◦ πh2 ) = T r(πh2 ◦ S).
Recall that all irreducible representations πk of a finite group G are unitarizable. The isotypic component Iπk of type (Uk , πk ) of ρ is isomorphic to
Uk ⊗ Uk∗ .
Theorem 1. Let T be the linear application of L2 (G) defined by T (ekij ) = ekji
for all ekij ∈ B and let σ
e be the homomorphism from G to Aut(L2 (G)) defined
by
σ
eg = (T ◦ ρg ◦ T )
4
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
for all g ∈ G. Then T is an involutive automorphism of L2 (G) and σ
e is a
representation of G such that:
i. ρg ◦ T = T ◦ σ
eg , g ∈ G.
ii. ρg ◦ σ
eh = σ
eh ◦ ρg , g, h ∈ G.
iii. T r(ρg ◦ T ) = χG (g), g ∈ G
Proof. Since
σ
eg = (T ◦ ρg ◦ T )
for all g ∈ G, we obtain that σ
eg is an automorphism of L2 (G) such that
ρg ◦ T = T ◦ σ
fg for each g ∈ G.
Furthermore, since
ρg (ekij ) =
nk
X
eklj (g)ekil
l=1
and
σ
eg (ekij )
for all
ekij
=
nk
X
ekli (g)eklj
l=1
∈ B, g ∈ G, we get that
(ρg ◦
σ
eh )(ekij )
=
nk
X
=
nk
X
ekli (h)
m=1
nk
X
=
ekmj (g)eklm
m=1
l=1
ekmj (g)
nk
X
nk
X
ekli (h)eklm
l=1
emj (g)e
σh (ekim )
m=1
= (e
σh ◦ ρg )(ekij ).
for g, h ∈ G and ekij ∈ B. Finally, since
(ρg ◦ T )(ekij ) =
nk
X
ekli (g)ekjl
l=1
for all g ∈ G and ekij ∈ B, then
nk
r
r X
X
X
k
χk (g) = χG (g).
(
eii (g) =
T r(ρg ◦ T ) =
k=1 i=1
k=1
Next we will prove that under certain conditions the central function
T r(ρ? ◦ T ) can constructed from an involutive anti-automorphism L of the
group G.
ON THE REALIZATION OF THE GELFAND CHARACTER OF A FINITE GROUP AS A TWISTED TRACE5
Theorem 2. Let L be an involutive antiautomorphism of G, such that L(g)
is conjugated to g, for all g ∈ G, and let L∗ be the automorphism of L2 (G)
defined by L∗ (f ) = f ◦ L. Then for all g ∈ G, we have
i.
ii.
iii.
iv.
v.
ρg ◦ L∗ = L∗ ◦ σL(g)−1
ρg ◦ σL(g)−1 = σL(g)−1 ◦ ρg
T r(ρg ◦ L∗ ) = |{h ∈ G : h−1 L(h)) = g}|
T r(ρg ◦ L∗ ) = T r(ρg−1 ◦ L∗ )
T r(ρg ◦ L∗ ) = Σrk=1 εk χk (g) where εk = ±1. .
Proof. The proof of i. and ii. is a straightforward calculation. By computing
the trace of ρg ◦ L∗ with respect to the canonical basis {δg : g ∈ G} where
δg (h) = δg,h , (h ∈ G) we obtain iii.
Let Ag = {h ∈ G : h−1 L(h) = g}. Taking inverses we see that h ∈ Ag if
and only if L(h) ∈ Ag−1 . Then iv. follows, since L is bijective.
Because of i. and ii. the representations (L2 (G), ρ) and (L2 (G), σ ∗ ) of G
and the involutive automorphism L∗ satisfy all conditions of proposition 1,
and so we deduce that the complex function T r(ρg ◦ L∗ ) is central and
∗
T r(ρg ◦ L ) =
r
X
εk χk (g),
k=1
for suitable complex numbers εk .
The antiautomorphism L induces an antiautomorphism L̃ on the complex
group algebra C[G]. Since χk (L(g)) = χk (g), g ∈ G, 1 ≤ k ≤ r, L̃ acts as
the identity on the centre and thereforeLinduces an antiautomorphism L̃k
on each simple component of C[G] ∼
=
1≤k≤r M (nk , C). Due to SkolemNoether theorem, L̃k is conjugated to the transposition map, i. e. there
t
exists bk ∈ GL(nk , C) such that L̃k (a) = b−1
k a bk for all a ∈ M (nk , C) .
−1 t −1
−1 t
t −1
t
Furthermore we have that a = L̃(L̃(a)) = b−1
k bk a(bk ) bk = bk bk a(bk bk )
−1 t
t
for all a ∈ M (nk , C), so bk bk belongs to the centre and therefore bk = εk bk
with εk = ±1, since ((bk )t )t = bk .
In this way, for each representation (Uk , πk ) of G a symmetric or a antisymmetric form bk exists, with respect to which the linear operators πk (g)
and L̃(πk (g)) are adjoint to each other. More precisely, if we consider the
bilinear form hu, vi = v t bk u, defined by bk , then
(2)
hL̃(πk (g))(u), vi = v t (bk L̃(πk (g)))u = v t (πk (g))t bu = hu, πk (g)(v)i
Let us suppose that for some k we have εk = 1. We choose then an
orthonormal basis Uk+ = {ui : 1 ≤ i ≤ nk } of Uk , with respect to the
symmetric form bk and we denote by ekij (g) = hπk (g)(uj ), ui i the matrix
coefficients of πk (g) with respect to this basis. Due to equation 2 we have the
following relations between the matrix coefficients of L̃(πk (g)) = πk (L(g))
and πk (g),
6
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
ekij (L(g)) = huj , πk (g)(ui )i = ekji (g).
(3)
Let E k = hekij : 1 ≤ i, j ≤ nk i and T rk (ρg ◦ L∗ ) denote the trace of the
restriction of ρg ◦ L∗ to the subspace E k of L2 (G). In order to compute
T rk (ρg ◦ L∗ ) we note that
[(ρg ◦ L∗ )(ekij )](h) = ekij (L(hg)) = ekij (L(g)L(h)) =
nk
X
ekil (L(g))eklj (L(h)) =
l=1
nk
X
ekli (g)ekjl (h).
l=1
Then
(ρg ◦ L∗ )(ekij ) =
nk
X
ekli (g)ekjl .
l=1
Since
T rk (ρg ◦ L∗ ) =
X
h(ρg ◦ L∗ )(ekij ), ekij i =
1≤i,j≤nk
X
nk
X
ekli (g)hekjl , ekij i
1≤i,j≤nk l=1
and hekjl , ekij i = δ(j,l),(i,j) , we get
∗
T rk (ρg ◦ L ) =
nk
X
ekii (g) = χk (g)
i=1
Let us suppose now that k is such that εk = −1. Let nk = 2mk ; then
we can find a symplectic basis Uk− = {ui , 1 ≤ i ≤ nk } of Uk such that
hui , ui+mk i = 1 and hui , uj+mk i = hui+mk , uj+mk i = hui , uj i = 0, i 6= j (1 ≤
i, j ≤ mk ).
Analogously, equation 1 gives us the following relations for the matrix
coefficients ekij (g) of πk (g) with respect to this basis:
hπk (g)(ui ), uj i = −ekj+mk i (g), j ≤ mk ,
hπk (g)(ui ), uj i = ekj−mk i (g), j > mk
Therefore
hπk (L(g))(ui ), uj i = −hπk (g)(uj ), ui i = eki+mk j (g), i ≤ mk ,
and
hπk (L(g))(ui ), uj i = −eki−mk j (g), i > mk
.
Taking into account these relations and equation 2 we get the following
relations between the matrix coefficients of L̃(πk (g)) = πk (L(g)) and πk (g) :
ON THE REALIZATION OF THE GELFAND CHARACTER OF A FINITE GROUP AS A TWISTED TRACE7
(4)
ekij (L(g)) = ekj+mk i−mk (g), j ≤ mk , i > mk
(5)
ekij (L(g)) = −ekj−mk i−mk (g), j > mk , i > mk
(6)
ekij (L(g)) = −ekj+mk i+mk (g), j ≤ mk , i ≤ mk
(7)
ekij (L(g)) = ekj−mk i+mk (g), j > mk , i ≤ mk
So we get:
For i, j ≤ mk
(8)
mk
nk
X
X
∗
k
k
k
(ρg ◦L )(eij ) =
−el+mk i+mk (g)ej+mk l+mk +
ekl−mk i+mk (g)ekj+mk l−mk .
l=mk +1
l=1
For i ≤ mk , j > mk
(9)
mk
nk
X
X
∗
k
k
k
(ρg ◦L )(eij ) =
−el+mk i+mk (g)ej−mk l+mk +
ekl−mk i+mk (g)(−ekj−mk l−mk ).
l=mk +1
l=1
For i > mk , j ≤ mk
(10)
mk
nk
X
X
(ρg ◦L∗ )(ekij ) =
ekl+mk i−mk (g)(−ekj+mk l+mk )+
−ekl−mk i−mk )(g)ekj+mk l−mk .
l=mk +1
l=1
And for i > mk , j > mk
(11)
mk
nk
X
X
∗
k
k
k
(ρg ◦L )(eij ) =
el+mk i−mk (g)ej−mk l+mk +
−ekl−mk i−mk (g)(−ekj−mk l−mk ).
l=mk +1
l=1
Notice that:
a) If i, j ≤ mk then ekj+mk l+mk 6= ekij and ekj+mk l−mk 6= ekij
b) If i ≤ mk , j > mk then ekj−mk l+mk = ekij if and only if j = mk + i > mk
and l + mk = j > mk , moreover ekj−mk l−mk 6= ekij
c) If i > mk , j ≤ mk then ekj+mk l−mk = ekij if and only if j + mk = i >
mk and l − mk = j < mk , moreover ekj+mk l+mk 6= ekij
d) If i > mk , j > mk then ekj−mk l+mk 6= ekij and ekj+mk l−mk 6= ekij
Therefore
nk
mk
X
X
∗
k
T rk (ρg ◦ L ) =
−ejj (g) +
−ekjj (g) = −χk (g).
j=mk +1
j=1
Hence, taking into account both cases, we get
8
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
T r(ρg ◦ L∗ ) =
r
X
T rk (ρg ◦ L∗ ) =
k=1
r
X
εk χk (g)
k=1
Definition 1. For any mapping L from G to itself, we denote by F ixG (L)
the fixed point set {g ∈ G : L(g) = g} of L.
Proposition 2. If L is an involutive antiautomorphism of G such that L(g)
is conjugated to g for all g ∈ G and
r
X
|F ixG (L)| =
nk ,
k=1
then
T r(ρg ◦ L∗ ) =
r
X
χk (g)
k=1
and
L∗ = T.
Proof. If we evaluate T r(ρg ◦ L∗ ) on e we obtain
T r(L∗ ) =
r
X
εk nk
k=1
but T r(L∗ ) = |F ixG (L)| and by hypothesis
r
X
nk .
|F ixG (L)| =
k=1
Since nk > 0,we conclude that εk = 1 for all k. Then using equation 3 it
follows that
L∗ (ekij )(g) = ekij (L(g)) = ekji (g),
hence L∗ = T.
Proposition 3. If L is an involutive antiautomorphism of G such that L(g)
is conjugate to g for all g ∈ G, then
1 X
χk (L(h)h−1 ))χk (g)
(12)
T rk (ρg ◦ L∗ ) = (
|G|
h∈G
Proof. To prove equation (12) we compute the Fourier coefficients λk , 1 ≤
k ≤ r of the central function
T r(ρg ◦ L∗ ) = T r(L∗ ◦ σL(g)−1 )
with respect to the basis
{χk =
nk
X
i=1
ekii : 1 ≤ k ≤ r}
ON THE REALIZATION OF THE GELFAND CHARACTER OF A FINITE GROUP AS A TWISTED TRACE9
of the centre Z of C[G]. First we observe that
(L∗ ◦ σL(g)−1 )(ekij )(h) = ekij (L(g)L(h)) =
nk
X
ekil (L(g))(eklj ◦ L)(h)
l=1
and
nk
1 XX
ekil (L(g))ekil (L(h))ekji (h−1 ).
|G|
h(L∗ ◦ σL(g)−1 )(ekij ), ekij i =
h∈G l=1
therefore
T r(L∗ ◦ σL(g)−1 ) =
r
X
nk
1 XX
ekil (L(g))eklj (L(h))ekji (h−1 )
|G|
X
k=1 1≤i,j≤nk
Since
nk
X
h∈G l=1
eklj (L(h))ekji (h−1 ) = ekli (L(h))h−1 )
j=1
we get that
T r(L∗ ◦ σL(g)−1 ) =
r
X
X
k=1 1≤i,l≤nk
and then, for 1 ≤
k′
r
X
λk ′ = h
1 X k
eli (L(h)h−1 )(ekil ◦ L)(g)
|G|
h∈G
≤ r,
X
k=1 1≤i,l≤nk
1 X k
eli (L(h)h−1 )(ekil ◦ L), χk′ i
|G|
h∈G
By hypothesis χk (L(g)) = χk (g), 1 ≤ k ≤ r, so
λk′ =
r
X
X
k=1 1≤i,l≤nk
But
1
|G|
P
nk ′
X
1 X k
1 X k
−1
eil (L(g))ekss′ (L(g)),
(
eli (L(h)h )
|G|
|G|
s=1
g∈G
h∈G
k
k′
g∈G eil (L(g))ess (L(g))
k′
= hekil , ess i, then
nk ′
X
1 X
1 X k′
ess (L(h)h−1 ) =
χk′ (L(h)h−1 ).
λk ′ =
|G|
|G|
s=1
h∈G
h∈G
Proposition 4. Let L be an involutive antiautomorphism of G such that
L(g) is conjugate to g for all g ∈ G. Then the following conditions are
equivalent
10
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
i.
|F ixG (L)| =
r
X
nk
k=1
ii.
1 X
χk (L(h)h−1 ) = 1, (1 ≤ k ≤ r)
|G|
h∈G
Proof. It follows from propositions 2 and 3
Proposition 5. Let L an involutive antiautomorphism
of G , τ : G → G
P
defined by τ (g) = L(g −1 ) and cτ (χ) = |G|−1 g∈G χ(gτ (g)) the twisted
Frobenius - Schur
indicator defined by Kawanaka and Matsuyama. Then
1 P
−1
cτ (χk ) = |G|
h∈G χk (hL(h ) = ±1 if χk (L(g)) = χk (g). Moreover if the
matrix representations Rk afforded by χk satisfy Rk (L(g)) = Rk (g), then
cτ (χk ) = 1.
Proof. Due to gτ (g)) = gL(g −1 ), we have
X
χ(gL(g −1 )) =
g∈G
X
χ(L(g)g −1 )
g∈G
and since
T r(ρg ◦ L∗ ) = T r(ρg ◦ L∗ )
we can write:
T r(ρg−1
r
X
1 X
◦L )=
χk (L(h)h−1 )χk (g −1 ),
|G|
∗
h∈G
k=1
and therefore
T r(ρg−1
r
X
1 X
◦L )=
χk (hL(h−1 ))χk (g).
|G|
∗
k=1
h∈G
Our proposition follows.
Theorem 3. If L is an involutive antiautomorphism of G such that:
i. L(g) is conjugated
Pr to g, for all g ∈ G;
ii. |F ixG (L)| = k=1 nk
then
∗
T r(ρg ◦ L ) =
r
X
k=1
χk (g).
ON THE REALIZATION OF THE GELFAND CHARACTER OF A FINITE GROUP AS A TWISTED TRACE
11
Proof. Since L(g) is conjugated to g, for all g ∈ G, aplying proposition 3 we
get
T r(ρg ◦ L∗ ) =
r
X
T rk (ρg ◦ L∗ ) =
r
X
1 X
χk (L(h)h−1 ))χk (g).
(
|G|
k=1
k=1
h∈G
Furthermore, it follows from ii), with the help of proposition 4, that
1 X
χk (L(h)h−1 ) = 1
|G|
h∈G
So
∗
T r(ρg ◦ L ) =
.
r
X
χk (g)
k=1
Remark 1. Notice that condition ii. above does not follow from condition
i. Indeed, for G = SL(2, Fq ) with q congruent to 1 mod 4 we have that i.
holds for L(g) = g −1 (g ∈ G) but L has only two fixed points in G.
3. Applying the twisted trace to test natural Gelfand Model
candidates
Let G be a group for which there exists an involutive antiautomorphism L
satisfying all the conditions of proposition 2. Then we can consider the G set X = F ixG (L) where the action of G in X is given by g · h = g −1 hL(g)−1 .
Let us call τ the associated natural representation L2 (X) of G.
Let Z denote the centre Z(G) of G; we note that if z ∈ Z then z ∈ X
because L(z) is conjugated to z by hypothesis. Furthermore since χτ (g) =
|{h ∈ G : ghL(g) = h}| for all h ∈ G, it follows that χτ (z) = |X| for all
involutions z ∈ Z. So the character χτ does not tell central involutions from
the identity.
Since the dimension of L2 (X) is equal to the sum of all irreducible representations of G, the representation (L2 (X), τ ) could be a Gelfand Model
for G. To check whether this is indeed the case, we can of course compare
its character χτ with the Gelfand character χG that we have described as a
twisted trace: χG (g) = T r(ρg ◦ L∗ ), g ∈ G.
We see then that a necessary condition for τ to be a Gelfand Model is
that
χG (z) = χτ (z) (= |X)|
for all central involutions z, i.e we need
T r(ρg (z) ◦ L∗ ) = |{g ∈ G : g −1 L(g) = z}| = |X|
.
12
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
3.1. The case of G = SL(2, Fq ), q odd. As an illustrative example we
consider here the case of G = SL(2, Fq ), q odd.
Recall that the sum of the dimensions of all irreducible representations of
G is q(q + 1).
0 , we may define L : G → G, L(g) = kg −1 k.
For k = 10 −1
We note that if g = ac db ∈ G then L(g) = ( dc ab ) , and
X = F ixG (L) = {( ac ab ) ∈ G : a2 − bc = 1}.
.
Proposition 6. L is an involutive antiautomorphism of G satisfying:
i. L(g) is conjugated with g, for all g ∈ G.
ii. |X| = q(q + 1)
iii. The action of G on X is transitive.
Proof. i) This is clear since L(g) has same trace and determinant as g ∈ G.
ii) For ( ac ab ) ∈ X, we must have bc = a2 − 1. Now, if a2 − 1 = 0, i. e.
a = ±1, we have 2q − 1 pairs (b, c) such that bc = a2 − 1. On the other hand,
if a2 − 1 6= 0, then there are q − 1 pairs (b, c) such that bc = a2 − 1. So
|X| = 2 · (2q − 1) + (q − 2) · (q − 1) = q(q + 1).
iii)We have that |OrbG (e)| = |G|/|StabG (e)| and StabG (e) = {g ∈ G :
L(g) = g −1 }. For g = ac db ∈ G, this means b = c = 0 and d = a−1
0
Therefore |StabG (e)| = |{ a0 a−1
: a ∈ F∗q }| = q − 1.
So |OrbG (e)| =
tive.
q(q 2 −1)
q−1
= q 2 + q = |X| and therefore the action is transi
In this way we get that L is an involutive antiautomorphism of G =
SL(2, Fq ) satisfying allconditions in Theorem 1 above. Hence for the central
0
involution z = −1
0 −1 , we have
But if h =
a b
c d
χG (z) = |{h ∈ G : h−1 L(h) = z}|.
∈ G satisfies h−1 L(h) = z then
−b
( dc ab ) = −a
−c −d
and that is equivalent to b = c = 0 and −a2 = 1.
So, if q ≡ 3 mod(4) then χG (z) = 0 and if q ≡ 1 mod(4) then χG (z) = 2.
In both cases χG (z) 6= χτ (z). Therefore (L2 (X), τ ) is not a Gelfand Model
for G.
4. Central involutions as an obstruction to the existence of
an involution fixed point Gelfand Model
Motivated by the previous example, we may notice the following.
Lemma 1. For every central element z ∈ G of order 2, we have.
χG (z) 6= χG (e).
ON THE REALIZATION OF THE GELFAND CHARACTER OF A FINITE GROUP AS A TWISTED TRACE
13
Proof. Indeed, notice that χπ (z) = ± dim(π) for all irreducible representations π of G and so for χG (z) = χG (e) to hold we need that χπ (z) =
dim(π) = χπ (e) for all irreducible representations π, but this is absurd since
the regular character χρ tells z from e, because χρ (e) = |G| =
6 0 = χρ (z).
Proposition 7. Let G a finite group and an involutive anti automorphism
L of G such that
i. L(g) is conjugated
to g, for all g ∈ G;
P
ii. |F ixG (L)| = rk=1 nk
Then if G has central (non trivial) involutions we conclude that the natural
representation (L2 (X), τ ) associated to the G− space X = F ixG (L) cannot
be a Gelfand Model for G.
Proof. This follows inmediately from the necessity of the condition χG (z) =
χτ (z) for all central involutions z and lemma 1.
Acknowledgements
The authors thank Pierre Cartier for helpful discussions related to this
work.
References
[1] A
(very
brief)
history
of
the
trace
formula,
retrieved
from:www.claymath.org/cw/arthur/pdf/HistoryTraceFormula.pdf
[2] D. Bump, D. Ginzburg, Generalized Frobenius-Schur Numbers, Journal of Algebra
278 (2004), 294-313.
[3] G. Frobenius, I. Schur, Über die rellen Darstellungen der endlichen Groupen, Sitz.Ber. Deutsch. Akad. Wiss. Berlin (1906), 186-208.
[4] R. Gow, Properties of the characters of the finite general linear groups related to the
transpose-inverse involution, Proc. London Math. Soc (3) 47 (1983), 493-506.
[5] K. Hashimoto, Twisted trace formula of the Brandt matrix, Proc. Japan Acad., 53,
Ser. A (1977), 98 - 102.
[6] N. Kawanaka , H. Matsuyama, A twisted version of the Frobenius -Schur indicator
and multiplicity free permutation representation, Hokkaido Math. J., 19 (1990), 495508.
[7] J. Soto-Andrade, Geometrical Gel’fand Models, Tensor Quotients and Weil Representations, Proc. Symp. Pure Math., 47 (1987), Amer. Math. Soc., 305-316.
[8] A. Terras, Fourier Analysis on Finite Groups and Applications, Cambridge U. Press,
Cambridge, UK, 1999.
[9] M. F. Yañez, A weakly geometrical Gel’fand model for GL(n, q) and a realization of
the Gel’fand character of a finite group. C. R. Acad. Sci. Paris Série I Math. 316 (11)
(1993), 1149–1154.
14
JORGE SOTO-ANDRADE AND M. FRANCISCA YÁÑEZ
Dep. Mat. Fac. Ciencias, Univ. de Chile
Escuela de Pregrado, Fac. Cs. Quı́micas y Farmacéuticas, Univ. de Chile
| 4math.GR
|
arXiv:1802.09121v1 [cs.CC] 26 Feb 2018
Limits on representing Boolean functions
by linear combinations of simple functions:
thresholds, ReLUs, and low-degree polynomials
R. Ryan Williams∗
Abstract
We consider the problem of representing Boolean functions exactly by “sparse” linear combinations
(over R) of functions from some “simple” class C . In particular, given C we are interested in finding
low-complexity functions lacking sparse representations. When C is the set of PARITY functions or the
set of conjunctions, this sort of problem has a well-understood answer; the problem becomes interesting
when C is “overcomplete” and the set of functions is not linearly independent. We focus on the cases
where C is the set of linear threshold functions, the set of rectified linear units (ReLUs), and the set of
low-degree polynomials over a finite field, all of which are well-studied in different contexts.
Building on the new easy witness lemma of Cody Murray and the author, we provide generic tools
for proving lower bounds on representations of this kind. Applying these, we give several new lower
bounds for “semi-explicit” Boolean functions. Let α (n) be an unbounded function such that nα (n) is
time constructible (e.g. α (n) = log⋆ (n)). We show:
• Functions in NTIME[nα (n) ] that require super-polynomially many linear threshold functions to
represent (depth-two neural networks with sign activation function, a special case of depth-two
threshold circuit lower bounds).
• Functions in NTIME[nα (n) ] that require super-polynomially many ReLU gates to represent (depthtwo neural networks with ReLU activation function).
• Functions in NTIME[nα (n) ] that require super-polynomially many O(1)-degree F p -polynomials
to represent exactly, for every prime p (related to problems regarding Higher-Order Uncertainty
Principles). We also obtain a function in ENP requiring 2Ω(n) -size linear combinations.
• Functions in NTIME[npoly(log n) ] that require super-polynomially many ACC ◦ THR circuits of
polynomial size to represent exactly (further generalizing the recent lower bounds of Murray and
the author).
We also obtain “fixed-polynomial” lower bounds for functions in NP, for the first three representation
classes.
∗ EECS and CSAIL, MIT. Supported by NSF CCF-1553288. Any opinions, findings and conclusions or recommendations
expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
1 Introduction
Given f : {0, 1}n → {0, 1} and a class C of “simple” functions, when can f be represented exactly as a
short R-linear combination of functions from C ? When C forms a basis for Bn (the set of all Boolean
functions on n inputs) the question has a unique answer that is generally easy to obtain, by analyzing the
appropriate linear system (the cases where C is the set of all parity functions or the set of all conjunctions
are canonical examples). For |C | ≫ 2n , the situation becomes much more interesting, as there can be
many possible representations. The general problem of understanding which functions do and do not have
sparse representations for simple C arises in many different mathematical topics. Three relevant to TCS are
depth-two threshold circuits, depth-two neural networks with various activation functions, and higher-order
Fourier analysis. We use the notation
SUM ◦ C
to denote the class of R-linear combinations of C -functions; for example, SUM ◦ MOD2 denotes R-linear
combinations of PARITY functions. The relevant complexity measure for a “circuit” in SUM ◦ C is the
fan-in of the SUM gate, which we call the sparsity of the circuit.
Sums of Threshold Circuits. Let SUM ◦ THR be linear combinations of linear threshold functions
2
(LTFs).1 As there are 2Θ(n ) n-variate threshold functions [Win60], a function f : {0, 1}n → {0, 1} has
many possible representations as a SUM ◦ THR. Such circuits are also known in the machine learning
literature as depth-two neural networks with sign activation functions.
In 1994, Roychowdhury, Orlitsky, and Siu [ROS94] noted that no interesting size lower bounds were
known for computing Boolean functions with SUM ◦ THR circuits (beyond the few that are/were known
for THR ◦ THR [HMP+ 93, ROS94, KW16, CSS16, Tam16, ACW16]). The problem was raised again
more recently in CCC’10 by Hansen and Podolskii [HP10]. In particular, the following remains largely
unanswered:
Problem: Find an explicit f : {0, 1}⋆ → {0, 1} without polynomially-sparse SUM ◦ THR, i.e., every
linear combination of LTFs computing f on n-bit inputs needs nω (1) LTFs, for infinitely many n.
Because of prior lower bounds in weaker settings (such as majority-of-majority [HMP+ 93] and majorityof-thresholds [Nis94]), it is natural to think that correlation bounds against linear threshold functions should
help.2 Correlation bounds do imply lower bounds for SUM ◦ THR, but only when the weights in the linear
combination are not too large (i.e., the weights must be in [−2δ n , 2δ n ] for small δ < 1). However, if arbitrary
weights are allowed, interesting lower bounds on SUM ◦ THR (beyond Ω(n2.5 wires [KW16]) were open,
to the best of our knowledge. In Section 4, we prove arbitrary polynomial lower bounds for NP functions:
Theorem 1.1. For all k, there is an fk ∈ NP without SUM ◦ THR circuits of nk sparsity. Furthermore, for
every unbounded α (n) such that nα (n) is time constructible, there is a function in NTIME[nα (n) ] that does
not have SUM ◦ THR circuits of polynomial sparsity.
Note that for arbitrary circuits (even for THR ◦ THR circuits) the best known complexity for such functions without nk -size circuits (for fixed k) is MA/1 ([San09]) and S2p .
1 From
here on, “linear combination” means “R-linear combination”, unless otherwise specified.
is, one wants to show that a function cannot be (1/2 + ε (n))-approximated by a linear threshold function, for the tiniest
ε (n) > 0 possible.
2 That
1
Sums of ReLU Gates. A ReLU (rectified linear unit) gate is a function f : {0, 1}t → R+ such that there
is a vector w ∈ Rt and scalar a ∈ R such that for all x,
f (x) = max{0, hx, wi + a}.
It is important to note that ReLU gates might not be Boolean-valued, but they must output non-negative
numbers on all Boolean inputs. Linear combinations of ReLU gates are also known as depth-two neural
networks with ReLU activation functions, and they are intensely studied in machine learning. Several
lower bounds for Sums-of-ReLU functions (which for consistency we call SUM ◦ ReLU) have recently
been shown for functions with real-valued inputs and outputs (examples include [ES16, Tel16, ABMM16,
Dan17, SS17]) but none of the methods extend to Boolean functions, to the best of our knowledge. Recently,
Mukherjee and Basu [MB17] have proved Ω(n1−δ )-gate lower bounds for THR ◦ ReLU circuits computing
the Andreev function, extending ideas in [KW16].
Observing that for |hx, wi| ≥ 1 we have
max{0, hx, wi + 1} − max{0, hx, wi} = sign(hx, wi),
it follows that every SUM ◦ THR circuit can be simulated by a SUM ◦ ReLU circuit with only a doubling of
the sparsity. In Section 5 we extend our lower bounds to Sums-of-ReLU circuits:
Theorem 1.2. For all k, there is an fk ∈ NP without SUM ◦ ReLU circuits of nk sparsity. Furthermore, for
every unbounded α (n) such that nα (n) is time constructible, there is a function in NTIME[nα (n) ] that does
not have SUM ◦ ReLU circuits of polynomial sparsity.
Representing Boolean Functions With Higher-Order Polynomials. Higher-order Fourier analysis of
Boolean functions deals with representing Boolean functions by R-linear combinations of F2 -polynomials
of degree higher than one (see [HHL16] for a survey of some applications in CS theory). The question of
which (if any) explicit functions lack sparse representations, even for degree-two polynomials, has been
wide open. Letting MOD2 be the class of parity functions, this question asks to find lower bounds for
SUM ◦ MOD2 ◦ AND2 circuits (in our notation, ANDk denotes ANDs of fan-in at most k). Such lower
bound problems appear much more difficult than the degree-one case of SUM◦MOD2. Even understanding
the sparsity of the AND function in the quadratic (and in general, degree-O(1)) setting is a prominent open
problem:
Hypothesis 1 (Quadratic Uncertainty Principle [FHH+ 14]). There is an ε > 0 such that the AND function
on n variables does not have SUM ◦ MOD2 ◦ AND2 circuits of 2ε n sparsity.
Although it is believed that AND needs exponential sparsity, to our knowledge the only lower bound
known for an explicit function in SUM ◦ MOD2 ◦ AND2 was Ω(n)-sparsity. For completeness we include a proof provided to us by Lovett [Lov17]) in Appendix A. Again, when the weights in the linear combination are required to be small (magnitudes are 2ε n for small ε > 0), correlation bounds yield
some results: one example (among many) is the work of Green [Gre04] showing that a majority vote
of quadratic F3 -polynomials needs 2Ω(n) polynomials to compute PARITY. (Other works in this vein include [HG91, CGT96, Bou05, GT12]; see Viola [Vio09] for a survey.) However, for arbitrary weights, no
non-trivial lower bounds have been reported (to our knowledge).
In Section 6, we prove polynomial sparsity lower bounds for Boolean functions in NP and 2Ω(n) -size
lower bounds for ENP , against linear combinations of polynomials over any prime field with any constant
degree:
2
Theorem 1.3. For every integer k, d ≥ 1 and prime p, there is an fk ∈ NP without SUM ◦ MODp ◦ ANDd
circuits of nk sparsity. Furthermore, for every unbounded α (n) such that nα (n) is time constructible, there
is a function in NTIME[nα (n) ] that does not have SUM ◦ MODp ◦ ANDd circuits of polynomial sparsity.
Theorem 1.4. For every d ≥ 1 and prime p, there is an α > 0 and an f ∈ ENP without SUM ◦ MODp ◦
ANDd circuits of 2α n sparsity.
Note the “smallest” known complexity class for a function lacking 2Ω(n) -size circuits is EΣ2 P [MNW99],
and it is a longstanding open problem to reduce the complexity class for such a function, even against
depth-3 AC0 circuits.
1.1 Intuition
Here we give an overview of some of the ideas used to prove the lower bounds in this work. The lower
bounds of this paper follow the high-level strategy of proving circuit lower bounds by designing circuitanalysis (satisfiability) algorithms [Wil13, Wil14b, Wil14a]. However, in this work we must execute this
strategy differently. All previous lower bounds proved in this framework utilize the “polynomial method”
from circuit complexity in various ways (representing a circuit by a low-degree polynomial of some kind),
combined with fast matrix multiplication and/or fast polynomial evaluation. These approaches do not seem
to work for solving SAT on linear combinations of thresholds, low-degree polynomials, or ReLU gates. For
example, we do not know how to get a sparse (probabilistic or approximate) polynomial (over any field)
for computing an OR of many SUM ◦ THRs, and it is likely that any reasonable approach via polynomials
would fail to yield non-trivial results. However, we are able to adapt some bits of the polynomial method
to the setting of low-degree polynomials (see Section 6).
Another complication is that, in the prior lower bound arguments, a nondeterministic procedure guesses
a small circuit C of the kind one wishes to prove a lower bound against, and composes C with other Boolean
circuitry to form a SAT instance. In our case, if we guess some arbitrary SUM ◦ C circuit, we first need to
know if this circuit is actually computing a Boolean function; if not, then the satisfiability question itself
is not well-defined, and it will not be possible to meaningfully compose such a circuit with other Boolean
circuits. Thus we need a way to efficiently check whether a linear combination is Boolean-valued.
We give a generic way to “lift” non-trivial algorithms for counting SAT assignments to short products
of C circuits to non-trivial algorithms for detecting if a given SUM ◦ C circuit is Boolean-valued and
for counting SAT assignments. More precisely, we show that in order to prove lower bounds for linear
combinations of C -functions, it suffices to solve a certain sum-product task faster than exhaustive search:
Sum-Product over C : Given k functions f1 , . . . , fk from C , each on Boolean variables x1 , . . . , xn ,
compute
k
∑ ∏ fi (x).
x∈{0,1}n i=1
Note the Sum-Product is computed over R, and the task makes sense even if the functions f1 , . . . , fk output
non-Boolean values. Further note that if the functions f1 , . . . , fk are Boolean-valued, then the product of k
of them is simply the AND of k of them. In general, the Sum-Product problem will be NP-hard for most
interesting representation classes: for example, it is already equivalent to Subset Sum when C is the set of
exact threshold functions (see Section 2 for a definition). Our meta-theorem states that mild improvements
over exhaustive search for Sum-Product over C imply strong lower bounds for SUM ◦ C :
3
Theorem 1.5. Suppose every C ∈ C has a poly(n)-bit representation, where each C can be evaluated on
a given input in poly(n) time. Assume there is an ε > 0 and for k = 1, . . . , 4 there is an nO(1) · 2n−ε n -time
algorithm for computing the Sum-Product of k functions f1 (x1 , . . . , xn ), . . . , fk (x1 , . . . , xn ) from C . Then:
1. For every k, there is a function in NP that does not have SUM ◦ C circuits of sparsity nk .
2. For every unbounded α (n) such that nα (n) is time constructible, there is a function in NTIME[nα (n) ]
that does not have SUM ◦ C circuits of polynomial sparsity.
Theorem 1.5 is used to prove lower bounds against SUM ◦ THR, SUM ◦ ReLU, and SUM ◦ MOD2 ◦
ANDO(1) . For the ENP lower bounds, we use a closure property of SUM ◦ MOD2 ◦ ANDO(1) combined with
standard ideas from this line of work (see Theorem 3.3).
Theorem 1.5 (and its components) can also be used to easily “lift” existing circuit lower bounds to linear
combinations of those circuits:
b
Theorem 1.6. For every d, m ≥ 1, there is a b ≥ 1 and an f ∈ NTIME[nlog n ] that does not have SUM ◦
AC0d [m] ◦ THR circuits of na size, for every a.
That is, we obtain super-polynomial sparsity lower bounds on representing nondeterministic quasipolynomial-time functions with R-linear combinations of ACC ◦ THR circuits (each of polynomial size).
This applies the fact that we can solve the Sum-Product problem on ACC ◦ THR circuits (because we can
count SAT assignments to them), with an analogous running time as the best SAT algorithm. More details
on Theorem 1.6 can be found in Section 3.
Outline. The next section is the Preliminaries, which gives background knowledge. Section 3 proves Theorem 1.5. In Sections 4, 5, and 6, Sum-Product algorithms for THR, ReLU, and MODp ◦ ANDd (degree-d
F p -polynomials) are provided which beat exhaustive search. The algorithms for THR and ReLU (Theorems 4.1 and 5.1) build upon and extend old Subset-Sum algorithms (Theorem 2.1). The algorithm for
MODp◦ANDd (Theorem 6.1) uses tools from the polynomial method in a new way. Applying Theorem 1.5
to each of these algorithms, we obtain strong lower bounds for SUM ◦ C for all three classes C .
2 Preliminaries
Let C be a class of functions of the form f : {0, 1}n → R. Each member C ∈ C has a number of inputs n
and a size, which is the length of the representation of C in bits. For the classes THR, MOD2 ◦ ANDO(1) ,
and ReLU, the size |C| of a representation is poly(n) bits, without loss of generality; see Proposition 1. (For
classes such as MOD2 ◦ ANDlog2 (n) , a member of the class takes Ω(nlog n ) bits to represent, in the worst
case.) We assume that for all n, our class C contains the projection functions fi (x1 , . . . , xn ) = xi for all
i = 1, . . . , n. We also assume that C is evaluatable, meaning that there is a universal k ≥ 1 such that every
C ∈ C can be evaluated on a given input in O(|C|k ) time. All classes we consider have this property.
As is standard, we let ANYc denote the class of Boolean functions with c inputs (the class contains “any”
such function).
An arbitrary SUM ◦ C circuit C over n variables represents some function f : {0, 1}n → R. We say that
C is Boolean-valued if for all x ∈ {0, 1}n , the output of C on x is in {0, 1}. The following proposition is
useful to keep in mind, as it shows that every sparse linear combination of Boolean functions implementing
another Boolean function has an equivalent linear combination with “reasonable” coefficients.
4
Proposition 1. Let C be a class of functions with co-domain {0, 1}, and let C be an SUM ◦ C circuit of
sparsity s that is Boolean-valued. There is an equivalent SUM ◦ C circuit C′ such that every weight in the
linear combination of C′ has the form j/k, where both j and k are integers in [−ss/2 , ss/2 ].
Proof. (See also [MTT61, BHPS10].) Let C be a linear combination of s functions from C . WLOG, the
set of s Boolean functions from C is a linearly independent set (otherwise, we could obtain a smaller linear
combination representing the same function). The problem of finding coefficients for the Boolean-valued C
n
is equivalent to solving a certain linear system Ax = b in s unknowns over the rationals, where b ∈ {0, 1}2
n
and A ∈ {0, 1}s×2 . Take a linearly independent subsystem of s of these 2n equations. Since the determinant
of any s × s Boolean matrix is in [−ss/2 , ss/2 ] [Had93], the result follows from Cramer’s rule.
The relevant theorem for sums of ReLU gates is more involved, but Maass [Maa97] shows how the
weights for a circuit of size s need only poly(s, n) bits of precision. Such “analog-to-digital” results are
crucial for our work, as in our lower bound proofs we will need a discrete nondeterministic algorithm to
guess a SUM ◦ C circuit and check various properties of it.
Useful Results For Thresholds. We draw from several algorithms and representation theorems from past
work. For SUM ◦ THR, we eventually appeal to a classic result from exact algorithms:
Theorem 2.1 (Horowitz and Sahni [HS74]). The number of Subset Sum solutions to any arbitrary instance
of n items with integer weights of magnitude [−2W , 2W ] can be computed in 2n/2 · poly(W ) time.
Theorem 2.1 is usually stated in terms of finding a subset sum solution, but the algorithm can be easily
adapted to count solutions as well.
A Boolean function f is called an exact threshold function if there are real-valued α1 , . . . , αn and t such
that for all x,
f (x) = 1 ⇐⇒ ∑ αi xi = t.
i
Let ETHR be the class of exact threshold functions. For our SUM ◦ THR circuit results, the following
transformation is extremely useful:
Theorem 2.2 (Hansen and Podolskii [HP10]). Every linear threshold function in n variables can be represented as an linear combination of poly(n) exact threshold functions, each with coefficient 1.
It follows that every SUM ◦ THR of sparsity s has an equivalent SUM ◦ ETHR of sparsity poly(s). The
idea is that a THR function defines a set of points in the Boolean hypercube lying on one side of a given
hyperplane; we can “cover” all the points lying on one side by a disjoint sum of poly(n) “parallel” hyperplanes, which function as ETHR gates. Thus each coefficient in the linear combination is simply 1.
Another useful property of ETHR gates is that they are closed under AND:
Theorem 2.3 (Hansen and Podolskii [HP10]). Every conjunction of t exact threshold functions in n variables with integer weights in [−W,W ] can be converted in poly(t, n) time to an equivalent single exact
threshold gate, with weights in [−(nW )Θ(t) , (nW )Θ(t) ].
The idea is simple: if we multiply the ith exact threshold gate’s linear form by the factor (nW )i , no linear
form will “interfere” with the other sums, and we can determine if all of them are satisfied simultaneously
with one exact threshold.
5
Useful Results for Finite Field Polynomials. Two tools from the literature will be helpful for our results
on linear combinations of polynomials. The first is modulus-amplifying polynomials, which have been used
in Toda’s Theorem [Tod91], representations of ACC and ACC-SAT algorithms [BT94, Wil14b], algorithms
for All-Pairs Shortest Paths [CW16], and algorithms for solving polynomial systems [LPT+ 17]:
Lemma 2.1 (Beigel and Tarui [BT94]). For all ℓ ∈ Z+ , the degree-(2ℓ − 1) polynomial (over Z)
ℓ
Pℓ (y) = 1 − (1 − y)
ℓ−1
∑
j=0
ℓ+ j−1 j
y
j
has the property for all integers m ≥ 2,
• if y = 0 mod m then Pℓ (y) = 0 mod mℓ ,
• if y = 1 mod m then Pℓ (y) = 1 mod mℓ .
Furthermore, each coefficient in Fℓ has magnitude at most 2O(ℓ) .
Recall that a multivariate polynomial is multilinear if it contains no powers larger than one. The second
tool is a classic result on rapidly evaluating a multilinear polynomial on all points in the Boolean hypercube.
Theorem 2.4 (cf. [BHK09], Section 2.2). Given the 2n -coefficient vector of a multilinear polynomial p ∈
Z[x1 , . . . , xn ] where each coefficient is in [−W,W ], the value of p on all points in {0, 1}n can be computed
in 2n · poly(n, logW ) time.
The algorithm of Theorem 2.4 can be obtained by divide-and-conquer (as described in [Wil11]) or by
dynamic programming (as in [BHK09], Section 2.2).
Connections Between Nondeterministic Circuit UNSAT Algorithms and Circuit Lower Bounds. We
also appeal to several known connections between circuit UNSAT algorithms that beat exhaustive search
and circuit lower bounds against nondeterministic time classes, which build on prior work [Wil13, JMV15,
SW13, BSV14].
Theorem 2.5 ([MW17]). If there is an ε > 0 such that Circuit Unsatisfiability for (fan-in 2) circuits with n
inputs and 2ε n size is solvable in O(2n−ε n ) nondeterministic time, then for every k there is a function in NP
that does not have nk -size (fan-in 2) circuits.
Theorem 2.6 (Corollary 12 in Tell [Tel18], following [MW17]). If there is a δ > 0 and c ≥ 1 such that
Circuit Unsatisfiability for (fan-in 2) circuits with n variables and m gates is solvable in O(2n(1−δ ) · mc )
nondeterministic time, then for every unbounded α (n) such that nα (n) is time-constructible, there is a
function in NTIME[nα (n) ] that is not in P/poly.
Theorem 2.7 ([MW17]). If there is an ε > 0 such that Circuit Unsatisfiability for (fan-in 2) circuits with
ε
ε
n inputs and 2n size is solvable in O(2n−n ) nondeterministic time, then for every k there is a function in
k
NTIME[npoly(log n) ] that does not have nlog n -size (fan-in 2) circuits.
In fact, all of these algorithms-to-lower-bounds connections still hold when we replace Circuit Unsatisfiability with the promise problem of distinguishing unsatisfiable circuits from circuits with 2n−1 satisfying
assignments.
6
The Power of Linear Combinations of Low-Degree Polynomials. We note that classical work suggests
that R-linear combinations of higher-degree F2 -polynomials can be quite powerful. For example, applying
Valiant’s depth reduction [Val77] and using the representation of the AND function in the Fourier basis,
it is easy to show that every O(n)-size O(log n)-depth circuit can be represented by a linear combination
of 2O(n/ log log n) F2 -polynomials of degree O(nε ), for any desired ε > 0. Moreover, one can represent any
O(n)-size “Valiant series-parallel” circuit (see [Cal08]) by a linear combination of 2ε n F2 -polynomials of
O(1/ε )
degree 22
. Hence there is a natural barrier to proving exponential-sparsity lower bounds for linear
combinations of “somewhat-low” degree polynomials.
3 Meta-Theorem for Lower Bounds on Linear Combinations of Simple
Functions
In this section, we prove our generic theorem which is applied in subsequent sections to prove lower bounds
against linear combinations of threshold functions, ReLU gates, and constant-degree polynomials. Recall
(from the Introduction) the Sum-Product problem:
Sum-Product over C : Given k functions f1 , . . . , fk from C , each on Boolean variables x1 , . . . , xn ,
compute
k
∑ ∏ fi (x).
x∈{0,1}n i=1
Reminder of Theorem 1.5 Suppose every C ∈ C has a poly(n)-bit representation, where each C can
be evaluated on a given input in poly(n) time. Assume there is an ε > 0 and for k = 1, . . . , 4 there is an
nO(1) · 2n−ε n -time algorithm for computing the Sum-Product of k functions f1 (x1 , . . . , xn ), . . . , fk (x1 , . . . , xn )
from C . Then:
1. For every k, there is a function in NP that does not have SUM ◦ C circuits of sparsity nk .
2. For every unbounded α (n) such that nα (n) is time constructible, there is a function in NTIME[nα (n) ]
that does not have SUM ◦ C circuits of polynomial sparsity.
The remainder of this section is devoted to proving Theorem 1.5, and an extension to ENP in some cases.
We are able to use much of the earlier arguments [Wil13, Wil14b, MW17] as black boxes. However we
need several modifications.
The first new component needed is a method for checking that a given linear combination of C circuits
actually encodes a Boolean function (i.e. is Boolean-valued on all Boolean inputs). This is provided by the
following theorem:
Theorem 3.1. Assume there is an ε > 0 and for k = 1, . . . , 4 there is an nO(1) · 2n−ε n -time algorithm for
computing the Sum-Product of k functions f1 (x1 , . . . , xn ), . . . , fk (x1 , . . . , xn ) from C .
Then there is an 2n−ε n · poly(n, s)-time algorithm that, given f (x1 , . . . , xn ) which is an arbitrary linear
combination of s functions from C , determines whether or not f (a) ∈ {0, 1} for all a ∈ {0, 1}n .
Proof. Suppose we are given f = ∑si=1 αi ci , where αi ∈ R and ci ∈ C each have n inputs. Consider the
polynomial
h(x) := f (x)2 · (1 − f (x))2 = f (x)2 − 2 f (x)3 + f (x)4 .
Observe that:
• If f (a) ∈ {0, 1} for all a ∈ {0, 1}n , then h(a) = 0 for all a.
7
• f (b) ∈
/ {0, 1} implies h(b) > 0.
• For all a ∈ {0, 1}n , h(a) ≥ 0.
Therefore ∑a∈{0,1}n h(a) = 0 if and only if f (a) ∈ {0, 1} for all a ∈ {0, 1}n . By applying the distributive
law to each of f (x)2 , f (x)3 , f (x)4 , and exchanging the order of summation, we have
!
∑
a∈{0,1}n
h(a) =
∑ βi ,i
∑
1 2
i1 ,i2
+
∑
∑
∑
γi1 ,i2 ,i3
i1 ,i2 ,i3
+
fi1 (x) · fi2 (x)
a∈{0,1}n
!
fi1 (x) · fi2 (x) · fi3 (x)
a∈{0,1}n
∑
δi1 ,i2 ,i3 ,i4
i1 ,i2 ,i3 ,i4
!
fi1 (x) · fi2 (x) · fi3 (x) · fi4 (x)
a∈{0,1}n
for βi1 ,i2 = αi1 · αi2 , γi1 ,i2 ,i3 = −2αi1 · αi2 · αi3 , δi1 ,i2 ,i3 ,i4 = αi1 · αi2 · αi3 · αi4 .
Observe that each sum over a ∈ {0, 1}n on the RHS is precisely a Sum-Product task over C , with products
ranging from k = 2 to k = 4. Therefore we can check that the sum ∑a∈{0,1}n h(a) is zero with O(s4 ) calls to
Sum-Product over C . By assumption, this can be done in O(2n−ε n · poly(n, s)) time.
The second crucial component yields the ability to solve Circuit Unsatisfiability efficiently with nondeterminism, under the hypotheses (in fact, weaker hypotheses). This is provided by the following lemma,
which is similar to (but more complicated than) Lemma 3.1 in [Wil14b]:
Lemma 3.1. Assume:
• There is an ε > 0 and for k = 1, . . . , 4 there is an nO(1) · 2n−ε n -time algorithm for computing the
Sum-Product of k functions from C .
• The Circuit Evaluation problem has SUM ◦ C circuits of sparsity nk , for some k > 0.
Then there is a nondeterministic 2n−ε n · poly(n, s)-time algorithm for Circuit Unsatisfiability, on arbitrary
fan-in-2 circuits with n inputs and s gates.
Proof. Suppose we are given a circuit C with n inputs and s gates of fan-in 2, and wish to nondeterministically prove it is unsatisfiable. Let us index the gates in topological order, so that gates 1, . . . , n are the input
gates, and the s-th gate is the output gate.
Our nondeterministic algorithm begins by guessing a SUM◦ C circuit EVAL with n + O(log s) inputs and
sparsity at most (n + s)k+1 , which is intended to encode the Circuit Evaluation function:
EVAL(C, x, i) := Evaluate C on x, and output the value of the i-th gate of C.
(Note i is encoded as an O(log s)-bit string.) Let
D(x, i) := EVAL(C, x, i),
i.e., we think of C as hard-coded in the function, to simplify the notation. Applying Theorem 3.1, we can
check that D encodes a Boolean function in 2n−ε n · poly(s, n) time.
Next, we check that D(a, s) = 0 for all a ∈ {0, 1}n ; in other words, D claims that C outputs 0 on every
input. Suppose D has the form
(n+s)k+1
D(x, i) =
∑
j=1
8
α j · c j (x, i),
for some α j ∈ R and c j ∈ C . Since D has already been determined to be Boolean, it suffices to compute
∑a∈{0,1}n D(a, s) to know whether or not D(x, s) = 0 for all a. By exchanging the order of summation,
!
∑
∑
D(a, s) =
a∈{0,1}n
a∈{0,1}n
= ∑αj ·
j
∑ α j · c j (a, i)
j
∑
!
c j (a, i) .
a∈{0,1}n
Therefore we only need to make (n + s)k+1 calls to Sum-Product over C (with k = 1) to determine that
D(x, s) = 0 for all a ∈ {0, 1}n . This can be done in 2n−ε n · poly(n, s) time, by assumption.
Next, we have to check that for every gate i = 1, . . . , s, and every a ∈ {0, 1}n , D(a, i) correctly reports the
output of the i-th gate when C evaluates a. To check the input gates, we need to check that D(x, i) = xi for
all i = 1, . . . , n; we can do this by checking that
∑
(D(x, i) − xi )2 = 0,
a∈{0,1}n
which (by distributivity and re-arranging the order of summation, as in the proof of Theorem 3.1) can be
computed with O((n + s)2(k+1) ) calls to Sum-Product over C (with k = 2) in 2n−ε n · poly(n, s) time.
For all gates i other than the input gates, the ith-gate takes inputs from previous gates indexed by some
i1 < i and i2 < i, and computes a function of their two outputs. To check the consistency of gate i, we
can form a degree-3 polynomial pi (A, B,C) which outputs 0-1 values on all A, B,C ∈ {0, 1}, such that
pi (A, B,C) = 0 if and only if A is the output of gate i, given that B is the output of gate i1 and C is the output
of gate i2 .
Since D is Boolean-valued, we have reduced our problem to determining that
∑
p(D(a, i), D(a, i1 ), D(a, i2 )) = 0,
a∈{0,1}n
for each gate i = n + 1, . . . , s, and each gate i’s corresponding input gates i1 and i2 . Applying the distributive
law to the LHS and exchanging the order of summation (as before), this results in O((n + s)3(k+1) ) SumProduct-over-C computations with up to k = 3 products, computable in 2n−ε n · poly(n, s) time.
Our nondeterministic algorithm determines that the input circuit C is unsatisfiable if and only if all of
the above checks pass. If C is satisfiable, then every possible D guessed will fail some check. If C is
unsatisfiable, then under the hypotheses of the theorem, a SUM ◦ C circuit D simulating every gate of C
always exists. By guessing this D, and running the assumed Sum-Product algorithm, our nondeterministic
algorithm accepts.
After the above preparation, we turn back to the proof of Theorem 1.5. At this point, it is simply a matter
of applying the above Lemma 3.1 with the known algorithms-to-lower-bound connections:
Proof of Theorem 1.5. Suppose every C ∈ C has a poly(n)-bit representation, where each C can be
evaluated on a given input in poly(n) time. Recall the hypothesis of the theorem is:
(A) There is an ε > 0 and for k = 1, . . . , 4 there is an nO(1) · 2n−ε n -time algorithm for computing the
Sum-Product of k functions f1 (x1 , . . . , xn ), . . . , fk (x1 , . . . , xn ) from C .
Furthermore, recall that Lemma 3.1 states:
9
Assuming (A) and assuming Circuit Evaluation has SUM ◦ C circuits of sparsity nk for some k, there
is a nondeterministic 2n−ε n · poly(n, s)-time algorithm for Circuit Unsatisfiability, on arbitrary fan-in-2
circuits with n inputs and s gates.
We can then prove the lower bounds of the theorem readily, as follows.
(1) Assume every function in NP has SUM ◦ C circuits of nk sparsity circuits, for some fixed k. Then
both hypotheses of Lemma 3.1 are satisfied (note Circuit Evaluation is in P), and the conclusion
implies that there is an ε > 0 such that Circuit Unsatisfiability for (fan-in 2) circuits with n inputs
and 2ε n size is solvable in O(2n−ε n ) nondeterministic time. Therefore by Theorem 2.5, for every k
there is a function in NP that does not have nk -size (fan-in 2) circuits. This is a contradiction because
SUM ◦ C circuits of nk sparsity can be simulated with nck -size fan-in-2 circuits, for some universal c.
(2) The same argument as in (1) and (2) (but with Theorem 2.6 applied) shows that for every unbounded
α (n) such that nα (n) is time-constructible, there is a function in NTIME[nα (n) ] that does not have
SUM ◦ C circuits of polynomial sparsity.
A Note on Lower Bounds for Linear Combinations of ACC Circuits. Other lower bound consequences
of the arguments in Theorem 1.5 follow easily from combining known results. Here is an example:
b
Reminder of Theorem 1.6 For every d, m ≥ 1, there is a b ≥ 1 and an f ∈ NTIME[nlog n ] that does not
have SUM ◦ AC0d [m] ◦ THR circuits of na size, for every a.
This lower bound can be obtained as follows. First, the argument of Lemma 3.1 also shows:
Theorem 3.2. Assume
ε
• There is an ε > 0 and for k = 1, . . . , 4 there is an nO(1) · 2n−n -time algorithm for computing the
Sum-Product of k functions from C .
• The Circuit Evaluation problem has SUM ◦ C circuits of sparsity na , for some a > 0.
ε
Then there is a nondeterministic 2n−n · poly(n, s)-time algorithm for Circuit Unsatisfiability, on arbitrary
fan-in-2 circuits with n inputs and s gates.
Now we combine this theorem with the following two facts:
1. For every depth d and integer m ≥ 2, there is an ε > 0 such that the Sum-Product of O(1) AC0d [m] ◦
ε
ε
THR circuits of 2n size can be computed in 2n−n time. This simply applies the algorithm for
counting satisfying assignments of AC0d [m] ◦ THR circuits ([Wil14a]).
α
α
2. If for some α > 0 there is a nondeterministic 2n−n -time Circuit Unsatisfiability algorithm for 2n b
a
size circuits, then for every a ≥ 1, there is a b ≥ 1 such that NTIME[nlog n ] does not have nlog n -size
circuits (this is a theorem of Murray and Williams [MW17]).
b
Theorem 1.6 is immediate: Assuming NTIME[nlog n ] has SUM ◦ AC0d [m] ◦ THR circuits of na size for
some a ≥ 1, both hypotheses of Theorem 3.2 are satisfied for C = AC0d [m] ◦ THR, and the conclusion of
Theorem 3.2 combined with item 2 above yields a contradiction.
3.1 Lower Bounds for Exponential Time With an NP Oracle
For classes C with a natural closure property, the lower bounds can be extended to 2Ω(n) sparsity for a
function in ENP . Recall ANYc denotes the class of Boolean functions with c inputs (the class contains
“any” such function).
10
For an integer c ≥ 1, we say that C is efficiently closed under NC0c if there is a polynomial-time algorithm
A such that, given any circuit C of the form C ◦ ANY c , algorithm A outputs an equivalent circuit D from C
(which is only polynomially larger). We note this property is true of O(1)-degree polynomials:
Proposition 2. For every integer m ≥ 2 and c ≥ 1, the class C =
under NC0c .
S
d≥1 MODm ◦ ANDd
is efficiently closed
Proof. Every MODm ◦ ANDd ◦ ANYc circuit can be represented by an MODm ◦ ANDdc circuit. In particular, every Boolean function on c inputs has an exact representation as a sum (modulo m) of ANDs of
fan-in c; composing such a sum with a MODm ◦ AND circuit and applying the distributive law yields the
result.
Theorem 3.3. There is a universal c ≥ 1 satisfying the following. Suppose C is efficiently closed under
NC0c , and suppose every C ∈ C has a poly(n)-bit representation, where each C can be evaluated on a given
input in poly(n) time.
Assume there is an ε > 0 and for k = 1, . . . , 4 there is an nO(1) · 2n−ε n -time algorithm for computing the
Sum-Product of k functions f1 (x1 , . . . , xn ), . . . , fk (x1 , . . . , xn ) from C .
Then there is a function in ENP that does not have SUM ◦ C circuits of sparsity 2α n , for some α > 0.
The remainder of this section sketches the proof of Theorem 3.3; we give only a sketch, as the argument
closely resembles others [Wil14b, JMV15]).
Let ε ∈ (0, 1). Assume C is efficiently closed under NC0 , and
(A) There is an ε > 0 and an O(2n−ε n )-time algorithm for computing the Sum-Product of k functions
from C , and
(B) For all functions f ∈ TIME[2O(n) ]NP and all α > 0, f has SUM ◦ C circuits of sparsity 2α n .
We wish to establish a contradiction. In particular, we will show that assumptions (A) and (B) together
imply that every problem in NTIME[2n ] can be simulated by a nondeterministic o(2n )-time algorithm,
contradicting the (strong) nondeterministic time hierarchy theorem [SFM78, Žák83].
Let L ∈ NTIME[2n ]. On a given input x, our nondeterministic o(2n )-time algorithm for L has two parts:
(i) It guesses a witness for x of o(2n ) size.
(ii) It verifies that witness for x in o(2n ) time.
To handle (i), we use assumption (B) to show that one can nondeterministically guess a 2α n · poly(n)-size
SUM ◦ C circuit that encodes a witness for x, applying a simple “easy witness” lemma from [Wil13]:
Lemma 3.2 (Lemma 3.2 in [Wil13]). Let D be any class of circuits. If ENP has circuits of size S(n) from
class D, then for every L ∈ NTIME[2n ] and every verifier V for L, and every x ∈ L of length n = |x|, there
is a y of length O(2n ) such that V (x, y) accepts and the D-circuit complexity of y (construed as a function
f : {0, 1}n+O(1) → {0, 1}) is at most S(n).
In other words, assumption (B) implies that every yes-instance of L has S(n)-size “witness circuits”: a
witness of length O(2n ) that can be represented as an S(n)-size SUM ◦ C Boolean-valued circuit. Furthermore, this holds for every verifier for L.
To handle (ii), we choose an appropriate verifier, so that verifying witnesses becomes equivalent to a simple Sum-Product call. In particular we use the following extremely “local” reduction from L ∈ NTIME[2n ]
to 3SAT instances of 2n · poly(n) length:
11
Lemma 3.3 ([JMV15]). Every L ∈ NTIME[2n ] can be reduced to 3SAT instances of O(2n · n4 ) size. Moreover, there is an algorithm that, given an instance x of L and an integer i ∈ [O(2n · n4 )] in binary, reads
only O(1) bits of x and outputs the i-th clause of the resulting 3SAT formula, in O(n4 ) time.
Since in Lemma 3.3 each bit of the output is a function of some c ≤ O(1) inputs, each bit of the output is
a member of ANYc . So for every instance x of length n for the language L, we can produce (in deterministic
poly(n) time) a circuit Dx which is an ordered collection of O(n) functions from ANYc . The circuit Dx takes
n + O(log n) binary inputs, construes that input as an integer i, and outputs the i-th clause of a formula Fx
which is satisfiable if and only if x ∈ L.
Our nondeterministic algorithm for L guesses a 2O(α n) -sparse SUM ◦ C circuit Cx that takes n + O(log n)
inputs and is meant to encode a satisfying assignment for the formula Fx . We can check Cx is Booleanvalued on all 2n · poly(n) inputs in 2n−ε n/2 time, by applying Theorem 3.1 and letting α > 0 be sufficiently
small.
Composing Cx with the O(n) polynomials forming Dx , we obtain a 2O(α n) -sparse SUM◦C ◦ANY c circuit
E with n + O(log n) inputs (composed of three copies of Cx , and O(n) copies of Dx ) such that
E is unsatisfiable if and only if Cx encodes a satisfying assignment for Fx .
(We leave out the details, as they are provided in multiple other papers [Wil13, Wil14b].) To complete the
o(2n )-time algorithm for L, it suffices to check unsatisfiability of the resulting 2O(α n) -size circuit E in o(2n )
nondeterministic time. This would yield the desired contradiction.
Such a nondeterministic UNSAT algorithm is provided by first converting E into an SUM ◦ C circuit in
2O(α n) time (using the fact that C is efficiently closed under NC0 ). This yields a sum of 2O(α n) C -circuits.
Analogously to the proof of Lemma 3.1, checking the unsatisfiability of such an E can be reduced to 2O(α n)
calls to Sum-Product of C , by applying distributivity. Applying the Sum-Product algorithm of assumption
(A) that runs in O(2n−ε n ) time, and setting α > 0 to be sufficiently small, the running time is o(2n ).
This completes the proof of Theorem 3.3.
4 Sparse Combinations of Threshold Functions
We now turn to proving SUM ◦ THR lower bounds. Due to Lemma 1.5, it suffices to give a 2n−ε n -time
algorithm for the Sum-Product Problem over THR:
Sum-Product over THR: Given k linear threshold functions f1 , . . . , fk , each on Boolean variables
x1 , . . . , xn , compute
k
∑ ∏ fi (x).
x∈{0,1}n i=1
Putting together various pieces (described in the Preliminaries), there is a substantially faster-than-2n
time algorithm:
Theorem 4.1. The Sum-Product of k linear threshold functions on n variables (with weights in [−nn , nn ])
can be computed in 2n/2 · nO(k) time.
Note that having weights in [−nn , nn ] is without loss of generality (in our lower bound proofs, our nondeterministic algorithm can always guess an equivalent circuit with such weights, as described by Proposition 1).
12
Proof. Let f1 , . . . , fk be n-variable threshold functions. Applying Theorem 2.2, we can write each fi as a
sum of t = poly(n) exact threshold functions:
t
fi (x) = ∑ gi (x),
i=1
where each gi (x) is defined by some weights wi,1 , . . . , wi,n ∈ R and a threshold value t ∈ R. Therefore we
can write the product f1 · · · fk as
k
∏ fi =
i=1
∑
gi1 · · · gik .
(i1 ,...,ik )∈[t]k
Each term gi1 · · · gik is a conjunction of k exact thresholds. Applying Theorem 2.3, each such term can be
replaced with a single exact threshold gate, with weights of magnitude nO(kn) , i.e., each weight is representable with O(kn log n) bits. Thus
k
∏ fi =
i=1
∑
hi1 ,...,ik
(i1 ,...,ik )∈[t]k
for some exact threshold gates hi1 ,...,ik . The desired sum can therefore be written as
k
∑ ∏ fi(a) = ∑
a∈{0,1}n i=1
∑
hi1 ,...,ik (a)
a∈{0,1}n (i1 ,...,ik )∈[t]k
=
∑
∑
(i1 ,...,ik )∈[t]k
!
hi1 ,...,ik (a) .
a∈{0,1}n
Now observe that each sum ∑a∈{0,1}n hi1 ,...,ik (a) on the RHS is equivalent to an instance of #Subset Sum. In
particular, each such sum is counting the number of subsets of a given set of n weights in [−nΩ(kn) , nO(kn) ]
which sum to zero. By Theorem 2.1, this can be computed in poly(k, n) · 2n/2 time. Since there are nO(k)
such sums to compute in the outer sum, the total running time is nO(k) · 2n/2 .
The following are immediate from Theorem 1.5:
Reminder of Theorem 1.1 For all k, there is an fk ∈ NP without SUM ◦ THR circuits of nk sparsity. Furthermore, for every unbounded α (n) such that nα (n) is time constructible, there is a function
in NTIME[nα (n) ] that does not have SUM ◦ THR circuits of polynomial sparsity.
5 Sparse Combinations of ReLU Gates
Recall that a function f : {0, 1}n → R from the class ReLU is defined with respect to a weight vector w ∈ Rn
and a scalar a ∈ R, such that for all a ∈ {0, 1}n ,
f (x) = max{0, hw, xi + a}.
To prove SUM ◦ ReLU lower bounds, we give a 2n−ε n -time algorithm for the Sum-Product Problem over
ReLU:
Sum-Product over ReLU: Given k ReLU functions f1 , . . . , fk , each on Boolean variables x1 , . . . , xn ,
compute
k
∑ ∏ fi (x).
x∈{0,1}n i=1
13
Theorem 5.1. The Sum-Product of k ReLU functions on n variables (with weights in [−W,W ]) can be
computed in 2n/2 · nO(k) · poly(k, n, logW ) time.
The proof is similar in spirit to the algorithm for Sum-Product of threshold functions (Theorem 4.1),
except that complications arise due to the real-valued outputs of ReLU functions. We end up having to
solve a problem generalizing #Subset Sum, but which turns out to have a nice “split-and-list” 2n/2 -time
algorithm, analogously to #Subset Sum.
Proof. Let f1 , . . . , fk be n-variable ReLU functions, defined by weight vectors w1 , . . . , wk ∈ Rn and scalars
a1 , . . . , ak ∈ R, respectively. Our task is to compute
∑
max{0, hx, w1 i + a1 } · · · max{0, hx, wk i + ak }.
x∈{0,1}n
First, we note the above sum is equal to
∑
[hx, w1 i ≥ −a1 ] · (hx, w1 i + a1 ) · · · [hx, wk i ≥ −ak ] · (hx, wk i + ak ),
x∈{0,1}n
where we are using the Iverson bracket notation [P] to denote a function that outputs 1 if P is true and 0
otherwise. Applying Theorem 2.2, each of the threshold functions [hx, wi i ≥ −ai ] can be represented as
a linear combination of t = poly(n) exact threshold functions. In particular there are exact thresholds gi, j
such that the above sum equals
!
!
t
t
∑ ∑ g1, j (x)
x
· (hx, w1 i + a1 ) · · ·
∑ gk, j (x)
· (hx, wk i + ak ).
j=1
j=1
Applying the distributive law, the above sum equals
∑ ∑
g1, j1 (x) · · · gk, jk (x) · (hx, w1 i + a1 ) · · · (hx, wk i + ak ).
x j1 ,..., jk ∈[t]k
Re-arranging the summation order yields
∑
∑ g1, j1 (x) · · · gk, jk (x) · (hx, w1i + a1) · · · (hx, wk i + ak ) .
j1 ,..., jk ∈[t]k
x
Applying Theorem 2.3, each g1, j1 (x) · · · gk, jk (x) can be replaced by a single exact threshold h j1 ,..., jk (x).
Our task has been reduced to nO(k) computations of the form
∑
h j1 ,..., jk (x) · (hx, w1 i + a1 ) · · · (hx, wk i + ak ).
(1)
x∈{0,1}n
Without the (hx, w1 i + a1 ) · · · (hx, wk i + ak ) term, (1) would be exactly a #Subset Sum instance, as in Theorem 4.1. In this new situation, we need to count a “weighted” sum over the subset sum solutions, where the
weights are determined by a product of k inner products of the solution vectors with some fixed vectors.
Let us now describe how to solve the generalized problem given by (1). To keep the exposition clear, we
will walk through an attempted solution and fix it as it breaks.
Suppose the exact threshold function h j1 ,..., jk (x) of (1) is defined by weights α1 , . . . , αn ∈ R and threshold
value t ∈ R, so that
n
h j1 ,..., jk (x) = 1 ⇐⇒
∑ αi xi = t.
i=1
14
As with the Subset Sum problem, we begin by splitting the set of variables x into two halves, {x1 , . . . , xn/2 }
and {xn/2+1 , . . . , xn } (WLOG, assume n is even). Correspondingly, we split each of the k weight vectors
(1)
wi ∈ Rn of (1) into two halves, wi
respectively.
(2)
∈ Rn/2 and wi
∈ Rn/2 for the first and second halves of variables,
We list all 2n/2 partial assignments to the first half, and all 2n/2 partial assignments to the second. For
each partial assignment A = (A1 , . . . , An/2 ) to the first half of variables {x1 , . . . , xn/2 }, we compute a vector
vA , as follows:
n/2
• vA [0] := −t + ∑i=1 αi Ai ,
D
E
(1)
• for all j = 1, . . . , k, vA [ j] := a j + w j , (A1 , . . . , An/2 ) .
For each partial assignment A′ = (An/2+1 , . . . , An ) from the second half, we compute a vector wA′ :
• wA′ [0] := ∑ni=n/2+1 αi Ai ,
D
E
(2)
• for all j = 1, . . . , k, wA′ [ j] := w j , (An/2+1 , . . . , An ) .
Notice that vA [0] + wA′ [0] = 0 if and only if h j1 ,..., jk (A, A′ ) = 1. Thus in our sum, we only need to consider
pairs of vectors vA from the first half and vectors wA′ from the second half such that vA [0] + wA′ [0] = 0.
Moreover, note that for all j = 1, . . . , k,
vA [ j] + wA′ [ j] = x, w j + a j .
It follows that (1) equals
∑
(vA [1] + wA′ [1]) · · · (vA [k] + wA′ [k]).
(vA ,wA′ ) : vA [0]+wA′ [0]=0
The Subset-Sum algorithm of Horowitz and Sahni [HS74] shows how to efficiently find pairs (vA , wA′ )
with vA [0] + wA′ [0] = 0: sorting all vectors in the second half by their 0-th coordinate, for each vector
vA from the first half we can compute (in poly(n) time) the number of second-half vectors wA′ satisfying
vA [0] + wA′ [0] = 0 (even if there are exponentially many such vectors). However it is unclear how to
incorporate the odd-looking (vA [1] + wA′ [1]) · · · (vA [k] + wA′ [k]) multiplicative factors into a weighted sum.
To do so, we modify the vectors vA and wB as follows. Consider the expansion of ∏ki=1 (vA [i] + wA′ [i]) into
a sum of 2k products: it can be seen as the inner product of two 2k -dimensional vectors, where one vector’s
entries is a function solely of vA and the other vector’s entries is a function solely of wA′ . (Furthermore, note
that the number of bits needed to describe entries in these new vectors has increased only by a multiplicative
factor of k.)
Thus we can assign (2k + 1)-dimensional vectors v′A (in place of the vA ) and w′B (in place of the wB ) such
that v′A [0] = vA [0], w′A [0] = wA [0], and for all A, A′ we have
2k
(vA [1] + wA′ [1]) · · · (vA [k] + wA′ [k]) =
∑ v′A [ j] · w′A [ j].
′
j=1
Now our goal is to compute
2k
∑
(v′A ,w′A′ ) : v′A [0]+w′A′ [0]=0
∑
j=1
!
v′A [ j] · w′A′ [ j]
.
(2)
We can get a more efficient algorithm for the problem defined by (2), by preprocessing the second half of
vectors (i.e., the w′A′ vectors). For each distinct value e = w′A [0] ∈ R among the 2n/2 vectors in the second
half, we make a new (2k + 1)-dimensional vector We′ where:
15
• We′ [0] = e, and
• for all i = 1, . . . , 2k , We′ [i] = ∑w′A : w′A [0]=e w′A [i].
That is, the coordinates 1, . . . , 2k of We′ are obtained by component-wise summing all vectors w′A such that
w′A [0] = e. The preparation of the vectors We′ can be done in 2n/2 · poly(k, n, logW ) time, by partitioning all
2n/2 vectors w′A from the second half of variables into equivalence classes (where two vectors are equivalent
if their 0-coordinates are equal), then obtaining each We′ by summing the vectors in one equivalence class.
Finally, we can use the WA′ ′ vectors to compute the sum (2) in 2n/2 · 2k · poly(k, n, logW ) time. Have a
running sum that is initially 0. Iterate through each vector v′A from the first half of variables, look up the
corresponding second-half vector We′ (with v′A [0] = −We′ [0]) in poly(k, n, logW ) time, and add the inner
product
2k
∑ v′A [i] ·We′ [i]
i=1
to the running sum. Because each vector (We′ [1], . . . ,We′ [2k ]) is the sum of all
k
such that v′A [0] + w′A′ [0] = 0, each inner product ∑2i=1 v′A [i] ·We′ [i] contributes
!
2k
w′A′
∑ v′A [ j] · w′A[ j]
∑
:
vectors (w′A′ [1], . . . , w′A′ [2k ])
v′A [0]+w′A′ [0]=0
j=1
to the running sum. Therefore after iterating through all vectors v′A , our running sum has computed (2)
exactly, in only 2n/2 · 2k · poly(n, logW ) time.
From the algorithm of Theorem 5.1, we immediately obtain the SUM ◦ ReLU lower bounds of Theorem 1.2.
6 Sparse Combinations of Low-Degree Polynomials over Finite Fields
We can also prove lower bounds for linear combinations of low-degree F p -polynomials in n variables, for
any prime p, by giving a faster Sum-Product algorithm. In this context, the Sum-Product problem becomes:
Sum-Product over MOD p ◦ ANDd : Given k polynomials p1 , . . . , pk ∈ F p [x1 , . . . , xn ], each of degree
at most d, compute
!
k
∑
x∈{0,1}n
∏ pi (x)
,
i=1
where the sum over all x ∈ {0, 1}n is taken over the reals (or rationals).
That is, we treat each ∏ki=1 pi (x) as a function from {0, 1}n to {0, 1, . . . , p − 1} ⊂ Q, and wish to compute
the sum of these integers over all x ∈ {0, 1}n .
In related work, Lokshtanov et al. [LPT+ 17] showed how to (deterministically) count solutions in Fnp to
6/7
a system of ℓ degree-d F p -polynomials in pn+o(n)−n/O(d p ) · poly(ℓ) time. For our Sum-Product problem,
we need to compute a “weighted” sum (the terms can take on values in {0, . . . , p − 1}), and we need to
count the weighted sum over only Boolean assignments. We can achieve this, with a comparable runtime
savings involving k and p:
Theorem 6.1. The Sum-Product of k degree-d polynomials p1 , . . . , pk ∈ F p [x1 , . . . , xn ] can be computed in
p2k · (1.9n + 2n−n/(6d p) ) · poly(n) time.
16
Proof. Let p1 , . . . , pk be given. We wish to compute
k
∑
x∈{0,1}n
!
∏ pi (x)
i=1
,
(3)
where each product outputs an integer in {0, 1, . . . , p − 1}. We first convert the Sum-Product problem of (3)
to an equivalent sum where each “term” in the sum is a small system of polynomial equations.
We say that a function f : {0, 1}n → {0, 1} is an exact F p -polynomial function if there is a polynomial
p ∈ F p [x1 , . . . , xn ] and a ∈ F p such that for all x ∈ {0, 1}n ,
f (x) = 1 ⇐⇒ p(x) = a.
We use the notation [p(x) = a] to denote such an exact polynomial function. Let us replace each polynomial
pi (x) in the sum-product expression with an equivalent linear combination (over Z) of exact polynomial
functions. In particular, replace each pi (x) with the sum over the integers
∑ a · [pi (x) = a].
a∈F p
That is, we are replacing pi (a) with an equivalent integer-valued sum of p Boolean functions. Now the
desired sum (3) looks like:
!!
k
k
∑
x∈{0,1}n
∏ ∑
i=1
a · [pi (x) = a]
=
∑
∑
x∈{0,1}n (a1 ,...,ak )∈Fkp
a∈F p
=
∑
a1 · · · ak · ∏[pi (x) = ai ] (by distributivity)
a1 · · · ak ·
(a1 ,...,ak )∈Fkp
i=1
∑
x∈{0,1}n
!
[p1 (x) = a1 ] · · · [pk (x) = ak ] . (4)
Each inner sum in (4) counts the number of Boolean solutions to a system of polynomial equations p1 (x) =
a1 , . . . , pk (x) = ak . We can further reduce this problem to counting the number of Boolean solutions to one
equation, by applying a simple reduction (from [Wil18]). Namely, we have the equation
#!
# "
"
k
k
k
1
∑ ∏[pi (x) = ai ] = pk ∑ k ∑ n ∑ b j · (p j (x) − a j ) = 0 − ∑ b j · (p j (x) − a j ) = 1 .
j=1
j=1
x∈{0,1}n i=1
(b ,...,b )∈F x∈{0,1}
1
k
p
(5)
To see why (5) holds, let x ∈ {0, 1}n such that [p1 (x) = a1 ] · · · [pk (x) = ak ] = 1. Then for every (b1 , . . . , bk ) ∈
Fkp , we have [∑kj=1 b j · (p j (x) − a j ) = 0] = 1. So every solution x to the system of k equations is counted
for pk times in (5); since the result is divided by pk , each solution contributes 1 to (5). On the other hand,
if x is not a solution to the system, and [p1 (x) = a1 ] · · · [pk (x) = ak ] = 0, then for some j, p j (a) − a j 6= 0. It
follows that there are precisely pk−1 vectors (b1 , . . . , bk ) ∈ Fkp such that [∑kj=1 b j · (p j (x) − a j ) = 0] = 1, and
there are precisely pk−1 (other) vectors (b′1 , . . . , b′k ) ∈ Fkp such that [∑kj=1 b′j · (p j (x) − a j ) = 1] = 1. These
two equal counts cancel out in the sum of (5), so non-solutions to the system contribute 0 to the sum of (5).
Putting (4) and (5) together, the original Sum-Product problem (3) can now be reduced to the computation
of O(p2k ) sums, each of the form
∑ [q(x1 , . . . , xn ) = 0],
x∈{0,1}n
17
where q is an F p -polynomial of degree at most d. That is, to obtain (3), we only need to count the Boolean
roots of O(p2k ) polynomials q, and take the appropriate R-linear combination of these counts.
Let us now focus on counting roots to a single polynomial q(x1 , . . . , xn ) of degree d. Let Pℓ (z) be the
modulus-amplifying polynomial of degree 2ℓ − 1, from Theorem 2.1. Let δ ∈ (0, 1/2) be a parameter, and
consider the following “reduced” polynomial in n − δ n variables, over the integers:
Q(x1 , . . . , xn−δ n ) :=
∑
Pδ n (1 − q(x1 , . . . , xn−δ n , a1 , . . . , aδ n ) p−1 ).
a1 ,...,aδ n ∈{0,1}
Note that Q has degree less than 2d pδ n. Set δ = 1/(6d p), and note that 2d pδ n < (n − δ n)/2. Over F p ,
the polynomial 1 − q(x) p−1 equals 1 mod p if x is a root of q, and is 0 mod p otherwise. Applying the
modulus-amplifying properties of Pδ n , we have:
• If x is a root of q, then Pδ n (1 − q(x) p−1 ) = 1 mod pδ n .
• If x is not a root of q, then Pδ n (1 − q(x) p−1 ) = 0 mod pδ n .
As the sum in Q is over only 2δ n such Pδ (· · · ) terms, and p ≥ 2, we conclude that for all b1 , . . . , bn−δ n ∈
{0, 1}, the quantity (Q(b1 , . . . , bn−δ n ) mod pδ n ) equals the number of a1 , . . . , aδ n ∈ {0, 1} such that
q(b1 , . . . , bn−δ n , a1 , . . . , aδ n ) = 0.
Therefore if we evaluate the polynomial Q over all 2n−δ n Boolean assignments (b1 , . . . , bn−δ n ), compute
each value separately modulo pδ n , then sum those values over the integers, we will obtain the number of
Boolean roots of q.
Over Boolean assignments, we may assume without loss of generality that Q is multilinear (i.e. x2i = xi
for all i). Since 2d pδ n < (n − δ n)/2, standard properties of binomial coefficients imply that the number of
monomials of Q is
n−δn
O
.
2d pδ n
and
By constructing Q term-by-term (expanding each Pδ n (1 − q(x1 , . . . , xn−δ n , a1 , . . . , aδ n ) p−1 ) one-by-one,
n−δ n
+
adding them to a running sum, similar to [CW16, LPT 17]), we may represent Q as a sum of O 2d pδ n
n−δ n
monomials, constructed in poly(n) · 2d
pδ n time. Letting δ = 1/(6d p), the number of monomials of Q
n
is less than n/3
≤ 1.9n . Applying the fast polynomial evaluation algorithm of Theorem 2.4, Q can be
evaluated on all 2n−n/(6d p) Boolean assignments in time (1.9n + 2n−n/(6d p) ) · poly(n) time.
Therefore, for every fixed degree d and prime p, there is an ε > 0 such that the relevant Sum-Product
problem is in 2n−ε n · poly(n) time. This immediately implies the lower bounds of Theorems 1.3 and 1.4.
In particular, to prove 1.4 we apply Theorem 3.3. Fix an integer degree d, and let c ≥ 1 be the universal
constant (from Theorem 3.3) such that we need to solve Sum-Product for MODp ◦ ANDd ◦ ANY c circuits.
Converting to SUM ◦ MODp ◦ ANDdc , Theorem 6.1 says that the Sum-Product problem can be solved in
2n−n/O(dc) time (omitting low-order terms).
7 Conclusion
Applying old and new tools, we have established several strong new lower bounds for representing Boolean
functions in different regimes. Among the most interesting open problems remaining, we find the Quadratic
Uncertainty Principle (that AND requires a large R-linear combination of quadratic F2 -polynomials) to
18
be especially intriguing. Quadratic polynomials have special properties that higher degrees do not; for
example, one can count the roots of a given quadratic F p -polynomial in polynomial time (see [Wil18] for
a recent application of this phenomenon). Therefore in some cases, our 2n−ε n -time algorithms become
poly(n)-time algorithms. This should imply lower bounds for functions in P against linear combinations
of quadratic F2 -polynomials, perhaps even lower bounds against the AND function, but so far we have not
yet been able to prove such bounds.
A longstanding problem in circuit complexity—seemingly related to the Quadratic Uncertainty Principle—
is the Constant Degree Hypothesis of Barrington, Straubing, and Therien [BST90]:
Hypothesis 2 (Constant Degree Hypothesis (CDH)). For every constant d ≥ 1 and primes p, q, there is an
ε > 0 such that the AND function on n variables cannot be computed by MOD p ◦ MODq ◦ ANDd circuits
of 2ε n size.
The CDH is currently only known to be true for d = 1, and for p = q. Can the techniques of this paper
say anything about such problems, even for the case of d = 2?
Acknowledgements. I thank Lijie Chen, Pooya Hatami, Adam Klivans, Shachar Lovett, and Anirbit
Mukherjee for comments and discussions on the topics of this paper. In particular, I am grateful to Shachar
for noticing a gap in a lemma in an earlier version of this paper. I am also grateful to Brynmor Chapman
for his proofreading, and patience with my explanations regarding this paper.
References
[ABMM16] Raman Arora, Amitabh Basu, Poorya Mianjy, and Anirbit Mukherjee. Understanding deep
neural networks with rectified linear units. arXiv preprint arXiv:1611.01491, 2016.
[ACW16]
Josh Alman, Timothy M. Chan, and R. Ryan Williams. Polynomial representations of threshold functions and algorithmic applications. In FOCS, pages 467–476, 2016.
[BHK09]
Andreas Björklund, Thore Husfeldt, and Mikko Koivisto. Set partitioning via inclusionexclusion. SIAM J. Comput., 39(2):546–563, 2009.
[BHPS10]
László Babai, Kristoffer Arnsfelt Hansen, Vladimir V. Podolskii, and Xiaoming Sun. Weights
of exact threshold functions. In Mathematical Foundations of Computer Science, pages 66–
77, 2010.
[Bou05]
Estimation of certain exponential sums arising in complexity theory. C.R. Acad. Sci. Paris
Ser. I, 340:627–631, 2005.
[BST90]
David A. Mix Barrington, Howard Straubing, and Denis Thérien. Non-uniform automata over
groups. Inf. Comput., 89(2):109–132, 1990.
[BSV14]
Eli Ben-Sasson and Emanuele Viola. Short PCPs with projection queries. In ICALP, pages
163–173, 2014.
[BT94]
Richard Beigel and Jun Tarui. On ACC. Computational Complexity, pages 350–366, 1994.
[Cal08]
Chris Calabro. A lower bound on the size of series-parallel graphs dense in long paths. Electronic Colloquium on Computational Complexity (ECCC), 15(110), 2008.
19
[CGT96]
Jin-yi Cai, Frederic Green, and Thomas Thierauf. On the correlation of symmetric functions.
Mathematical Systems Theory, 29(3):245–258, 1996.
[CM17]
Arkadev Chattopadhyay and Nikhil S. Mande. Weights at the bottom matter when the top is
heavy. Electronic Colloquium on Computational Complexity (ECCC), 24:83, 2017.
[CSS16]
Ruiwen Chen, Rahul Santhanam, and Srikanth Srinivasan. Average-case lower bounds and
satisfiability algorithms for small threshold circuits. In CCC, pages 1:1–1:35, 2016.
[CW16]
Timothy M. Chan and Ryan Williams. Deterministic APSP, Orthogonal Vectors, and more:
Quickly derandomizing Razborov-Smolensky. In SODA, pages 1246–1255, 2016.
[Dan17]
Amit Daniely. Depth separation for neural networks. In Proceedings of COLT, pages 690–
696, 2017.
[ES16]
Ronen Eldan and Ohad Shamir. The power of depth for feedforward neural networks. In
Proceedings of COLT, pages 907–940, 2016.
[FHH+ 14]
Yuval Filmus, Hamed Hatami, Steven Heilman, Elchanan Mossel, Ryan O’Donnell, Sushant
Sachdeva, Andrew Wan, and Karl Wimmer. Real Analysis in Computer Science: A collection
of open problems, Simons Institute, 2014.
[Gre04]
Frederic Green. The correlation between parity and quadratic polynomials mod3. Journal of
Computer and System Sciences, 69(1):28–44, 2004.
[GT12]
Anna Gál and Vladimir Trifonov. On the correlation between parity and modular polynomials.
Theory Comput. Syst., 50(3):516–536, 2012.
[Had93]
Jacques Hadamard. Résolution d’une question relative aux déterminants. Bull. Sci. Math.,
17:30–31, 1893.
[HG91]
Johan Håstad and Mikael Goldmann. On the power of small-depth threshold circuits. Computational Complexity, 1:113–129, 1991.
[HHL16]
Hamed Hatami, Pooya Hatami, and Shachar Lovett. Higher-order fourier analysis and applications. Manuscript, 2016.
[HMP+ 93] András Hajnal, Wolfgang Maass, Pavel Pudlák, Mario Szegedy, and György Turán. Threshold
circuits of bounded depth. J. Comput. Syst. Sci., 46(2):129–154, 1993.
[HP10]
Kristoffer Arnsfelt Hansen and Vladimir V Podolskii. Exact threshold circuits. In CCC, pages
270–279, 2010.
[HS74]
Ellis Horowitz and Sartaj Sahni. Computing partitions with applications to the knapsack
problem. JACM, 21(2):277–292, 1974.
[JMV15]
Hamid Jahanjou, Eric Miles, and Emanuele Viola. Local reductions. In Proceedings of
ICALP, pages 749–760, 2015.
[KW16]
Daniel M. Kane and Ryan Williams. Super-linear gate and super-quadratic wire lower bounds
for depth-two and depth-three threshold circuits. In STOC, pages 633–643, 2016.
[Lov17]
Shachar Lovett. Personal communication, 2017.
20
[LPT+ 17]
Daniel Lokshtanov, Ramamohan Paturi, Suguru Tamaki, R. Ryan Williams, and Huacheng
Yu. Beating brute force for systems of polynomial equations over finite fields. In SODA,
pages 2190–2202, 2017.
[Maa97]
Wolfgang Maass. Bounds for the computational power and learning complexity of analog
neural nets. SIAM Journal on Computing, 26(3):708–732, 1997.
[MB17]
Anirbit Mukherjee and Amitabh Basu. Lower bounds over Boolean inputs for deep neural
networks with ReLU gates. ArXiv e-prints, November 2017.
[MNW99]
Peter Bro Miltersen, N. V. Vinodchandran, and Osamu Watanabe. Super-polynomial versus
half-exponential circuit size in the exponential hierarchy. In COCOON, Springer LNCS 1627,
pages 210–220, 1999.
[MTT61]
S. Muroga, I. Toda, and S. Takasu. Theory of majority decision elements. Journal of the
Franklin Institute, 271:376–418, 1961.
[MW17]
Cody Murray and Ryan Williams. Circuit lower bounds for nondeterministic quasi-polytime:
An easy witness lemma for NP and NQP. Electronic Colloquium on Computational Complexity (ECCC), TR17-188, 2017.
[Nis94]
Noam Nisan. The communication complexity of threshold gates. In Proceedings of “Combinatorics, Paul Erdos is Eighty”, pages 301–315, 1994.
[ROS94]
Vwani P. Roychowdhury, Alon Orlitsky, and Kai-Yeung Siu. Lower bounds on threshold and
related circuits via communication complexity. IEEE Transactions on Information Theory,
40(2):467–474, 1994.
[San09]
Rahul Santhanam. Circuit lower bounds for Merlin–Arthur classes.
39(3):1038–1061, 2009.
[SFM78]
Joel Seiferas, Michael Fischer, and Albert Meyer. Separating nondeterministic time complexity classes. Journal of the ACM, 25(1):146–167, January 1978.
[SS17]
Itay Safran and Ohad Shamir. Depth-width tradeoffs in approximating natural functions with
neural networks. In International Conference on Machine Learning, pages 2979–2987, 2017.
[SW13]
Rahul Santhanam and Ryan Williams. On medium-uniformity and circuit lower bounds. In
IEEE Conf. Computational Complexity, pages 15–23, 2013.
[Tam16]
Suguru Tamaki. A satisfiability algorithm for depth two circuits with a sub-quadratic number of symmetric and threshold gates. Electronic Colloquium on Computational Complexity
(ECCC), 23:100, 2016.
[Tel16]
Matus Telgarsky. benefits of depth in neural networks. In Proceedings of COLT, pages 1517–
1539, 2016.
[Tel18]
Roei Tell. Proving that prBPP=prP is as hard as “almost” proving that P 6= NP. Electronic
Colloquium on Computational Complexity (ECCC), 18(3), 2018.
[Tod91]
S. Toda. PP is as hard as the polynomial-time hierarchy. SIAM Journal on Computing,
20(5):865–877, 1991.
21
SIAM J. Comput.,
[Val77]
L. G. Valiant. Graph-theoretic arguments in low-level complexity. In J. Gruska, editor, MFCS,
volume 53 of LNCS, pages 162–176, Tatranská Lomnica, Czechoslovakia, September 1977.
Springer.
[Vio09]
Emanuele Viola. Guest column: correlation bounds for polynomials over {0, 1}. SIGACT
News, 40(1):27–44, 2009.
[Wil11]
Ryan Williams. A casual tour around a circuit complexity bound. SIGACT News, 42(3):54–76,
2011.
[Wil13]
Ryan Williams. Improving exhaustive search implies superpolynomial lower bounds. SIAM
Journal on Computing, 42(3):1218–1244, 2013.
[Wil14a]
Ryan Williams. New algorithms and lower bounds for circuits with linear threshold gates. In
STOC, pages 194–202, 2014.
[Wil14b]
Ryan Williams. Nonuniform ACC circuit lower bounds. JACM, 61(1):2, 2014.
[Wil18]
Ryan Williams. Counting solutions to polynomial systems via reductions. In Raimund Seidel,
editor, 1st Symposium on Simplicity in Algorithms (SOSA 2018), pages 6:1–6:15, 2018.
[Win60]
R. O. Winder. Threshold Logic. PhD thesis, Princeton University, 1962. Preliminary version
in FOCS’60.
[Žák83]
Stanislav Žák. A Turing machine time hierarchy. Theoretical Computer Science, 26(3):327–
333, 1983.
A
Linear Lower Bound for AND With Sums of Quadratic Polynomials
For reference, we report a folklore Ω(n) lower bound on representing AND with linear combinations of
quadratic F2 -polynomials (recall it is conjectured that the sparsity lower bound is 2Ω(n) ). The below proof
was communicated to us by Shachar Lovett.
Theorem A.1 (Lovett [Lov17]). The AND function on n inputs does not have SUM◦MOD2◦AND2 circuits
of sparsity less than n/2.
Proof. Let f : {0, 1}n → {0, 1} be the NOR function (which by DeMorgan’s laws has the same sparsity as
AND). Suppose we can write
s
f (x) = ∑ αi (−1)qi (x) ,
i=1
where the qi (x) are quadratic F2 -polynomials, and all αi ∈ R. Note that without loss of generality we may
assume qi (0) = 0 for all i (if qi (0) = 1, then replacing αi by −αi and qi (x) by qi (x) + 1 yields an equivalent
expression). If s < n/2, then by the ChevalleyWarning theorem, the number of common roots of {q1 , ..., qr }
is divisible by 2. But then there is another common root x⋆ , so f (0) = f (x⋆ ), contradicting the definition of
NOR.
22
| 9cs.NE
|
Noname manuscript No.
(will be inserted by the editor)
Positive-instance driven dynamic programming for treewidth ⋆
arXiv:1704.05286v2 [cs.DS] 12 Jan 2018
Hisao Tamaki
Received: date / Accepted: date
Abstract Consider a dynamic programming scheme for a decision problem in which all
subproblems involved are also decision problems. An implementation of such a scheme is
positive-instance driven (PID), if it generates positive subproblem instances, but not negative
ones, building each on smaller positive instances.
We take the dynamic programming scheme due to Bouchitté and Todinca for treewidth
computation, which is based on minimal separators and potential maximal cliques, and design a variant (for the decision version of the problem) with a natural PID implementation.
The resulting algorithm performs extremely well: it solves a number of standard benchmark
instances for which the optimal solutions have not previously been known. Incorporating
a new heuristic algorithm for detecting safe separators, it also solves all of the 100 public instances posed by the exact treewidth track in PACE 2017, a competition on algorithm
implementation.
We describe the algorithm, prove its correctness, and give a running time bound in terms
of the number of positive subproblem instances. We perform an experimental analysis which
supports the practical importance of such a bound.
Keywords treewidth · tree decomposition · dynamic programming · positive-instance
driven
1 Introduction
Suppose we design a dynamic programming algorithm for some decision problem, formulating subproblems, which are decision problems as well, and recurrences among those subproblems. A standard approach is to list all subproblem instances and scan the list from
“small” ones to “large” , deciding the answer, positive or negative, to each instance by means
of these recurrences. When the number of positive subproblem instances are expected to be
⋆ A preliminary and an abridged version of this paper was presented at the 25th European Sysmposium on
Algorithms
H. Tamaki
Kawasaki, 214-8571, Japan, Meiji University, 1-1-1 Higashi-Mata, Tama, Kawasaki, 214-8571, Japan,
Tel.: +81-44-934-7478
E-mail: [email protected]
2
Hisao Tamaki
much smaller than the total number of subproblem instances, a natural alternative is to generate positive instances only, using recurrences to combine positive instances to generate a
“larger” positive instance. We call such a mode of dynamic programming execution positiveinstance driven or PID for short. One goal of this paper is to demonstrate that PID is not
simply a low-level implementation strategy but can be a paradigm of algorithm design for
some problems.
The decision problem we consider is that of deciding, given graph G and positive integer k, if the treewidth of G is at most k. This graph parameter was introduced by Robertson
and Seymour [17] and has had a tremendous impact on graph theory and on the design of
graph algorithms (see, for example, a survey [7]). The treewidth problem is NP-complete
[1] but fixed-parameter tractable: it has an f (k)nO(1) time algorithm for some fixed function
f (k) as implied by the graph minor theorem of Robertson and Seymour [18], and an explicit
O( f (k)n) time algorithm was given by Bodlaender [3]. A classical dynamic programming
algorithm due to Arnborg, Corneil, and Proskurowsky (ACP algorithm) [1] runs in nk+O(1)
time. Bouchitté and Todinca [9] developed a more refined dynamic programming algorithm
(BT algorithm) based on the notions of minimal separators and potential maximal cliques,
which lead to algorithms running in O(1.7549n ) time or in O(n5 ⌈(2n+k+8)/3⌉
) time [11,
k+2
12]. Another important approach to treewidth computation is based on the perfect elimination order (PEO) of minimal chordal completions of the given graph. PEO-based dynamic
programming algorithms run in O∗ (2n ) time with exponential space and in O∗ (4n ) time with
polynomial space [4], where O∗ ( f (n)) means O(nc f (n)) for some constant c.
There has been a considerable amount of effort on implementing treewidth algorithms
to be used in practice and, prior to this work, the most successful implementations for exact
treewidth computation are all based on PEO. The authors of [4] implemented the O∗ (2n )
time dynamic programming algorithm and experimented on its performance, showing that
it works well for small instances. For larger instances, PEO-based branch-and-bound algorithms are known to work well in practice [13]. Recent proposals for reducing treewidth
computation to SAT solving are also based on PEO [19, 2]. From the PID perspective, this
situation is somewhat surprising, since it can be shown that each positive subproblem instance in the PEO-based dynamic programming scheme corresponds to a combination of
an indefinite number of positive subproblem instances in the ACP algorithm, and hence the
number of positive subproblem instances can be exponentially larger than that in the ACP algorithm. Indeed, a PID variant of the ACP algorithm was implemented by the present author
and has won the first place in the exact treewidth track of PACE 2016 [10], a competition
on algorithm implementations, outperforming other submissions based on PEO. Given this
success, a natural next step is to design a PID variant of the BT algorithm, which is tackled
in this paper.
The resulting algorithm performs extremely well, as reported in Section 8. It is tested on
DIMACS graph-coloring instances [14], which have been used in the literature on treewidth
computation as standard benchmark instances [13, 5, 15, 19, 4, 2]. Our implementation of the
algorithm solves all the instances that have been previously solved (that is, with matching upper and lower bounds known) within 10 seconds per instance on a typical desktop
computer and solves 13 out of the 42 previously unsolved instances. For nearly half of the
instances which it leaves unsolved, it significantly reduces the gap between the lower and
upper bounds. It is interesting to note that this is done by improving the lower bound. Since
the number of positive subproblem instances are much smaller when k < tw(G) than when
k = tw(G), the PID approach is particularly good at establishing strong lower bounds.
We also adopt the notion of safe separators due to Bodlaender and Koster [5] in our
preprocessing and design a new heuristic algorithm for detecting safe separators. With this
Positive-instance driven dynamic programming for treewidth ⋆
3
preprocessing, our implementation also solves all of the 100 public instances posed by PACE
2017 [16], the successor of PACE 2016. It should be noted that these test instances of PACE
2017 are much harder than those of PACE 2016: the winning implementation of PACE 2016
mentioned above, which solved 199 of the 200 instances therein, solves only 62 of these 100
instances of PACE 2017 in the given time of 30 minutes per instance.
Adapting the BT algorithm to work in PID mode has turned out non-trivial. Each subproblem instance in the BT algorithm for given graph G and positive integer k takes the
form of a connected set C of G such that NG (C), the open neighborhood of C in G, is a
minimal separator of G with cardinality at most k. For each such C, we ask if C is feasible,
in the sense that there is a tree decomposition of the subgraph of G induced by C ∪ NG (C)
of width at most k that has a bag containing NG (C) (see Section 2 for the definition of a
tree decomposition of a graph). The difficulty of making the BT algorithm PID comes from
the fact that the recurrence for deciding if C is feasible may involve an indefinite number of
connected sets C ′ such that C ′ ⊂ C. Thus, even if the number of positive instances is small,
there is a possibility that the running time is exponential in that number. We approach this
issue by introducing an auxiliary structure we call O-blocks (see Section 3) and formulate a
recurrences that are binary: a combination of a feasible connected set and a feasible O-block
may yield either a larger feasible connected set or a larger feasible O-block. Due to this
binary recurrence, we obtain an upper bound on the running time of our algorithm which
is sensitive to the number of subproblem instances (Observation 1 in Section 5). To support
the significance of such a bound, we perform an experimental analysis which shows the existence of huge gaps between the actual number of combinatorial objects corresponding to
subproblems and the known theoretical upper bounds.
The rest of this paper is organized as follows. In Section 2, we introduce notation, define
basic concepts and review facts in the literature. In Section 3, we precisely define the subproblems in our dynamic programming algorithm and formulate recurrences. We describe
our algorithm and prove its correctness in Section 4 and then analyze its running time in
Section 5. In Section 6, we describe our experimental analysis. In Section 7, we describe
some implementation details. Finally, in Section 8, we give details of the performance results sketched above.
2 Preliminaries
In this paper, all graphs are simple, that is, without self loops or parallel edges. Let G be
a graph. We denote by V (G) the vertex set of G and by E(G) the edge set of G. For each
v ∈ V (G), NG (v) denote the set of neighbors of v in G: NG (v) = {u ∈ V (G) | {u, v} ∈ E(G).
For U ⊆ V (G), the open neighborhood of U in G, denoted by NG (U), is the set of vertices
S
adjacent to some vertex in U but not belonging to U itself: NG (U) = ( v∈U NG (v)) \U. The
closed neighborhood of U in G, denoted by NG [U], is defined by NG [U] = U ∪ NG (U). We
also write NG [v] for NG [{v}] = NG (v) ∪ {v}. We denote by G[U] the subgraph of G induced
by U: V (G[U]) = U and E(G[U]) = {{u, v} ∈ E(G) | u, v ∈ U}. In the above notation, as
well as in the notation further introduced below, we will often drop the subscript G when
the graph is clear from the context.
We say that vertex set C ⊆ V (G) is connected in G if, for every u, v ∈ C, there is a path
in G[C] between u and v. It is a connected component of G if it is connected and is inclusionwise maximal subject to this condition. A vertex set C in G is a component associated with
S ⊆ G, if C is a connected component of G[V (G) \ S]. For each S ⊆ V (G), we denote by
CG (S) the set of all components associated with S. A vertex set S ⊆ V (G) is a separator of
4
Hisao Tamaki
G if |CG (S)| > CG (0)|,
/ that is, if its removal increases the number of connected components
of G. A component C associated with separator S of G is a full component if NG (C) = S. A
separator S is a minimal separator if there are at least two full components associated with
S. This term is justified by this fact: if S is a minimal separator and a, b vertices belonging
to two distinct full components associated with S, then for every proper subset S′ of S, a
and b belong to the same component associated with S′ ; S is a minimal set of vertices that
separates a from b. A block is a pair (S,C), where S is a separator and C is a component
associated with S; it is a full block if C is a full component, that is, S = N(C).
Graph H is chordal if every induced cycle of H has length exactly three. H is a minimal
chordal completion of G if it is chordal, V (H) = V (G), E(G) ⊆ E(H), and E(H) is minimal
subject to these conditions. A vertex set Ω ⊆ V (G) is a potential maximal clique of G, if Ω
is a clique in some minimal chordal completion of G.
A tree-decomposition of G is a pair (T, X ) where T is a tree and X is a family
{Xi }i∈V (T ) of vertex sets of G such that the following three conditions are satisfied. We
call members of V (T ) nodes of T and each Xi the bag at node i.
1. i∈V (T ) Xi = V (G).
2. For each edge {u, v} ∈ E(G), there is some i ∈ V (T ) such that u, v ∈ Xi .
3. The set of nodes Iv = {i ∈ V (T ) | v ∈ Xi } of V (T ) induces a connected subtree of T .
S
The width of this tree-decomposition is maxi∈V (T ) |Xi | − 1. The treewidth of G, denoted by
tw(G) is the minimum width of all tree-decompositions of G. We may assume that the bags
Xi and X j are distinct from each other for i 6= j and, under this assumption, we will often
regard a tree-decomposition as a tree T in which each node is a bag.
We call a tree-decomposition T of G canonical if each bag of T is a potential maximal
clique of G and, for every pair X, Y of adjacent bags in T , X ∩ Y is a minimal separator of
G. The following fact is well-known. It easily follows, for example, from Proposition 2.4 in
[8].
Lemma 1 Let G be an arbitrary graph. There is a tree-decomposition T of G of width
tw(G) that is canonical.
The following local characterization of a potential maximal clique is crucial. We say
that a vertex set S ⊆ V (G) is cliquish in G if, for every pair of distinct vertices u and v in S,
either u and v are adjacent to each other or there is some C ∈ C (S) such that u, v ∈ N(C).
In other words, S is cliquish if completing N(C) for every C ∈ C (S) into a clique makes S a
clique.
Lemma 2 (Theorem 3.15 in [8]) A separator S of G is a potential maximal clique of G if
and only if (1) S has no full-component associated with it and (2) S is cliquish.
It is also shown in [8] that if Ω is a potential maximal clique of G and S is a minimal separator contained in Ω , then there is a unique component CS associated with S that contains
Ω \ S. We need an explicit way of forming CS from Ω and S.
Let K ⊆ V (G) be an arbitrary vertex set and S an arbitrary proper subset of K. We say
that a component C ∈ C (K) is confined to S if N(C) ⊆ S; otherwise it is unconfined to S.
Let unconf(S, K) denote the set of components associated with K that are unconfined to S.
S
Define the crib of S with respect to K, denoted by crib(S, K), to be (K \ S) ∪ C∈unconf(S,K) C:
it is the union of K \ S and all those components associated with K that have neighborhoods
intersecting K \ S.
The following lemma relies only on the second property of potential maximal cliques,
namely that they are cliquish, and will be applied not only to potential maximal cliques but
also to separators with full components, which are trivially cliquish.
Positive-instance driven dynamic programming for treewidth ⋆
5
Lemma 3 Let K ⊆ V (G) be a cliquish vertex set. Let S be an arbitrary proper subset of K.
Then, crib(S, K) is a full component associated with S.
Proof Let C = crib(S, K). We first show that G[C] is connected. Suppose K \ S has two
distinct vertices u and v. Since K is cliquish, either u and v are adjacent to each other or there
is some component C ′ ∈ C (K) such that u, v ∈ N(C ′ ). In the latter case, as C ′ is unconfined
to S, we have C ′ ⊆ C. Therefore, u and v belong to the same connected component of G[C].
As this applies to every pair of vertices in K \ S, K \ S is contained in a single connected
component of G[C]. Moreover, each component C ′ ∈ C (K) contained in C is unconfined to
S, by the definition of crib(S, K), and hence has a neighbor in K \ S. Therefore, we conclude
that G[C] is connected. Each vertex v not in S ∪ C belongs to some component in C (K)
that is confined to S and hence does not have a neighbor in C. Therefore, C is a component
associated with S.
To see that C is a full component, let u ∈ S and v ∈ K \ S be arbitrary. Since K is cliquish,
either u and v are adjacent to each other or there is some C ′ ∈ C (K) such that u, v ∈ N(C ′ ).
As such C ′ is unconfined to S in the latter case, we conclude that u ∈ N(C) in either case.
Since this holds for arbitrary u ∈ S, we conclude that C is a full component associated with
S.
⊔
⊓
Remark 1 As crib(S, K) contains K \ S, it is clear that it is the only component associated
with S that intersects K. Therefore, the above mentioned assertion on potential maximal
cliques is a corollary to this Lemma.
3 Recurrences on oriented minimal separators
In this section, we fix graph G and positive integer k that are given in the problem instance:
we are to decide if the treewidth of G is at most k. We assume that G is connected.
For connected set C ⊆ V (G), we denote by GhCi the graph obtained from G[N[C]] by
completing N(C) into a clique: V (GhCi) = N[C] and E(GhCi) = E(G[N[C]]) ∪ {{u, v} |
u, v ∈ N(C), u 6= v}. We say C is feasible if tw(GhCi) ≤ k. Equivalently, C is feasible if
G[N[C]] has a tree-decomposition of width k or smaller that has a bag containing N(C).
Let us first review the BT algorithm [8] adapting it to our decision problem. We first
list all minimum separators of cardinality k or smaller and all potential maximal cliques
of cardinality k + 1 or smaller. Then, for each pair of a potential maximal clique Ω and a
minimal separator S such that S ⊂ Ω , place a link from S to Ω . To understand the difficulty
of formulating a PID variant of the algorithm, it is important to note that the pair (Ω , S) to
be linked is easy to find from the side of Ω , but not the other way round. Then, we scan the
full blocks (N(C),C) of minimal separators in the increasing order of |C| to decide if C is
feasible, using the following recurrence: C is feasible if and only if there is some potential
maximal clique Ω such that N(C) ⊂ Ω , C = crib(N(C), Ω ), and every component D ∈
unconf(N(C), Ω ) is feasible. Finally, we have tw(G) ≤ k if and only if there is a potential
maximal clique Ω with |Ω | ≤ k +1 such that every component associated with Ω is feasible.
To facilitate the PID construction, we orient minimal separators as follows. We assume a
total order < on V (G). For each vertex set U ⊆ V (G), the minimum element of U, denoted by
min(U), is the smallest element of U under <. For vertex sets U and W , we say U precedes
W and write U ≺ W if min(U) < min(W ).
We say that a connected set C is inbound if there is some full block associated with
N(C) that precedes C; otherwise, it is outbound. Observe that if C is inbound then N(C) is
6
Hisao Tamaki
a minimal separator, since N(C) has another full component associated with it and, contrapositively, if N(C) is not a minimal separator then C is necessarily outbound. We say a full
block (N(C),C) is inbound (outbound) if C is inbound (outbound, respectively).
Lemma 4 Let K be a cliquish vertex set and let A1 , A2 be two components associated with
K. Suppose that A1 and A2 are outbound. Then, either N(A1 ) ⊆ N(A2 ) or N(A2 ) ⊆ N(A1 ).
Proof Let K, A1 , and A2 be as above and suppose neither of N(A1 ) and N(A2 ) is a subset
of the other. For i = 1, 2, let Ci = crib(N(Ai ), K). Since N(A2 ) \ N(A1 ) is non-empty and
contained in K \ N(A1 ), A2 is contained in C1 . We have A1 ≺ C1 as A1 is outbound and hence
A1 ≺ A2 . A contradiction, since similarly we have A2 ≺ A1 .
⊔
⊓
Let K be a cliquish vertex set. Based on the above lemma, we define the outlet of
K, denoted by outlet(K), as follows. If no non-full component associated with K is outbound, then we let outlet(K) = 0.
/ Otherwise, outlet(K) = N(A), where A is a non-full component associated with K that is outbound, chosen so that N(A) is maximal. We define
support(K) = unconf(outlet(K), K), the set of components associated with K that are not
confined to outlet(K). By Lemma 4, every member of support(K) is inbound.
We call a full block (N(C),C) an I-block if C is inbound and |N(C)| ≤ k. We call it an
O-block if C is outbound and |N(C)| ≤ k.
We say that an I-block (N(C),C) is feasible if C is feasible. We say that an O-block
S
(N(A), A) is feasible if N(A) = C∈C N(C) for some set C of feasible inbound components.
Note that this definition of feasibility of an O-block is somewhat weak in the sense that we
do not require every inbound component associated with N(A) to be feasible.
We say that a potential maximal clique Ω is feasible if |Ω | ≤ k + 1 and every C ∈
support(Ω ) is feasible.
In order to formulate mutual recurrences among feasible I-blocks, O-blocks, and potential maximal cliques, we need the following auxiliary notion of buildable potential maximal
cliques.
Let Ω be a potential maximal clique with |Ω | ≤ k + 1. For each C ∈ support(Ω ), block
(N(C),C) is an I-block, since C is inbound as observed above and we have |N(C)| ≤ k by
our assumption that |Ω | ≤ k + 1. We say that Ω is buildable if |Ω | ≤ k + 1 and either
1. Ω = N[v] for some v ∈ V (G),
S
2. there is some subset C of support(Ω ) such that Ω = D∈C N(D) and every member of
C is feasible, or
3. Ω = N(A) ∪ (N(v) ∩ A) for some feasible O-block (N(A), A) and a vertex v ∈ N(A).
It will turn out that every feasible potential maximal clique is buildable (Lemma 9).
Lemma 5 We have tw(G) ≤ k if and only if G has a feasible potential maximal clique Ω
with outlet(Ω ) = 0.
/
/
Proof Suppose first that G has a feasible potential maximal clique Ω with outlet(Ω ) = 0.
Note that support(Ω ) = C (Ω ), as every C ∈ C (Ω ) is unconfined to an empty set. For each
component C ∈ support(Ω ), let TC be the tree-decomposition of GhCi of width k or smaller,
which exists since C is feasible by the definition of a feasible potential maximal clique. Let
XC be a bag of TC such that N(C) ⊆ XC . Combine these tree-decompositions into a tree T
by adding bag Ω and letting each XC in TC be adjacent to Ω . That T satisfies the first two
conditions for tree decomposition is trivial. The third condition is also satisfied, since, if a
vertex v appears in N[C] for two or more members C in support(Ω ), then v appears in XC
Positive-instance driven dynamic programming for treewidth ⋆
7
for each such C and in Ω . Therefore, T is a tree decomposition of G of width k or smaller
and hence tw(G) ≤ k.
For the converse, suppose the treewidth of G is k or smaller. Let T be a canonical treedecomposition of G of width k or smaller: each bag of T is a potential maximal clique and
the intersection of each pair of adjacent bags of T is a minimal separator. Orient each edge of
T as follows. Let X and Y be adjacent bags in T and let S = X ∩Y . Let C be the outbound full
component associated with the minimal separator S. Then, C intersects exactly one of X and
Y . If C intersects X then we orient the edge between X and Y from Y to X; otherwise from
X to Y . Since T is a tree, the resulting directed tree has a sink X0 . Then, each component
C associated with X0 is inbound and hence outlet(X0 ) = 0.
/ We show that each such C is
moreover feasible. Indeed, the required tree-decomposition of GhCi may be obtained from
T by taking intersection of every bag with N[C]: the resulting tree is a tree-decomposition
of G[N(C)] and contains the bag X0 ∩ N[C] ⊇ N(C). The width of the tree-decomposition is
not greater than that of T and hence is k or smaller. Therefore, I-block (N(C),C) for each
component C associated with X0 is feasible and hence the potential maximal clique X0 is
feasible.
⊔
⊓
Lemma 6 Let C be a connected set of G such that N(C) is a minimal separator. Let Ω be a
potential maximal clique of GhCi. Then, Ω is a potential maximal clique of G.
Proof For each component D associated with N(C), let HD be a minimal chordal completion
of GhCi. In particular, choose HC so that Ω is a clique in HC . Let H be the union of these
S
graphs: V (H) = V (G) and E(H) = D∈C (N(C)) E(HD ). It is clear that H is chordal. Let H ′
be a minimal chordal completion of G contained in H. It is well-known that every minimal
separator is a clique in every chordal completion and hence N(C) is a clique in H ′ . Therefore,
the minimality of HD for each D implies that H ′ = H. As Ω is a clique in HC , it is a clique
in H and hence is a potential maximal clique of G.
⊔
⊓
The following is our oriented version of the recurrence in the BT algorithm described in
the beginning of this section.
Lemma 7 An I-block (N(C),C) is feasible if and only if there is some feasible potential
S
maximal clique Ω with outlet(Ω ) = N(C) and D∈support(Ω ) D = C.
Proof Suppose first that there is a feasible potential maximal clique Ω as in the lemma.
For each component D ∈ support(Ω ), let TD be a tree-decomposition of GhDi of width k
or smaller and XD be a bag in TD containing N(D). Combine these tree-decompositions TD ,
D ∈ support(Ω ), into a tree T by adding bag Ω and let it be adjacent to XD for each D ∈
support(Ω ). We confirm that T is a tree-decomposition of G[N[C]]. Every vertex v ∈ N[C]
appears in some bag of T since C is the union of D for all D ∈ support(Ω ) and the bag Ω
contains N(C). Every edge of G[N[C]] appears in some bag of T for the same reason. The
third condition for T being a tree decomposition is also satisfied, since, if a vertex v appears
in N[D] for two or more members D in support(Ω ), then v appears in XD for each such D
and in Ω . Therefore, T is a tree decomposition of G[N[C]] of width k or smaller and hence
the bag Ω in T contains N(C), T attests the feasibility of the I-block (N(C),C).
For the converse, suppose that I-block (N(C),C) is feasible. Let T be a canonical treedecomposition of GhCi of width k or smaller. Orient the edges of T as in the proof of
Lemma 5: orient the edge from X to Y if and only if Y intersects the outbound full component
associated with X ∩Y . We need to stress here that the notion of outbound components used
in this orientation is with respect to the entire graph G and not with respect to GhCi, the
8
Hisao Tamaki
graph of which T is a tree-decomposition. As N(C) is a clique in GhCi, T contains a bag
that contains N(C). In the subtree of T induced by those bags containing N(C), let X0 be
a sink with respect to the above orientation. As T is canonical, X0 is a potential maximal
clique of GhCi and hence of G by Lemma 6. We show below that X0 is feasible.
Let A be the outbound full component associated with N(C). As N(C) ⊆ X0 and A ∩
N[C] = 0,
/ A is a component associated with X0 . We claim that N(C) = outlet(X0 ). Suppose
otherwise that there is some outbound component A′ associated with X0 such that N(C) is
a proper subset of N(A′ ). Then, as A′ is not confined to N(C), C = crib(N(C), X0) contains
A′ . Therefore, there is some bag X adjacent to X0 in T such that X ∩ A′ 6= 0.
/ Since N(C)
is a minimal separator that separates A from A′ , X must contain N(C). But, since A′ is an
outbound component associated with X0 , the edge between X0 and X is oriented from X0 to
X. This contradicts the choice of X0 and we conclude that N(C) = outlet(X0 ).
It remains to verify that each D ∈ support(X0 ) is feasible. This is true since the tree of
bags obtained from T by intersecting each bag with N[D] is a tree-decomposition of GhDi
required for the feasibility of D.
⊔
⊓
Lemma 8 Let K be a cliquish vertex set, C a non-empty subset of support(K), and S =
S
C∈C N(C). If S is a proper subset of K then crib(S, K) is outbound.
Proof Let K, C and S be as in the lemma. Since K is cliquish, crib(S, K) is a full component
associated with S that contains K \ S, by Lemma 3. To show that it is outbound, it suffices
to show that no other full component associated with S is outbound. Let A be an arbitrary
full component associated with S that is distinct from crib(S, K). As A does not intersect
K, it is a component associated with K. Let C be an arbitrary member of C . Then, C is
confined to S by the definition of S. On the other hand C is not confined to outlet(K) since
C ∈ support(K). Therefore, S is not a subset of outlet(K). A cannot be outbound, since it
would imply that S = N(A) ⊆ outlet(K). Therefore, A is inbound and, since this holds for
every full component associated with S other than crib(S, K), crib(S, K) is outbound.
⊔
⊓
The following lemma is crucial for our PID result: the algorithm described in the next
section generates all buildable potential maximal cliques and we need to guarantee all feasible maximal cliques to be among them.
Lemma 9 Let Ω be a feasible potential maximal clique. Then, Ω is buildable.
Proof Let S = C∈support(Ω ) N(C).
Suppose first that S ∪ outlet(Ω ) 6= Ω and let v be an arbitrary member of Ω \ (S ∪
outlet(Ω )). Since Ω is cliquish and v is not in N(C) for any component C associated with
Ω , v is adjacent to every other vertex in Ω . Therefore, Ω ⊆ N[v]. Let C be an arbitrary component associated with Ω . If C is confined to outlet(Ω ) then v 6∈ N(C) since v 6∈ outlet(Ω ).
Otherwise, C ∈ support(Ω ) and hence v 6∈ N(C) as v 6∈ S. Therefore, N(v) \ Ω is empty and
hence we have Ω = N[v]. Thus, Ω is buildable, the first case of buildability.
Suppose next that S ∪ outlet(Ω ) = Ω . We have two cases to consider: S = Ω and S 6= Ω .
Consider the case where S = Ω . Let C0 be an arbitrary minimal subset of support(Ω )
S
such that C∈C0 N(C) = Ω . Since Ω does not have a full component associated with it, C0
has at least two members. Let C0 be an arbitrary member of C0 and let C1 = C0 \ {C0 }.
S
From the minimality of C0 , S1 = C∈C1 N(C) is a proper subset of Ω . By Lemmas 3 and 8,
A1 = crib(S1 , Ω ) is a full component associated with S1 and is outbound. Therefore, (S1 , A1 )
is an O-block and is feasible since every member of C1 ⊆ support(Ω ) is feasible as potential
maximal clique Ω is feasible. Thus, the second case in the definition of feasible potential
maximal cliques applies.
S
Positive-instance driven dynamic programming for treewidth ⋆
9
Finally, suppose that S 6= Ω . Let A = crib(S, Ω ). Then, A is a full component associS
ated with S and is outbound, by Lemmas 3 and 8. Since S = C∈support(Ω ) N(C) and Ω is
feasible, the O-block (S, A) is feasible. Let x be an arbitrary vertex in Ω \ S. Since we are
assuming that S ∪ outlet(Ω ) = Ω we have x ∈ outlet(Ω ) \ S. Let v be an arbitrary vertex
in Ω \ outlet(Ω ). Observe that there is no component C associated with Ω such that N(C)
contains both x and v: x 6∈ N(C) for every C ∈ support(Ω ) and v 6∈ N(C) for every C that is
confined to outlet(Ω ). Since Ω is cliquish, it follows that x and v are adjacent to each other.
Therefore, we have Ω \ S ⊆ N(v). Moreover, A contains Ω \ S by Lemma 3. Finally, A \ Ω
is disjoint from N(v), since every component D associated with Ω such that v ∈ N(D) is not
confined to outlet(Ω ) and hence contained in C. Therefore, we have Ω = S ∪ (N(v) ∩ A),
and the third case in the definition of buildable potential maximal cliques applies.
⊔
⊓
4 Algorithm
Given graph G and positive integer k, our algorithm generates all I-blocks, O-blocks, and
potential maximal cliques that are feasible. In the algorithm description below, the following
variables, with suffixes, are used: I for listing feasible I-blocks, O for feasible O-blocks,
P for buildable potential maximal cliques, and S for feasible potential maximal cliques.
We note that each member of I and O is actually the component part of an I- or O-block.
Algorithm PID-BT
Input: Graph G and positive integer k
Output: “YES” if tw(G) ≤ k; “NO” otherwise
Procedure:
Let I0 = 0/ and O0 = 0.
/
Initialize P0 and S0 to 0.
/
Set j = 0.
For each v ∈ V (G), if N[v] is a potential maximal clique with |N[v]| ≤ k + 1 then add
N[v] to P0 and if, moreover, support(N[v]) = 0/ then do the following.
(a) Add N[v] to S0 .
(b) If outlet(N[v]) 6= 0/ then let C = crib(outlet(N[v]), N[v]) and, provided that C 6= Ch
for 1 ≤ h ≤ j, increment j and let C j = C.
5. Set i = 0.
6. Repeat the following and stop repetition when j is not incremented during the iteration
step.
(a) While i < j, do the following.
i. Increment i and let Ii be Ii−1 ∪ {Ci }.
ii. Initialize Oi to Oi−1 , Pi to Pi−1 , and Si to Si−1 .
iii. For each B ∈ Oi−1 such that Ci ⊆ B and |N(Ci ) ∪ N(B)| ≤ k + 1, let K = N(Ci ) ∪
N(B) and do the following.
A. If K is a potential maximal clique, then add K to Pi .
B. If |K| ≤ k and there is a full component A associated with K (which is
unique), then add A to Oi .
iv. Let A be the full component associated with N(Ci ) and add A to Oi .
v. For each A ∈ Oi \Oi−1 and v ∈ N(A), let K = N(A)∪(n(v)∩A) and if |K| ≤ k +1
and K is a potential maximal clique then add K to Pi .
1.
2.
3.
4.
10
Hisao Tamaki
vi. For each K ∈ Pi \ Si−1 , if support(K) ⊆ Ii then add K to Si and do the
following: if outlet(K) 6= 0/ then let C = crib(outlet(K), K) and, provided that
C 6= Ch for 1 ≤ h ≤ j, increment j and let C j = C.
7. If there is some K ∈ S j such that outlet(K) = 0,
/ then answer “YES”; otherwise, answer
“NO”.
Theorem 1 Algorithm PID-BT, given G and k, answers “YES” if and only if tw(G) ≤ k.
Proof We show that SJ computed by the algorithm, where J denotes the final value of j,
is exactly the set of feasible potential maximal cliques for the given G and k. The theorem
then follows by Lemma 5.
In the following proof, Oi , Pi , and Si for each i stand for the final values of these
program variables.
We first show by induction on i that the following conditions are satisfied.
1.
2.
3.
4.
5.
For every 1 ≤ h ≤ i, (N(C j ),C j ) is a feasible I-block.
Ii = {Ch | 1 ≤ h ≤ i}.
For every A ∈ Oi , (N(A), A) is a feasible O-block.
Every K ∈ Pi is a buildable potential maximal clique.
Every K ∈ Si is a feasible potential maximal clique.
Consider the base case i = 0. Condition 1 vacantly holds. Conditions 2 and 3 also hold
since I0 = O0 = 0.
/ Condition 4 holds: N[v] is confirmed to be a potential maximal clique
before it is added to P0 and is buildable by the definition of buildability (case 1). Condition
5 holds since support(N[v]) = 0/ implies that the potential maximal clique N[v] is feasible.
Suppose i > 0 and that the above conditions are satisfied for smaller values of i.
1. When Ci is defined, there is some i′ < i and K ∈ Si′ such that outlet(K) 6= 0/ and Ci =
crib(outlet(K), K). By the induction hypothesis, K is a feasible potential maximal clique
and hence, by Lemma 7, (N(Ci),Ci ) is a feasible I-block.
2. As Ii−1 = {Ch | 1 ≤ h ≤ i − 1} and Ii = Ii−1 ∪ {Ci }, Ii = {Ch | 1 ≤ h ≤ i} holds.
3. Let A ∈ Oi \ Oi−1 . Then there is some B ∈ Oi−1 such that A is outbound, |N(A)| ≤ k, and
N(A) = N(Ci ) ∪ N(B). From the first two conditions, (N(A), A) is an O-block. By the
S
induction hypothesis, (N(B), B) is a feasible O-block and hence N(B) = D∈C N(D) for
some set C of feasible inbound components. As Ci is feasible by 1 above and N(A) =
S
D∈C ∪{Ci } N(D), O-block (N(A), A) is feasible.
4. Let K ∈ Pi \ Pi−1 . Then, K is added to Pi either at step 6-(a)-iii-A or at step 6-(a)-v.
Consider the first case, Then, K = N(B) ∪ N(Ci) where (N(B), B) is a feasible O-block
S
and hence N(B) = D∈C N(D) for some set C of feasible inbound components. As Ci
is feasible, K satisfies all the conditions in the second case of the definition of buildable
potential maximal cliques. Consider next the second case, K is obtained at step 6-(a)-v.
Then, K = N(A) ∪ (n(v) ∩ A), where (N(A), A) is a feasible O-block, and the third case
in the definition of buildable potential maximal cliques applies.
5. Let K ∈ Si \ Si−1 . Then, K ∈ Pi and is a buildable potential maximal clique by 4
above. The confirmed condition support(K) ⊆ Ii ensures that K is feasible, since every
member of Ii is feasible by 1 and 2 above.
We conclude that every member of SJ is a feasible potential maximal clique.
In showing the converse, the following observation is crucial. Let (N(A), A) be a feasible
S
O-block such that N(A) = C∈C N(C) for some set C of feasible components and suppose
C ⊆ Ii . Then, A ∈ Oi . The proof is a straightforward induction on i.
The proof of the converse consists in showing the following by induction on m.
Positive-instance driven dynamic programming for treewidth ⋆
11
1. For each feasible I-block (N(C),C), with |C| = m, there is some i such that C = Ci .
2. For each feasible O-block (N(A), A) with |A| = |V (G)| − m, there is some i such that
A ∈ Oi .
S
3. For each buildable potential maximal clique Ω such that | C∈support(Ω ) C| = m, there is
some i such that Ω ∈ Pi .
S
4. For each feasible potential maximal clique Ω such that | C∈support(Ω ) C| = m, there is
some i such that Ω ∈ Si .
The base case m = 0 is vacantly true. Suppose m > 0 and the statements hold for smaller
values of m.
1. Let (N(C),C) be a feasible I-block with |C| = m. Then, by Lemma 7, there is some
feasible potential maximal clique Ω such that N(C) = outlet(Ω ) and C = crib(N(C), Ω ).
S
We have | C∈support(Ω ) C| < m, since this union is a subset of C \ (Ω \ N(C)). Therefore,
by the induction hypothesis, there is some i such that Ω ∈ Si . Therefore, C is constructed
as C j either at step 4-(b) or at step 6-(a)-vi.
2. Let (N(A), A) be a feasible O-block with |A| = |V (G)| − m. Let C be a set of feasible
S
components such that N(A) = C∈C N(C) and let C be an arbitrary member of C . As C,
A, and N(C) are pairwise disjoint, we have |C| < m. Therefore, there is some iC such that
CiC = C. Set i = max{iC | C ∈ C }. Then, C ⊆ Ii and hence A ∈ Oi , by the observation
above.
S
3. Let Ω be a buildable potential maximal clique with | C∈support(Ω ) C| = m. In the first
case of the definition of buildability, Ω is added to P0 at step 4. In the second case,
S
we have Ω = C∈C N(C) for some C ⊆ support(Ω ) such that every member of C is
feasible. Choose C to be minimal subject to these conditions. Let C be an arbitrary
member of C . As |C| ≤ m, by the induction hypothesis and 1 above, there is some iC
such that C ⊆ IiC . Choose C ∈ C so that iC is the largest and let the chosen be D. Let
S
C ′ = C \ {D} and let S = C∈C ′ N(C). By the minimality of C , S is a proper subset of
Ω . Therefore, crib(S, Ω ) is a full component associated with S and there is an outbound
full component A associated with S. As all members of C ′ is feasible and |S| ≤ k, (S, A)
is a feasible O-block. By the choice of D, we have C ′ ⊆ IiD −1 and hence A ∈ OiD −1 by
the observation above. At step 6-(a)-iii-A in the iteration for i = iD , Ω is put into PiD .
S
4. Let Ω be a feasible potential maximal clique with | C∈support(Ω ) C| = m. Then, by
3 above, there is some i1 such that Ω ∈ Pi1 . Furthermore, as every member C of
support(Ω ) is feasible and |C| ≤ m, there is some i2 such that support(Ω ) ⊆ Ii2 , by
1 above. At step 7 in the iteration for i = max{i1 , i2 }, Ω is put into Si .
We conclude that every feasible potential maximal clique is in SJ . This completes the proof.
⊔
⊓
5 Running time analysis
The running time of our algorithm is stated in terms of the the number of positive subproblem instances. Given G and k > 0, let IGk denote the set of feasible I-blocks and OGk the set
of feasible O-blocks.
Observation 1 Given G and k > 0, algorithm PID-BT runs in O∗ (|IGk | · |OGk |) time.
Proof The number of iteration in step 6, where i is incremented each time, is |IGk |. In each
iteration step, every computation step may be charged to each element of Oi−1 and the total
12
Hisao Tamaki
number of steps charged to a single element of Oi−1 is nO(1) . Since |Oi−1 | ≤ |OGk |, we have
the claimed time bound.
⊔
⊓
The bound in this observation is incomparable to the previous bounds on non-PID versions of the BT algorithm, which run in O∗ (|ΠG |) time when ΠG , the set of potential
maximal cliques in G, is given. In [12], in addition to a combinatorial bound of |ΠG | =
O(1.7549n ), it was shown that ΠG can be computed in O∗ (ΠG ) time.
It should be emphasized, however, that it is not known whether the decision problem
version of the treewidth problem with given k can be solved in O∗ (|ΠGk+1 |) time, where ΠGk
is the set of potential maximal cliques of cardinality at most k in G. The bottleneck here is
the time to list all members of ΠGk+1 . Although a nontrivial upper bound on |ΠGk+1 | in terms
of n and k, together with a running time bound based on it, is given in [12], a huge gap
between the actual value |ΠGk+1 | and the upper bound is observed in practice, as shown in
the next section. This is the gap that makes the bound in Obseravation 1 interesting.
6 Experimental analysis
To study the strength of the running time bound of Observation 1 from a practical view point,
we have performed some experiments, in which we count the number of combinatorial objects involved in the treewidth computation. We first compare the actual number of relevant
potential maximal cliques (that is, of cardinality at most k + 1 where k is the treewidth) with
n
the theoretical uppser bounds on that number: the naive bound of k+1
and an assymptot
⌈(2n+k+4)/2⌉
)
given
in
[12].
Table 1 shows the
+
ically stronger bound of n( ⌈(2n+k+7)/3⌉
k+1
k+2
results on some random instances, where the number of vertices n is 20, 30, 40, or 50, the
number of edges m is 2n, 3n, 4n or 5n, and the graph for each pair (n, m) is chosen uniformly
at random from the set of all graphs with n vertices and m edges. Huge gaps between the
actual number and the upper bounds are apparent.
n = |V |
20
20
20
20
30
30
30
30
40
40
40
40
50
50
50
50
|E|
40
60
80
100
60
90
120
150
80
120
160
200
100
150
200
250
k = tw
6
8
11
11
7
11
14
16
8
14
18
20
10
16
20
24
PMCs (≤ k + 1)
115
96
121
37
559
682
1137
768
5341
10372
17360
6820
6029
48068
36388
47729
n
k+1
n(
⌈(2n+k+7)/3⌉
+ ⌈(n+k+4)/2⌉
)
k+2
k+1
77520
167960
125970
125970
5852925
86493225
155117520
119759850
273438880
40225345056
131282408400
131282408400
37353738800
9847379391150
67327446062800
126410606437752
Table 1 The numbers of relevant potetntial maximal cliques and their upper bounds
1003860
2076360
1921680
1921680
67393950
352580340
430361970
426140550
2705471600
91260807600
135562547400
157012867200
201991095800
10332510412500
53246262826500
52230760068000
Positive-instance driven dynamic programming for treewidth ⋆
13
Since the running time bound in Observation 1 involves the quantity |OGk | which is not
theoretically upper-bounded by a function of |ΠGk+1 |, the gaps observed in Table 1 alone
may not be sufficient to support the importance of this running time bound. To address this
issue, we have counted more combinatorial objects involved in our PID computation on the
same graph instances: in addition to relevant potential maximal cliques counted above, all
potential maximal cliques, relevant minimal separators, all minimal separators, feasible Iblocks, feasible O-blocks and feasible potential maximal cliqeus. Here, the input k to the
decision problem is set to the treewidth of the graph.
Table 2 shows the result. We see that the number of feasible O-blocks is smaller than
the number of relevant potential mmaximal cliques, as far as these instances are concerend.
This, together with what we have observed in Table 1, provides an evidence that the running
time bound of Observation 1 is more relevant from a practical point of view than the running
time bounds of known theoretical algorithms.
We also see that the number of all potential maximal cliques grows much faster than the
number of relevant potential maximal cliques. This shows the advantage of our algorithm
which avoids generating all potential maximal cliques.
To summarize, our PID algorithm has advantages over the standard BT algorithms because the running time upper bounds of those algorithms are either in terms of a combinatorial upper bound on the number of relevant potential maximal cliques or in terms of
the actual number of all potential maximal cliques: our experiments reveal huge gaps between the actual number of relevant potential maximal cliques and both of these quantities.
Note that, if there is an efficient method of generating relevant potential maximal cliques, a
non-PID version of the BT algorithm might outperform our PID version.
|V |
20
20
20
20
30
30
30
30
40
40
40
40
50
50
50
50
|E|
40
60
80
100
60
90
120
150
80
120
160
200
100
150
200
250
tw
6
8
11
11
7
11
14
16
8
14
18
20
10
16
20
24
minimal separators
all
≤ tw
98
51
191
48
185
122
107
25
535
185
2983
247
2713
376
1913
281
14842
1070
164773
2356
134485
3952
52182
1790
96499
1361
1792713
9152
2130811
7878
1452449
10571
PMCs
all
≤ tw + 1
376
115
796
96
698
376
354
37
3122
559
20154
682
16736
1137
10535
768
178661
5341
1740644
10372
1251656
17360
423691
6820
1123621
6029
>2000000
48068
>2000000
36388
>2000000
47729
feasible objects
I-blocks
O-blocks
19
26
46
108
121
158
24
32
114
170
228
708
352
804
240
498
840
2965
2080
8637
3289
10023
1502
4749
779
2171
8099
36881
6956
28247
8949
30834
PMCs
37
93
370
36
334
618
1055
647
4154
8577
13646
5347
2914
39803
29842
37115
Table 2 The numbers of principal objects in treewidth computation
7 Implementation
In this section, we sketch two important ingredients of our implementation. Although both
are crucial in obtaining the result reported in Section 8, our work on this part is preliminary
and improvements are the subject of future research.
14
Hisao Tamaki
7.1 Data structures
The crucial elementary operation in our algorithm is the following. We have a set O of feasible O-blocks obtained so far and, given a new feasible I-block (N(C),C), need to find all
members (N(A), A) of O such that C ⊆ A and |N(C) ∪ N(A)| ≤ k + 1. As the experimental analysis in the previous section shows, there is only a few such A on average for the
tested instances even though O is usually huge. To support an efficient query processing, we
introduce an abstract data structure we call a block sieve.
Let G be a graph and k a positive integer. A block sieve for graph G and width k is a data
structure storing vertex sets of V (G) which supports the following operations.
store(U) : store vertex set U in in the block sieve.
supersets(U) : return the list of entries W stored in the block sieve such that U ⊆ W and
|N(U) ∪ N(W)| ≤ k + 1.
Data structures for superset query have been studied [20]. The second condition above on
the retrieved sets, however, appears to make this data structure new. For each U ⊆ V (G), we
define the margin of U to be k + 1 − |N(U)|. Our implementation of block sieves described
below exploits an upper bound on the margins of vertex sets stored in the sieve.
We first describe how such block sieves with upper bounds on margins are used in our
algorithm. Let O be the current set of O-blocks. We use t block sieves B1 , . . . , Bt , each
Bi having a predetermined upper bound mi on the margins of the sets stored. We have
0 < m1 < m2 < . . . < mt = k. We set m0 = 0 for notational ease below. In our implementation,
we choose roughly t = log2 k and mi = 2i for 0 < i < t. For each (N(A), A) in O, A is stored
in Bi such that the margin k + 1 − |N(A)| is mi or smaller but larger than mi−1 . When we are
given an I-block (N(C),C) and are to list relevant blocks in O, we query all of the t blocks
with the operations supersets(C). These queries as a whole return the list of all vertex sets A
such that (N(A), A) ∈ O, C ⊆ A, and |(N(A) ∪ N(C))| ≤ k + 1.
We implement a block sieve by a trie T . The upper bound m on margin is not used
in the construction of the sieve; it is used in the query time. In the following, we assume
V (G) = {1, . . . , n} and, by an interval [i, j], 1 ≤ i ≤ j ≤ n, we mean the set {v : i ≤ v ≤ j}
of vertices. Each non-leaf node p of T is labelled with a non-empty interval [s p , f p ], such
that sr = 0 for the root r, s p = fq + 1 if p is a child of q, and f p = n if p is a parent of a leaf.
Each edge (p, q) which connects node p and a child q of p, is labelled with a subset S(p,q)
of the interval [s p , f p ]. Thus, for each node p, the union of the labels of the edges along the
path from the root to p is a subset of the interval [1, s p − 1], or [1, n] when p is a leaf, which
we denote by S p . The choice of interval [s p , f p ] for each node p is heuristic. It is chosen so
that the number of descendants of p is not too large or too small. In our implementation, the
interval size is adaptively chosen from 8, 16, 32, and 64.
Each leaf q of trie T represents a single set stored at this leaf, namely Sq as defined
above. We denote by S(T ) the set of all sets stored in T . Then, for each node p of T , the
set of sets stored under p is {U | U ∩ [1, p] = S p }.
We now describe how a query is processed against this data structure. Suppose query
U is given. The goal is to visit all leaves q such that U ⊆ Sq and |N(U) ∪ N(Sq)| ≤ k + 1.
This is done by a depth-first traversal of the trie T . When we visit node p, we have the
invariant that U ∩ [1, f p ] ⊆ S p , since otherwise no leaf in the subtree rooted at p stores a
superset of U. Therefore, we descend from p to a child p′ of p only if this invariant is
maintained. Moreover, we keep track of the quantity i(p,U) = |N(U) ∩ S p | in order to make
further pruning of search possible. For each leaf q below p such that U ⊆ Sq , we have
i(q,U) ≥ i(p,U). Combining this with eauality |N(U) \ N(Sq )| = |N(U) ∩ Sq | = i(q,U), we
Positive-instance driven dynamic programming for treewidth ⋆
15
have |N(U) ∪ N(Sq)| ≥ |N(Sq )| + i(p,U). Since we know an upper bound m on the margin
k + 1 − |N(Sq )| of Sq , or lower bound k + 1 − m on |N(Sq )|, we may prune the search under
node p if i(p,U) > m, since this inequality implies |N(U) ∪ N(Sq )| > k + 1 for every leaf q
under p. When we reach a leaf q, we test if |N(U) ∪ N(Sq)| ≤ k + 1 indeed holds.
7.2 Safe separators
The notion of safe separators for treewidth was introduced by Bodlaender and Koster [5]: a
separator S of G is safe if completing S into a clique does not change the treewidth of G. If
we find a safe separator S then the problem of deciding tree width of G reduces to that of
deciding the treewidth of GhCi for each component C associated with S. Preprocessing G
into such independent subproblems is highly desirable whenever possible.
The above authors observed that a powerful sufficient condition for safeness can be
formulated based on graph minors. A labelled minor of G is a graph obtained from G by
zero or more applications of the following operations. (1) Edge contraction: choose an edge
{u, v}, replace u and v by a single new vertex and let all neighbors of u and v be adjacent
to this new vertex; name the new vertex as either u or v. (2) Vertex deletion: delete a vertex
together with all incident edges. (3) Edge deletion.
Lemma 10 ([5]) A separator S of G is safe if, for every component C associated with S,
G[V (G) \C] contains clique S as a labelled minor.
Call a separator minor-safe if it satisfies the sufficient condition for safeness stated in this
lemma. Bodlaender and Koster [5] showed that if S is a minimal separator and is an almost
clique (deleting some single vertex makes it a clique) then S is minor-safe and moreover that
the set of all almost clique minimal separators can be found in O(n2 m) time, where n is the
number of vertices and m is the number of edges.
We aim at capturing as many minor-safe separators as possible, at the expense of theoretical running time bounds on the algorithm for finding them. Thus, in our approach, both the
algorithm for generating candidate separators and the algorithm for deciding minor-safeness
are heuristic. For candidate generation, we use greedy heuristic for treewidth such as minfill and min-degree: the separators in the resulting tree-decomposition are all candidates for
safe separators.
When we apply our heuristic decision algorithm for minor-safeness to candidate separator S, one of the following occurs.
1. The algorithm answers “YES”. In this case, a required labelled clique minor has been
found for every component associated S and hence S is minor-safe.
2. The algorithm answers “DON’T KNOW”. In this case, the algorithm has failed to find
a labelled clique minor for at least one component, and hence it is not known if S is
minor-safe or not.
3. The algorithm aborts, after reaching the prescribed number of execution steps.
Our heuristic decision algorithm works in two phases. Let S be a separator, C a component associated with S, and R = V (G) \ (S ∪ C). In the first phase, we contract edges in
R and obtain a graph B on vertex set S ∪ R′ , where each vertex of R′ is a contraction of
some vertex set of R and B has no edge between vertices in R′ . For each pair u, v of distinct
vertices in S, let N(u, v) denote the common neighbors of u and v in graph B. The contractions are performed with the goal of making |N(u, v) ∩ R′ | large for each missing edge {u, v}
in S. In the second phase, for each missing edge {u, v}, we choose a common neighbor
16
Hisao Tamaki
w ∈ N(u, v) ∩ R′ and contract either {u, w} or {v, w}. The choice of the next missing edge
to be processed and the choice of the common neighbor are done as follows. Suppose the
contractions in the second phase are done for some missing edges in S. For each missing
edge {u, v} not yet “processed”, let N ′ (u, v) be the set of common neighbors of u and v that
are not yet contracted with any vertex in S. We choose {u, v} with the smallest |N ′ (u, v) ∩ R′ |
to be processed next. Tie-breaking when necessary and the choice of the common neighbor
w in N ′ (u, v) ∩ R′ to be contracted with u or v is done in such a way that the minimum of
|(N ′ (x, y) ∩ R′ ) \ {w}| is maximized over all remaining missing edges {x, y} in S.
The performance of these heuristics strongly depends on the instances. For PACE 2017
public instances, they work quite well. Table 3 shows the preprocessing result on the last 10
of those instances. See Section 8 for the description of those instances and the computational
environment for the experiment. For each instance, the number of safe separators found
and the maximum subproblem size in terms of the number of vertices, after the graph is
decomposed by the safe separators found, are listed. The results show that these instances,
which are deemed the hardest among all the 100 public instances, are quickly decomposed
into manageable subproblems by our preprocessing.
name
ex181
ex183
ex185
ex187
ex189
ex191
ex193
ex195
ex197
ex199
|V |
109
265
237
240
178
492
1391
216
303
310
|E|
732
471
793
453
4517
1608
3012
382
1158
537
tw(G)
18
11
14
10
70
15
10
10
15
9
safe separators found
18
173
142
138
6
184
791
114
176
157
max subproblem
89
76
52
81
161
132
119
84
56
131
time(secs)
0.078
0.031
0.046
0.031
0.062
0.171
3.17
0.015
0.062
0.046
Table 3 Safe separator preprocessing on PACE 2017 instances
On the other hand, these heuristics turned out useless for most of the DIMACS graph
coloring instances: no safe separators are found for those instances. We suspect that this
is not the limitation of the heuristics but is simply because those instances lack minor-safe
separators. We need, however, further study to get a firm conclusion.
8 Performance results
We have tested our implementation on two sets of instances. The first set comes from the
DIMACS graph coloring challenge [14] and has served as a standard benchmark suite for
treewidth in the literature [13, 5, 15, 19, 4, 2]. The other is the set of public instances posed
by the exact treewidth track of PACE 2017 [16].
The computing environment for the experiment is as follows. CPU: Intel Core i7-7700K,
4.20GHz; RAM: 32GB; Operating system: Windows 10, 64bit; Programming language:
Java 1.8; JVM: jre1.8.0 121. The maximum heap space size is 6GB by default and is 24GB
where it is stated so. The implementation is single threaded, except that multiple threads
may be invoked for garbage collection by JVM. The time measured is the CPU time, which
includes the garbage collection time.
Positive-instance driven dynamic programming for treewidth ⋆
17
To determine the treewidth of a given instance we use our decision procedure with k
being incremented one by one, starting from the obvious lower bound, namely the minimum
degree of the graph. Binary search is not used because the cost of overshooting the exact
treewidth can be huge. We do not feel the need of using stronger lower bounds either, since
the cost of executing the decision procedure for k below such lower bounds is usually quite
small.
Table 4 shows the results on DIMACS graph coloring instances. Each row shows the
name of the instance, the number of vertices, the number of edges, the exact treewidth computed by our algorithm, CPU time in seconds, and the previously best known upper and
lower bounds on the treewidth. Rows in bold face show the newly solved instances. For all
but three of them, the previous best upper bound has turned out optimal: only the lower
bound was weaker. In this experiment, however, no knowledge of previous bounds are used
and our algorithm independently determines the exact treewidth.
The results on “queen” instances illustrate how far our algorithm has extended the practical limit of exact treewidth computation. Queen7 7 with 49 vertices is the largest instance
previously solved, while queen10 10 with 100 vertices is now solved. Also note that all previously solved instances are fairly easy for our algorithm: all of them are solved within 10
seconds per instance and many of them within a second.
18
Hisao Tamaki
name
|V |
|E|
tw
time(secs)
prev UB
prev LB
anna
138
493
12
0.078
12
12
david
87
406
13
0.031
13
13
DSJC125.5
125
3891
108
459
108
56
DSJC125.9
125
6961
119
0.062
119
119
DSJC250.9
250
27897
243
0.44
243
212
DSJC500.9
500
112437
492
14
492
433
DSJR500.5
500
58862
246
546
DSJR500.1c
500
121275
485
2.12
485
485
fpsol2.i.1
496
11654
66
3.30
66
66
fpsol2.i.2
451
8691
31
5.66
31
31
fpsol2.i.3
425
8688
31
5.68
31
31
games120†
120
638
32
94738
32
24
homer†
561
1628
30
2765
31
26
huck
74
301
10
0.012
10
10
inithx.i.1
864
18707
56
8.10
56
56
inithx.i.2
645
13979
31
8.14
31
31
inithx.i.3
621
13969
31
10
31
31
jean
80
254
9
0.031
9
9
miles250
128
387
9
0.000
9
9
miles500
128
1170
22
0.11
22
22
miles750
128
2113
36
0.23
36
35
miles1000
128
3216
49
0.33
49
49
miles1500
128
5198
77
0.45
77
77
mulsol.i.1
197
3925
50
1.41
50
50
mulsol.i.2
188
3885
32
1.77
32
32
mulsol.i.3
184
3916
32
1.80
32
32
mulsol.i.4
185
3946
32
1.78
32
32
mulsol.i.5
186
3973
31
1.80
31
31
myciel2
5
5
2
0.000
2
2
myciel3
11
20
5
0.000
5
5
myciel4
23
71
10
0.015
10
10
myciel5
47
236
19
0.33
19
19
myciel6
95
755
35
419
35
29
queen5 5
25
160
18
0.000
18
18
queen6 6
36
290
25
0.031
25
25
queen7 7
49
476
35
0.19
35
35
queen8 8
64
728
45
4.16
45
25
queen9 9
81
1056
58
274
58
35
queen8 12
96
1368
65
649
39
queen10 10
100
1470
72
20934
72
39
zeroin.i.1
211
4100
50
1.09
50
50
zeroin.i.2
211
3541
32
1.64
32
32
zeroin.i.3
206
3540
32
1.55
32
31
Previous upper bounds from [13] and [15]; previous lower bounds from [13] and [6].
† 24GB heap space is used for these instances.
Table 4 Results on the DIMACS graph coloring instances
Table 5 shows the lower bounds obtained by our algorithm on unsolved DIMACS graph
coloring instances. Lower bound entries in bold face are improvements over the previously
known lower bounds. Computation time of the previously best lower bounds ranges from a
few minutes to a week [6]. Detailed comparison of lower bound methods, which requires
the normalization of machine speeds, is not intended here. Rather, the table is meant to show
the potential of our algorithm as a lower bound procedure.
Positive-instance driven dynamic programming for treewidth ⋆
19
For many of the instances the improvements are significant. It can also be seen from
this table that our algorithm performs rather poorly on relatively sparse graphs with a large
number of vertices.
lower bounds computed
previous bounds
name
|V |
|E| 1 sec
1 min 30 min lower
upper
DSJC125.1
125
736
25
30
36
20
60
DSJC250.1
250
3218
45
57
66
43
167
DSJC250.5
250
15668
180
197
211
114
229
DSJC500.1
500
12458
94
115
87
409
DSJC500.5
500
62624
360
388
231
479
DSJC1000.1
1000
49629
172
189
183
896
DSJC1000.5
1000
249826
724
742
469
977
DSJC1000.9
1000
449449
983
987
872
991
le450 5a
450
5714
29
50
59
79
243
le450 5b
450
5734
49
57
246
le450 5c
450
9803
84
100
106
265
le450 5d
450
9757
94
99
265
le450 15a
450
8168
24
40
49†
94
262
le450 15b
450
8169
23
32
47†
258
le450 15c
450
16680
114
132
139
350
le450 15d
450
16750
112
131
353
le450 25a
450
8260
11
23
25†
96
216
le450 25b
450
8263
16
26
30†
219
le450 25c
450
17343
43
89
109
144
320
le450 25d
450
17425
93
112
327
myciel7
191
2360
22
31
35
52
66
queen11 11
121
1980
61
70
77
40
87
queen12 12
144
2596
71
76
84
55
103
queen13 13
169
3328
70
82
91
51
121
queen14 14
196
4186
74
87
98
55
140
queen15 15
225
5180
78
93
104
73
162
queen16 16
256
6320
83
99
110
79
186
school1
385
19095
73
112
125
149
178
school1 nsh
352
14612
78
105
118
132
152
Previous upper bounds from [15]; previous lower bounds from [6].
†
out of memory before time out
Table 5 New lower bounds on the treewidth of unsolved DIMACS graph coloring instances
Table 6 shows the results on PACE 2017 instances. The prefix “ex” in the instance names
means that they are for the exact treewidth track. Odd numbers mean that they are public
instances disclosed prior to the competition for testing and experimenting. Even numbered
instances, not in the list, are secret and to be used in evaluating submissions. The time
allowed to be spent for each instance is 30 minutes. As can be seen from the table, our
algorithm solves all of the public instances with a large margin in time.
20
Hisao Tamaki
name
ex001
ex003
ex005
ex007
ex009
ex011
ex013
ex015
ex017
ex019
ex021
ex023
ex025
ex027
ex029
ex031
ex033
ex035
ex037
ex039
ex041
ex043
ex045
ex047
ex049
ex051
ex053
ex055
ex057
ex059
ex061
ex063
ex065
ex067
ex069
ex071
ex073
ex075
ex077
ex079
ex081
ex083
ex085
ex087
ex089
ex091
ex093
ex095
ex097
ex099
|V |
262
92
377
137
466
465
56
177
330
291
318
690
92
274
238
219
363
247
272
56
205
279
600
1854
117
136
218
197
281
298
158
103
50
235
235
253
712
111
237
314
188
213
229
380
318
193
454
220
286
616
|E|
648
2113
597
451
662
1004
280
669
571
752
572
1355
472
715
411
382
541
804
615
280
341
513
865
21118
332
254
383
813
9075
780
1058
582
175
424
441
434
1085
360
423
4943
638
380
370
5790
576
336
664
555
4079
923
tw
10
44
7
12
7
9
29
15
9
11
9
8
20
11
9
8
7
14
10
32
9
9
7
21
13
10
9
18
117
10
22
34
25
10
9
9
7
8
10
42
6
10
8
47
9
9
7
11
48
7
time (secs)
1.48
8.92
14
0.046
13
0.50
15
0.046
1.11
40
2.80
0.91
1.61
51
1.33
12
50
3.60
3.43
58
0.63
3.34
7.80
140
0.078
0.62
1.98
0.078
0.093
0.47
9.59
4.76
79
2.70
1.43
2.42
15
0.28
2.70
1.64
0.55
3.05
11
46
11
31
27
0.59
2.01
88
Table 6 Results on the PACE 2017 public instances
name
ex101
ex103
ex105
ex107
ex109
ex111
ex113
ex115
ex117
ex119
ex121
ex123
ex125
ex127
ex129
ex131
ex133
ex135
ex137
ex139
ex141
ex143
ex145
ex147
ex149
ex151
ex153
ex155
ex157
ex159
ex161
ex163
ex165
ex167
ex169
ex171
ex173
ex175
ex177
ex179
ex181
ex183
ex185
ex187
ex189
ex191
ex193
ex195
ex197
ex199
|V |
1038
237
1038
166
1212
395
93
963
77
84
204
122
320
228
737
292
522
2822
196
334
226
130
48
101
698
279
772
758
260
582
1046
244
222
509
3706
647
536
227
227
187
109
265
237
240
178
492
1391
216
303
310
|E|
291034
419
291037
396
1794
668
488
419877
181
479
1164
635
8862
527
2826
1386
1296
129474
1098
568
1168
660
96
606
2604
733
11654
11580
467
2772
3906
445
742
969
42236
2175
1011
1000
759
346
732
471
793
453
4517
1608
3012
382
1158
537
tw
540
10
540
12
7
9
14
908
13
23
34
35
70
10
14
18
11
87
19
9
34
35
12
16
12
12
47
47
9
18
12
10
14
10
22
14
10
17
14
10
18
11
14
10
70
15
10
10
15
9
time (secs)
12
3.01
12
1.44
43
4.33
0.046
18
18
16
76
14
8.19
0.20
0.97
0.17
3.94
49
0.34
8.34
117
52
18
0.093
0.75
210
57
103
6.42
2.37
2.84
4.69
0.23
7.96
530
0.77
5.05
113
0.23
14
0.20
8.61
0.33
2.80
3.59
21
3.80
6.11
0.36
23
Positive-instance driven dynamic programming for treewidth ⋆
21
Acknowledgment
The author thanks Hiromu Ohtsuka for his help in implementing the block sieve data structure. He also thanks Yasuaki Kobayashi for helpful discussions and especially for drawing
the author’s attention to the notion of safe separators. This work would have been nonexistent if not motivated by the timely challenges of PACE 2016 and 2017. The author is
deeply indebted to their organizers, especially Holger Dell, for their dedication and excellent
work.
References
1. S. Arnborg, D. G. Corneil, and A. Proskurowski: Complexity of finding embeddings in a k-tree. SIAM
Journal on Algebraic Discrete Methods 8, 277-284, 1987
2. J. Berg and M. Järvisalo: SAT-based approaches to treewidth computation: an evaluation. Proceedings of
the IEEE 26th International Conference on Tools with Artificial Intelligence, 328-335, 2014
3. H. L. Bodlaender: A linear-time algorithm for finding tree-decompositions of small treewidth. SIAM
Journal on Computing 25(6), 1305-1317, 1996
4. H. L. Bodlaender, F. V. Fomin, A. M. C. A. Koster D. Kratsch, and D. M. Thilikos: On exact algorithms
for treewidth. ACM Transactions on Algorithms 9(1), 12, 2012
5. H. L. Bodlaender and A. M. C. A. Koster: Safe separators for treewidth. Discrete Mathematics 306(3),
337-350, 2006
6. H. L. Bodlaender, T. Wolle, and A. M. C. A. Koster: Contraction and Treewidth Lower Bounds. Journal
of Graph Algorithms and Applications 10(1), 5-49, 2006
7. H. L. Bodlaender and A. M. C. A. Koster: Combinatorial Optimization on Graphs of Bounded Treewidth.
The Computer Journal 51(3), 255-269, 2008
8. V. Bouchitté and I. Todinca: Treewidth and minimum fill-in: Grouping the minimal separators. SIAM
Journal on Computing 31(1), 212-232, 2001
9. V. Bouchitté and I. Todinca: Listing all potential maximal cliques of a graph. Theoretical Computer
Science 276, 17-32, 2002
10. H. Dell, T. Husfeldt, B. M. Jansen, P. Kaski, C. Komusiewicz, and F. A. Rosamond: The First Parameterized Algorithms and Computational Experiments Challenge LIPIcs-Leibniz International Proceedings
in Informatics 63, 2017.
11. F. V. Fomin, D. Kratsch, I. Todinca, and Y. Villanger: Exact algorithms for treewidth and minimum
fill-in. SIAM Journal on Computing, 38(3), 1058-1079, 2008
12. F. Fomin and Y. Villanger: Treewidth computation and extremal combinatorics. Combinatorica 32(3),
289-308, 2012
13. V. Gogate and R. Dechter: A complete anytime algorithm for treewidth. Proceedings of the 20th conference on Uncertainty in artificial intelligence, AUAI Press, 2004
14. D. S. Johnson and M. A. Trick (eds.): Cliques, coloring, and satisfiability: second DIMACS implementation challenge. Series in Discrete Mathematics and Theoretical Computer Science, American Mathematical
Society, Vol. 26. American Mathematical Society, 1996
15. N. Musliu: An iterative heuristic algorithm for tree decomposition. Recent Advances in Evolutionary
Computation for Combinatorial Optimization, 133-150, 2008
16. PACE 2017 website: https://pacechallenge.wordpress.com/
17. N. Robertson and P. D. Seymour: Graph minors. II. Algorithmic aspects of tree-width. Journal of
Algorithms 7, 309-322, 1986
18. N. Robertson and P. D. Seymour: Graph minors. XX. Wagner’s conjecture. Journal of Combinatorial
Theory, Series B 92(2), 325-357, 2004
19. M. Samer and H. Veith: Encoding treewidth into SAT. Proceedings of International Conference on
Theory and Applications of Satisfiability Testing, 45-50, 2009
20. I. Savnik: Index data structure for fast subset and superset queries. Proceedings of International Conference on Availability, Reliability, and Security, 134-148, 2013
21. Github repository: https://github.com/TCS-Meiji/PACE2017-TrackA
| 8cs.DS
|
arXiv: arXiv:1510.01188
arXiv:1510.01188v2 [math.ST] 28 Apr 2017
Analytic Posteriors for Pearson’s
Correlation Coefficient
Alexander Ly∗ , Maarten Marsman∗ and Eric-Jan Wagenmakers
∗
University of Amsterdam
Weesperplein 4
1018 XA Amsterdam
The Netherlands
[email protected]
Abstract: Pearson’s correlation is one of the most common measures of
linear dependence. Recently, Bernardo (2015) introduced a flexible class
of priors to study this measure in a Bayesian setting. For this large class
of priors we show that the (marginal) posterior for Pearson’s correlation
coefficient and all of the posterior moments are analytic. Our results are
available in the open-source software package JASP.
MSC 2010 subject classifications: Primary 62H20, 62E15, 62F15.
Keywords and phrases: Bivariate normal distribution, hypergeometric
functions, reference priors.
1. Introduction
Pearson’s product-moment correlation coefficient ρ is a measure of the linear
dependency between two random variables. Its sampled version, commonly denoted by r, has been well-studied by the founders of modern statistics such
as Galton, Pearson, and Fisher. Based on geometrical insights Fisher (1915,
1921) was able to derive the exact sampling distribution of r, and established
that this sampling distribution converges to a normal distribution as the sample size increases. Fisher’s study of the correlation has lead to the discovery of
variance-stabilizing transformations, sufficiency (Fisher, 1920), and, arguably,
the maximum likelihood estimator (Fisher, 1922; Stigler, 2007). Similar efforts were made in Bayesian statistics which focus on inferring the unknown
ρ from the data that were actually observed. This type of analysis requires
the statistician to (i) choose a prior on the parameters, thus, also on ρ, and
to (ii) calculate the posterior. Here we derive analytic posteriors for ρ given
a large class of priors that include the recommendations of Jeffreys (1961),
Lindley (1965), Bayarri (1981), and, more recently, Berger and Sun (2008) and
Berger, Bernardo and Sun (2015). Jeffreys’s work on the correlation coefficient
can also be found in the second edition of his book (Jeffreys, 1961), originally
∗ This work was supported by the starting grant “Bayes or Bust” awarded by the European
Research Council (Grant number: 283876). The authors thank Christian Robert, Fabian Dablander, Tom Koornwinder and an anonymous reviewer for helpful comments that improved
an earlier version of this manuscript.
1
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
2
published in 1948; see Robert, Chopin and Rousseau (2009) for a modern reread of Jeffreys’s work. An earlier attempt at a Bayesian analysis of the correlation coefficient can be found in Jeffreys (1935). Before presenting the results, we
first discuss some notations and recall the likelihood for the problem at hand.
2. Notation and Result
Let (X1 , X2 )′ have a bivariate normal distribution with mean ~µ = (µ1 , µ2 )′ and
covariance matrix
2
σ1
ρσ1 σ2
Σ=
,
ρσ1 σ2
σ22
where σ12 , σ22 are the population variances of X1 and X2 , and where ρ is
ρ=
Cov(X1 , X2 )
E(X1 X2 ) − µ1 µ2
=
.
σ1 σ2
σ1 σ2
(2.1)
Pearson’s correlation coefficient ρ measures the linear association between X1
and X2 . In brief, the model is parametrized by the five unknowns θ = (µ1 , µ2 , σ1 , σ2 , ρ).
Bivariate normal data consisting of n pairs of observations can be sufficiently
summarized as y = (n, x̄1 , x̄2 , s1 , s2 , r), where
Pn
j=1 (x1j − x̄1 )(x2j − x̄2 )
,
r=
ns1 s2
Pn
is the sample correlation coefficient, x̄i = n1 j=1 xij the sample mean and
P
n
s2i = n1 j=1 (xij − x̄i )2 the average sums of squares. The bivariate normal model
implies that the observations y are functionally related to the parameters by the
following likelihood function
p
(x̄1 −µ1 )2
−n
2
n
2 −µ2 )
2)
− 2ρ (x̄1 −µσ11)(x̄
f (y | θ) = 2πσ1 σ2 1 − ρ2
+ (x̄2 −µ
exp − 2(1−ρ
2)
σ2
σ12
σ22
h
s2 2
s1 2
n
1 s2
× exp − 2(1−ρ
.
(2.2)
− 2ρ rs
2)
σ1
σ1 σ2 + σ2
For inference, we use the following class of priors
β
πη (θ) ∝ (1 − ρ2 )α−1 (1 + ρ2 ) 2 σ1γ−1 σ2δ−1 ,
|
{z
} | {z } | {z }
πα,β (ρ)
(2.3)
πγ (σ1 ) πδ (σ2 )
where η denotes the hyperparameters, that is, η = (α, β, γ, δ). This class of
priors is inspired by the one José Bernardo used in his talk on reference priors
for the bivariate normal distribution at the 11th International Workshop on
Objective Bayes Methodology in honor of Susie Bayarri. This class of priors
contains certain recommended priors as special cases.
If we set α = 1, β = γ = δ = 0 in Eq. (2.3), we retrieve the prior that Jeffreys recommended for both estimation and testing (Jeffreys, 1961, pp. 174–179
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
3
and 289–292). This recommendation is not the prior derived from Jeffreys’s
rule based on the Fisher information (e.g., Ly et al., 2016), as discussed in
Berger and Sun (2008). With α = 1, β = γ = δ = 0, thus, a uniform prior on ρ,
Jeffreys showed that the marginal posterior for ρ is proportional to ha (n, r | ρ),
where
ha (n, r | ρ) = (1 − ρ2 )
n−1
2 (1
− ρr)
3−2n
2 ,
represents the ρ dependent part of the likelihood Eq. (2.2) with θ0 = (µ1 , µ2 , σ1 , σ2 )
integrated out. For n large enough, the function ha is a good approximation to
the true reduced likelihood hγ,δ , given below.1
If we set α = β = γ = δ = 0 in Eq. (2.3), we retrieve Lindley’s reference prior
for ρ. Lindley (1965, pp 214–221) established that the posterior of tanh−1 (ρ)
is asymptotically normal with mean tanh−1 (r) and variance n−1 , which relates
the Bayesian method of inference for ρ to that of Fisher. In Lindley’s (1965, p.
216) derivation it is explicitly stated that the likelihood with θ0 integrated out
cannot be expressed in terms of elementary functions. In his analysis, Lindley
approximates the true reduced likelihood hγ,δ with the same ha that Jeffreys
used before. Bayarri (1981) furthermore showed that with the choice γ = δ = 0
the marginalization paradox (Dawid, Stone and Zidek, 1973) is avoided.
In their overview, Berger and Sun (2008) showed that for certain a, b with
α = b/2 − 1, β = 0, γ = a − 2 and δ = b − 1 the priors in Eq. (2.3) correspond
to a subclass of the generalized Wishart distribution. Furthermore, a rightHaar prior (e.g., Sun and Berger, 2007) is retrieved when we set α = β = 0,
γ = −1, δ = 1 in Eq. (2.3). This right-Haar prior then has a posterior that can be
constructed through simulations. That is, by simulating from a standard normal
distribution and two chi-squared distributions (Berger and Sun, 2008, Table 1).
This constructive posterior also corresponds to the fiducial distribution for ρ
(e.g, Fraser, 1961, Hannig, Iyer and Patterson, 2006). Another interesting case
is given by α = 0, β = 1, γ = δ = 0, which corresponds to the one-at-a-time
reference prior for σ1 and σ2 , see also Jeffreys (1961, p. 187).
The analytic posteriors given below follow directly from exact knowledge
of the reduced likelihood hγ,δ (n, r | ρ) rather than its approximation used in
previous work. As we have not encountered this proof in earlier work, full details
are given below.
Theorem 2.1 (The Reduced Likelihood hγ,δ (n, r | ρ)). If |r| < 1, n > γ + 1
and n > δ + 1, then the likelihood f (y | θ) times the prior Eq. (2.3) with θ0 =
(µ1 , µ2 , σ1 , σ2 ) integrated out is a function fγ,δ that factors into
fγ,δ (y | ρ) = pγ,δ (y0 )hγ,δ (n, r | ρ).
(2.4)
The first factor is the marginal likelihood with ρ fixed at zero, which does not
1 We
thank an anonymous reviewer for clarifying how Jeffreys derived this approximation.
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
4
depend on r nor on ρ, that is,
Z Z Z Z
pγ,δ (y0 ) =
f (y | θ0 , ρ = 0)πγ (σ1 )πδ (σ2 )dµ1 dµ2 dσ1 dσ2
=2
−γ−δ−4
2
1+γ−n
1+δ−n
π 1−n
(ns21 ) 2 (ns22 ) 2 Γ
n
n−γ−1
2
Γ
n−δ−1
2
,
(2.5)
where y0 = (n, x̄1 , x̄2 , s1 , s2 ). We refer to the second factor as the reduced likelihood, a function of ρ which is given by a sum of an even and an odd function,
that is, hγ,δ = Aγ,δ + Bγ,δ where
n−γ−δ−1
n−γ−1 n−δ−1 1
2
, 2 ; 2 ; r 2 ρ2 ,
2F 1
2
n−γ−δ−1
n−δ 3
2
− ρ2 )
Wγ,δ (n)2 F 1 n−γ
2 , 2 ; 2 ;
Aγ,δ (n, r | ρ) =(1 − ρ2 )
Bγ,δ (n, r | ρ) =2rρ(1
h
n−δ ih
where Wγ,δ (n) = Γ n−γ
Γ
Γ 2
2
notes Gauss’ hypergeometric function.
n−γ−1
2
i
n−δ−1
Γ
2
(2.6)
r 2 ρ2 ,
(2.7)
and where 2 F1 de⋄
Proof. To derive fγ,δ (y | ρ) we have to perform three integrals: (i) with respect
to π(µ1 , µ2 ) ∝ 1, (ii) πγ (σ1 ) ∝ σ1γ−1 , and (iii) πδ (σ2 ) ∝ σ2δ−1 .
(i) The integral with respect to π(µ1 , µ2 ) ∝ 1 yields
1−n
√
s2
2π 1−ρ2 σ1 σ2
s2
rs1 s2
−n
1
exp 2(1−ρ
+ σ22 , (2.8)
f (y | σ1 , σ2 , ρ) =
2 ) σ 2 − 2ρ σ σ
n
1
2
1
R∞ R∞
where we abbreviated f (y | σ1 , σ2 , ρ) = −∞ −∞ f (y | θ0 , ρ)dµ1 dµ2 . The factor
pγ,δ (y0 ) follows directly by setting ρ to zero in Eq. (2.8) and two independent
gamma integrals with respect to σ1 and σ2 resulting in Eq. (2.5). These gamma
integrals cannot be used when ρ is not zero. For fγ,δ (y | ρ) which is a function
of ρ, we use results from special functions theory.
(ii) For the second integral, we collect only that part of Eq. (2.8) that involves
σ1 into a function g, that is,
Z ∞
Z ∞
ns2
1
1 s2
σ1γ−n exp − 2(1−ρ1 2 ) σ12 + σ2ns
g(y | σ1 )πγ (σ1 )dσ1 =
(1−ρ2 ) rρ σ1 dσ1 .
1
0
0
The assumption n > γ + 1 and the substitution u = σ1−1 allows us to solve this
integral using Lemma A.1, which we distilled from the Bateman manuscript
ns2
ns1 s2
project (Erdélyi et al., 1954), with a = 2(1−ρ1 2 ) , b = − (1−ρ
2 )σ rρ and c = n −
2
γ − 1. This yields
Z ∞
i
n−γ−3
n−γ−1 h
1−ρ2
2
(2.9)
Åγ + B̊γ ,
g(y | σ1 )πγ (σ1 )dσ1 = 2 2
ns2
1
0
where
Åγ =Γ n−γ−1
2
r
B̊γ =
1F 1
n−γ−1
2
;
1
2
;
ns22 (rρ)2 1
2(1−ρ2 ) σ22 ,
2ns22 (rρ)2 −1
n−γ
n−γ
1F 1
(1−ρ2 ) σ2 Γ
2
2
;
3
2
;
ns22 (rρ)2 1
2(1−ρ2 ) σ22 ,
(2.10)
(2.11)
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
5
and where 1 F 1 denotes the confluent hypergeometric function. The functions
Åγ and B̊γ are the even and odd solution of Weber’s differential equation in the
ns2
variable z = (rρ)2 2(1−ρ22 )σ2 respectively.
R 2∞
(iii) With fγ (y | σ2 , ρ) = 0 f (y | σ1 , σ2 , ρ)πγ (σ1 )dσ1 , we see that fγ,δ (y | ρ) follows from integrating σ2 out of the following expression
fγ (y | σ2 , ρ)πδ (σ2 ) = 2
−n−γ−1 1−n
1+γ−n
2
π
2
2
(1
n (ns1 )
− ρ2 )
−γ
2
where
Ăγ =Γ
n−γ−1
2ns2
2
1
Ăγ (y | σ2 , ρ) + B̆γ (y | σ2 , ρ) ,
k(n,r | ρ,σ2 )
z
}|
ns22
1
−
2
2
σ2δ−n e 2(1−ρ ) σ2 1 F 1 n−γ−1
2
B̆γ =( 1−ρ22 ) 2 rρΓ
n−γ
2
σ2δ−n−1 e
|
ns22
1
− 2(1−ρ
2 ) σ2
2
{
; 21 ; (rρ)2 2(1−ρ2 ) σ12 ,
1F 1
ns22
(2.12)
2
n−γ
2
;
3
2
{z
l(n,r | ρ,σ2 )
ns2
; (rρ)2 2(1−ρ2 2 ) σ12 .
2
}
Hence, the last integral with respect to σ2 only involves the functions k and l
ns2
in Eq. (2.12). The assumption n > δ + 1 and the substitution t = 2(1−ρ2 2 ) σ2−2 ,
q
3
ns2
thus, dσ2 = − 12 2(1−ρ2 2 ) t− 2 dt allows us to solve this integral using Eq. 7.621.4
from Gradshteyn and Ryzhik (2007, p. 822) with s = 1, k̃ = (rρ)2 . This yields
Z ∞
n−δ−3
n−δ−1 n−δ−1
1−ρ2
n−γ−1 n−δ−1 1
2
Γ
, 2 ; 2 ; r 2 ρ2 ,
k(n, r | ρ, σ2 )dσ2 =2 2
2
2F 1
2
2
ns2
0
Z ∞
n−δ−2
n−δ n−δ
n−γ n−δ 3
1−ρ2
2 2
2 Γ
l(n, r | ρ, σ2 )dσ2 =2 2
2
2F 1
2
2 , 2 ; 2 ; r ρ .
ns
2
0
After we combine the results we see that fγ,δ (y | ρ) = Ãγ,δ (y | ρ) + B̃γ,δ (y | ρ),
where
n−γ−δ−1
Ãγ,δ (y | ρ)
2
=(1 − ρ2 )
2F 1
pγ,δ (y0 )
n−γ−1 n−δ−1
, 2
2
n−γ−δ−1
B̃γ,δ (y | ρ)
2
Wγ,δ (n)2 F 1
=2rρ(1 − ρ2 )
pγ,δ (y0 )
;
1
2
; r 2 ρ2 ,
n−γ n−δ
2 , 2
Hence, fγ,δ (y | ρ) is of the asserted form. Note that Aγ,δ =
while
B̃γ,δ (y | ρ)
pγ,δ (y0 )
;
3
2
; r 2 ρ2 .
Ãγ,δ (y | ρ)
pγ,δ (y0 )
is even,
is an odd function of ρ.
This main theorem confirms Lindley’s insights; hγ,δ (n, r | ρ) is indeed not
expressible in terms of elementary functions and the prior on ρ is updated by
the data only through its sampled version r and the sample size n. As a result,
the marginal likelihood for data y then factors into pη (y) = pγ,δ (y0 )pγ,δ
α,β (n, r),
R
γ,δ
where pα,β (n, r) = hγ,δ (n, r | ρ)πα,β (ρ)dρ is the normalizing constant of the
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
6
marginal posterior of ρ. More importantly, the fact that the reduced likelihood
is the sum of an even and an odd function allows us to fully characterize the
posterior distribution of ρ for the priors Eq. (2.3) in terms of its moments. These
moments are easily computed, as the prior πα,β (ρ) itself is symmetric around
zero. Furthermore, the prior πα,β (ρ) can be normalized as
β
πα,β (ρ) =
(1 − ρ2 )α−1 (1 + ρ2 ) 2
,
1
B 2 , α 2 F 1 − β2 , 12 ; 21 + α ; −1
(2.13)
where B(u, v) = Γ(u)Γ(v)
Γ(u+v) denotes the beta function. The case with β = 0 is
also known as the (symmetric) stretched beta distribution on (−1, 1) and leads
to Lindley’s reference prior when we ignore the normalization constant, i.e.,
B( 21 , α), and, subsequently, let α → 0.
Corollary 2.1 (Characterization of the Marginal Posteriors of ρ). If n > γ +δ−
2α+1, then the main theorem implies that whenever the marginal likelihood with
all the parameters integrated out factors as pη (y) = pγ,δ (y0 )pγ,δ
α,β (n, r), where
pγ,δ
α,β (n, r) =
Z
1
−1
hγ,δ (n, r | ρ)πα,β (ρ)dρ =
Z
1
−1
Aγ,δ (n, r | ρ)πα,β (ρ)dρ,
(2.14)
defines the normalizing constant of the marginal posterior for ρ. Observe that the
integral involving Bγ,δ is zero, because Bγ,δ is odd on (−1, 1). More generally,
the kth posterior moment of ρ is
R1 k
γ,δ 1
ρ Aγ,δ (n, r | ρ)πα,β (ρ)dρ if k is even,
pα,β (n,r)
−1
(2.15)
E(ρk | n, r) =
R1 k
1
pγ,δ (n,r) ρ Bγ,δ (n, r | ρ)πα,β (ρ)dρ if k is odd.
α,β
−1
These posterior moments define the series
n−δ−1
n−γ−1
∞
P
1
2
2
m
m
C
ak,m r2m
1
α,β
m!
m=0
k
2m
E(ρ | n, r) =
n−γ
n−δ
∞
2Wγ,δ (n) P
2
2
m
m
bk,m r2m+1
C
3
α,β
m=0
2
m
m!
if k is even,
(2.16)
if k is odd,
1
1
where Cα,β = B( 21 , α)2 F 1 ( −β
2 , 2 ; α + 2 ; −1) is the normalization constant of
the prior Eq. (2.13), Wγ,δ (n) is the ratios of gamma functions as defined under
= x(x + 1)(x + 2) . . . (x + m − 1) refers to the
Eq. (2.7) and (x)m = Γ(x+m)
Γ(x)
Pochhammer symbol for rising factorials. The terms ak,m and bk,m are
n−γ−δ−1
−β k+2m+1 k+2m+2α+n−γ−δ
;
; −1 ,
ak,m = B 21 + k+2m
2F 1 2 ,
2 ,α +
2
2
2
−β k+2m+2 k+2m+2α+n−γ−δ+1
bk,m = B 12 + k+2m+1
, α + n−γ−δ−1
;
; −1 .
2F 1 2 ,
2
2
2
2
The series defined in Eq. (2.16) are hypergeometric when β is a non-negative
integer.
⋄
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
7
Proof. The series E(ρk | n, r) result from term-wise integration of the hypergeometric functions in Aγ,δ and Bγ,δ . The assumption n > γ + δ − 2α + 1 and
the substitution x = ρ2 allows us to solve these integrals using Eq. (3.197.8)
in Gradshteyn and Ryzhik (2007, p. 317) with their α̃ = 1, u = 1, λ = β2 ,
and ν = 12 + k+2m
when k is even, while we use ν = 12 + k+2m+1
µ = α+ n−γ−δ−1
2
2
2
when k is odd. A direct application of the ratio test shows that the series converge when |r| < 1.
3. Analytic Posteriors for the Case β = 0
For most of the priors discussed above we have β = 0, which leads to the
following simplification of the posterior.
Corollary 3.1 (Characterization of the Marginal Posteriors of ρ, when β = 0).
If n > γ + δ − 2α + 1 and |r| < 1, then the marginal posterior for ρ is
π(ρ | n, r) =
(1 − ρ2 )
2α+n−γ−δ−3
2
pγ,δ
α (n, r)B
(3.1)
1
2, α
h
× 2 F 1 ( n−γ−1
, n−δ−1
;
2
2
1
2
n−δ
; r2 ρ2 ) + 2rρWγ,δ (n)2 F 1 ( n−γ
2 , 2 ;
3
2
i
; r 2 ρ2 ) ,
where pγ,δ
α (n, r) refers to the normalizing constant of the (marginal) posterior
of ρ, which is given by
n−γ−δ−1
n−γ−1 n−δ−1
1
, 2 ; α + n−γ−δ
; r2 /B 12 , α .
pγ,δ
2F 1
α (n, r) = B 2 , α +
2
2
2
More generally, when β = 0, the kth posterior moment is
k+1 n−γ−1 n−δ−1 1 k+2α+n−γ−δ
B 12 + k2 , α + n−γ−δ−1
, 2 ; 2,
; r2
3F 2
2
2 ,
2
2
,
n−γ−1 n−δ−1 2α+n−γ−δ
B 12 , α + n−γ−δ−1
, 2 ;
; r2
2F 1
2
2
2
when k is even, and
2rWγ,δ (n)
B
1
2
when k is odd.
+
n−γ−δ−1
k+1
k+2 n−γ n−δ
3F 2
2 ,α+
2
2 , 2 , 2
n−γ−δ−1
n−γ−1 n−δ−1
1
, 2
B 2, α +
2F 1
2
2
; 32 , k+2α+n−γ−δ+1
; r2
2
,
; 2α+n−γ−δ
; r2
2
⋄
Proof. The assumption n > γ + δ − 2α + 1 and the substitution x = ρ2 allows
us to use Eq. (7.513.12) in Gradshteyn and Ryzhik (2007, p. 814) with µ =
and ν = 21 + k2 when k is even, while we use ν = 21 + k+1
α + n−γ−δ−1
2
2 when
k is odd. The normalizing constant of the posterior pγ,δ
α (n, r) is a special case
with k = 0.
The marginal posterior for ρ updated from the generalized Wishart prior,
the right-Haar prior and Jeffreys’s recommendation then follow from a direct
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
8
substitution of the values for α, γ and δ as discussed under Eq. (2.3). Lindley’s
reference posterior for ρ is given by
n n 3
n−1 n−1 1
2 2
2 2
n−3
2 F1
2 , 2 ; 2 ; r ρ + 2rρW0,0 (n) 2 F1 2, 2 ; 2 ; r ρ
(1 − ρ2 ) 2 ,
n−1 n−1 n
2
,
;
;
r
F
B 12 , n−1
2 1
2
2
2
2
which follows from Eq. (3.1) by setting γ = δ = 0 and, subsequently, letting
α → 0.
Lastly, for those who wish to sample from the posterior distribution, we
suggest the use of an independence-chain Metropolis algorithm (IMH; Tierney,
1994) using Lindley’s normal approximation of the posterior of tanh−1 (ρ) as the
proposal. This method could be used when Pearson’s correlation is embedded
within a hierarchical model, as the posterior for ρ will then be a full conditional
distribution within a Gibbs sampler. For α = 1, β = γ = δ = 0, n = 10
observations and r = 0.6, the acceptance rate of the IMH algorithm was already
well above 75%, suggesting a fast convergence of the Markov chain. For n larger,
the acceptance rate further increases. The R code for the independence-chain
Metropolis algorithm can be found on the first author’s home page. In addition,
this analysis is also implemented in the open-source software package JASP (url:
https://jasp-stats.org/).
Appendix A: A Lemma distilled from the Bateman Project
Lemma A.1. For a, c > 0 the following equality holds
Z ∞
i
ch
uc−1 exp − au2 − bu du = 2−1 a− 2 Å(a, b, c) + B̊(a, b, c) ,
(A.1)
0
that is, the integral is solved by the functions
Å(a, b, c) =Γ
B̊(a, b, c) = −
c
b2
c 1
2 1 F 1 2 ; 2 ; 4a ,
√b Γ c+1 1 F 1 c+1
2
2
a
(A.2)
b
; 32 ; 4a
,
2
which define the even and odd solutions to Weber’s differential equation in the
⋄
variable z = √b2a respectively.
Proof. By Erdélyi et al. (1954, p 313, Eq. 13) we note that,
Z ∞
−c
b2
D−c
uc−1 exp − au2 − bu dv = (2a) 2 Γ(c) exp 8a
0
√b
2a
,
(A.3)
where Dλ (z) is Whittaker’s (1902) parabolic cylinder function (Abramowitz and Stegun,
1992). By virtue of Eq. 4 on p. 117 of Erdélyi et al. (1981), we can decompose
Dλ (z) into a sum of an even and odd function. Replacing this decomposition for
Dλ (z) in Eq. (A.3) and an application of the duplication formula of the gamma
function yields the statement.
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
9
References
Abramowitz, M. and Stegun, I. A., eds. (1992). Handbook of mathematical
functions with formulas, graphs, and mathematical tables. Dover Publications,
Inc., New York Reprint of the 1972 edition. MR1225604 (94b:00012)
Bayarri, M. J. (1981). Bayesian inference on the correlation coefficient of a
bivariate normal population. Trabajos de Estadı́stica y de Investigación Operativa 32 18–31. MR697200 (84i:62047)
Berger, J. O., Bernardo, J. M. and Sun, D. (2015). Overall objective
priors. Bayesian Anal. 10 189–221.
Berger, J. O. and Sun, D. (2008). Objective priors for the bivariate normal
model. The Annals of Statistics 36 963–982. MR2396821 (2009b:62058)
Bernardo, J. M. (2015). An overall prior for the five-parameter normal distribution. 11th International Workshop on Objective Bayes Methodology dedicated to Susie Bayarri.
Dawid, A. P., Stone, M. and Zidek, J. V. (1973). Marginalization paradoxes
in Bayesian and structural inference. Journal of the Royal Statistical Society.
Series B. Methodological 35 189–233. With discussion by D. J. Bartholomew,
A. D. McLaren, D. V. Lindley, Bradley Efron, J. Dickey, G. N. Wilkinson, A.
P.Dempster, D. V. Hinkley, M. R. Novick, Seymour Geisser, D. A. S. Fraser
and A. Zellner, and a reply by A. P. Dawid, M. Stone, and J. V. Zidek.
MR0365805 (51 ##2057)
Erdélyi, A., Magnus, W., Oberhettinger, F. and Tricomi, F. G. (1954).
Tables of integral transforms. Vol. I. McGraw-Hill Book Company, Inc.,
New York-Toronto-London Based, in part, on notes left by Harry Bateman.
MR0061695 (15,868a)
Erdélyi, A., Magnus, W., Oberhettinger, F. and Tricomi, F. G. (1981).
Higher transcendental functions. Vol. II. Robert E. Krieger Publishing Co.,
Inc., Melbourne, Fla. Based on notes left by Harry Bateman, Reprint of the
1953 original. MR698780 (84h:33001b)
Fisher, R. A. (1915). Frequency distribution of the values of the correlation
coefficient in samples from an indefinitely large population. Biometrika 507–
521.
Fisher, R. A. (1920). A Mathematical Examination of the Methods of Determining the Accuracy of an Observation by the Mean Error, and by the Mean
Square Error. Mon. Not. R. Astron. Soc. 80 758–770.
Fisher, R. A. (1921). On the “Probable Error” of a Coefficient of Correlation
Deduced from a Small Sample. Metron 1 3–32.
Fisher, R. A. (1922). On the Mathematical Foundations of Theoretical Statistics. Phil. Trans. R. Soc. A 222 309–368.
Fraser, D. A. S. (1961). On fiducial inference. Annals of Mathematical Statistics 32 661–676. MR0130755 (24 ##A614)
Gradshteyn, I. S. and Ryzhik, I. M. (2007). Table of integrals, series, and
products, Seventh ed. Elsevier/Academic Press, Amsterdam Translated from
the Russian, Translation edited and with a preface by Alan Jeffrey and
Daniel Zwillinger, With one CD-ROM (Windows, Macintosh and UNIX).
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
Ly, Marsman, & Wagenmakers/Analytic Correlation Posterior
10
MR2360010 (2008g:00005)
Hannig, J., Iyer, H. and Patterson, P. (2006). Fiducial generalized confidence intervals. J. Amer. Statist. Assoc. 101 254–269.
Jeffreys, H. (1935). Some tests of significance, treated by the theory of probability. Pro. Cambridge Philos. Soc. 31 203–222.
Jeffreys, H. (1961). Theory of probability, third ed. Oxford Science Publications. The Clarendon Press, Oxford University Press, New York.
MR745621 (85f:60005)
Lindley, D. V. (1965). Introduction to probability and statistics from a
Bayesian viewpoint. Part II: Inference. Cambridge University Press, New
York. MR0168084 (29 ##5349)
Ly, A., Marsman, M., Verhagen, A. J., Grasman, R. P. P. P. and Wagenmakers, E. J. (2016). A tutorial on Fisher information. Manuscript submitted for publication at the Journal of Mathematical Psychology.
Robert, C. P., Chopin, N. and Rousseau, J. (2009). Harold Jeffreys’s theory
of probability revisited. Statistical Science. A Review Journal of the Institute
of Mathematical Statistics 24 141–172. MR2655841 (2011c:62002)
Stigler, S. M. (2007). The epic story of maximum likelihood. Statistical Science. A Review Journal of the Institute of Mathematical Statistics 22 598–620.
MR2410255
Sun, D. and Berger, J. O. (2007). Objective Bayesian analysis for the multivariate normal model. In Bayesian statistics 8. Oxford Sci. Publ. 525–562.
Oxford Univ. Press, Oxford. MR2433206 (2009m:62019)
Tierney, L. (1994). Markov chains for exploring posterior distributions. The
Annals of Statistics 22 1701–1762. With discussion and a rejoinder by the
author. MR1329166 (96m:60150)
Whittaker, E. T. (1902). On the Functions associated with the Parabolic
Cylinder in Harmonic Analysis. Proceedings of the London Mathematical Society S1-35 417. MR1577011
imsart-generic ver. 2014/10/16 file: sunBergerCorrieR1.tex date: May 1, 2017
| 10math.ST
|
A Max-Sum algorithm for training discrete neural networks
Carlo Baldassi
DISAT, Politecnico di Torino, Corso Duca Degli Abruzzi 24, 10129 Torino and
arXiv:1505.05401v2 [cond-mat.dis-nn] 13 Aug 2015
Human Genetics Foundation, Via Nizza 52, 10124 Torino
Alfredo Braunstein
DISAT, Politecnico di Torino, Corso Duca Degli Abruzzi 24, 10129 Torino
Human Genetics Foundation, Via Nizza 52, 10124 Torino and
Collegio Carlo Alberto, Via Real Collegio 1, Moncalieri
We present an efficient learning algorithm for the problem of training neural networks
with discrete synapses, a well-known hard (NP-complete) discrete optimization problem.
The algorithm is a variant of the so-called Max-Sum (MS) algorithm. In particular, we
show how, for bounded integer weights with q distinct states and independent concave a
priori distribution (e.g. l1 regularization), the algorithm’s time complexity can be made to
scale as O (N log N ) per node update, thus putting it on par with alternative schemes, such
as Belief Propagation (BP), without resorting to approximations. Two special cases are of
particular interest: binary synapses W ∈ {−1, 1} and ternary synapses W ∈ {−1, 0, 1} with
l0 regularization. The algorithm we present performs as well as BP on binary perceptron
learning problems, and may be better suited to address the problem on fully-connected twolayer networks, since inherent symmetries in two layer networks are naturally broken using
the MS approach.
2
CONTENTS
I. Introduction
2
II. The network model
5
III. The Max-Sum algorithm
7
A. Max Convolution
9
B. The binary case
13
IV. Numerical results
16
V. Conclusions
18
Acknowledgments
20
Appendix
20
Details of the computation of the cavity fields.
References
20
23
I.
INTRODUCTION
The problem of training an artificial, feed-forward neural network in a supervised way is a well-known
optimization problem, with many applications in machine learning, inference etc. In general terms, the
problem consists in obtaining an assignment of “synaptic weights” (i.e. the parameters of the model) such
that the device realizes a transfer function which achieves the smallest possible error rate when tested on
a given dataset of input-output examples. Time is usually assumed to be discretized. In a single-layer
network, the transfer function is typically some non-linear function (e.g. a sigmoid or a step function) of
the scalar product between a vector of inputs and the vector of synaptic weights. In multi-layer networks,
many single-layer units operate in parallel on the same inputs, and their outputs provide the input to
3
other similar (with a varying degree of similarity) units, until the last layer is reached.
The most popular and successful approaches to these kind of optimization problems are typically
variants of the gradient descent algorithm, and in particular the back-propagation algorithm [1]. On
single-layer networks with simple non-linearities in their output functions these algorithms can even be
shown to achieve optimal results in linear time [2]; on multi-layer networks these algorithms suffer from
the usual drawbacks of gradient descent (mostly the presence of local minima, and slow convergence under
some circumstances).
On the other hand, gradient descent can only be applied to continuous problems. If the synaptic
weights are restricted to take only discrete values, the abovementioned family of methods can not be
applied; in fact, it is known that even the simplest version of the problem (classification using a singlelayer network) becomes computationally hard (NP-complete) in the worst-case scenario [3, 4]. However,
some theoretical properties of the networks, such as the storage capacity (i.e. the amount of information
which can be effectively stored in the device by setting the synaptic weights), are only slightly worse in
the case of discrete synapses, and other properties (e.g. robustness to noise and simplicity) would make
them an attractive model for practical applications. Indeed, some experimental results [5–7], as well as
arguments from theoretical studies and computer simulations [8–11], suggest that long term information
storage may be achieved by using discrete — rather than continuous — synaptic states in biological neural
networks.
Therefore, the study of neural network models with discrete weights is interesting both as a hard
combinatorial optimization problem and for its potential applications, in practical implementations as
well as for modeling biological networks. On the theoretical side, some light has been shedded upon the
origin of the computational hardness in these kind of problems by the study of the space of the solutions
by means of methods derived from Statistical Physics approaches [12, 13]: in brief, most solutions are
isolated, i.e. far from each other, and the energy landscape is riddled with local minima which tend to
trap purely local search methods, which thus show very poor performance. On the application side, a
family of heuristic algorithms, derived from the cavity method, have been devised, which exhibit very
good performance on random instances, both in terms of solution time and in terms of scaling with the
4
size of the problem.
In particular, it was first shown in [14] that a version of the Belief Propagation (BP) algorithm [15]
with the addition of a reinforcement term was able to efficiently solve the problem of correctly classifying
αN random input-output associations using a single-layer network, or a tree-like two-layer network, with
N synapses, up to a value of α close to the theoretical upper bound. For the single-layer case, the
theoretical bound for binary synapses is αc ' 0.83 [12], while the algorithmic bound as estimated from
extensive simulations up to N = 106 is αBP ' 0.74. Two more algorithms, obtained as crudely simplified
versions of the reinforced BP, were later shown [16, 17] to be able to achieve very similar performances,
despite being simpler and working in an on-line fashion. The time complexity of all these algorithms was
√
measured to be of order N log N per pattern; the BP algorithm in particular achieves this performance
thanks to a Gaussian approximation which is valid at large N .
When considering multi-layer networks, the original BP approach of [14] can only effectively deal with
tree-like network structures; fully-connected structures (such as those commonly used in machine learning
tasks) can not be addressed (at least not straightforwardly) with this approach, due to strong correlations
arising from a permutation symmetry which emerges in the second layer.
In this paper, we present a new algorithm for addressing the problem of supervised training of network
with binary synapses. The algorithm is a variant of the so-called Max-Sum algorithm (MS) [15] with an
additional reinforcement term (analogous to the reinforcement term used in [14]). The MS algorithm is
a particular zero-temperature limit of the BP algorithm; but it should be noted that this limit can be
taken in different ways. In particular, the BP approach in [14] was applied directly at zero temperature
as patterns had to be learned with no errors. In the MS approach we present here, in addition to hard
constraints imposing that no errors are made on the training set, we add external fields with a temperature
that goes to zero in a second step. Random small external fields also break the permutation symmetry
for multi-layer networks.
In the MS approach, the Gaussian approximation which is used in the BP approach can not be used,
and a full convolution needs to be computed instead: this in principle would add a factor of N 2 to the
time complexity, but, as we shall show, the exploitation of the convexity properties of the problem allows
5
to simplify this computation, reducing the additional factor to just log N .
This reinforced MS algorithm has very similar performance to the reinforced BP algorithm on single
layer networks in terms of storage capacity and of time required to perform each node update; however,
the number of updates required to reach a solution scales polynomially with N , thus degrading the overall
scaling. On fully-connected multi-layer networks, the MS algorithm performs noticeably better than BP.
The rest of the paper is organized as follows: in Section II we present the network model and the
mathematical problem of learning. In Section III we present the MS approach for discrete weights. We
show how the inherent equations can be solved efficiently thanks to properties of the convolution of concave
piecewise-linear functions, and describe in complete detail the implementation for binary weights. Finally,
in Section IV we show simulation results for the single and two-layer case.
II.
THE NETWORK MODEL
We consider networks composed of one or more elementary “building blocks” (units), each one having
a number of discrete weights and a binary transfer function, which classify binary input vectors. Units
can be arranged as a composed function (in which the output from some units is considered as the input
of others) in various ways (also called architectures) that is able to produce a classification output from
each input vector.
We denote the input vectors as ξ µ = {ξiµ }i=1,...,N ∈ {−1, +1}N (where µ is a pattern index) and the
weights as W k = Wik i=1,...,N (where k is a unit index). In the following, the Wik are assumed to take q
evenly spaced values; we will then explicitly consider the cases q = 2 with Wik ∈ {−1, 1} and q = 3 with
Wik ∈ {−1, 0, 1}. The output of the unit is given by:
σ
kµ
= sign
N
X
!
Wik ξiµ
(1)
i=1
with the convention that sign (0) = 1.
We will consider two cases: a single layer network and two-layer comitee machine. In single-layer
networks, also called perceptrons [18], there is a single unit, and therefore we will omit the index k. Fully
6
connected two-layer networks consist of K units in the second layer, each of which receives the same input
vector ξ µ , and the output function of the device is
σ µ = sign
K
X
!
σ kµ
k=1
This kind of architecture is also called a committee or consensus machine [19]. When K = 1, this
reduces to the perceptron case. In a tree-like committee machine the input vectors would not be shared
among the units; rather, each unit would only have access to a subset of the input vectors, without
overlap between the units. For a given N , the tree-like architectures are generally less powerful (in terms
of computational capabilities or storage capacity) than the fully-connected ones, but are easier to train
[20]. Intermediate situations between these two extremes are also possible. In fully-connected committee
machines there is a permutation symmetry in the indices k, since any two machines which only differ by
a permutation of the second layer’s indices will produce the same output.
Throughout this paper we will consider supervised contexts, in which each pattern µ has an associated
µ
desired output σD
.
In classification (or storage) problems, M = αN K association pairs of input vectors ξ µ and correµ
sponding desired outputs σD
are extracted from some probability distribution, and the goal is to find a
k
µ
.
set of weights W k such that ∀µ ∈ {1, . . . , αN K} : σ µ = σD
In random generalization problems, the input patterns ξ µ are still extracted from some probability
µ
distribution, but the desired outputs σD
are computed from some rule, usually from a teacher device
(teacher-student problem). The goal then is to learn the rule itself, i.e. to achieve the lowest possible
error rate when presented with a pattern which was never seen during the training phase. If the teacher’s
architecture is identical to that of the student device, this can be achieved when the student’s weights
match those of the teacher (up to a permutation of the units’ indices in the fully-connected case).
In the following, we will always address the problem of minimizing the error function on the training
7
patterns:
E
αN
n
o X
n
o X
Wk
=
Eµ W k
−
Γki Wik
k
k
µ=1
=
αN
X
Θ
µ=1
µ
−σD
sign
(2)
i
K
X
k=1
sign
N
X
!!!
Wik ξiµ
−
X
i=1
Γki Wik
i
where Θ (x) is the Heaviside step function Θ (x) = 1 if x > 0 and 0 otherwise. The term Γki Wik has
the role of an external field, and can be used e.g. to implement a regularization scheme; in the following,
we will always assume it to be concave. For example we can implement l1 regularization by setting
Γki (Wik ) = −λ|Wik | where λ > 0 is a parameter. The first term of expression (2) therefore counts the
number of misclassified patterns and the second one favours sparser solutions.
Throughout the paper, all random binary variables are assumed to be extracted from an unbiased
i.i.d. distribution.
Under these conditions, it is known that in the limit of N 1 there are phase transitions at particular
values of α. For single units (perceptrons) with binary ±1 synapses, for the classification problem, the
minimum number of errors is typically 0 up to αc ' 0.83. For the generalization problem, the number of
devices which are compatible with the training set is larger than 1 up to αT S ' 1.245, after which the
teacher perceptron becomes the only solution to the problem.
III.
THE MAX-SUM ALGORITHM
Following [14], we can represent the optimization problem of finding the zeros of the first term of
eq. (2) on a complete bipartite factor graph. Starting from the single-layer case, the graph has N vertices
(variable nodes) representing the Wi values and αN factor nodes representing the error terms Eµ (W).
The standard MS equations for this graph involve two kind of messages associated with each edge of
the graph; we indicate with Φtµ→i (Wi ) the message directed from node µ to variable i at time step t, and
with Ψti→µ (Wi ) the message directed in the opposite direction.
These messages represent a certain zero-temperature limit of BP messages, but have also a direct interpretation in terms of energy shifts of modified systems. Disregarding an insubstantial additive constant,
8
message Φµ→i (Wi ) represents the negated energy (2) restricted to solutions taking a specific value Wi
for variable i, on a modified system in which the energy depends on Wi only through the factor node
µ, i.e. in which all terms Wi ξiν for all ν 6= µ are removed from the energy expression (2). Similarly,
message Ψti→µ (Wi ) represents an analogous negated energy on a modified system in which the term Eµ
is removed. For factor graphs that are acyclic, the MS equations can be thought of as the dynamic
programming exact solution. In our case, the factor graph, being complete bipartite, is far from being
acyclic and the equations are only approximate. For BP, the approximation is equivalent to the one of
the Thouless-Anderson-Palmers equations [21] and is expected to be exact in the single-layer case below
the critical capacity [12]. For a complete description of the MS equations and algorithm, see [15].
The MS equations [15] for energy (2) are:
Φt+1
µ→i (Wi ) =
max
{Wj }j6=i : Eµ (W)=0
Ψti→µ (Wi ) = Γi (Wi ) +
X
X
t+1
Ψtj→µ (Wj ) − Zµ→i
(3)
j6=i
t
Φtµ0 →i (Wi ) − Zi→µ
(4)
µ0 6=µ
t
t
where Zµ→i
and Zi→µ
are normalization scalars that ensure
P
Wi
Ψti→µ (Wi ) =
P
Wi
Ψtµ→i (Wi ) = 0 and
can be computed after the rest of the RHS. At any given time t, we can compute the single-site quantities
Ψti (Wi ) = Γi (Wi ) +
X
Φtµ→i (Wi ) − Zit
(5)
µ
and use them to produce an assignment of the W’s:
Wit = argmaxWi Ψti (Wi )
(6)
The standard MS procedure thus consists in initializing the messages Ψ0i→µ (Wi ), iterating eqs. (3) and
(4) and, at each time step t, computing a vector W t according to eqs. (5) and (6) until either E W t = 0
(in the absence of prior terms, i.e. when λ = 0), or the messages converge to a fixed point, or some
maximum iteration limit is reached.
Strictly speaking, standard MS is only guaranteed to reach a fixed point if the factor graph is acyclic,
which is clearly not the case here. Furthermore, if the problem has more than one solution (ground state),
9
the assignment in eq. (6) would not yield a solution even in the acyclic case. In order to (heuristically)
overcome these problems, we add a time-dependent reinforcement term to eqs. (4) and (5), analogously
to what is done for BP [14]:
Ψti→µ (Wi ) = r t Ψit−1 (Wi ) + Γi (Wi ) +
X
t
Φtµ0 →i (Wi ) − Zi→µ
(7)
µ0 6=µ
Ψti (Wi ) = r t Ψit−1 (Wi ) + Γi (Wi ) +
X
Φtµ0 →i (Wi ) − Zit
(8)
µ0
where r > 0 controls the reinforcement speed. This reinforcement term in the case of standard BP
implements a sort of “soft decimation” process, in which single variable marginals are iteratively pushed
to concentrate on a single value. For the case of MS, this process is useful to aid convergence: on a system
in which the MS equations do not converge, the computed MS local fields still give some information about
the ground states and can be used to iteratively “bootstrap” the system into one with very large external
fields, i.e. fully polarized on a single configuration [22]. The addition of this term introduces a dependence
on the initial condition. Experimentally, by reducing r this dependence can be made arbitrarily small,
leading to more accurate results (see Sec. IV), at the cost of increasing the number of steps required for
convergence; our tests show that the convergence time scales as r−1 .
Furthermore, in order to break symmetries between competing configurations, we add a small
symmetry-breaking concave noise Γ0i (Wi ) 1 to the external fields Γi (Wi ); this, together with the
addition of the reinforcement term, is sufficient to ensure — for all practical purposes — that the argmax
in (6) is unique at every step of the iteration.
A.
Max Convolution
While Eq. (7) can be efficiently computed in a straightforward way, the first term of Eq. (3) involves
a maximum over an exponentially large set. The computation of Eq. (3) can be rendered tractable by
P
adding 0 = max∆ L ∆, j6=i ξjµ Wj where L (x, y) is equal to 0 if x = y and −∞ otherwise, which leads
to the following transformations:
10
t+1
Φt+1
µ→i (Wi ) + Zµ→i =
X
max
{Wj }j6=i : Eµ (W)=0
Ψtj→µ (Wj )
j6=i
X µ
X
=
max
max L ∆,
ξj W j +
Ψtj→µ (Wj )
{Wj }j6=i : Eµ (W)=0 ∆
j6=i
j6=i
X
t
max
Ψ
(W
)
=
max
j
P
j→µ
µ
µ
∆:σD
(∆+ξiµ Wi )≥0 {Wj }j6=i : j6=i ξj Wj =∆
j6=i
=
max
µ
∆:σD
(∆+ξiµ Wi )≥0
t
Fµ→i
(∆) ,
(9)
t
where in the last step above Fµ→i
(∆) is defined as:
t
Fµ→i
(∆) =
max
P
{Sj }j6=i :
j6=i Sj =∆
X
Ψtj→µ Sj ξjµ .
(10)
j6=i
The right-hand side of (10) is usually called a “Max-Convolution” of the functions fj (Sj ) =
Ψj→µ Sj ξjµ for j 6= i, and is analogous to the standard convolution but with operations (max, +)
substituting the usual (+, ×). As standard convolution, the operation is associative, which allows to
compute the convolution of the N − 1 functions in a recursive way. As the convolution of two functions
with discrete domains {0, . . . , q1 } and {0, . . . , q2 } respectively can be computed in q1 q2 operations and
has domain in {0, . . . , q1 + q2 }, it follows that (10) can be computed in O N 2 operations. In principle,
this computation must be performed N times for each pattern µ to compute all Φµ→i messages, in a total
of time O N 3 .
A technique like the one described in [23–25] can be employed to reduce this by a factor N , coming
back again to O N 2 operations per pattern update, as follows. Precomputing the partial convolutions
Ln and Rn of f1 , . . . , fn and fn , . . . , fN (respectively) for every n in 1, . . . , N can be done using N 2
operations in total; then the convolution of {fj }j6=i can be computed as the convolution of Li−1 and Ri+1 .
Computing this convolution would require O(N 2 ) operations but fortunately this will not be needed; the
11
computation of (9) can proceed as:
t+1
Φt+1
µ→i (Wi ) + Zµ→i =
maxµ
µ
∆:σD
(∆+ξi Wi )≥0
n
o
max Li−1 (z) + Ri+1 (∆ − z)
z
(
)
= max Li−1 (z) +
z
max
µ
∆:σD
(∆+ξiµ Wi )≥0
Ri+1 (∆ − z)
n
o
µ
σD
= max Li−1 (z) + Ri+1
(z + ξiµ Wi )
(11)
z
where we defined Riσ (x) = max∆:σ(∆+x)≥0 Ri (∆). As the vectors Riσ can be pre-computed recursively
in a total time of O N 2 and (11) requires time O (N ), we obtain a grand total O N 2 operations per
pattern update, or O(M N 2 ) per iteration. Unfortunately, this scaling is normally still too slow to be of
practical use for moderately large values of N and we will thus proceed differently by exploiting convexity
properties. However, note that the above scaling is still the best we can achieve for the general case in
which regularization terms are not concave.
At variance with standard discrete convolution, in general Max-Convolution does not have an analogous
to the Fast Fourier Transform, that would allow a reduction of the computation time of a convolution of
functions with O(N ) values from N 2 to N log N . Nevertheless, for concave functions the convolution can
be computed efficiently, as we will show below. Note that for this class of functions, an operation that is
analogous to the Fast Fourier Transform is the Legendre-Fenchel transform [26], though it will be simpler
to work with the convolution directly in the original function space.
First, let us recall well-known results about max-convolution of concave piecewise-linear functions in
def
the family C = {f : R≥0 → R ∪ {−∞}} [26]. First, the max-convolution f12 = f1 > f2 of f1 , f2 ∈ C
def
belongs to C. Moreover, f1 > f2 can be built in an efficient way from f1 and f2 . Start with x12
=
1
1
2
i
inf {x : (f1 > f2 ) (x) > −∞}, which is easily computed as x12
1 = x1 + x1 with x1 = inf {x : fi (x) > −∞}.
Moreover, (f1 > f2 ) x12
= f1 x11 + f2 x21 . Then, order the set of linear pieces from f1 and f2 in
1
12
decreasing order of slope and place them in order, starting from x12
to form a piecewise-linear
1 , f x1
continuous function. The method is sketched in Fig. 1. In symbols, let us write each concave piecewiselinear function fi (x) when x ≥ xi1 as:
fi (x) =
ki
X
j=1
x − xij
+
aij + fi xi1
i = 1, 2
12
with aij ∈ [0, ∞] for j = 2, . . . , ki and xi1 < xi2 < · · · < xiki . Here we used the notation[27] y+ = 21 (|y| + y).
h
i
P
This function is concave, as for x ∈ xij , xij+1 the slope is bij = jk=1 aik that is clearly decreasing with
j. To compute the convolution of f1 and f2 , just order the slopes bij ; i.e. compute a one to one map
π : (i, j) 7→ c from couples i ∈ {1, 2},1 ≤ j ≤ ki to integers 1 ≤ c ≤ k1 + k2 such that π (i, j) < π (i0 , j 0 )
0
implies bij > bij 0 . The max convolution for x ≥ x12
1 is still concave and piecewise-linear, and thus it can
be written as:
(f1 > f2 ) (x) =
k12
X
x − x12
c
c=1
+
12
a12
c + (f1 > f2 ) x1
where k12 = k1 + k2 − 1. For each c we can retrieve (i (c) , j (c)) = π −1 (c); with this, the parameters of
i(1)
i(c+1)
i(c)
i(c)
i(c)
12
12
12
the convolution are a12
1 = bj(1) , ac+1 = bj(c+1) − bj(c) and xc+1 = x1 + xj(c)+1 − xj(c) .
For more details about the max-convolution of piecewise-linear concave functions, see e.g. [26, Part
II].
We now consider the case of functions defined on a discrete domain. Let f, g be concave discrete
functions in
D = {f : {0, . . . , q − 1} → R ∪ {−∞}}
We will define the continuous extension fˆ ∈ C as the piecewise-linear interpolation of f , with value −∞
for arguments in (−∞, 0) ∪ (q − 1, ∞). This can be written as:
fˆ (x) =
q
X
(x − j + 1)+ aj + f (0)
j=1
with a1 = f (1) − f (0), aj+1 = f (j + 1) − 2f (j) + f (j − 1) (implying aq = −∞). It is easy to see that
h = fˆ > ĝ coincides with the discrete convolution of f and g in its (discrete) domain; the reason is simply
that h is also piecewise-linear, with kinks only in discrete values ,{0, . . . 2 (q − 1)}.
When computing the convolution of N functions f1 , . . . , fN with domain in {0, . . . , q − 1}, one can
clearly order all slopes together in one initial pass, using N q log (N q) comparisons. It is also easy to see
that, if one has the full convolution, it is simple to compute a “cavity” convolution in which one function fi
is omitted, in O (q) time: this is achieved by simply removing the q slopes of fi from the full convolution.
13
a
b
d
e
c
d
a
b
e
c
FIG. 1. Sketch of the discrete max-convolution f > g of piecewise-linear concave functions f, g. The result is simply
obtained by sorting the pieces of the two functions in descending order of slope.
In order to apply this to eq. (10) the only remaining step is a trivial affine mapping of the functions
arguments Wj ξjµ on a domain {A + tB : t ∈ {0, . . . , q − 1}} to the domain {0, . . . , q − 1}. In the following,
we will show explicitly how to do this for the binary case q = 2, but the argument can be easily generalized
to arbitrary q. Note that, while in the binary case the Ψ functions are linear and thus trivially concave, in
the general case we need to ensure that both the initial values Φ0i→µ (Wi ) and the external fields Γi (Wi )
are concave; in such case, the iteration equations (3), (7) and (8) ensure that the concavity property holds
for all time steps t > 0.
B.
The binary case
We will show explicitly how to perform efficiently the computations for the binary case. In this
case we can simplify the notation by parametrizing the message with a single scalar, i.e. we can write
t
Φtµ→i (Wi ) = Wi φtµ→i and Ψti→µ (Wi ) = Wi ψi→µ
and Γi (Wi ) = Wi γi . Eqs. (3) and (7) then become:
φt+1
µ→i
1
=
2
X
max
W: Wi =1∧Eµ (W )=0
t
ψi→µ
= r t ψit−1 + γi +
X
µ0 6=µ
t
Wj ψj→µ
j6=i
φtµ0 →i
−
max
W: Wi =−1∧Eµ (W )=0
X
t
Wj ψj→µ
(12)
j6=i
(13)
14
Correspondingly, eqs. (8) and (6) simplify to:
ψit = r t ψit−1 + γi +
X
φtµ→i
(14)
µ
Wit = signψit
(15)
In order to apply the results of the previous sections, and perform efficiently the trace over all possible
assignments of W of eq. (12), we first introduce the auxiliary quantities
Fµt (S) =
X
t
Si ξiµ ψi→µ
(16)
i
Fµt (∆) =
max
Fµt (S)
P
{S: i Si =∆}
(17)
For simplicity of notation, we will temporarily drop the indices µ and t. We will also assume that
t
all values ψi→µ
are different: as remarked above, the presence of term Γi is sufficient to ensure that
this is the case, and otherwise we can impose an arbitrary order without loss of generality. With this
assumption, the function F, which is defined over ∆ = {−N, −N + 2, . . . , N − 2, N }, has a single absolute
maximum, and is indeed concave. The absolute maximum is obtained with the special configuration
t
S̃ = argmax{Si }i F (S), which is trivially obtained by setting S̃i = ξiµ signψi→µ
for all i. This configuration
P
˜ =
corresponds to a value ∆
i S̃i . Any variable flip with respect to this configuration, i.e. any i for
t
which Si = −S̃i , adds a “cost” ∆Fi = 2 ψi→µ
in terms of F (S). Therefore, if we partition the indices
n
o
i in two groups S+ and S− defined by S± = i : S̃i = ±1 , and we sort the indices within each group
in ascending order according to ∆Fi , we can compute the function F (∆) for each ∆ by constructively
˜ , then
computing the corresponding optimal configuration S, in the following way: we start from F ∆
proceed in steps of 2 in both directions subtracting the values ∆Fi in ascending order, using the variable
˜ and those in S− for ∆ > ∆.
˜
indices in S+ for ∆ < ∆
This procedure also associates a “turning point” Ti to each index i, defined as the value of ∆ for which
the optimal value of Wi changes sign, or equivalently such that F Ti + S̃i − F Ti − S̃i = ∆Fi . This
also implies that:
X
˜ −
F (∆) = F ∆
Θ S̃i (Ti − ∆) ∆Fi
i
(18)
15
We can also bijectively associate an index to each value of Ti , by defining jk such that Tjk = k.
Next, consider the same quantity where a variable i is left out of the sum (see eq. (10))
X
t
F (i) S(i) =
Sj ξjµ ψj→µ
(19)
j6=i
F
(i)
(∆) =
max
P
{Sj }j6=i :
j6=i
F (i) S(i)
(20)
Sj =∆
Clearly, one gets the same overall picture as before, except with a shifted argmax, and shifted turning
points. The shifts can be easily expressed in terms of the previous quantities, and the expressions used
for computing eq. (12) as:
φt+1
µ→i
ξµ
= i
2
!
max
µ
∆: σD
∆>0
F
(i)
(∆ − 1) −
max
µ
∆: σD
∆>0
F
(i)
(∆ + 1)
(21)
The full details of the computation are provided in the Appendix, Sec. V. Here, we report the end
result:
µ
µ
˜ + S̃i + 1 (Θ (Ti − 1) hj + Θ (−Ti + 1) hj ) +
φt+1
Θ σD
Θ −∆
0
2
µ→i = ξi
µ
˜ − S̃i + 1 Θ (Ti + 1) hj + Θ (−Ti − 1) hj
Θ ∆
+ Θ −σD
−2
0
(22)
t
where hj = −ξjµ ψj→µ
. From this expression, we see that we can update the cavity fields φµ→i very
efficiently for all i, using the following procedure:
• We do one pass of the whole array of hi by which we determine the S̃i values, we split the indices
˜ This requires O (N ) operations (all of which are trivial).
j into S+ and S− and we compute ∆.
• We separately partially sort the indices in S+ and S− and get j−2 , j0 and j2 and the turning points
Ti . This requires at most O (N log N ) operations. Note that we can use a partial sort because we
˜ and so we know how many indices we need to sort, and from which set S± , until we
computed ∆,
get to the ones with turning points around 0; also, we are only interested in computing Θ (Ti − 1)
and Θ (Ti + 1) instead of all values of Ti . This makes it likely for the procedure to be significantly
less computationally expensive than the worst case scenario.
• For each i we compute φt+1
µ→i from the equation above. This requires O (1) operations (implemented
in practice with three conditionals and a lookup).
16
IV.
NUMERICAL RESULTS
We tested extensively the binary case q = 2 with Wi ∈ {−1, +1} and the ternary case q = 3 with
Wi ∈ {−1, 0, 1}, for single layer networks.
We start from the binary case. Fig. 2 shows the probability of finding a solution when fixing the
reinforcement rate r, for different values of r and α. Reducing r allows to reach higher values of α; the
shape of the curves suggest that in the limit r → 0 there would be sharp transitions at critical values of
α’s. In the classification case, Fig. 2A, the transition is around α ' 0.75, while the theoretical critical
capacity is αc = 0.83. This value is comparable to the one obtained with the reinforced BP algorithm
of [14]. In the generalization case, there are two transitions: the first one occurs around α ' 1.1, before
the first-order transition at αT S = 1.245 where, according to the theory, the only solution is the teacher;
the second transition occurs around α ' 1.5. This second transition is compatible with the end of the
meta-stable regime (see e.g. [2]); indeed, after this point the algorithm is able to correctly infer the teacher
perceptron.
A second batch of experiments on the same architecture, in the classification case, is shown in Fig. 3.
In this case, we estimated the maximum value of r which allows to find a solution, at different values of
N and α; i.e. for each test sample we started from a high value of r (e.g. r = 10−1 ) and checked if the
algorithm was able to find a solution; if the algorithm failed, we reduced r and tried the same sample
again. In the cases shown, the solution was always found eventually. The results indicate that the value
of r required decreases with N , and the behaviour is well described by a power low, i.e. r = aN b with
a < 0 and b < 0, where the values of a and b depend on α. Since the number of iterations required is
inversely proportional to r (not shown), this implies that the overall solving time of the MS algorithm is
of O N 1−b log (N ) , i.e. it is worse than the reinforced BP in this respect. The value of b is between 0
and −0.5 up to α = 0.6, after which its magnitude decreases abruptly (see Fig. 3B). The behaviour for
large α seems to be reasonably well fit by a curve b (α) =
c
αU −α ,
suggesting the presence of a vertical
asymptote at αU = 0.755 ± 0.004, which is an estimate of the critical capacity of the algorithm in the
limit of large N .
In the ternary single layer case, we tested learning of {−1, 1} random patterns with ternary {−1, 0, 1}
17
A
1
B
0.8
solving probability
solving probability
0.8
1
0.6
0.4
reinforcement r
10⁻¹
10⁻²
10⁻³
10⁻⁴
0.2
0.6
0.4
reinforcement r
10⁻¹
10⁻²
10⁻³
10⁻⁴
0.2
0
0
0
0.1
0.2
0.3
0.4
0.5
patterns per synapse
0.6
0.7
0.8
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
patterns per synapse
FIG. 2. Solving probability. Probability of finding a solution for different values of α, in the binary perceptron
case with N = 1001, with different values of the reinforcement rate parameter r. Performance improves with lower
values of r. A. Classification case, 100 samples per point. The theoretical capacity is αc = 0.83 in this case. B.
Generalization case, 20 samples per point. In this case, the problem has multiple solutions up to αT S = 1.245, after
which the only solution is the teacher.
weights and concave bias (i.e. prior). In practice, we use the function Γi (Wi ) = λδ (Wi ) + Γ0i (Wi ) (where
Γ0 is the symmetry-breaking noise term and λ is sufficiently large) to favour zero weights, so solutions
with a minimimal number of zeros are searched, i.e. we add an l0 regularization term. Results (See
Fig. 4) are qualitatively similar to the {−1, 1} case with a larger capacity (around α = 1; the critical
capacity is αc = 1.17 in this case). The average non-zero weights in a solution grows when getting closer
to the critical α up to a value that is smaller than 2/3 (the value that makes the entropy of unconstrained
{−1, 0, 1} perceptrons largest).
In the fully-connected multi-layer case, the algorithm does not get as close to the critical capacity
as for the single-layer case, but it is still able to achieve non-zero capacity in rather large instances.
For example, in the classification case with binary synapses, N = 1001 inputs, K = 3 hidden units,
the algorithmic critical capacity is α ' 0.33 when r = 10−5 (tested on 20 samples), corresponding to
storing M = 1001 patterns with 3003 weights (thus demonstrating a greater discriminatory power than
18
0.1
0.2
0.3
0.4
0.5
0.6
0.625
0.65
0.675
0.7
max. reinforcement rate r
10⁻¹
10⁻²
B
-0.2
-0.4
-0.6
r exponent
A
-0.8
-1
10⁻³
-1.2
10⁻⁴
1001
2001
4001
8001
16001
32001
-1.4
0.1
0.2
0.3
number of synapses N
0.4
0.5
0.6
0.7
patterns per synapse
FIG. 3. Maximum reinforcement rate. A. Average value of the maximum reinforcement rate r which allows
to find a solution, in the binary perceptron classification case, at various values of N and α, in log-log scale. The
reinforcement rate decreases with N and α. Error bars show the standard deviation of the distributions. Black lines
show the result of fits of the form r (α, N ) = a (α) N b(α) , one for each value of α. The number of samples varied
between 100 for N = 1001 and 10 for N = 32001. B. The fitted values of the exponents b (α) in panel A. The
continuous curve shows a fit of the data for α ≥ 0.6 by the function b (α) =
c
αU −α .
The fit yields c = −0.063 ± 0.002
and αU = 0.755 ± 0.004. The value of αU is an estimate of the critical capacity of the algorithm.
the single-layer case with the same input size). The reason for the increased difficulty in this case is not
completely clear: we speculate that it is both due to the permutation symmetry between the hidden units
and to replica-symmetry-breaking effects: these effects tend to trap the algorithm — in its intermediate
steps — in states which mix different clusters of solutions, making convergence difficult. Still, the use
of symmetry-breaking noise helps achieving non trivial results even in this case, which constitutes an
improvement with respect to the standard BP algorithm.
V.
CONCLUSIONS
Up to now, the large N limit could be exploited on BP equations for the learning problem with discrete synapses to obtain an extremely simple set of approximated equations that made the computation
19
1
r=10⁻²
r=10⁻³
r=10⁻⁴
⟨|W|⟩ (r=10⁻⁴)
0.6
⟨|W|⟩
solving probability
0.8
0.4
0.2
0
0.8
0.9
1
1.1
patterns per synapse
FIG. 4. Learning of random {−1, 1} patterns with a ternary w = {−1, 0, 1} perceptron, with dilution (regularization)
prior term; N = 1001, 20 samples per point. For solved instances with r = 10−4 , the average fraction of non-zero
weights is also shown (standard deviations smaller than point size).
of an iteration to scale linearly with the problem size M N . For the MS equations however, those approximations cannot be made and a naive approximation scales as M N 3 which is normally too slow for
most practical purposes. In this work, we showed that MS equations can be computed exactly (with no
approximations) in time M N log N , rendering the approach extremely interesting in practice. A word
is in order about the MS equations with reinforcement term, which we propose as a valid alternative to
Belief Propagations-based methods. Although we cannot claim any universal property of the reinforced
equations from theoretical arguments and we only tested a limited number of cases, extensive simulations for these cases and previous results obtained by applying the same technique to other optimization
problems of very different nature [22, 28–30] have confirmed the same qualitative behaviour; that is, that
the number of iterations until convergence scales as r−1 and that results monotonically improve as r
decreases. As an additional advantage of MS, inherent symmetries present in the original system are naturally broken thanks to ad-hoc noise terms that are standard in MS. The MS equations are additionally
20
computationally simpler because they normally require only sum and max operations, in contrast with
hyperbolic trigonometric functions required by BP equations. Extensive simulations for discrete {−1, 1}
and {−1, 0, 1} weights show that the performance is indeed very good, and the algorithm achieves a
capacity close to the theoretical one (very similar to the one of Belief Propagation).
ACKNOWLEDGMENTS
CB acknowledges the European Research Council for grant n° 267915.
APPENDIX
Details of the computation of the cavity fields.
In this section we provide the full details of the computation leading to eq. (22).
As noted in the main text, the expression of the cavity quantities F (i) (∆) (see eq. (20)) is analogous
˜ (i) = ∆
˜ − S̃i ,
to that of the non-cavity counterpart F (∆) (eq. (18)), where the argmax has changed to ∆
and the turning points have changed:
X
(i)
˜ (i) −
Θ S̃j Tj − ∆ ∆Fj
F (i) (∆) = F (i) ∆
(23)
j6=i
We need to express the relationship between the old turning points and the new ones: having omitted
the variable i, it means that there is a global shift of −S̃i , and that the turning points to the left (right)
of Ti have shifted to the right (left) if S̃i = 1 (S̃i = −1, respectively):
(i)
Tj = Tj − S̃i + 2S̃i Θ S̃i (Ti − Tj )
= Tj + sign Ti − Tj − S̃i
(24)
(note that we chose to use the convention that Θ (0) = 0).
Therefore we obtain:
X
˜ (i) −
F (i) (∆) = F (i) ∆
Θ S̃j Tj + sign Ti − Tj − S̃i − ∆ ∆Fj
j6=i
(25)
21
Next, we consider the cavity quantity:
Ci (∆, Si ) =
max
P
{Sj }j6=i :
j
Sj ξjµ =∆
X
t
Sj ξ µ ψj→µ
j
j6=i
= F (i) (∆ − Si )
(26)
which allows us to write eq. (12) as
φt+1
µ→i
!
ξµ
= i
2
max
µ
∆: σD
∆>0
Ci (∆, ξiµ )
−
max
µ
∆: σD
∆>0
Ci (∆, −ξiµ )
!
ξµ
= i
2
max
µ
∆: σD
∆>0
F
(i)
(∆ − 1) −
max
µ
∆: σD
∆>0
F
(i)
(∆ + 1)
(27)
(this is eq. (21) in the main text).
˜ (i) = ∆
˜ − S̃i . Using this fact, and eq. (25), we
Note that F (i) (∆) is concave and has a maximum at ∆
can derive explicit formulas for the expressions which appear in the cavity field, by considering the two
µ
separately, and simplifying the result with simple algebraic manipulations afterwards:
cases for σD
max
µ
∆: σD ∆>0
µ
F (i) (∆ − 1) = Θ σD
˜ − S̃i + Θ −∆
˜ − S̃i + 1 F (i) ∆
˜ + S̃i − 1 F (i) (0) +
Θ ∆
µ
˜ − S̃i + Θ ∆
˜ − S̃i + 1 F (i) (−2)
˜ + S̃i − 1 F (i) ∆
+Θ −σD
Θ −∆
µ
˜ − S̃i +
˜ − S̃i + 1 F (i) ∆
∆
= Θ σD
µ
˜
+Θ σD Θ −∆ + S̃i − 1 F (i) (0) +
µ
˜ − S̃i + 1 F (i) (−2)
+Θ −σD
Θ ∆
max
µ
∆: σD ∆>0
µ
F (i) (∆ + 1) = Θ σD
˜ − S̃i − 1 F (i) ∆
˜ − S̃i + Θ −∆
˜ + S̃i + 1 F (i) (2) +
Θ ∆
µ
˜ + S̃i + 1 F (i) ∆
˜ − S̃i + Θ ∆
˜ − S̃i − 1 F (i) (0)
+Θ −σD
Θ −∆
µ
˜ − S̃i − 1 F (i) ∆
˜ − S̃i +
= Θ σD
∆
µ
˜ + S̃i + 1 F (i) (2) +
+Θ σD
Θ −∆
µ
˜ − S̃i − 1 F (i) (0)
+Θ −σD
Θ ∆
22
Plugging these back in the expression for the cavity field, we can reach — again by simple algebraic
manipulations — an expression which only uses F (i) (−2), F (i) (0) and F (i) (2):
ξiµ µ ˜
µ
˜ − S̃i − 1
˜ − S̃i +
Θ σD ∆ − S̃i + 1 − Θ σD
∆
F (i) ∆
2
µ
˜ + S̃i − 1 F (i) (0) − Θ −∆
˜ + S̃i + 1 F (i) (2) +
+Θ σD
Θ −∆
µ
˜ − S̃i + 1 F (i) (−2) − Θ ∆
˜ − S̃i − 1 F (i) (0)
+ Θ −σD
Θ ∆
ξiµ ˜
µ (i)
δ ∆, S̃i sign σD
F (0) +
=
2
µ
˜ S̃i F (i) (2) + Θ −∆
˜ + S̃i − 1 F (i) (0) − F (i) (2) +
+Θ σD
−δ ∆,
µ
˜ S̃i F (i) (−2) + Θ ∆
˜ − S̃i − 1 F (i) (−2) − F (i) (0)
δ ∆,
+ Θ −σD
ξµ ˜
µ
µ
+
F (i) (0) − F (i) 2σD
= i δ ∆,
W̃i ξiµ sign σD
2
µ
˜ + W̃i ξ µ − 1 F (i) (0) − F (i) (2) +
Θ −∆
+Θ σD
i
µ
˜ − W̃i ξ µ − 1 F (i) (−2) − F (i) (0)
+ Θ −σD Θ ∆
i
ξµ
µ
˜ + S̃i + 1 F (i) (0) − F (i) (2) +
= i Θ σD
Θ −∆
2
µ
˜ − S̃i + 1 F (i) (−2) − F (i) (0)
+ Θ −σD
Θ ∆
φt+1
µ→i =
These expressions can be further simplified, since the differences between the values of F (i) at neighboring values only depends on the “steps” induced by the spins which are associated with turning points
in that region:
F (i) (0) − F (i) (2) =
X
Θ S̃j Tj + sign Ti − Tj − S̃i − 2 − Θ S̃j Tj + sign Ti − Tj − S̃i
∆Fj
j6=i
=−
X
S̃j δ Tj , 1 − sign Ti − Tj − S̃i ∆Fj
j6=i
= −S̃j0 (1 − δ (i, j0 )) Θ Ti − S̃i ∆Fj0 − S̃j2 (1 − δ (i, j2 )) Θ −Ti + 2 + S̃i ∆Fj2
= −S̃j0 Θ (Ti − 1) ∆Fj0 − S̃j2 Θ (−Ti + 1) ∆Fj2
23
where in the last step we used the Kronecker deltas to get rid for the differences between the two cases
for S̃i . The other case is very similar:
F (i) (−2) − F (i) (0) =
X
− Θ S̃j Tj + sign Ti − Tj − S̃i + 2
∆Fj
Θ S̃j Tj + sign Ti − Tj − S̃i
j6=i
=−
X
S̃j δ Tj , −1 − sign Ti − Tj − S̃i ∆Fj
j6=i
= −S̃j−2 (1 − δ (i, j−2 )) Θ Ti + 2 − S̃i ∆Fj−2 − S̃j0 (1 − δ (i, j0 )) Θ S̃i − Ti ∆Fj0
= −S̃j−2 Θ (Ti + 1) ∆Fj−2 − S̃j0 Θ (−1 − Ti ) ∆Fj0
t
Going back to the cavity fields, and defining hj = − 12 S̃j ∆Fj = −ξjµ ψj→µ
, we finally get eq. (22):
µ
µ
˜ + S̃i + 1 (Θ (Ti − 1) hj + Θ (−Ti + 1) hj ) +
φt+1
Θ σD
Θ −∆
0
2
µ→i = ξi
µ
˜ − S̃i + 1 Θ (Ti + 1) hj + Θ (−Ti − 1) hj
+ Θ −σD
Θ ∆
−2
0
[1] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams.
Learning representations by back-
propagating errors. Nature, 323(6088):533–536, October 1986.
[2] A Engel and C. van den Broeck. Statistical mechanics of learning. Cambridge University Press, Cambridge,
UK; New York, NY, 2001.
[3] Edoardo Amaldi, Eddy Mayoraz, and Dominique de Werra. A review of combinatorial problems arising in
feedforward neural network design. Discrete Applied Mathematics, 52(2):111–138, 1994.
[4] Avrim L Blum and Ronald L Rivest. Training a 3-node neural network is np-complete. Neural Networks,
5(1):117–127, 1992.
[5] Carl CH Petersen, Robert C Malenka, Roger A Nicoll, and John J Hopfield. All-or-none potentiation at ca3-ca1
synapses. Proceedings of the National Academy of Sciences, 95(8):4732–4737, 1998.
[6] Daniel H O’Connor, Gayle M Wittenberg, and Samuel S-H Wang. Graded bidirectional synaptic plasticity is
composed of switch-like unitary events. Proceedings of the National Academy of Sciences of the United States
of America, 102(27):9679–9684, 2005.
24
[7] Thomas M Bartol, Cailey Bromer, Justin P Kinney, Michael A Chirillo, Jennifer N Bourne, Kristen M Harris,
and Terrence J Sejnowski. Hippocampal spine head sizes are highly precise. bioRxiv, 2015.
[8] Upinder S Bhalla and Ravi Iyengar. Emergent properties of networks of biological signaling pathways. Science,
283(5400):381–387, 1999.
[9] William Bialek. Stability and noise in biochemical switches. Advances in neural information processing systems,
pages 103–109, 2001.
[10] Arnold Hayer and Upinder S Bhalla. Molecular switches at the synapse emerge from receptor and kinase
traffic. PLoS computational biology, 1(2):e20, 2005.
[11] Paul Miller, Anatol M Zhabotinsky, John E Lisman, and Xiao-Jing Wang. The stability of a stochastic camkii
switch: dependence on the number of enzyme molecules and protein turnover. PLoS biology, 3(4):e107, 2005.
[12] Werner Krauth and Marc Mézard. Storage capacity of memory networks with binary couplings. Journal de
Physique, 50(20):3057–3066, 1989.
[13] Haiping Huang and Yoshiyuki Kabashima. Origin of the computational hardness for learning with binary
synapses. Physical Review E, 90(5):052813, November 2014.
[14] Alfredo Braunstein and Riccardo Zecchina. Learning by message passing in networks of discrete synapses.
Physical Review Letters, 96(3):030201, January 2006.
[15] Marc Mézard and Andrea Montanari. Information, Physics, and Computation. Oxford University Press,
January 2009.
[16] Carlo Baldassi, Alfredo Braunstein, Nicolas Brunel, and Riccardo Zecchina. Efficient supervised learning in
networks with binary synapses. Proceedings of the National Academy of Sciences, 104(26):11079–11084, June
2007.
[17] Carlo Baldassi. Generalization learning in a perceptron with binary synapses. Journal of Statistical Physics,
136, September 2009.
[18] Frank Rosenblatt. The perceptron: a probabilistic model for information storage and organization in the brain.
Psychological review, 65(6):386, 1958.
[19] Nils J Nilsson. Learning machines. New York, McGrawHill, 1965.
[20] A Engel, HM Köhler, F Tschepke, H Vollmayr, and A Zippelius. Storage capacity and learning algorithms for
two-layer neural networks. Physical Review A, 45(10):7590, 1992.
[21] Yoshiyuki Kabashima. A CDMA multiuser detection algorithm on the basis of belief propagation. Journal of
Physics A: Mathematical and General, 36(43):11111, October 2003.
[22] M. Bailly-Bechet, C. Borgs, A. Braunstein, J. Chayes, A. Dagkessamanskaia, J.-M. François, and R. Zecchina.
25
Finding undetected protein associations in cell signaling by belief propagation. Proceedings of the National
Academy of Sciences, 108(2):882–887, January 2011.
[23] Alfredo Braunstein, Roberto Mulet, and Andrea Pagnani. Estimating the size of the solution space of metabolic
networks. BMC Bioinformatics, 9(1):240, May 2008.
[24] A. Braunstein, F. Kayhan, and R. Zecchina. Efficient data compression from statistical physics of codes over
finite fields. Physical Review E, 84(5):051111, November 2011.
[25] Alfredo Braunstein, Farbod Kayhan, and Riccardo Zecchina. Efficient LDPC codes over GF(q) for lossy data
compression. In IEEE International Symposium on Information Theory, 2009. ISIT 2009, Seul, Korea, 2009.
[26] Jean-Yves Le Boudec and Patrick Thiran. Network Calculus: A Theory of Deterministic Queuing Systems for
the Internet. Springer Science & Business Media, July 2001.
[27] We allow aiki = −∞ , in which case we conventionally define aiki y+ = 0 if y ≤ 0.
[28] Anthony Gitter, Alfredo Braunstein, Andrea Pagnani, Carlo Baldassi, Christian Borgs, Jennifer Chayes, Riccardo Zecchina, and Ernest Fraenkel. Sharing information to reconstruct patient-specific pathways in heterogeneous diseases. In Pacific Symposium on Biocomputing 2014, pages 39–50. World Scientific, 2013.
[29] F. Altarelli, A. Braunstein, L. Dall’Asta, and R. Zecchina. Optimizing spread dynamics on graphs by message
passing. Journal of Statistical Mechanics: Theory and Experiment, 2013(09):P09011, September 2013.
[30] F. Altarelli, A. Braunstein, L. Dall’Asta, J. R. Wakeling, and R. Zecchina. Containing Epidemic Outbreaks
by Message-Passing Techniques. Physical Review X, 4(2):021024, May 2014.
| 9cs.NE
|
IEEE TRANSACTIONS ON FUZZY SYSTEMS
1
Generic Evolving Self-Organizing Neuro-Fuzzy
Control of Bio-inspired Unmanned Aerial Vehicles
arXiv:1802.00635v1 [cs.SY] 2 Feb 2018
Md Meftahul Ferdaus, Student Member, IEEE, Mahardhika Pratama, Member, IEEE, Sreenatha G.
Anavatti, Matthew A. Garratt, and Yongping Pan
Abstract—At recent times, with the incremental demand of the
fully autonomous system, a huge research interest is observed in
learning machine based intelligent, self-organizing, and evolving
controller. In this work, a new evolving and self-organizing
controller namely Generic-controller (G-controller) is proposed.
The G-controller that works in the fully online mode with very
minor expert domain knowledge is developed by incorporating
the sliding model control (SMC) theory based learning algorithm
with an advanced incremental learning machine namely Generic
Evolving Neuro-Fuzzy Inference System (GENEFIS). The controller starts operating from scratch with an empty set of fuzzy
rules, and therefore, no offline training is required. To cope
with the plant’s vulnerable behavior, the controller can add, or
prune the rules on demand. Control law and adaptation laws
for the consequents are derived from the SMC algorithm to
establish a stable closed-loop system, where the stability of the
G-controller is guaranteed using the Lyapunov function. The
uniform asymptotic convergence of tracking error to zero is
witnessed through the implication of an auxiliary robustifying
control term. In addition, the implementation of the multivariate
Gaussian function helps the controller to handle the non-axis
parallel data from the plant and consequently enhances the
robustness against the uncertainties and environmental perturbations. Finally, the controller’s performance has been evaluated
by observing the tracking performance in controlling simulated
plants of unmanned aerial vehicle namely bio-inspired flapping
wing micro air vehicle (BIFW MAV) and hexacopter for a variety
of trajectories.
Index Terms—Data-driven controller, Evolving, GENEFIS,
Model-free, Neuro-fuzzy, Self-constructing, Sliding mode control
I. I NTRODUCTION
N AUTONOMOUS unmanned vehicle systems, to obtain
an accurate first principle based model is considerably
arduous due to their highly non-linear and over-actuated or
under-actuated behavior. Besides, various uncertainty factor
like impreciseness in the obtained data from sensors, induced
noise by the sensors, outdoor environmental uncertainties like
I
Md Meftahul Ferdaus is with the School of Engineering and Information
Technology, University of New South Wales at the Australian Defence Force
Academy, Canberra, ACT 2612, Australia, e-mail: [email protected].
Mahardhika Pratama is with the School of Computer Science and Engineering, Nanyang Technological University, Singapore, 639798, Singapore,
e-mail: [email protected].
Sreenatha G. Anavatti is with the School of Engineering and Information
Technology, University of New South Wales at the Australian Defence Force
Academy, Canberra, ACT 2612, Australia, e-mail: [email protected].
Matthew A. Garratt is with the School of Engineering and Information
Technology, University of New South Wales at the Australian Defence Force
Academy, Canberra, ACT 2612, Australia, e-mail: [email protected].
Yongping Pan is with the Department of Biomedical Engineering, National
University of Singapore, Singapore, e-mail: [email protected].
wind gust, motor degradation, etc. are strenuous or impossible to integrate into the first principle models. In such
circumstances, approaches without the necessity of accurate
mathematical models of the system under control, are much
appreciated. Being a model-free approach, the Neural Network
(NN) and Fuzzy Logic system (FLS) based controllers have
been successfully implemented in many control applications
[1], [2], [3], [4] over the past few years. Recently, systems
with an amalgamation of FLS and NN namely Fuzzy Neural
Network (FNN) based controllers are becoming popular in a
variety of engineering application, such as in controlling robot
manipulator [5], anti-lock braking systems [6], temperature
control [7], controlling motor drive [8], direct current converter
[9] etc.
To develop an FLS, NN, or an FNN controller with a better
accuracy, is challenging. There exist numerous methods to
develop FLS, NN, or FNN controllers [10], [11], [12]. When
the system dynamics and various characteristics of a plant to
be controlled are known, then the information is utilized to
train the controller. It constructs a fixed-structured controller
with a certain number of rules, membership functions, neurons,
and layers. Due to the fixed structure, these controllers cannot
cope with changing plant dynamics. Similarly, during online
control application, the plant dynamics and other system
information are not readily available. Therefore, the fixedstructured FLS, NN, or FNN controller become unreliable.
Furthermore, the characteristic of real-world plants is nonstationary. This should not be handled by a fixed-size controller. To handle uncertainties in control system, researchers
have tried to combine the FLS, NN, FNN system with sliding
mode control (SMC) [13], H∞ control, back-stepping, etc.
Such amalgamation empowers the FLS, NN, FNN controller
with the feature of tuning parameters, which provides a more
robust and adaptive control structure. It assists them to mitigate
the adverse effects of various uncertainties and perturbations.
However, such adaptive FNN control structures are not able
to evolve their structures by adding or pruning rules. It forces
the controller to determine the number of rules a priori,
where a selection of few fuzzy rules may hinder to achieve
adequate and desired control performance. On the other hand,
consideration of too many rules may create complexity, or
make it impossible to implement in real time.
The problems mentioned above can be circumvented by
implementing evolving FLS, NN, FNN controllers that can
evolve their structure by adding, or deleting rules through selforganizing techniques. In recent time, researchers are trying
to develop evolving FLS, NN, FNN controllers by following
IEEE TRANSACTIONS ON FUZZY SYSTEMS
various approaches to add or delete the rules [14]. In the rule
growing mechanism in [15], −completeness and system error
was measured to add a rule, and the concept of error reduction
ratio (ERR) was utilized to prune a rule. Therefore, their
developed self-organizing FNN controller can adapt both the
parameters and the structure. Nevertheless, the computation of
larger matrix in each step, and storing of all previous inputoutput data makes their proposed technique computationally
costly. Sometimes, the spatial proximity between a particular
data-point and all other points, namely potential is used to
add or replace a rule in an evolving fuzzy controller as
described in [16]. However, the controller proposed in [16]
cannot prune any rules, which makes them computationally
costly in real-time control, especially in case of controlling
highly nonlinear over-actuated complex systems. Researchers
have also tried in [17] to employ a max-min method to add
and prune layers, and error backpropagation to tune evolving
neuro-fuzzy controller’s parameters like weight, center, and
width of Gaussian membership function. Nonetheless, the
controller needed some knowledge about the plant like the
bounded nominal matrix, which may not be known during
control operation.
Besides, an evolving TS fuzzy controller can be developed
by combining it with a fixed TS fuzzy FLS as exposed in
[18], where the evolving TS FLS worked online, and no
previous information was needed. Though such design of an
evolving controller is simple, it requires some information
about the plant due to the utilization of static TS fuzzy system.
Researchers have tried to develop a cloud-based evolving
fuzzy controller in [19]. The advantage of their cloud-based
evolving fuzzy controller was the parameter-free antecedent
part of a rule, whereas the consequent part was expressed
in the form of a PID controller. The three parameters of
their PID like evolving controller were tuned online through
Lyapunov theory based stable adaptation law. The controller
can add rules or clouds by measuring local data density. Their
evolving structure helped them to track the desired trajectory
properly. However, they were not able to delete rules, or
clouds. An evolving fuzzy controller is possible to develop by
using Fuzzy granular computation in explained [20], which can
model and control an unknown non-stationary system in online
without any experts knowledge. A fuzzy Lyapunov function
was utilized to evaluate the closed-loop stability of their
controller. Data uncertainty was handled and incorporated into
knowledge domain by Fuzzy granular computation. However,
diversity in data among different granules, and unmeasurable
state variables was not considered.
Furthermore, an evolving neuro-fuzzy controller can also
be developed by utilizing model predictive control methods
as described in [21], [22], [23]. A radial basis function neural
network based self-organizing controller was developed in [21]
where the evolving neuro-fuzzy system was utilized to bound
the modeling error uniformly for nonlinear systems. Besides,
a fast gradient method was employed to minimize the computational cost of their online evolving model predictive control
method. The Lyapunov theory demonstrated the closed-loop
stability of their control method. However, their controller was
not implemented in real-time application.
2
Until now, all the discussed evolving controllers have utilized univariate Gaussian function, which does not expose the
scale-invariant property. Besides, they are not effective in dealing with non-axis parallel data distribution. To mitigate these
shortcomings, we utilized a multivariate Gaussian functions
based incremental learning machine algorithm called generic
evolving neuro-fuzzy inference system (GENEFIS) [24]. In
this work, GENEFIS is amalgamated with SMC technique
to develop a self-evolving controller namely G-controller. In
this work, main features of the proposed G-controller are as
follows:
1) The controller’s performance does not depends upon the
plant dynamics or any other features of the plant.
2) No previous information or off-line training is required.
Thus, the controller starts self-construction from scratch
with only one rule at the beginning, and then it adds,
deletes, or merge rules to follow the desired trajectory.
Besides, the application of a fast kernel-based metric
approach helps to capture fuzzy set and rule level
redundancy.
3) Integration of the Generalized Adaptive Resonance Theory+ (GART+) helps to upgrade the premise parameters
with respect to input data distributions, and utilization
of multivariate Gaussian function aids the controller to
handle a variety of data generated from the sudden
change in plants, or from uncertainties, environmental
perturbations.
4) Adaptation law for the GENEFIS based G-controller’s
consequent parameters are derived from the SMC learning theory, which confirms a stable closed-loop control
system. A robustifying auxiliary control term ensures
uniform asymptotic convergence of tracking error to
zero. Finally, the stability of the G-controller is proved
using the Lyapunov function.
5) Successful evaluation of the proposed G-controller
through implementing it into the simulated BIFW MAV,
and hexacopter plant.
The above-mentioned characteristics of the proposed selfevolving G-controller make them an appropriate candidate to
control autonomous vehicles like BIFW MAV, quadcopter,
hexacopter, octocopter with better accuracy than a standalone first principle based controller. Furthermore, the whole
controller is developed using C programming language to
make it compatible with all types of hardware, where their
implementation is made publicly available in [25]. In the next
subsection I-A, the recent implementation of FLS, NN, FNN
based controller in autonomous vehicles are summarized.
A. Related Work
Autonomous vehicles express a complex mathematical
model, where the incorporation of various uncertainties is
very difficult, or not possible in some cases. The situation
becomes worse, when those vehicles are highly nonlinear,
under-actuated, or over-actuated like bio-inspired flapping
wing micro air vehicle (BIFW MAV), quadcopter, hexacopter,
octocopter unmanned aerial vehicles [26]. It is due to the
associated uncertainties, and rarely available information or
IEEE TRANSACTIONS ON FUZZY SYSTEMS
no information at all about the autonomous vehicles. As a
consequence, in recent time research interest is increasing
in employing model-free adaptive and self-evolving fuzzy
and neuro-fuzzy controller to autonomous vehicles. A selforganizing adaptive fuzzy controller was developed in [27]
to control a complex surface vehicle with uncertainties, and
external perturbation. A robustifying auxiliary control term
contributed to ensure closed-loop system stability of their
controller. From simulation studies, better performance was
observed from their fixed adaptive control techniques. However, a higher number of rules were generated to achieve
desired tracking accuracy. In [28], an adaptive fuzzy-PID
controller was developed to control the attitude and altitude of
a quadrotor. Though the TS-fuzzy system can approximate the
plant’s dynamics with uncertainties and unknown disturbances,
their dependency of a static PID controller affected their
performance to cope with sudden changes in plant dynamics.
An evolving fuzzy reference controller was proposed in
[29], which can partition the input-output space. It was able
to identify the local controllers gain. The stability of their
controller is yet to prove. In [30], self-tuning of a twodegrees-of-freedom control algorithm was proposed based on
an evolving fuzzy model. Their control algorithm consisted
of a feedforward, and feedback loop. The feedforward part
was derived from inversion of TS fuzzy model, which helped
the system to follow the trajectory closely. An evolving fuzzy
system with the ability to add, prune, merge, or split the
clusters was utilized in the feedback loop of their controller.
Nonetheless, their controller was applicable for single-inputsingle-output (SISO) systems only. In [31], a Taylor Series
Neuro Fuzzy (TaSe-NF) model based online self-evolving
neuro fuzzy controller is developed. The local online learning
of consequent parameters provided them a proper control
mechanism, and the self-evolving structure provided satisfactory performance with minimum rules and no prior information about the plant. However, their controller’s stability yet
to be proved.
A direct evolving neural controller was developed in [32] by
implementing a self-constructing radial basis function neural
network (RBFNN), which can add or prune neurons online
to adapt to plant condition. The closed-loop stability of their
controller was guaranteed by Lyapunov stability theory. A
self-organizing fuzzy controller is proposed in [33], where
the consequents are adapted at each selected topology, and
membership functions were added online. After analysing
the complete operating region of the plant the topology was
modified, which raised their controllers robustness. However,
their controller had a dependency of plants sampling period,
which made them impractical in implementing real embedded
UAV plants.
An interval type-2 fuzzy neural network based gain adaptive
sliding mode controller was developed in [34]to control the
attitude micro aerial vehicle. The sliding mode gain adaptive law was utilized to tune the parameters of the interval
type-2 fuzzy neural network online. Better tracking accuracy
was observed than the conventional adaptive sliding mode
controller. However, the fuzzy controller can only tune its
parameters, and unable to evolve their structure by adding or
3
deleting rules. A self-organizing interval type-2 fuzzy neural
network (SOIT2FNN) based controller was developed in [35],
and combined with a PD controller to control the attitude
of a quadcopter Micro Aerial Vehicle (MAV), where the
SOIT2FNN was able to learn the inverse model of quadcopter
online. Besides, it could minimize the model errors and
external disturbances. Nonetheless, due to their dependency
on static PD controller, they were affected by sudden changes
in plant dynamics.
To deal with the uncertainties exists in the nonlinear
complex FW MAV model, a radial basis function neural
network (RBFNN) was utilized in [36]. The stability of their
closed loop control system was verified through Lyapunov
stability theorem. By selecting appropriate control variables,
a good trajectory tracking performance was observed from
their simulation results. However, their control technique was
not self-evolving in nature. A fuzzy controller namely hybrid
adaptive fuzzy controller (HAFC) was developed in [14] to
control a dragonfly-like FW MAV. Their HAFC was able
to tune its parameters by using the hybrid adaptive law to
minimize the tracking error. The simulation results of the
HAFC were compared with a direct adaptive (DA) method
based fuzzy controller (DAFC). Better tracking performance
was observed from the HAFC than the DAFC. However,
their controller was constructed by the batch learning process, therefore static in nature. A Spiking Neural Network
(SNN) based controller was developed for an FW MAV
called RoboBee in [37], where they had utilized a rewardmodulated Hebbian plasticity mechanism to adopt a leaky
integrate-and-fire spiking neural network in flight. In [38],
an online self-organizing neural controller was combined with
a conventional controller to track the desired trajectory of a
simulated helicopter plant, while performing highly nonlinear
maneuvers. In their dynamic radial basis function network,
a Lyapunov based adaptation law was integrated with the
neuron growing and pruning mechanism, which ensured the
closed-loop stability and desired tracking accuracy. A selfconstructing neuro-fuzzy controller was proposed in [39] to
track the desired attitude of a quadcopter, which can add, and
delete rules online. Integration of a projection based adaptation
law handled the drift, and singularity problem in antecedent
parameters. However, due to their rule growing mechanism
comparatively higher number of rules were generated online,
which made them computationally costly.
To adapt the consequent parameters of various self-evolving
neuro-fuzzy controllers, sometimes gradient-based algorithms
are utilized. These gradient-based algorithms perform better
only with those plants, where a slow variation in their dynamics is observed. Besides, some gradient-based algorithms
namely dynamic backpropagation includes partial derivatives,
which may slow down the convergence speed, especially in
case of complex search space. In addition, the tuning procedure may be trapped into a local minimum in backpropagation
methods as explained in [40]. As a solution to these problems,
evolutionary approaches are proposed in [41]. However, the
stability of such controller is questionable. To overcome these
issues, sliding mode control (SMC) theory based algorithms
were proposed to tune the consequent parameter of various
IEEE TRANSACTIONS ON FUZZY SYSTEMS
FLS, NN, FNNs in [42], [43], [44], [45]. An SMC theory
based adaptive and online neuro-fuzzy-PID controller was
developed in [46], which exhibited an improved performance
than standalone PID controller. Since the adaptation laws
are derived from the PID controller part, without the PID
controller, the neuro-fuzzy controller was unable to tune its
parameters.
To mitigate the shortcomings of the existing self-evolving
neuro-fuzzy controller, in this work a novel self-evolving
controller namely G-controller is developed in guiding various
autonomous vehicles. The evolving architecture of the controller is inspired by the incremental learning algorithm called
GENEFIS [24], and the consequent parameters are adapted
utilizing SMC theory based learning algorithm. Unlike the
evolving neuro-fuzzy controllers discussed in the literature,
the integration of GART+ with a four-stage checking to evolve
rules provide a very quick response, and reduce the computational complexity by pruning unnecessary rules. Instead of
predefined values, the sliding parameters in the SMC theory is
also self-organizing in this work. To the best of our knowledge,
such evolving sliding parameters are never utilized before in
any of the existing evolving neuro-fuzzy controllers.
The organization of rest of this paper is as follows. Section
II describes the problem associated with two autonomous
vehicles namely BIFW MAV, and hexacopter in tracking the
trajectory accurately. The self-evolving architecture of the
GENEFIS based G-controller is explained in Section III. Section IV represents SMC learning algorithm based adaptation
of the proposed G-controller. The results are summarized,
and analysed in Section V. Finally, the paper ends with the
concluding remarks encompassed in Section VI.
II. P ROBLEM S TATEMENT
Modelling and control of BIFW MAV is one of the latest
research topics related to the field of autonomous Unmanned
Aerial Vehicles (UAVs). BIFW MAV exhibits some advanced
characteristics like fast flight, vertical take-off and landing,
hovering, and quick turn, and enhanced manoeuvrability when
compared to similar sized fixed and rotary wing UAVs. To
observe these features from a BIFW MAV, an advanced
control mechanism is a must need. To take the challenge,
our proposed G-controller is implemented in a bio-inspired
(BI) dragonfly-like four wings simulated MAV plant developed
by the Defence Science and Technology Group (DSTG) as
described in [47]. The simulated BIFW MAV plant saves the
time and expenses to set-up for experimental flight test. The
high level architecture of the BIFW MAV flight simulator is
exposed in Fig. 1.
In the flight simulator, four actuators are denoted by ’Actuator 1’, ’Actuator 2’, ’Actuator 3’, and ’Actuator 4’ as shown in
Fig. 1. These actuators are utilized to control four wings of the
BIFW MAV. Each actuator is controlled by eight (8) flapping
parameters indicated by F1 to F8 in ’flapping parameters’
block, where F1 indicates the stroke plane angle (in rad), F2
is the flapping frequency (in Hz), F3 represents the flapping
4
Table I
E FFECTS OF FLAPPING PARAMETERS IN DIFFERENT MANOEUVRING OF
FW MAV(φ0 : F LAPPING AMPLITUDE ( DEGREE ); Ψ: P HASE ( DEGREE ))
Actuators with corresponding flapping parameter
Actuator: 1, 2, 3, 4; φ0 : 90
Actuator: 1, 2; φ0 : 90 and Actuator: 3, 4; φ0 : 60
Actuator: 1, 2; φ0 : 60 and Actuator: 3, 4; φ0 : 90
Actuator: 1, 3; φ0 : 90 and Actuator: 2, 4; φ0 : 60
Actuator: 1, 3; φ0 : 60 and Actuator: 2, 4; φ0 : 90
Actuator: 1, 4; Ψ: 90 and Actuator: 2, 3; Ψ: 60
Actuator: 1, 4; Ψ: 60 and Actuator: 2, 3; Ψ: 90
Action
Take-off
Roll-right
Roll-left
Pitch-up
Pitch-down
Yaw-right
Yaw-left
amplitude (in rad), F4 presents the mean angle of attack (in
rad), F5 is the amplitude of pitching oscillation (in rad), F6
is the phase difference between the pitching and plunging
motion, F7 is the time step (in sec), F8 is the kappa set as
zero in the plant. By altering these flapping parameters, from
this flight simulator various mission profile or manoeuvrability
such as take-off, rolling, pitching, and yawing of an BIFW
MAV can be analysed. To find the dominant flapping parameter, a parametric analysis is accomplished in this work. It is
found that the flapping amplitude is the most dominant among
the eight parameters to control the BIFW MAV. By tuning the
flapping amplitude the take-off, rolling, and pitching motion
is observed. Only for the yawing motion flapping phase needs
to be tuned. Tuning of these parameters and their effect on
various manoeuvring is summarized in TABLE I.
The force generated by each actuator (Fri ) is supplied to
the ’Forces & Moments’ block of the Fig. 1. Besides the
four forces from four actuators (Fr1 , Fr2 , Fr3 , Fr4 ), the
mass and Direction Cosine Matrix (DCM) is also fed to the
’Forces & Moments’ block. This block supplies the required
force (F) and moment (M) to the rigid body dynamics of the
simulator based on the relative airflow acting on each wing
and the commanded actuator speed. The wing kinematics were
modelled utilizing the derivation explained in [48], [49], [50],
where it was considered that the wing was flapping in an
inclined stroke plane with a certain angle. In the simulator,
to express the MAV’s flapping profile the flapping angle (φ)
is presented in a sinusoidal form as follows:
φ(t) = φa cos(πf t)
(1)
where φa is the flapping amplitude in radian, f is the flapping
frequency in Hz, t is the time is second. Besides, the angle of
attack (Aaoa ) can be expressed as:
Aaoa = Amn − Ap sin(ωdt + Ψ)
(2)
where Amn is mean angle of attack in radian, Ap is amplitude
of pitching oscillation in radian, dt is time step in seconds, and
Ψ is the phase difference between the pitching and plunging
motion. All the four wings of the BIFW MAV follows the
same flapping profile.
Now in the ’Forces & Moments’ block M is actually a
summation of four momentums of four different actuators,
which can be expressed as:
M = M1 + M2 + M3 + M4
(3)
IEEE TRANSACTIONS ON FUZZY SYSTEMS
5
Flapping amplitude
F1 to F8 for
Actuator 1
Command to
Actuator 1
F1 to F8 for
Actuator 1
Command to
Actuator 2
F1 to F8 for
Actuator 1
Command to
Actuator 3
Command to
F1 to F8 for
Actuator 4
Actuator 1
Flapping parameter
to set various
mission profile
Force
Actuator 1
F
Force
Actuator 2
Rotational &
Translational
displacement
, velocity and
acceleration
Force
Ref
Body States
Zref
Zb
Altitude
Control
G-controller
Error
Actuator 3
M
Force
DCM
Actuator 4
Rigid body
dynamics
Forces & Moments
Mass
Figure 1. High level architecture of the over-actuated FWMAV plant
Each of these moments (M1 , M2 , M3 , M4 ) is actually
the torque induced in each wing based on the relative airflow
acting on each of them. The individual momentum of each
wing can be expressed as:
M1 = Fr1 × (CG − CP1 )
(4)
M2 = Fr2 × (CG − CP2 )
(5)
M3 = Fr3 × (CG − CP3 )
(6)
M4 = Fr4 × (CG − CP4 )
(7)
where, CG = [0 0 0]; and CP1 = [0.08 0.05 0];
CP2 = [0.08 0.05 0]; CP3 = [0.08 − 0.05 0]; CP4 =
[−0.08 − 0.05 0]; and × is presenting (3 × 3) cross product.
Each of the forces (Fri ) are three dimensional where the
elements are [Fxr Fyr Fzr ]. On the other hand, F is actually
a summation of four forces of four different actuators along
with the gravitational force, which can be expressed as:
F = Fr1 + Fr2 + Fr3 + Fr4 + (mg × DCM )
(8)
where, m is the mass, g is the acceleration due to gravity.
Finally, in the ’Rigid Body Dynamics’ block the forces and
moments are converted into the body coordinate system,
and all the required body states like three dimentional
rotational displacements (roll(φ), pitch(θ), yaw(ψ)),
velocities (ωbx , ωby , ωbz ), and accelerations (αbx
=
dωby
dωbz
dωbx
,
α
=
,
α
=
)
and
translational
by
bz
dt
dt
dt
displacements (Xb , Yb , Zb ), velocities (vbx , vby , vbz ),
dv
and accelerations (abx = dvdtbx , aby = dtby , abz = dvdtbz ) are
obtained, and the MAV states are updated. The complete
mathematical model of the BIFW MAV dynamics is written
in C code and MATLAB m file, where the C code is
called using Simulink S-function. The complete BIFW MAV
simulator with the necessary files describing the dynamics is
made publicly available in [25]. In addition, the development
of hardware construction of the above described BIFW MAV
plant is going on at this moment by the DSTG group. Their
developed flapping wing-actuation system is shown in the
supplementary document. After successful completion of the
construction, our proposed G-controller will be implemented
in hardware, since the G-controller is compatible with their
hardware.
Besides the BIFW MAV, another autonomous vehicle
namely hexacopter, a rotary wing UAV is considered in this
work. The simulated hexacopter plant is developed by UAV
laboratory of the UNSW at the Australian Defence Force
Academy. The model is of medium fidelity and contains
both full 6 degrees of freedom (DOF) rigid body dynamics
and non-linear aerodynamics. The hexacopter simulated plant
introduces two extra degrees of freedom which are obtained
by shifting two masses using two aircraft servos with each
mass sliding along its own rail aligned in longitudinal and
lateral directions respectively, which makes the plant an overactuated system. The top-level diagram of this over-actuated
simulated plant is exhibited in Fig. 2.
The ’attitude controller’ block consists of inner loop and
outer loop controllers. In outer loop, our proposed G-controller
is utilized and PID is used in inner loop. The altitude or
height, the moving-mass based shifting of center of gravity
in X and Y axis (CGX , and CGY ) is controlled by the
G-controller. Therefore, G-controllers are utilized in ’outer
loop controller’, ’Height controller’, ’CGX controller’, ’CGY
controller’ blocks in Fig. 2. The ’control mixing’ block converts the attitude (roll and pitch), yaw, and thrust commands
coming from the controller to motor speed commands. This
is done using a simple linear mixing arrangement based on
the relevant positions of each rotor. The ’forces and moments’
block calculates the thrust and torque of each rotor based on
the relative airflow acting on each rotor and the commanded
motor speed. The thrusts and torque of each rotor are then
added together to give the total vertical force (F z) and yawing
torque (N ) of the hexacopter. The thrust of each rotor is
multiplied by the appropriate moment arms to calculate the
rolling (L) and pitching moments (M ) acting on the airframe.
The thrust, yawing torque and rolling and pitching moments
are then fed to the rigid body dynamics block so that the
hexacopter state can be updated. The above explained plant is
also being converted into hardware at the UAV laboratory of
UNSW@ADFA. The airframe components and basic structure
of the hexacopter is exhibited in Fig. 3. After the accomplishment of the construction, our proposed controller will be
instrumented in that hardware.
Deriving a proper mathematical model of such highly
nonlinear, complex, and over-actuated micro air vehicles are
IEEE TRANSACTIONS ON FUZZY SYSTEMS
Ref
6
Vx
Error
Ref
Vy
Roll
Cmd
Outer loop
Controller
Pitch
Cmd
Error
Roll
W1
W1
W2
W2
W3
W3
W4
W4
W5
W5
W6
W6
Error
Inner loop
Controller
Error
Pitch
Attitude Controller
Ref
Error
Yaw
controller
Yaw
F&M
Rigid Body
Dynamics
Yaw
Ref
Z
Error
Height
controller
Thrust
Body
States
Control
Mixing
Ref
Vx
Error
Ref
Error
Vy
CG_X
controller
CG_X
CG_Y
controller
CG_Y
Forces &
Moments
Figure 2. Top-level diagram of the over-actuated simulated Hexacopter plant
where Ri represents the i − th rule (membership function)
constructed from a concatenation of fuzzy sets and epitomizing
a multidimensional kernel, k denotes the dimension of input
feature, Z is an input vector of interest, ai is the consequent
parameter, ζk is the k − th input feature. The predicted output
of the self-evolving model can be expressed as:
η̂ =
j
X
j
P
ψi (ζ)ηi (ζ) =
i=1
)LJXUH &RPSRQHQWV RI WKH $LUIUDPH
Pj
= Pi=1
j
T
exp(−(Z − Θi )Σ−1
i (Z − Θi ) )ηi
i=1
III. A RCHITECTURE OF THE S ELF -E VOLVING
G-C ONTROLLER
The self organizing mechanism of the G-controller is
adopted from GENEFIS developed in [24]. GENEFIS is a TS
FLS that features multidimensional membership functions in
the input space where the contours are ellipsoid in arbitrary positions. Each estimated one dimensional membership functions
represent a portion in the input space partition by assigning
the Gaussian function’s own center and width. Concurrently,
our GENEFIS based G-controller is generating 1st order firstorder polynomials as consequent parts of the fuzzy rules. In
G-controller, a typical fuzzy rule can be expressed as follows:
IF Z is Ri , then ηi = a0i + a1i ζ1 + a2i ζ2 + ... + aki ζk (9)
Ri
i=1
Figure 3. Airframe components and basic structure of the Hexacopter
developed at UAV Laboratory of UNSW@ADFA
extremely difficult, where insertion of uncertainties and unknown disturbance is more challenging, or impossible in some
cases. Thus, a controller that performs accurately with a
minimum or no knowledge about the system is much needed
in dealing with such autonomous vehicles. Being model-free
)LJXUH &RPSRQHQWV RI WKH $LUIUDPH
and self-evolving data-driven,
our developed G-controller is
an appropriate candidate. Besides, the computational cost of
our proposed controller is very low, where it maintains better
tracking performance. The self-evolving architecture of our
proposed G-controller is elaborated in the next section.
Ri η i
i=1
j
P
T
exp(−(Z − Θi )Σ−1
i (Z − Θi ) )
(10)
In Eq. 10, Θi is the centroid of the i − th fuzzy rule Θi ∈
<1×j , Σi is a non-diagonal covariance matrix Σi ∈ <k×k
whose diagonal components are expressing the spread of the
multivariate Gaussian function, and k is the number of fuzzy
rules.
A. Statistical Contribution Based Rule Growing Mechanism
The Datum Significance (DS) method developed in [51] is
utilized as a rule growing mechanism in G-controller. In our
work, the original DS method is geared into the multivariate
Gaussian membership function and polynomial consequents,
which is the crux of GENEFIS [24]. The integration of the
multivariate Gaussian membership function into the original
DS method can be observed in our work as follows:
(Z − Zn )Σ−1 (Z − Zn )T
= |ern |
exp −
(Z − Θ)Σ−1 (Z − Θ)T
Z
Z
1
dz
H(Z)
(11)
where Dsgn denotes the the significance of the n − th datum,
Zn is the current input to the controller in a closed-loop control
system Zn ∈ <, and H(Z) is the range of input Z. In a
closed-loop control system, error (e) indicates the difference
the desired reference and plant’s output, which is usually fed
as input to the controller. Whereas, error ern mentioned in Eq.
11 is difference from the input error (e). The error ern can be
expressed as:
Dsgn
IEEE TRANSACTIONS ON FUZZY SYSTEMS
7
B. Statistical Contribution Based Rule Pruning Mechanism
|ern | = |trn − ηn |
(12)
where trn is the target, which is plant’s output in our proposed
G-controller, and ηn is control output from the G-controller at
n − th episode. After applying k−fold numerical integration
to Eq. 11, the following is obtained:
Dsgn = |ern |
det(Σj+1 )
H(Z)
k
(13)
The Extended Rule significance (ERS) method was put
forward by [51]. The ERS concept appraises the statistical contribution of the fuzzy rules when the number of observations is
approaching infinity. The default ERS theory is not possible
to integrate directly into GENEFIS’s learning platform due
to the incompatibility of default ERS concept with the NN.
Numerous modifications are made in ERS theory to fit them
with GENEFIS based G-controller. In this work, the concept
of statistical contribution of the fuzzy rules can be expressed
mathematically as follows:
When the statistical contribution of the datum is higher
the existing rules it becomes an appropriate candidate to
be a new rule. Therefore, the DS criterion can be amended
mathematically as follows:
(14)
When a sample lies far away from the nearest rule, a high
value of Dsgn may obtain from Eq. 14 even with a small
value of ern . In such situation, generalisation capability of the
self-evolving neuro-fuzzy controller remains good without the
addition of any new rules. Therefore, a high value of Dsgn
does not always indicate the necessity of a rule evolution.
On the other hand, a high value of ern may obtain in case
of an overfitting phenomenon. In such case, the addition of
a new rule may worsen the overfitting phenomenon. Thus, a
separation is needed in Eq. 14 to cover two above mentioned
discernible situation.
To overcome an overfitting scenario, it is important to
monitor the effect of a newly injected sample on ern , since
the structural learning is not occurring in every observation. In
other words, the rule growing mechanism is probably turned
on when the rate of change of ern is positive. In this work, the
mean and variance of ern is measured by recursively updating
ern and standard deviation [52] as follows:
n−1
1
ērn−1 + ērn
n
k
(15)
1
n−1 2
σ̄rn−1 + (ērn − ērn−1 )
n
k
(16)
ērn =
2
σ̄rn
=
2
2
When ērn + σ̄rn
− (ērn−1 + σ̄rn−1
) > 0, the DS criterion is
simplified in our work as follows:
k+1
X
|ηi |
(18)
1
dz
H(Z)
(19)
i=1
Z
n T
exp(−(Z − Θni )Σ−1
i (Z − Θi ) )
Ei =
k
det(Σj+1 )
Dsgn = |ern | Pj+1
k
i=1 det(Σi )
E(i, n) = |δi |Ei , where |δi | =
Z
From Eq. 18, it can be anticipated that the contribution of the
fuzzy rules is a summary of the total contribution of input
and output parts of the fuzzy rules, where Ei is expressing
the modified version of original input contribution explained
in [51], [53], and δi is explicating the contribution of output
parameters. Usually, inverse covariance matrix Σ−1
in Eq.
i
19 has a smaller size than that of Z, which necessitates an
amendment in Eq. 19 as follows:
Ei ≈
1
H(Z)
Z
2
∞
exp −
0
Z2
det(Σi )
k
dz
(20)
By using the k−fold numerical integration, the final version
of ERS theory can be expressed as:
i
Einf
=
j+1
X
det(Σi )k
ηi Pj
k
i=1 det(Σi )
i=1
(21)
i
When Einf
≤ ke , it is presumed that the clusters cannot
capture the latest incoming data to the G-controller in the
closed-loop control cycle. It can be deduced that the hypervolume of the triggered cluster indicates the significance of
the fuzzy rule. Thus, when the volume of the ith cluster is
much lower than the summation of volumes of all cluster, that
rule is considered as inconsequential. Such a rule is pruned to
protect the rule base evolution from its adverse effect. In this
work, ke exhibits a plausible trade-off between compactness
and generalization of the rule base. The allocated value for δ
is δ = [0.0001, 1], and ke = 10% of δ.
C. Adaptation of Rule Premise Parameters
k
det(Σj+1 )
Dsgn = Pj+1
k
i=1 det(Σi )
(17)
The condition in expanding the rule base utilizing Eq. 17
is Dsgn ≥ g, where g is a predefined threshold. Eq. 17
represents an encouraging generalization and summarization
of the datum, since a new rule can omit possible overfitting
effects. Besides, this DS criterion can predict the probable
contribution of the datum during its lifetime.
Generalized Adaptive Resonance Theory+ (GART+) [54]
is used in G-controller as a technique of granulating input
features and adapting premise parameters. It is observed that
GART [55] and its successor improved GART (IGART) [56]
suffer from a category growing problem. In GRAT the compatibility measure is done utilizing the maximal membership
degree of a new datum to all available rules. In first round if the
selected category expresses a higher membership degree than a
predefined threshold ρa , then it is declared as winning category
IEEE TRANSACTIONS ON FUZZY SYSTEMS
8
and the match tracking mechanism is executed. However,
the first round winning category fails to beat the match
tracking threshold ρb , which deactivates that category and
increases the value of threshold ρa to find a better candidate.
A larger width is required in the next selected category to
cope with the increased value of ρa . Otherwise, it fabricates
a new category. Nonetheless, a category with larger radii may
contain more than one distinguishable data clouds and thereby
marginalizing the other clusters in every training episode.
In incremental learning environment this effect is known as
cluster delamination [57], pictorially exhibited in Fig. 4. To
relieve from the cluster delamination effect, in GENEFIS
based G-controller the size of fuzzy rule are constrained by
using GART+, which allows a limited grow or shrink of a
category.
1) Improved Selection procedure of winning Category: To
determine the most compatible or winning category Bayes’s
decision theory is utilized in GART+ [54]. To stimulate a more
appropriate selection of the winning category, the Bayesian
concept does not only consider the proximity of a category
to the inserted datum, but also the dominance of the category
with respect to the other categories through the category prior
probability. That is, the category prior probability can count
the number of samples falling in the outreach of the category,
and is expressed as follows:
Ni
P̂r (ψi ) = Pj
i=1
p̂r (Z|ψi ) =
where Vi determines the estimated hyper-volume of feature
space covered by the ith category, which can be expressed as:
Vi = det(Σi )
where Ni indicates the number of times that ith category or
fuzzy rule wins the competition.
When a new input datum to the controller finds two
categories with almost similar distance but with different
population number, Bayes’s decision theory assists to select
the category with more data points and declared it as a winning
category. In the proposed G-controller, posterior probability of
the ith category can be represented as follows:
p̂r (Z|ψi )P̂r (ψi )
P̂r (ψi |Z) = Pj
i=1 p̂r (Z|ψi )P̂r (ψi )
(23)
where p̂r (ψi |Z) and P̂r (ψi ) represents the likelihood and the
prior probability correspondingly. The likelihood can also be
elaborated as follows:
Case I: Rwin ≥ ρa , Vwin ≤ Vmax
Input feature Z 2
Input feature Z 2
Extraction of one cluster from two distinct clouds
Θnew
win =
old
Z − Θold
Nwin
win
Θold
win +
old
old + 1
Nwin + 1
Nwin
(27)
−1
−1
Σold
α
win
Σnew
=
+
win
1−α
1−α
T
−1
old−1
new
new
Σwin (Z − Θwin ) Σold
(Z
−
Θ
)
win
win
old
new
1 + α (Z − Θnew
win ) Σwin (Z − Θwin )
new
old
Nwin = Nwin + 1
T
(28)
(29)
where α can be expressed as follows:
α=
One
cluster
Input feature Z 1
(26)
where Rwin is the membership degree of the winning rule
to seize the latest datum. More importantly, the condition in
Eq. 26 is indicating the capability of the selected category
to accommodate the newest datum and emphasizing on the
limited size of a category. In our proposed G-controller ρa
is set close to 1. Contrarily, the value of ρb is set as [0.0001,
0.1]. Then the adaptation mechanism of focal point Θi , and the
dispersion matrix Σi is generated by the equations as follows:
−1
Data of cluster 1
Data of cluster 2 with dissimilar
characteristics
(25)
The Bayesian concept presented in Eq. 23 is implemented
in G-controller, which can be interpreted as follows:
1) When a new sample is adjacent to existing categories,
it causes a higher likelihood expressed by Eq. 24.
2) A category with a large volume is forced to divide its
volume in Eq. 24, as a consequent it delivers a lower
value of the posterior probability 23. This is important
particularly to avoid large span clusters and to decrease
the likelihood of cluster delamination effects.
3) According to 22, categories surrounded by more incoming data samples are more worthwhile, which inflicts a
high value of posterior probability.
2) Vigilance Test: There are two goals to perform the
vigilance test. The first goal concerns about the capability
of the winning category to accommodate a new datum. The
second goal is to reduce the size of the category, where a rule is
not allowed to have a volume higher
Pjthan the threshold Vmax ,
that is calculated from Vmax ≡ ρb i=1 Vi . The vigilance test
is a way to rule deletion, update, or evolution where it needs
to be satisfied four different conditions as presented below:
(22)
Ni
1
T
exp(−(Z − Θi )Σ−1
i (Z − Θi ) ) (24)
(2πVi )1/2
Input feature Z 1
Figure 4. Cluster delamination effect (adapted from [24] with proper
permission)
1
old + 1
Nwin
(30)
old
where Nwin
denotes the number of training samples populating the winning cluster.
Besides, a major advantage of utilizing Eq. 28 is the
prompter update of the dispersion matrix (inverse covariance
IEEE TRANSACTIONS ON FUZZY SYSTEMS
9
matrix), since a direct adjustment of the dispersion matrix is
occurring without the necessity to re-inverse the dispersion
matrix [58]. Concerning the conditions in 26, some pertinent
likelihoods may emerge in the rehearsal process and they are
outlined as follows:
Case II: Rwin < ρa , Vwin > Vmax
In this circumstance, the input data to the G-controller
cannot be touched by any existing rules of the controller, since
the inserted input data is hardly covered by any rules. The
statistical contribution of the datum needs to be calculated by
DS-criterion. When both conditions are satisfied, a new rule
is generated and its parameters are assigned as follows:
Θj+1 = Z
max((Θi − Θi−1 ), (Θi − Θi+1 ))
q
diag Σj+1 =
1
In()
(31)
(32)
Sssr (ug , u) = uARC (t) = ug (t) + u(t)
where the value of is 0.5. Equation 32 ensures a sufficient
coverage of the newly added rule, which is proved in [59]. It
helps GENEFIS to explore untouched regions in the feature
space fitting a superfluous cluster at whatever point a relatively
unexploited region or knowledge is fed, which is a mandatory
element to confronting possible non-stationary and evolving
qualities of the self-evolving control system. Note that proper
initialization of inverse covariance matrix plays a crucial role
to the success of multivariate Gaussian fuzzy rule. Although
it meets the −completeness criterion, Equation 32 requires
re-inversion phase which sometimes leads to instability when
the covariance matrix is not full-rank. As an alternative, the
inverse covariance matrix can be initialized as:
Σ−1
0 = kf s I
(33)
where kf s is a user-defined parameter, and I is an identity
matrix.
Case III: Rwin ≥ ρa , Vwin > Vmax
This situation is indicating the capability of existing rule
base to cover the current data easily. However, the width of the
chosen cluster is oversized. This datum creates a redundancy
when added to the rule base. To mitigate the adverse impact,
one of the solutions is to replace the selected cluster merely
by this datum. Then the fuzzy region is eased as follows:
Θwin = Z
Σnew
win
−1
(34)
1
−1
Σold
(35)
kwin win
where kwin is a constant with a value of 1.1, and the width of
the cluster is reduced until a desirable fuzzy region is obtained
while satisfying Vwin ≤ Vmax .
Case IV: Rwin < ρa , Vwin ≤ Vmax
The same action is taken as in Case I, i.e., the adjustment
process is executed to stimulate the category to move towards
the incoming input data.
=
and adapt premise parameters of the proposed G-controller,
where the integration of multivariate Gaussian function and
GART+ method helps the controller to reduce the structural
complexity and to adapt with sharp changes in autonomous
vehicle’s plant dynamics. On the other hand, being robust
enough to guarantee the robustness of a system against external
perturbations, parameter variations, unknown uncertainties, the
SMC theory is applied in our work to adapt the consequent
parameters of the G-controller. In SMC scheme, the motion
of a system is restricted to a plane known as sliding surface.
In this work, the SMC learning theory-based adaptation laws
are developed to establish a stable closed-loop system. By
following the regulations of SMC scheme as explained in [60],
[61], [62], the zero dynamics of the learning error coordinate
is defined as time-varying sliding surface as follows:
IV. SMC T HEORY-BASED A DAPTATION OF
G-C ONTROLLER
In our work, an advanced self-evolving neuro-fuzzy system
called GENEFIS is utilized to build the evolving structure
(36)
The sliding surface for the highly nonlinear over-actuated
autonomous vehicles namely FW MAV, and hexacopter plant
to be controlled is expressed as:
Z
sH = e + λ1 ė + λ2
t
e(τ )dτ
(37)
0
α3
α2
, λ2 = α
,e is the error which is the difference
where, λ1 = α
1
1
between the actual displacement from the plant and desired
position in case of altitude control. In this work, in case the
BIFW MAV plant, the sliding parameter α1 has initialized
with a small value 1 × 10−2 , whereas α2 has initialized with
1 × 10−3 , and α3 ≈ 0. Each of the parameters is then evolved
by using learning rates. These learning rates are set in such
a way so that the sliding parameters can achieve the desired
value in the shortest possible time to create a stable closedloop control system. A higher initial value of the sliding
parameters is avoided, since it may cause a big overshoot at
the beginning of the trajectory. It can be abstracted that, to
make our proposed G-controller absolutely model free, these
sliding parameters are self-organizing rather than predefined
constant values.
Def inition : After a certain time tk a sliding motion
will be developed on the sliding manifold Sssr (ug , u) =
uARC (t) = 0, where the state Sssr (t)Ṡssr (t) =
uARC (t)u̇ARC (t) < 0 to be satisfied for the whole time period
with some nontrival semi-open sub-interval of time expressed
as [t, tk ) ⊂ (0, tk ).
It is expected to produce such online adaptation of consequent parameters of the proposed G-controller that the sliding
mode condition of the aforestated definition is enforced. The
adaptation process of the proposed method is summarized
below.
Theorem 1. The adaptation laws for the consequent parameters of the G-controller are chosen as:
ω̇(t) = −α1 G(t)ψ(t)sH (t), where ω(0) = ω0 ∈ <nR×1
(38)
where the term G(t) can be updated recursively as follows:
IEEE TRANSACTIONS ON FUZZY SYSTEMS
ė
∫e Sliding
e surface
Zd
Error
calculation
Auxiliary
robustifying
uARC
control
Adaptive
law
e
10
∑
sum
e
ė Self-evolving GZ
u Autonomous
vehicles
plant
∇ω J(t) = 0
Z
Z t
⇒ ψ(τ )u(τ )dτ + ω(t)
ψ(τ )ψ T (τ )dτ = 0
Z
0
Z
ug
−1 Z
t
T
⇒ ω(t) =
t
ψ(τ )ψ (τ )dτ
ψ(τ )u(τ )dτ
0
0
Z t
ψ(τ )u(τ )dτ
⇒ ω(t) = −G(t)
0
Z t
ψ(τ )u(τ )dτ
⇒ G−1 (t)ω(t) = −
controller
Proposed self-evolving
control system
(44)
(45)
(46)
0
Figure 5. Self-evolving G-controller based closed-loop control system
where,
Z
G(t) =
T
Ġ(t) = −G(t)ψ(t)ψ (t)G(t), where G(0) = G0 ∈ <
uARC (t) = α1 sH
(40)
The robustifying auxiliary control term uARC may suffer
from high-frequency oscillations in the control input. It is
an undesirable phenomenon in sliding mode controller and
known as chattering effect. Due to simplicity, saturation or
sigmoid functions are mostly used to reduce the chattering
effect. In this work, a saturation function is utilized to mitigate
the adverse effect of chattering.
The G-controller’s final output signal can be expressed as
follows:
ug (t) = ψ T (t)ω(t)
(42)
The cost function can be defined as:
Z
−1
G
(47)
Z
(t) =
t
ψ(τ )ψ T (τ )dτ
(48)
The derivative of Eq. (48) is as follows:
G−1 (t)Ġ(t)G−1 (t) = −ψ(t)ψ T (t)
Ġ(t) = −G(t)ψ(t)ψ T (t)G(t)
(49)
From Eq. (49), it is observed that Ġ(t) is a negative definite
and G(t) is decreasing over time, therefore G(t) ∈ l∞ . Now
executing the time derivative of Eq. (45) and utilizing Eq. Fig.
6(a)(40), (41), (42), and (46) the following is obtained:
ω̇(t) = Ġ(t)G−1 (t)ω(t) − G(t)ψ(t)u(t)
= −G(t)ψ(t)ψ T (t)ψ(t) − G(t)ψ(t)u(t)
= −G(t)ψ(t) ψ T (t)ω(t) + u(t)
= −α1 G(t)ψ(t)sH (t)
(50)
1) Stability Analysis: Def inition : FLS is known as a
general function approximator. Therefore, in this work it is
assumed that without loss of generality there exists a ω ∗ such
that:
u(t) = ψ T ω ∗ (t) + ε∗f (z)
(51)
where ε∗f (z) = [ε∗f 1 , ε∗f 1 , ..., ε∗f 1 ]T ∈ <k is the minimal
functional approximator error. In this work, the following is
defined:
ω̃(t) = ω(t) − ω ∗
(52)
sH (t) = ψ T ω
e (t)
(53)
In addition:
J(t) =
The gradient of J with respect to ω is as follows:
ψ(τ )ψ (τ )dτ
0
t
s2H (τ )dτ
0
Z t
1
= 2
(u(τ ) + ug (τ ))2 dτ
α1 0
Z t
1
= 2
(u(τ ) + ψ T (t)ω(τ ))2 dτ
α1 0
−1
(41)
The overall control signal as observed in Fig.(5) can be
obtained as follows:
u(t) = uARC (t) − ug (t)
T
0
nR×nR
(39)
where n is the number of inputs to the controller, and R is the
number of generated rules. These adaptation laws guarantee
a stable closed-loop control system, where the plants to be
controlled can be of various order.
P roof : The sliding parameter dependent robustifying auxiliary control term of the proposed controller can be expressed
as follows:
t
(43)
Lemma 1 :
d(G−1 (t)ω̃(t))
= −G−1 (t)Ġ(t)G−1 (t)ω̃(t) + G−1 (t)ω
e˙ (t)
dt
= ψ(t)ψ T (t)ω̃(t) − ψ(t)sH (t)
= ψ(t)sH (t) − ψ(t)sH (t)
=0
(54)
IEEE TRANSACTIONS ON FUZZY SYSTEMS
11
This is indicating that G−1 (t)ω̃(t) is not altering with
respect to time, and therefore G−1 (t)ω̃(t) = G−1 (0)ω̃(0),
∀t > 0.
limt→∞ ω̃(t) = limt→∞ G(t)G−1 (0)ω̃(0)
(55)
Since G(t) is decreasing andω̃(t) ∈ l∞ ,ω(t) ∈ l∞ . In this
work the following Lyapunov function is considered:
1 T
ω
e (t)G−1 (t)ω̃(t)
(56)
2
The time derivative of the Lyapunov function is as follows:
V (t) =
1 T
1 T
˙
ω
e (t)G−1 ω̃(t)
+ ω
e (t)Ġ−1 ω̃(t)
2
2
1 T
= −e
ω T (t)ψ(t)sH (t) − ω
e (t)ψ(t)ψ T (t)ω̃(t)
2
1
= −s2H (t) − s2H (t)
2
3
(57)
= − s2H (t) ≤ 0
2
From Eq. 56, and Eq. 57, it is observed that V (t) > 0, and
V̇ (t) ≤ 0. In addition, Eq. 57 shows that V̇ (t) = 0, if and
only if e(t) = 0. It is indicating that the global stability of the
system is guaranteed by the Lyapunov theorem. By utilizing
Barbalat’s lemma [63], it can also be observed that e(t) → 0
as t → ∞. It is ensuring the asymptotic stability of the system.
Thus, a convergence of the system’s tracking error to zero is
witnessed.
V̇ (t) =
V. R ESULTS AND D ISCUSSION
In this work, the self-evolving generic neuro-fuzzy controller namely G-controller is attempted to control of autonomous vehicles online. As mentioned in Section II that
the G-controller is utilized in controlling a BIFW MAV and a
hexacopter UAV. In case of the BIFW MAV, various altitude
trajectory tracking is observed to evaluate the controller’s performance, whereas, in hexacopter, not only the altitude but also
the attitude tracking is witnessed. Being an evolving controller,
the G-controller can evolve both the structure and parameters.
The observed evaluation procedure from the G-controller’s
performance is explained in the following subsection V-A.
A. Evaluation Procedure
The proposed G-controller has the capability of evolving
the structure by adding or pruning the rules like many other
evolving controllers discussed in the subsection I-A. However,
unlike the existing evolving controller, GRAT+, multivariate
Gaussian function, SMC learning theory based adaptation laws
are combined in the G-controller. From the amalgamation of
such advanced features, the faster self-evolving mechanism
is recorded with a lower computational cost. In controlling
the altitude and attitude of the highly nonlinear and complex
autonomous vehicles discussed in Section II, the activation
of only the rule growing mechanism was sufficient. Due
to the evolving nature of the G-controller, the fuzzy rules
are generated in different time steps for different reference
signals. This rule generation of the G-controller with respect
to various desired altitude of BIFW MAV are compiled in a
table provided in the supplementary document. To understand
graphically, the number of generated rules for various trajectories of BIFW MAV and Hexacopter are plotted and disclosed
in Fig. 12.
B. Results
The G-controller’s performance is observed with respect
to various reference signals, and the results are compared
with a TS fuzzy controller [64], and a Proportional Integral
Derivative (PID) controller. Our source codes are made publicly available in [25]. In case of the BIFW MAV, variety of
desired trajectories were utilized in the closed-loop control
system to evaluate controllers performance for 100 seconds,
such as: 1) a constant altitude of 10 meters expressed as
Zd (t) = 10; 2) three different step functions, where one of
them is varying its amplitude from 0 m to 10 m, another
is from 5 m to 10 m, and the other one is varying from 5 m to 5m, presented as Zd (t) = 10u(t − 20), Zd (t) =
5u(t)+5u(t−20), Zd (t) = −5u(t)+10u(t−20) respectively;
3) three different square wave function with a frequency of 0.1
Hz, and amplitude of 1 m, 4 m, and 10 m correspondingly;
4) two square wave function with a frequency of 1 Hz, and
amplitude of 1 m and 4 m; 5) a customized trajectory, where
the amplitude varies from 0 to 2 m; 6) a sawtooth wave
function with an amplitude of 1 m and frequency of 1 Hz; 7)
a sine wave function with an amplitude of 1 m and frequency
of 1 Hz. For all these trajectories, the performance of our
proposed G-controller, TS fuzzy controller, and PID controller
is observed and compared, where higher accuracy is obtained
from the G-controller. For clearer understanding, some of these
observations are presented pictorially from Fig. 6 to Fig. 9.
The performance of various controllers for a step function
Zd (t) = 5u(t) + 5u(t − 20) is observed in Fig. 6, where
our proposed G-controller outperformed the PID, and TS
fuzzy controller. The performance of the trajectory like square
wave pulse with an amplitude of 4 m and frequency of
1 Hz is observed in Fig. 7, where comparatively improved
performance is witnessed from our developed G-controller. In
case of this trajectory the TS fuzzy controller fails to follow
the trajectory. Therefore, the comparison only between the
PID and G-controller is exposed in Fig. 7, where the Gcontroller has beaten the PID controller in terms of accuracy.
In case of all the trajectories for BIFW MAV exposed in
this paragraph, superior performance is visualized by our
proposed G-controller. To observe the controllers performance
deeply, the root mean square error (RMSE), rising time,
and settling time of all the controllers for various reference
signals are measured and summarized in TABLE II, where
the lowest RMSE is inspected from the G-controller. Since
the G-controller starts operating from scratch with an empty
fuzzy set, the rising time is comparatively higher than the
PID controller. However, comparatively lower settling time
is indicating the proposed controllers ability to back to the
desired trajectory sharply.
A wind gust is added to the BIFW MAV plant dynamics
to check the robustness of our proposed G-controller against
IEEE TRANSACTIONS ON FUZZY SYSTEMS
utilized before in any of the existing evolving neuro-fuzzy
controllers. It makes the proposed G-controller a fully selfevolving controller.
Overshoot
observed
in PID
controller
Flapping amplitude
10
8
Overshoot
observed
in TS fuzzy
controller
6
4
G-Controller
Reference
PID
TS-Fuzzy
Due to empty rule base
slower response at the
beginning in G-controller, but
recovered sharply and
outperformed others
2
0
0
20
40
60
80
100
time (sec)
Figure 6. Performance observation of various controllers in altitude tracking
of BIFW MAV when the trajectory is a step function Zd (t) = 5u(t)+5u(t−
20)
10
Flapping amplitude
unknown perturbations and uncertainties. This simulated wind
gust has a maximum velocity of 40 ms−1 and is applied to the
plant after 2 seconds. In the presence of the wind gust, some
of the trajectory tracking performances of the controllers are
manifested in Fig. 9, where an insignificant deterioration in
tracking is observed at the beginning. However, this adverse
effect has been minimized very sharply by the G-controller.
The RMSE by considering the effect of wind gust are also
tabulated in TABLE II.
Furthermore, the G-controller has been utilized to control
altitude, the outer loop of attitude (roll and pitch) of the
simulated over-actuated hexacopter plant developed in ADFA
and results are compared with a PID controller. In case of
controlling the altitude, the controllers are employed to control
the thrust of the control-mixing box of the plant. Due to the
addition of the moving mass, the rolling motion is not only
controlled by the velocity in Y-axis (vy ) generated by the
motors, but also the mass moving in the Y direction due to
their Center of Gravity (CG) shifting capability. Our proposed
controller has been employed in both facts to control the
rolling motion. Similarly, to control the pitching motion of
the hexacopter, the G-controller has been used to control both
the velocity in X-axis (vx ) and the mass moving in the X
direction. The altitude tracking performance of hexacopter for
various trajectories has been observed in Fig. 10, whereas
the tracking of rolling and pitching is exhibited in Fig. 11.
In all cases, better tracking has been monitored from the Gcontroller than that of PID controller. The RMSE, rise time,
and settling time has been calculated for all the trajectories
of the hexacopter and outlined in a TABLE provided in the
supplementary document, where lower RMSE is perceived
from the proposed G-controller. Besides, the settling time of
the G-controller is much lower than that of the PIDs, which
clearly indicates their improvement over the PID controller.
12
G-Controller
Reference
PID
5
0
-5
Higher overshoot observed in
PID controller more frequently
-10
0
20
40
60
80
100
time (sec)
Figure 7. Performance observation of various controllers in altitude tracking
of BIFW MAV when the trajectory is a square wave function with an
amplitude of 4 m and frequency 1 Hz
C. Discussion
4
Flapping amplitude
Unlike the PID and TS fuzzy controller, the G-controller
starts the self-construction online with an empty fuzzy set at
the beginning of the closed-loop control system. Whereas,
the PID and TS fuzzy controller start operating with their
pre-set control parameters. In case of the PID controller,
control parameters (proportional gain Kp , integral gain Ki ,
and differential gain KD ) are obtained offline before starting
the closed-loop control operation. The TS fuzzy controller
consists of five rules, where univariate Gaussian membership
functions are utilized in each rule. To obtain the antecedents
and consequent parameters of the rules, the fuzzy controller is
trained with the PID controller’s input and output datasets. It is
clear that in both PID and TS fuzzy controllers the parameters
are fixed before the starting of the closed-loop control system.
On the contrary, in G-controller not only the GENEFIS but
also the parameters of the sliding surface are evolving. Those
sliding parameters are initialized with a very small value,
then evolved to a desired value by using different learning
rates. These rates are varied with respect to the corresponding
plants, and desired actions. To the best of our knowledge,
this approach of evolving the sliding parameters is never
G-Controller
Reference
PID
TS-Fuzzy
3
Higher overshoot
in PID controller
2
1
TS fuzzy controller is unable
to follow the trajectory
between 2 to 7 seconds
0
0
2
4
6
8
10
time (sec)
Figure 8. Performance observation of various controllers in BIFW MAV in
case of tracking a customized trajectory
IEEE TRANSACTIONS ON FUZZY SYSTEMS
13
Table II
M EASURED RMSE, RISE AND SETTLING TIME OF VARIOUS CONTROLLERS IN OPERATING THE BIFW MAV
Desired Trajectory
(Zd )
Maximum
amplitude
(meter)
Constant height
10
Zd (t) =
5u(t) +
5u(t − 20)
Zd (t) =
10u(t − 20)
Step function
Zd (t) =
−5u(t) +
10u(t − 20)
1
Square wave
function
(f = 0.1 Hz)
4
10
1
Square wave
function(f =
1 Hz)
4
Customized wave
function
2
Sawtooth wave
function
1
Sine wave function
1
Measured
features
PID
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
RMSE
Rise time (sec)
Settling time (sec)
0.6630
0.9040
8.2595
0.4000
21.242
28.1
0.6266
21.03
29.50
0.6695
21.035
29.5
0.2039
0.38
10.11
0.8909
0.45
6.75
2.7381
23.30
42.5
0.6719
3.43
22.6
3.1435
3.752
12.4
0.2856
2.0
5.5
0.5235
0.24
4.1
0.2096
0.824
4.1
Without gust
TSGFuzzy
controller
0.5693
0.6460
0.7442
1.2583
4.9305
3.9525
0.4023
0.3866
21.094
21.37
23.01
21.50
0.6701
0.5677
20.91
20.746
22.05
22.07
0.7188
0.6535
20.95
21.53
22.5
22.1
0.2493
0.1746
0.48
0.83
3.8
2.75
0.9059
0.8337
0.524
0.914
5.6
2.5
511.9071
2.5406
N/A
23.37
N/A
34.5
362.9960
0.5893
N/A
3.72
N/A
8.5
573.3459
2.7275
N/A
4.06
N/A
7.81
4.7793
0.1033
1.30
1.98
7.25
2.8
325.4397
0.4781
N/A
0.96
N/A
1.5
0.0737
0.0356
1.19
1.01
3.4
2.7
With gust
TSFuzzy
0.5708
0.7442
3.6
0.4002
21.05
23.6
0.6616
20.81
24.99
0.7109
20.81
24.59
0.2552
63.43
69.50
0.9025
0.52
5.14
1629.4883
N/A
N/A
363.3044
N/A
N/A
1258.0926
N/A
N/A
4.7708
0.93
7.25
325.4111
N/A
N/A
0.0764
1.19
3.32
PID
0.6635
0.9040
10.1
0.3962
21.23
27.5
0.6216
21.02
32.02
0.6648
21.02
31.56
0.2020
63.18
69.52
0.8892
0.49
8.92
2.7339
0.97
8.37
0.6720
3.435
18.1
3.1336
4.13
6.58
0.2846
2.16
7.66
0.5240
0.9039
5.68
0.1880
0.7405
5.23
Gcontroller
0.6461
0.7984
2.6
0.3865
21.12
21.80
0.5611
20.75
21.66
0.6477
20.79
21.03
0.1710
63.45
67.05
0.8311
0.91
2.15
2.5376
1.26
2.25
0.5735
4.17
4.4
2.6766
4.62
4.74
0.1026
1.96
2.75
0.4776
0.6499
0.65
0.0395
1.08
2.11
2.5
G-Controller
Reference
PID
TS-Fuzzy
Effect of
wind gust
1
0
2
Altitude (m)
Flapping amplitude
2
-1
Higher overshoot
observed in sharp
edges from TS
fuzzy and PID
controller
-2
-3
-4
1.5
20
40
60
80
G-Controller
Reference
PID
0.5
0
0
Comparatively slower response
observed in PID controller
1
100
0
5
10
15
20
25
30
35
40
45
50
time (sec)
time (sec)
Figure 9. Performance observation of various controllers in altitude tracking
of BIFW MAV by considering wind gust as environmental uncertainty
VI. C ONCLUSIONS
The G-controller developed in this work is an entirely
model-free approach and self-evolving in nature, i.e. it can
alter its structure, and system parameters online to cope with
changing dynamics of the plant to be controlled. Besides, the
synthesis of SMC theory based adaptation laws improve its
robustness against various internal and external uncertainties.
These desirable features make the G-controller a suitable
candidate for highly nonlinear autonomous vehicles. In this
Figure 10. Performance observation of a PID and proposed G-controller in
tracking various altitude of a hexacopter
work, our proposed control algorithm is developed using C
programming language considering the compatibility issues to
implement directly in hardware of a variety of autonomous
vehicles like BIFW MAV, quadcopter, hexacopter, etc. The
controller’s performance has been evaluated by observing the
tracking performance of an over-actuated BIFW MAV and an
over-actuated hexacopter’s plant with respect to a variety of
trajectories. The performances are compared to that of a PID,
and a TS fuzzy controller to observe the improvements in our
proposed online evolving G-controller. The G-controller starts
IEEE TRANSACTIONS ON FUZZY SYSTEMS
14
0.6
Pitch (rad)
0.4
aerial vehicles.
G-Controller
Reference
PID
Comparatively higher overshoot
observed in PID controller
ACKNOWLEDGEMENT
0.2
0
-0.2
0
5
10
15
20
25
30
35
40
45
50
time (sec)
(a)
0.4
Roll (rad)
0.3
0.2
In G-controller the response is quicker than
PID controller with better accuracy
R EFERENCES
0.1
G-Controller
Reference
PID
0
-0.1
0
5
10
15
20
25
30
35
40
45
50
time (sec)
(b)
3
6
2.5
5
Number of rules
Number of rules
Figure 11. Performance observation of a PID and proposed G-controller in
tracking desired (a) pitching and (b) rolling motion of a hexacopter
2
1.5
1
4
3
2
1
0.5
0
0
0
2
4
6
8
10
0
0.2
0.4
0.6
0.8
1
0.8
1
time (sec)
time (sec)
(a)
(b)
3
4
Number of rules
Number of rules
2.5
2
1.5
1
3
2
1
0.5
0
0
The authors would like to thank the Australian Defense
Science and Technology Group for providing the simulated
BIFW MAV plant, Unmanned Aerial Vehicle laboratory of
the UNSW at the Australian Defense Force Academy for
supporting with the hexacopter plants, and the computational
support from the Computational Intelligence Laboratory of
Nanyang Technological University (NTU) Singapore. This
work is fully supported by NTU start-up grant and MOE tier-1
grant.
2
4
6
8
10
0
0
0.2
0.4
0.6
time (sec)
time (sec)
(c)
(d)
Figure 12. Generated rules of the self-evolving G-controller at various
trajectories of BIFW MAV and hexacopter where the trajectories are (a) step
function altitude for BIFW MAV, (b) customized altitude for BIFW MAV, (c)
pitching for Hexacopter, (d) rolling for Hexacopter
building the structure from scratch with an empty fuzzy set in
the closed-loop system. It causes a slow response at the very
beginning of the loop, which is a common phenomenon in
any self-evolving controller. However, due to the integration of
GRAT+, multivariate Gaussian function, SMC learning theory
based adaptation laws, the self-evolving mechanism of the
G-controller is faster with a lower computational cost. In
addition, wind gust has been added to the BIFW MAV plant
as environmental uncertainties to evaluate the G-controller’s
robustness against unknown perturbations, where satisfactory
tracking of the desired trajectory proves the proposed controller performance to eliminate various uncertainties. Thus,
the G-controller’s stability is confirmed by both the Lyapunov
theory and experiments. In future, the controller will be executed through hardware-based flight test of various unmanned
[1] H. Gomi and M. Kawato, “Neural network control for a closed-loop
system using feedback-error-learning,” Neural Networks, vol. 6, no. 7,
pp. 933–946, 1993.
[2] Y.-J. Liu and S. Tong, “Adaptive fuzzy identification and control for
a class of nonlinear pure-feedback mimo systems with unknown dead
zones,” IEEE Transactions on Fuzzy Systems, vol. 23, no. 5, pp. 1387–
1398, 2015.
[3] J. R. Noriega and H. Wang, “A direct adaptive neural-network control
for unknown nonlinear systems and its application,” IEEE Transactions
on Neural Networks, vol. 9, no. 1, pp. 27–34, 1998.
[4] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its
applications to modeling and control,” in Readings in Fuzzy Sets for
Intelligent Systems. Elsevier, 1993, pp. 387–403.
[5] R.-J. Wai and P.-C. Chen, “Intelligent tracking control for robot manipulator including actuator dynamics via tsk-type fuzzy neural network,”
IEEE Transactions on Fuzzy Systems, vol. 12, no. 4, pp. 552–560, 2004.
[6] S. L. Perić, D. S. Antić, M. B. Milovanović, D. B. Mitić, M. T. Milojković, and S. S. Nikolić, “Quasi-sliding mode control with orthogonal
endocrine neural network-based estimator applied in anti-lock braking
system,” IEEE/ASME Transactions on Mechatronics, vol. 21, no. 2, pp.
754–764, 2016.
[7] C.-J. Lin, “A ga-based neural fuzzy system for temperature control,”
Fuzzy Sets and Systems, vol. 143, no. 2, pp. 311–333, 2004.
[8] F.-J. Lin, C.-H. Lin, and P.-H. Shen, “Self-constructing fuzzy neural
network speed controller for permanent-magnet synchronous motor
drive,” IEEE transactions on fuzzy systems, vol. 9, no. 5, pp. 751–759,
2001.
[9] R.-J. Wai and L.-C. Shih, “Adaptive fuzzy-neural-network design for
voltage tracking control of a dc–dc boost converter,” IEEE transactions
on power electronics, vol. 27, no. 4, pp. 2104–2115, 2012.
[10] Y. Pan, T. Sun, and H. Yu, “Peaking-free output-feedback adaptive neural
control under a nonseparation principle,” IEEE transactions on neural
networks and learning systems, vol. 26, no. 12, pp. 3097–3108, 2015.
[11] Y. Pan and H. Yu, “Biomimetic hybrid feedback feedforward neuralnetwork learning control,” IEEE transactions on neural networks and
learning systems, vol. 28, no. 6, pp. 1481–1487, 2017.
[12] Y. Pan, T. Sun, Y. Liu, and H. Yu, “Composite learning from adaptive
backstepping neural network control,” Neural Networks, vol. 95, pp.
134–142, 2017.
[13] T. Sun, H. Pei, Y. Pan, H. Zhou, and C. Zhang, “Neural network-based
sliding mode adaptive control for robot manipulators,” Neurocomputing,
vol. 74, no. 14, pp. 2377–2384, 2011.
[14] M. Pratama, S. G. Anavatti, P. P. Angelov, and E. Lughofer, “Panfis:
A novel incremental learning machine,” IEEE Transactions on Neural
Networks and Learning Systems, vol. 25, no. 1, pp. 55–68, 2014.
[15] Y. Gao and M. J. Er, “Online adaptive fuzzy neural identification and
control of a class of mimo nonlinear systems,” IEEE Transactions on
Fuzzy Systems, vol. 11, no. 4, pp. 462–477, 2003.
[16] P. Angelov, “A fuzzy controller with evolving structure,” Information
Sciences, vol. 161, no. 1, pp. 21–35, 2004.
[17] C.-M. Lin and T.-Y. Chen, “Self-organizing cmac control for a class
of mimo uncertain nonlinear systems,” IEEE Transactions on Neural
Networks, vol. 20, no. 9, pp. 1377–1384, 2009.
[18] J.-C. de Barros and A. L. Dexter, “Evolving fuzzy model-based adaptive
control,” in Fuzzy Systems Conference, 2007. FUZZ-IEEE 2007. IEEE
International. IEEE, 2007, pp. 1–5.
IEEE TRANSACTIONS ON FUZZY SYSTEMS
[19] I. Škrjanc, S. Blažič, and P. Angelov, “Robust evolving cloud-based pid
control adjusted by gradient learning method,” in Evolving and Adaptive
Intelligent Systems (EAIS), 2014 IEEE Conference on. IEEE, 2014, pp.
1–8.
[20] D. Leite, R. M. Palhares, V. C. Campos, and F. Gomide, “Evolving
granular fuzzy model-based control of nonlinear dynamic systems,”
IEEE Transactions on Fuzzy Systems, vol. 23, no. 4, pp. 923–938, 2015.
[21] H.-G. Han, X.-L. Wu, and J.-F. Qiao, “Real-time model predictive
control using a self-organizing neural network,” IEEE Transactions on
Neural Networks and Learning Systems, vol. 24, no. 9, pp. 1425–1436,
2013.
[22] H. Han and J. Qiao, “Nonlinear model-predictive control for industrial
processes: An application to wastewater treatment process,” IEEE Transactions on Industrial Electronics, vol. 61, no. 4, pp. 1970–1982, 2014.
[23] H.-G. Han, L. Zhang, Y. Hou, and J.-F. Qiao, “Nonlinear model
predictive control based on a self-organizing recurrent neural network,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 27,
no. 2, pp. 402–415, 2016.
[24] M. Pratama, S. G. Anavatti, and E. Lughofer, “Genefis: toward an
effective localist network,” IEEE Transactions on Fuzzy Systems, vol. 22,
no. 3, pp. 547–562, 2014.
[25] M. M. Ferdaus and M. Pratama, “G controller code,” 2018. [Online].
Available: https://www.ntu.edu.sg/home/mpratama/Publication.html
[26] M. Pratama, S. G. Anavatti, M. Garratt, and E. Lughofer, “Online identification of complex multi-input-multi-output system based on generic
evolving neuro-fuzzy inference system,” in Evolving and Adaptive
Intelligent Systems (EAIS), 2013 IEEE Conference on. IEEE, 2013,
pp. 106–113.
[27] N. Wang, M. J. Er, J.-C. Sun, and Y.-C. Liu, “Adaptive robust online
constructive fuzzy control of a complex surface vehicle system,” IEEE
Transactions on Cybernetics, vol. 46, no. 7, pp. 1511–1523, 2016.
[28] M. M. Ferdaus, S. G. Anavatti, M. A. Garratt, and M. Pratama,
“Fuzzy clustering based nonlinear system identification and controller
development of pixhawk based quadcopter,” in Advanced Computational
Intelligence (ICACI), 2017 Ninth International Conference on. IEEE,
2017, pp. 223–230.
[29] D. Dovžan, S. Blažič, and I. Škrjanc, “Towards evolving fuzzy reference
controller,” in Evolving and Adaptive Intelligent Systems (EAIS), 2014
IEEE Conference on. IEEE, 2014, pp. 1–8.
[30] A. Zdešar, D. Dovžan, and I. Škrjanc, “Self-tuning of 2 dof control
based on evolving fuzzy model,” Applied Soft Computing, vol. 19, pp.
403–418, 2014.
[31] A. B. Cara, L. J. Herrera, H. Pomares, and I. Rojas, “New online selfevolving neuro fuzzy controller based on the tase-nf model,” Information
Sciences, vol. 220, pp. 226–243, 2013.
[32] H. Han, W. Zhou, J. Qiao, and G. Feng, “A direct self-constructing neural
controller design for a class of nonlinear systems,” IEEE Transactions on
Neural Networks and Learning systems, vol. 26, no. 6, pp. 1312–1322,
2015.
[33] A. B. Cara, H. Pomares, and I. Rojas, “A new methodology for the online
adaptation of fuzzy self-structuring controllers,” IEEE Transactions on
Fuzzy Systems, vol. 19, no. 3, pp. 449–464, 2011.
[34] D. LI, X. Cheng, and Z. Xu, “Gain adaptive sliding mode controller
for flight attitude control of mav,” Optics and Precision Engineering,
vol. 21, no. 5, pp. 1183–1191, 2013.
[35] X. Chen, D. Li, Z. Xu, and Y. Bai, “Robust control of quadrotor
mav using self-organizing interval type-ii fuzzy neural networks (soitiifnns) controller,” International Journal of Intelligent Computing and
Cybernetics, vol. 4, no. 3, pp. 397–412, 2011.
[36] W. He, Z. Yan, and C. Sun, “Trajectory tracking control of a flapping
wing micro aerial vehicle via neural networks,” in Advanced Robotics
and Mechatronics (ICARM), International Conference on. IEEE, 2016,
pp. 443–448.
[37] T. S. Clawson, S. Ferrari, S. B. Fuller, and R. J. Wood, “Spiking neural
network (snn) control of a flapping insect-scale robot,” in Decision and
Control (CDC), 2016 IEEE 55th Conference on. IEEE, 2016, pp. 3381–
3388.
[38] S. Suresh and N. Sundararajan, “An on-line learning neural controller
for helicopters performing highly nonlinear maneuvers,” Applied Soft
Computing, vol. 12, no. 1, pp. 360–371, 2012.
[39] C. Dong, N. Wang, and M. J. Er, “Self-organizing adaptive robust fuzzy
neural attitude tracking control of a quadrotor,” in Control Conference
(CCC), 2016 35th Chinese. IEEE, 2016, pp. 10 724–10 729.
[40] A. V. Topalov and O. Kaynak, “Online learning in adaptive neurocontrol
schemes with a sliding mode algorithm,” IEEE Transactions on Systems,
Man, and Cybernetics, Part B (Cybernetics), vol. 31, no. 3, pp. 445–450,
2001.
15
[41] A. V. Topalov, K.-C. Kim, J.-H. Kim, and B.-K. Lee, “Fast genetic
on-line learning algorithm for neural network and its application to
temperature control,” in Evolutionary Computation, 1996., Proceedings
of IEEE International Conference on. IEEE, 1996, pp. 649–654.
[42] G. G. Parma, B. R. Menezes, and A. P. Braga, “Sliding mode algorithm
for training multilayer artificial neural networks,” Electronics Letters,
vol. 34, no. 1, pp. 97–98, 1998.
[43] S. Yu, X. Yu, and Z. Man, “A fuzzy neural network approximator with
fast terminal sliding mode and its applications,” Fuzzy Sets and Systems,
vol. 148, no. 3, pp. 469–486, 2004.
[44] G. Cascella, F. Cupertino, A. Topalov, O. Kaynak, and V. Giordano,
“Adaptive control of electric drives using sliding-mode learning neural
networks,” in Industrial Electronics, 2005. ISIE 2005. Proceedings of the
IEEE International Symposium on, vol. 1. IEEE, 2005, pp. 125–130.
[45] M. O. Efe, O. Kaynak, and X. Yu, “Sliding mode control of a three
degrees of freedom anthropoid robot by driving the controller parameters
to an equivalent regime,” Journal of Dynamic Systems, Measurement,
and Control, vol. 122, no. 4, pp. 632–640, 2000.
[46] E. Kayacan, E. Kayacan, H. Ramon, and W. Saeys, “Adaptive neurofuzzy control of a spherical rolling robot using sliding-mode-controltheory-based online learning algorithm,” IEEE Transactions on Cybernetics, vol. 43, no. 1, pp. 170–179, 2013.
[47] J. Kok and J. Chahl, “A low-cost simulation platform for flapping
wing mavs,” in SPIE Smart Structures and Materials+ Nondestructive
Evaluation and Health Monitoring. International Society for Optics
and Photonics, 2015, pp. 94 290L–94 290L.
[48] Z. J. Wang, “The role of drag in insect hovering,” Journal of Experimental Biology, vol. 207, no. 23, pp. 4147–4155, 2004.
[49] ——, “Dissecting insect flight,” Annu. Rev. Fluid Mech., vol. 37, pp.
183–210, 2005.
[50] Z. J. Wang and D. Russell, “Effect of forewing and hindwing interactions
on aerodynamic forces and power in hovering dragonfly flight,” Physical
review letters, vol. 99, no. 14, p. 148101, 2007.
[51] G.-B. Huang, P. Saratchandran, and N. Sundararajan, “An efficient
sequential learning algorithm for growing and pruning rbf (gap-rbf)
networks,” IEEE Transactions on Systems, Man, and Cybernetics, Part
B (Cybernetics), vol. 34, no. 6, pp. 2284–2292, 2004.
[52] P. Angelov, “Fuzzily connected multimodel systems evolving autonomously from data streams,” IEEE Transactions on Systems, Man,
and Cybernetics, Part B (Cybernetics), vol. 41, no. 4, pp. 898–910, 2011.
[53] G.-B. Huang, P. Saratchandran, and N. Sundararajan, “A generalized
growing and pruning rbf (ggap-rbf) neural network for function approximation,” IEEE Transactions on Neural Networks, vol. 16, no. 1, pp.
57–67, 2005.
[54] R. J. Oentaryo, M. J. Er, L. San, L. Zhai, and X. Li, “Bayesian art-based
fuzzy inference system: A new approach to prognosis of machining
processes,” in Prognostics and Health Management (PHM), 2011 IEEE
Conference on. IEEE, 2011, pp. 1–10.
[55] K. S. Yap, C. P. Lim, and I. Z. Abidin, “A hybrid art-grnn online learning
neural network with a ε insensitive loss function,” IEEE transactions
on neural networks, vol. 19, no. 9, pp. 1641–1646, 2008.
[56] K. S. Yap, C. P. Lim, and M. T. Au, “Improved gart neural network
model for pattern classification and rule extraction with application to
power systems,” IEEE transactions on neural networks, vol. 22, no. 12,
pp. 2310–2323, 2011.
[57] E. Lughofer, “A dynamic split-and-merge approach for evolving cluster
models,” Evolving Systems, vol. 3, no. 3, pp. 135–151, 2012.
[58] ——, Evolving fuzzy systems-methodologies, advanced concepts and
applications. Springer, 2011, vol. 53.
[59] S. Wu, M. J. Er, and Y. Gao, “A fast approach for automatic generation
of fuzzy rules by generalized dynamic fuzzy neural networks,” IEEE
Transactions on Fuzzy Systems, vol. 9, no. 4, pp. 578–594, 2001.
[60] E. Kayacan, O. Cigdem, and O. Kaynak, “Sliding mode control approach
for online learning as applied to type-2 fuzzy neural networks and its
experimental evaluation,” IEEE transactions on industrial electronics,
vol. 59, no. 9, pp. 3510–3520, 2012.
[61] V. I. Utkin, Sliding modes in control and optimization. Springer Science
& Business Media, 2013.
[62] E. Kayacan and R. Maslim, “Type-2 fuzzy logic trajectory tracking
control of quadrotor vtol aircraft with elliptic membership functions,”
IEEE/ASME Transactions on Mechatronics, vol. 22, no. 1, pp. 339–348,
2017.
[63] J.-J. E. Slotine, W. Li et al., Applied nonlinear control. Prentice hall
Englewood Cliffs, NJ, 1991, vol. 199, no. 1.
[64] M. M. Ferdaus, S. G. Anavatti, M. A. Garratt, and M. Pratama, “Fuzzy
Clustering based Modelling and Adaptive Controlling of a Flapping
IEEE TRANSACTIONS ON FUZZY SYSTEMS
Wing Micro Air Vehicle,” in Computational Intelligence (IEEE SSCI),
2017 IEEE Symposium Series on. IEEE, 2017, pp. 1914–1919.
Md Meftahul Ferdaus All about you and the what
your interests are.
C:/Users/N1705348C/examples/CV-image
Mahardhika Pratama M
16
| 3cs.SY
|
arXiv:1609.02377v1 [math.GT] 8 Sep 2016
Boundaries of Kleinian groups
Peter Haı̈ssinsky, Luisa Paoluzzi, and Genevieve Walsh
September 9, 2016
Abstract
We review the theory of splittings of hyperbolic groups, as
determined by the topology of the boundary. We give explicit
examples of certain phenomena and then use this to describe
limit sets of Kleinian groups up to homeomorphism.
AMS classification: Primary 20F67, 30F40; Secondary 57N16.
1
Introduction and Background
A Kleinian group G is a discrete subgroup of PSL2 (C). It acts properly
discontinuously on the hyperbolic 3-space H3 via orientation-preserving
b via Möbius transforisometries and it acts on the Riemann sphere C
mations. The latter action is usually not properly discontinuous: there
is a canonical and invariant partition
b = ΩG ⊔ ΛG
C
where ΩG denotes the ordinary set, which is the largest open set of
b on which G acts properly discontinuously, and where ΛG denotes
C
b
the limit set, which is the minimal G-invariant compact subset of C.
A Kleinian group G is convex-cocompact if there is a convex subset C
We would like to thank CIRM (Centre International de Recontres
Mathématiques) for funding the 2-week “Research in Pairs” program where this
project was initiated. P. Haı̈sinsky and L. Paoluzzi were partially supported by ANR
project GDSous/GSG no. 12-BS01-0003-01. G. Walsh was partially supported by
NSF grant 1207644.
1
invariant under G such that the restriction of the action of G to C is
cocompact.
We wish to address the following general problem.
Problem. — Classify (convex-cocompact) Kleinian groups from the
topology of their limit set. In particular, what can be said of two convexcocompact Kleinian groups which have homeomorphic limit sets?
We will provide some examples showing that there are no obvious
answers and that certainly an interesting answer might depend on the
specific topology of the limit sets we are considering. We observe below in Section 3.1 that a particular type (graph-Kleinian) of Kleinian
group can be detected via the boundary and we expect that a different
type (mixed Kleinian surface) is not detected via the boundary. This
question may be posed more generally for word hyperbolic groups: we
will start reviewing their definition and main properties with respect
to this problem and similar ones.
1.1
Hyperbolic spaces and groups in the sense of Gromov
Mirroring a key property of Hn , a geodesic metric space X is called
Gromov-hyperbolic or δ-hyperbolic if there exists a δ ≥ 0 such that for
every geodesic triangle in X, the third side is contained the union of the
δ-neighborhoods of the other two. Analogous to Fuchsian and Kleinian
groups, hyperbolic groups are groups G that by definition act geometrically (properly discontinuously, co-compactly and by isometries) on
some Gromov-hyperbolic metric space X. Since any two geodesic metric spaces on which G acts geometrically are quasi-isometric [6, p 141]
and any geodesic metric space that is quasi-isometric to a hyperbolic
space is hyperbolic [6, p 409-410], this notion is well-defined. We will
abuse notation and say that G is quasi-isometric to a space X. For
example, if G is the fundamental group of a surface of genus g ≥ 2,
then G is hyperbolic and quasi-isometric to H2 .
When X is a hyperbolic metric space, we can define the boundary
of X, ∂X, which is a topological space. An important property of
this boundary is that when X is hyperbolic, and X ′ is quasi-isometric
to X, the boundaries ∂X and ∂X ′ are homeomorphic. Therefore, we
2
consider this to be the boundary of G when G acts geometrically on
X. We denote this boundary ∂G ∼
= ∂X, when X is any geodesic space
on which G acts geometrically.
There are many equivalent definitions of the boundary of a hyperbolic space X. For our purposes we will consider the geodesic boundary.
The points in this space consist of equivalence classes of geodesic rays
in X, where two rays are equivalent if they have bounded Hausdorff
distance. We can define a natural topology on the boundary given by
the following neighborhood basis. For p ∈ ∂X,
V (p, r) = {q ∈ ∂X : for some geodesic rays γ1
and γ2 , [γ1 ] = p, [γ2 ] = q, lim inf (γ1 (s), γ2 (t))x ≥ r}
s,t→∞
where (y, z)x = 12 (d(x, y) + d(x, z) − d(y, z)). Then as r → ∞, these
nested sets form a basis for the topology of ∂X.
Since the topological type of the boundary is an invariant of the
quasi-isometry class of a hyperbolic group, it is natural to ask if hyperbolic groups with homeomorphic boundaries are quasi-isometric. This
is false. There are examples of hyperbolic buildings given by Bourdon [2] which are not quasi-isometric but which have homeomorphic
boundaries. Their associated isometry groups act geometrically on the
buildings giving counter-examples for groups. It should be noted that
the boundary carries a canonical quasiconformal structure which determines the group up to quasi-isometry [22]. A natural numerical
quasi-isometry invariant coming from the quasiconformal structure is
the conformal dimension of the boundary [21]. In Bourdon’s examples, this invariant suffices to distinguish buildings from one another.
We will give examples of groups in the sequel which all have the same
conformal dimension, equal to one.
1.2
Convex co-compact Kleinian groups
An important class of hyperbolic groups are convex co-compact Kleinian
groups. A Kleinian group is a discrete subgroup of Isom(H3 ). When
the quotient of H3 by a Kleinian group G is a manifold H3 /G, we
call it a Kleinian manifold. Regarding H3 in the Poincaré ball model,
∂H3 ∼
= S 2 , and H3 ∪ ∂H3 is naturally a closed ball. Isometries of H3
extend to (and are extensions of) conformal transformations of S 2 . If
3
G is a Kleinian group, the closure in S 2 of an orbit of a point doesn’t
depend on the point. This set in S 2 is called the limit set of G, Λ(G).
Let C(G) denote the convex hull of Λ(G) in H3 ∪∂H3 . A Kleinian group
G leaves its limit set invariant, and therefore acts properly discontinuously and by isometries on C(G) \ Λ(G). Since H3 is a δ-hyperbolic
space, so is any convex subset of H3 . When G acts co-compactly on
C(G) \ Λ(G), G is a hyperbolic group and ∂G ∼
= Λ(G). In this case we
say G is a convex co-compact Kleinian group. For example, the fundamental groups of closed hyperbolic 3-orbifolds, the fundamental groups
of hyperbolic orbifolds with totally geodesic boundary, free groups, and
the fundamental groups of closed surfaces of genus greater than 1 all
have actions as convex co-compact Kleinian groups. The fundamental
groups of hyperbolic knot complements do not, since they are not hyperbolic groups. When G is a convex co-compact Kleinian group and
(C(G)\Λ(G))/G is a manifold, we call the quotient a convex co-compact
Kleinian manifold.
The class of convex co-compact Kleinian groups exhibits both rigidity and non-rigidity with respect to boundary homeomorphism. For
example, any hyperbolic group with boundary homeomorphic to two
points is virtually cyclic and any hyperbolic group with boundary a
Cantor set is virtually free [15, § 8]. Due to the work of many authors
[12, 9, 4] any hyperbolic group with boundary homeomorphic to S 1
is virtually a Fuchsian group. Thus for groups with these boundaries,
∂G ∼
= ∂G′ implies that G is quasi-isometric to G′ .
On the other hand, the situation for Kleinian groups with Sierpiński
carpet boundary is somewhat different. The Sierpiński carpet is the
unique planar, 1-dimensional, connected, locally connected, compact
topological space without cut points or local cut points. The Sierpiński
carpet can be realized as the complement of a union of round open
discs in S 2 . When Γ is the fundamental group of a hyperbolic 3manifold with totally geodesic boundary, Γ can be realized as a convex
co-compact Kleinian group G where Λ(G) (and hence ∂G) is homeomorphic to the Sierpiński carpet. The stabilizers of the discs are precisely the conjugates of the surface groups corresponding to the totally
geodesic boundary components. The boundaries of the discs are the
boundaries of these surface subgroups. These 3-manifold groups have
homeomorphic boundaries and may not be quasi-isometric.
Frigerio [11] has proven that the fundamental groups of hyperbolic
4
manifolds with totally geodesic boundary are quasi-isometric exactly
when the groups are commensurable. It is easy to construct noncommensurable hyperbolic manifolds with totally geodesic boundary:
Every manifold M with 1 totally geodesic boundary component contains a knot k such that M \ K is a hyperbolic manifold with 1 cusp
and 1 totally geodesic boundary component, by Myers [19]. Then highenough Dehn filling on this cusp will produce hyperbolic manifolds
whose volumes approach the volume of the cusped manifold. Since
their boundaries have the same genus, any commensurable manifolds
in this set must have the same volume. As Dehn filling produces manifolds with infinitely many different volumes, there are infinitely many
commensurability classes in this set. Then by Frigerio, there are infinitely many quasi-isometry classes in this set. Bourdon and Kleiner
[3] give a completely different example of infinitely many groups with
Sierpiński carpet boundaries where the groups have different conformal
dimensions which shows they are not quasi-isometric.
A less restrictive rigidity question is: If ∂G ∼
= ∂G′ , does G act
′
geometrically on the same type of space as G ? (Here type might
mean different things.) In the category of convex co-compact Kleinian
groups, there are several important open questions in this spirt.
Conjecture 1.1 (Cannon). [7] If G is a hyperbolic group with ∂G ∼
=
2
3
S , G acts geometrically on H .
Conjecture 1.2 (Kapovich and Kleiner). [16] If G is a hyperbolic
group with ∂G ∼
= S, where S is the Sierpiński carpet, then G acts geometrically on a convex subset of H3 bounded by totally geodesic planes.
2
Review of Bowditch’s splittings
We now briefly review Bowditch’s theory of an important relationship
between the topology of the boundary of a group and the group’s algebraic structure. We remark that there are prior and related theories
of group splittings which generalize the splittings of a manifold along
essential spheres, tori, and cylinders ([23]).
Theorem 2.1 (Theorem 0.1, [5]). Suppose that Γ is a one-ended hyperbolic group, which is not a co-compact Fuchsian group. Then there
is a canonical splitting of Γ as a finite graph of groups such that each
5
edge group is two-ended, and each vertex group is of one of the following
three types:
1. a two-ended subgroup,
2. a maximal “hanging fuchsian” subgroup, or
3. a non-elementary quasiconvex subgroup not of type (2).
These types are mutually exclusive, and no two vertices of the same
type are adjacent. Every vertex group is a full quasiconvex subgroup.
Moreover, the edge groups that connect to any given vertex group of
type (2) are precisely the peripheral subgroups of that group.
Note: the peripheral subgroups are the (orbi-)boundaries of the
surfaces (orbifolds) in type (2)
The graph of group splitting can be seen from the topology of the
boundary as follows. The local cut points of the boundary M = ∂G are
the points p ∈ M such that M \ p has more than one end. The number
of ends is called the valency of p, V al(p). Two local cut points x and
y are said to be equivalent if x = y or if V al(x) = V al(y) = n where
M \ {x, y} has n components. Bowditch proves that all equivalence
classes of local cut points contain either 2 or infinitely many elements.
The stabilizers of pairs of equivalent local cut points are 2-ended groups
and hence virtually cyclic. Such a pair {x, y} will correspond to a
splitting of the group where the vertex groups are the stabilizers of
the the closures of the components of M \ {x, y}, and another copy of
Z for the pair {x, y}. The infinite equivalence classes are all valence
2, and each infinite equivalence class σ comes equipped with a cyclic
order. The closure σ̄ is a cyclically ordered Cantor set, whose stabilizer
is a free group subgroup of G. The equivalent pairs of points in σ̄ \ σ
correspond to 2-ended groups, which are edge groups in the graph of
groups splitting. The stabilizer of σ̄ is a virtually a surface group which
is “hanging onto” the rest of the group via its boundary subgroups. The
rigid pieces will arise as follows. Each splitting along a two-ended group
will result in several components, where the stabilizers of the closures
of these components are subgroups of G. The intersections of these
subgroups (corresponding to intersections of the components) may be
a hanging Fuchsian group as above, which can be detected via the local
cut points. On the other hand, they could have no further splittings,
6
in which case these group are the rigid subgroups. We give such an
example below.
3
Examples
We consider examples of one-ended convex-cocompact Kleinian groups
where each piece in its JSJ decomposition is a free group. It follows that
their conformal dimensions are all equal to one so conformal dimension
cannot be used to distinguish quasi-isometry classes [8].
3.1
Books of I-bundles
A graph-Kleinian manifold is a convex co-compact Kleinian manifold
that admits a decomposition along essential annuli so that each piece is
a solid torus or an I-bundle over a compact surface with boundary. For
each I-bundle over S, we require that ∂S × I is exactly its intersection
with the annuli. In terms of the Bowditch decomposition above, the
vertex groups are all of type (1) or (2). A graph-Kleinian manifold is
exactly a Kleinian manifold with boundary whose double is a graphmanifold. These are often called “Books of I-bundles” as in [10, p
286].
Observation 3.1. Let G be a hyperbolic group with boundary homeomorphic to the boundary of a graph-Kleinian group. Then G is virtually
a graph-Kleinian group.
The proof is a special case of [13, Theorem 1.2], although the statement is slightly different. We give a brief outline of the proof in our
situation. Let G be a group whose boundary is homeomorphic to the
boundary of a graph-Kleinian group K. By Bowditch’s JSJ splitting
theorem above, G is a amalgamated product of virtually Fuchsian
groups, the vertex groups, amalgamated over edge groups which are
their boundary subgroups. Since this decomposition forms a malnormal quasi-convex hierarchy, the group is virtually special, and hence
quasi-convex subgroups of G are separable by [25]. Therefore, there is
a finite index subgroup G′ of G with a graph of groups decomposition
where the vertex groups are torsion free Fuchsian groups and the edge
groups are cyclic peripheral subgroups. Furthermore in this decomposition of G′ the generators of the peripheral subgroup are primitive in
7
G′ . Therefore, we can form a 3-manifold with fundamental group G′
which is a union of thickened surfaces S × I glued along ∂S × I to solid
tori. The gluing pattern is determined by the graph of groups decomposition given by Bowditch. Then applying Thurston’s Uniformization
Theorem, G′ can be realized as a convex co-compact Kleinian group.
Finally, Bowditch’s characterization of the hanging Fuchsian groups
implies that G′ is a graph-Kleinian group.
3.2
One rigid piece
Now we consider the case when the Kleinian manifold is made out of
surfaces, but some of the pieces may be rigid. A mixed Kleinian surface
group is a one-ended convex co-compact Kleinian group Γ such that
the Kleinian manifold C(Γ)/Γ admits a decomposition along essential
annuli so that each piece is a solid torus or an I-bundle over a compact
surface with boundary (so ∂S × I is contained in the set of annuli).
Furthermore, there must be at least one rigid piece in the Bowditch
decomposition of Γ. In contrast with graph-Kleinian groups above, it
is not at all clear that an analog of Observation 3.1 holds for mixed
Kleinian surface groups. We give a simple example, which seems to be
very special.
Example 3.2. (The ABC example) Consider a genus two handlebody
H with standard generators a and b. Let c = [a, b]. Note that H is
homeomorphic to T × [0, 1], where T is a one-holed torus. Now embed
curves representing a, b and c as follows. The curve a lies in T × 0,
b lies in T × 1, and c is the boundary curve on T × 21 . See Figure
1. Now attach 3 copies of S × I, where S is a one-holed surface, to
regular annular neighborhoods of a, b and c along ∂S × I. We denote
the manifold by Mabc and its fundamental group by Γabc .
Proposition 3.3. The group Γabc is a mixed Kleinian surface group.
Proof. The group can be realized as a convex co-compact Kleinian
group by applying Thurston’s uniformization theorem to Mabc . Since
the boundary of Mabc is incompressible, Γabc is one-ended. The Bowditch
decomposition will have cyclic vertex groups corresponding to regular
solid toric neighborhoods of the three attaching curves a, b and c. There
8
c
b
a
a
b
c
Figure 1: The curves a, b and c, on T × [0, 1] and on H, respectively
will be three “maximally hanging Fuchsian” vertex groups corresponding to the three copies of S × I.
It remains to show that the manifold T × I, pared along annular
neighborhoods of a, b and c (denoted Na , Nb , Nc ) is acylindrical, and
thus rigid. Indeed, consider T ×I, doubled along ∂(T ×I)\(Na ∪Nb ∪Nc ).
We claim this double W is homeomorphic to the exterior B of the
Borromean rings in S 3 . It is well-known that the complement of the
Borromean rings admits a hyperbolic metric of finite volume and this
implies that W is atoroidal and the pared manifold is acylindrical.
That W is homeomorphic to B is well-known and was first noticed in
Hodgson’s Thesis [14, Section 4].
In Figure 3, opposite sides of the cube should be identified to obtain
the double, so that W can be seen also as the exterior of a link in the 3torus. A standard computation shows that the fundamental group of W
admits the following presentation: ha, b, c | [a, [B, c]], [b, [C, a]], [c, [A, b]]i
(where capital letters denote inverses). It is easy to see that this group
is isomorphic to the fundamental group of B. Note that W has a cover
W̃ corresponding to the universal cover of the 3-torus: this is R3 minus the tubular neighborhoods of infinitely many lines parallel to the
standard axes. This is also a cover of the Borromean rings exterior. Indeed, the Borromean rings’ exterior is obtained as the quotient of this
cover by the action of the group generated by π-rotations of R3 about
the lines whose neighborhoods have been removed. Note also that a
fundamental domain for W is compact. Therefore, the complement of
the link in T 3 in Figure 3 admits a complete finite volume hyperbolic
metric, induced by that of the cover of the Borromean rings comple9
ment. By Mostow rigidity, using the fact that π1 (W ) = π1 (B), W is
homeomorphic to the Borromean rings exterior.
Therefore, Γabc has the property that all the vertex groups in the
graph of groups decomposition are (free) surface groups, and at least
one of them is a rigid group.
Note that, since the vertex group corresponding to H is rigid, there
is a hyperbolic structure on a genus 2 handlebody, where the curves
a, b, and c are parabolic. After the curves are pinched to parabolics,
there are two three-cusped spheres as the boundary, and these are
totally geodesic. We denote this Kleinian group (which is unique up to
conjugation) by HW , as it is the fundamental group of a handlebody
where some of the boundary curves are pinched to parabolics.
Lemma 3.4. The limit set of HW , Λ(HW ), is the Apollonian gasket.
Proof. The Kleinian group HW is the fundamental group of the genus
two handlebody, pared along the generators a and b and their commutator c. This admits a hyperbolic structure with totally geodesic
boundary. The group HW is a free group on two generators a and b
where a, b and the commutator ABab are all parabolic. (Capital letters
denote inverses.) Up to conjugacy, we may assume a : z → z + 1 and
the fixed point of b is zero. Solving the equations so that the commutator is parabolic we obtain: b : z → 1/(2iz + 1). Then the fixed point
of ABab is (−1 + i)/2. The limit set is pictured in Figure 2, and can
easily be seen to be conjugate to the Apollonian gasket in Figure 7.5
of [18], which is also the limit set of a free group of rank 2.
Theorem 3.5. Any hyperbolic group G with ∂G homeomorphic to
∂Γabc is a virtually mixed Kleinian surface group. Furthermore, each
rigid vertex group in the Bowditch decomposition of G is commensurable with HW .
Proof. Since ∂G is homeomorphic to ∂Γabc , G is one-ended and ∂G
has the same Bowditch decomposition as ∂Γabc . In particular, G has
a graph of groups decomposition where the edge groups are virtually
cyclic, and the vertex groups are either maximally hanging Fuchsian
or rigid. Let VR be an arbitrary rigid vertex group of G. Then VR
is virtually free, since its boundary is homeomorphic to the boundary
10
Figure 2:
Part of the limit set of HW .
This picture
was made with Curt McMullen’s LIM program, available at
http://www.math.harvard.edu/~ctm/programs/index.html
of the rigid vertex group of Γabc , which is a Cantor set. We continue
to denote the finite index free subgroup of VR by VR . The boundary
∂VR has the property that identifying the endpoints of the edge groups
results in a planar set, since this space is homeomorphic to the limit
set of HW , Λ(HW ). Then, by Otal [20], VR is the fundamental group
of a handlebody H and there are essential curves {c1 , ..., cn } on H
such that each edge subgroup incident to VR is a conjugate of some
ci . Then, as VR is rigid, VR admits a hyperbolic structure with totally
geodesic boundary, where the ci are parabolic. Call this structure HG .
Then the limit set Λ(HG ) is homeomorphic to Λ(HW ). By Lemma 3.4,
Λ(HG ) is the Apollonian gasket. Any space consisting of round circles
homeomorphic to the Apollonian gasket is conjugate to the Apollonian
gasket, by taking the centers of three tangent circles to the centers of
three tangent circles. Then, since HW and HG are both finite index
in the discrete group of maximal symmetries of the limit set, HW is
commensurable to HG which implies HG is a surface group. As above
in Observation 3.1, we have a finite index subgroup G′ of G which is a
mixed Kleinian surface group.
11
b
c
a
Figure 3: A decomposition of the double of the rigid piece. Parallel
faces are identified by translations as for the 3-torus.
4
Homeomorphism types of limit sets
The concept of a tree of metric compacta was introduced by Świa̧tkowski
[24] as a method for understanding more exotic boundaries of hyperbolic groups, but it is also useful in understanding the boundaries of
Kleinian groups.
Definition 4.1. (Świa̧tkowski) A tree system of metric compacta is a
tuple Ω = (T, {Kt }, {Σe }, {φe }) such that
1. T is a countable tree
2. to each t ∈ VT , there is an associated compact metric space Kt
3. to each e ∈ ET , there is an associated non-empty compact subset
Σe ⊂ Kα(e) , and a homeomorphism φe : Σe → Σē such that
φē = φe −1 .
4. For each t ∈ VT , the family {Σe : e is an oriented edge emanating
from t} has the property that for every ǫ > 0, only finitely many
sets in the family have diameter > ǫ.
12
The limit of a tree system of metric compacta is the natural topological object obtained by identifying the vertex spaces via the φe along
with the ends of the tree. The precise definition is given in [24, Section
1.C].
Theorem 4.2. Let M be the boundary of a hyperbolic group, and suppose that M ⊂ S 2 . Then M is the limit of a tree system of metric
compacta, where each vertex space is either:
1. The 2-sphere S 2 (in which case the tree must be trivial)
2. The Sierpiński carpet.
3. The circle S 1
4. not connected.
5. itself the limit of a tree system of metric compacta
Proof. Let M be planar with M ∼
= ∂G. Then suppose M has no local
cut points. If M has covering dimension 2, it contains an open subset
of S 2 . Since G acts minimally on G (there are no non-trivial closed
invariant subsets) and M is closed, M has empty frontier M̄ \ int(M ).
Thus M must be all of S 2 . Now suppose M is one-dimensional. If it is
connected, it is locally connected. Therefore, in this situation we have a
compact, connected, locally connected subset of the plane with no local
cut point, which must be the Sierpiński carpet [26]. If M is connected
and 0-dimensional, it is a point. This cannot happen, since a group
cannot act geometrically on a hyperbolic space with one equivalence
class of geodesic rays.
If M is connected and has local cut points, then by Bowditch [5]
M is either homeomorphic to S 1 or there is a canonical splitting of
G over Z where G acts on a tree, where the stabilizer of each vertex
corresponds to a subgroup Gv of G; cf. Theorem 2.1. Then M is a tree
of metric compacta, where the compacta are the limit sets of the vertex
groups. Similarly, each vertex group has a boundary which is either
(2), (3), or (4) above, or has local cut points, and hence a canonical
splitting over Z. We only need to repeat this process finitely many
times, by [17].
13
We note that Benoist and Hulin have a characterization of the sets
in the Euclidean S 2 which are the limit sets of convex cocompact
Kleinian groups, [1]. They are exactly the conformally autosimilar
closed subsets of this sphere.
References
[1] Y. Benoist and D. Hulin, Conformal autosimilarity, preprint,
2015.
[2] M. Bourdon. Immeubles hyperboliques, dimension conforme et
rigidité de Mostow, Geom. Funct. Anal. 7 (1997) 245–268.
[3] M. Bourdon and B. Kleiner, Some applications of lpcohomology to boundaries of Gromov hyperbolic spaces. Groups
Geom. Dyn. 9 (2015) 435–478.
[4] B. H. Bowditch, A topological characterisation of hyperbolic
groups, J. Amer. Math. Soc. 11 (1998) 643–667.
[5] B. H. Bowditch, Cut points and canonical splittings of hyperbolic groups. Acta Math. 180 (1998) 145–186.
[6] M. R. Bridson and A. Haefliger, Metric spaces of non-positive
curvature, Grundlehren der Mathematischen Wissenschaften,
319 Springer-Verlag, Berlin, 1999.
[7] J. W. Cannon, The theory of negatively curved spaces and
groups. Ergodic theory, symbolic dynamics, and hyperbolic spaces (Trieste, 1989) 315–369, Oxford Sci. Publ., Oxford Univ. Press, New York, 1991.
[8] M. Carrasco, On the conformal gauge of a compact metric
space. Ann. Sci. Éc. Norm. Supér. (4) 46 (2013) 495–548.
[9] A. Casson and D. Jungreis, Convergence groups and Seifert
fibered 3-manifolds, Invent. Math. 118 (1994) 441–456.
[10] M. Culler and P.B. Shalen, Volumes of hyperbolic Haken manifolds, I, Invent. Math. 118 (1994) 285–329.
14
[11] R. Frigerio, Commensurability of hyperbolic manifolds with
geodesic boundary, Geom. Dedicata 118 (2006), 105–131.
[12] D. Gabai, Convergence groups are Fuchsian groups, Ann.
Math. 136 (1992), 447–510.
[13] P. Haı̈ssinsky. Quasi-isometric rigidity of the class of convex
cocompact Kleinian groups. Proceedings of the VIth AhlforsBers Colloquium, to appear.
[14] C. D. Hodgson, Degeneration and Regeneration of Geometric Structures on 3-manifolds, Thesis, Princeton University
(1986).
[15] I. Kapovich, and N. Benakli. Boundaries of hyperbolic groups,
Combinatorial and geometric group theory, 296 New York,
2002.
[16] M. Kapovich and B. Kleiner, Hyperbolic groups with lowdimensional boundary, Ann. Sci. Ecole Norm. Sup. 33 (2000),
647–669.
[17] L. Louder and N. Touikan. Strong accessibility for finitely
presented groups. arXiv:1302.5451, 2013.
[18] D. Mumford, C. Series, and D. Wright, Indra’s Pearls: the
Vision of Felix Klein. Cambridge University Press, 2002.
[19] R. Myers. Excellent 1-manifolds in compact 3-manifolds,
Topology Appl. 49 (1993), 115–127.
[20] J.-P. Otal. Certaines relations d’équivalence sur l’ensemble des
bouts d’un groupe libre. J. London Math. Soc. (2) 46 (1992),
123–139.
[21] P. Pansu. Dimension conforme et sphère à l’infini des variétés
à courbure négative. Ann. Acad. Sci. Fenn. Ser. A I Math.
14 (1989), 177–212.
[22] F. Paulin Un groupe hyperbolique est déterminé par son bord,
J. London Math. Soc. (2) 54 (1996), 50–74.
15
[23] P. Scott and G. A. Swarup Regular neighbourhoods and canonical decompositions for groups. Astérisque No. 289 (2003), 1–
233.
[24] J. Świa̧tkowski, Trees of metric compacta and trees of manifolds arXiv:1304.5064v2 (2016) 1–45.
[25] D. T. Wise. The structure of groups with a quasiconvex hierarchy. preprint, Oct. 2012.
[26] G. T. Whyburn. Topological characterization of the Sierpiński
curve. Fund. Math. 45 (1958), 320–324.
Peter Haı̈ssinsky
Aix-Marseille Université, CNRS, Centrale Marseille, I2M,
UMR 7373,
13453 Marseille, France
[email protected]
Luisa Paoluzzi
Aix-Marseille Université, CNRS, Centrale Marseille, I2M,
UMR 7373,
13453 Marseille, France
[email protected]
Genevieve S. Walsh
Tufts University
Medford MA 02155
[email protected]
16
| 4math.GR
|
Multidimensional Predicates for Prolog
Günter Khyo
arXiv:1603.06266v1 [cs.PL] 20 Mar 2016
March 22, 2016
Abstract
In 2014, Ungar et al. proposed Korz, a new computational model for structuring adaptive (objectoriented) systems [UOK14]. Korz combines implicit parameters and multiple dispatch to structure
the behavior of objects in a multidimensional space. Korz is a simple yet expressive model which does
not require special programming techniques such as the Visitor or Strategy pattern to accommodate
a system for emerging contextual requirements.
We show how the ideas of Korz can be integrated in a Prolog system by extending its syntax and
semantics with simple meta-programming techniques. We developed a library, called mdp (multidimensional predicates) which can be used to experiment with multidimensional Prolog systems. We
highlight its benefits with numerous scenarios such as printing debugging information, memoization,
object-oriented programming and adaptive GUIs.
In particular, we point out that we can structure and extend Prolog programs with additional
concerns in a clear and concise manner. We also demonstrate how Prolog’s unique meta-programming
capabilities allow for quick experimentation with syntactical and semantical enhancement of the new,
multidimensional model.
While there are many open concerns, such as efficiency and comprehensibility in the case of larger
systems, we will see that we can use the leverage of mdp and Prolog to explore new horizons in the
design of adaptive systems.
1
Introduction
Consider the problem of finding a path in a directional graph. The graph is represented by the
edge(N1, N2)-relation. In anticipation of changing requirements, we decide to use mdp from the
beginning and define the transitive relation path(From, To) as follows:
[] # edge (a , b ). [] # edge (b , c ). % ...
[] # path (A , B ) : - ? edge (A , B ).
[] # path (A , C ) : ? edge (A , B ) ,
? path (B , C ).
As a multidimensional relation, path/2 consists of two parts separated by a hashtag: (1) a list
of concerns (its context) and (2) the name of the predicate. Currently, there are no concerns and
therefore all context specifications are empty. For implementation reasons, when evaluating a multidimensional relation the ? operator has to be supplied explicitly.
Suppose we want to add debugging output to our program. We add a new variant of edge/2 by
adding a debugging dimension to its context:
[ debug : P ] # edge (A , B ) : [ - debug ] ? edge (A , B ) , % Remove debug dimension from implicit context
apply (P , [( A , B )]).
We get the desired behavior by executing path/2 within a debug context, i.e.,
[debug: writeln] ? path(X, Y). The context is implicitly passed along the call chain. If there
are multiple definitions of a predicate, the definitions which have the highest number of matching
dimension/value pairs will be chosen. Thus, our augmented definition of edge/2 will be preferred
over the plain definition.
In addition, the programmer may add arbitrary Prolog expressions to a context specification.
These then serve as preconditions which have to hold for the matching result to count. For example,
we could constrain our debugging context as follows:
[ debug : P , is_io_predi ca te (P , IODevice ) , ready ( IODevice )] # edge (A , B ) : - ...
% definitions for is_io_predic at e /2 and ready /1 ...
(where the programmer has to provide definitions for is io predicate and ready which serve as
illustration purposes only in this example).
If any of the preconditions do not hold, the default variant with no debugging will be executed.
If we do not want this behavior but throw an exception instead, we have to move the preconditions
inside the body of edge and throw an exception.
While we have a clean solution for debugging, we have to write our debugging code for every
predicate, which will soon become tedious. We can generalize our debugging concern for arbitrary
predicates by defining an anonymous multidimensional rule:
[ predicate : P ,
debug : D ,
is_io_predic at e (P , IODevice ) , ready ( IODevice )] : [ - debug ] ? P ,
apply (D , [ P ]).
The dimension predicate is part of every context and is supplied by mdp automatically. Note that
it is important that we remove our specialized debugging rule for edge/2, otherwise, we execute the
2
generalized definition and the specialized rules since both have equal matching scores 1 .
Looking back at the problem of finding paths, as experienced programmers we know immediately
that our solution will only work for acyclic graphs. To accommodate for cycles, we have to keep
record of already visited nodes. However, instead of simply adapting our implementation to the most
general case, we decide to keep the simpler and more efficient implementation for acyclic graphs. We
make following arrangements:
[] # path (A , B ) : - ? edge (A , B ).
[ graph_type : acyclic ] # path (A , C ) : ? edge (A , B ) ,
? path (B , C ).
We then add following lines:
[ graph_type : cylic , nodes : _ ] # path (A , B ) : - ? edge (A , B ).
[ graph_type : cylic , nodes : L ] # path (A , C ) : ? edge (A , B ) ,
\+ member (B , L ) ,
append ([ A ] , L , LWithA ) ,
[ nodes : LWithA ] ? path (B , C ).
[ graph_type : cyclic ] # path (A , C ) : ? edge (A , B ) ,
[ nodes : []] ? path (B , C ).
When we state the query
? - graph_type ( T ) , [ graph_type : T ] ? path (X , Y )
the most efficient of the available implementations will be chosen.
However, we have to be careful: If we do not provide any context information to our query, our
program may yield unpredictable results or even hang (since all path definitions are equally specific
and get ”‘mixed-up”’). For the same reason, it is also crucial that the first line also includes the
nodes dimension even if it is not needed in the base-case.
To fix this, in the default case, if we know nothing about the structure of a graph, it is safe to
execute the predicate which can handle cycles. We add following line:
[] # path (A , B ) : - [ graph_type : cyclic ] ? path (A , B ).
There are many other aspects that we can improve on. For example, we can add another dimension
which describes the type of data-structure for holding the visited nodes. Instead of using a list, we
could use a set or any data structure and let mdp select the most efficient predicates for inserting and
testing for members. Expanding the example is straightforward and left as an exercise to the reader.
In summary, mdp allows us to add concerns to a prolog program with relative ease. There are
many other useful applications for mdp as shown in Section 2, where we will see how we can model
objects, context-sensitive GUIs and memoization. Details on the implementation and guidance on
how to extend mdp are found in Section 3.
1 The programmer may also influence this behavior by assigning weights to context specifications (an example is shown
in Section 2.2.2).
3
2
Further mdp-Examples
In this section we will see more examples that motivate the use of mdp. We will address following
issues: memoization, object-oriented programming and how we can augment a GUI with additional
concerns.
2.1
Memoization
Memoization is a common technique used to cache expensive computations. In this section, we will
explore how we can orthogonally add the concern of memoization to prime number generation. We
will then derive a generic solution to memoize arbitrary predicates.
Suppose we have a naive implementation of a prime number test:
[] # is_prime ( P ) : P > 1,
UpperTestLimi t is floor ( sqrt ( P )) ,
forall ( between (2 , UpperTestLimit , Divisor ) , \+ 0 is P mod Divisor ).
Fortunately we know that, for some reason, there are only a few unknown numbers we are testing
so we might get away with memoization. We add following lines:
: - dynamic memoized_i s_ p ri m e /2.
[ memoize : _ ] # is_prime ( P ) : memoized_is _p r im e (P , IsPrime ) ->
call ( IsPrime ) ;
(
[ - memoize ] ? is_prime ( P ) ->
assertz ( memoized_is _p r im e (P , true )) ;
assertz ( memoized_is _p r im e (P , false ))
).
This works fine, but if we have to copy and paste the memoization code for every other predicate we
would like to enhance with memoization. We remove our specialized implementation and generalize
as follows:
: - dynamic memoized /2.
[ memoize : _ , predicate : P , memoized (P , TorF )]
: - call ( TorF ).
[ memoize : _ , predicate : P , \+ memoized (P , _ )] : forall ([ - memoize ] ? P , assertz ( memoized (P , true ))) ,
(
\+ [ - memoize ] ? P ->
assertz ( memoized (P , false )) , fail ;
memoized (P , _ )
).
This solution is slightly more complex, since it also handles memoization for non-deterministic
predicates. Now, for every multidimensional predicate P we can state the query:
? - [ memoize : _ ] ? P ( Args )
to memoize. Our solution can be further improved by extending assert and retract to invalidate
memoization data, which is left as an exercise to the reader.
4
2.2
Object-oriented Programming with mdp
In this section, we will explore how mdp can be used to implement a purely object-oriented system,
i.e., no classes. To this end, we built a simple meta-object-protocol mop on top of mdp. In the
following sections, we explore how we can represent messages and state in mdp and implement a
model for subtyping, polymorphism and protecting data.
2.2.1
Objects, Messages and State
In mop, as in Korz, an object is a globally unique identifier. New objects (identifiers) are created with
the predicate new_oid(+NewOID). Messages can be sent to objects with the ! operator (inspired by Erlang), for example, Display ! render(Object) which is equivalent to [rcvr: Display] ? render(Object).
Thus, a method is just a multidimensional predicate which includes a receiver dimension, containing
the id of the object.
The attribute of an object can be inspected and changed with OID ! read(Name, Value) and
OID ! write(Name, NewValue), respectively. Also, for every object there are predicates to clone the
object (which is a shallow copy of its attributes) and to query its type (which, internally, is just an
attribute).
The meta-object protocol is defined as follows:
: - dynamic data /3. % data ( OID , Name , Value )
: - op (0101 , xfx , !). % Send a message
[ rcvr : OID ] # write ( Name , Value ) : retractall ( data ( OID , Name , _ )) ,
assertz ( data ( OID , Name , Value )).
[ rcvr : OID ] # read ( Name , Value )
: - data ( OID , Name , Value ).
[ rcvr : OID ] # type ( T ) : - data ( OID , type , T ).
[ rcvr : OID ] # clone ( OIDClone ) : new_oid ( OIDClone ) ,
forall ( data ( OID , Name , Value ) , OIDClone ! write ( Name , Value )).
The message-send operator is implemented using a hook provided by mdp (see Section 3):
% Translate Receiver ! P ( Args ) to [ rcvr : Receiver ] ? P ( Args )
hook_mdp_ter m (
ImplicitContext ,
Receiver ! Predicate ,
?( ImplicitContext , [ rcvr : Receiver ] , Predicate )).
% Translate [ Context ] ? Receiver ! P ( Args ) to
%
[ rcvr : Receiver , Context ] ? Receiver ! P ( Args )
hook_mdp_ter m (
ImplicitContext ,
Receiver ! (?( GivenContext , Predicate )) ,
?( ImplicitContext , [ rcvr : Receiver | GivenContext ] , Predicate )).
The last definition allows us to provide a context to a message-send (we will see examples of this
in Section 2.2.3 and Section 2.3). (There are other hook-definitions to implement further syntactic
variations which can be found in the mdp distribution.)
5
2.2.2
Subtyping and Polymorphism
Let us consider the canonical example of modeling objects and behavior for representing geometric
shapes. On the type level, we define following subtype-relations:
subtype ( shape , rectangle ).
subtype ( rectangle , special_rec t an gl e ).
subtype ( shape , circle ).
We define the polymorphic predicate (method) representation/1 as follows:
% OID < shape reads : OID is a subtype of shape
[ rcvr : Shape , Shape < shape ] # representatio n ( _ ) : throw ( ’ Abstract method : No implementati on ’ ).
[ rcvr : Rectangle , Rectangle < rectangle ] # representatio n ( R ) : Rectangle ! read ( width , W ) ,
Rectangle ! read ( height , H ) ,
R = rectangle (W , H ).
[ rcvr : Rectangle , Rectangle < special_rec ta n gl e ] # representati on ( R ) : Rectangle ! read ( width , W ) ,
Rectangle ! read ( height , H ) ,
R = special_re ct a ng l e (W , H ).
[ rcvr : Circle , Circle < circle ] # representati on ( R ) : Circle ! read ( radius , Radius ) ,
R = circle ( Radius ).
Internally, the subtype-relation < is translated to a form mdp can understand, i.e.,
[ rcvr : OID , OID < rectangle ] # representatio n ( R ) : - ...
% Translates to
[ rcvr : OID ,
[ rcvr : OID ] ? type ( ActualType ) ,
@ ( type_affinity ( Subtype , ObjectType , D ) , D )
] # representatio n ( R ) : - ...
We can specify this translation by defining a rule for a hook supplied by mdp:
h o o k _ c o n t e x t _ r u l e _ m d p _ t e r m( Context , OID < Subtype , Translation ) : Translation = (
[ rcvr : OID ] ? type ( ObjectType ) ,
@ ( type_affinity ( Subtype , ObjectType , D ) , D )
).
where the predicate type_affinity measures how close the supplied type constraint fits the actual
type of the objects. The result is then used to give weight to the rule, using the @-annotation, to
ensure proper matching.
type_affinit y (T , S , N ) : max_type_di s ta n ce ( D ) ,
type_distance (T , S , DistanceTS ) ,
N is D - DistanceTS + 1.
6
type_distanc e (T , T , 1).
type_distanc e (T , S , N ) : T \= S ,
subtype ( Parent , S ) ,
type_distance (T , Parent , N1 ) ,
N is N1 + 1.
max_type_d i st an c e ( D ) : findall (D , (
subtype (T , _ ) ,
subtype (_ , S ) ,
type_distanc e (T , S , D )) , Distances ) ,
max_list ( Distances , D ).
While this is a highly inefficient implementation for a subtype check, we can use memoization to
make the type check efficient and we have the advantage of a readable specification.
We can now run the query AnyShapeType ! representation(R) and we get the representation
corresponding to the type of AnyShapeT ype. Since there are no classes, prototypes have to be defined
and cloned to create new instances. For example, the rectangle prototype can be created as follows:
?-
new_oid ( Rectangle ) ,
Rectangle ! write ( type , rectangle ) ,
Rectangle ! write ( width , 100) ,
Rectangle ! write ( height , 100) ,
Rectangle ! representatio n ( R ).
R = rectangle (100 , 100).
There are many improvements we can make to our meta-object protocol. For example, we could
check whether an attribute has been declared before writing to it.
2.2.3
Protecting Data
In many class-based object-oriented languages, the programmer may restrict access to certain attributes or methods with visibility modifiers. For example, consider following game:
% Overrides default mop read method
[ rcvr : OID , OID < guessing_game ] # read ( secret_number , S ) : throw ( ’ No funny stuff ’ ).
[ rcvr : OID , OID < guessing_game , modifier : private ] # read ( secret_number , S ) : number ( S ) ,
data ( OID , secret_number , S ).
[ rcvr : OID , OID < guessing_game ] # guess ( S ) : [ modifier : private ] ? OID ! read ( secret_number , S ).
We could improve our solution by protecting access to data/3 and providing a cryptographic
context which cannot be tempered with. The reader is encouraged to experiment with more secure
models.
The key point is that mdp provides all facilities to model a variety of concerns and no special
language constructs such as visibility modifiers are needed. In fact, mdp offers even more flexibility,
for instance, we can also imagine role-based visibility.
7
2.2.4
But why are there no classes?
Most object-oriented languages rely on concepts such as classes, meta-classes, subtyping and visibility
modifiers for data protection. In accordance to the research of Self [US87, CUS95], we do not believe
that there should be a separate language concept for classes (or a fixed class model) for following
reasons:
1. Classes as abstract entities break with the real-world metaphor of having tangible objects.
Humans are better dealing with the concrete, rather than abstract, non-tangible ideas.
2. Classes introduce the problem of infinite meta-regress [US87]. If there is a class, there must be
a meta-class, and if there is a meta-class, there must be a meta-meta-class and so on.
3. A purely object-based system can be used to implement a class-based system. For example, a
Smalltalk interpreter can be written in Self.
4. Just as classes, objects can serve as namespaces, protect access to attributes and do everything
a class description can do, even more.
mdp gives us the flexibility to create objects and contexts that serve as class descriptions or in
general, ontologies of any kind. The reader is encouraged to experiment with new ideas using mdp
and mop.
2.3
An Adaptive GUI
In this section we will explore how we can model an adaptive graphical user interface system. (Note
that the example is only a sketch and serves for demonstrating the potential of mdp. We will provide
full-fledged examples of adaptive GUIs in future works).
The system executes within a dynamic context which can be manipulated by the user or other
parts of the system. For the moment, we do not make any assumptions how the context is structured,
i.e., which dimensions of user-experience it contains. In this example, we focus on displaying objects
only.
In some part of the system, assume we have a predicate refresh(Display, Object) which is evaluated
whenever there is a change in the system context or one of its graphical objects:
refresh ( Display , Object ) : system_conte xt ( Context ) ,
Context ? Object ! representati o n ( R ) ,
Display ! present ( R ).
Different types of objects have different graphical representations. For the moment, we assume
that all objects are represented as colored bitmaps. We use a third-party library for rasterization.
There might be following implementations for the predicate representation.
[ rcvr : G , G < text ] # representati on ( R ) : G ! text_contents ( T ) ,
G ! text_color ( Color ) ,
text_pixmap (T , Color , R ). % Third party library
[ rcvr : G , G < box ] # representati on ( R ) : G ! bounds ( Width , Height ) ,
G ! color ( Color ) ,
box_pixmap ( Width , Height , Color , R ). % Third party library
8
% Specification s for other shape types ...
% If we do not recognize a new type of object , use a default representati on
[ rcvr : G , G < object ] # representati on ( R ) : d e f a u l t _ r e p r e s e n t a t i o n( R ).
Many modern GUIs such as browsers accept different types of graphical representations, for example vector graphics. Assuming we find out that our display object is capable of rendering vector
graphics, we do not have to change any of the existing code to support this. We just add the dimension render type to our system (and provide means of render-type-selection for the user) and add a
corresponding predicate definition:
[ rcvr : G ,
G
G
R
G
!
!
=
< box , render_type : svg ] # representati on ( R ) : bounds ( Width , Height ) ,
color ( Color ) ,
svg ( shape = box , color = Color ).
[ rcvr : G , G < box , render_type : pixmap ] # representatio n ( R ) : % Fallback to original implementatio n
[ - render_type ] ? G ! representatio n ( R ).
Now suppose that there is a light-sensor hooked to our system and we would like to take advantage
of this by adapting the colors of the objects to the ambient light. For example, when the room is
dark, we want to display our objects in a midnight blue theme. Now, we could add following lines:
[ rcvr : G , G < text , ambient_light : dark ] # representatio n ( R ) : G ! set_color ( gray ) ,
[ - ambient_light ] ? G ! representatio n ( R ).
[ rcvr : G , G < box , ambient_light : dark ] # representatio n ( R ) : G ! set_color ( midnight_blue ) ,
[ - ambient_light ] ? G ! representatio n ( R ).
Unfortunately, this will introduce non-determinism if the context contains both an ambient light
and a render type dimension. The reason is that, given a graphical object, the light-sensitive and the
render-type sensitive implementations of representation match the context as none is more specific
than the other. mdp will evaluate both variants, in the order of their definitions.
? - box_prototype ( T ) , [ ambient_light : dark , render_type : svg ] ? T ! representati on ( R ).
R = svg ( shape = box , color = midnight_blue ) ;
R = svg ( shape = box , color = original_colo r ).
We can fix this in two ways: We can either add the render_type dimension to our definitions or
we can assign a higher score to the context:
[ rcvr : G , G < box , ambient_light : dark , render_type : _ ] # representatio n ( R ) : G ! set_color ( midnight_blue ) ,
[ - ambient_light ] ? G ! representatio n ( R ).
% Alternative implementat io n
dimension_ we ig h t ( ambient_light , 2).
9
[ rcvr : G , G < box ,
ambient_ligh t : dark ,
dimension_we i gh t ( ambient_light , W ) @W ] # representati o n ( R ) : G ! set_color ( midnight_blue ) ,
[ - ambient_light ] ? G ! representatio n ( R ).
However, both solutions have the problem that all variants of representation have to be examined
to safely add new variants. Nonetheless, the latter solution of assigning weights to dimension, gives
us more flexibility as it allows us to organize dimension as a separate concern. Nonetheless, the
programmer might forget that there is a weight assignment for a specific dimension and manually
adding a weight annotation to each dimension is too tedious. However, it is straightforward extend
mdp with automatic annotations.
The bigger issue is how the programmer can be assisted in structuring a multidimensional predicate
space. Also, Ungar et al. stress the importance of having equal dimensions (”‘no dimension holds
sway over another”’). Future research and experimentation with larger systems is needed to assess
this problem.
10
3
Implementation
The implementation consists of two parts: an extensible transformer which translates mdp specifications into ordinary Prolog rules and a dispatcher which is responsible for evaluating multidimensional
queries by selecting and evaluating the most specific predicates w.r.t. a given a context.
Both the transformer and the dispatcher are implemented in standard Prolog. The transformer
implements the predicate term expansion/3 which is offered by Prolog systems for rewriting consulted
programs.
3.1
Transforming mdp-Code
For every mdp rule, the transformer generates two parts, an implementation for the predicate and a
signature which contains various meta-data used by the dispatcher to determine which implementation
to select.
For example, consider following mdp rule:
[ debug : P , weight ( debug , D ) @D ] # edge (A , B ) : [ - debug ] ? edge (A , B ) , % Remove debug dimension from implicit context
call (P , (A , B )).
The transformer will generate following code:
mdp_signatur e (
edge /2 ,
mdp_implem en t at i on ( edge /815) ,
% where 815 is a generated identifier to distinguish mdp predicates
% that have the same functor and arity
context_spec (
% Context variable
Ctx ,
% Defined dimensions
[ debug ] ,
% Translated context specification ( used by the dispatcher )
[ ctx_member ( Ctx , debug , P ) , weight ( debug , D )] ,
% Matching - score variables ( summed up by the dispatcher ) ,
[D]
)
).
mdp_implem e nt a ti o n ( edge /815 , Ctx , A , B ) : ctx_member ( Ctx , debug , P ) ,
? ( Ctx , [ - debug ] , edge (A , B ) , % invoke dispatcher
call (P , (A , B )).
The basic transition rules for the transformer and the semantics of mdp are shown in Table 1.
The transformer is applied to every mdp definition and executes within an environment Γ. Γ
comprises four kinds of information: Γcontext , Γdimensions , Γscores and Γrules . Γcontext denotes a
fresh variable which holds the context of an mdp rule, Γdimensions and Γscores are initially empty lists
that are extended by the transition rules to contain the required dimension names and the weights
11
of an mdp rule. The weights are either variables or integer constants. Finally, Γrules contains the
translated code of a context-specification. For readability reasons its derivation is omitted.
The rules can be extended to accommodate for syntactical and semantical enhancements such as
we have seen for object-oriented programming (see Section 2.2). For this purpose, the transformer
provides two hooks that allow the extension of the basic rules.
The first hook hook context rule mdp term(+Context, +Rule, -Term) is applied to every term in
a context specification and can be extended by the programmer to add transformation rules. T erm
can be either ordinary Prolog code or an mdp query.
The second hook hook mdp term(+Context, +Term1, -Term1) is applied after the first hook and
transforms the context specification terms and the bodies of predicates into ordinary prolog terms.
3.2
Predicate Selection and Evaluation
When given a multidimensional query, the dispatcher will manipulate the implicit context (such as
remove or change dimensions) and select the most specific predicates (MSPs).
The MSPs are determined by following steps:
1. Collect the list of all defined mdp rules that match the functor and arity of a query.
2. Add all anonymous rules to the list.
3. Remove all rules which have failed preconditions.
4. Remove all rules which require more dimensions than are present in the implicit context.
5. Determine the score of every rule by counting one for each matched dimension and adding up
all weights.
6. Select all predicates with the highest score.
Listing 1 shows how the dispatcher works in detail.
12
% G i v e n C o n t e x t refers to the e x p l i c i t c o n t e x t
% given in a m u l t i d i m e n s i o n a l query
d i s p a t c h( I m p l i c i t Co nte xt , GivenContext , P r e d i c a t e) : % Update c o n t e x t i n f o r m a t i o n ( remove or update dimensions ,
%
see source d i s t r i b u t i o n for d e t a i l s)
updated_context(
I m p l i c i t C on text , GivenContext , Predicate , U p d a t e d C o n t e x t ) ,
P r e d i c a t e =.. [ _ | A r g u m e n t s] ,
m o s t _ s p e c i f i c _ p r e d i c a t e ( U p d a t e d C onte xt , Predicate , MSP ) ,
e v a l u a t e( MSP , U p d a t e d C onte xt , [ U p d a t e d C o n t e x t | A r g u m e n t s ]).
m o s t _ s p e c i f i c _ p r e d i c a t e ( Ctx , Predicate , MSP ) : f u n c t o r( Predicate , Name , Arity ) ,
f i n d a l l( ScoreKey - MSP ,
(
(
m d p _ s i g n a t u r e ( Name / Arity , MSP , S ) ;
m d p _ s i g n a t u r e ( ’ $ a n o n y m o u s _ r u l e ’/_ , MSP , S )
),
p r e d i c a t e _ s c o r e ( Ctx , S , Score ) ,
S c o r e K e y is - Score
),
M S P B y S c o r e) ,
k e y s o r t( MSPByScore , [ -( Score , _ )| _ ]) ,
member ( Score - MSP , M S P B y S c o r e ).
p r e d i c a t e _ s c o r e ( Context , Signature , Score ) : Signature =
c o n t e x t _ s p e c ( Context , R e q u i r e d D i m en si on s , ContextRules , Vars ) ,
c t x _ k e y s( Context , G i v e n D i m e n s i o n s ) ,
i n t e r s e c t i o n ( R e q u i r e d D i m en sio ns , G i v e n D i m e n s i o n s , R e q u i r e d D i m e n s i o n s ) ,
length ( R e q u i r e d D i m en si ons , N u m M a t c h i n g D i m e n s i o n s ) ,
call ( C o n t e x t R u l e s ) , % E v a l u a t e p r e c o n d i t i o n s
s u m _ l i s t( Vars , V a r S c o r e ) ,
Score is V a r S c o r e + N u m M a t c h i n g D i m e n s i o n s .
Listing 1: Dispatcher Implementation
13
(Phase 1) The following rules are applied to every term in an mdp rule
multidimensional query
term extension
Γ ⊢ Context?P (Args)
Γ ⊢ dispatch(Γcontext , Context, P (Args))
Γ ⊢ T, hook mdp term(Γcontext , T, T ′ )
Γ ⊢ T′
(Phase 2) Translation of mdp rules to single dimensional rules
dimension
Γ ⊢ [Dim : Coord | R] # P (Args) :– Body
Γ, Γdimensions ∪ {Dim} ⊢ R # P (Args) :– ctx member(Γcontext , Dim, Coord), Body
precondition
weight
spec extension
Γ ⊢ [P re(Argsp ) | R] # P (Args) :– Body
Γ ⊢ R # P (Args) :– P re(Argsp ), Body
Γ ⊢ [ @W | R] # P (Args) :– Body
Γ, Γscores ∪ {W } ⊢ R # P (Args) :– Body
Γ ⊢ [SpecP art | R] # P (Args) :– Body
hook context rule mdp term(Γcontext , SpecP art, SpecP art′ )
Γ ⊢ R # P (Args) :– SpecP art′ , Body
(might require re-application
of Phase 1)
rule generation
Γ ⊢ [] # P (Args) :– Body
mdp implementation(P/N ewID, Args, Body)
mdp signature(P/Arity, mdp implementation(P/N ewID),
context(Γcontext , Γdimensions , Γrules , Γscores ))
fact extension
anonymous rule
Γ ⊢ ContextSpec # P (Args)
Γ ⊢ ContextSpec # P (Args) :– true
Γ ⊢ ContextSpec :– Body
Γ ⊢ ContextSpec # ’$anonymous rule’(Args) :– true
Table 1: Overview of mdp transformation rules
14
4
Related Work
Numerous methods and techniques such as Aspect-Oriented Programming (AOP), Subject-oriented
programming [HO93, LRN11] and special programming techniques for one-dimensional systems in
the form of architectural and design patterns have been conceived to handle concerns in adaptive
systems. The problems with these techniques are that they either require the use of special tools and
languages, such as aspect description languages and compilers, or reprogramming of the mechanisms
mdp already provides for every problem instance. We further argue that patterns in particular reduce
the readability of context-sensitive systems since they offer no real separation between specification
and implementation of concerns at the language level.
Korz, on the other hand, has been designed with multidimensionality from the start and firmly
builds on the principles of uniformity and minimality. As such, the system is simple yet very expressive
and readable. However, it has to be pointed out that techniques such as AOP, are in general, more
expressive (but not necessarily as comprehensible) than Korz.
While mdp rests on a Prolog, a single-dimensional language, its unique properties of uniformity,
minimality and homoiconicity provide a flexible foundation for embedding multidimensionality and
experimenting with new features. This allowed us to quickly implement one of Ungar et al.’s suggested
improvements to Korz, selector-based matching (anonymous rules). Also, unlike Korz, multiple
methods (predicates) can be executed if they have the same specificity. We made this design choice
because non-determism is an important and desired feature of Prolog programming and it naturally
extends to multidimensional predicates. Another notable difference is mdp’s reliance on weights. The
initial design did not include weights, but when introducing subtyping, it became apparent that there
must be a kind of weight dimension. Also, we can exploit Prolog’s unification capabilities to perform
pattern matching on dimension coordinates which will be explored in future works.
5
Conclusion and Future Work
We have seen that we can use mdp for many purposes. We can use multidimensional predicates to do
object-oriented programming and to extend a program with additional concerns. Using techniques
such as memoization, we obtain the benefit of having readable and efficient specifications. Yet we
have barely scratched the surface and we will present more scenarios in future works.
In particular, there are following issues that need attention:
• Efficiency of Implementation
A major drawback of mdp is its efficiency. The dispatching mechanism for multidimensional
predicates is very inefficient and has been programmed with quick adaptability and readability
in mind. Its performance is unlikely to be satisfactory in production systems. For satisfactory
performance, a Prolog VM could be extended with special instructions for multidimensional
dispatching.
However, it is not clear at this stage whether the mechanism is sufficiently expressive for largescale Prolog systems. To this end, more scenarios have to be investigated to reveal possible
shortcomings.
• Debugging and Comprehensibility
Currently, there is no support for debugging mdp programs. When debugging an mdp program,
the programmer has to trace the call-chain of the mdp implementation to get information such
as the context and predicate selection. A usable debugger has to hide these implementation
details from the user and present the desired information in an easily accessible way.
Also, the programmer has to be very careful in organizing multidimensional predicates. When
programming in the large-scale, a development environment that helps a programmer organizing
dimensions and point out at possible mistakes is mandatory.
The design of such a development environment with a suitable debugger will be an interesting
challenge.
15
• Weight Dimension and Pattern Matching on Dimension Values
We have seen that in many cases such as implementing subtyping, a weight dimension is necessary to ensure selection of the most specific subtype. However, instead of using annotations, a
cleaner solution would be to explicitly define a weight dimension and use pattern matching to
determine the weights of rules:
[ weight : 1] # p .
[ weight : 2] # p .
? - [ weight : X ] ? p
X = 1 ;
X = 2.
An anonymous could then do weight-based predicate selection. This would further simplify the
matching rules and the dispatcher. However, one problem is that in our example both variants
are executed. In our case this is not desirable, but there might be cases in which we want
to match on dimensions and on arguments of the predicate. Experiments have to provide an
answer which approach is practical.
• Extending the mop
The meta-object protocol we presented is far from complete. For example, we can enrich mop
with syntactical sugar for defining OOP-style properties and context-oriented visibility modifiers. Also, the syntax for matching subtypes is a bit verbose, i.e., [rcvr: OID, OID < Subtype]
can be shortened to [OID < Subtype] (like in Korz). Modifying mop accordingly is straightforward.
• Restructuring Prolog Libraries
mdp gives us many new ways to structure Prolog code and much incentive to restructure existing
code bases. For example, the ordsets-library provides special predicates for manipulating and
querying data types in ordered sets such as ord intersection/3 or ord memberchk/2. With mdp
we can give the programmer a more consistent experience and have overloaded predicate names
such as member and, depending on the context, dynamically choose the predicates to deal with
special data types.
Another interesting experiment would be to rewrite XPCE [WA02], a GUI framework for SWIProlog, using the techniques we have presented in this paper.
• Correctness
Currently, mdp is an experimental stage, so bugs are to be expected even though we have made
sure that the presented examples work. To make mdp ready for production systems, rigorous
testing has to be applied along with formal verification, where we will take advantage of Prolog’s
well-defined language semantics.
We welcome suggestions, bug reports and improvements to the implementation which can be
obtained at: https://bitbucket.org/nexialist/mdp.
References
[CUS95] Bay-Wei Chang, David Ungar, , and Randall B. Smith. Getting Close to Objects: ObjectFocused Programming Environments. In Visual Object-Oriented Programming, Margaret
Burnett, Adele Goldberg, and Ted Lewis, eds., Prentice-Hall, 1995.
[HO93]
William Harrison and Harold Ossher. Subject-oriented programming: A critique of pure
objects. SIGPLAN Not., 28(10):411–428, October 1993.
[LRN11] Daniel Langone, Jorge Ressia, and Oscar Nierstrasz. Unifying subjectivity. In Objects, Models, Components, Patterns - 49th International Conference, TOOLS 2011, Zurich, Switzerland, June 28-30, 2011. Proceedings, pages 115–130, 2011.
16
[UOK14] David Ungar, Harold Ossher, and Doug Kimelman. Korz: Simple, symmetric, subjective,
context-oriented programming. In Proceedings of the 2014 ACM International Symposium
on New Ideas, New Paradigms, and Reflections on Programming & Software, Onward!
2014, pages 113–131, New York, NY, USA, 2014. ACM.
[US87]
David Ungar and Randall B. Smith. Self: The power of simplicity. In Conference Proceedings on Object-oriented Programming Systems, Languages and Applications, OOPSLA ’87,
pages 227–242, New York, NY, USA, 1987. ACM.
[WA02]
Jan Wielemaker and Anjo Anjewierden. Programming in XPCE/Prolog, 2002. Available
online: http://info.ee.pw.edu.pl/Prolog/Download/userguide.pdf, Accessed: 07.03.2016.
17
| 6cs.PL
|
arXiv:1512.06412v3 [stat.ME] 4 Mar 2017
Graphical Model Selection for Gaussian
Conditional Random Fields in the Presence of
Latent Variables
Benjamin Frot∗
Department of Statistics, University of Oxford
and
Luke Jostins
Wellcome Trust for Human Genetics, University of Oxford
and
Gilean McVean
Department of Statistics, University Of Oxford
March 7, 2017
Abstract
We consider the problem of learning a conditional Gaussian graphical model in
the presence of latent variables. Building on recent advances in this field, we suggest
a method that decomposes the parameters of a conditional Markov random field into
the sum of a sparse and a low-rank matrix. We derive convergence bounds for this
estimator and show that it is well-behaved in the high-dimensional regime as well
as “sparsistent” (i.e. capable of recovering the graph structure). We then show how
proximal gradient algorithms and semi-definite programming techniques can be employed to fit the model to thousands of variables. Through extensive simulations, we
BF gratefully acknowledges the EPSRC and Amazon Web Services. LJ is supported by a Wellcome
Trust grant (098759/Z/12/Z) and Christ Church, Oxford. GM is funded by the Wellcome Trust grant
100956/Z/13/Z. This work was jointly supervised by LJ and GM.
∗
1
illustrate the conditions required for identifiability and show that there is a wide range
of situations in which this model performs significantly better than its counterparts,
for example, by accommodating more latent variables. Finally, the suggested method
is applied to two datasets comprising individual level data on genetic variants and
metabolites levels. We show our results replicate better than alternative approaches
and show enriched biological signal.
Keywords: Genetics, Metabolites, Low-Rank plus Sparse, Conditional Markov random
field, Multivariate Analysis.
2
1
Introduction
The task of performing graphical model selection arises in many applications in science
and engineering. There are several factors that make this problem particularly challenging.
First, it is common that only a subset of the relevant variables are observed and estimators
that do not account for hidden variables are therefore prone to confounding. On the other
hand, modelling latent variables is itself difficult because of identifiability and tractability
issues. Second, the number of variables being modelled is often greater than the number of
samples. It is well known that, in such a scaling regime, obtaining a consistent estimator
is usually impossible without making further assumptions about the model, e.g. sparsity
or low-dimensionality. Finally, modelling the joint distribution over all observed variables
is not always relevant. It is sometimes preferable to learn a graphical model over a number
of variables of interest while conditioning on the rest of the collection.
These problems are encountered in many fields of application. In genetics, for example,
one might model a gene expression network conditional on the samples’ combinations of
DNA variants: the variables of interest are the expression levels, while the DNA variants
are included because of their predictive power and capacity to explain some of the observed
correlations between genes (Stearns, 2010). As genotype is not causally influenced by gene
expression levels (i.e. the direction of effect only goes genotype to expression), we would
like to model expression levels conditional on genotype. For another example, consider the
task of modelling stock returns conditional on sentiment analysis data. The variables that
encode sentiment about the stocks have value (Li et al., 2014), but modelling their joint
distribution might be difficult and unnecessary, hence the need for conditioning. Moreover,
a number of unmeasured variables (e.g. energy prices) might impact many stocks and
should be modelled for better predictive accuracy (Chandrasekaran, Parrilo and Willsky,
3
2012).
The problem of learning a Gaussian graphical model in the presence of latent variables
was considered by Chandrasekaran, Parrilo and Willsky (2012). They suggest estimating
an inverse covariance matrix which is the sum of a sparse and a low-rank matrix. Another
partial solution to our problem was introduced independently by Sohn and Kim (2012)
and Wytock and Kolter (2013) who defined the concept of a sparse Gaussian conditional
random field : a regularised maximum-likelihood estimator that learns a Gaussian graphical
model over a subset of the variables (X, say) while conditioning on the remaining variables
(Z, say).
Chandrasekaran, Parrilo and Willsky (2012), Sohn and Kim (2012) and Wytock and Kolter
(2013) made significant advances to the problem of model selection in general graphical models, but there exist many situations where we may wish to allow for latent variables and condition on some of those measured. Here, we suggest learning a Gaussian
conditional random field in the presence of latent variables and introduce a novel regularised maximum-likelihood estimator which fits into the “low-rank plus sparse” framework
(Chandrasekaran et al., 2009; Candès et al., 2011). In our setting, inputs (variables in Z)
are allowed to act on the outputs (X) in both a sparse and a low-rank fashion, while the
inverse covariance matrix over X is estimated conditional on Z and on the marginalised
latent variables. As will be shown later, this approach allows us to correctly recover graphs
that are typically denser and with more hidden variables than the ones that can be handled
by other methods.
From both a theoretical and a computational point of view, modelling latent variables
with a conditional random field gives rise to a number of complications (e.g. the proximal
operator is not defined in closed form) that we address in this paper. In particular, we derive
convergence bounds for our estimator and show that under suitable identifiability conditions
4
it is consistent in the high-dimensional regime as well as “sparsistent” (i.e. capable of
recovering the graph structure). We then show how the alternating direction method of
multipliers (Boyd, 2010) and semi-definite programming techniques can be employed to
fit the model to thousands of variables. Through extensive simulations, we illustrate the
conditions required for identifiability and show that there is a wide range of situations in
which this model performs significantly better than its counterparts. In order to show how
our model behaves in a realistic setting, we apply the present estimator to two datasets
comprising genetic variants and metabolite levels. Both replication and a test statistic
constructed using an independent source of validation suggest that our estimates have
more biological relevance than the results obtained via other methods.
2
2.1
Problem Statement
Setup
Throughout, we consider n independent, identically distributed realisations of a zero-mean
random vector Y ∈ Rm+p+h . Y is indexed by disjoint subsets of {1, . . . , m + p + h}, denoted
Z, X, H and with respective cardinality m, p and h. They correspond to the variables we
wish to condition on, the variables we wish to model and the hidden variables. We write
YZ (resp. YX and YH ) for the subvector
of Y indexed by Z (resp. X and H). Our main
YX
assumption is that the distribution of ∈ Rp+h conditional on YZ ∈ Rm is normal
YH
and that its mean is a linear combination of the inputs YZ . More precisely, we assume a
Gaussian conditional random field parametrised as follows:
−1
−1
∗
∗
∗ T
∗
∗
MX
MXH
M
MX
MXH
Y
ZX YZ ,
X |YZ ∼ N −
,
∗ T
∗ T
∗ T
MZH
MXH
MH∗
MXH
MH∗
YH
5
where we have used partitioned matrices to show the contributions of the observed and
∗
∗
hidden variables. Thus, MX∗ ∈ Rp×p , MZX
∈ Rm×p , MXH
∈ Rp×h , . . .. The superscript −∗
is used to indicate that these matrices are nominal parameters of the model, as opposed to
estimates. Note that there are no distributional assumptions about YZ .
Finally, we assume that variables indexed by H are unobserved. Accordingly, we compute the marginal distribution YX |YZ , which yields
∗
∗ T
∗
YX |YZ ∼ N − (SX
− L∗X )−1 SZX
− L∗ZX T YZ , (SX
− L∗X )−1 ,
(2.1)
∗
∗
∗ T
∗
∗
where we have defined SX
, MX∗ , L∗X , MXH
MX∗ −1 MXH
, SZX
, MZX
and L∗ZX ,
∗
∗ T
MZH
MH∗ −1 MXH
. This expression follows straightforwardly from the formula for the in-
verse of a partitioned matrix (the full derivation is given in the supplementary materials). From Equation (2.1), the log-likelihood function can be expressed in terms of the
P
P
sample covariance matrices ΣnZ , n1 i (YZ )i (YZ )Ti , ΣnX , n1 i (YX )i (YX )Ti and ΣnZX ,
P
1
T
i (YZ )i (YX )i :
n
ℓ(SX , LX , SZX , LZX ; ΣnZ , ΣnX , ΣnZX ) = log det (SX − LX ) − T r (ΣnX (SX − LX )) −
2T r ΣnZX (SZX − LZX )T − T r ((SX − LX )−1 (SZX − LZX )T ΣnZ (SZX − LZX ) . (2.2)
For clarity, all terms related to a given subset will be dropped from the expression when
the subset is empty. For example, whenever Z = H = ∅ the log-likelihood becomes
ℓ(SX ; ΣX ) = log det SX − T r(ΣnX SX ).
Note that our assumption about the Gaussianity of X, H plays an important role in the
∗
interpretation of the nominal parameters (MX∗ , MXH
, . . .). Under this assumption, it is well
known that the structure of the conditional Gaussian graphical model (GGM) over X, H
can be read-off these matrices directly by looking at the location of their non-zero entries
(Lauritzen, 1996). Briefly, a graphical model is a statistical model defined according to a
6
graph whose nodes are random variables and whose edges encode conditional independence
statements between variables (Lauritzen, 1996). Thus, (MX∗ )i,j = (MX∗ )j,i = 0 if and
∗
only if Xi ⊥
⊥ Xj |Z, X \ {Xi , Xj }, H. Likewise, (MZH
)i,j = 0 if and only if Zi ⊥
⊥ Hj |Z \
Zi , X, H \ {Hj }. Note that since the conditional mean vector is a linear transformation of
∗
∗
YZ , this interpretation of the non-zero entries of MZH
and MZX
holds irrespective of YZ ’s
distribution.
2.2
Goal
∗
In typical applications such as the ones mentioned in the introduction, SX
is the target.
∗
Since it encodes the structure of the graphical model over X, recovering SX
can provide
insight into the causal mechanisms underpinning the data but, in general, hidden variables
make it impossible to access this parameter directly. Instead, it follows from Equations
∗
∗
(2.1) and (2.2) that only the sum SX
− L∗X can be inferred (similarly, only SZX
− L∗ZX
is accessible). The maximiser of the log-likelihood (2.2) is not unique and the problem is
fundamentally misspecified.
We are therefore facing two related, but distinct, problems:
• identifiability: under which conditions does the problem admit a unique solution?
Ideally, these conditions ought to be as broad as possible so that they will be met in
realistic situations. Notice that unlike the breakdown caused by the high-dimensional
regime, this kind of non-identifiability is more fundamental and remains no matter
how large the number of samples.
• consistency: provided there exists a unique solution, can we derive a consistent,
∗
∗
tractable estimator which is capable of recovering (SX
, L∗X , SZX
, L∗ZX )?
7
∗
Here, we chose to focus on SX
because it fits our application but there might be situa∗
tions in which other parameters are of interest, e.g. SZX
in Zhang and Kim (2014).
2.3
Previous Work
In practice, model selection in the context of GGMs is often performed using ℓ1 -regularised
maximum likelihood estimators (MLEs) such as the ones introduced in Banerjee et al.
(2008); Yuan and Lin (2007), and the so-called graphical lasso (Friedman et al., 2008).
The ℓ1 -norm is the convex envelope of the ℓ0 unit ball and is therefore a natural convex relaxation to learn sparse matrices. Building on the success of the graphical lasso, estimators
of the form “log-likelihood” + “non-Euclidian convex penalty” have received considerable
interest (Chandrasekaran, Recht, Parrilo and Willsky, 2012). A relevant example is the
use of the nuclear norm (i.e. the sum of the singular values) as a convex relaxation for
learning low-rank models (Bach, 2008). Beyond their attractive computational properties,
the ℓ1 and nuclear norm regularised MLEs enjoy strong theoretical guarantees (Bach, 2008;
Ravikumar et al., 2011).
Using penalised MLEs, the questions raised above (Section 2.2) have been solved in
some special cases of model (2.1).
Sparse Gaussian Conditional Markov Random Field: H = ∅
When H is empty, (2.1) reduces to
∗ −1 ∗ T
∗ −1
YX |YZ ∼ N −SX
SZX YZ , SX
.
The log-likelihood associated with this model is convex and maximum-likelihood estimates
can be obtained in closed form. In order to increase the interpretability of the estimates and
8
cope with high-dimensionality, Sohn and Kim (2012); Wytock and Kolter (2013) suggested
∗
∗
the following estimator of (SX
, SZX
):
(ŜX , ŜZX ) =
−ℓ(SX , SZX ; ΣnZ , ΣnX ) + λn (||SX ||1 + ||SZX ||1 ),
arg min
SX
∈Rp×p ,S
ZX
∈Rm×p ,S
X ≻0
with λn > 0. The entries of both SX and SZX are being shrunk in order to jointly learn
a pair of sparse matrices describing the direct effects of Z on X and the graph over X.
Wytock and Kolter (2013) studied the theoretical properties of this estimator and derived
∗
∗
a set of sufficient conditions for the correct recovery of SX
and SZX
. Among other results,
they showed that this approach often outperforms the graphical lasso in terms of predictive
power and model selection accuracy. Alternative parametrisations and approaches have
been suggested in the multivariate linear regression literature. We refer the reader to
Yin and Li (2011); Sohn and Kim (2012) and references therein for more details on these
estimators and their relative performances.
Low-Rank Plus Sparse Decomposition: Z = ∅
The presence of latent variables (H 6= ∅) is a substantial complication. As explained earlier,
∗
the marginal precision SX
− L∗X is then the sum of two matrices and the problem is fun-
damentally misspecified. However, following the seminal work of Candès et al. (2011) and
Chandrasekaran et al. (2009), Chandrasekaran, Parrilo and Willsky (2012) showed that it
∗
is sometimes possible to correctly decompose SX
−L∗X into its summands. Loosely speaking,
∗
this is the case if SX
is sparse and there are few hidden variables with an effect spread over
most of the observed variables. As a result, Chandrasekaran, Parrilo and Willsky (2012)
introduced an estimator which penalises the ℓ1 -norm of SX and the nuclear norm of LX as
follows:
(ŜX , L̂X ) = arg min −ℓ(SX , LX ; ΣX ) + λn (γ||SX ||1 + ||LX ||∗ ),
SX ,LX ∈Rp×p
9
(2.3)
subject to SX − LX ≻ 0, LX 0., with λn , γ > 0. Here ||LX ||∗ denotes the nuclear norm of
LX (i.e. the sum of its singular values). Among other useful results, Chandrasekaran, Parrilo and Willsky
(2012) showed that this estimator is, under suitable conditions, sparsistent and “ranksistent”: the sign patterns of both the entries of S and the spectrum of L can be recovered
exactly.
2.4
Suggested Estimator
As hinted in the introduction, there are many cases where one might want to both condition
and allow for latent variables. In such cases, neither the sparse Gaussian conditional Markov
random field nor the low-rank plus sparse approach would be optimal. Building on these
estimators, we propose decomposing the parameters of a Gaussian conditional Markov
random field into the sum of a low-rank and a sparse matrix. To that end, we suggest
optimising the following regularised MLE
(ŜX , L̂X , ŜZX , L̂ZX ) =
arg min
SX ,LX
∈Rp×p ;S
ZX ,LZX
−ℓ(SX , LX , SZX , LZX ; ΣnZ , ΣnX , ΣnZX ) + λn (γ kSk1 + kLk∗ )
L
SX
, L = X . (2.4)
s.t. SX − LX ≻ 0, LX 0 and S =
LZX
SZX
∈Rm×p
Solving (2.4) amounts to minimising a function which is jointly convex in its parameters
over a convex constraint set (proofs are in the supplementary materials, along with other
elementary properties of the likelihood). As mentioned earlier, this likelihood is structured
around two parameters, SZX and SX , accounting respectively for the direct (i.e. conditional
on all variables) effects of Z on X and the structure of the graph over X. However, because
we penalise the rank of L, the effect of all latent variables is modelled jointly and a single
10
set of latent factors is learned. No distinction is being made between the variables that
“mediate” the action of Z and the ones that act as confounders on X. On the other hand,
the parameters SX and SZX retain their interpretability.
3
Theoretical Analysis
According to our assumptions, we assume here that each sample is generated according to
the model
∗
∗
∗
YX |YZ ∼ N −(SX
− L∗X )−1 (SZX
− L∗ZX )T YZ , (SX
− L∗X )−1 ,
(3.1)
and ask under what circumstances Estimator (2.4) correctly recovers the parameters S ∗ , L∗
∗
∗
(as built by stacking SX
, SZX
and L∗X , L∗ZX ) with overwhelming probability.
We analyse this problem in the framework of Chandrasekaran, Parrilo and Willsky
(2012) and therefore our proofs often mirror theirs. However, because of the form taken
by the likelihood and because we do not limit ourselves to square matrices, the analysis is
significantly more involved.
As mentioned earlier, modelling latent variables by decomposing the parameters into
a sum of two matrices raises identifiability issues: given samples drawn from (3.1), when
is it possible to exactly decompose the sum S − L (where S, L are defined as before)
into its summands? This is a problem which has been tackled in great generality in
Chandrasekaran, Parrilo and Willsky (2012) and their results directly apply to the present
situation: they are expressed in terms of the Fisher Information Matrix but do not explicitly involve the likelihood function. For that reason, key definitions, as well as assumptions
necessary for our result to hold, are deferred to the supplementary materials. Here we focus
on the original contributions of this paper by giving an intuition for these conditions before
formally stating the consistency of the estimator defined by (2.4).
11
3.1
Identifiability
Until now, it was repeatedly mentioned that a “low-rank plus sparse decomposition” is
possible when S is sparse and L is low-rank. However, it is clear that imposing conditions on the sparsity of S and the rank of L is not sufficient. For example, consider a
matrix with a single entry: it is at the same time sparse and low-rank and there is, therefore, no unique way of decomposing it into the sum of a low-rank and a sparse matrix.
Chandrasekaran et al. (2009) introduce the notion of rank-sparsity incoherence and define
quantities that make it possible to express the conditions under which such a problem is
well-posed, even for arbitrary matrices. Two concepts are particularly important (precise
mathematical statements and explanations can be found in the supplementary materials):
• ξ(T (L∗)): a small ξ(T (L∗ )) guarantees that no single latent variable will have a strong
effect on only a small set of the observed variables. It is closely related to the concept
of incoherence introduced in Candès et al. (2011).
• µ(Ω(S ∗ )) quantifies the diffusivity of S’s spectrum. It can be shown that matrices
with few non-zero entries per row/column (and thus sparse) have a small µ.
A sufficient condition for identifiability can be expressed in terms of ξ, µ by requiring that
their product be small enough (ξ(T (L∗ ))µ(Ω(S ∗ )) ≤ 16 C 2 ) and that the tuning parameter
i
h
∗
C
γ be chosen within a given range (γ ∈ 3ξ(TC(L )) , 2µ(Ω(S
∗ )) ), for some constant C which
depends on the Fisher Information Matrix (FIM). In other words, there must be a small
number of latent variables acting on many observed ones and S ∗ must not have too many
non-zero entries in any given row or column. This is a condition on the nominal parameters
S ∗ , L∗ and it is related to the problem of decomposing the sum of two matrices. Moreover,
it can be shown that natural classes of matrices satisfy these assumptions. In particular,
the degree of S ∗ and number of latent variables h are allowed to increase as a function of
12
the problem size p, m (Chandrasekaran et al., 2009). We call these restrictions on ξ, µ and
γ Assumption 1.
Another issue is that one does not directly observe S ∗ − L∗ but samples generated from
(3.1). All lasso-type methods face this problem and conditions on the FIM are usually
imposed (the so-called irrepresentability condition) (Ravikumar et al., 2011). Similar assumptions about the FIM are made here and detailed in the supplementary materials. This
is Assumption 2.
3.2
Consistency
We can now present our main result and state the consistency of Estimator (2.4) (see
supplementary materials for the proof). First, let us recall that for any matrix P , ||P ||2
denotes its largest singular value and ||P ||∞ is its largest entry in magnitude. We can then
define the following quantities:
∗
∗
∗
ψZ = ||ΣnZ ||2, ψX
= ||(SX
− L∗X )−1 ||2 , φ∗ZX = ||SZX
− L∗ZX ||2 ,
s
2
3 ∗
9 ∗ ∗
1 φ∗ZX Q2
ψZ
ψ = ψX 1 + 2 ∗ 1 + ψX φZX , W = Q1 min
, ∗ 2 .
,
∗
2
ψX
4
6ψX
4 ψX
ψ
q
q
∗ 2 pM
256ψX
Q3
ψZ
m 2
.
Finally, for M = max 1, 4ψ∗ (1 + p ) , let λn = ξ(T (L∗ ))
n
X
We prove the following theorem in the supplementary materials (Q1 to Q6 are constants
whose definitions are deferred for clarity):
Theorem 1. (Algebraic Consistency)
Suppose that Assumptions 1 and 2 hold and that we are given n samples drawn according
to (3.1). Further assume that the following hold:
∗ 2
256ψX
(a) n ≥ ξ(TpM
max
2,
.
(L∗ ))4
W2
13
(b) (σmin condition) Let the minimum non-zero singular value σ of L∗ be such that
σ≥
Q4 λn
.
ξ(T (L∗ ))2
(c) (θmin condition) Let the minimum magnitude non-zero entry θ of S ∗ be such that
θ≥
Q5 λn
.
µ(Ω(S ∗ ))
ψ∗
Then, with probability greater than 1 − max 2 exp(−pM), exp(−4 ψXZ pM) we have
1. sign(Ŝ) = sign(S ∗ ) and rank(L̂) = rank(L∗ ).
2.
max
1
Ŝ − S ∗
γ
∞
, L̂ − L∗
2
∗
Q6 ψX
≤
ξ(T (L∗ ))
r
pM
.
n
Seen at a high-level, Theorem 1 is analogous to the result obtained by Chandrasekaran, Parrilo and Will
(2012) for the low-rank plus sparse (LR+S) estimator. In particular, this result holds when
both the dimension of the problem (parametrised by m and p) and the number of samples
(n) grow. Moreover, through their dependencies on ξ and µ, both the degree of S and the
rank of L are allowed to scale with n, which is essential to study connected graphs (see
(Chandrasekaran et al., 2009) for examples of scaling regimes). To compare these results
further, a few points are worth considering.
First, we do not make any distributional assumptions about YZ and there are therefore
many scenarios in which only Theorem 1 applies.
For the sake of comparison, we can assume that YZ follows a normal distribution so that
the consistency theorem of LR+S is applicable. Since LR+S does not model conditional
distributions, Z and X are modelled jointly. The estimated matrices, (ŜLR+S , L̂LR+S ), are
of size (p+m)×(p+m) and, to obtain ŜX , ŜZX , . . ., the relevant sub-matrices are extracted
14
from the larger (p + m) × (p + m) estimates. Then, the algebraic consistency of LR+S
holds with a probability πLR+S of at least 1 − 2 exp(−(p + m)) provided the number of
samples nLR+S satisfies nLR+S &
p+m
ξ(T (L∗LR+S ))4
(Chandrasekaran, Parrilo and Willsky, 2012,
Theorem 4.1). Now, note that irrespective of the value of M, our convergence regimes
are very similar since we require n &
pM
ξ(T (L∗ ))4
for consistency to hold with a probability π
satisfying π ≥ 1 − c exp(−pM), for some c. Since we have both nLR+S 6= n and πLR+S 6= π,
a direct comparison is not obvious. There are special cases in which it is easier. For
example, when M = 1 one proves a result similar to Corollary X (suppl. materials) showing
that n &
p+m
ξ(T (L∗ ))4
is required for π to be at least 1 − 2 exp(−(p + m)), thus recovering a
convergence rate similar to LR+S’s.
Finally, µ and ξ play an identical role in both Theorem 1 and (Chandrasekaran, Parrilo and Willsky,
2012, Theorem 4.1), namely through Assumption 1 and conditions a), b) and c). How∗
ever, these quantities are usually different (i.e. µ(Ω(S ∗ )) 6= µ(Ω(SLR+S
)), ξ(T (L∗ )) 6=
ξ(T (L∗LR+S ))), which has interesting implications. An obvious consequence is the one stated
in the previous section: since µ, ξ define the acceptable range for γ, its span can vary widely
across methods. More importantly, one shows that conditions a), b) and c) are driven by
the lower-end of that range. Should it be assumed instead that γ =
C
2µ(Ω(S ∗ ))
(the upper-
end), all three conditions would be relaxed (Chandrasekaran, Parrilo and Willsky, 2012,
Corollary 4.2). Thus, the smaller the value of ξ(T (L∗ )), the wider the acceptable range
and the more likely Theorem 1 is to hold.
4
Optimisation
Optimising (2.4) in the high-dimensional setting is a challenging problem. For example,
some of the constraints are hard to accommodate (e.g. SX − LX ≻ 0, LX 0) and the
15
penalty terms are non-smooth. Fortunately, (2.4) has similarities with (2.3) (the estimator
of Chandrasekaran, Parrilo and Willsky (2012)) and we can rely on algorithms that have
proven effective on (2.3), namely the Alternative Direction Method of Multipliers (ADMM)
(Boyd, 2010; Ma et al., 2013; Ye et al., 2011) and approaches relying on Semi-Definite
Programming (SDP) (Vandenberghe and Boyd, 1996; Wang et al., 2010; Tütüncü et al.,
2003). The general theory behind both ADMM and SDP is applicable to the problem at
hand but features that are specific to (2.4) prevent a straightforward application of existing
algorithms. SDP is an active field of research and recasting (2.4) within that framework
makes it easier for the reader to use existing software and even benefit from future advances
in that field. On the other hand, our ADMM implementation is tailored to the problem
at hand but converges to a reasonable accuracy quickly. This is why we discuss both
strategies. Technical details and step-by-step derivations are given in the supplementary
materials.
4.1
The Alternating Direction Method of Multipliers
The Alternating Direction Method of Multipliers (ADMM) is a first-order optimisation
procedure which is well-suited to the minimisation of large-scale convex functions. It proceeds by decomposing the original problem into more amenable subproblems which are then
solved iteratively (Boyd, 2010). It is sometimes possible to obtain closed-form solutions
to these subproblems but this is not required for ADMM to converge: even inexact iterative methods can be employed (Eckstein and Bertsekas, 1992; Goldstein and Osher, 2009).
Morover, only a few tens of iterations are necessary for ADMM to converge to an accuracy
which is sufficient for most applications1 (Boyd, 2010). ADMM (and related algorithms such
as Bregman iterations and Douglas–Rachford splitting) has been celebrated as an efficient
1
However, converging to a very high accuracy can be slow in comparison to second-order methods.
16
and robust general-purpose algorithm for ℓ1 -regularised problems (Goldstein and Osher,
2009).
More recently, (Ye et al., 2011; Ma et al., 2013) used ADMM to solve (2.3) and showed
that it can be optimised by iteratively solving four smaller subproblems (Ye et al., 2011).
A similar decomposition is applicable to the problem at hand but, in the case of (2.4), one
of the subproblems requires the computation of a so-called proximal operator which does
not admit a closed-form solution. Consequently, we derived an algorithm which iteratively
converges to this proximal operator. In practice, we found that only a few iterations
(typically less than 10) of this subprocedure are necessary to obtain a good approximation
to the proximal operator.
4.2
Recasting the objective function as a Semi-Definite Program
The solvers made available in the MATLAB R packages SDPT3 and Logdet-PPA are capable of solving problems of the form (Tütüncü et al., 2003; Wang et al., 2010):
arg min T r(X1C1T ) + T r(X2 C2T ) + . . . + a1 log det(X1 )
(4.1)
X1 ,X2 ,...
subject to a number of linear, quadratic and positive semi-definite constraints 2 . Our goal is
then to recast (2.4) as a problem of the same form as (4.1). We show in the supplementary
2
This is only a subset of the problems that can be tackled by such packages. See references for a
formulation of this problem in its full generality.
17
materials that (2.4) admits the following SDP reformulation:
1
n
T
arg min
T r(KΣO ) − log det SX + λn γ1 F 1 + (T r(H1) + T r(H2))
2
SX ,LX ,SZX ,LZX ,W,F,H1 ,H2
H1 L
0, − Fij ≤ Sij ≤ Fij , ∀i, j;
subject to K 0, SX ≻ 0, LX 0,
T
L H2
SX
LX
W
SZX − LZX
, S =
, L =
. (4.2)
where K =
T
T
SZX
LZX
SZX − LZX SX − LX
(4.2) can easily be implemented in e.g. YALMIP and solved using LogdetPPA or
SDPT3 (Löfberg, 2004; Wang et al., 2010; Tütüncü et al., 2003). We remark that the
objective function is now smooth (as opposed to (2.4)) but contains many more variables
and constraints.
5
Simulations
We now study the properties of the proposed model on synthetic data and compare its
performances to the three other methods introduced earlier: the graphical lasso (GLASSO)
(Friedman et al., 2008), the sparse conditional Gaussian graphical model (SCGGM) (Sohn and Kim,
2012; Zhang and Kim, 2014; Wytock and Kolter, 2013) and the low-rank plus sparse decomposition (LR+S) (Chandrasekaran, Parrilo and Willsky, 2012). The suggested approach
will henceforth be referred to as LSCGGM (i.e. Latent Sparse Conditional Gaussian
Graphical Model).
In Section 3, it was established that assumptions about both the nominal parameters
(S ∗ , L∗ ) and the Fisher information matrix are necessary to guarantee the identifiability of
the problem and, subsequently, the applicability of Theorem 1. In particular, we recalled
the key role played by the maximum degree of S ∗ and the incoherence of L∗ . To better
18
understand when these assumptions are expected to hold, we simulate data from a set of
graphical models that span the range of possible latent structures and measure the ability
of the different methods to recover the underlying graphs.
5.1
Graphical Structures and Methods
The set of graphical structures we simulate from is constructed in such a way that only two
integers, dZ and dH , describe the relevant properties (rank, sparsity, incoherence, degree)
∗
of SZX
− L∗ZX and L∗X , respectively. Thus, dZ controls the relationship between inputs (Z)
∗
and outputs (X) while dH encodes the behaviour of L∗X . The remaining parameter, SX
,
remains unchanged throughout. We now briefly describe how the graphs are constructed
but defer technical details to the supplementary materials (e.g. distribution of effect sizes).
The code used to generate the data and fit our model is made available with this paper.
For all simulations, each observation is generated according to a model of the form
−1
−1
∗
∗
∗ T
∗
∗
Y
S
MXH
M
S
MXH
X |YZ ∼ N − X
ZX YZ , X
,
YH
M∗ T M∗
0
M∗ T M∗
XH
H
XH
H
with YZ a random vector of size p whose entries are drawn independently from a tdistribution with 4 degrees of freedom. YX is also of size p. Here, YX and YH are drawn
jointly from a conditional random Markov field but only YX and YZ are observed, which
∗
∗ T
∗
∗
implies that L∗X = MXH
MH∗ −1 MXH
. The matrices SX
, L∗X and MZX
are constructed as
follows.
∗
The non-zero pattern of the p × p matrix SX
is identical across all simulations and is
similar to the one adopted by Wytock and Kolter (2013): the graph over X is a chain of p
∗
variables in which one link out of five has been removed. The non-diagonal entries of SX
∗
are such that SX
ij 6= 0, if and only if i = j + 1 and i 6≡ 0 (mod 5).
19
As stated above, the rank/sparsity of L∗X is described by a single integer, dH . Specifically, we assume that p is an integer of the form p = 2k and pick dH ∈ {0, 1, . . . , k}. Then,
∗
for a fixed value of dH , MH∗ and MZX
are random matrices constructed so that: a) there
are exactly 2dH confounders, i.e. the rank of L∗X is 2dH ; b) each of the 2dH confounders
impacts exactly p/2dH outputs; c) each output is connected to exactly one latent variable.
Thus, when dH = k, there is effectively no confounding since latent variables and outputs
are in a one-to-one correspondence. When dH is much smaller than k, there are few confounders with an effect spread over many observed variables. When dH is set close to k,
there are many hidden variables, each affecting a handful of outputs – a gross violation of
the identifiability assumptions.
∗
Likewise, dZ accounts for the structure of MZX
. Here again, we assume p = 2k and
∗
∗
pick dZ ∈ {0, 1, . . . , k}. Then, MZX
is designed to satisfy: a) rk(MZX
) = 2dZ ; b) each
∗
row/column of MZX
has exactly p/2dZ non-zero entries. The effect of dZ is easily interpreted.
For example, dZ = k is an ideal situation where inputs and output are in a one-to-one
∗
correspondence. As dZ goes from k to 0, MZX
becomes denser and increasingly incoherent.
∗
When dZ is close to k, MZX
is estimated as a sparse matrix. When dZ is small, its
decomposition is a single low-rank matrix.
Finally, since neither GLASSO nor LR+S model conditional distributions, we use these
estimators as described in Section 3, i.e. by first modelling Z and X jointly and then
extracting submatrices of the estimates.
5.2
Results
In our simulations, we set p = 32, n = 3000 and let (dZ , dH ) take values in {2, 3, 4, 5}2.
Each of these 16 designs is replicated 20 times, for a total of 320 distinct datasets.
∗
Here, we are interested in recovering the structure of SX
and we use precision/recall
20
curves as a metric, thus ignoring the rank of the latent component. LR+S and LSCGGM
both have two tuning parameters (λ and γ). For each value of γ, one obtains a distinct
precision/recall curve by varying λ. For each of the 320 simulated datasets, we computed
the paths corresponding to 15 distinct values of γ and subsequently selected γ so as to
maximise the Area Under the Curve (AUC). Figure 1 shows the average precision/recall
curves obtained by applying this procedure.
First, we see that known methods behave as expected: GLASSO behaves best when
there is no confounding and Z acts in a sparse fashion (dH = dZ = 5) ; SCGGM is
more robust to changes in dZ , but this is restricted to situations in which there is not
confounding (dH = 5); LR+S performs best when dH = 5 or when there is low-rank, diffuse
confounding (dH = 2). In a number of cases, the method proposed here is better than any
of the alternative methods and, in the worst cases, it offers comparable performances.
Specifically, it outperforms LR+S significantly when both inputs and hidden variables act
on the outputs through a relatively low-rank mechanism (dZ = 2, 3; dH = 2, 3). Two factors
might explain this behaviour: a) the inputs are not normally distributed, which violates the
assumptions of LR+S; b) the data is generated according to a conditional random Markov
field, which is not assumed by LR+S, and may result in a violation of its identifiability
assumptions.
dH = 4 corresponds to the extreme situation in which each latent variable confounds
exactly two random variables. None of the methods performs well but LR+S and LSCGGM
behave better than GLASSO and SCGGM in scenarios where one would not expect any
differences (e.g. dZ = dH = 5). This is because LR+S and LSCGGM have two tuning
parameters, one of which (γ) is chosen with perfect knowledge: it improves the AUC of
these methods but causes L̂X to be non-zero. Additional simulations made available in
the supplementary materials show that when γ is chosen with cross-validation, the selected
21
dZ = 2
dZ = 3
dZ = 4
dZ = 5
1.00
dH = 2
0.75
0.50
0.25
1.00
dH = 3
0.75
0.25
1.00
0.75
dH = 4
Precision
0.50
0.50
0.25
1.00
dH = 5
0.75
0.50
0.25
0.2 0.4 0.6 0.8 1.0
0.2 0.4 0.6 0.8 1.0
0.2 0.4 0.6 0.8 1.0
0.2 0.4 0.6 0.8 1.0
Recall
Method
GLASSO
SCGGM
LR+S
LSCGGM
Figure 1: Comparison of the suggested estimator (LSCGGM) to other published methods.
Along the x-axis (resp y-axis), dZ (resp dH ) varies from 2 to 5. More precisely, in the
bottom row (dH = 5), there is no confounding at all. In the second row from the bottom
(dH = 4), hidden variables act in a very sparse fashion. In the top row (dH = 2), there
are 4 hidden variables and we are in the range of applicability of the low-rank plus sparse
method. The second row (dH = 3) corresponds to an intermediate regime in which there
are 8 latent variables. Settings: p = 32, n = 3000. For each dataset, the value of the
tuning parameter γ was chosen so as to maximise the Area Under the Curve (AUC). Each
of the 16 designs is repeated 20 times. We report average precisions at fixed recalls of
{0.1, 0.2, . . . , 1}.
22
value of γ is indeed often too small.
Both LR+S and LSCGGM have two tuning parameters (λ, γ): λ controls the overall shrinkage on the sparsity/rank of the estimates, γ accounts for the trade-off between
sparse and low-rank components. To better understand the role of γ, we look at the precision/recall curves obtained for various values of this tuning parameter. As suggested in
Chandrasekaran et al. (2009), the penalty term is reformulated as λ(γ||S||1 + (1 − γ)||L||∗ )
with γ ranging from 0 to 1 instead of (0, +∞). By analogy to the AUC metric, we report
the “Volume Under the Surface” (VUS) which accounts for the effect of both regularisation
parameters.
In Figure 2, the surfaces obtained for (dH = dZ = 2) and (dH = 5, dZ = 3) are plotted.
They show that the suggested approach is less sensitive to γ than LR+S, thus making
it easier to pick a sensible value in real-world applications. Figure 2 b) illustrates what
happens when both methods offer comparable performances according to Figure 1 (which
is obtained by choosing γ perfectly): compared to LSCGGM, there are actually very few
values of γ for which LR+S achieves its best AUC. Here, only two of the 16 possible surface
plots are shown, but Figure 2 c) indicates that LSCGGM is less sensitive to this tuning
parameter across all simulation designs, as measured by the VUS. In particular, we have
consistently observed that upper-end of the acceptable range for γ is higher for LSCGGM
than LR+S. The next simulations illustrate the implications of this property.
In these simulations, our main concern was to illustrate how methods differ in terms
of identifiability and consistency. Setting p and m to a relatively small value (32) made it
possible to capture most scenarios with only 16 graphical structures. In the supplementary
materials, we simulate from larger graphs (p = m = 27 = 128, n = 3000) and obtain results
that are similar to the ones showed here. We also report the estimation errors for the other
∗
parameters L∗ , L∗X along with the precision/recall curves for SZX
. Finally, we look at the
23
effect of choosing γ using cross-validation. In the next section, we show how one can select
λ and γ when some control over the number of falsely discovered edges is expected.
a)
b)
1
Precision
Precision
1
0.5
1
0.5
1
0
0
0
0
Recall
0.8
c)
Volume under Surface
Gamma
Recall
0.8
0
1
Gamma
dZ = 2
dZ = 3
dZ = 4
1
0
dZ = 5
1.00
0.75
0.50
0.25
0.00
2
3
4
5
2
3
4
5
2
3
4
5
2
3
4
5
dH
Method
LR+S
LSCGGM
Figure 2: Sensitivity to the tuning parameter γ. Here, an alternative parametrisation of
the regularisation term is used: λn (γ||S||1 + (1 − γ)||L||∗), so that γ ∈ (0, 1) instead of
(0, +∞). a) Precision/recall surface for dZ = dH = 2 (i.e. each input acts on 8 random
outputs and there are 4 confounding variables). b) Precision/recall surface for dZ = 3 and
dH = 5 (there are no confounders, each input acts on 4 random outputs). c) Volume under
surface across all 16 simulation designs.
24
6
Application: Using genetic information to detect relationships between human metabolites
To illustrate the value of our new approach, we now apply it to a dataset combining human
metabolite levels and genetic markers. Here, metabolites play the role of the variables
indexed by X while genetic variants are the inputs, Z. For comparison purposes, we also
report the results obtained with the Low-Rank plus Sparse method (LR+S)3 .
6.1
The Avon Longitudinal Study of Parents and Children (ALSPAC)
The Avon Longitudinal Study of Parents and Children (ALSPAC) is a cohort study of
children born in the county of Avon during 1991 and 1992 (Boyd et al., 2012; Fraser et al.,
2012). More details about this study and data preparation are available in the supplementary materials. Here, only key features of this dataset are reported4 .
The data at our disposal contains genetic and phenotypic measurements on approximately 8,000 children and their mothers. We first performed our entire analysis on the
children’s cohort (called "Child cohort" throughout) and then independently applied the
same procedure to the mothers’ cohort (Mother cohort). We modelled the levels of 39
metabolites. Measurements for all 39 variables were available without missing data for
5,242 children and 2,770 mothers. In each cohort, independent genetic variants were selected based on their predictive power with respect to any of the 39 traits under study: 133
3
The other two methods (graphical lasso and sparse conditional graphical model) arise as special cases
by setting γ close to 0.
4
Please note that the study website contains details of all the data that is available through a fully searchable data dictionary (http://www.bristol.ac.uk/alspac/researchers/access/). Ethical approval for
the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics
Committees.
25
and 44 variants were selected in the Child and Mother cohorts, respectively. Metabolite levels being continuous variables, they were quantile normalised and standardised. Genotypes,
on the other hand, were encoded as ternary variables (0/1/2).
In summary, for the Child cohort (resp. Mother cohort) we have: n = 5242, p = |X| =
39, m = |Z| = 133 (resp. n = 2770, p = 39, m = 44).
6.2
Methods
Since both the suggested approach (LSCGGM) and the LR+S method have two tuning
parameters (λ, γ), some procedure is required in order to set these parameters to appropriate values. As shown by both theoretical results and simulations, solutions are expected
to be identical for a range of values of γ. Consequently, we do not select a single value of
γ but consider instead 30 values within the range (0.02, 0.98)5. To each γ corresponds a
regularisation path: a graph along each path is selected using “pointwise” complementary
pairs stability selection (Meinshausen and Bühlmann, 2010; Shah and Samworth, 2013).
Following the approach used in Meinshausen and Bühlmann (2010), the threshold on the
inclusion probabilities is chosen by requiring that the expected number of falsely discovered
edges be at most one: E(V ) ≤ 1 (using their notations). Thus, for each method and each
cohort we obtain a collection of thirty graphical structures.
In order to measure how similar two graphical structures are, we consider their edge sets.
For any pair of undirected graphs G1 = (V1 , E1 ), G2 = (V2 , E2 ) we define their similarity
by their Jaccard Index
J(G1 , G2 ) =
|E1 ∩ E2 |
.
|E1 ∪ E2 |
This measure has two uses: 1) it makes it possible to select γ by measuring how the
5
The penalty is parametrised as λ(γ||S||1 + (1 − γ)||L||∗ ), so that γ ∈ (0, 1).
26
estimates relate to each other as γ varies from 0 to 1; 2) it allows us to measure how well
the findings are replicated across cohorts.
Another important step is assessing the biological relevance of the estimates using an
external source of information. We used ChEBI: an ontology of small chemical entities
of biological interest (Hastings et al., 2012). We manually matched all 39 metabolites to
their ChEBI IDs and annotated them using the ontology. Using such annotations, one
can compute an “enrichment statistic” reflecting whether a given graph contains edges
between related metabolites more often than would be expected in a random graph with
a similar topology (such a graph has an expected statistic of 1). We defer the definition
of this statistic to the supplementary materials but remark that this method is close to
the ontology analyses frequently encountered in computational biology (Wang et al., 2011).
By randomly permuting the annotations, empirical p-values for this statistic can also be
computed.
6.3
Results
First, we can ask how sensitive the estimates are to the tuning parameter γ. Indeed, as
pointed out earlier, one would expect to see a “stable region”: a range of values of γ for which
there is little variation. One would typically select a graph estimated with a γ within this
(γ)
(γ)
region. Let ĜLSCGGM,Ch (resp. ĜLR+S,Ch ) denote the graph returned by LSCGGM (resp.
LR+S) for a given value of γ in the Child cohort. For every pair (γ1 , γ2 ), Figure 3a shows
(γ )
(γ )
1
2
how similar the estimates are to each other (as computed by J(ĜLSCGGM,Ch
, ĜLSCGGM,Ch
)).
In the range 0.6 ≤ γ1 , γ2 ≤ 0.9, they are very close to each other. For small values of γ, the
graphical structures returned by LSCGGM vary smoothly with γ. The regime γ ≤ 0.05
corresponds to the case in which the rank of the latent component is 0: LSCGGM behaves
like a sparse conditional graphical model. Similar figures can be found for the LR+S
27
method and the Mother cohort in the supplementary materials.
Having established that both methods exhibit a stable region, we look at how close
(γ )
(γ )
1
2
the estimates found in these regions are. To that end, we plot J(ĜLSCGGM,Ch
, ĜLR+S,Ch
)
for all pairs (γ1 , γ2 ) (Figure 3b). For small values of γ, LR+S and LSCGGM appear
indistinguishable. However, for γ1 , γ2 > 0.5 their Jaccard Index drops to reach values
around 0.3 - 0.4. But the range γ > 0.5 covers precisely the stable regions of both LSCGGM
and LR+S, thus indicating that the methods’ “best guesses” are different. Figure 4 shows
in what way the graphs found in those stable regions differ, with LR+S inferring more
connections between amino-acids. Here again, a similar result was obtained in the Mother
0.98
0.95
0.91
0.88
0.85
0.81
0.78
0.75
0.72
0.68
0.65
0.62
0.58
0.55
0.52
0.48
0.45
0.42
0.38
0.35
0.32
0.28
0.25
0.22
0.19
0.15
0.12
0.09
0.05
0.02
b)
Jacc.
index1.00
LR+S
a)
0.75
0.50
0.25
0.98
0.95
0.91
0.88
0.85
0.81
0.78
0.75
0.72
0.68
0.65
0.62
0.58
0.55
0.52
0.48
0.45
0.42
0.38
0.35
0.32
0.28
0.25
0.22
0.19
0.15
0.12
0.09
0.05
0.02
Jacc.
index1.00
0.75
0.50
0.25
LSCGGM
Figure 3: Sensitivity of LSCGGM and LR+S to the tuning parameter γ. For any two
graphs, their similarity is computed using the Jaccard Index of their edge sets. (a) Similarities between the edges sets of the graphs returned by LSCGGM in the Child cohort,
as a function of γ (for 30 values of γ ∈ (0.02, 0.98)). (b) Similarities between the graphs
returned by LSCGGM and LR+S in the Child cohort.
28
cohort (see suppl. mat.). The supplementary materials also contains the full name of the
metabolites being modelled.
Given that two cohorts are at our disposal, one way of assessing the quality of our
results is to look at how well they replicate across datasets. In Figure 5a we plot the
similarity between graphs estimated at the same value of γ (see suppl. mat. for a plot of
this similarity for all possible pairs γ1 , γ2 .). First, it can be seen that higher replication
6
LA
w
FA
FA
PU
LSCGGM
C
m
ru
Se tC
Es
LR+S
C
ee G
o
tP
To tCh
To
Fr
PC tFA
To
A
tC
SF oB an
Ap mn
e
R w3
FA A 1
H
D
oA
Ap UFA TG
M rum
Se
G
DA
p
G
LA
C
r
Py
c
La
a
Al
it
C
e
Ac
is
H
b
Al
ln
G
lc
G
r
Ty
e
Ph
Ile u
Le l
e
Va Ac ut
Ac HB
n
Le
FA ea
r
C
bO
PUFA
FAw6
LA
SerumC
EstC
FreeC
TotPG
TotCho
PC
TotFA
SFA
ApoB
RemnantC
FAw3
DHA
ApoA1
MUFA
SerumTG
DAG
Gp
CLA
Pyr
Lac
Ala
Cit
Ace
His
Alb
Gln
Glc
Tyr
Phe
Ile
Leu
Val
AcAce
bOHBut
FALen
Crea
Figure 4: Adjacency matrices of the graphs returned by the LSCGGM and LR+S methods
for γ = 0.81 and γ = 0.68 respectively.
29
values are achieved in the stable regions of their respective methods, with Jaccard Indices
at 0.6 or above. We also see that LSCGGM’s edge set replicates better than LR+P’s.
Moreover, the suggested estimator retrieves more edges under the condition E(V ) ≤ 1 (see
suppl. mat.).
a)
Jaccard Index Ch. vs Mo.
Finally, we use the “enrichment statistic” defined earlier. In our attempt to assess the
0.7
0.6
Method
LR+S
0.5
LSCGGM
0.4
0.3
1.00
0.75
0.50
0.25
0.00
Gamma
b)
Enrichement Statistic
Child Cohort
Mother Cohort
15
Method
10
LR+S
LSCGGM
5
1.00
0.75
0.50
0.25
0.00
1.00
0.75
0.50
0.25
0.00
Gamma
Figure 5: (a) Comparing estimates across cohorts. For each value of γ and each method, we
plot the similarity between the estimate obtained in one cohort against the one obtained
in the other. We limit ourselves to values of γ for which the estimates in both cohorts
comport 15 edges or more. (b) Enrichment statistic, as a function of the tuning parameter
γ.
30
quality of our estimates and their biological relevance, this metric is useful as it makes it
possible to score graphs using an external source of information. Figure 5b shows the value
taken by this statistic across cohorts and methods. Associated p-values can be found in
the see supplementary materials. Here again it is clear that, irrespective of the dataset,
higher values are achieved within the stable regions of their respective methods. Just like
in the case of the replication measure, LSCGGM achieves the highest values. Given that
the Child cohort contains twice as many samples as the Mother cohort, it is surprising to
observe better performances in the Mother dataset. This might be due to the fact that this
cohort is more homogeneous: there are women only, measurements were taken the same
number of months after pregnancy, etc...
7
Discussion
We discussed the problem of estimating a conditional Gaussian graphical model in the
presence of latent variables. Building on the framework introduced by the authors of
Chandrasekaran, Parrilo and Willsky (2012), we suggested an estimator which decomposes
the parameters of a sparse conditional Gaussian graphical model into the sum of a low-rank
and a sparse matrix. Among other theoretical results, we established that the proposed
approach is well-behaved in the high-dimensional regime. Through simulations and an
application to a modern dataset comprising genetic and metabolic measurements, we compared the performances of this approach to alternative methods. In particular, we showed
how such a conditional graphical model leads to better replication of the results across
cohorts and to estimates that are more biologically relevant.
The rise of high-throughput genetics, along with progress in data linkage, biobanking
and functional genomics projects, has dramatically increased the number of datasets that
31
include both genetic and multivariate phenotypic data. The data application we present
in this paper, using genotype data to draw biological conclusions about the relationships
between human traits, is thus becoming one of the most rapidly growing statistical challenges in human genetics. Conditional graphical models are particularly well-suited to
such problems as they rely on an assumption we know to be true (namely, that genotype
impacts phenotype and not vice-versa). Moreover, genetic measurements are discrete in
nature and it is therefore difficult to model them alongside continuous measurements. To
the best of our knowledge, there are no approaches capable of learning a joint distribution
over continuous and discrete data in the presence of latent variables.
When it comes to lasso-type estimators, choosing an appropriate value of the tuning
parameters can also be challenging. In simulations, our method seems to be less sensitive
to the value of the tuning parameter γ, which makes it easier to set it to a suitable value in
real life applications. Moreover, the use of (complementary pairs) stability selection makes
the estimates less sensitive to the value of λ while providing some form of error control.
Another limitation of such estimators comes from the fact that consistency/identifiability
conditions are highly likely to be violated in real world applications. While this is true,
a more realistic take is to regard our method as a means to generate “causal” hypotheses
from a high-dimensional dataset. Paired with stability selection, such an approach can
realistically be used to generate a high-quality set of putative causal relationships that can
then further be investigated using hypothesis testing driven approaches (e.g. instrumental
variables). As shown in our application this is an achievable goal.
Naturally, the method suggested here also suffers from a number of limitations and more
work is required. For example, assuming that the latent variables are normally distributed
appears quite restrictive when compared to the flexibility offered by instrumental variable
methods. The question of learning discrete graphical models is also important but it is not
32
yet clear how the present work can be extended to such models.
Acknowledgements
We are extremely grateful to all the families who took part in the ALSPAC study, the
midwives for their help in recruiting them, and the whole ALSPAC team, which includes
interviewers, computer and laboratory technicians, clerical workers, research scientists,
volunteers, managers, receptionists and nurses. The UK Medical Research Council and
the Wellcome Trust (Grant ref: 102215/2/13/2) and the University of Bristol provide core
support for ALSPAC. This publication is the work of the authors and Gil McVean will
serve as guarantor for the contents of this paper. GWAS data was generated by Sample
Logistics and Genotyping Facilities at the Wellcome Trust Sanger Institute and LabCorp
(Laboratory Corporation of America) using support from 23andMe. This research was
specifically funded by the Wellcome Trust grant 100956/Z/13/Z (GM); the Wellcome Trust
grant 098759/Z/12/Z (LJ) and the EPSRC grant EP/F500394/1 and an Amazon Web
Services research grant (BF).
References
Bach, F. (2008), ‘Consistency of trace norm minimization’, J. Mach. Learn. Res. 8, 1019–
1048.
Banerjee, O., El Ghaoui, L. and d’Aspremont, A. (2008), ‘Model selection through sparse
maximum likelihood estimation for multivariate gaussian or binary data’, J. Mach. Learn.
Res. 9, 485–516.
33
Boyd, A. et al. (2012), ‘Cohort profile: The children of the 90s–the index offspring of the
avon longitudinal study of parents and children’, International Journal of Epidemiology
42(1), 111–127.
Boyd, S. (2010), ‘Distributed optimization and statistical learning via the alternating direction method of multipliers’, Foundations and Trends in Machine Learning 3(1), 1–122.
Candès, E. J., Li, X., Ma, Y. and Wright, J. (2011), ‘Robust principal component analysis?’,
J. ACM 58(3), 1–37.
Chandrasekaran, V., Parrilo, P. A. and Willsky, A. S. (2012), ‘Latent variable graphical
model selection via convex optimization’, The Annals of Statistics 40(4), 1935–1967.
Chandrasekaran, V., Recht, B., Parrilo, P. A. and Willsky, A. S. (2012), ‘The convex geometry of linear inverse problems’, Foundations of Computational Mathematics 12(6), 805–
849.
Chandrasekaran, V., Sanghavi, S., Parrilo, P. A. and Willsky, A. S. (2009), ‘Rank-sparsity
incoherence for matrix decomposition’, SIAM Journal on Optimization 21(2), 572–596.
Eckstein, J. and Bertsekas, D. P. (1992), ‘On the douglas—rachford splitting method and
the proximal point algorithm for maximal monotone operators’, Mathematical Programming 55(1-3), 293–318.
Fraser, A. et al. (2012), ‘Cohort profile: The avon longitudinal study of parents and children: ALSPAC mothers cohort’, International Journal of Epidemiology 42(1), 97–110.
Friedman, J., Hastie, T. and Tibshirani, R. (2008), ‘Sparse inverse covariance estimation
with the graphical lasso’, Biostatistics 9(3), 432–41.
34
Goldstein, T. and Osher, S. (2009), ‘The split bregman method for l1-regularized problems’,
SIAM Journal on Imaging Sciences 2(2), 323–343.
Hastings, J. et al. (2012), ‘The ChEBI reference database and ontology for biologically
relevant chemistry: Enhancements for 2013’, Nucleic Acids Research 41(D1), D456–
D463.
Lauritzen, S. (1996), Graphical Models, Clarendon Press.
Li, X., Xie, H., Chen, L., Wang, J. and Deng, X. (2014), ‘News impact on stock price
return via sentiment analysis’, Knowledge-Based Systems 69, 14–23.
Löfberg, J. (2004), Yalmip : A toolbox for modeling and optimization in matlab, in ‘Proceedings of the CACSD Conference’, pp. 284 – 289.
Ma, S., Xue, L. and Zou, H. (2013), ‘Alternating direction methods for latent variable
gaussian graphical model selection’, Neural Computation 25(8), 2172–2198.
Meinshausen, N. and Bühlmann, P. (2010), ‘Stability selection’, Journal of the Royal Statistical Society. Series B: Statistical Methodology 72(4), 417–473.
Ravikumar, P., Wainwright, M. J., Raskutti, G. and Yu, B. (2011), ‘High-dimensional
covariance estimation by minimizing l1-penalized log-determinant divergence’, Electronic
Journal of Statistics 5.
Shah, R. D. and Samworth, R. J. (2013), ‘Variable selection with error control: Another
look at stability selection’, Journal of the Royal Statistical Society. Series B: Statistical
Methodology 75(1), 55–80.
35
Sohn, K.-A. and Kim, S. (2012), Joint estimation of structured sparsity and output structure in multiple-output regression via inverse-covariance regularization, in ‘Conference
on Artificial Intelligence and Statistics’.
Stearns, F. W. (2010), ‘One hundred years of pleiotropy: A retrospective’, Genetics
186(3), 767–773.
Tütüncü, R. H., Toh, K. C. and Todd, M. J. (2003), ‘Solving semidefinite-quadratic-linear
programs using sdpt3’, Mathematical Programming, Series B 95(2), 189–217.
Vandenberghe, L. and Boyd, S. (1996), ‘Semidefinite programming’, SIAM Review
38(1), 49–95.
Wang, C., Sun, D. and Toh, K.-C. (2010), ‘Solving log-determinant optimization problems by a newton-cg primal proximal point algorithm’, SIAM Journal on Optimization
20(6), 2994–3013.
Wang et al., . (2011), ‘NOA: a novel network ontology analysis method’, Nucleic Acids
Research 39(13).
Wytock, M. and Kolter, J. Z. (2013), Sparse gaussian conditional random fields: Algorithms, theory, and application to energy forecasting, in ‘Proceedings of the 2013 International Conference on Machine Learning’, pp. 1265–1273.
Ye,
G.-B.,
Wang,
Y. and Xie,
X. (2011),
‘Efficient latent variable graphi-
cal model selection via split bregman method’,
Available on the arXiv at
http://arxiv.org/pdf/1110.3076v1.pdf .
Yin, J. and Li, H. (2011), ‘A sparse conditional gaussian graphical model for analysis of
genetical genomics data.’, The annals of applied statistics 5(4), 2630–2650.
36
Yuan, M. and Lin, Y. (2007), ‘Model selection and estimation in the gaussian graphical
model’, Biometrika 94(1), 19–35.
Zhang, L. and Kim, S. (2014), ‘Learning gene networks under snp perturbations using eqtl
datasets.’, PLoS computational biology pp. 1–20.
37
| 10math.ST
|
Turing Kernelization for Finding Long Paths
in Graph Classes Excluding a Topological Minor∗
Bart M. P. Jansen†, Marcin Pilipczuk‡, and Marcin Wrochna‡
arXiv:1707.01797v1 [cs.DS] 6 Jul 2017
Abstract
The notion of Turing kernelization investigates whether a polynomial-time algorithm can
solve an NP-hard problem, when it is aided by an oracle that can be queried for the answers
to bounded-size subproblems. One of the main open problems in this direction is whether
k-Path admits a polynomial Turing kernel: can a polynomial-time algorithm determine
whether an undirected graph has a simple path of length k, using an oracle that answers
queries of size k O(1) ?
We show this can be done when the input graph avoids a fixed graph H as a topological
minor, thereby significantly generalizing an earlier result for bounded-degree and K3,t -minorfree graphs. Moreover, we show that k-Path even admits a polynomial Turing kernel when
the input graph is not H-topological-minor-free itself, but contains a known vertex modulator
of size bounded polynomially in the parameter, whose deletion makes it so. To obtain our
results, we build on the graph minors decomposition to show that any H-topological-minorfree graph that does not contain a k-path, has a separation that can safely be reduced after
communication with the oracle.
1
Introduction
Suppose that Alice is a polynomial-time agent faced with an input to an NP-hard problem
that she wishes to solve exactly. To facilitate her in this process, she can ask questions to an
all-knowing oracle. These will be answered truthfully and instantly, but the oracle is memory-less
and will not take previous questions into account when answering the next one. How large do
these questions have to be, to allow Alice to find the answer to her problem? Clearly, the answer
can be established by sending the entire input to the oracle, who determines the answer and
sends it to Alice. Could there be a more clever strategy? Alice can attempt to isolate a small
but meaningful question about the behavior of her input, such that after learning its answer,
she can reduce to a smaller input without changing the outcome. Iterating this process solves
her problem: when it has become sufficiently small, it can be posed to the oracle in its entirety.
Such problem-solving strategies can be rigorously analyzed using the notion of Turing
kernelization that originated in parameterized algorithmics. The parameter makes it possible to
express how the size of the questions that Alice asks, depends on properties of the input that
she is given. (See Section 3 for a formal definition.)
Understanding the power of Turing kernelization is one of the main open research horizons
in parameterized algorithmics. There is a handful of problems for which a nontrivial Turing
∗
This work was supported by the Netherlands Organization for Scientific Research (NWO) Veni grant
639.021.437 “Frontiers in Parameterized Preprocessing” and Gravitation grant 024.002.003 “Networks”. Marcin
Pilipczuk is supported by the “Recent trends in kernelization: theory and experimental evaluation” project,
carried out within the Homing programme of the Foundation for Polish Science co-financed by the European
Union under the European Regional Development Fund. Marcin Wrochna is supported by the National Science
Centre of Poland grant number 2013/11/D/ST6/03073 and by the Foundation for Polish Science (FNP) via the
START stipend programme.
†
Eindhoven University of Technology, The Netherlands
‡
University of Warsaw, Poland
1
kernelization is known [1, 3, 4, 6, 14, 17, 19, 20, 26, 28]. On the other hand, there is a hierarchy
of parameterized complexity classes which are conjectured not to admit polynomial Turing
kernels [16]. Arguably, the main open problem (cf. [5, 4, 16]) in this direction is to determine
whether the k-Path problem (determine whether an undirected graph has a simple path of
length k) has a polynomial Turing kernel. In earlier work [18], the first author showed that
k-Path indeed admits polynomial Turing kernels on several graph classes. In this work, we
develop Turing kernels for k-Path in a much more general setting.
Our results Our algorithmic contributions are twofold. First of all, we extend the Turing
kernelization for k-Path to much broader families of sparse graphs. Whereas the earlier work
could only deal with K3,t -minor-free graphs, claw-free graphs, and bounded-degree graphs, we
show that a Turing kernelization exists on H-minor-free graphs for all fixed graphs H. We
even lift the kernelization to H-topological-minor-free graphs, thereby capturing a common
generalization of the bounded-degree and K3,t -minor-free cases.
Theorem 1.1. For every fixed graph H, the k-Path problem, restricted to graphs excluding H
as a topological minor, admits a polynomial Turing kernel. Furthermore, the kernel runs in time
k OH (1) n2 m and invokes k OH (1) · n calls to the oracle.
Our second contribution is the following theorem. By a novel algorithmic approach, we
obtain a Turing kernelization even when the input graph does not belong to the desired restricted
graph class itself, but contains a small known vertex modulator whose deletion places the graph
in such a graph class.
Theorem 1.2. For every fixed graph H, the k-Path problem, on instances consisting of a graph
G, integer k, and a modulator M ⊆ V (G) such that G − M is H-topological-minor-free, admits
a polynomial Turing kernel, when parameterized by k and |M |.
Techniques To explain our approach, we briefly recall the idea behind the Turing kernelization for k-Path on planar graphs. At the core lies a win/win: there is a polynomial-time
algorithm that either (i) establishes that a planar graph G has a k-path (a simple path on
k vertices), or (ii) finds a separation (A, B) in G with the following property: the size of A
is polynomially bounded in k, but large enough that after marking a witness structure for
each reasonable way in which a k-path might intersect A, some vertex remains unmarked.
Using bounded-size oracle queries to mark the witness structures, this allows the problem to be
simplified by removing an unmarked vertex from A without changing the answer.
Theorem 1.1 is established by lifting this win/win approach to H-(topological)-minor-free
graphs. This requires an adaptation of the decomposition theorems of Robertson and Seymour [25]
(for minors) and of Grohe and Marx [15] (for topological minors), to obtain the following. Every
H-free graph that does not have a k-path, has a tree decomposition of constant adhesion and
width poly(k). A reducible separation can be found by inspecting this tree decomposition. To
establish this result, we exploit known theorems stating that triconnected n-vertex graphs that
exclude K3,t as a minor for some t [8], contain paths of length Ω(nε ) for some ε > 0. Roughly
speaking, this allows us to infer the existence of a k-path if there is a large embedded part in
the nearly-embeddable graph corresponding to a bag of the graph minors decomposition, since
graphs embeddable in a fixed surface are K3,t -minor-free for some t. We use lower bounds on
the circumference of graphs of bounded degree [7, 27] to achieve a similar conclusion from the
existence of a large bounded-degree bag in the topological-minor-free decomposition. Several
technical steps are needed to translate this into the desired win/win, due to the existence of
vortices, virtual edges, and the lack of a direct polynomial-time algorithm to compute the
decomposition.
To prove Theorem 1.2, we introduce a new algorithmic tool for finding irrelevant vertices for
the k-Path problem in the presence of a modulator M in the input graph G. Since Theorem 1.1
2
can be applied to find a k-path in G − M if one exists, the challenge is to detect a k-path in G
that jumps between M and G − M several times. The absence of a k-path in G − M implies
it has a tree decomposition of width poly(k) and constant adhesion. Using Theorem 1.1 as a
subroutine, along with a packing argument, we can compute a vertex set X of size polynomial
in k + |M | with the following guarantee. If there is a k-path, then there is a guarded k-path P in
which each successive pair of vertices in M ∩ P are connected by a subpath through G − M that
intersects X. Using the tree decomposition of G − M , the standard ancestor-marking technique
allows us to identify a vertex subset C of G − (M ∪ X) that is adjacent to constantly many
vertices from X. Unless G is already small, we can find such a set C that is sufficiently large to
be reducible but small enough that we may invoke the oracle for questions about it. We can
then reduce the graph without losing the existence of a guarded k-path, by marking a witness
for each sensible way in which a constant-size subset from M can connect to prescribed vertices
in X through C. The fact that C only has constantly many neighbors in X implies that there
are only polynomially many relevant choices. We may then safely remove the unmarked vertices.
Organization After preliminaries in Section 2, we give a generic Turing-style reduction
rule for k-Path in Section 3. In Section 4 we show that an H-minor-free graph either has a
k-path or a separation that is suitable for reduction. In Section 5 we extend this to topological
minors. Finally, in Section 6 we present a Turing kernel applicable when the input graph has a
small modulator to a suitable graph class.
2
Preliminaries
Notation. All graphs we consider are finite, simple, and undirected. A separation of a graph
G is a pair (A, B), A, B ⊆ V (G) such that A ∪ B = V (G) and there are no edges between
A \ B and B \ A. The order of the separation (A, B) is |A ∩ B|. A graph is triconnected if it is
connected and cannot be disconnected by deleting fewer than three vertices. When referring to
the size of a graph in our statements, we mean the number of vertices.
A tree decomposition of a graph G is a pair (T, X ) where T is a rooted tree and X is a
function that assigns to every node t ∈ V (T ) a subset X (t) of V (G) called a bag such that:
S
• t∈V (T ) X (t) = V (G);
• for each edge uv ∈ E(G), there is a node t ∈ V (T ) with u, v ∈ X (t);
• for each v ∈ V (G), the nodes {t | v ∈ X (t)} induce a (connected) subtree of T .
The width of (T, X ) is maxt∈V (T ) |X (t)| − 1. Its adhesion is maxtt0 ∈E(T ) |X (t) ∩ X (t0 )|. We also
call the set X (t) ∩ X (t0 ) the adhesion of tt0 , for every edge tt0 of T . For a decomposition (T, X ) of
G and a node t ∈ V (T ), the torso, denoted torso(G, X (t)), is the graph obtained from G[X (t)]
by adding an edge between each pair of vertices in X (t) ∩ X (t0 ), for every neighbor t0 of t in T
(so each adhesion induces a clique in the torso). Added edges
virtual
S not present in G are called
0
0
0
edges. For a subtree T ⊆ T we write X (T ) for the union t∈V (T 0 ) X (t) of bags in T .
For an edge t1 t2 ∈ E(T ), let Ti be the connected component of T − {t1 t2 } that contains ti .
Let Vi = X (Ti ). Observe that the properties of a tree decomposition imply that (V1 , V2 ) is a
separation with V1 ∩ V2 = X (t1 ) ∩ X(t2 ).
A decomposition (T, X ) is connected if for every t ∈ V (T ) and its child t0 , if Tt0 is the subtree
of T rooted at t0 , we have (i) that G[X (Tt0 ) \ X (t)] is connected, and (ii) that X (Tt0 ) \ X (t) has
edges to every vertex of the adhesion X (t)∩X (t0 ). It is straightforward to turn any decomposition
into a connected one without increasing its width nor adhesion. For (i), as long as there exists a
pair (t, t0 ) violating the requirement, make a distinct copy TtC0 of Tt0 for each connected component
C of G[X (Tt0 ) \ X (t)], restrict the bags of TtC0 to the vertices of C ∪ X (t) only, and attach TtC0 as
a subtree with the root being a child of t. For (ii), while there is a vertex v ∈ X (t) ∩ X (t0 ) that
has no neighbors in X (Tt0 ) \ X (t), simply remove v from all bags in Tt0 .
3
We will also need the following non-standard complexity measure of a tree decomposition
(T, X ). For every t ∈ V (T ), the number of distinct adhesions X (t) ∩ X (t0 ) for t0 ∈ NT (t) is called
the adhesion degree of t. The maximum adhesion degree over all nodes t is the adhesion degree
of the decomposition (T, X ). Observe that if a tree decomposition (T, X ) has width less than `
and adhesions of size at most h, then its adhesion degree is at most
h
X
`
i=0
i
≤ (1 + `)h .
However, in sparse graph classes we can prove a much better bound on the adhesion degree due
to linear bounds on the number of cliques in such graphs; cf. Lemma 4.3.
A path decomposition is a tree decomposition where T is a path; we will denote it simply as
a sequence of bags Z1 , . . . , Z|V (T )| .
For an integer θ, a tangle of order θ in a graph G is a family T of separations of order at most
θ such that, for every separation (X, Y ) of order at most θ in G, exactly one of the separations
(X, Y ) and (Y, X) belongs to T . If (X, Y ) ∈ T we call X the small side of the separation and
Y the large side. Furthermore, we require that for every three separations (X1 , Y1 ), (X2 , Y2 ),
(X3 , Y3 ) in T , we have G[X1 ] ∪ G[X2 ] ∪ G[X3 ] 6= G.
3
Turing kernels
In this section we introduce a general toolbox and notation for proving our Turing kernel bounds.
3.1
Definitions and the auxiliary problem
For a parameterized problem Π and a computable function f , a Turing kernel of size f is an
algorithm that solves an input instance (x, k) of Π in polynomial time, given access to an oracle
that solves instances (x0 , k 0 ) of Π with |x0 |, k 0 ≤ f (k). A Turing kernel is a polynomial one if f is
a polynomial.
If we are only interested in distinguishing between NP-complete problems admitting a
polynomial Turing kernel from the ones that do not admit such a kernel, we can assume that
the oracle solves an arbitrary problem in NP, not necessarily the k-Path problem. Indeed, note
that by the definition of NP-completeness, an oracle to a problem in NP can be implemented
with an oracle to k-Path with only polynomial blow-up in the size of the passed instances.
In our work, it will be convenient to reduce to the Auxiliary Linkage problem, defined as
follows. The input consists of an undirected graph G0 , an integer k 0 , a set of terminals S ⊆ V (G0 ),
and a number of requests R1 , R2 , . . . , Rr ; a request is a set of at most two terminals. A path
Pi in G is said to satisfy a request Ri if V (Pi ) ∩ S = Ri and every vertex of V (Pi ) ∩ S is an
endpoint of Pi . With such an input, the Auxiliary Linkage problem
S asks for a sequence of r
paths P1 , P2 , . . . , Pr such that Pi satisfies Ri for every 1 ≤ i ≤ r, | ri=1 V (Pi )| = k 0 , and every
vertex of V (G) \ S is contained in at most one path Pi (i.e., the paths Pi are vertex-disjoint,
except that they may share an endpoint, but only if the requests ask them to do so).
We remark that Auxiliary Linkage is a more general problem than k-Path: an instance
with G0 = G, k 0 = k, S = ∅, r = 1, and R1 = ∅ asks precisely for a k-path in G.
Clearly, the decision version of the Auxiliary Linkage problem belongs to the class NP.
By using its self-reducibility (cf. [18, Lemma 2]), we assume that the oracle returns a sequence
of paths (Pi )ri=1 in case of a positive answer. That is, in all subsequent bounds on the number
of Auxiliary Linkage oracle calls, the bound adheres to the number of calls to an oracle that
returns the actual paths Pi ; if one wants to use a decision oracle, one should increase the bound
by the blow-up implied by the self-reducibility application (i.e., at most |E(H)| for calls on a
graph H).
4
Z
N (A)
2
3
A
1
Figure 1: A set A (blue) and a path with three A-traverses (bold). The path is guarded
w.r.t. Z ⊆ N (A) (the red set), since each A-traverse has an endpoint in it. (A path fully
contained in A (with one A-traverse) or disjoint from A (with no A-traverses) would also be
guarded.)
3.2
Generic reduction rule
We now show a generic reduction rule for the k-Path problem. We start with a few definitions.
Definition 3.1. For a graph G, a subset A ⊆ V (G), and a simple path P in G, an A-traverse
of P is a maximal subpath of P that contains at least one vertex of A and has all its internal
vertices in A.
Note that if Q is an A-traverse of P , then every endpoint of Q is either an endpoint of P or
lies in NG (A). See Figure 1.
Definition 3.2. Let G be a graph, A ⊆ V (G), and let k be an integer. A set Z ⊆ N (A) is
called a k-guard of A if the following implication holds: if G admits a k-path, then there exists
a k-path P in G that is either contained in A or such that every A-traverse of P has at least
one endpoint in Z.
Given a graph G, a set A ⊆ V (G), and a k-guard Z ⊆ N (A) of A, a k-path P satisfying
properties as in the above definition is called guarded (w.r.t. k, A, and Z). If the integer k and
the set A are clear from the context, we call such a set Z simply a guard.
Observe that Z = N (A) is always a guard, but sometimes we will be able to find smaller
ones. Of particular interest will be guards of constant size, as our kernel sizes will depend
exponentially on the guard size. To describe our single reduction rule, we show how solutions to
Auxiliary Linkage can be used to preserve the existence of guarded k-paths.
Assume we are given a graph G, a set A ⊆ V (G), an integer k, and a k-guard Z ⊆ N (A)
of A. Let h = |Z| and ` = |N (A)|. Furthermore, assume that G admits a k-path, and
let P be a guarded one w.r.t. A and Z. Let (Q1 , Q2 , . . . , Qr ) be the A-traverses of P , let
Ri = VS(Qi ) \ A = V (Qi ) ∩ N (A) for 1 ≤ i ≤ r, let G0 = G[N [A]], S = N (A), and let
k 0 = | ri=1 V (Qi )|. Observe that (Q1 , Q2 , . . . , Qr ) is a feasible solution to the Auxiliary
Linkage instance IP := (G[N [A]], k 0 , S, (Ri )ri=1 ); the instance IP is henceforth called induced
by P and A. Furthermore, it is easy to see that if (Q01 , Q02 , . . . , Q0r ) is a different feasible solution
to IP , then a path P 0 obtained from P by replacing every subpath Qi with Q0i is also a guarded
k-path in G.
The crucial observation is that a small guard limits the number of A-traverses.
Lemma 3.3. The number r of traverses of the guarded k-path P is bounded by max(1, 2|Z|).
Proof. Every vertex of Z can be an endpoint of at most two traverses. If r > 1, then none of
the traverses Qi are contained in G[A], and thus every traverse has at least one endpoint in the
guard Z.
5
Lemma 3.3 in turn limits the number of possible instances I that can be induced by a guarded
k-path, for a fixed set A and guard Z. Note that we have 0 ≤ k 0 ≤ k and 0 ≤ r ≤ max(1, 2|Z|).
Furthermore, unless r = 1 and R1 = ∅, we have Ri ⊆ N (A), |Ri | ∈ {1, 2}, and every set Ri
needs to have at least one element of Z; there are at most |Z| · (|N (A)| + 1) = h(` + 1) choices
for such a set Ri . Consequently, the number of possibilities for the instance I is at most
!
2h
X
r
r
(k + 1) · 1 +
h (` + 1) ≤ (k + 1) · (h(` + 1))2h+1 =: p(k, `, h).
(1)
r=0
Reduction rule. If |A| > k ·p(k, `, h), then we can apply the following reduction rule. For each
Auxiliary Linkage instance I out of at most p(k, `, h) reasonable instances for A-traverses
of a guarded k-path in G, we invoke an oracle on the instance I, and mark the vertices of the
solution if the oracle finds one. The whole process will mark at most k · p(k, `, h) < |A| vertices,
thus at least one vertex of |A| will remain unmarked. We delete any such vertices.
The observation that on a guarded k-path P one can replace a solution to the instance IP
induced by P and A by a different solution provides safeness of this reduction. Finally, note
that the reduction invokes at most p(k, `, h) calls to the oracle; each call operates on a subgraph
of the graph G[N [A]] with k 0 ≤ k and r ≤ 2|Z|.
We shall apply the Reduction Rule for a medium-sized set A and a guard set Z of constant
size formed from adhesions of a tree decomposition. For most of the paper we will use Z = N (A)
with ` = h = |Z| a constant (depending on the excluded (topological) minor, in the results of
Sections 4 and 5). Only in Section 6, when dealing with a modulator M such that G − M has
an appropriate structure, it will be important to consider N (A) potentially containing all of M ,
with a guard set Z of constant size disjoint from M .
3.3
Separation oracles
The natural way of using our reduction rule is to find in a graph a large (but not too large)
part of the graph with a small (preferably, constant) boundary. Let us first make an abstract
definition of an algorithm finding such a separation.
Definition 3.4. For a graph class G, a constant h, and a computable coordinate-wise nondecreasing function q : Z≥0 × Z≥0 → Z≥0 , an algorithm S is called a (h, q, TS )-separation oracle if,
given a graph G ∈ G and integers k and p, in time TS (|G|, k, p) it finds a separation (A, B) in G
of order at most h with p < |A| ≤ q(k, p), or correctly concludes that G contains a k-path.
For all considered graph classes, we will be able to provide a separation oracle with q being
a polynomial. This, in turn, allows the following generic Turing kernel.
Lemma 3.5. Let S be a (TS , h, q)-separation oracle for a hereditary graph class G. Take
ĥ := (2h)4h+3 . Then, the k-Path problem restricted to graphs from G can be solved:
- in time O TS (|G|, k, k 2 ĥ) · |V (G)| + k ĥ · |V (G)| · |E(G)| ,
- using at most k ĥ · |V (G)| calls to Auxiliary Linkage
- each call on an induced subgraph of the input graph of size at most q(k, k 2 ĥ).
Proof. Let
p = k · p(k, h, h) ≤ k(k + 1)(h(h + 1))2h+1 ≤ 2k 2 (2h)4h+2 ≤ k 2 ĥ.
As long as |V (G)| > p, we proceed as follows. Invoke algorithm S on G. If S claims that G
admits a k-Path, we simply output the answer yes. Otherwise, let (A0 , B 0 ) be the separation
output by S. Apply the Reduction Rule for k, A := A0 \ B 0 , and Z = N (A) ⊆ A0 ∩ B 0 . Note that
6
as |Z| ≤ h, the Reduction Rule deletes at least one vertex of A. Furthermore, the Reduction
Rule invokes at most
p(k, h, h) ≤ k ĥ
calls to the oracle, each call on an induced subgraph of G of size at most
|A0 | ≤ q(k, p) = q(k, k 2 ĥ).
Once we obtain |V (G)| ≤ p, we solve the instance using a single call to Auxiliary Linkage
with k 0 = k, r = 1, and R1 = ∅. The bounds follow, as there are at most |V (G)| applications of
the Reduction Rule, and each call to the oracle takes O(|E(G)|) time to prepare the instance
and parse the output.
Note that for any graph class where separations as in Definition 3.4 exist, there exists a trivial
separation oracle which finds them, running in time nh+O(1) : one iterates over every candidate
for A ∩ B and, for fixed set A ∩ B, a straightforward knapsack-type dynamic programming
algorithm checks if one can assemble A \ B of the desired size from the connected components
of G − (A ∩ B).
However, this running time bound is unsatisfactory, as it greatly exceeds the number of used
oracle calls. For all considered graph classes we prove a much stronger property than just merely
the prerequisites of Lemma 3.5, in particular providing a more efficient separation oracle. We
provide necessary definitions in the next section.
3.4
Decomposable graph classes
The following definition captures the key concept of this section.
Definition 3.6. For a constant h and a computable nondecreasing function w : Z≥0 → Z≥0 , a
graph class G is called (w, h)-decomposable if for every positive integer k and every G ∈ G that
does not admit a k-path, the graph G admits a tree decomposition of width less than w(k) and
adhesions of size at most h.
A standard argument shows that in a decomposable graph class, given the decomposition
with appropriate parameters, it is easy to provide a separation oracle.
Lemma 3.7. Assume we are given a graph G and a tree decomposition (T, X ) of G of width
less than w, adhesion at most h, and adhesion degree at most a ≥ 2. Then,
Pgiven an integer p
such that |V (G)| > p, one can in time hO(1) · (|V (G)| + |E(G)| + |V (T )| + t∈V (T ) |X (t)|) find
a separation (A, B) of order at most h such that
p < |A| ≤ w + p · a.
Proof. Root the tree T in an arbitrary node, and for t ∈ V (T ) let Tt be the subtree of T rooted
in t. Let t0 be the lowest node of T such that |X (V (Tt ))| > p; such a node can be computed in
linear time in the size of G and (T, X ).
Group the children t0 of t according to their adhesions X (t0 ) ∩ X (t). Due to the bound on
the adhesion degree, there are at most a groups. For every adhesion S, let Xt,S be the set of the
children of t with S = X (t0 ) ∩ X (t). Define
[
VS =
X (Tt0 ).
t0 ∈Xt,S
We consider now two cases. First, assume that |VS | ≤ p for every adhesion S. Then, by the
adhesion degree bound, we have
|X (Tt )| ≤ |X (t)| + ap ≤ w + ap.
7
Consequently, we can return the separation (A, B) with A = X (Tt ) and B = X (T − V (Tt )).
In the other case, there exists an adhesion S with VS > p. We greedily take a minimal subset
Yt,S ⊆ Xt,S such that
[
VS0 :=
X (Tt0 )
t0 ∈Yt,S
is of size greater than p. By the minimality of t, for every t0 ∈ Xt,S we have |X (Tt0 )| ≤ p
and, consequently |VS0 | ≤ 2p. Thus, we can return the separation (A, B) for A = VS0 and
B = NG [V (G) \ VS0 ], as then A ∩ B ⊆ S.
A critical insight is that the decomposition used by Cygan et al. [9] to solve the Minimum
Bisection problem in fact provides an approximate decomposition in a decomposable graph
class. Let us first recall the main technical result of [9].
Definition 3.8. A vertex set X ⊆ V (G) of a graph G is called (q, h)-unbreakable if every
separation (A, B) of order at most h satisfies |(A \ B) ∩ X| ≤ q or |(B \ A) ∩ X| ≤ q.
Theorem 3.9 ([9]). There is an algorithm that given a graph G and integer h runs in time
2
2O(h ) |V (G)|2 |E(G)| and outputs a connected tree decomposition (T, Y) of G such that: (i) for
each t ∈ V (T ), the bag Y(t) is (2O(h) , h)-unbreakable in G, and (ii) for each tt0 ∈ E(T ) the
adhesion Y(t) ∩ Y(t0 ) has at most 2O(h) vertices and is (2h, h)-unbreakable in G.
Lemma 3.10. Let G be a graph and suppose there exists a decomposition (T, X ) of G of width
less than w, adhesion h, and adhesion degree a. Let (T 0 , Y) be a tree decomposition of G such
that for each t ∈ V (T 0 ), the bag Y(t) is (2O(h) , h)-unbreakable in G. Then |Y(t)| ≤ w + a · 2O(h) .
Proof. Consider a bag Y = Y(t) of the second decomposition, or any (2O(h) , h)-unbreakable set
Y ⊆ V (G). For every edge tt0 ∈ E(T ) of the h-decomposition, removing it partitions T into
subtrees Tt0 t and Ttt0 containing t0 and t, respectively. The induced separation (X (Ttt0 ), X (Tt0 t ))
on G has order at most h, so one of the sides, say X (Tt0 t ) \ X (t), contains at most 2O(h) vertices
of Y , by definition of unbreakability. Let us orient the edge tt0 away from t0 (the ‘smaller’ side).
By orienting each edge of T this way, we find a single node t0 of T such that all incident edges
point to it. That is, for every neighbor t0 of t0 , we have |Y ∩ X (Tt0 t0 ) \ X (t0 )| ≤ 2O(h) .
While t0 can have many neighbors, we can group those neighbors t0 according the adhesion
X (t0 ) ∩ X (t0 ) to which they correspond. By the adhesion degreeSbound, there are at most a such
groups. If for any such group U ⊆ N (t0 ) ⊆ V (T ), the union t0 ∈U X (Tt0 t0 ) \ X (t0 ) contained
more than 3 · 2O(h) vertices of Y , then the group can be partitioned into two parts with more than
2O(h) vertices of Y each. This would give a separation of order h of G with too many vertices of
Y on both sides, contradicting its unbreakability. Therefore, there are at most a groups, each
containing at most 3 · 2O(h) vertices of Y , thus the size of Y is bounded by |X (t0 )| + a · 3 · 2O(h)
Thus, the decomposition computed by the algorithm of Theorem 3.9 approximates the desired
decomposition of a decomposable graph class.
Corollary 3.11. Let G be a (w, h)-decomposable graph class. Then, for every G ∈ G and every
2
integer k, one can in 2O(h ) |V (G)|2 |E(G)| time either correctly conclude that G admits a k-path,
or find a tree decomposition of G of width at most (w(k) + 1)O(h) and adhesion at most 2O(h) .
Let us now combine all the above. That is, given an integer k and a graph G from a hereditary
(w, h)-decomposable graph class G, we start by computing the tree decomposition of Corollary 3.11
(or conclude there is a k-path). In general this approximated decomposition has adhesion degree
O(h)
(w(k) + 1)2
. We use this decomposition to find separations of any induced subgraphs of G
using the algorithm of Lemma 3.7 in time 2O(h) times linear in the size of G and the computed
O(h)
decomposition. This gives a (h, q, T )-separation oracle with q(k, p) = p · (w(k) + 1)2
and
O(h)
O(h)
T (n, k, p) = 2
· n · (w(k) + 1)
, for any hereditary (w, h)-decomposable graph class. By
plugging it into Lemma 3.5, we obtain the following.
8
Corollary 3.12. Let G be a hereditary (w, h)-decomposable graph class. Then, the k-Path
O(h)
O(h)
problem, restricted to graphs from G, can be solved in time 22
|V (G)|2 |E(G)| using 22
kn
O(h)
2
2
calls to Auxiliary Linkage on induced subgraphs of the input graph of size k (1 + w(k))
.
In the next section, we prove that H-minor-free graphs are (OH (k 22 ), OH (1))-decomposable
by analyzing the Global Structure Theorem of H-minor-free graphs due to Robertson and
Seymour. A subsequent section provides an analogous result for graphs excluding a fixed
topological minor. In both cases we also get better bounds on the adhesion degree of the
approximate decomposition outputted by Theorem 3.9, improving the bounds in the final kernel.
We would like to remark that we do not want to claim in this paper the idea that, in the
context of H-(topological)-minor-free graphs, the decomposition of Theorem 3.9 should be
related to the decomposition of the Global Structure Theorem via an argument as in the proof
of Lemma 3.10. In particular, this observation appeared previously in a work of the second
author with Daniel Lokshtanov, Michal Pilipczuk, and Saket Saurabh [21].
4
Excluding a minor
In this section we tackle proper minor-closed graph classes, that is, we prove Theorem 1.1 for
graph classes excluding a fixed minor, by proving the following.
Theorem 4.1. For every fixed graph H, the k-Path problem restricted to H-minor-free graphs
can be solved in time OH (n2 m) using OH (kn) calls to Auxiliary Linkage on instances being
induced subgraphs of the input graph of size OH (k 24 ).
Our main technical result is the following:
Theorem 4.2. For every graph H, the class of H-minor-free graphs is (w, h)-decomposable for
w(k) = OH (k 22 ) and h = OH (1).
By plugging the above into Corollary 3.12, we obtain the desired polynomial Turing kernel,
but with worse bounds than promised by Theorem 4.1. To obtain better bounds, we need to
recall the folklore bound on the adhesion degree in sparse graph classes; for completeness, we
provide a full proof in Appendix A.1.
Lemma 4.3. Let G be a graph not containing H as a topological minor, and let (T, X ) be a
connected tree decomposition of G of width less than ` and adhesion h. Then the adhesion degree
of (T, X ) is bounded by f (h, H) · ` for some integer f (h, H) depending only on h and H.
This way, we conclude that H-minor-free graphs without k-paths have tree decompositions
of width OH (k 22 ), adhesion OH (1) and adhesion degree OH (k 22 ). We can use the algorithm of
Theorem 3.9 for h to find (by Lemma 3.10) an approximate decomposition of width OH (k 22 ),
adhesion OH (1), and thus, again using Lemma 4.3, of adhesion degree OH (k 22 ). Theorem 4.1
follows from Lemma 3.5 if we find separations using the algorithm of Lemma 3.7 applied to this
decomposition.
Thus, it remains to prove Theorem 4.2. For the proof, we use the graph minors structure
theorem, decomposing an H-minor-free graph G into parts ‘nearly embeddable’ in surfaces
(precise definitions are given in the next subsection). By carefully analyzing details of the
structure, we either find a large triconnected embedded part, which must contain a long path
by the following theorem of Chen et al. [8], or we tighten the graph structure to give a tree
decomposition where all parts are small (polynomial in k) and adhesions (‘boundaries’) between
them are of constant size.
Theorem 4.4 ([8]). There is a constant ε > 0 such that for every integer t, every triconnected
graph on n ≥ 3 vertices embeddable in a surface of (Euler) genus g contains a cycle of length at
2
least nε /2(2g+3) .
9
We note that Chen et al. phrase the theorem (more generally) for K3,t -minor-free graphs,
but a folklore edge-counting argument shows that graphs embeddable on a surface of genus g
are K3,t -minor-free for t = 2g + 3 (see e.g. [2]).
Two intertwined problems that arise with this approach is that torsos of decompositions are
not necessarily triconnected, and long paths in them do not necessarily imply long paths in the
original graph, because of virtual edges added in torsos. Torsos can be made triconnected if
their near-embeddings include cycles or paths around each vortex, but these may use virtual
edges in essential ways. On the other hand, the decomposition can be modified so that virtual
edges can be replaced with paths in the original graph, but this requires changes that remove
virtual edges, hence potentially removing paths around vortices and destroying triconnectedness.
Because of that, we need to go a little deeper and use a local, strong version of the structure
theorem from Graph Minors XVII [25]. For the same reason we cannot use existing algorithms
for finding the graph minors decompositions. Instead, we only prove the existence of a tree
decomposition of bounded adhesion, small width, and with nearly embeddable bags.
Global and local graph minor structure theorems
We now define near-embeddability and the graph minors α-decomposition.
Definition 4.5. For an integer α, an α-near embedding of G consists of:
(i) a set A of at most α vertices (called the apex set);
(ii) a family G0 ∪ G1 ∪ · · · ∪ Gα0 = G \ A of edge-disjoint subgraphs of G \ A, where:
•
•
•
•
G0 is called the embedded part,
V = {G1 , . . . , Gα0 } for some α0 ≤ α are called (large) vortices,
the intersection Ωi := V (G0 ) ∩ V (Gi ) is called the society of vortex Gi ,
vortices are pairwise vertex-disjoint,
(iii) an embedding of G0 in a surface of genus at most α such that for i ∈ [α0 ], the society Ωi is
embedded on the boundary of a disk whose interior is empty (i.e., does not intersect the
embedding or other disks), called the disk accomodating Gi ;
(iv) a linear ordering w1 , . . . , w|Ωi | of each vortex society Ωi , corresponding to its natural
ordering around its disk (for some choice of direction and starting point);
(v) for each large vortex Gi ∈ V, a path decomposition Z1 , . . . , Z|Ωi | of Gi of width at most α
such that wj ∈ Zj , for j ∈ [|Ωi |].
We denote such an α-near embedding as (A, G0 , V), with the embedding and path decompositions
only implicit in the notation.
Definition 4.6. A (graph minors) α-decomposition of a graph G consists of:
• a rooted tree decomposition (T, Y) of G of adhesion at most α;
• for each t ∈ V (T ), an α-near embedding (At , Gt0 , Vt ) of torso(G, X (t)).
In this subsection we prove the following variant of the Global Structure Theorem which
implies Theorem 4.2.
Theorem 4.7. For every graph H, there is a constant α = α(H) such that the following
holds, for any integer k: any graph G excluding H as a minor and without a k-path has an
α-decomposition of width at most α · k 22 .
We deduce Theorem 4.7 from a similarly modified variant of the Local Structure Theorem.
An α-near local embedding of a graph G is defined similarly to an α-near embedding, but we
allow an arbitrary number of ‘small vortices’ W and we allow the path decompositions of all
vortices to have only bounded adhesion instead of bounded width; thus arbitrarily complicated
graphs can hide behind vortices. We additionally require each large vortex to be surrounded by
a certain path. Formally:
10
Definition 4.8. A comb is a union of a path P with some mutually vertex-disjoint (possibly
trivial) paths that have their first vertex and no other vertex on P . The last vertices of those
paths are called the teeth of the comb, they are naturally ordered by P .
Definition 4.9. For an integer α, an α-near local embedding of G consists of:
(i) a set A of at most α vertices (called the apex set);
(ii) a family G0 ∪ G1 ∪ · · · ∪ Gn = G \ A of edge-disjoint subgraphs of G \ A, where:
•
•
•
•
•
•
•
G0 is called the embedded part,
V = {G1 , . . . , Gα0 } for some α0 ≤ α are called large vortices,
W = {Gα0 +1 , . . . , Gn } are called small vortices,
vortices intersect only in G0 : V (Gi ) ∩ V (Gj ) ⊆ V (G0 ) for i 6= j ∈ [n],
the intersection Ωi := V (G0 ) ∩ V (Gi ) is called the society of vortex Gi ,
large vortices are pairwise vertex-disjoint,
small vortices have societies of size ≤ 3;
(iii) an embedding of G0 in a surface of genus at most α such that for i ∈ [n], the society Ωi is
embedded on the boundary of a disk whose interior is empty (i.e., does not intersect the
embedding or other disks), called the disk accomodating Gi ;
(iv) a linear ordering w1 , . . . , w|Ωi | of each society Ωi , corresponding to its natural ordering
around its disk (for some choice of direction and starting point);
(v) for each large vortex Gi ∈ V, a path decomposition Z1 , . . . , Z|Ωi | of Gi of adhesion at most
α such that Zj ∩ Ωi = {wj−1 , wj } for 1 < j ≤ |Ωi | and Z1 S
∩ Ωi = {w1 }.
(vi) for each large vortex Gi ∈ V, a comb Ci in Gi ∪ G0 ∪ W whose teeth are vertices
w1 , . . . , w|Ωi | of Ωi , in the same order.
We denote such an α-near local embedding as (A, G0 , V, W), with the embedding and path
decompositions only implicit in the notation. The near embedding is said to respect a tangle T
if the large side of a separation in T \ A is never contained in a vortex W ∈ W or in a bag of
the decomposition of a vortex V ∈ V.
Definition 4.10. In a graph G, a tangle T controls an H-minor if there is a minor model in G
(defined by branch sets (Bh )h∈V (H) inducing vertex-disjoint connected subgraphs in G and with
an edge between Bh and Bh0 whenever hh0 ∈ E(H)) such that no branch set is fully contained
in a small side of a separation in T .
The following Local Structure Theorem follows from [25], as explained in [12] (we note that
while [12] assumes that G is H-minor-free, the original statement in [25] only requires that T
controls no H-minor – we will need this stronger version when dealing with topological minors).
Since this is crucial for our approach, we stress that the comb Ci contains a path of length at
least |Ωi | (a scrupulous reading of [25] implies even stronger statements, but this will suffice).
Theorem 4.11 ([12]). For every graph H there exist integers α, θ such that: for every graph
G and every tangle T in G of order ≥ θ that controls no H-minor, there is an α-near local
embedding of G which respects T .
We now improve the statement of the Local Structure Theorem 4.11 in steps: first requiring
small vortices to be ‘well attached’, then making the torsos triconnected, and finally deducing
bounds on width in terms of longest path length.
Lemma 4.12. In Theorem 4.11, we can additionally assume that:
(vii) for each small vortex Gi ∈ W and every v1 , v2 ∈ Ωi , there is a path in Gi between v1 and
v2 , with no internal vertices in Ωi .
11
Proof. Property (vii) is guaranteed by (9.1) in [25]. (As explained in [12], the border cells
of [25] are translated into bags of large vortex decompositions, so small vortices arise only from
internal cells; for an internal cell c0 , α(c0 )∗ is translated into a small vortex Gi , while γ(c̃0 ) is
translated into Ωi , giving exactly the statement we want. We also note that the proof of (9.1)
simply partitions a vortex with no v1 -v2 path into two vortices with at most two vertices in their
societies).
Definition 4.13. For an α-near local embedding (A, G0 , V, W) of G, define G∗0 to be the graph
obtained from G0 by adding an edge between every two consecutive vertices in each society Ωi ,
and a new vertex for each society of a large vortex, with edges to every vertex of the society.
The embedding of G0 is naturally extended to an embedding of G∗0 (the new vertices and edges
embedded in place of the accommodating disks).
Lemma 4.14. In Lemma 4.12, we can additionally assume that:
(viii) G∗0 is triconnected.
Proof. Suppose G∗0 has a separation (X, Y ) of order ≤ 2. Since societies of small vortices induce
cliques in G∗0 , they are contained in X or Y . Similarly for each large vortex Gi ∈ V, its society
Ωi , together with the new vertex in G∗0 with Ωi as neighborhood, induces a wheel in G∗0 , which
is triconnected, hence contained in X or Y . In the language of [25], this translates back into a
partition I, J of cells, whose intersection corresponds to at most two embedded vertices v, v 0
(that is, γ(I) ∩ γ(J) = {v, v 0 }). But this contradicts (11.1) in [25], which states that such an
intersection has size at least 3.
Lemma 4.15. In Lemma 4.14, we can additionally assume that there is a constant α0 depending
on H only such that:
(ix) for any integer k, if G does not contain a path of length ≥ k, then |V (G0 )| < α0 · k 22 .
Proof. For a given graph H, let α be the constant given by Theorem 4.11. By Theorem 4.4, there
is a universal constant ε such that any triconnected graph G0 embeddable in a surface of genus
2
2
at most α contains a cycle of length at least |G0 |ε /2(2α+2) . Let c := 2ε and α0 := (α · 2(2α+2) )1/ε .
For an H-minor-free graph G, suppose the α-near local embedding given by Lemma 4.14 has
|V (G0 )| ≥ α0 · k c . Then G∗0 is a triconnected graph (by property (viii)) embedded in a surface of
genus at most α (by Definition 4.13) with at least as many vertices, so G∗0 must contain a cycle
2
C of length at least (α0 k c )ε /2(2α+2) = α · k 2 .
If the society Ωi of any vortex of the α-near local embedding of G has at least k vertices,
then G contains a path of at least that length (in the comb from property (vi)), in which case
the lemma follows. Otherwise, there are at most α large vortices, hence at most α · (k − 1)
vertices in their societies, and at most α virtual vertices that were inserted as the centers of the
wheels. Let P be the longest subpath of C between any two such vortex-related vertices (or any
subpath not visiting any such vertices, if there are less than two). Then P is a path of length
2
∗
∗
at least α·k
α·k = k in G0 whose edges in E(G0 ) \ E(G0 ) came only from small vortices, by the
definition of G∗0 (Definition 4.13). By property (vii), these edges can be replaced with paths
(each of length at least 2) in the corresponding small vortices, giving a path in G. Only one or
two consecutive edges can come from the same small vortex (since their societies have at most 3
vertices), so the resulting path has length at least k.
The constant from Theorem 4.4 given by Chen et al. [8] is ε = log1729 2, hence the universal
constant here is c = 2ε = 2 log2 (1729) < 22.
Finally, we deduce Theorem 4.7 (the Global Structure Theorem) from Theorem 4.11 (the
Local Structure Theorem) by a standard induction, exactly as done by Diestel et al. [10, Theorem
4]. (Note that in case the excluded graph H is planar, we can already conclude the theorem
12
trivially from the fact that there is a tree-decomposition of width OH (1) [23]). The only difference
is that in the Local Structure Theorem we add the bound on the size of the embedded part
from Lemma 4.15. In the proof from [10], every bag of the created decomposition is either
constructed as a set of size bounded as OH (1), or is constructed from an α-near local embedding
by taking the vertices of: the embedded part G0 , the set of apices A, and for each large vortex,
the intersection of every two consecutive bags of its path decomposition, each of size at most
α = OH (1). The number of bags in a path decomposition of a large vortex is equal to the size
of its society, and large vortex societies are disjoint subsets of V (G0 ). Therefore every tree
decomposition bag constructed in the proof has size at most |V (G0 )| + α + α · |V (G0 )| = OH (k 22 ),
by Lemma 4.15. This proves the additional condition we require in Theorem 4.7.
5
Excluding a topological minor
In this section we tackle graph classes excluding a topological minor, that is, we prove Theorem 1.1
by proving the following.
Theorem 5.1. For every fixed graph H, the k-Path problem restricted to H-topological-minorfree graphs can be solved in time OH (n2 m) using OH (kn) calls to Auxiliary Linkage on
instances being induced subgraphs of the input graph of size k OH (1) .
This follows as before from the following decomposability theorem. Note the exponent in the
polynomial bound on width (bag size) now depends on H.
Theorem 5.2. For every graph H, the class of H-topological-minor-free graphs is (w, h)decomposable for w(k) = k OH (1) and h = OH (1).
We proceed with the proof of the above theorem. Grohe and Marx [15] proved that when
excluding a topological minor, graphs admit a similar structure as for excluding a minor, but
apart from nearly embeddable parts, one needs to consider parts that have bounded degree
except for a bounded number of vertices:
Definition 5.3. A graph G has (a, d)-bounded degree (for a, d ∈ N) if ≤ a vertices of G have
degree > d in G.
The main ‘Global Structure Theorem’ of [15] is that graphs excluding a topological minor H
admit a tree decomposition of bounded adhesion, where the torso of every bag is either nearly
embeddable (as when excluding a minor), or has (h, d)-bounded degree, for d = OH (1). We still
need the slightly stronger notion of near embeddability (with combs) for our proofs, which is
why we will work with the following definition of ‘Local Structure’, analogous to α-near local
embeddings.
A star decomposition of a graph is a (rooted) tree decomposition (T , X ) where T is a star:
a root node troot with only leaf nodes attached. We call the bag X (troot ) the root bag and the
remaining ones leaf bags. The root torso is torso(G, X (troot )). A star decomposition respects a
tangle T if the large side of any separation in T is not fully contained in any single leaf bag.
Grohe and Marx [15, Lemma 6.12] proved the following Local Structure Theorem.1
Theorem 5.4 ([15]). For every integer k there exist integers θ, d, k 0 such that: for every graph G
excluding Kk as a topological minor and every tangle T of order ≥ θ that controls a Kk0 -minor,
there is a star decomposition of G that respects T , has adhesion < θ, and such that the root
torso has (k, d)-bounded degree.
1
Grohe and Marx [15] use a slightly stronger definition of respecting a tangle, which is however equivalent to
the conjunction of this one and the assertion that all adhesions in the star decomposition are strictly smaller than
the order of T , which we prefer to state separately. They also assume a slightly weaker condition than that of a
tangle controlling a minor (which they call not being removed by the tangle), but it is trivially implied.
13
Corollary 5.5. For every graph H there exist integers θ, α, d such that: for every graph G
excluding H as a topological minor and every tangle T in G of order ≥ θ, G has one of the
following:
• a star decomposition that respects T , has adhesion α, and such that the root torso has
(|V (H)|, d)-bounded degree, or
• an α-near local embedding that respects T , satisfying properties (vii)-(ix).
Proof. For k = |V (H)| let the constants given by Theorem 5.4 be θ0 , d0 , k 0 . Let α00 , θ00 be the
constants given by Theorem 4.11 for graphs excluding Kk0 as a minor. We prove the claim for
θ = max(θ0 , θ00 ), α = max(θ0 , α00 ) and d = d0 . If the tangle T of G controls no Kk0 -minor, then
the claim follows from Theorem 4.11. Otherwise T does control an Kk0 -minor and Theorem 5.4
applies.
The main theoretical tool we use, besides the Grohe and Marx decomposition in the form of
Corollary 5.5, is the following lower bound on cycles in graphs of bounded degree by Shan [27]
(we note the constant ∆ − 1 is asymptotically optimal; for small ∆, previous results by Chen et
al. [7] give a slightly better bound).
Theorem 5.6 ([27]). If G is a triconnected graph with maximum degree at most ∆ ≥ 425, then
G has a cycle of length at least n1/ log2 (∆−1) /4 + 2.
With this in hand, we can proceed to bounding the bags of almost bounded degree.
Lemma 5.7. In Corollary 5.5, we can additionally assume that in the first outcome, for any
integer k, if G has no k-path, then the root bag has size k OH (1) .
Proof. Without loss of generality assume that θ ≥ α(|V (H)|+2), in the statement of Corollary 5.5
(otherwise increase θ). Let us first ensure that the star decomposition (T, X ) is connected.
Claim 5.8. We can additionally assume that
For every leaf node t ∈ V (T ), the set X (t) \ X (troot ) is connected, and
(C1)
X (t) \ X (troot ) has edges to each vertex in the adhesion X (t) ∩ X (troot )
(C2)
Proof. If (C1) fails we can partition the leaf bag into many leaf bags. If (C2) fails we can remove
the adhesion vertex in question from the leaf bag.
y
We need to define a variant of the torso that is similarly connected as the root torso, but
has no virtual edges (so that long paths in it imply long paths in G). Construct G∗root from G as
follows: contract X (t) \ X (troot ) to a single vertex xt for each leaf t ∈ T and merge any such
vertices xt , xt0 that have equal neighborhoods (corresponding to the same adhesion) into one.
Let X be the set of all those new vertices.
Claim 5.9. If G has no k-path, then neither does G∗root .
Proof. For each leaf node t ∈ V (T ), the set X (t)\X (troot ) is connected by (C1) and can therefore
be contracted to a single vertex xt by repeated edge contractions. Since G∗root is a subgraph of
the graph obtained by these edge contractions, it follows that G∗root is a minor of G. As minor
operations do not increase the length of a longest path, the claim follows.
y
Claim 5.10. G∗root has (|V (H)|, max(α, d + 2d ))-bounded degree.
14
Proof. Observe that G∗root can be obtained from the root torso of (T, X ) by removing virtual
edges and adding the vertices of X. Each vertex in X has degree equal to the size of some
corresponding adhesion, which is ≤ α. Each vertex v in the torso gains as neighbors only vertices
in X, which correspond to different (as vertex subsets) adhesions containing v; since these
adhesions induced cliques containing v in the torso (which had (|V (H)|, d)-bounded degree),
their number is bounded by 2d (unless v was already one of the ≤ |V (H)| vertices of high
degree).
y
Claim 5.11. If G∗root does not contain a path of length at least k, then G∗root has a tree decomposition of adhesion ≤ |V (H)| + 2 and width k OH (1) .
Proof. Assume that G∗root does not contain a path of length at least k. By the previous
claim, in G∗root one can delete ≤ |V (H)| vertices to obtain a subgraph of degree bounded by
∆ := max(α, d + 2d ); name this subgraph G†root . Consider the Tutte decomposition [29] of G†root .
In modern terms (cf. [18, Theorem 1]), it is a tree decomposition of G†root of adhesion at most
two, whose torsos are triconnected topological minors of G†root . Consequently, each torso again
has degree bounded by ∆ and cannot contain a path of length at least k. By Theorem 5.6, if n
denotes the size of such a bag, then nε /4 + 2 < k, where ε = 1/ log2 (max(425, α, d + 2d ) − 1).
Thus the size of each bag is n < (4(k − 2))1/ε = k O(d+log2 α) = k OH (1) . Adding the ≤ |V (H)|
deleted vertices back to every bag, we obtain a tree decomposition of G∗root as claimed.
y
Denote the root torso of (T, X ) as Groot .
Claim 5.12. For every tree decomposition of G∗root , one can find a tree decomposition of Groot
with adhesion and maximum bag size at most α times larger than the adhesion and maximum
bag size of the original tree decomposition.
Proof. Let (S, Y) be a tree decomposition of G∗root . Define Y 0 (s) as Y(s) with vertices xt ∈ X
replaced by all of NG∗root (xt ). We claim (S, Y 0 ) is a tree decomposition of Groot . Every edge of
Groot is either an edge of G∗root − X, still contained in the same bag, or a virtual edge between two
vertices of an adhesion, hence contained in NG∗root (xt ) for some xt ∈ X and thus in a bag Y 0 (s)
(for an s such that Y(s) contained xt ). To show that the subset Y 0−1 (v) := {s ∈ S : v ∈ Y 0 (s)}
is connected, for any given vertex v ∈ V (Groot ), observe that it is the union of Y −1 (v) with
Y −1 (xt ) for all xt ∈ X adjacent to v in G∗root ; since such sets Y −1 (v) and Y −1 (xt ) are connected
in S and intersect (as xt is adjacent to v), their union is connected. Each vertex in each bag has
been replaced by at most |NG∗root (xt )| vertices, which is the size of an adhesion in (T, X ) and
hence at most α. It follows that the maximum bag size increases by a factor at most α.
To bound any adhesion, observe similarly that for s1 , s2 ∈ V (S), the adhesion equals
[
Y 0 (s1 ) ∩ Y 0 (s2 ) = ((Y(s1 ) ∩ Y(s2 )) \ X) ∪
N (xt1 ) ∩ N (xt2 ).
xt1 ∈Y(s1 ),xt2 ∈Y(s2 )
For any vertex v ∈ N (xt1 ) ∩ N (xt2 ) (for some xti ∈ Y(si )), the set Y −1 (v) is connected and
must intersect both Y −1 (xt1 ) and Y −1 (xt2 ). Hence either Y −1 (xt1 ) contains both s1 and s2 , or
Y −1 (xt2 ) does, or Y −1 (v) does. This means xt1 , xt2 or v is contained in Y(s1 ) ∩ Y(s2 ). Therefore
[
Y 0 (s1 ) ∩ Y 0 (s2 ) ⊆ ((Y(s1 ) ∩ Y(s2 )) \ X) ∪
N (xt ).
xt ∈Y(s1 )∩Y(s2 )
This again implies each adhesion increases at most α times.
y
By the last three claims, assuming G has no k-path, Groot has a tree decomposition (S, Y) of
adhesion α0 := α(|V (H)| + 2) and width k OH (1) . We focus this decomposition on the unique bag
that, in some sense, respects the tangle T . Let us first see how T orients separations of Groot .
15
For a separation (A, B) of Groot , every adhesion X (t) ∩ X (troot ) of (T, X ) induces a clique in
Groot , hence it is fully contained in A or in B; thus (A, B) can be extended to a separation of G
by adding all of X (t) to A or B, correspondingly, with an arbitrary choice for X (t) such that
X (t) ∩ X (troot ) ⊆ A ∩ B. We call such separations of G extensions of the separation (A, B) of
Groot . For a separation (A, B) of order ≤ θ of Groot , observe that either all extensions of (A, B)
and no extensions of (B, A) belong to T , or vice-versa: this follows from tangle axioms and the
fact that (T, X ) respects T (that is, each X (t) such that X (t) ∩ X (troot ) ⊆ A ∩ B is small w.r.t.
T , so putting it one side of an extension (A0 , B 0 ) instead of the other cannot change whether A0
is small, as otherwise the two small sides together with X (t) would cover all of G).
We now orient edges of the decomposition (S, Y) according to T , to find a (unique) bag we
can focus on. For each edge s1 s2 ∈ E(S), if S1 , S2 denote the two connected components of
S −s1 s2 containing s1 , s2 , respectively, then (Y(S1 ), Y(S2 )) is a separation with |Y(S1 )∩Y(S2 )| =
|Y(s1 ) ∩ Y(s2 )| ≤ α0 ≤ θ, hence either all of its extensions or all extensions of (Y(S2 ), Y(S1 ))
belongs to the tangle T . We direct the edge s1 s2 to either s2 or s1 accordingly (towards the
side with large extensions). After repeating this for every edge of S, we find a node s0 with no
outgoing edges. This means that if we root S at s0 and S1 , S2 , . . . , S` are the subtrees of S − s0 ,
then all extensions of (Y(Si ), Y(S − Si )) belong to T , for i ∈ [`].
We can now forget about some details of the tree decomposition and define the following
star decomposition (S 0 , Y 0 ) of Groot : its nodes are sroot := s0 and its direct children s1 , . . . , s` in
S, while its bags are Y 0 (sroot ) := Y(s0 ) and Y 0 (si ) := Y(Si ) for i ∈ [`]. By construction (S 0 , Y 0 )
is a star decomposition of Groot of adhesion α0 , with a root bag of size k OH (1) , which satisfies
the following:
all extensions of (Y(si ), Y(S − si )) are in T , for i ∈ [`].
(2)
The following claim thus concludes the proof of the lemma, by extending this star decomposition of Groot to a star decomposition of G respecting T , with a root torso of almost bounded
degree and size k OH (1) .
Claim 5.13. If Groot has a star decomposition (S, Y) of adhesion α0 ≤ θ satisfying (2), then
G has a star decomposition of adhesion max(α, α0 ) respecting T where the root torso has
(|V (H)|, d + α0 · d)-bounded degree and its vertex set is equal to the root bag of (S, Y).
Proof. Let (S, Y) be a star decomposition of Groot satisfying (2). Without loss of generality
assume that for every leaf s ∈ V (S), every vertex of the adhesion Y(s) ∩ Y(sroot ) has a neighbor
in Y(s) \ Y(sroot ) (otherwise we can remove it from Y(s)). Every adhesion of the (larger) star
decomposition (T, X ) of G induces a clique in Groot by definition, hence we can choose for each
t ∈ V (T ) a node s(t) ∈ V (S) such that X (t) ∩ X (troot ) ⊆ Y(s(t)). We construct a new star
decomposition (T 0 , X 0 ) of G from (T, X ) as follows: the root bag becomes Y(sroot ), leaves t with
s(t) = sroot and their bags remain unchanged, the remainingSleaves are deleted, and each leaf s
of (S, Y) is added as a new leaf of T 0 , with X 0 (s) := Y(s) ∪ t : s(t)=s X (t).
It is straightforward to check that (T 0 , X 0 ) is a star decomposition of G of adhesion max(α, α0 ),
as follows. Every edge is covered by a bag in X 0 , because it was covered by a bag in X , which is
either a subset of a bag in X 0 , or in X (troot ), and hence in a bag of Y, which are again subsets of
bags in X 0 . Every vertex occurs in bags X 0 of a connected subtree of T 0 , because if a vertex occurs
in two different leaf bags of X 0 , then either it occurs in two different leaf bags of Y and hence in
Y(sroot ) = X 0 (troot ), or in two leaf bags X (t) and X (t0 ) with s(t) 6= s(t0 ) and hence in X (troot ),
which means it occurs in the adhesions of X (t) and X (t)0 ; these adhesion are contained in Y(s(t))
and Y(s(t0 )) respectively, so the vertex must be contained in Y(sroot ) = X 0 (troot ) as well. Finally
to check the sizes of adhesions, observe that each adhesion of (T, X 0 ) is either (for leaf nodes
t coming from T ) a subset of an adhesion in X , namely X 0 (t) ∩ X 0 (troot ) ⊆ X (t) ∩ X (troot ), or
16
(for leaf nodes s coming from S) equal to an adhesion of (S, Y):
[
X (t) ∩ Y(sroot ) = Y(s) ∩ Y(sroot ),
X 0 (s) ∩ X 0 (troot ) = Y(s) ∪
t : s(t)=s
since every vertex in X (t) ∩ Y(sroot ) ⊆ X (t) ∩ X (troot ) is contained in Y(s(t)).
To show that the new decomposition respects T , suppose to the contrary that some leaf
0
bag contains the large side B of a separation (A, B) ∈ T . Then either this
S bag is X (t) = X (t)
0
for t ∈ V (T ), contradicting that (T, X ) respects T , or X (s) = Y(s) ∪ t : s(t)=s X (t) for some
S
s ∈ V (S). Then since (X 0 (s), s0 6=s∈V (S) X 0 (s0 )) is a separation induced by the adhesion at X 0 (s),
of order ≤ α0 (as shown in the previous paragraph), X 0 (s) (which contains B) S
must be the large
side of this separation in T . But this is an extension of the separation (Y(s), s0 6=s∈V (S) Y(s0 )),
contradicting (2).
The part that is not entirely straightforward is bounding the degrees in the root torso of
the new star decomposition (T 0 , X 0 ). Observe that this new torso can be obtained from the
root torso Groot of (T, X ) by taking the subgraph induced by X 0 (troot ) = Y(sroot ), removing
certain virtual edges (from adhesions corresponding to deleted leaves) and adding virtual edges
between every two vertices of each new adhesion X 0 (s) ∩ X 0 (troot ), s ∈ V (S). We need to bound
the number of such edges added to a vertex v ∈ Y(sroot ), assuming it was not already one of
the |V (H)| high-degree vertices of the root torso of (T, X ). Thus v had at most d neighbors in
Groot . Since for each new adhesion X 0 (s) ∩ X 0 (troot ) = Y(s) ∩ Y(sroot ) (as shown above) which
contains v we assumed that v has a neighbor in Y(s) \ Y(sroot ) (and these are disjoint sets, for
distinct s), it follows that v received new edges from at most d distinct new adhesions, at most
α0 edges from each. Therefore, the degree of all but |V (H)| vertices in the root torso of (T 0 , X 0 )
is bounded by d + α0 · d, that is, Corollary 5.5 holds with this constant in place of d.
y
This concludes the proof of Lemma 5.7.
As before, the global decomposition stated in Theorem 5.2 follows from the Local Structure
Theorem in Corollary 5.5 improved by the bound on bag size in Lemma 5.7. The construction
only changes in that it can take root bags from the local structure (the star decomposition) given
by the first outcome of Corollary 5.5 and use them as bags of the global decomposition, without
any modifications. We note this can also be seen as an instance of a general theorem (11.1)
by Robertson and Seymour [24] which allows to turn ‘local decompositions’ (with respect to a
tangle) into global ones; however, giving a precise translation would involve too much notational
overhead.
6
Adding a modulator
In this section we prove Theorem 1.2 in a more general setting of Section 3. More precisely,
Theorem 1.2 follows directly from the following theorem via Theorems 4.2 and 5.2.
Theorem 6.1. One can solve in polynomial time a given k-Path instance (G, k), given access
to a set M ⊆ V (G) such that G − M admits a tree decomposition of width less than w and
adhesion h = O(1), and an oracle that solves the Auxiliary Linkage problem for instances
(G0 , k 0 , S, (Ri )ri=1 ) with G0 being a subgraph of G, r, k 0 ≤ k, |S| ≤ |M | + O(1), and |V (G0 )| being
bounded polynomially in k, w, and |M |.
Contrary to Section 3, in this section we are not precise with the polynomial dependency
on the parameters k, w, and |M |, as well as the dependency on the adhesion h = O(1) of the
decomposition. This is due to the fact that the latter dependency on h is substantial, most
notably in the exponent of the polynomial bound on the size of the oracle calls. Therefore, the
17
result of this section is a purely theoretical result classifying the aforementioned parameterization
as admitting a polynomial Turing kernel.
Decomposition. We apply the algorithm of Theorem 3.9 to the graph G − M and separation
size h. Since h = O(1), the algorithm runs in polynomial time. By Lemma 3.10, the output
decomposition (T0 , X0 ) has width wO(1) and adhesion O(1).
We modify the decomposition (T0 , X0 ) as follows. For every node t ∈ V (T ) we group the
children t0 of t according to the adhesions X (t) ∩ X (t0 ). For every group U ⊆ NT (t), we add a
new child tU of t with X (tU ) = X (t) ∩ X(t0 ), t0 ∈ U , and reattach the nodes of U as children of
tU . Then, we duplicate tU into max(1, |U | − 1) copies, arranged in a binary tree, with the nodes
of U being different leaves of the binary tree. Let (T, X ) be the final decomposition. In this
manner, we maintain the property that (T, X ) has width wO(1) and adhesion O(1), while every
node has a polynomial in w number of children.
Marking. We continue the proof of Theorem 6.1 with a somewhat standard marking process.
For an integer 0 ≤ k 0 ≤ k − 2 and two vertices u, v ∈ M , define a (u, v, k 0 )-path in G as a (simple)
path with k 0 + 2 vertices: two endpoints u and v and exactly k 0 internal vertices, all in G − M .
Note that a (u, v, 0)-path is a path consisting of an edge uv. Similarly, for 0 ≤ k 0 ≤ k − 1 and a
vertex u ∈ M , a (u, k 0 )-path in G is a path with k 0 + 1 vertices: one endpoint being u and all
other vertices in G − M .
For every 0 ≤ k 0 ≤ k − 2 and every u, v ∈ M , we compute an inclusion-wise maximal family
0
Pu,v,k0 of internally vertex-disjoint (u, v, k 0 )-paths. Similarly, for every 0 ≤ k 0 ≤ k − 1 and u ∈ M
0
0
we compute an inclusion-wise maximal family Pu,k
0 of internally vertex-disjoint (u, k )-paths.
The computation can be done via the previously established kernel for decomposable graph
classes.
0
0
Lemma 6.2. The families Pu,v,k
0 and Pu,k 0 can be computed in polynomial time with the access
to the assumed Auxiliary Linkage oracle.
0
Proof. We focus on Pu,v,k
0 ; the proof for the second family is analogous. Since the required
family is inclusion-wise maximal, it suffices to find its paths one-by-one. That is, we show how
to find, in a given induced subgraph of G − (M \ {u, v}) obtained by removing the internal
vertices of the previously-found paths, a next path of the desired length k 0 ≤ k between u and v
(or conclude that none exists and the packing is maximal). For that, we use a slight modification
of the previously established kernel in Section 3 for decomposable graph classes (Corollary 3.12).
More precisely, first observe that the Reduction Rule is still correct when we are looking for
0
a k -path between fixed endpoints u and v (as opposed to any k 0 -path) as long as the endpoints
do not lie in the set A. Thus, we can (just as in Corollary 3.12) iteratively use the decomposition
(T, X ) of G − M , reintroduce u and v to every bag, and find a separation (A0 , B 0 ) by Lemma 3.7
such that the Reduction Rule applies, and iterate this until we find the k 0 -path between u and v
or conclude there is none, as in Lemma 3.5.
0
We define Pu,v,k0 to be a subfamily of k + 1 arbitrarily chosen paths from Pu,v,k
0 , or let
0
Pu,v,k0 = Pu,v,k0 if |Pu,v,k0 | < k + 1. We define Pu,k0 similarly.
Let |M | = s. Let X1 be the set of all vertices of G − M that lie on some path in any of
the sets Pu,v,k0 or Pu,k0 (u, v ∈ M , k 0 ≤ k). Note that, as we picked at most k + 1 paths to the
families Pu,v,k0 and Pu,k0 , each consisting of at most k − 1 vertices from G − M , for each of at
most k choices of the length k 0 ≤ k, and 2s + s ≤ 2s2 choices for the endpoint(s), we have
|X1 | < (k + 1) · (k − 1) · k · 2s2 .
For every x ∈ X1 , pick one node t(x) ∈ V (T ) such that x ∈ X (t(x)). Let B1 = {t(x) | x ∈ X1 }.
Let B2 ⊆ V (T ) consist of all of B1 , the root of T , and the lowest common ancestor of every pair
18
(t1 , t2 ) ∈ B1 × B1 . Let X2 = X (B2 ). It is folklore that |B2 | ≤ 2|B1 |, hence:
|B2 | ≤ 2(k + 1) · (k − 1) · k · 2s2 ,
|X2 | = wO(1) k 3 s2 .
Let C be the set of connected components of T − B2 . As B2 is closed under taking lowest
common ancestors of nodes, we have that |NT (C)| ≤ 2 for any C ∈ C: one of the elements of
NT (C) is a parent of the root of C (if it exists), and additionally one vertex of C may have a
child in NT (C).
Reducing a large component. Our main goal is to show that if the bags of any component
C contain in total more than m = poly(k, w, s) vertices of G, that is, |X (C)| > m, then one can
use the Reduction Rule from Section 3 to reduce the graph G by at least one vertex. Fix such a
component C.
Let t0 ∈ V (C) be the lowest node in C such that, if D is the set of nodes of C in the subtree
of T rooted in t0 , then |X (D)| > m.
Let Z = X (D) ∩ X2 . Recall that |NT (C)| ≤ 2 due to the fact that B2 is closed under taking
lowest common ancestors. By the properties of the tree decomposition, every vertex of Z lies in
X (t) for some t ∈ NT (C). Since the adhesions are of constant size, we have that |Z| = O(1).
Let A = X (D) \ Z. Note that A ∩ X2 = ∅. Since in (T, X ) every node has a number
of children bounded polynomially in w, the choice of t0 ensures that |X (D)| ≤ poly(w) · m.
Furthermore, note that NG (A) ⊆ Z ∪ M ; in particular, |NG (A)| ≤ |M | + O(1).
The marking process ensures the following crucial property.
Lemma 6.3. The set A is guarded by Z.
Proof. Let P be a k-path in G that contains a maximum possible number of vertices from X1 ;
we claim that P is guarded by Z in A. Assume the contrary, let Q be a traverse of P through
A for which neither endpoint belongs to Z. Since P is not completely contained in G[A], and
NG (A) ⊆ Z ∪ M , we have that Q is either a (u, v, k 0 )-path for some u, v ∈ M and 1 ≤ k 0 ≤ k − 2
or a (u, k 0 )-path for some u ∈ M and 1 ≤ k 0 ≤ k − 1. Assume the first case; the proof for the
second case is analogous.
0
Since A does not contain any vertex of X2 ⊇ X1 , from the maximality of Pu,v,k
0 we infer
0
that Pu,v,k0 ( Pu,v,k0 (since it contains Q) and, consequently, |Pu,v,k0 | = k + 1. Thus, at least one
(u, v, k 0 )-path Q0 ∈ Pu,v,k0 has no internal vertices on the k-path P . Consequently, by replacing
Q with Q0 on P , we obtain a simple k-path P 0 with strictly more vertices of X1 than P . This is
a contradiction.
Thus, if m = poly(k, w, s) is large enough, we can apply the Reduction Rule for the set A
and the guard Z; recall that |Z| = O(1) while m − O(1) < |A| ≤ poly(w) · m. The Reduction
Rule deletes at least one vertex after invoking a number of calls to the Auxiliary Linkage
oracle on the subgraph G[N [A]]; note that N [A] ⊆ A ∪ M ∪ Z.
Wrap up. Recall that we have shown that one can apply the Reduction Rule if there exists a
component C ∈ C with |X (C)| > m for some m bounded polynomially in k, w, and s. However,
since |B2 | ≤ 2(k + 1) · (k − 1) · k · 2s2 and a node of T has poly(w) children, there are at most
poly(w)k 3 s2 components. Consequently, if the reduction is not applicable, we have |V (G)|
bounded polynomially in k, w, and s. Such an instance can be solved with a single call to
Auxiliary Linkage with k 0 = k, S = ∅, r = 1, and R1 = ∅. This finishes the proof of
Theorem 6.1 and, consequently, of Theorem 1.2.
19
7
Conclusions
We significantly extended the graph classes on which k-Path has a polynomial Turing kernel.
In addition, we showed that even an instance that does not belong to such a class, but has a
small vertex modulator whose deletion makes it so, can be solved efficiently using small queries
to an oracle. A subdivision-based argument (cf. [13]) shows that we cannot generalize much
beyond H-topological-minor-free graphs without settling the problem in general. In particular,
the existence of a polynomial Turing kernel for graphs of bounded expansion implies its existence
in general graphs.
While our narrative focused on k-Path, after small modifications our techniques can also be
applied to prove analogues of Theorems 1.1 and 1.2 for the k-Cycle problem of detecting a
simple cycle of length at least k. The main difficulty in adapting our arguments to k-Cycle is
the fact that, a priori, the only cycles of length at least k may be arbitrarily much larger than k.
However, this issue can easily be resolved in the following way. Since a cycle is contained within
a single biconnected component, a Turing kernelization can decompose its input into biconnected
components and solve the problem independently in each of them. We then start by testing for
the existence of a path with k 2 vertices using the algorithms developed in the paper. If there is
a path of length k 2 in a biconnected component, then by a classic theorem of Dirac [11] there
is a cycle of length at least k, and we are done. If no such path exists, then the longest cycle
in G has length less than 2k, and we can continue under the guarantee that the cycle we are
looking for has length at least k and less than 2k. In this setting, our arguments can be easily
adapted. In particular, the absence of a path of length k 2 implies the existence of suitable tree
decompositions from which reducible separations can be extracted.
A significant portion of the technical work in this paper was devoted to modifying the graph
minors decomposition to obtain the win/win that either answers the problem or finds a reducible
separation. In this way, the algorithmic question has driven a challenging graph-theoretic project.
It would be interesting to find more problems amenable to such an approach. We conclude with
some concrete open problems. Does k-Path have a polynomial Turing kernel on chordal graphs?
How about Induced k-Path or Directed k-Path, on planar graphs?
References
[1] A. M. Ambalath, R. Balasundaram, C. R. H., V. Koppula, N. Misra, G. Philip, and M. S. Ramanujan.
On the kernelization complexity of colorful motifs. In Proc. 5th IPEC, pages 14–25, 2010. doi:
10.1007/978-3-642-17493-3_4.
[2] D. Archdeacon. Topological graph theory. a survey. Congressus Numerantium, 115(5-54):18, 1996.
[3] F. Barbero, C. Paul, and M. Pilipczuk. Exploring the complexity of layout parameters in tournaments
and semi-complete digraphs. In Proc. 44th ICALP, 2017. In press. arXiv:1706.00617.
[4] D. Binkele-Raible, H. Fernau, F. V. Fomin, D. Lokshtanov, S. Saurabh, and Y. Villanger. Kernel(s)
for problems with no kernel: On out-trees with many leaves. ACM Trans. Algorithms, 8(4):38, 2012.
doi:10.1145/2344422.2344428.
[5] H. L. Bodlaender, E. D. Demaine, M. R. Fellows, J. Guo, D. Hermelin, D. Lokshtanov, M. Müller,
V. Raman, J. v. Rooij, and F. A. Rosamond. Open problems in parameterized and exact computation
- IWPEC 2008. Technical Report UU-CS-2008-017, Utrecht University, 2008.
[6] H. L. Bodlaender, B. M. P. Jansen, and S. Kratsch. Kernelization lower bounds by cross-composition.
SIAM J. Discrete Math., 28(1):277–305, 2014. doi:10.1137/120880240.
[7] G. Chen, Z. Gao, X. Yu, and W. Zang. Approximating longest cycles in graphs with bounded
degrees. SIAM J. Comput., 36(3):635–656, 2006. doi:10.1137/050633263.
[8] G. Chen, X. Yu, and W. Zang. The circumference of a graph with no K3,t -minor, II. J. Comb.
Theory, Ser. B, 102(6):1211–1240, 2012. doi:10.1016/j.jctb.2012.07.003.
[9] M. Cygan, D. Lokshtanov, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Minimum bisection is fixed
parameter tractable. In Proc. STOC 2014, pages 323–332. ACM, 2014. doi:10.1145/2591796.
2591852.
20
[10] R. Diestel, K. Kawarabayashi, T. Müller, and P. Wollan. On the excluded minor structure theorem
for graphs of large tree-width. J. Comb. Theory, Ser. B, 102(6):1189–1210, 2012. doi:10.1016/j.
jctb.2012.07.001.
[11] G. A. Dirac. Some theorems on abstract graphs. Proceedings of the London Mathematical Society,
s3-2(1):69–81, 1952. doi:10.1112/plms/s3-2.1.69.
[12] J. Fröhlich and T. Müller. Linear connectivity forces large complete bipartite minors: An alternative
approach. J. Comb. Theory, Ser. B, 101(6):502–508, 2011. doi:10.1016/j.jctb.2011.02.002.
[13] J. Gajarský, P. Hlinený, J. Obdrzálek, S. Ordyniak, F. Reidl, P. Rossmanith, F. S. Villaamil, and
S. Sikdar. Kernelization using structural parameters on sparse graph classes. J. Comput. Syst. Sci.,
84:219–242, 2017. doi:10.1016/j.jcss.2016.09.002.
[14] V. Garnero and M. Weller. Parameterized certificate dispersal and its variants. Theor. Comput.
Sci., 622:66–78, 2016. doi:10.1016/j.tcs.2016.02.001.
[15] M. Grohe and D. Marx. Structure theorem and isomorphism test for graphs with excluded topological
subgraphs. SIAM J. Comput., 44(1):114–159, 2015. doi:10.1137/120892234.
[16] D. Hermelin, S. Kratsch, K. Soltys, M. Wahlström, and X. Wu. A completeness theory for polynomial
(Turing) kernelization. Algorithmica, 71(3):702–730, 2015. doi:10.1007/s00453-014-9910-8.
[17] F. Hüffner, C. Komusiewicz, and M. Sorge. Finding highly connected subgraphs. In Proc. 41st
SOFSEM, pages 254–265, 2015. doi:10.1007/978-3-662-46078-8_21.
[18] B. M. P. Jansen. Turing kernelization for finding long paths and cycles in restricted graph classes.
J. Comput. Syst. Sci., 85:18–37, 2017. doi:10.1016/j.jcss.2016.10.008.
[19] B. M. P. Jansen and D. Marx. Characterizing the easy-to-find subgraphs from the viewpoint of
polynomial-time algorithms, kernels, and Turing kernels. In Proc. 26th SODA, pages 616–629, 2015.
doi:10.1137/1.9781611973730.42.
[20] S. Kolay and F. Panolan. Parameterized algorithms for deletion to (r, `)-graphs. In Proc. 35th
FSTTCS, pages 420–433, 2015. doi:10.4230/LIPIcs.FSTTCS.2015.420.
[21] D. Lokshtanov, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Manuscript, 2017.
[22] J. Nešetřil and P. Ossona de Mendez. Sparsity: Graphs, Structures, and Algorithms, volume 28 of
Algorithms and Combinatorics. Springer, 2012. doi:10.1007/978-3-642-27875-4.
[23] N. Robertson and P. D. Seymour. Graph minors. V. Excluding a planar graph. J. Comb. Theory,
Ser. B, 41(1):92–114, 1986. doi:10.1016/0095-8956(86)90030-4.
[24] N. Robertson and P. D. Seymour. Graph minors. X. Obstructions to tree-decomposition. J. Comb.
Theory, Ser. B, 52(2):153–190, 1991. doi:10.1016/0095-8956(91)90061-N.
[25] N. Robertson and P. D. Seymour. Graph minors: XVII. Taming a vortex. J. Comb. Theory, Ser. B,
77(1):162–210, 1999. doi:10.1006/jctb.1999.1919.
[26] A. Schäfer, C. Komusiewicz, H. Moser, and R. Niedermeier. Parameterized computational complexity of finding small-diameter subgraphs. Optim. Lett., 6(5):883–891, 2012. doi:10.1007/
s11590-011-0311-5.
[27] S. Shan. Homeomorphically Irreducible Spanning Trees, Halin Graphs, and Long Cycles in 3connected Graphs with Bounded Maximum Degrees. PhD thesis, Georgia State University, 2015.
URL: http://scholarworks.gsu.edu/math_diss/23/.
[28] S. Thomassé, N. Trotignon, and K. Vuskovic. A polynomial Turing-kernel for weighted independent set in bull-free graphs. In Proc. 40th WG, pages 408–419. Springer, 2014. doi:
10.1007/978-3-319-12340-0_34.
[29] W. T. Tutte. Connectivity in graphs. Mathematical expositions. University of Toronto Press, 1966.
21
A
A.1
Omitted proofs
Proof of Lemma 4.3
Consider a node t ∈ V (T ), let Tt be the subtree of T rooted in t, and let Gt = G[X (Tt )]. Note
that (Tt , X ) is a tree decomposition of Gt with t being a root; with regards to this decomposition,
we consider the torso Ht := torso(Gt , X (t)). Note that the connectivity of (T, X ) implies that
Ht can be obtained from Gt [X (t)] by turning the neighborhood NGt (C) into a clique for every
connected component C of Gt − X (t). Hence every adhesion to a child of t corresponds to a
clique in Ht , and we may bound the adhesion degree by bounding the number of distinct cliques
in Ht .
To show the bound on the number of cliques, we show a longer, more elaborate process that
turns Gt into Ht . Let C be the family of connected components of Gt − X (t). For every C ∈ C
and every vertex v ∈ N (C) ⊆ X (t), pick one neighbor xv,C ∈ N (v) ∩ C. Let SC be a minimal
connected subgraph of C that contains every vertex xv,C for v ∈ N (C). Note that by minimality,
SC is a tree with at most |N (C)| ≤ h leaves and thus less than h vertices of degree at least 3.
Delete from Gt all vertices and edges of G[C] that do not lie in SC , and suppress every node
of degree 2 in SC that is not of the form xv,C for v ∈ N (C). By performing this operation for
every component C ∈ C, we obtain a topological minor G1t of Gt with the following property:
the graph Gt [X (t)] remains untouched in G1t , while every C ∈ C is turned into its topological
minor C 1 with N (C) = N (C 1 ) and |C 1 | < 2|N (C)| ≤ 2h. Since G does not contain H as a
topological minor, neither does G1t .
We now switch to the theory of graphs of bounded expansion; for an introduction and more
notation we refer to the textbook [22]. A graph class G is of bounded expansion if there exists a
function ∇ : Z≥0 → Z≥0 such that for every radius r, every graph H ∈ G, and every r-shallow
minor H 0 of H, we have |E(H 0 )|/|V (H 0 )| ≤ ∇(r).
Let G2t be a graph obtained from G1t by contracting every component C 1 for C ∈ C into
a single vertex xC . Then, as |C 1 | < 2h, G2t is an 2h-shallow minor of G1t . Let G3t be a graph
obtained from G2t by replacing every vertex xC for C ∈ C with a clique KC of |N (C)| copies of
xC . Finally, note that Ht can be obtained from G3t by contracting every vertex of KC onto a
distinct vertex of N (C), that is, Ht is a 1-shallow minor of G3t .
Since taking an O(1)-shallow-minor or replacing every vertex with a clique of constant size in
a graph from a class of bounded expansion gives a graph also from a class of bounded expansion
(but possibly worse function ∇) [22], we have that there exists a graph class G of bounded
expansion (with the function ∇ depending on H and h) such that Ht ∈ G. Since graphs of
bounded expansion have bounded degeneracy and therefore have linearly many cliques [22,
Lemma 3.1], we have that Ht contains at most f 0 (h, H) · |X (t)| cliques for some constant f 0 (h, H)
depending on h and H. Thus, in Lemma 4.3 we can take f (h, H) = 1 + f 0 (h, H) to accommodate
for the additional adhesion corresponding to the parent of t.
22
| 8cs.DS
|
Generative Sensing: Transforming Unreliable Sensor Data
for Reliable Recognition
Lina J. Karam, Tejas Borkar, Yu Cao, Junseok Chae
School of ECEE, Arizona State University, Tempe, Arizona, USA
{karam, tsborkar, Yu.Cao, Junseok.Chae}@asu.edu
Abstract
This paper introduces a deep learning enabled
generative sensing framework which integrates low-end
sensors with computational intelligence to attain a high
recognition accuracy on par with that attained with highend sensors. The proposed generative sensing framework
aims at transforming low-end, low-quality sensor data
into higher quality sensor data in terms of achieved
classification accuracy. The low-end data can be
transformed into higher quality data of the same modality
or into data of another modality. Different from existing
methods for image generation, the proposed framework is
based on discriminative models and targets to maximize
the recognition accuracy rather than a similarity measure.
This is achieved through the introduction of selective
feature regeneration in a deep neural network (DNN). The
proposed generative sensing will essentially transform
low-quality sensor data into high-quality information for
robust perception. Results are presented to illustrate the
performance of the proposed framework.
Keywords: generative sensing; quality; resolution; visible;
infrared; NIR; IR; image; multi-modal; recognition;
classification; deep learning.
1. Introduction
Recent advances in machine learning coupled with the
accessibility of compact and low-cost multimedia sensors,
such as visible image and infrared (IR) cameras, have
fundamentally altered the way humans live, work and
interact with each other. However, the performance of
available sensors is still significantly constrained by the
quality/cost tradeoff of the sensors themselves and by the
sensitivity of deep learning algorithms to variations in
sensor data quality and sensor modality. For example,
image quality may be impacted by environmental factors,
i.e., lighting conditions, and specifications of sensors, i.e.,
number of pixels, dynamic range, noise, etc. Under lowlighting conditions, the visible wavelength image suffers
while the IR image may not largely suffer from the dark
Figure 1. Effect of image quality on DNN predictions when
using an AlexNet model [4] trained on high-resolution
images. Degradation severity increases from left to right for
white additive noise (a) and Gaussian blur (b). It can be seen
that, while the quality degradation does not hinder the human
ability to classify the images, the object class label predicted
by the DNN changes significantly even at low degradation
levels.
environment; the quality of both (visible and IR) may
suffer due to low-end sensor performance.
Low-power low-profile (thin) sensors are often limited
to produce low-resolution (blurry) and low SNR (noisy)
images. Dodge and Karam [1] showed that such
degradations, even at relatively small levels, can cause a
significant drop in the classification performance of stateof-the-art DNNs which are typically trained using highresolution images. Before the work of Dodge and
Karam [1], the image quality was an aspect very often
overlooked while designing deep learning based image
classification systems. Dodge and Karam also showed very
recently [2][3] that, although DNNs perform better than or
on par with humans on pristine high-resolution images, the
DNN classification performance is still significantly much
lower than human performance on images with
degradations in quality. Fig. 1 illustrates that, while
humans are capable of recognizing with a reasonable
accuracy objects in low-resolution blurred images or in low
SNR noisy images, DNNs that are trained on pristine
images predict incorrect class labels even at a relatively
low-level of perceivable degradation.
A visible wavelength image sensor, denoted as image
sensor, is a core part of digital cameras and smart
phones [5][6]. The cost of image sensors has been
aggressively scaled largely due to the high-volume
consumer market products. In addition to the cost, the
technical specification of image sensors, including the
number of pixels, color contrast, dynamic range, and power
consumption, meet almost all demands of consumer market
products. IR sensors, on the other hand, have been mostly
used for specific needs such as surveillance- and trackingbased applications in the military domain, yet some started
penetrating the consumer market recently [7][8]. IR
sensors typically cost significantly higher than image
sensors to produce IR images at an equivalent resolution
primarily due to the relatively low volume market.
Generally those sensors show a trend of “the higher the cost
is, the better the delivered performance.”
This work develops a deep learning based framework,
which we refer to as generative sensing, that enables
attaining the classification accuracy of a higher quality,
high-end sensor while only using a low-quality, low-end
sensor. Furthermore, the high-end sensor and low-end
sensor can be of different types, in which case the proposed
framework can be seen as transforming one type of sensor
(e.g., near infrared (NIR) or IR sensor) into another type of
sensor (e.g., visible wavelength image sensor) in terms of
achieved classification performance. This is performed
through feature regeneration for improved classification.
This is important for enabling low-power and low-cost
sensing platforms without compromising the recognition
performance.
This paper is organized as follows. Section 2 discusses
related work. Section 3 describes the proposed generative
sensing framework. Illustrative performance results are
presented in Section 4 and a conclusion is given in
Section 5.
2. Related Work
Existing methods [9]-[14] were proposed for image
generation with applications to colorization (converting a a
grayscale image to a color image) [9]-[11] including IR
colorization [11], artistic style transfer [12][13], and
dataset augmentation [14]. These existing methods are
mainly concerned with the generation of new data samples
that are similar to existing reference data samples. To
achieve this goal, they optimize an objective function that
maximizes a similarity measure between the generated data
and the reference data. In contrast, our proposed generative
sensing is based on discriminative models and optimizes a
target-oriented objective function (e.g., a regularized
categorical cross-entropy loss function) to maximize the
classification accuracy rather than a similarity measure.
In [15], we presented a method (DeepCorrect) to
identify the convolutional filters that are most sensitive to
degradations in image quality and to correct the degraded
activations of these filters. We showed that correcting only
a fraction of the most susceptible filter activations using
small low complexity convolutional filter blocks results in
a significant performance improvement on popular datasets
including ImageNet [16]. We also showed that our
DeepCorrect model can achieve a classification accuracy
higher than fine-tuning (which retrains all network
parameters) while only training a fraction of the network
parameters, and that it can also train faster than fine-tuning.
Our proposed generative sensing framework builds on and
extends our DeepCorrect work [15] to sensors under
varying conditions (e.g.,
varying illumination,
environment, and acquisition characteristics in addition to
resolution) and varying modalities (e.g., NIR, IR,
Ultrasound, RADAR, in addition to visible wavelength).
3. Proposed Generative Sensing
A block diagram of the proposed generative sensing
framework is shown in Fig. 2. In the following description,
the term high-end (low-end) sensor data refers to sensor
data that result in a relatively high (low) classification
accuracy when using a pre-trained deep neural network
(DNN) 𝜙𝜙.
During the training phase (Fig. 2a), given high-end
sensor data of type X and low-end sensor data of type Y,
where types X and Y can be the same (X=Y) or different
(X≠Y), feature maps are obtained by applying DNN 𝜙𝜙 to
each set of data separately. Let Փhigh and Փlow denote,
respectively, the set of feature maps resulting from the
high-end sensor data and low-end sensor data. Using a
distance measure to quantify differences between colocated features in Փhigh and Փlow, feature difference maps
ΔՓ are obtained. We are mainly interested in locating
feature differences that result in a significant drop in
classification accuracy (e.g., a drop greater than a certain
user-defined value) by finding features in Փlow that are
significantly different from co-located features in Փhigh
based on the change in classification accuracy. One
approach for locating such features is by measuring the
classification accuracy drop when a feature in Փhigh is
replaced by its co-located Փlow feature while keeping all
other features in Փhigh unchanged [15]. Another approach
for determining significant feature differences can be based
on maximizing the drop in classification accuracy by
replacing clusters of features in Փhigh with clusters of
features in Փlow. Classification-based feature difference
maps ΔՓ can thus be obtained with values corresponding
to the classification accuracy drop. ΔՓ can also be
thresholded to produce a binary significance difference
tensor (significance mask) with a value of 1 denoting a
significant feature difference at a location and a value of 0
Figure 2. Proposed Generative Sensing Framework.
denoting an insignificant difference. Selective feature
regeneration is performed by learning transformations to
be applied only to features in Փlow corresponding to
significant ΔՓ values, while leaving all other features in
Փlow unchanged. Such transformations can be learned by
using relatively small residual learning units as in [15] or
by adopting other learning models. We refer to such
transformations as generative transformation units or
simply as generative units for short. The resulting
generative sensing network 𝜙𝜙𝑔𝑔𝑔𝑔𝑔𝑔 consists of the original
pre-trained DNN 𝜙𝜙 augmented with generative units that
are applied to a select limited number of significant
features. During the testing phase (Fig. 2b), low-end type
Y sensor data is classified using the generative sensing
DNN 𝜙𝜙𝑔𝑔𝑔𝑔𝑔𝑔 to obtain an improved classification accuracy.
In this way, our proposed generative sensing aims to
convert the low-end sensor into a high-end sensor in terms
of matched classification accuracy.
When the generative unit takes the form of a multilayer network with trainable parameters, the generative
transform can be estimated by determining the trainable
parameters Wgen of each generative unit so as to minimize
a target-oriented loss function such as [15]:
𝑀𝑀
1
𝐸𝐸�𝐖𝐖𝒈𝒈𝒈𝒈𝒈𝒈 � = 𝜆𝜆 𝜌𝜌�𝐖𝐖𝒈𝒈𝒈𝒈𝒈𝒈 � + � ℒ( 𝜙𝜙𝑔𝑔𝑔𝑔𝑔𝑔 (𝐱𝐱 𝑖𝑖 ), 𝑦𝑦𝑖𝑖 )
𝑀𝑀
𝑖𝑖=1
where ℒ(. , . ) is a classification loss function, 𝑦𝑦𝑖𝑖 is a target
output label for input 𝐱𝐱 𝑖𝑖 , 𝜙𝜙𝑔𝑔𝑔𝑔𝑔𝑔 (. ) is the output of the
network with the selectively applied generative units,
which we refer to as generative sensing network, 𝜌𝜌(. ) Is
a regularization term (e.g., ℓ1 or ℓ2 norm), 𝜆𝜆 is a
regularization parameter, and 𝑀𝑀is the total number of data
samples in the training set. The generative units are
designed to have a relatively very small number of
parameters as compared to DNN 𝜙𝜙.
4. Performance Results
In order to show the performance of our proposed
generative sensing framework and its ability to generalize
to different tasks (e.g., face recognition, scene recognition),
different input modalities (e.g., RGB, NIR, and IR), and
different sensor sizes/resolutions, we adopt for the baseline
pre-trained DNN 𝜙𝜙 of Fig. 2 the baseline AlexNet [4] that
has been pre-trained for object recognition on the relatively
high-quality visible-wavelength (RGB) ImageNet dataset
[16]. For varying tasks and input modalities, we make use
of the RGB-IR SCface face recognition dataset [17] and the
EPFL RGB-NIR Scene recognition dataset [18]. Unlike the
task of face recognition, where the aim is to assign the face
in a test image to one of the known subjects in a database,
the goal of scene recognition is to classify the entire scene
of the image.
The SCface dataset [17] is primarily designed for
surveillance-based face recognition. It consists of images
acquired in the visible (RGB) as well as infrared spectrum
(IR) for 130 subjects, with one frontal mugshot image per
subject, for each input modality in addition to other nonfrontal mugshots. We make use of the frontal mugshot
images. The EPFL RGB-NIR Scene dataset consists of 9
scene categories with at least 50 images per class, for both
visible (RGB) and near-infrared spectra (NIR) [18].
We simulate the effect of decreasing the input sensor
resolution by blurring the original images by a Gaussian
kernel, where the size of the blur kernel is set to 4 times the
blur standard deviation σb. Using a blur standard deviation
σb ∈ {0, 1, 2, 3, 4, 5, 6} for Gaussian blur, we simulate 7
successively decreasing levels of sensor resolution, with
σb = 0 representing the original high resolution sensor and
σb = 6 representing the lowest sensor resolution.
We use the output of the second fully-connected layer
in the baseline AlexNet model as a deep feature extractor
Table 1: Top-1 accuracy for images of the SCface dataset [17] acquired under different sensor resolutions and different input
modalities, respectively. For each modality (RGB and IR), bold numbers show best accuracy for each resolution level.
Method-Modality
Baseline-RGB
Generative Sensing- RGB
Baseline-IR
Generative Sensing-IR
Sensor Resolution: Blur Level
σb = 0
σb = 1
σb = 2
σb = 3
σb = 4
σb = 5
σb = 6
0.9923
0.9538
0.9769
0.9000
0.7538
0.9461
0.7923
0.8777
0.4384
0.9000
0.4769
0.8077
0.3230
0.8692
0.1076
0.7538
0.1461
0.7692
0.0461
0.6538
0.1000
0.6846
0.0076
0.5077
0.0770
0.6461
0.0076
0.4692
Avg
0.4043
0.8241
0.3450
0.7098
Table 2: Top-1 accuracy for images of the RGB-NIR Scene dataset [18] acquired under different sensor resolutions and different
input modalities, respectively. For each modality (RGB and NIR), bold numbers show best accuracy for each resolution level.
Method-Modality
Baseline-RGB
Generative Sensing- RGB
Baseline-NIR
Generative Sensing-NIR
σb = 0
0.9444
0.9333
0.7629
0.7518
σb = 1
0.8466
0.8555
0.6733
0.7200
Sensor Resolution: Blur Level
σb = 2
σb = 3
σb = 4
0.7644
0.6177
0.4622
0.8511
0.8555
0.8333
0.5911
0.4000
0.3088
0.7177
0.6977
0.6622
for the face recognition and scene recognition tasks with
images acquired under different sensor resolutions (high to
low) and modalities (RGB, NIR, IR). Since the baseline
AlexNet model has not been trained on the SCface [17] and
EPFL RGB-NIR Scene dataset [18], for each dataset and
for each modality, we train a new fully connected layer
with a softmax activation on top of deep features that are
extracted only from the original dataset images (i.e., σb =
0) and evaluate the performance of such a deep feature
extractor for different levels of sensor resolution (i.e., σb
∈{0, 1, 2, 3, 4, 5, 6}) and for each modality. The new fully
connected layer acts like a linear classifier and has a
dimensionality equal to the number of classes in the
corresponding dataset.
Next we replace the baseline AlexNet DNN 𝜙𝜙 (layers 1
to 7) with our generative sensing DNN 𝜙𝜙𝑔𝑔𝑔𝑔𝑔𝑔 (i.e., baseline
AlexNet augmented with generative units) while using the
same previously learnt final fully connected layer with
softmax activation on top of it as a linear classifier. The
generative units were only trained using the ImageNet
dataset [16] augmented by adding degraded versions of the
ImageNet images with varying levels of blur to simulate
varying sensor resolutions [15]. From the results shown in
Tables 1 and 2 it can be concluded that, although no images
from the SCface [17] and RGB-NIR Scene [18] datasets
were used for training, the produced generative units
generalize well.
Tables 1 and 2 present the performance results on all
levels of sensor resolution for the SCface dataset [17] and
the RGB-NIR Scene dataset [18], respectively. It should be
noted that we do not learn a new classifier to act on top of
our generative sensing deep feature extractor, but instead
just use the one trained for the baseline AlexNet deep
feature extractor.
σb = 5
0.3511
0.8355
0.2488
0.6377
σb = 6
0.2911
0.8200
0.2200
0.6222
Avg
0.6110
0.8548
0.4578
0.6870
As shown in Tables. 1 and 2, for both the visible
spectrum and the near-infrared/infrared spectrum, the
sensor resolution significantly affects the accuracy of the
baseline feature extractor, with a 59% and 64% drop in
respective average accuracies for the SCface dataset [17],
a 35% and 40% drop in respective average accuracies for
the RGB-NIR Scene dataset [18] relative to the original
sensor resolution (i.e., σb = 0).
For the SCface dataset [17], the generative sensing
feature extractor significantly outperforms the baseline
feature extractor with a 103% and 105% relative
improvement in mean accuracy for the visible spectrum
and infrared spectrum, respectively. Similarly, for the
RGB-NIR Scene dataset [18], the generative sensing
feature extractor significantly outperforms the baseline
feature extractor with a 40% and 50% relative
improvement in mean accuracy for the visible spectrum
and the near-infrared spectrum, respectively. The large
performance gap between the generative sensing feature
extractor and the baseline AlexNet feature extractor
highlights the generic nature of our modality-invariant and
sensor resolution-invariant features that are learnt by our
generative sensing models.
5. Conclusion
This work presents a deep learning based generative
sensing framework for attaining increased classification
accuracy on par with the classification accuracy of a highend sensor while only using a low-end sensor. The highand low-end sensors can be of different types, in which
case the proposed framework can be seen as transforming
one type of sensor (e.g., NIR or IR sensor) into another type
of sensor (e.g., visible spectrum image sensor) in terms of
classification performance. This is achieved through
learned transformations that perform selective feature
regeneration for improved classification. This is important
for enabling robust low-power and low-cost sensing
platforms that can work under varying conditions without
compromising the recognition performance.
6. References
[1] S. Dodge and L. Karam, “Understanding how image quality
affects deep neural networks,” International Conference on
Quality of Multimedia Experience (QoMEX), 6 pages, 2016.
[2] S. Dodge and L. Karam, “A study and comparison of human
and deep learning recognition performance under visual
distortions,”
International
Conference
on
Computer
Communications and Networks, 7 pages, Jul.-Aug. 2017.
[3] S. Dodge and L. Karam, “Can the early human visual system
compete with deep neural networks?” ICCV Workshop on Mutual
Benefits of Cognitive and Computer Vision, 7 pages, October
2017.
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet
classification with deep convolutional neural networks,”
Advances in Neural Information Processing Systems, pp. 1097–
1105, 2012.
[5] E. Fossum, “CMOS image sensors: Electronic camera-on-achip,” IEEE Transactions on Electron Devices, vol. 44, no. 10,
pp. 1689-1698, 1997.
[6] M. Bigas, E. Cabruja, J. Forest, and J. Salvi., “Review of
CMOS image sensors,” Microelectronics Journal, vol. 37, no. 5,
pp. 433-451, 2006.
[7] J. Sobrino, F. Del Frate, and M. Drusch, “Review of thermal
infrared applications and requirements for future high-resolution
sensors,” IEEE Transactions on Geoscience and Remote Sensing,
vol. 54, no. 5, pp. 2963-2972, 2016.
[8] A. Rogalski, P. Martyniuk, and M. Kopytko, “Challenges of
small-pixel infrared detectors: a review,” Reports on Progress in
Physics, vol. 79, no. 4, pp. 046501, 2016.
[9] A. Deshpande, J. Rock, and D. Forsyth, "Learning LargeScale Automatic Image Colorization," IEEE International
Conference on Computer Vision (ICCV), pp. 567-575, 2015.
[10] R. Zhang, P. Isola, and A.A. Efros, "Colorful Image
Colorization," European Conference on Computer Vision
(ECCV), In: Leibe B., Matas J., Sebe N., Welling M.(eds)
Computer Vision – ECCV 2016. Lecture Notes in Computer
Science, vol 9907, pp. 649-666, Springer, Cham.
[11] M. Limmer and H.P.A. Lensch, "Infrared Colorization Using
Deep Convolutional Neural Networks," IEEE International
Conference on Machine Learning and Applications (ICMLA), pp.
61-68, 2016.
[12] L. A. Gatys, A.S. Ecker, and M. Bethge, "A Neural
Algorithm of Artistic Style," arXiv:1508.06576 [cs.CV], 2015.
[13] L. A. Gatys, A.S. Ecker, and M. Bethge, "Image Style
Transfer Using Convolutional Neural Networks," IEEE CVPR,
pp. 2414-2423, 2016.
[14] V. V. Kniaza, V. S. Gorbatsevicha, and V. A. Mizginova,
"THERMALNET: A Deep Convolutional Network for Synthetic
Thermal Image Generation," 2nd International ISPRS Workshop
on PSBB, pp. 41-45, May 2017.
[15] T. Borkar and L. Karam, " DeepCorrect: Correcting DNN
models against Image Distortions," arXiv:1705.02406 [cs.CV],
2017.
[16] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S.
Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg,
and L. Fei-Fei, “ImageNet Large Scale Visual Recognition
Challenge,” International Journal of Computer Vision (IJCV),
vol. 115, no. 3, pp. 211–252, 2015.
[17] M. Grgic, K. Delac, S. Grgic, SCface - surveillance cameras
face database, Multimedia Tools and Applications Journal, vol.
51, No. 3, pp. 863-879, February 2011.
[18] M. Brown and S. Süsstrunk, Multispectral SIFT for Scene
Category Recognition, IEEE International Conference on
Computer Vision and Pattern Recognition (CVPR), pp. 177-184,
2011.
| 1cs.CV
|
Multimodal and Multiscale Deep Neural Networks for
the Early Diagnosis of Alzheimer’s Disease using
structural MR and FDG-PET images
Donghuan Lu1,* , Karteek Popuri1 , Weiguang Ding1 , Rakesh Balachandar1 , Mirza Faisal
Beg1 , and for the Alzheimer’s Disease Neuroimaging Initiative+
1 School
of Engineering Science, Simon Fraser University, Burnaby, V5A 1S6, Canada
arXiv:1710.04782v1 [cs.CV] 13 Oct 2017
* [email protected]
+ Data
used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and
implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete
listing of ADNI investigators can be found at: http:
//adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf
ABSTRACT
Alzheimer’s Disease (AD) is a progressive neurodegenerative disease. Amnestic mild cognitive impairment (MCI) is a common
first symptom before the conversion to clinical impairment where the individual becomes unable to perform activities of daily
living independently. Although there is currently no treatment available, the earlier a conclusive diagnosis is made, the earlier
the potential for interventions to delay or perhaps even prevent progression to full-blown AD. Neuroimaging scans acquired from
MRI and metabolism images obtained by FDG-PET provide in-vivo view into the structure and function (glucose metabolism) of
the living brain. It is hypothesized that combining different image modalities could better characterize the change of human
brain and result in a more accuracy early diagnosis of AD. In this paper, we proposed a novel framework to discriminate normal
control(NC) subjects from subjects with AD pathology (AD and NC, MCI subjects convert to AD in future). Our novel approach
utilizing a multimodal and multiscale deep neural network was found to deliver a 85.68% accuracy in the prediction of subjects
within 3 years to conversion. Cross validation experiments proved that it has better discrimination ability compared with results
in existing published literature.
Introduction
Alzheimer’s disease (AD), the most common dementia, affecting 1 out of 9 people over the age of 65 years1 . Alzheimer’s
diseases involves progressive cognitive impairment, commonly associated with early memory loss, requiring assistance for
activities of self care during advanced stages. Alzheimer’s is posited to evolve through a prodromal stage which is commonly
referred to as the mild cognitive impairment(MCI) stage and 10 - 15% of individuals with MCI, progress to AD2 each year.
With improved life expectancy, it is estimated that about 1.2% of global population will develop Alzheimer’s disease by 20463
thereby affecting millions of individuals directly, as well as many more indirectly through the effects on their families and
caregivers. Current Alzheimer’s research targets reliable prodromal identification of patients harbouring Alzheimer’s pathology,
for reasons that early intervention could potentially change the course of illness. Clinical diagnosis involves rigorous evaluation
to rule out non-Alzheimer’s causes for cognitive decline but this is however limited by specificity of identifying prodromal AD.
Hence, we see the necessity of a tool to reliably detect and identify prodromal Alzheimer’s disease.
Efforts to understand AD pathology in the past resulted in identifying neuroimaging as one of the promising tool for
prodromal diagnosis4 . Neuroimaging involving magnetic resonance imaging (MRI)5 and fluorodeoxyglucose positron emission
tomography (FDG-PET)6 were the unique imaging modalities recognized as useful tools to identify individuals with prodromal
AD. MRI offers structural details such as texture, thickness, density and shape of various brain regions, while FDG-PET
measures the resting state glucose metabolism7 , reflecting the functional activity of the underlying tissue6 . FDG-PET and
MRI are frequently employed neuroimaging techniques in computer-aided diagnosis of neurodegenerative diseases. There has
been considerable efforts to use structural MRI8–10 or FDG-PET11, 12 or a combination with other biomarkers13, 14 to develop
automated computer aided tools for prodromal diagnosis of Alzheimer’s disease.
Deep neural networks have been studied extensively and proven to have the best performance for many recognition tasks15 .
The application of deep neural networks in recognition of AD-related patterns has also attracted interests in its application for
prodromal AD16–18 . By applying deep neural network to extract features, such as stacked autoencoder(SAE) or Deep Boltzmann
Machine(DBM), these approaches outperform other popular machine learning methods, e.g., support vector machine (SVM)
and random forest techniques. However, one of the major hurdles preventing applications of deep learning in neuroimaging
is that it requires a large data set to train the model, while the available number of images are limited to several hundred or
thousand which are much less than feature dimension of data sample. To overcome this challenge, one popular approach is to
segment images into patches and extract features from each patch16–18 . However, down sampling the input data can result in the
loss of discriminative information which could be a potential reason why previous methods haven’t achieved satisfying accuracy
for this diagnostic task. A common extension to pattern mining and feature extraction for image recognition is multiscale
processing19, 20 . By extracting features at different resolutions or scales, multiscale features can better characterize images for
classification task and a recent study indicates it can also improve the classification performance of deep neural networks21 .
Therefore, we are proposing a novel approach to combine multiscale and multimodal processing with deep neural network
for the early diagnosis of AD. Through cross validation experiments with more than 1000 subjects, we demonstrated that 1) our
network can learn hidden patterns from multiscale and multimodal features for the detection of AD pathology; 2) our approach
outperform previous methods regarding the discriminative task of potential AD subjects; and, 3) our network can identify the
subjects who are going to convert to AD in 3 years with an accuracy of 85.68% which is a promising result.
Methods
There are two major steps in the proposed framework: 1)image preprocessing: segment both MRI scans and FDG-PET images
into patches, and extract features from each patch; and, 2)classification: train a deep neural network to learn the patterns that
discriminate AD pathology, and then use it to classify individuals with AD pathology.
Data
Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database
(adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator
Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, PET, other biological markers, and
clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI)
and early Alzheimer’s disease (AD).
For a comprehensive validation of the proposed method, it is emphasized that all the available ADNI subjects(N = 1242)
with both a T1-weighted MRI scan and FDG-PET image at the time of preparation of this manuscript were used in this study.
These subjects were categorized into 5 groups: 1) stable Normal controls (sNC): 360 subjects diagnosed to be NC at baseline
and remained the same at the time of preparation of this manuscript; 2) stable MCI (sMCI): 409 subjects diagnosed to be MCI
at all time points(at least for 2 years); 3) progressive NC (pNC): 18 subjects evaluated to be NC at baseline but have progressed
to probable AD at the time of preparation of this manuscript; 4) progressive MCI (pMCI): 217 subjects evaluated to be MCI at
baseline and progressed to probable AD; 5) stable Alzheimer’s disease (sAD): 238 subjects diagnosed to be AD for all available
time points. Demographic and clinical information of the subjects are shown in Table 1. Numbers in brackets are the number of
male and female subjects in the second row,, while in the rest 3 rows the two number represent the minimum and maximum
value of age, education year and MMSE(Mini–Mental State Examination) score. It was worth mentioning that each subject
could have multiple scans at different time points. In total there were 2402 FDG-PET scans and 2402 MRI images used in
this study. Detailed descriptions of the ADNI subject cohorts, image acquisition protocols procedures and post-acquisition
preprocessing procedures can be found at http://www.adni-info.org.
Mean(min-max)
sNC
sMCI
pNC
pMCI
sAD
Count (M/F)
Age
Education
MMSE
360(167/193)
73.4(60-94)
16.5(6-20)
29.1(24-30)
409(239/170)
74(56-91)
15.8(7-20)
28.0(22-30)
18(11/7)
77(68-84)
15.7(12-20)
29.4(27-30)
217(126/91)
74(55-89)
16.0(8-20)
26.5(9-30)
238(141/97)
75(55-90)
15.3(4-20)
23.2(18-27)
Table 1. Subject Demographics. In the second row, the first number in each cell is the total number of subjects and the
numbers in brackets represents the number of male and female. In the last 3 rows, the first number in each cell is the mean and
the two numbers in brackets are the minimum and maximum value of age, education year and MMSE score, respectively.
Image Processing
Unlike typical image recognition problems where deep learning has shown to be effective, our data set is relatively small. Hence
directly using this smaller database of images to train the deep neural network is not likely to deliver high classification accuracy.
2/12
However, contrary to typical image recognition tasks, where the images contain large heterogeneity, the images in this database
are all human brain images acquired with similar pose and scale which show relatively much less heterogeneity in comparison.
Therefore we applied following processing steps to extract patch-wise features as shown in Figure 1: FreeSurfer 5.122 was used
to segment each T1 structural MRI image into gray matter and white matter followed by subdivision of the gray matter into
87 anatomical regions of interest (ROI). The Freesurfer segmentation were quality controlled by an expert neuroanatomist
and any errors noted were manually corrected. Then, for a standard T1 MRI image, a voxel-wise k-means clustering based
on spatial coordinates was performed to segment each ROI into patches based on its spatial information23 . The size of patch
was predefined as 500, 1000 and 2000 voxels in this study and resulted into 1488, 705 and 343 patches, respectively. It was
designed to keep enough detailed information as well as avoiding too large feature dimension considering the limited number
of data samples were available in this study. Subsequently, each ROI of the standard template MRI was registered to the same
ROI of every target image via a high-dimensional non-rigid registration method (LDDMM24 ). The registration maps were
then applied to the patch-wise segmentation of the standard template. This transformed the template segmentation into each
target MRI space so the target images were subdivided into the same number of patches. It worth mentioning that after the
transformation, the size of a template patch in different images is not the same due to non-rigid registration encoding local
expansion/contraction and hence is one of the features used to represent the regional information of a given structural brain
scan. Then, for each target subject, the FDG-PET image of the subject was co-registered to its skull-stripped T1 MRI scan
with a rigid transformation using FSL-FLIRT program25 based on normalized mutual information. The degrees of freedom
(DOF) was set as 12 and Normalized correlation was used as cost function. The mean intensity in the brainstem region of the
FDG-PET image was the chosen reference to normalize the voxel intensities in that individual brain metabolism image, because
brainstem region was most unlikely to be affected by AD. The mean intensity of each patch was used as an element to form the
feature vector to represent the metabolism activity, and the volume of each patch was used to represent the brain atrophy.
Figure 1. Flowchart of extracting patch-wise features from MRI scans and FDG-PET images. Each subject was segmented
into patches through registration to a segmented template. Patch volume and mean intensity of FDG-PET were extracted as the
feature to represent each patch.
Multiscale Deep Neural Network
With the features extracted from MRI and FDG-PET images, we trained a Multimodal Multiscale Deep Neural Network
(MMDNN) to perform the classification. As shown in Fig 2, the network consists of two parts. The first parts was 6 independent
deep neural network (DNN) corresponding to each scale of a single modality. The second part was another DNN used to fuse
the features extracted from these 6 DNNs. The input data of this DNN was the concatenated latent representation learned from
each single DNN. The DNNs in two parts shared the same structure. For each DNN, the number of nodes for each hidden
layer were set as 3N, 34 N and 100 respectively, where N denotes the dimension of input feature vector. The number of nodes
was chosen to explore all possible hidden correlation across features from different patches in the first layer and gradually
reduce the number of features in the following layers to avoid over-fitting. We trained each DNN with two steps, unsupervised
3/12
pre-training and supervised fine-tuning, respectively. Then all the parameters of MMDNN was tuned together.
Figure 2. Multimodal Multiscale Deep Neural Network. 1488, 705 and 343 are input feature dimension(number of patches)
extracted from different scales. For each layer, its number of nodes was denoted by the number on its top left. For each scale of
each image modality, its patch-wise biomarkers were feed to a single DNN. The features from these 6 DNNs were fused by
another DNN to generate the final classification result.
• Unsupervised Pre-training For unsupervised pre-training step, each DNN was trained as a stacked-autoencoder(SAE).
Autoencoder is an artificial neural network used for unsupervised learning of non-linear hidden patterns from input data.
It consists of three layers, input layer, hidden layer and output layer, for which two nearby layers are fully-connected.
Three functions are used to define an autoencoder, encoding function, decoding function and loss function. In this study,
encoding function is defined as: y = s(W1 x + b1 ), where x is the input data, y is the latent representation, W1 is the weight
matrix, b1 is the bias term and s is the activation function for which we used rectified linear function max(0, x). Similarly,
decoding function can be represented as: z = s(W2 y + b2 ), where we constrained it with tied weight W1 = W T and z is
the reconstructed data which is supposed to be close to input x. Squared error 12 ||x − z||2 is applied as loss function to
optimize the network. The hypothesis is that the latent representation can capture the main factors of variation in the
data. Comparing with another popular unsupervised feature learning method principle component analysis (PCA), the
activation function enables the network to capture non-linear factors of data variation, especially when multiple encoders
and decoders are stacked to form a SAE. To fully train the network, we applied greedy layer wise training26 approach
where every hidden layer was trained separately.
• Supervised Fine-tuning After pre-training, the first three layers of a DNN were initialized with the parameters of
encoders from pre-trained SAE followed by a softmax output layer. At first we trained the output layer independently
while fixing the parameters of first 3 layers. Then we fine-tuned the whole network as Multilayer Perceptron (MLP) with
subject labels for criterion. The network outputs the probabilities of a subject belongs to each class and the class with
highest probability determines the output label of the subject. If we use xi , yi to represent the input feature vector and
label of the ith sample, respectively, the loss function based on cross entropy can be displayed as:
H(i) = −
1 N 2
∑ ∑ [1{yi = j}log(h(xi ) j ]
N i=1
j=1
(1)
where N is the number of input samples, j represents the class of samples, and h represents the network function.
• Optimization of Network Every training step of the networks were performed via back propagation with Adam
algorithm27 . It is a first-order gradient-based optimization algorithm which has been proven to be computational efficient
and appropriate for training deep neural network. During training stage, the training set was randomly split into mini
4/12
batches28 where each of them contains 50 samples in this study. At every iteration, only a single mini batch was used for
optimization. After every batch has been used once, the training set was reordered and randomly divided again so that
each batch would have different samples in different epoch.
• Dropout In order to prevent deep neural network from overfitting, regularization is necessary to reduce its generalization
error. In this study, we used dropout29 to learn more robust features and prevent overfitting. In the dropout layer, some
units were randomly dropped, providing a way to combine exponentially many different neural networks. In this study,
we inserted dropout layers after every hidden layer. In each iteration of training stage, only half of hidden units were
randomly selected to feed the results to the next layer, while in the testing stage all hidden units were kept to perform
the classification. By avoiding training all hidden units on every training sample, this regularization technique not only
prevented complex co-adaptations on training data and decrease overfitting, but also reduced the amount of computation
and improved training speed.
• Early Stopping Another approach we used to prevent overfitting is early stopping. Because deep architecture were
trained with iterative back propagation, the network were prone to be more adaptive to training data after every epoch. At
a certain point, improving the network’s fit to the training set will start to decrease generalization accuracy. In order
to terminate the optimization algorithm before over-fitting, early stopping was used to provide guidance for how many
iterations are needed. In the cross validation experiment, after dividing the data set into training and testing, we further
split the training samples into a training set and a validation set. The networks were trained only with data in the former
set, while samples in the latter set was used to determine when to stop the algorithm: while the network has the highest
generalization accuracy for validation set. In actual training, we stopped the optimization if the validation accuracy had
ceased to increase for 50 epochs.
• Ensemble Classifiers Although early stopping has proven to be useful in most deep learning problems, relatively small
data set limited the number of samples we could use for validation. And a small validation set may not able to represent
the whole data set resulting in a biased network. Therefore, we resorted to ensemble multiple classifiers to perform more
stable and robust classification. Instead of selecting a single validation set, we randomly divided the training set into 10
sets and used them to train 10 different networks to ’vote’ for the classification. At the training stage, for network i, set i
would be used for validation while the rest 9 sets were used for training. At the testing stage, the test samples were feed
into all these networks resulting in 10 sets of probabilities. For each sample, the probabilities from 10 networks was
added and the class with highest probability was the classification result of this sample. Although the performance of
ensemble classifiers may not be better than single network in every occasion, this strategy can statistically improve the
classification accuracy as well as the robustness and stability of the classifier.
Results and Discussion
Experiments Setup
The proposed deep neural network was built with Tensorflow30 , an open source deep learning toolbox provided by Google.
First, to compare the discriminant ability with state-of-the-art methods, 10-fold cross validation experiments were applied
to classify sMCI and pMCI subjects. Then we performed three experiments with different training sets to test whether the
images of pNC and pMCI contain AD pathology or not. For these experiments, 4 data sets: sNC, sAD, pNC and pMCI, were
divided into two groups in 3 different ways: 1) subjects of sNC and sAD were considered as group 1, and subjects of pNC and
pMCI belonged to group 2; 2) subjects of pMCI, sNC and sAD belonged to group1, and pNC were considered as group2; 3) all
subjects were considered as group1. For each experiment, we applied a 10-fold cross validation on group1. The subjects of
group 1 were randomly divided into 10 subsets, with 9 sets used for training and the rest set combined with subjects of group 2
used for testing. As detailed in the Methods Section, 10% of training subjects were randomly selected as validation set for early
stopping to prevent overfitting. 10 networks with different validation set were trained to ’vote’ for the final classification result.
Noticing it was not images but subjects we were splitting, so images from different time point of the same subject won’t be
used for both training and testing.
Compare with State-of-the-Art Methods
Researchers in the past have worked on classification of subjects with progressive cognitive decline and those with stable
cognitive impairment. pMCI were recognized as individuals with high risk of AD, while the sMCI were considered as those with
no risk or low risk of AD in these studies. To evaluate the performance of our approach, we compared the classification accuracy
of pMCI vs. sMCI with previous methods using the same data modality, i.e. T1-weighted MRI and FDG-PET13, 18, 31, 32 . The
proposed network outperformed the state-of-the-art method in classifying pMCI and sMCI individuals, irrespective of using
single or multimodal imaging, as shown in Table 2. It is worth to mention that in the study of Chen et.al32 , they performed
5/12
domain transfer learning to exploit the auxiliary domain data(AD/NC subjects) to improve the classification. Even though, the
acuccracy of our methods without auxiliary knowledge was 3.5% accuracy than theirs.
Method
Young et al.
Liu et al.
Suk et al.
Cheng et al.
Proposed
Young et al.
Liu et al.
Suk et al.
Cheng et al.
Proposed
Young et al.
Liu et al.
Suk et al.
Cheng et al.
Proposed
Modality
MRI
MRI
MRI
MRI
MRI
PET
PET
PET
PET
PET
MRI+PET+APOE
PET+MRI
PET+MRI
PET+MRI+CSF
PET+MRI
# Subjects
143
234
204
99
626
143
234
204
99
626
143
234
204
99
626
Accuracy
64.3
68.8
72.42
73.4
75.44
65.0
68.8
70.75
71.6
81.53
69.9
73.5
75.92
79.4
82.93
Sensitivity
53.2
64.29
36.7
74.3
77.27
66.0
57.14
25.45
76.4
78.20
78.7
76.19
48.04
84.5
79.69
Specificity
69.8
74.07
90.98
72.1
76.19
64.6
82.41
96.55
67.9
82.47
65.6
70.37
95.23
72.7
83.84
Table 2. Accuracy(%), Sensitivity(%), and Specificity(%) of the proposed network comparing with state-of-the-art methods.
The third row is the number of subjects used in the experiments
AD Pathology Classification
One problem of sMCI subjects was that we only know they remained stable at the time of preparation this manuscript, but they
could still progress to AD or other mental disease in the future. Although the the sMCI vs. pMCI experiment were commonly
used to test the discriminate ability of classifiers in recent studies, the classification result of sMCI subjects may not be very
accurate. Therefore, we performed the second experiment involved classifying individuals with only known Alzheimer’s
progression (pNC, pMCI and sAD) and normal controls (sNC). We investigated the performance of the classifier by using
various combinations of samples during training phase. At a very basic level, we trained the classifier by discriminating sAD
and sNC, at the next level pMCI and sAD were combined to represent the Alzheimer’s group and trained to discriminate them
from the sNC group. In the last level we combined pNC, pMCI and sAD to represent the Alzheimer’s group to discriminate
from the sNC group. The features extracted by the deep neural network are displayed in Fig 3. We observed the accuracy
and sensitivity of the classifier progressively improved by additionally training with pMCI and pNC, while the specificity
decreased, as displayed in Table 4. Further, the classifier performance was marginally better with the combination of FDG-PET
and structural measurements compared to the performances with individual modalities. Interestingly, the classifier performance
of structural imaging measurements were inferior to that of FDG-PET measurements. Supporting, the fact that FDG-PET, a
measure of neuronal activity is a better tool to identify prodromal Alzheimer’s as compared to structural images33, 34 .
Multiscale Classification
The classification accuracy of features extracted with different scales are listed in Table 3. We could not recognize any trend of
increasing or decreasing classification performance with the changes in patch size. Therefore, features with higher resolution do
not necessarily cover the discriminative information of lower resolution features. However, fusing multiscale features yielded
superior accuracy as compared to uniscale, suggesting the network has the ability to capture discriminative information across
the coarse to fine resolutions.
Early Diagnosis
We also investigated the classifier’s ability to identify individuals with high risk of acquiring Alzheimer’s, prior to disease onset.
The classifier trained with the combined sample of Alzheimer’s trajectory (pNC, pMCI and sAD) was superior, as compared
to the classifier trained with sAD alone. As the network classifier was trained with patterns of AD trajectory using pNC and
pMCI, the network was able to achieve exceptional classification performances in identifying individuals with AD risk, i.e
the classifier recognized individuals with AD risk with a of 90.08%, 85.61% and 81.20%, approximately at 1, 2, and 3 years
prior to disease onset respectively. Studies in the past have predicted AD onset, using unimodal or multimodal investigation.
Few studies have used PET as a single modality or in combination with structural MRI, CSF or cognitive measures to predict
6/12
Figure 3. Features fed to the output layer. From left to right the training set is sNC vs sAD, sNC vs sAD and pMCI, and sNC
vs sAD, pMCI and pNC, respectively. The y axis represents the units of the second from last layer, while x axis denotes the
different data samples. This figure shows the different patterns as distilled by the deep learning network from the sNC, pNC,
pMCI and AD images.
Training Set
sAD vs. sNC
sAD and pMCI
vs. sNC
sAD, pMCI
and pNC vs. sNC
Modality
FDG-PET
MRI
FDG-PET
MRI
FDG-PET
MRI
500
84.29
81.27
85.34
82.18
85.43
81.69
1000
83.76
81.58
84.80
82.69
85.28
82.04
2000
83.89
81.01
84.87
82.10
84.93
81.64
Multiscale
84.46
81.89
85.46
82.77
85.89
82.45
Table 3. Accuracy(%) using features at different scales of different modality.
AD onset13, 32, 35–38 . The accuracy of 3 year prediction in the present network analysis was superior than those reported in the
quoted studies. Studies predicting the illness onset, using structural MRI as a standalone tool or in addition to other clinical
variables, have reported accuracy values inferior than to those obtained using PET5, 39–49 .
Figure 4. Classification accuracy of different training sets. From left to right the training set is sNC vs sAD, sAD and pMCI,
sAD, pMCI and pNC, respectively. The y axis represents accuracy, while x axis denotes time(year) to conversion in which ’0’
means the subjects are already diagnosed as AD at the imaging visit. The number in legend is the classification accuracy of all
time points. Noticing the accuracy increases with more data used for training.
Deep neural network is a strong tool for accurate recognition of objects, by a-priori training of images with well characterized
objects. Hence the basic requirement for accurate classification using DNN tool are providing large number of images (usually in
millions) and well characterized objects during supervised training phase15 . Therefore a compromise in the a-priori knowledge
of the objects (features of Alzheimer’s) provided during training would limit the accuracy of the subsequent classification.
As our current understanding of AD pathogenesis and its precise characteristics in FDG-PET and structural MRI images are
limited, DNN suffered jeopardy in achieving 100% accurate classification. The clinical criteria for the diagnosis of AD involves
a series of evaluations to provide near precise diagnosis. Despite rigorous evaluations, clinically diagnosed individuals with
AD are not 100% accurate and hence the FDG-PET and structural MRI characters can overlap with conditions other than AD,
7/12
FDG-PET
Training Set
sAD vs. sNC
sAD and pMCI vs. sNC
sAD, pMCI and pNC vs. sNC
Acc
84.46
85.46
85.89
Sen
79.89
85.01
85.62
Volume
Spe
91.90
86.19
86.32
Acc
81.89
82.77
82.45
Sen
75.49
79.76
80.23
Multimodal
Spe
92.30
87.65
86.06
Acc
84.59
85.96
86.44
Sen
80.17
85.65
86.52
Spe
91.77
86.45
86.32
Table 4. Accuracy(%), sensitivity(%) and specificity(%) of different modality using different training sets.
including NC. Therefore the DNN trained with less accurately characterized images (FDG-PET and structural MRI) was unable
to achieve 100% accurate classification. We propose an improvement in characterization of AD features by either upgrading
FDG-PET and structural imaging methods or an increment in the understanding of AD specific pathogenesis, would positively
impact the classification accuracy of the DNN classifier tool in future studies.
Conclusion
In summary, we proposed a deep neural network to identify individuals at risk of developing Alzheimer’s disease. We trained the
classifier with patterns hidden in different resolutions and different modalities to distinguish subjects with Alzheimer’s trajectory
(pNC, pMCI and sAD) and those without cognitive deficits (sNC). Our results show the classifier’s ability to successfully
distinguish individuals with AD pathology from sNC with a remarkable accuracy of 82.93% using cross validation experiments.
We observed the performance of network classifier built by the combination of FDG-PET and structural MRI images was better
than those built with either structural MRI or FDG-PET alone. Further the classifier trained with the combined sample of
pNC, pMCI and sAD (Alzheimer’s trajectory) was found to yield the highest classification accuracy. Lastly, our experiment to
recognize individuals with AD pathology, prior to illness onset demonstrated a sensitivity of 85.68% in 3 years earlier to illness
onset. Hence, the proposed deep neural network classifier can be a potential tool of choice in the future for early prediction of
AD pathology. The number of pNC subjects was limited in this study resulting in a relatively low accuracy for pNC, as more
data is accumulated in the future, we expect better accuracy in the prediction of NC subjects with AD pathology.
References
1. Association, A. et al. 2011 alzheimer’s disease facts and figures. Alzheimer’s & dementia: journal Alzheimer’s Assoc. 7,
208 (2011).
2. Petersen, R. C. et al. Mild cognitive impairment: ten years later. Arch. neurology 66, 1447–1455 (2009).
3. Brookmeyer, R., Johnson, E., Ziegler-Graham, K. & Arrighi, H. M. Forecasting the global burden of alzheimer’s disease.
Alzheimer’s & dementia 3, 186–191 (2007).
4. Weiner, M. W. et al. Recent publications from the Alzheimer’s Disease Neuroimaging Initiative: Reviewing progress
toward improved AD clinical trials. Alzheimers Dement 13, e1–e85 (2017).
5. Davatzikos, C., Bhatt, P., Shaw, L. M., Batmanghelich, K. N. & Trojanowski, J. Q. Prediction of mci to ad conversion, via
mri, csf biomarkers, and pattern classification. Neurobiol. aging 32, 2322–e19 (2011).
6. Landau, S. M. et al. Associations between cognitive, functional, and fdg-pet measures of decline in ad and mci. Neurobiol.
aging 32, 1207–1218 (2011).
7. Mosconi, L. et al. Pre-clinical detection of alzheimer’s disease using fdg-pet, with or without amyloid imaging. J.
Alzheimer’s Dis. 20, 843–854 (2010).
8. Farhan, S., Fahiem, M. A. & Tauseef, H. An ensemble-of-classifiers based approach for early diagnosis of alzheimer’s
disease: Classification using structural features of brain images. Comput. mathematical methods medicine 2014 (2014).
9. Korolev, S., Safiullin, A., Belyaev, M. & Dodonova, Y. Residual and plain convolutional neural networks for 3d brain mri
classification. arXiv preprint arXiv:1701.06643 (2017).
10. Payan, A. & Montana, G. Predicting alzheimer’s disease: a neuroimaging study with 3d convolutional neural networks.
arXiv preprint arXiv:1502.02506 (2015).
11. Gray, K. R. et al. Multi-region analysis of longitudinal fdg-pet for the classification of alzheimer’s disease. NeuroImage
60, 221–229 (2012).
8/12
12. Toussaint, P.-J. et al. Resting state fdg-pet functional connectivity as an early biomarker of alzheimer’s disease using
conjoint univariate and independent component analyses. Neuroimage 63, 936–946 (2012).
13. Young, J. et al. Accurate multimodal probabilistic prediction of conversion to alzheimer’s disease in patients with mild
cognitive impairment. NeuroImage: Clin. 2, 735–745 (2013).
14. Zhang, D. et al. Multimodal classification of alzheimer’s disease and mild cognitive impairment. Neuroimage 55, 856–867
(2011).
15. Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. In Advances
in neural information processing systems, 1097–1105 (2012).
16. Liu, S. et al. Multimodal neuroimaging feature learning for multiclass diagnosis of alzheimer’s disease. IEEE Transactions
on Biomed. Eng. 62, 1132–1140 (2015).
17. Liu, S. et al. Early diagnosis of alzheimer’s disease with deep learning. In Biomedical Imaging (ISBI), 2014 IEEE 11th
International Symposium on, 1015–1018 (IEEE, 2014).
18. Suk, H.-I., Lee, S.-W., Shen, D., Initiative, A. D. N. et al. Hierarchical feature representation and multimodal fusion with
deep learning for ad/mci diagnosis. NeuroImage 101, 569–582 (2014).
19. Zhang, W., Zelinsky, G. & Samaras, D. Real-time accurate object detection using multiple resolutions. In Computer Vision,
2007. ICCV 2007. IEEE 11th International Conference on, 1–8 (IEEE, 2007).
20. Lowe, D. G. Distinctive image features from scale-invariant keypoints. Int. journal computer vision 60, 91–110 (2004).
21. Tang, Y. & Mohamed, A.-r. Multiresolution deep belief networks. In AISTATS, 1203–1211 (2012).
22. Dale AM, S. M., Fischl B. Cortical surface-based analysis. ii: Inflation, flattening, and a surface-based coordinate system.
Neuroimage 9(2), 195–207 (1999).
23. Raamana, P. R. et al. Thickness network features for prognostic applications in dementia. Neurobiol. aging 36, S91–S102
(2015).
24. Younes, M. B. M. M. A. T. L. Computing large deformation metric mappings via geodesic flows of diffeomorphisms. Int.
journal computer vision 61.2, 139–157 (2005).
25. Jenkinson, M., Bannister, P., Brady, M. & Smith, S. Improved optimization for the robust and accurate linear registration
and motion correction of brain images. Neuroimage 17, 825–841 (2002).
26. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H. et al. Greedy layer-wise training of deep networks. Adv. neural
information processing systems 19, 153 (2007).
27. Kingma, D. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).
28. Bengio, Y. Practical recommendations for gradient-based training of deep architectures. In Neural networks: Tricks of the
trade, 437–478 (Springer, 2012).
29. Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to prevent neural
networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014).
30. Abadi, M., Agarwal, A., Barham, P., Brevdo, E. & et al. TensorFlow: Large-scale machine learning on heterogeneous
systems (2015). URL http://tensorflow.org/. Software available from tensorflow.org.
31. Liu, K., Chen, K., Yao, L. & Guo, X. Prediction of mild cognitive impairment conversion using a combination of
independent component analysis and the cox model. Front. human neuroscience 11 (2017).
32. Cheng, B., Liu, M., Zhang, D., Munsell, B. C. & Shen, D. Domain transfer learning for mci conversion prediction. IEEE
Transactions on Biomed. Eng. 62, 1805–1817 (2015).
33. Kawachi, T. et al. Comparison of the diagnostic performance of FDG-PET and VBM-MRI in very mild Alzheimer’s
disease. Eur. J. Nucl. Medicine Mol. Imaging 33, 801–809 (2006).
34. Chételat, G. et al. Direct voxel-based comparison between grey matter hypometabolism and atrophy in Alzheimer’s disease.
Brain 131, 60–71 (2008).
35. Xu, L., Wu, X., Chen, K. & Yao, L. Multi-modality sparse representation-based classification for alzheimer’s disease and
mild cognitive impairment. Comput. methods programs biomedicine 122, 182–190 (2015).
36. Cabral, C. et al. Predicting conversion from mci to ad with fdg-pet brain images at different prodromal stages. Comput.
biology medicine 58, 101–109 (2015).
9/12
37. Zhang, D. & Shen, D. Predicting future clinical changes of MCI patients using longitudinal and multimodal biomarkers.
PLoS ONE 7, e33182 (2012).
38. Shaffer, J. L. et al. Predicting cognitive decline in subjects at risk for Alzheimer disease by using combined cerebrospinal
fluid, MR imaging, and PET biomarkers. Radiol. 266, 583–591 (2013).
39. Eskildsen, S. F., Coupe, P., Garcia-Lorenzo, D., Fonov, V. & et al. Prediction of Alzheimer’s disease in subjects with mild
cognitive impairment from the ADNI cohort using patterns of cortical thinning. Neuroimage 65, 511–521 (2013).
40. Moradi, E. et al. Machine learning framework for early mri-based alzheimer’s conversion prediction in mci subjects.
Neuroimage 104, 398–412 (2015).
41. Korolev, I. O., Symonds, L. L., Bozoki, A. C., Initiative, A. D. N. et al. Predicting progression from mild cognitive
impairment to alzheimer’s dementia using clinical, mri, and plasma biomarkers via probabilistic pattern classification. PloS
one 11, e0138866 (2016).
42. Misra, C., Fan, Y. & Davatzikos, C. Baseline and longitudinal patterns of brain atrophy in MCI patients, and their use in
prediction of short-term conversion to AD: results from ADNI. Neuroimage 44, 1415–1422 (2009).
43. Ye, J. et al. Sparse learning and stability selection for predicting MCI to AD conversion using baseline ADNI data. BMC
Neurol 12, 46 (2012).
44. Gaser, C., Franke, K., Kloppel, S., Koutsouleris, N. & Sauer, H. BrainAGE in Mild Cognitive Impaired Patients: Predicting
the Conversion to Alzheimer’s Disease. PLoS ONE 8, e67346 (2013).
45. Cuingnet, R. et al. Automatic classification of patients with Alzheimer’s disease from structural MRI: a comparison of ten
methods using the ADNI database. Neuroimage 56, 766–781 (2011).
46. Wolz, R., Julkunen, V., Koikkalainen, J., Niskanen, E. & et al. Multi-method analysis of MRI images in early diagnostics
of Alzheimer’s disease. PLoS ONE 6, e25446 (2011).
47. Chupin, M., Gerardin, E., Cuingnet, R., Boutet, C. & et al. Fully automatic hippocampus segmentation and classification
in Alzheimer’s disease and mild cognitive impairment applied on data from ADNI. Hippocampus 19, 579–587 (2009).
48. Cho, Y., Seong, J. K., Jeong, Y., Shin, S. Y. & et al. Individual subject classification for Alzheimer’s disease based on
incremental learning using a spatial frequency representation of cortical thickness data. Neuroimage 59, 2217–2230 (2012).
49. Coupe, P., Eskildsen, S. F., Manjon, J. V., Fonov, V. S. & Collins, D. L. Simultaneous segmentation and grading of
anatomical structures for patient’s classification: application to Alzheimer’s disease. Neuroimage 59, 3736–3747 (2012).
Consortia
for the Alzheimer’s Disease Neuroimaging Initiative
Michael Weiner2 , Paul Aisen3 , Ronald Petersen4 , Cliford Jack5 , William Jagust6 , John Trojanowki7 , Arthur Toga8 ,
Laurel Beckett9 , Robert Green10 , Andrew Saykin11 , John Morris12 , Leslie Shaw12 , Jefrey Kaye13 , Joseph Quinn13 ,
Lisa Silbert13 , Betty Lind13 , Raina Carter13 , Sara Dolen13 , Lon Schneider8 , Sonia Pawluczyk8 , Mauricio Beccera8 , Liberty Teodoro8 , Bryan Spann8 , James Brewer14 , Helen Vanderswag14 , Adam Fleisher14 , Judith Heidebrink15 , Joanne
Lord15 , Sara Mason5 , Colleen Albers5 , David Knopman5 , Kris Johnson5 , Rachelle Doody16 , Javier Villanueva-Meyer16 ,
Munir Chowdhury16 , Susan Rountree16 , Mimi Dang16 , Yaakov Stern17 , Lawrence Honig17 , Karen Bell17 , Beau Ances12 , John Morris12 , Maria Carroll12 , Mary Creech2 , Erin Franklin12 , Mark Mintun12 , Stacy Schneider12 , Angela
Oliver12 , Daniel Marson18 , Randall Grifth18 , David Clark18 , David Geldmacher18 , John Brockington18 , Erik Roberson18 , Marissa Natelson Love18 , Hillel Grossman19 , Efe Mitsis19 , Raj Shah20 , Leyla deToledo-Morrell20 , Ranjan Duara21 ,
Daniel Varon21 , Maria Greig21 , Peggy Roberts21 , Marilyn Albert22 , Chiadi Onyike22 , Daniel D’Agostino22 , Stephanie
Kielb22 , James Galvin23 , Brittany Cerbone23 , Christina Michel23 , Dana Pogorelec23 , Henry Rusinek23 , Mony de
Leon23 , Lidia Glodzik23 , Susan De Santi23 , P. Doraiswamy24 , Jefrey Petrella24 , Salvador Borges-Neto24 , Terence
Wong24 , Edward Coleman24 , Charles Smith25 , Greg Jicha25 , Peter Hardy25 , Partha Sinha25 , Elizabeth Oates25 , Gary
Conrad25 , Anton Porsteinsson26 , Bonnie Goldstein26 , Kim Martin26 , Kelly Makino26 , M. Ismail26 , Connie Brand26 ,
Ruth Mulnard27 , Gaby Thai27 , Catherine Mc-Adams-Ortiz27 , Kyle Womack28 , Dana Mathews28 , Mary Quiceno28 ,
Allan Levey29 , James Lah29 , Janet Cellar29 , Jefrey Burns30 , Russell Swerdlow30 , William Brooks30 , Liana Apostolova31 , Kathleen Tingus31 , Ellen Woo31 , Daniel Silverman31 , Po Lu31 , George Bartzokis31 , Neill Graf-Radford32 ,
Francine Parftt32 , Tracy Kendall32 , Heather Johnson32 , Martin Farlow11 , Ann Marie Hake11 , Brandy Matthews11 ,
Jared Brosch11 , Scott Herring11 , Cynthia Hunt11 , Christopher Dyck33 , Richard Carson33 , Martha MacAvoy33 , Pradeep
Varma33 , Howard Chertkow34 , Howard Bergman34 , Chris Hosein34 , Sandra Black35 , Bojana Stefanovic35 , Curtis
Caldwell35 , Ging-Yuek Robin Hsiung36 , Howard Feldman36 , Benita Mudge36 , Michele Assaly36 , Elizabeth Finger37 ,
10/12
Stephen Pasternack37 , Irina Rachisky37 , Dick Trost37 , Andrew Kertesz37 , Charles Bernick38 , Donna Munic38 , MarekMarsel Mesulam39 , Kristine Lipowski39 , Sandra Weintraub39 , Borna Bonakdarpour39 , Diana Kerwin39 , Chuang-Kuo
Wu39 , Nancy Johnson39 , Carl Sadowsky40 , Teresa Villena40 , Raymond Scott Turner41 , Kathleen Johnson41 , Brigid
Reynolds41 , Reisa Sperling42 , Keith Johnson42 , Gad Marshall42 , Jerome Yesavage43 , Joy Taylor43 , Barton Lane43 ,
Allyson Rosen43 , Jared Tinklenberg43 , Marwan Sabbagh44 , Christine Belden44 , Sandra Jacobson44 , Sherye Sirrel44 ,
Neil Kowall45 , Ronald Killiany45 , Andrew Budson45 , Alexander Norbash45 , Patricia Lynn Johnson45 , Thomas Obisesan46 , Saba Wolday46 , Joanne Allard46 , Alan Lerner47 , Paula Ogrocki47 , Curtis Tatsuoka47 , Parianne Fatica47 , Evan
Fletcher48 , Pauline Maillard48 , John Olichney48 , Charles DeCarli48 , Owen Carmichael48 , Smita Kittur49 , Michael Borrie50 , T-Y Lee50 , RobBartha50 , Sterling Johnson51 , Sanjay Asthana51 , Cynthia Carlsson51 , Steven Potkin52 , Adrian
Preda52 , Dana Nguyen52 , Pierre Tariot53 , Anna Burke53 , Nadira Trncic53 , Adam Fleisher53 , Stephanie Reeder53 ,
Vernice Bates54 , Horacio Capote54 , Michelle Rainka54 , Douglas Scharre55 , Maria Kataki55 , Anahita Adeli55 , Earl
Zimmerman56 , Dzintra Celmins56 , Alice Brown56 , Godfrey Pearlson57 , Karen Blank57 , Karen Anderson57 , Laura
Flashman58 , Marc Seltzer58 , Mary Hynes58 , Robert Santulli58 , Kaycee Sink59 , Leslie Gordineer59 , Jef Williamson59 ,
Pradeep Garg59 , Franklin Watkins59 , Brian Ott60 , Henry Querfurth60 , Geofrey Tremont60 , Stephen Salloway61 , Paul
Malloy61 , Stephen Correia61 , Howard Rosen62 , Bruce Miller62 , David Perry62 , Jacobo Mintzer63 , Kenneth Spicer63 ,
David Bachman63 , Nunzio Pomara64 , Raymundo Hernando65 , Antero Sarrael64 , Norman Relkin65 , Gloria Chaing65 ,
Michael Lin65 , Lisa Ravdin65 , Amanda Smith66 , Balebail Ashok Raj66 & Kristin Fargher66 .
2 Magnetic Resonance Unit at the VA Medical Center and Radiology, Medicine, Psychiatry and Neurology, University of
California, San Francisco, USA. 3 San Diego School of Medicine, University of California, California, USA. 4 Mayo Clinic,
Minnesota, USA. 5 Mayo Clinic, Rochester, USA. 6 University of California, Berkeley, USA. 7 University of Pennsylvania,
Pennsylvania, USA. 8 University of Southern California, California, USA. 9 University of California, Davis, California, USA.
10 MPH Brigham and Women’s Hospital/Harvard Medical School, Massachusetts, USA. 11 Indiana University, Indiana, USA.
12 Washington University St. Louis, Missouri, USA. 13 Oregon Health and Science University, Oregon, USA. 14 University of
California–San Diego, California, USA. 15 University of Michigan, Michigan, USA. 16 Baylor College of Medicine, Houston,
State of Texas, USA. 17 Columbia University Medical Center, South Carolina, USA. 18 University of Alabama – Birmingham,
Alabama, USA. 19 Mount Sinai School of Medicine, New York, USA. 20 Rush University Medical Center, Rush University,
Illinois, USA. 21 Wien Center, Florida, USA. 22 Johns Hopkins University, Maryland, USA. 23 New York University, NY, USA.
24 Duke University Medical Center, North Carolina, USA. 25 University of Kentucky, Kentucky, USA. 26 University of Rochester
Medical Center, NY, USA. 27 University of California, Irvine, California, USA. 28 University of Texas Southwestern Medical
School, Texas, USA. 29 Emory University, Georgia, USA. 30 University of Kansas, Medical Center, Kansas, USA. 31 University
of California, Los Angeles, California, USA. 32 Mayo Clinic, Jacksonville, Jacksonville, USA. 33 Yale University School of
Medicine, Connecticut, USA. 34 McGill University, Montreal-Jewish General Hospital, Montreal, Canada. 35 Sunnybrook
Health Sciences, Ontario, USA. 36 U.B.C. Clinic for AD & Related Disorders, Vancouver, BC, Canada. 37 Cognitive Neurology St. Joseph’s, Ontario, USA. 38 Cleveland Clinic Lou Ruvo Center for Brain Health, Ohio, USA. 39 Northwestern University,
San Francisco, USA. 40 Premiere Research Inst (Palm Beach Neurology), west Palm Beach, USA. 41 Georgetown University
Medical Center, Washington DC, USA. 42 Brigham and Women’s Hospital, Massachusetts, USA. 43 Stanford University,
California, USA. 44 Banner Sun Health Research Institute, Sun City, AZ 85351, USA. 45 Boston University, Massachusetts,
USA. 46 Howard University, Washington DC, USA. 47 Case Western Reserve University, Ohio, USA. 48 University of California,
Davis – Sacramento, California, USA. 49 Neurological Care of CNY, Liverpool, NY 13088, USA. 50 Parkwood Hospital,
Pennsylvania, USA. 51 University of Wisconsin, Wisconsin, USA. 52 University of California, Irvine – BIC, USA. 53 Banner
Alzheimer’s Institute, Phoenix, AZ 85006, USA. 54 Dent Neurologic Institute, NY, USA. 55 Ohio State University, Ohio,
USA. 56 Albany Medical College, NY, USA. 57 Hartford Hospital, Olin Neuropsychiatry Research Center, Connecticut, USA.
58 Dartmouth-Hitchcock Medical Center, New Hampshire, USA. 59 Wake Forest University Health Sciences, North Carolina,
USA. 60 Rhode Island Hospital, state of Rhode Island, Providence, RI 02903, USA. 61 Butler Hospital, Providence, Rhode
Island, USA. 62 University of California, San Francisco, USA. 63 Medical University South Carolina, Charleston, SC 29425,
USA. 64 Nathan Kline Institute, Orangeburg, New York, USA. 65 Cornell University, Ithaca, New York, USA. 66 USF Health
Byrd Alzheimer’s Institute, University of South Florida, Tampa, FL 33613, USA.
Acknowledgements
This work was supported by National Science Engineering Research Council (NSERC), Canadian Institutes of Health Research
(CIHR), Michael Smith Foundation for Health Research (MSFHR), Brain Canada, Genome BC and the Pacific Alzheimer
Research Foundation (PARF). Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging
Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number
W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging
and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s
11/12
Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.;
Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its
affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research &
Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co.,
Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation;
Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes
of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by
the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California
Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the
University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of
Southern California.
Author contributions statement
Donghuan Lu and Weiguang Ding built the deep neural network. Donghuan Lu and Karteek Popuri processed the neuroimage
data. Rakesh Balachander and Mirza Faisal Beg designed the experiments and interpreted the results. All authors reviewed the
manuscript.
Additional information
Competing Interest: The authors declare that they have no competing interest.
12/12
| 1cs.CV
|
A multiset model of multi-species evolution to
solve big deceptive problems
Luís Correia1, António Manso1,2
1- LabMAg - Laboratório de Modelação de Agentes, Faculdade de Ciências da Universidade
de Lisboa, Campo Grande, 1749-016 Lisboa – Portugal
2 - Instituto Politécnico de Tomar, Quinta do Contador - Estrada da Serra, 2300-313 Tomar –
Portugal
Abstract This chapter presents SMuGA, an integration of symbiogenesis with the
Multiset Genetic Algorithm (MuGA). The symbiogenetic approach used here is
based on the host-parasite model with the novelty of varying the length of
parasites along the evolutionary process. Additionally, it models collaborations
between multiple parasites and a single host. To improve efficiency, we
introduced proxy evaluation of parasites, which saves fitness function calls and
exponentially reduces the symbiotic collaborations produced. Another novel
feature consists of breaking the evolutionary cycle into two phases: a symbiotic
phase and a phase of independent evolution of both hosts and parasites. SMuGA
was tested in optimization of a variety of deceptive functions, with results one
order of magnitude better than state of the art symbiotic algorithms. This allowed
to optimize deceptive problems with large sizes, and showed a linear scaling in the
number of iterations to attain the optimum.
Keywords: Genetic Algorithm; Multisets; Symbiogenesis; Deceptive
optimization problems
Introduction
Computational models of coevolution can be used to study both natural settings
and artificial scenarios. Moreover, they can solve optimization problems.
Computational models are an effective tool configurable to model different types
of multi-species evolution: parasitism, commensalism, mutualism and cooperative
interactions. Competitive multi-species evolution has been useful in optimization
applications due to it providing better results when compared with a single
problem solver population. Coevolution of a solver population with a problem
creator population pushes both populations to increasingly better solutions, a
phenomenon called arms-race (Rosin and Belew, 1997).
On the other hand, symbiosis is a form of cooperative coevolution, which has
been gaining relevance in biology (Daida et al., 1996). In artificial systems,
symbiogenetic coevolution has been shown to improve evolutionary optimization
algorithms by a specialization of the different components of the symbiotic
2
collaboration (Wallin et al., 2005). In this case of cooperative coevolution there is
a kind of division of labor between the different types of symbionts. Each host is
combined with a set of parasites forming a collaboration. Each collaboration is
evaluated as a solution to the optimization problem. This is repeated for different
hosts and parasites. Artificial symbiogenetic evolution is proving useful in solving
deceptive problems (Wallin et al., 2005), a class of functions that is especially
difficult to optimize due to the fact that the optimum is surrounded by regions of
low quality solutions.
Artificial evolutionary models are inspired by nature, but when used as
engineering tools they do not need to maintain a strict correspondence with their
natural counterparts. The main goal of engineering is to obtain efficient tools, in
this case designed to solve optimization problems. Taking this into account, we
further explore different approaches of evolutionary algorithms (EA) and their
operators that one may consider unrealistic by comparison to nature. The multiset
representation of populations is one of those examples and in previous work we
have used that representation to support the evolutionary algorithms populations
(Manso and Correia, 2009). That algorithm is called Multiset Genetic Algorithm
(MuGA) and is successful in the optimization of various kinds of problems. The
populations are represented by multisets and the operators that are used explore
the representation in order to make the evolutionary process more efficient and
effective in a optimization of difficult problems.
In this work we present the Symbiogenetic MuGA Algorithm (SMuGA), which
uses natural inspiration of symbiogenesis to solve large deceptive problems that
are not solved by the common version of MuGA.
In the next section we present the base algorithm MuGA. In the following
section the symbiogenetic approach used is detailed. In particular we have two
different evolutionary processes, one for the hosts and another for the parasites,
and we describe each one separately and then aggregated. Next, we present results
obtained in several types of deceptive functions. The final section of this chapter
presents conclusion and proposes future work.
MuGA - A Multiset Genetic Algorithm
MuGA is a genetic algorithm that explores the features of a multiset to represent
populations of evolutionary algorithms and to improve their performance. The
traditional representation of populations used in evolutionary algorithms raises
two types of problems: the loss of genetic diversity during the evolutionary
process and evaluation of redundant individuals. These problems can be alleviated
when using multisets to represent populations.
Multiset population is not a representation that can be found in the natural
world, but it works well for optimization of difficult engineering problems.
3
Populations represented be multisets
A multiset (or multiple memberships set) is a collection of objects, called
elements, which are allowed to repeat. We can define the multiset as a set of
ordered pairs <copies, element> where copies are the cardinality associated to the
element. MuGA is a genetic algorithm in which populations represented by
multisets are called MultiPopulations (MP) and individuals represented by pairs
<copies, genotype> are called MultiIndividuals (MI).
Fig. 1 shows a simple population (SP) with eight individuals of OnesMax
problem and the equivalent MP with four MI. A multiset representation of
populations contains characteristics that make it a good alternative to the
collections that are usually used:
– MI has always different genotypes and the size of MP corresponds to the
genotype diversity at the genotypic level;
– The number of copies of MI may be used to control the selection pressure
in favor of the best fit individuals;
– The compact representation needs less computational effort to store the
population and avoids evaluation of identical individuals.
Individual
Fitness
MultiIndividual
Fitness
11111110
7
< 3, 11111110 >
7
11111110
7
< 2, 11110000 >
4
11111110
7
< 2, 10001000 >
3
11110000
4
< 1, 10000000 >
1
11110000
4
b)
10001001
3
10001001
3
10000000
1
a)
Fig. 1. a) Simple Population of 8 individuals; b) MultiPopulation of 4 MultiIndividuals.
The introduction of individuals in a MP is done either by incrementing the
number of copies of corresponding MI if the genotype exists in the population or
by introducing a new pair <1,genotype>. The elimination is done by decrementing
the number of copies of corresponding MI if the number of copies is greater than
one, or otherwise by removing the MI.
4
MuGA Algorithm
In evolutionary algorithms, populations are traditionally represented as a
collection of individuals. To minimize the issues such models raise, we developed
MuGA (Algorithm 1), whose most distinctive feature is that it represents
populations by multisets.
The algorithm starts by randomly generating and evaluating n individuals of the
problem to be optimized, while assuring that the base population, MP0, contains n
different genotypes. The design of the MuGA algorithm is prepared to preserve
the genetic diversity by maintaining the dimension of MP0 across generations.
The evolutionary process starts by selecting m individuals from MP0. These m
individuals are stored in MP1 and the number of MI is less than or equal to m. The
process continues with the recombination of MP1 and subsequent mutation of
MP2, generating MP3. MP4 is produced by the application of the replacement
operator on MP0 and MP3 to select n MI from the two populations. This operator
maintains the number of MI as a constant across generations. The evolutionary
process tends to produce many copies of good individuals. To reduce the number
of copies in MP4 the rescaling operator is applied and produces a new population
(MP0) for the iterative evolutionary process.
MuGA (n , m , problem)
MP0 = generate n MultiIndividuals from problem
Evaluate MP0
Repeat
MP1 = Select m Individuals from MP0
MP2 = Recombine the Individuals of MP1
MP3 = Mutate the Individuals of MP2
Evaluate MP3
MP4 = Select n MultiIndividuals from MP3 and MP0
MP0 = Rescale the number of copies of MP4
Until stop criteria
End Function.
Algorithm 1- MuGA - Multiset Genetic Algorithm
Multipopulations enable the execution of traditional genetic operators and
allow the design of new operators using the extra information, a set of unique
genotypes and associated number of copies, to extend operators that benefit from
such information. Next, we briefly describe the behavior of genetic operators
using MPs.
5
Multiset Selection
This operator chooses, from the base population, the parents that will be
reproduced to generate new individuals. We first expand the MP to an SP, Fig. 1,
so that MI with multiple copies has higher probability of being selected. We can
then use traditional selection operators (tournament selection, proportional
selection or ranking selection) or any improved selection operator (Sivaraj and
Ravichandran, 2011). When the operator allows the selection of the same
individual several times over, the mating population will contain MI and the
number of copies will reflect the degree of fitness of the genotype. The number of
copies of the fittest individuals tends to be larger than the remaining elements and
can be explored by the subsequent genetic operators.
Multiset Recombination
The recombination operator is responsible for the combination of chromosomes to
produce offspring that share genetic material of both parents. There is a great
variety of recombination operators in accordance with the representation of the
genes and chromosomes (e.g. binary strings, vectors of real numbers or trees) of
individuals and the type of problem to be solved, e.g. optimization of real
functions (Herrera et al., 2003), permutations (Otman and Jaafar, 2011) or
combinatorial (Spears and Anand, 1991). All these operators can be used in
MuGA through equivalence between MP and SP in terms of genotype
representation. Nevertheless we can design new operators using the number of
copies to make a genotype associated with the various parameters of the genetic
algorithm such as the probability of application, the number of cutting points, the
strength of individuals to spread their genes, etc. A wide range of possibilities is
available to explore the usefulness of this information and (Manso and Correia,
2011) presents a multiset recombination operator applied to the optimization of
real coded functions.
Multiset Mutation
The mutation operator in EA mimics what occurs in nature and randomly changes
a (usually small) part of the genome. The main function of this operator is the
introduction of new genes, enabling exploration of new areas in the search space
that are not attainable by the recombination of parental characteristics. Like the
recombination operator, mutation is also dependent on the type of problem and
representation of the individuals (Abdoun et al., 2012), (Droste et al., 2002). A
new operator that uses multiset information to optimize deceptive binary
6
functions, called Multiset Wave Mutation (MWM) is presented in (Manso and
Correia, 2013) and another one used to optimize real coded functions is presented
in (Manso and Correia, 2011).
Multiset Replacement
After recombination and mutation, the evolutionary algorithm has two populations
of individuals: the main population and the offspring generated by genetic
operators. The replacement operator selects which individuals will continue in the
evolutionary process. The generational strategy replaces the parents with their
children and the steady-state strategy replaces only a few parents with offspring
(Lozano et al., 2008). The operator must maintain the genetic diversity in the main
population so that the genetic operators can circumvent local optima and avoid
premature convergence (Yu and Suganthan, 2010), (Jayachandran and Corns,
2010). A new operator that uses multiset information to replace populations in a
steady state strategy, called Multiset Decimation Replacement (MDR), is
presented in (Manso and Correia, 2013).
Multiset Rescaling
The introduction of repeated elements in the MP tends to increase the number of
copies of the best fit MI if nothing is done to oppose it.
The rescaling operator was proposed to avoid that the best individuals get too
many copies (Manso and Correia, 2009). In order to control the number of
repeated elements, the rescaling operator divides the number of copies of each MI
by a factor, controlling in this way the pressure exhibited by the fittest individuals.
The operator ensures that each MI has at least one copy and that the total number
of individuals in the MP is not greater than a constant. An adaptive form of this
operator, called Adaptive Rescaling (AR) calculates in each iteration the value of
the reduction factor to maintain approximately the desired number of individuals.
SMuGA – A Symbiogenetic Multiset Genetic Algorithm
Symbiosis is set of natural theories that try to explain the natural relationship
between individuals that live together and how that relationship is vital to the
survival of the group. In nature symbiosis occurs and involves a relationship that
is constant and intimate between dissimilar species (Daida et al., 1996). That
relationship is more than the ecological interaction and includes mutualism, where
both individuals gain advantages from the alliance; commensalism, in which one
7
individual gains advantages and the other doesn’t have any inconvenience; and
parasitism, where one individual gains advantages and the other is harmed by the
relation.
Symbiosis theory provides an additional genetic operator to the artificial
evolutionary process and is successfully applied to solve a wide range of hard
problems. See (Heywood and Lichodzijewski, 2010) for a review of
symbiogenesis as a mechanism to build complex adaptive systems.
The Symbiogenetic MuGA algorithm (SMuGA) is inspired by the
Symbiogenetic Coevolutionary Algorithm (SCA), proposed by Wallin et al. in
2005, which explores a host-parasite relationship for optimization of concatenated
deceptive functions. Although the names “hosts” and “parasites” suggest a
parasitic relationship, the interaction between two species is benign and the gains
of parasites are not harmful to the hosts. SCA is successfully used to optimize
concatenated deceptive functions and MuGA by itself has proved to be an efficient
algorithm in the optimization of such functions with a moderate size (Manso and
Correia, 2013).
However, when the size of the problems increases MuGA experiences
difficulties in its optimization. In this paper, we apply the concept of symbiosis to
increase the efficiency of the MuGA. SMuGA is an algorithm that uses two
cooperative species, hosts and parasites, which evolve together in a mutualistic
relationship. The parasites are composed of a tuple <position, genome>, where the
position represents the parasite genome location where the parasite acts, and the
genome represents the genetic material of the parasite. In SMuGA the host
genome is replaced by the genome of the parasite in the location defined by the
position attribute (Fig. 2). The parasite considers the host genome as a circle,
which means that when the copy of the parasite genome to the host reaches the
limit of the host genome, the copy continues in the beginning. In Fig. 2, parasite
p1 is applied in host genome alleles 4, 5 and 6 and parasite p2 is applied in the
host genome alleles 9 and 0. The collaboration is the combination of host genes
and the genes introduced by parasites p1 and p2.
Index
Host
0
1
2
3
4
5
6
7
8
9
0
0
0
0
0
0
0
0
0
0
Parasite p1
4
1
1
1
9 1
p2
Collaboration 1 0
1
0
p2
1
Parasite p2
0
0
p1 p1 p1
1 1 1 0
Fig. 2. Collaboration formed by the symbiosis of a host and a parasite.
SCA has some deficiencies identified by the authors. The size of the parasites is
static and defined as a parameter, and collaboration is from one parasite to one
host, where each host can only be infected by a parasite at a time. The best results
8
obtained by the algorithm are when the parasite genome size is similar to the size
of the functions to be optimized, the building blocks (BB), and the performance
degrades quickly as the size of the parasites deviates from the size of the BB.
Another weakness of the SCA algorithm is that the collaboration is one to one,
which limits its applicability to separable problems.
The SMuGA was designed to suppress these two shortcomings by combining
the concept of symbiosis with the potential that the populations based on multisets
present on the optimization of this kind of functions. In the next section we
present the representation and evolution of parasite populations, the evolution of
host populations and the interaction between them with Symbiogenetic Multiset
Genetic Algoritm (SMuGA). In the design of the SMuGA some choices are made
with the objective of enhancing the success of the algorithm in the optimization of
problems and contouring the shortcomings that SCA presents.
Evolution of Parasites
In order to avoid having a human choice interfere significantly in performance, we
eliminate the need to specify the size of the parasites. As mentioned earlier the
work of Wallin et al (2005) showed that there was a very strong dependence of
performance relative to the size of the parasite. When the size of parasites
approaches BB size the performance is good, however it decays very quickly with
deviations from the ideal dimension.
In our approach the user does not have to know the size of BB because the
algorithm adapts the parasite’s length as necessary. To the best of our knowledge,
this is the first model of parasites that may vary their length along the evolutionary
process. This system is important for solving problems in which the size of BB is
not known or the BB has a variable size. The size of the parasites is changed by
genetic operators of recombination and mutation. The selection operator gives
opportunity to parasites that have a good performance in the host population to
reproduce and to pass on their genetic material and position to their descendants.
According to the theory of survival of the fittest, the parasites with a good
genome, which includes the position of application and the genetic material, will
spread their genes to subsequent generations, discovering and optimizing
simultaneously the position, the size and alleles of the parasites.
Parasite recombination
The following four situations can occur when two parasites recombine:
1. The parasites do not share positions in the genome of the host;
2. The parasites occupy consecutive positions in the genome of the host;
3. The parasites share some positions in the host;
4. All positions of one of the parasites occupy positions of the other.
9
In the first case, as the parasites infect different regions of the host genome,
recombination between the two parasites cannot take place. In all other cases the
idea underlying this operator is not only to recombine genetic material but also to
introduce different genome lengths. We selected the recombination of parasite
genomes as the principal operator to grow and shrink the length of the parasites.
a)
p1 4
1
1
1
1
0
2
3
4
5
6
7
1
1
1
1
1
b)
c)
o1 4
1
1
1
1
p2 8
0
0
0
0
8
9
10
11
12
13
0
0
0
0
0
0
0
0
14
Fig. 3. Recombination by the union of consecutive parasites: a) Selected parasites; b) Positions
occupied by parasites in the genome; c) Result of p1 and p2 recombination.
In the second case, Fig. 3, in which the parasites occupy consecutive locations
in the host genome, we determine that recombinant parasites are the union of
genomes generating a single parasite. The offspring o1, Fig. 3 c), has a genome
whose size is the sum of the size of the parental genomes. This type of
reproduction connects the parasites, and increases the length of the parasite
genome.
a)
p1 4
1
1
1
1
1
p2 6
0
0
0
0
0
0
0
2
3
4
5
6
7
8
9
10
11
12
13
14
p1 1
1
1
1
1
p2 0
0
0
0
0
0
Recombination Mask 1
0
1
12
13
14
1
0
1
1
b)
0
1
2
c)
d)
o1 4
1
3
4
5
6
7
8
9
10
11
o1 1
1
0
1
0
0
0
0
o2 1
0
1
1
0
0
0
o2 6
1
0
0
10
Fig. 4. Recombination by the share of some positions in the host: a) Selected parasites; b)
Positions occupied by parasites in the genome; c) Result of p1 and p2 recombination; d)
offspring parasites.
In case 3, Fig. 4, in which the parasites share some positions in the host
genome, alleles in the overlapping zone are combined using uniform crossover.
Furthermore, the offspring will have different genome sizes compared to their
parents. In Fig. 4 b) we illustrate uniform crossover. A recombination mask is
randomly obtained to perform an exchange of the parental alleles in the
overlapping zone. The symbol 1 in the mask means that there is an exchange of
alleles in the overlapping zone and the symbol 0 means the opposite. Fig. 4 c) and
d) show the recombination result of parents p1 and p2. The offspring o1 inherits
from both parents the parts that are not common between them, as well as the
recombined genome produced by the recombination mask. The offspring o2
inherits only the recombined common part with a dual mask. The offspring o1 is
longer than the parents and o2 is shorter.
a)
p1 4
1
1
1
1
1
1
1
p2 6
0
0
0
0
0
2
3
4
5
6
7
8
9
10
11
12
13
14
p1 1
1
1
1
1
1
1
p2 0
0
0
0
Recombination Mask 1
0
1
0
11
12
13
14
0
1
0
1
1
b)
0
1
2
c)
d)
o1 4
1
3
4
5
6
7
8
9
o1 1
1
0
1
0
1
o2 1
0
1
0
1
1
1
o2 6
1
1
0
0
10
Fig. 5. Recombination when one of parasites occupies all the positions of other: a) Selected
parasites; b) Positions occupied by parasites in the genome; c) Result of p1 and p2
recombination; d) offspring parasites.
In case 4, Fig. 5, where one of the parasites, p1, occupies all the positions of
the other, p2, in the genome of the host, the overlapping zone is also recombined
using uniform crossover. As in the previous case the genetic material is exchanged
in the overlapping zone through a recombination mask, Fig. 5 b) generated from a
uniform distribution. Fig. 5 d) shows the result of the recombination and the
complete offspring. Individual o1 inherits from the parent p1 the first part not
common to both parents, and the recombined common part, and the offspring o2
inherits the dual recombined common part, and the last not common part of p1. In
this case the small parasites act as cutting knives of larger parasites.
11
Parasite Mutation
The mutation operator randomly changes features of a parasite. These features
include the position, length and their genetic material. We use three types of
parasite mutation:
1. Change in anchoring position;
2. Change in the genome;
3. Parasite genome splitting whereby two new parasites are formed.
In the first situation, parasites change the position of host infection. In Fig. 6 a)
the parasite p1 that infects the fourth position generates the mutant m1 infecting
position 10 with the same genotype. Note that the parasite m1 affects the host
genome in a circular way where the last three bits of the parasite infect the first
three positions of the host.
a)
0
1
2
b)
c)
d)
0
1
2
0
0
0
p1 4
1
1
1
1
0
0
0
0
3
4
5
6
7
8
9
10
11
12
13
14
p1 1
1
1
1
0
0
0
0
5
6
7
8
9
10
11
12
13
14
m1 1
1
1
1
0
0
0
3
4
m1 10 1
1
1
1
0
0
Fig. 6. Mutation by changing position: a) Original parasite; b) Positions occupied by original; c)
Positions occupied by mutant parasite; d) Mutant parasite.
In the second case, the value of the alleles is changed by a probability
distribution that generates the mutation mask shown in Fig. 7 b. At the positions
where the mask has the value 1 the bit value of parasite is flipped. In this situation,
only the value of the parasite's genome is modified, which enhances the
appearance of parasites in the population with new genomes.
12
a)
p1 4
1
1
1
1
0
0
0
0
2
4
5
6
7
8
9
10
11
p1 1
Mutation Mask 0
1
0
1
1
1
0
0
0
0
0
0
1
0
0
0
4
5
6
7
8
9
10
11
m1 1
1
0
1
0
0
1
0
1
0
1
0
0
1
0
0
b)
1
1
2
b)
d)
3
3
m1 4
1
12
13
14
12
13
14
Fig. 7. Mutation by changing genome: a) Original parasite; b) Positions occupied by original and
mutation mask; c) Positions occupied by mutant parasite; d) Mutant parasite.
In the latter situation the parasite genome is split into two parts, originating into
two new parasites. The probability to split a genome is proportional to its length in
bits.
a)
0
1
p1 4
1
0
1
1
0
0
0
0
2
4
5
6
7
8
9
10
11
0 1 1
Break Point
0
0
0
1
8
9
10
11
m2 0
0
0
1
m2 8
0
0
b)
p1 1
0
1
2
c)
d)
3
m1 4
3
4
5
6
7
m1 1
0
1
1
1
1
1
1
12
13
14
12
13
14
0
1
Fig. 8. Mutation by breaking genome: a) Original parasite; b) Positions occupied by original
parasite and the break point; c) Positions occupied by mutant parasites; d) Mutant parasites.
Equation 1) shows the formula to calculate the probability of parasite splitting.
Parameter k controls the dimension from which the splitting of a parasite is
inevitable, i.e when the ratio is greater than one; Parameter n controls the shape of
distribution probability of splitting in other cases. The genotype split point is
selected by a uniform probability distribution over the genotype of the parasite.
13
(
)
(1)
This type of mutation avoids disproportionate growth of parasite length and
possible subsumption of the host genome. In Fig. 8 parasite p1 creates two
parasites, m1 and m2, where m2 position corresponds to the location splitting
point of the parasite genome p1.
Evaluation of parasites
The evaluation of the population of parasites is obtained indirectly through the
genomes present in the population of hosts. This feature allows the parasites to be
evaluated without the need to apply them to the hosts and then call the fitness
function to evaluate the collaboration. In this way, we replace fitness function
calls by a proxy consisting of simply checking if the parasite genome is present in
the host genome and using the host fitness rank. Therefore, we significantly save
function fitness calls as well as computational resources that would be spent on
testing and generating collaborations.
We defined three goals for the parasites:
1. Promoting the emergence of parasites with new genetic material, necessary
for the evolution of the combined population and prevention of its
stagnation;
2. Promoting the dissemination of parasites with good genotypes in the host
population so that all individuals have the parasite;
3. Promoting the variability of the anchoring point of good parasites in the
host genome in order to allow different regions to be infected.
The last two goals are incompatible with the first since it involves the
destruction of the original genetic material. Also, the evaluation function should
promote growth of the parasite length to speed up the evolutionary process to
discover large BB, and therefore we made the value of parasite fitness directly
proportional to its size.
In addition, the evaluation function of the parasites must be independent from
the scale of the fitness values in the hosts. To accomplish this, hosts are sorted
with a descending rank and parasites use those ranks to compute their evaluation.
The parasite evaluation algorithm sums the ranks of the hosts that have the
parasite in their genome. If the host rank is defined in the interval [1, n], where n
is the rank of fittest host and 1 the worst, parasites that infected the entire
population have maximum fitness value. Their contribution to diversification of
the population is zero, contrary to goal 1, nevertheless they are good candidates
for dissemination, goals 2 and 3. To circumvent this obstacle we shifted the rank
of the hosts to the interval [-n/2-1, n/2] where n is the size of the population. This
shift in ranking of the population provides a number of significant advantages.
First of all, the fitness of parasites that infect the entire population is zero;
14
parasites present only in the best individuals have positive fitness, and by
opposition, parasites that are present only in worst individuals have negative
fitness.
In order to reward individuals with a large genome, the value of the sum of
ranks is multiplied by the size of the parasite. Thus, if a parasite has a positive sum
of ranks its size is rewarded; otherwise its size contributes to the decrease of its
fitness. Such evaluation makes the discovery of a good parasite to be valuable at
the beginning, thereby promoting its spreading, and as it infects the population
through successive generations, its interest fades because the population has
already assimilated its genome. This parasite evaluation is very efficient because it
does not use a single call to the fitness function.
When evolution discovers a new parasite, whose genotype does not exist in the
population, the evaluation function should reward its discovery with a fitness that
allows it to survive and reproduce if it is a good parasite. On the other hand the
length of a new parasite should decrease its fitness to prevent the emergence of
large parasites with random genomes that contrast with large parasites evolved
from good BB. We decided to assign the new parasite a fitness value equal to the
population size divided by its length in bits, as a reward for the discovery of new
parasite genomes. The evaluation function allows small parasites with new
genotypes to appear in the population and to recombine themselves with existing
ones thereby promoting their growth if they contain useful genetic material for
evolution.
Algorithm of parasites evolution
The evolution of parasites is done by Algorithm 2. The algorithm receives as
parameters the population of parasites to evolve, pPop, the population of hosts to
perform the evaluation of the parasites, hPop, and the number of parasites that
will be selected to evolve, n.
ParasiteEvolution (pPop, hPop, n)
selectPop = select n parasites from pPop
offspringPop = recombine selectedPop
while offspringPop.size < pPop.size
Select random parasite from offspringPop
Mutate a clone of parasite
Insert mutated parasite clone in offspringPop
End while
Evaluate offspring in hPop
pPop = select pPop.size parasites from pPop and
offspringPop
End Function.
15
Algorithm 2- Parasite Evolution Algorithm
The algorithm starts by selecting n parasites from pPop. It continues with the
recombination of the selected population giving rise to the population
offspringPop. This step recombines genetic material of selected parasites and
changes the length of the offspring with the rules described above. The population
offspringPop is constructed by removing a pair of individuals from the selected
population, applying the recombination algorithm to the parents and inserting the
offspring in offspringPop population. The algorithm continues completing
offsringPop through successive mutations of clones of randomly selected
individuals in offsringPop. One of the three types of mutation described above,
genomic mutation, position mutation and genome splitting is randomly applied
with uniform probability. This way of completing a population allows a parasite to
undergo several mutations in a single generation, because a mutant parasite can be
selected and cloned several times.
The population offspringPop is evaluated through the genes of individuals of
the population hPop. The algorithm terminates with the calculation of a new
population through replacement operator applied to the original pPop and to the
population of its descendants, the offspringPop.
Evolution of hosts
A population of hosts is evolved with a MuGA algorithm, Algorithm 1, that uses
some genetic operators adapted to multipopulations (MP). The adaptation of
genetic operators to use the number of copies is critical to MuGA being able to
solve difficult problems. MuGA uses standard operators of selection and
recombination and an adapted form of mutation and replacement operators. In the
next section we describe the adaptions made in operators to take advantages of the
number of copies present in MI of MuGA populations.
MWM - Multiset Wave Mutation
To solve problems where the solution cannot be found by a recombination of
parent genes, the mutation operator performs a critical mission to introduce new
genes into the population. Mutation introduces random changes in the genome of
the individuals. Usually the operator introduces small changes in the genome of
the individual and the new features acquired are propagated in the population
through generations. A high rate of mutation is required if the changes to escape
from local maxima include many alleles but it is harmful if this assumption does
not happen. MI in multiset populations represents a set of clones of the same
genotype on which we apply different mutation rates.
16
[
]
(2)
Equation 2 presents a waveFunction formula to calculate the probability of
mutation from each clone of the MI that produces values between 0 and 1 (Fig. 9).
When the mutation value reaches the value 1, all the bits are changed and that
feature is very important to optimize deceptive functions where, usually, the
optimum is the complement of the local maxima.
Multiset Wave Mutation
1,0
0,9
0,8
Mutation Probability
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0,0
1
2
3
Copy number
4
5
6
7
8
9
10
Roughness =2 Thinness =3
11
12
13
14
15
16
Fig. 9. Graph of wave function with roughness=2 and thinness=3
MultisetWaveMutation (MI, minProb , mutOperator)
mutants = empty Multipopulation
For copy = 1 to MI.copies
probabilty = min ( minProb + waveFunction(copy), 1)
individual = MI.genotype
mutOperator( individual , probability )
mutants.add( individual )
next copy
return mutants
End Function.
Algorithm 3- Multiset Mutation Algorithm
Multiset Wave Mutation Algorithm 3, fully explained in (Manso and Correia,
2013) was designed to apply a traditional mutation operator, mutOperator, to a
multi-individual, MI, using the waveFunction to calculate the probability of
mutation of each clone. The probability is calculated adding a minimal
probability, minProb, to the result of waveFunction and truncating the result to 1
if the sum is greater than 1. Mutation in the offspring population is brought about
by applying Algorithm 3 to every MI present in the population.
MDR - Multiset Decimation Replacement
17
The replacement operator has the task of forming the population that will continue
the evolutionary process. This operator selects from parents and offspring MP
which individuals are selected to continue the evolutionary process.
MultisetDecimation (parentsPop,offspingPop , n)
parentsSize = parentsPop.size
parentsPop = parentsPop + offspringPop
while parentsPop.size > parentsSize
tournament = select n random MultiIndividuals
from parentsPop
selected = weakest MultiIndividual in tournament
remove selecteded from parentsPop
end while
End Function.
Algorithm 4- Multiset Decimation Algorithm
Multiset Decimation Replacement operator (MDR), Algorithm 4, was designed
to replace the parents population with an offspring population in a steady state
approach maintaining the multiset characteristics of MI present in both
populations. MDR joins the offspring population with the parents population and
the individuals with the same genotype increase their number of copies. The
algorithm then selects a group of random MI and removes the weakest. This
procedure is repeated until the parent population is reduced to the same number of
MI of the original population.
Co-evolution of hosts and parasites
The algorithm SMuGA is an evolutionary algorithm that uses two cooperating
populations to solve difficult problems: the host population that contains solutions
of the problem, and the parasite population that helps the first to reach the best
solution. Parasite populations evolve to achieve good genes that represent partial
solutions, and infect hosts through the incorporation of those genes.
The interaction between hosts and parasites produces a new population using
symbiosis that mimics what occurs in the natural world. We define collaboration
as the result of a host infected by one or more parasites using symbiosis.
Collaboration
A collaboration is obtained by copying the alleles of the parasite into the host. In
this case the alleles of the host are replaced by those of the parasite.
18
h
* * 1 1 0 0 * * *
h
* * 1 1 0 0 * * *
p1 0 0 1 1
c
p2 1 1 0 0
* * 0 0 1 1 * * *
c
* * 1 1 0 0 * * *
a)
b)
Fig. 10. Collaboration between one host and one parasite: a) successful collaboration b)
collaboration rejected.
A collaboration of a parasite with a host is only allowed if the host does not
have all the bits of the parasite, Fig. 10 a). This means that a parasite can infect a
host only once, Fig. 10 b). This detail allows the elimination of collaborations that
do not add anything new, and clears space for collaborations that do modify
something in the host.
h * * * * * * * * *
p1 0 1 1
h * * * * * * *
p1 0 1 1
p2 0 1 1
c * 0 1 1 * 0 1 1 *
a)
h *
p1 0 1 1
p3 1 1 1 1
c * 0 1 1 1 1 *
b)
* * * * * *
p4 0 1 1 1
c *
* * * * * *
c)
Fig. 11. Infection of a host by two parasites: a) non overlapping parasites, b) compatible
overlapping parasites c) incompatible overlapping parasites
We restrict the application of multiple parasites to cases where parasites do not
have incompatible bits. This means that the parasites may overlap, provided that
the overlapping segment does not contain different bits.
In Fig. 11 a) parasites p1 and p2 can infect host h because they infect disjoint
regions. Fig. 11 b) parasites p1 and p3 can infect the host h because, although they
share two genes, they have the same value and therefore the infection causes no
ambiguity. In Fig. 11 c) parasites p1 and p2 cannot be used simultaneously
because they overlap in two genes, one of which has distinct alleles. In this case,
the host can be infected by any of them but not by both simultaneously.
Algorithm 5 controls the formation of collaborations among a population of
hosts and a population of parasites. Algorithm 5 takes as parameters a host
multipopulation, sortedHostPop, sorted in descending order, a parasite population,
parasitePop and a parameter n that controls the probability of infection. The order
of the population is important because the index of the host in a population
determines the probability of the host receiving parasites. The algorithm continues
with the definition of the population resulting from the collaboration, symbPop,
among populations that are passed as a parameter. Afterwards, the hosts are
selected sequentially and the probability of infection is calculated. As hosts are
MI, the algorithm proceeds to expand into clones and applies parasites to each one
19
of them independently. Individuals with a higher ranking are those that usually
make more copies and thus may suffer various combinations of parasites.
After selecting a host and calculating a probability of infection, the algorithm
continues with the application of parasites to each of its clones. The parasites are
randomly arranged within the population of parasites to ensure no preference in its
application. In the next step the algorithm tries to apply each parasite to the host
selected using the compatibility rules of Fig. 11. In order to preserve the good
individuals of the population from a generalized infection, and hence the sudden
change of its genome, parasites are applied in a probabilistic manner. A host is
particularly vulnerable to parasites when its rank in the population is smaller. This
allows the fittest individuals to receive few parasites, thereby preserving their
genes, and lower-ranked individuals are subject to a generalized infection
accommodating several parasites. This process is similar to that described in
(Dumeur, 1996).
(
)
(3)
Equation 3) shows the formula to calculate the probability of a parasite
infecting a host, h, contained within a population, pop. The rank function returns
the rank of the individual within the population, in descending order of fitness and
pop.size represents the number of hosts that the population has. Parameter n
controls the shape of the ratio described above.
Collaboration (parasitesPop, sortedHostsPop, n)
symPop = empty MultiPopulation
for index = 1 to sortedHostsPop.size
host = sortedHostsPop.get(index)
pInfection = (index / hosts.size)^n
for copy = 1 to host.numberOfCopies
symbiosis = host.genotype
randomize parasites in parasitesPop
foreach parasite in parasitesPop
if compatible(parasite, symbiosis) and
uniformRandom(0,1) < pInfection
symbiosis = symbiosis + parasite
add symbiosis.clone to symbPop
end if
next parasite
next copy
next index
return symbPop
End Function.
Algorithm 5- Collaboration between Hosts and Parasites
20
The symbiosis population is built by the infection of selected parasites into the
host genomes. When a parasite is applied to the host, the genome of the parasite is
copied to the genome of the host generating a new individual through symbiosis.
A clone of that collaboration is added to the population of symbiosis, and the
symbiosis continues the process of being infected by other parasites.
SMuGA - Symbiogenetic Multiset Genetic Algorithm
SMuga, Algorithm 6, uses multipopulations to represent the populations of hosts
and parasites. This representation enables the use of multiset-adapted genetic
operators in both populations to help the evolutionary process. The use of
multipopulations is required to optimize deceptive problems, and every
challenging problem has a degree of deception (Whitley, 1991). This algorithm
has two phases: the collaboration phase, where the parasites infect the hosts; and
the evolution phase, where hosts and parasites evolve using coevolution.
SMuGA (h, p, problem, iterations, k, n)
hPop = generate h MultiIndividuals from problem
Evaluate hPop
pPop = generate p MultiParasites from problem
Evaluate pPop with hPop
Repeat
/* Collaboration phase */
selPop = Select k hosts from hPop
symbPop = collaboration( pPop, selPop, n)
hPop = Select h hosts from symbPop and hPop
/* Evolution phase */
Repeat iterations times
Evolve hPop
Evolve pPop
End repeat
Until stop criteria
End Function.
Algorithm 6- SMuGA – Symbiogenetic Multiset Genetic Algorithm
The algorithm has six parameters: h represents the size of the host population; p
the size of the parasite population; problem the problem to be solved; iterations
the number of iterations that hosts and parasites evolve without collaboration; k
the number of hosts selected to participate in the collaboration; and n that controls
the probability of hosts infection.
21
Hosts
Hosts
...
Evaluate
Evaluate
Evaluate
Evaluate
...
Parasites
...
Hosts
Collaboration
Hosts
Parasites
...
Collaboration
Evaluate
Parasites
Hosts
Fig. 12. Interaction between populations in SMuGA algorithm.
The algorithm starts by generating and evaluating a host population, hPop, with
h hosts of problem, and a parasite population, pPop, with p parasites. The only
information about the problem needed by parasites is the size of the host to
perform mutations. The evaluation of pPop is done using hPop. Fig. 12 show the
interaction between sPop and hPop.
The evolutionary iterative process starts with the collaboration phase followed
by the evolution phase until a stop criteria is reached.
Collaboration phase is performed by Algorithm 5 between populations of
parasites, pPop, and the k selected hosts in the host population, selPop, using the
parameter n to control the infection probability of hosts. The result of Algorithm 5
is a symbiosis population, symbPop that contains the selected host clones infected
by the parasites. Because one host may be infected by many parasites and the
algorithm saves clones when a host is infected by one parasite, the number of
symbiosis is huge when compared to the number of parasites and number of hosts.
This phase is computationally expensive. That effort is relieved by the use of
multipopulations since the collaboration algorithm produces symbiosis with
repeated genotypes and the multiset representation helps in its storage and
evaluation. The collaboration phase ends with the selection of h hosts from the
union of host population, hPop and symbiosis population, symbPop.
The evolution phase starts with the evolution of hPop using MuGA algorithm,
Algorithm 1, and the evolution of pPop using Algorithm 2. Both populations
evolve for iteration generations without establishing new collaborations. This
phase is used to stabilize the individuals in the populations and to assimilate, in
the hosts, genetic material introduced by the collaboration phase. The host
population evolves on its own, however, the parasite population still uses hosts,
since parasites are evaluated using the genes of the host population as a proxy for
fitness evaluation. When hosts evolve and change their genes, the fitness value of
parasites may change too.
Experimental study
To examine the influence of symbiosis in the solutions of hard problems we
conducted a set of experiments with the SMuGA algorithm and compared the
22
results with the standard MuGA. We compared, also, the results of SMuGA with
SCA in order to assess the scalability of the algorithm to big deceptive problems.
Experimental Setup
MuGA was configured with 128 MI in the main population. Selection is made by
tournaments with size 3. The operator selects 256 individuals for the mating pool
and in this way MI with copies are guaranteed for the following operators.
Recombination is made by one point crossover operator with probability 0.6.
Mutation is made by the multiset wave mutation, MWM, configured with
roughness 2 and thinness 3 (Fig. 9). The minimal mutation probability, parameter
minProb of Algorithm 3, is equal to 1/l , where l represents the size in bits of the
genome of the individual. Rescaling was applied to maintain a maximum total of
copies in the main population of twice the number of MI.
SMuGA is configured with 32 MI in the host population and 32 MI in the
parasite Populations. In this case, we can use a smaller population than with
MuGA, due to the increased genetic variety introduced by parasites. The size of
the population selected to make collaboration is 16 MI, and the parameter that
controls the probability of infection, parameter n in Algorithm 5, has value 1. The
number of iterations of the evolution phase in Algorithm 6 is set to 16. The
evolution of hosts uses tournament selection with tournament size 3 and selects 32
individuals. Recombination is done by uniform crossover with probability 0.6 and
mutation, replacement and rescaling are performed in the same way as in MuGA.
Table 1 shows evolutionary parameters of MuGA and SMuGA.
Table 1- Configuration of MuGA and SMuGA
MuGA
Parameter
SMuGA
Settings
Size of Population
Individuals
128
Selection
Recombination
Mutation
Replacement
Rescalling
Tournament size 3
Crossover 1 cut
MWM
Decimation
Adaptive
256
0.6
2,3
2
2
Parameter
Hosts
Parasites
Tournament size 3
Uniform Crossover
MWM
Decimation
Adaptive
Settings
32
32
32
0.6
2,3
2
2
To obtain statistical confidence we performed 128 independent runs for each
experiment. In each run, random initial populations were generated for
individuals, hosts and parasites. The stop criteria used in the simulations is the
number of evaluation function calls, and due to the varied difficulty of the
problems that limit is adjusted to allow the success of the evolutionary process.
For each experiment we compute the average of the number of evaluations to find
23
the optimum. We assign the maximum number of evaluations to the experiments
where the optimum is not found. We also compute what we consider a more
revealing result, which is the success rate, meaning the percentage of runs that
reach the optimum.
To compare the algorithms we use pair-wise Student T tests with 95%
confidence interval for the means. Due to the large number of simulations we
assume the normality of the variables. For each problem we also compare results
with other previously referred algorithms, when available, which means only for
smaller genome lengths. However, results published for these problems are not
always precise. In some cases only logarithmic graphs are printed and the results
here presented are best effort readings. And they never present the percentage of
runs that reach the optimum.
Experimental results with deceptive functions
The key to the success of Evolutionary Algorithms (EA) is their combination of
low-order building blocks (BB) to form higher-order BB, which eventually leads
to the optimum. When the solution cannot be built through this incremental
combination of BB, we are in the presence of deceptive problems and we need to
improve the artificial evolutionary process in order to solve those problems. The
concept of deception was first introduced by Goldberg (Goldberg, 1987) and much
work has been done in addressing this class of problems. MuGA and SCA are two
evolutionary algorithms that are able to optimize deceptive functions. In the next
sections we present experimental results on different deceptive benchmark
functions, for SMuGA, MuGA and SCA.
Fully deceptive F3 Function
Goldberg (Goldberg, 1989), devised a 3-bit function, F3, presented in Equation 4,
that is fully deceptive since building blocks of order n are deceptive to build
blocks of order n+1.
F3(000) = 28,
F3(001) = 26,
F3(010) = 22,
F3(011) = 0
F3(100) = 14,
F3(101) = 0,
F3(110) = 0,
F3(111) = 30
(4)
Fully deceptive function F3 is easily solved by EA due to is size of three bits.
To get a changeling problem we define the function F3 10 as ten consecutive
copies of F3. This procedure is usual in the optimization in this kind of deceptive
problems and is adequate to be solved using symbiogenesis present in SMuGA.
Optimization of F3 10 was successfully solved by the two algorithms
(SMuGA, MuGA), Table 2, and the symbiotic approach speeds up the
evolutionary process. Fig. 15 shows the evolution of the success rate of the
24
algorithms in the first 30,000 evaluation function calls, and Fig. 14 presents a
statistical view of the number of evaluation function calls needed to reach the
optimum in both algorithms. Results of SMuGA in function F3 10 are more than
on order of magnitude better than those presented by (Yang, 2004) and (Chen et
al., 2008).
Table 2- Statistics of SMuGA and MuGA result in F3 10 function.
SMuGA
F3 10
Mean
MuGA
Std
Mean
Std
Evals. to find Best
3309.79
1273.36
6074.30
2516.68
Best value found
300.00
0.00
300.00
0.00
Sucess rate (%)
100.00
0.00
100.00
0.00
Fig. 13 shows in more detail the evolution of the success rate, observing only
the first 6,000 evaluation function calls. In that figure we can clearly see, in the
major steps, the effect of the periodic incorporation of parasites in hosts, when
new collaborations are formed and integrated into the host population. The
evolution of the isolated host population over a few generations allows spreading
of good genetic material introduced by symbionts through the population. The
parasite population evolves in parallel, in this case taking into account the host
population to estimate the fitness of parasites. This process is very economical in
the number of collaborations generated, and subsequent calls to the fitness
function.
25
Fig. 13. Detail of the evolution of the success rate in the optimization of 10 copies of F3
Function with SMuGA solver. Blue line represents the collaboration event between hosts and
parasites.
Function F3 10 is solved by SMuGA due the use of symbiosis between hosts and
parasites. If one parasite that represents a BB of the function is found, it may be
copied to the position where another BB starts and the fitness of the collaboration
is sharply increased. The search for the BB and their positions is not easy because
no information about the function landscape is provided to SMuGA. Remarkably
SMuGA finds adequate length BB and their positions and uses symbiosis in a very
efficient way.
Fig. 14. Box-plots of the evaluation function calls to find the best of 10 copies of F3 Function
26
Fig. 15. Evolution of the success rate in the optimization of 10 copies of F3 Function.
Table 3- Statistics of SMuGA evolution result in optimization of concatenated F3 function with
different lengths.
SMuGA
F3
Evals. to find best
Mean
Std
Sucess %
Mean
Std
30 bits
3088.69
1363.68
100.00
0.00
60 bits
5054.30
1495.08
100.00
0.00
120 bits
8457.08
3021.80
100.00
0.00
240 bits
17500.25
12182.22
98.44
12.40
480 bits
24960.44
13143.26
95.31
21.14
In order to verify the scalability of the algorithm SMuGA to big genome
problems we performed a set of tests with the composition of 10, 20, 40, 80, and
160 fully deceptive F3 functions corresponding to problems with 30, 60, 120, 240
and 480 bits respectively. For these tests we only present results for SMuGA
since, in large problems, MuGA does not achieve solutions in reasonable time,
and other algorithms do not present results.
27
Fig. 16. SMuGA: Box-plots of the evaluation function calls to find the best value in 10(30),
20(60), 40(120), 80(240) and 160(480) copies(bits) of F3 function. Notice that vertical axis is
linear while horizontal axis is exponential.
Fig. 17. SMuGA: Evolution of the size of building blocks of 10, 20, 40, 80 and 160 copies of F3
function
Table 3 and Fig. 16 show the evolution statistics in the optimization of the
concatenated F3 function with different lengths using SMuGA after 75,000
function evaluations calls. The algorithm scales in a linear way in this kind of
functions due to its ability in finding good BB, assembling them with
recombination, Fig. 3, and thus forming larger BB which can be moved to other
locations in the genome, Fig. 6. This feature allows the solution of problems with
long genomes of concatenated functions in a very efficient way. Fig. 17 shows the
28
evolution of the size of building blocks in that experiment. As we can see,
problems with long genomes are solved by parasites also with long genomes,
which will eventually incorporate a collaboration, speeding up the evolution of
hosts. Again we note that the algorithm does not receive any information about
BB.
Fig. 18 shows the evolution of the success rate. The decrease of success in
optimization of F3 with 240 bits, 98%, and 480 bits, 95 %, can be explained by the
small size of the parasite population (32 individuals) for a very large genome of
the hosts. In that case, the probability of assembling useful BB in parasites
decreases due to the large space that they explore.
For the functions analyzed next, we notice similar behavior to the one depicted
in Fig. 13 in the step growth of success rate; and also a similar result to the one
depicted in Fig. 17, regarding the evolution of the length of parasites as a function
of the size of the problem. Therefore, we do not present such graphs for the
remaining functions.
Fig. 18. SMuGA: Evolution of the success rate in the optimization of 10, 20, 40, 80 and 160
copies of F3 function
Maximally separated fully deceptive F3 Function
The composition of functions in a sequential way is solved by SMuGA using the
mobility property of parasites present in the algorithm. The application of one
good parasite, which represents a BB, in a position where other BB starts,
contributes to the success of the algorithm due to the nature of the function
composition.
The problem becomes difficult when the bits of each function are separated.
The most difficult case of separation is when they are uniformly and maximally
29
distributed in the chromosome. We call these functions F3S N, where N
represents the number of F3 functions in the chromosome. In case of F3S 10 each
bit of one function is located in positions i, i+10, and i+20.
Table 4- Statistics of SMuGA and MuGA result in F3S 10 function
SMuGA
F3 Separated
Mean
MuGA
Std
Mean
Std
Evals. to Find Best
9419.20
5604.92
34063.77
18018.62
Best value found
300.00
0.00
299.95
0.30
Sucess rate (%)
100.00
0.00
99.22
8.80
Fig. 19. Evolution of the success rate in the optimization of F3S 10 function
These functions are difficult because the problem is not separable and the
formation of BB is not possible with a naïve strategy. In this way the bits of the
functions are spread and the application of one parasite in different positions is not
enough to solve the problem. SMuGA escapes this situation by combining several
parasites in a single host. With this experiment we verify SMuGA’s effectiveness
in non-separable problems as well.
30
Table 4 presents the results of the optimization of F3S 10. Both SMuGA and
MuGA solve the function with notable efficacy and, again, symbiogenesis speeds
up the evolutionary process. Fig. 19 shows the evolution of the success rate of the
algorithms in the first 100,000 evaluation function calls, and Fig. 14 presents a
statistical view of the number of function evaluations needed to reach the optimum
in both algorithms.
Fig. 20. Box-Plot of the evaluations to find the best in the optimization of F3S 10 function
Table 5- Statistics of SMuGA evolution result in optimization of separated F3S with different
lengths.
SMuGA
F3S
Evals. to find best
Mean
Std
Sucess %
Mean
Std
30 bits
7651.08
5049.29
100.00
0.00
60 bits
26429.98
19408.04
100.00
0.00
120 bits 184549.02 145359.55
95.31
21.14
240 bits 381741.92 148172.39
51.56
49.98
1.56
12.40
480 bits 496565.75
27516.57
Fig. 21 and Table 5 shows the statistics of the optimization of the composition
of 10, 20, 40, 80 and 160 F3S function in the chromosome after 500,000
evaluation function calls. As previously stated, the bits of F3S N functions are
maximally spread over the chromosome, and big genomes separate the bits of one
function with large distances. SMuGA fully succeeds in the optimization of 10
and 20 F3S N functions. In the optimization of 40 F3S, whose chromosome has
31
120 bits and the bits of each F3S function are separated by 40 bits, SMuGA
succeeds in 95% of simulations and needs more generations to fully succeed. In
the larger simulations, the small population of parasites and the large genome of
the hosts hinder the optimization, and the parameters must be adjusted.
Fig. 21. SMuGA: Box-plots of the evaluations to find the best value in 10, 20, 40, 80 and 160
copies of F3S function.
Fig. 22. SMuGA: Evolution of the success rate in the optimization of 10, 20, 40, 80 and 160
copies of F3 function.
Deceptive functions
32
Deceptive functions, also referred to as trap functions, were introduced by Ackley
(Ackley, 1987) and are defined in the unitation space. In this space, only the
number of ones in the chromosome counts, regardless of the order. Equation 5
presents the formula of a deceptive function where x is the chromosome, u(x) is
the number of ones in the chromosome x and l represents the length of
chromosome x. Fig. 23 presents a deceptive function with four bits in the unitation
space. This allows us to test the algorithm with a larger function and for which
there are other models with published results.
{
(5)
Deceptive 4
5
fitness value
4
3
2
1
0
0
1
2
3
4
unitation value
Fig. 23. Deceptive function with four bits in the unitation space.
Table 6- Statistics of SMuGA and MuGA results in Deceptive 4 function with 16 copies.
SMuGA
Deceptive 16 4
Mean
MuGA
Std
Mean
Std
Evals. to Find Best
5431,04
3724,07
84749,20
30440,50
Best value found
80,00
0,00
78,34
1,43
Sucess rate (%)
100,00
0,00
27,34
44,57
In this experiment we use a concatenated 16 blocks of four bits deceptive
function, Fig. 23, representing a chromosome with 64 bits. Table 6 shows the
results of MuGA and SMuGA in the optimization of the function after 100,000
evaluation function calls. SMuGA optimizes all the experiments with very little
evaluation function calls when compared to MuGA. Fig. 24 shows the evolution
of the success rate of both algorithms in evolution. MuGA experiences several
difficulties in optimizing deceptive functions with large genomes.
33
Comparing the results with SCA presented in (Wallin et al., 2005), where SCA
needs hundreds of thousands of function evaluations, we conclude that SMuGA is
significantly better. The ability of SMuGA to manipulate the size of the parasite
genomes is the key to solve this kind of problems. SCA do not have that property,
and the static size of the parasites slows down the evolution.
Fig. 24. Evolution of the success rate in the optimization of 16 copies of Deceptive 4 function
Table 7- Statistics of SMuGA evolution result in optimization of deceptive 4 function with
different lengths.
SMuGA
Deceptive 4
Evals. to find best
Mean
Std
Sucess %
Mean
Std
64 bits
4243.00
1645.40
100.00
0.00
128 bits
7061.89
3215.18
100.00
0.00
256 bits
11180.36
4897.55
100.00
0.00
512 bits
21458.61
13696.15
96.88
17.40
Table 7 and Fig. 25 show the statistics of evolution after 75,000 function
evaluation calls for the problems composed by 16, 32, 64 and 128 deceptive 4
functions that represent genomes with 64, 128, 256 and 512 bits. SMuGA was
successful in all the simulations. However in a simulation with problems
composed by 512 bits, SMuGA experiments some difficulties in the optimization
34
due to the large genome of the host and more generations are needed to optimize
all the problems as we can see in Fig. 26.
Fig. 25. SMuGA: Box-plots of the number of evaluation function calls for SMuGA to find the
best value in 16, 32, 64 and 128 copies of deceptive 4 function.
Fig. 26. SMuGA: Evolution of the success rate in the optimization of 16, 32, 64 and 128 copies
of deceptive 4 function.
SMuGA scales up very well to optimize large deceptive 4 problems and results
are again over one order of magnitude better than those presented in (Wallin et al.,
2005) using SCA and (Thierens, 2010) using Linkage Tree Genetic Algorithm
(LTGA). Table 8 shows the number of function evaluations to solve Deceptive 4
35
function with different lengths provided by our best effort to read the graphics
supplied in the papers.
Table 8- Number of functions evaluation calls to to solve Deceptive4 function using SCA and
LTGA algorithm (aprox.).
Algorithm
SCA
Size
64
128
Evals.
100000
200000
Algorithm
Size
60
100
LTGA
Evals.
40000
75000
Intertwined Deceptive functions
The Pair-Intertwined function proposed by Wallin and colleagues (Wallin et al.,
2005) is defined as two deceptive functions where the bits are intertwined in the
same function, Fig. 27. The Pair-Intertwined function was many local optima
introduced by the combination of the pair of deceptive functions. In this
experiment, we use as building block two deceptive functions of four bits each
composing a deceptive intertwine function, D4PI, with eight bits.
a)
d
d
d
d
b)
D
D
D
D
c)
d
D
d
D
d
D
d
D
Fig. 27. Intertwined pair deceptive functions: a) deceptive function d; b) deceptive function D; c)
Intertwined deceptive function dD.
Table 9- Statistics of SMuGA and MuGA result in deceptive 4 pair intertwined function with 8
copies
SMuGA
D4PI 8
Mean
MuGA
Std
Mean
Std
Evals. to Find Best
12401.04
12114.49
97294.48
12106.18
Best value found
80.00
0.00
76.88
1.61
Sucess rate (%)
100.00
0.00
6.25
24.21
Table 9 presents the statistics of the optimization of 8 D4PI functions,
amounting to 64 bits, after 100,000 evaluation function calls. SMuGA optimizes
all the problems with a small number of evaluation function calls due to the
36
capability, provided by the parasites, to discover the BB of the D4TI function and
the ability to concatenate BB and move them along the chromosome. The success
of MuGA in this experiment is very limited due to the large length of the BB and
the long genome of the individuals, Fig. 28. Comparing results with SCA
presented in (Wallin et al., 2005), Table 10, SMuGA is more than one order of
magnitude better.
Fig. 28. Evolution of the success rate in the optimization of 8 copies of D4PI function
Table 10- Number of functions evaluation calls to solve Deceptive Pair Intertwined function
using SCA algorithm (aprox.).
Algorithm
SCA
Size
64
128
Evals.
150000
250000
Table 11- Statistics of SMuGA evolution result in optimization of D4PI with different lengths.
SMuGA
D4PI
Evals. to Find best
Mean
Std
Sucess %
Mean
Std
64 bits
9246.20
9045.93
100.00
0.00
128 bits
20050.59
21544.04
100.00
0.00
256 bits
56673.89
84180.82
100.00
0.00
512 bits 204222.19 206614.44
79.69
40.23
37
Fig. 29 and Table 11 show the statistics of evolution after 500,000 function
evaluation calls for the problems composed by 8, 16, 32 and 64 D4PI functions
which represent chromosomes with 64, 128, 256 and 512 bit. SMuGA was
successful in all the simulations. However, in the 512 bit problems SMuGA
experiments some difficulties in the optimization due the large genome of the
host. More generations would allow to optimize these problems as we can infer
from Fig. 30, but adjusting the parameters for the 512 bit problem would
supposedly increase convergence.
Fig. 29. SMuGA: Box-plots of the evaluation function calls to find the best value in 8, 16, 32 and
64 copies of D4PI function.
Fig. 30. SMuGA: Evolution of the success rate in the optimization of 8, 16, 32 and 64 copies of
D4PI function
38
Deceptive Intertwined Pair 0-1 function
To assess the ability of SMuGA to evolve building blocks with optima that are not
all ones or all zeroes we defined a new intertwined function, DeceptivePI01,
where one function is evaluated by equation (5) and other by the equation (6). In
the deceptiveZ function, equation (6), z(x) counts the number of zeroes in the
string x. The optimum of function DeceptivePI01 is composed by a string with
alternating zeroes and ones and the translocations of the building blocks done by
the parasites need alignment in the host.
{
(6)
Fig. 25 show the evolution of success rate along the 1,000,000 function
evaluation calls for the problems composed by 16, 32, 64 and 128 DeceptivePI01
functions that represent genomes with 64, 128, 256 and 512 bits.
Using parameters of Table 1 MuGA again shows a poor performance. SMuGA
in most simulations optimizes the DeceptivePI01 function composed by eight bit
blocks, four of equation 5 and four of equation 6 interleaved. One reason for the
failures could be explained by the small number of parasites in the parasite
population.
The need for parasite alignment with the host requests a larger population of
parasites to avoid local maxima introduced by the bit pattern of the DeceptivePI0
functions. The two local maxima, all ones and all zeroes, are more attractive to the
parasites because that pattern does not need alignment and that parasites are easily
assimilated by the hosts.
39
Fig. 31. SMuGA: Evolution of the success rate in the optimization of 8, 16, 32 and 64 copies of
DeceptivePI01 function with 8 bits.
Fig. 32 shows the effect of the size of parasite population in the optimization of
8 copies of DeceptivePI01 with 8 bits. As can be seen, the increase of the number
of parasites in the symbiotic system increases the robustness of the solver. The
increase of parasite population increases the computational complexity of the
algorithm, but parasite population can evolve in parallel to the host population
exploring the multicore resources of the computers.
Fig. 32. SMuGA: Evolution of the success rate in the optimization of 8 copies of DeceptivePI01
function with 8 bits with solver with 16, 32, 64 and 128 parasites in the parasite population.
Table 12- Statistics of SMuGA evolution result in optimization of D4PI01 with different lengths.
SMuGA
D4PI01
64 bits
Evals. to Find best
Mean
Mean
Std
29939.75
100.00
0.00
54285.36 111617.89
100.00
0.00
256 bits 107131.61 187231.57
100.00
0.00
512 bits 350228.05 421266.94
73.44
44.17
128 bits
61204.72
Std
Sucess %
Fig. 33 and Table 12 present the same situation of Fig. 31 but now with 128
elements in the parasite population, instead of 32. The success of the algorithm is
40
increased and simulations evolving functions with 64, 128 and 256 bit are always
successfully optimized. The rate of success of simulation with 512 bits also
increases although not attaining 100% success. Further parameter tuning is one
possible solution to achieve perfect score.
These results show that a large size of the parasite population makes SMuGA
more robust in the evolution of difficult functions. Complex bit patterns impose
difficulties to SMuGA in the alignment of parasites but these seem to be
circumvented by larger parasite populations.
Fig. 33. SMuGA: Evolution of the success rate in the optimization of 8, 16, 32
and 64 copies of DeceptivePI01 function with 8 bits intertwined with 128 parasites
solvers.
Conclusions
This chapter presented the Symbiogenetic MuGA (SMuGA), an extension of the
Multiset Genetic Algorithm (MuGA) with a novel approach to artificial
symbiogenesis where a host receives genetic material from multiple parasites of
variable length. This is the first evolutionary model where parasites do not have a
fixed length. Rather their length varies along the evolutionary process.
The model proposed also introduced a two-phased step of evolution. In one
phase, symbiotic collaborations are generated and compete with previous hosts to
form the next generation host population. In the other phase, host and parasite
populations evolve on their own for a few generations, but parasites use hosts’
fitness as proxys to compute their own. Proxy parasite evaluation significantly
saves fitness function calls and avoids the need to generate an exponential number
41
of collaborations. The phase of separate evolution of both hosts and parasites
allows to simultaneously stabilize host population and to foster exploration by the
parasite population.
Results obtained have largely surpassed previous symbiogenetic models,
allowing us to solve very large deceptive problems. It should be noted in spite of
MuGA obtaining good results, it is only SMuGA that achieves solutions to very
large problems, by integrating symbiogenesis in MuGA, with two-phase evolution
and proxy evaluation of parasites.
In fact, SMuGA turned out to be so efficient as to show a linear scaling with the
length of the deceptive problems used for testing. The variation of the parasites’
length allows evolution to find adequate length building blocks (BB) for the
problem at hand. Accumulating multiple parasites in a single host provides the
opportunity of using parasite combinations, which prove to be important for more
complex problems.
In future work, we want to test more operators in the parasites. In particular,
inversion might be important to hierarchical deceptive problems. We also need to
explore different types of problems with SMuGA. Those used in this paper are
repeated concatenations of the same function. Also, the flexibility of this model
indicates that it is adequate for dynamic fitness functions, and we should test it on
dynamic problems. The symbiotic system can also be taken as a new operator
introducing new parameters in the evolutionary process. Consequently the new
parameters can be tuned to increase the effectiveness of SMuGA and in the future
we will make an effort in optimization and automation of these parameters.
Acknowledgments
The authors thank Mel Todd, Guida Manso and Nathalie Gontier for the
precious revisions that made this text more clear and readable.
References
Abdoun, O., Abouchabaka, J., Tajani, C., 2012. Analyzing the Performance of
Mutation Operators to Solve the Travelling Salesman Problem.
arXiv:1203.3099.
Ackley, D.H., 1987. A Connectionist Machine for Genetic Hillclimbing. Kluwer
Academic Publishers, Norwell, MA, USA.
Baghshah, M.S., Shouraki, S.B., Halavati, R., Lucas, C., 2007. Evolving fuzzy
classifiers using a symbiotic approach, in: IEEE Congress on
Evolutionary Computation, 2007. CEC 2007. Presented at the IEEE
Congress on Evolutionary Computation, 2007. CEC 2007, pp. 1601–
1607.
42
Chen, Y., Hu, J., Hirasawa, K., Yu, S., 2008. Solving deceptive problems using a
genetic algorithm with reserve selection, in: IEEE Congress on
Evolutionary Computation, 2008. CEC 2008. (IEEE World Congress on
Computational Intelligence). Presented at the IEEE Congress on
Evolutionary Computation, 2008. CEC 2008. (IEEE World Congress on
Computational Intelligence), pp. 884 –889.
Daida, J.M., Grasso, C.S., Stanhope, S.A., Ross, S.J., 1996. Symbionticism and
Complex Adaptive Systems I: Implications of Having Symbiosis Occur
in Nature, in: Proceedings of the Fifth Annual Conference on
Evolutionary Programming. The MIT Press, pp. 177–186.
Droste, S., Jansen, T., Wegener, I., 2002. On the analysis of the (1+ 1)
evolutionary algorithm. Theor. Comput. Sci. 276, 51–81.
Dumeur, R., 1996. Evolution through cooperation: The Symbiotic Algorithm, in:
Alliot, J.-M., Lutton, E., Ronald, E., Schoenauer, M., Snyers, D. (Eds.),
Artificial Evolution, Lecture Notes in Computer Science. Springer Berlin
Heidelberg, pp. 145–158.
Goldberg, D.E., 1987. Simple Genetic Algorithms and the Minimal, Deceptive
Problem. Genet. Algorithms Simulated Annealing L. Davis, editor, San
Mateo, CA: Morgan Kaufmann, 74–88.
Goldberg, D.E., 1989. Genetic Algorithms and Walsh Functions: Part I, A Gentle
Introduction. Complex Syst. 3, 129–152.
Herrera, F., Lozano, M., Sanchez, A.M., 2003. A taxonomy for the crossover
operator for real-coded genetic algorithms: An experimental study. Int. J.
Intell. Syst. 18, 309–338.
Heywood, M.I., Lichodzijewski, P., 2010. Symbiogenesis as a Mechanism for
Building Complex Adaptive Systems: A Review, in: Chio, C.D.,
Cagnoni, S., Cotta, C., Ebner, M., Ekárt, A., Esparcia-Alcazar, A.I., Goh,
C.-K., Merelo, J.J., Neri, F., Preuß, M., Togelius, J., Yannakakis, G.N.
(Eds.), Applications of Evolutionary Computation, Lecture Notes in
Computer Science. Springer Berlin Heidelberg, pp. 51–60.
Jayachandran, J., Corns, S., 2010. A comparative study of diversity in
evolutionary algorithms. IEEE, pp. 1–7.
Lozano, M., Herrera, F., Cano, J., 2008. Replacement strategies to preserve useful
diversity in steady-state genetic algorithms. Inf. Sci. 178, 4421–4433.
Manso, A., Correia, L., 2009. Genetic algorithms using populations based on
multisets, in: New Trends in Artificial Intelligence, EPIA 2009. Luís
Seabra Lopes and Nuno Lau and Pedro Mariano and Luís Rocha (Eds.),
Universidade de Aveiro, pp. 53–64.
Manso, A., Correia, L., 2011. A multiset genetic algorithm for real coded
problems, in: Proceedings of the 13th Annual Conference Companion on
Genetic and Evolutionary Computation - GECCO ’11. Presented at the
the 13th annual conference companion, Dublin, Ireland, p. 153.
Manso, A., Correia, L., 2013. A multiset genetic algorithm for the optimization of
deceptive problems, in: Proceeding of the Fifteenth Annual Conference
43
on Genetic and Evolutionary Computation Conference, GECCO ’13.
ACM, New York, NY, USA, pp. 813–820.
Otman, A., Jaafar, A., 2011. A Comparative Study of Adaptive Crossover
Operators for Genetic Algorithms to Resolve the Traveling Salesman
Problem. IJCA (0975-8887).
Paredis, J., 1995. The Symbiotic Evolution of Solutions and Their
Representations, in: Proceedings of the 6th International Conference on
Genetic Algorithms. Morgan Kaufmann Publishers Inc., San Francisco,
CA, USA, pp. 359–365.
Rosin, C.D., Belew, R.K., 1997. New Methods for Competitive Coevolution. Evol
Comput 5, 1–29.
Sivaraj, R., Ravichandran, T., 2011. A review of selection methods in genetic
algorithm. Int. J. Eng. Sci. Technol. 3, 3792–3797.
Spears, W., Anand, V., 1991. A Study Of Crossover Operators In Genetic
Programming.
Thierens, D., 2010. Linkage Tree Genetic Algorithm: First Results, in:
Proceedings of the 12th Annual Conference Companion on Genetic and
Evolutionary Computation, GECCO ’10. ACM, New York, NY, USA,
pp. 1953–1958.
Wallin, D., Ryan, C., Azad, R.M.A., 2005. Symbiogenetic coevolution, in: The
2005 IEEE Congress on Evolutionary Computation, 2005. Presented at
the The 2005 IEEE Congress on Evolutionary Computation, 2005, pp.
1613–1620 Vol. 2.
Whitley, L.D., 1991. Fundamental Principles of Deception in Genetic Search.
Found. Genet. Algorithms 1, 221–241.
Yang, S., 2004. Adaptive group mutation for tackling deception in genetic search.
WSEAS Trans. Syst. 3, 107–112.
Yu, E.L., Suganthan, P.N., 2010. Ensemble of niching algorithms. Inf. Sci. 180,
2815–2833.
| 9cs.NE
|
Does a Daily Deal Promotion Signal a Distressed
Business? An Empirical Investigation of Small Business
Survival
Ayman Farahat
Nesreen K. Ahmed
Adobe Inc.
Purdue University
Rice University
[email protected]
[email protected]
[email protected]
arXiv:1211.1694v1 [cs.CE] 7 Nov 2012
ABSTRACT
In the last four years, daily deals have emerged from nowhere
to become a multi-billion dollar industry world-wide. Daily
deal sites such as Groupon and Livingsocial offer products
and services at deep discounts to consumers via email and
social networks. As the industry matures, there are many
questions regarding the impact of daily deals on the marketplace. Important questions in this regard concern the
reasons why businesses decide to offer daily deals and their
longer-term impact on businesses. In the present paper, we
investigate whether the unobserved factors that make marketers run daily deals are correlated with the unobserved
factors that influence the business, In particular, we employ
the framework of seemingly unrelated regression to model
the correlation between the errors in predicting whether a
business uses a daily deal and the errors in predicting the
business’ survival. Our analysis consists of the survival of
985 small businesses that offered daily deals between January and July 2011 in the city of Chicago. Our results
indicate that there is a statistically significant correlation
between the unobserved factors that influence the business’
decision to offer a daily deal and the unobserved factors that
impact its survival. Furthermore, our results indicate that
the correlation coefficient is significant in certain business
categories (e.g. restaurants).
Categories and Subject Descriptors
D.2.8 [Computational Advertising]: Economics of Daily
Deals
Keywords
Daily deals, Consumer ratings, Seemingly unrelated regression
1.
INTRODUCTION
Daily deal sites such as Groupon represent a novel approach to Internet marketing that tap into local markets,
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
Utpal Dholakia
and based on the massive scale and rapid growth of such
sites, the business model has gained rapid adoption from a
wide range of businesses.
Despite this success, a vocal contrarian view of the daily
deals model has emerged. Its chief criticism is probably
skepticism about the value of daily deals to the merchants
whose goods and services are promoted. To be sure, enthusiastic advocates are easy to find; Groupon has claimed that
97% [11] of businesses using its service want to be featured
again. But an independent study estimates repeat intent at
only 48.1%. Some anecdotal reports are surprisingly harsh,
including a highly publicized blog posting by the owner of
a New York bakery cafe, who described her Groupon promotion as ”the single worst decision I have ever made as a
business owner”.
The diverging views about the profitability and long term
impact of the multi-billion daily deals industry calls for a
thorough study and more details than previous attempts.
Any study pertaining to evaluate the impact of daily deals
on key business metrics needs to isolate the causal impact
of daily deals from other factors that might be correlated
with daily deal adoption and also impact business metrics.
In this paper, we examine the impact of daily deals on
the business from the lens of modern econometrics. It is
well known that the gold standard for evaluating a treatment effect such as daily deals is through randomized experiments. In practice however, it is not possible to run the
experiments needed to isolate the impact of daily deals from
other confounding factors. Therefore, most studies including
ours will have to contend with observational data. Working
with observational data poses its own set of challenges and
it is exactly these challenges that call for tools from modern
econometrics.
One of the main challenges of working with observational
data is unobserved heterogeneity. While two businesses might
look identical along the observed dimensions such as category, location..etc, they may differ along dimensions that
can have a significant impact on the business. For example, a struggling business with a high staff turnover will be
more likely to fail than a similar business. Also, struggling
businesses might be more tempted to use a daily deal to
help shore sales. A naive application of standard statistical techniques will show daily deal adoption to be a factor
in business failure while in reality it is staff turnover that
contributed to the failure. To alleviate the problems of unobserved heterogeneity, we treat the daily deal adoption as
a dependent variable. A business’s decision to adopt a daily
deal will depend on a number of observed independent vari-
ables such as the business category, how popular are deals in
category, etc. The daily deal decision will also be impacted
by unobserved factors. Similarly, business failure will depend on observed dependent variables and unobserved factors. For example, consider a business that runs a daily deal
despite of having no apparent reason to do so. Does that
signal anything about the unobserved reasons that might
lead the business to fail down the line? If the unobserved
factors that make a business run a daily deal are correlated
with the unobserved factors that lead to business failure then
indeed the daily deal adoption conveys additional information about the business and should be taken into account.
On the other hand, if there is no correlation between the
unobserved factors in daily deal and failure, then knowing
whether a business offered a daily deal does not convey additional information.
We resort to techniques from modern econometrics to help
test whether the unobserved factors in daily deal adoption
and failure are correlated. In particular, we use the framework of multiple equation models. One equation in our setting models the business’s decision to offer a daily deal and
the second equation models the business failure. In order
to model both equations, we need to identify the dependent
variables (daily deal adoption and failure) and a set of covariates that best describe each model.
In this work, we tried to insure that our results were statistically significant, robust to modeling assumptions and scalable. To help achieve statistical significance, we compiled a
large data set that had information on daily deal adoption,
business failure and information about the businesses. To
test robustness of our models, we used two frameworks that
make different modeling assumptions. To insure scalability,
we developed a data driven approach that does not require
expensive interviews with customers or lab experiments. In
particular, we crawled Yelp to get information about the
business and whether a business was closed. For daily deal
adoption, we used a data set from [14].
In this paper we develop a model for business failure in
Chicago where both the failure data and business information are derived from Yelp. We also develop a separate
model for daily deal adoption where we join the data set
from [14] with Yelp data. We decided to use Chicago because
in addition to being the third largest city in the U.S, it is
the home town of the largest daily deal provider “Groupon”
and had a large number of business adopting daily deals. We
then developed joint models of daily deal adoption and business failure. Our results show that a joint model of business
failure and daily deal adoption does a better job in explaining the data than the two separate models. We tested the
robustness of our results by specifying two different modeling paradigms with different model assumption ; bivariate
probit and seemingly unrelated regression.
Our results show that models of business failure and daily
deals adoption using Yelp based features provide good performance. Furthermore, we find a positive and statistically
significant correlation between the unobserved factors that
make a business offer a daily deal and the unobserved factors that contribute to failure. Our results indicate that the
correlation is strongest in case of Restaurants 0.281 and
smaller in case of Spas 0.24. These results are consistent
with results of Gupta [11] that found that deals are generally good for Spas and bad for Restaurants.
In summary, the main contributions of this paper are the
following:
• We conducted a data driven large scale study to test
and quantify whether the unobserved factors that make
a business decide to make daily deals are correlated to
the unobserved factors that impact the business survival.
• We developed and analyzed business survival and daily
deal adoptions models based on business features collected from Yelp.
• We conducted statistical tests and two different econometrics frameworks to insure statistical significance of
results.
• Our results are consistent with findings from previous
research work that used labor intensive surveys.
This paper is organized as follows. In Section 2, we survey
related work. In Section 3, we describe our data and the
methods employed to collect it. In Section 4, we describe
the necessary background from econometrics that is needed
to develop our models. In Section 5, we describe our experiments, results and evaluations. Finally, in Section 6 we
summarize our findings and give directions for future work.
2.
RELATED WORK
Recently, there has been an increasing number of both
empirical and theoretical research on daily deals or what is
sometimes called voucher discounting. Most of this work
focused on studying Groupon and LivingSocial discounted
deals. For example, Dholakia studied the question of whether
Groupon promotions are profitable for businesses and which
businesses fare the best and worst after offering a Groupon
promotion [7, 8]. He found mixed results where some business owners reported their Groupon promotion was profitable and others regretted making the promotion based on
their experience of lower spending and return rates from
Groupon users. In another study, Arabshahi provided a detailed analysis that explains the Groupon business model
and its underlying principles [3]. In his paper, he explained
that the main challenge facing merchants lies in identifying
price-sensitive potential customers and offering them discounts. Therefore, Groupon can help merchants to apply
price discrimination through the ”highly discounted deals”
provided to a massive scale of price-sensitive subscribers. In
a similar work, Edelman et al. provided a theoretical study
of the economics of Groupon deals from the perspective of
participating merchants rather than from the perspective
of the deal service provider [9]. Their results indicate that
voucher discounts are naturally good fits for certain types
of merchants, and poor fits for others.
Similarly, Gupta et al. [11] investigated when are daily
deals profitable for business by interviewing over 2000 business that offered daily deal through “Groupon”. They found
that the success of a daily deal is far from certain and that
the return on investment varies widely. They identified the
types of businesses as a reliable predictor of profitability and
that daily deals are good for spas but bad for restaurants.
The work presented in [7, 8] and [11] can be viewed as
complementary to our work; while we focused on a data
driven scalable approach, [7, 8] and [11] focused on a more
labor intensive interview process. These two approaches are
not exclusive with the findings from one guiding the other.
Business Category
Restaurants & Bars
Beauty & Spas
Active Life
Shopping
Fitness & Instruction
Food
Health & Medical
Nightlife
Hair Salons
Arts & Entertainment
Total No.
337
189
151
116
87
84
77
73
73
72
Multiple Deals (%)
7.1
10
9.2
6.4
5.3
7.1
12.1
5.5
9
6.8
Closed (%)
8.3
4.7
1.9
0.9
0.0
11.9
0.0
1.4
6.5
1.4
Table 1: Statistics of Groupon Businesses in Chicago (Jan-July 2011)
Also, Ye et al. studied the group purchasing behavior of
daily deals in Groupon and LivingSocial and they proposed
a predictive dynamic model for group buying behavior [20].
Their model was able to predict the popularity of group
deals as function of time. They also found that the different
incentive mechanisms applied in Groupon and Living Social
(individual threshold versus collective threshold) lead to different propagation behavior, which finally lead to different
predictability.
While studying daily deals is interesting in itself, another
trend of research started to study the marriage between daily
deal sites and the growing consumer phenomena such as
Yelp. Byers et al. initiated the study of how daily deal
sites affect the reputation of the business and in particular
the business Yelp reviews [5]. In their first research paper,
the authors studied the interplay between social networks
and daily deal sites. They found that daily deal sites benefit
from significant word-of-mouth effects during sales events.
They also studied the effects of daily deals on the long-term
reputation of merchants, based on their Yelp reviews before
and after they run a daily deal. They found that the Yelp
ratings of Groupon-bearing consumers were on average 10%
lower than those of their peers.
In another study, Byers et al. rigorously evaluated various hypotheses about underlying consumer and merchant
behavior to understand the Groupon effect on businesses
[6]. They examined a number of hypotheses to justify the
Groupon effect. For example, they illustrated a poor business behavior, and Groupon user experimentation to be possible root causes of the Groupon effect. They also found an
evidence that on average Groupon users are no more critical
than their peers.
Similarly, Zervas tried to establish basic facts regarding
the evolving quality of the deals that Groupon offers [22].
He used Yelp ratings as a proxy for measuring the quality.
Using simple regression analyses, Zervas found a statistically
significant negative correlation between the time deals that
have been offered and the Yelp ratings of the merchants
who offered them. Further, he discussed some possibilities
that might cause these trends. For example, as Groupon
is expanding the number of deals it offers, it has to work
with some lower-rated merchants. Also, it is possible that
better-rated merchants dropping out of running Groupon
deals, and Groupon has to substitute them for merchants
with some lower-rated merchants.
Our work builds on [6] and [22] by explicitly modeling the
decision to offer a daily deal and leveraging the error in the
model to explain part of the unobserved factors in modeling
business performance.
Business reviews collected from Yelp have also been studied by Luca [16]. Luca evaluated the impact of Yelp reviews
on restaurant’s quarterly earnings in Seattle using the framework of regression discontinuity. Luca finds that the observed response to Yelp rating are consistent with bayesian
learning. Under the bayesian hypothesis, reactions to signals
are stronger when the signal is more precise (i.e., the Yelp
average rating contains more information when the number
of reviews is high). Moe precisely, a change in a restaurant’s
average rating has 50% more impact when the restaurant
has at least 50 reviews (compared to a restaurant with fewer
than 10 reviews).
Luca [16] also tests whether restaurants are gaming the
rating system using the McCrary [17] test. The intuition of
the test is as follows. Suppose that restaurants were gaming Yelp in a way that would bias the results. Then, one
would expect to see a disproportionately large number of
restaurants just above the rounding thresholds. Luca [16]
finds that this is not the case. The results presented in [16]
are related to our work in two ways. First, we use the concept of “Bayesian Learning” to derive statistically significant
predictors of both daily deal adoption and business failure.
Second, the McCrary test [17] suggests that the reviews on
Yelp truly reflect the opinion of the Yelp community and are
not being manipulated by the businesses on Yelp.
Pindyck and Rubinfeld [18] model the relation between
private school attendance and voting for property tax increases that are used in part to finance public schools. In
this application, the variables are whether children attend
private school, number of years the family has been at the
present residence , log of property tax , log of income and
whether the head of the household voted for an increase in
property taxes. Pindyck and Rubinfeld [18] wanted to test
the hypothesis that parents of children who attended private
school will have no incentive from an increase in property
taxes that finance public schools and will vote against any
such increases. Pindyck and Rubinfeld [18] model the bivariate outcomes of whether children attend private school and
whether the head of the household voted for other covariates. They conclude that the two outcomes are independent
and that the voting patterns of parents of children attending private schools do not differ from parents of children
attending public schools. Our work is related to [18] in that
we test for the independence of two binary outcomes; daily
deal adoption and business failure.
3.
DATA COLLECTION & ANALYSIS
Our dataset collection has two major components. First,
Business Category
Restaurants & Bars
Shopping
Food
Beauty & Spas
Health & Medical
Nightlife
Active Life
Arts & Entertainment
Hair Salons
Fitness & Instruction
Total No.
8490
4961
3259
2692
2666
1851
1301
1267
948
740
Closed (%)
17.7
10.6
14.9
5.9
2.14
16.0
6.3
6.0
5.3
7.6
Table 2: Statistics of Yelp Businesses in Chicago as
of July 2012
we collected data from Groupon as one of the top deal sites
that offers daily deals in Chicago. Second, we collected data
from Yelp for all the businesses in Chicago.
Groupon Data.
We used the Groupon data set compiled by Byers et al.
[5] which includes the web links of 16, 692 deals offered by
Groupon in 20 U.S. cities between January and July 2011.
In this paper we focus only on the subset of Groupon deals
offered in Chicago. We selected Chicago not just because it
is the third largest city in U.S but also it the home town of
Groupon. When the Groupon business is featured on Yelp,
Groupon occasionally uses that information to promote the
deal by including a link to the Yelp site as well as other
information (e.g. star rating and selected customer reviews).
However, in some cases Groupon does not mention the Yelp
link on the deal page even if the business has a Yelp link.
We are interested in Yelp since it provides a wide range
of information about the business. For Example, Yelp provides business location, number of reviews, date of review,
star ratings, review text and other features such as alcohol
license and price range. Moreover, Yelp indicates whether
the business is still in operation or whether the business has
closed by adding the string ”CLOSED” next to the business
name. Previous research has shown the potential of Yelp
to indicate business parameters and performance [16] as we
explained in section 2.
Groupon provides a convenient API 1 to collect information about the deals, however, we decided to develop our
web crawler to extract features that are not supported by
the API. For example, whether there is a link to Yelp or not.
We initially had 1861 Groupon deals from chicago, with approximately 60% of them had their Yelp links listed. For the
deals without Yelp links, we used the Yelp search feature to
find a match for the Groupon business on Yelp. Specifically,
we searched Yelp by the business name and the zip code
listed on the deal webpage. Typically, Yelp return search
results for relevant matches within the given zip code and
other nearby zip codes. However, we report only the query
results that exactly matched both the business name and
its zip code. By the end of this matching process, we successfully associated 1184 Groupon deals with Yelp links, we
call them ”GrouponDealsWithYelp”. We also observed that
some businesses offered multiple deals while others offered
only one deal. While these deals are supposed to be all in
Chicago, we found few cases where the deal zip code was
outside Chicago (for other branches of the business in other
states). Since we focus only on Chicago business we decided
1
http://www.groupon.com/pages/api
to filter these cases. Finally, we developed our web crawler
to extract the Yelp information for the businesses in the
set ”GrouponDealsWithYelp”. After filtering the businesses
that had zero reviews (since they don’t provide any information about the business), we observed 985 businesses with
Groupon deals and Yelp links. We call this set GrouponBusiness”. Table 1 provides statistics of the set ”GrouponBusiness” for the top 10 business categories.
Yelp Data.
We crawled the Yelp site to collect all the businesses that
appear in Chicago (regardless whether they offer a deal or
not). Yelp uses a structured format that arranges business
names by alphabetical order. We initially had 38, 000 businesses listed in Yelp. After filtering all the cases that have
zero reviews, we had 32, 424 with approximately 9% failed
businesses (closed) and approximately 4% offered Groupon
between January and July 2011. We refer to this set as our
Yelp Population to represent the real population of businesses in Chicago. However, we expect Yelp data would
represent certain business categories (e.g. restaurants) more
than others (e.g. Insurance). Therefore, we decided to analyze only the top business categories. In Table 2 we provide
statistics of the dataset we collected from Yelp for the top
10 business categories.
In the next sections, We proceeded to build a model for
predicting failure using Yelp data. The details of the model
are shown in Section 5.2. To build the bivariate model of
Groupon adoption and business failure, we had to restrict
our analysis to the businesses that did not offer a Groupon
but were operating during the same period as the ”GrouponBusiness”. We use the date of the last review posted for the
business as a proxy for the closing date. We refer to the set
that includes all of the businesses that did not offer Groupon
deals and did not fail before January 2011 as ”nonGrouponBusiness”.
Analysis.
Table 1 provides statistics of the set ”GrouponBusiness”
for the top 10 business categories. Also, Table 2 provides
statistics of our data Yelp Population which includes both
the two sets ”GrouponBusiness” and ”nonGrouponBusiness”.
From Table 1 and Table 2, we observe some interesting patterns. First, we observe a difference in the ranks of the
business categories between the total population of Yelp
data and the ”GrouponBusiness” except for the first category
”Restaurants & Bars”. We conjecture that some business
categories could be popular on Yelp but they don’t have the
incentives to offer daily deals. We also observe that the highest percentage of closed businesses comes from the categories
”Restaurants & Bars”, ”Food”, and ”Nightlife”. In addition,
we observed that some business in the set ”GrouponBusiness” had the incentives to offer multiple daily deals during
the six month period we analyzed.
As we discussed in section 2, previous research work emphasized the potential of Yelp as a proxy measure for business key performance indicators (e.g. survival, consumer
appeal, and revenue) [16, 5, 6, 22]. Luca in [16] found
that 69% of restaurants in Seattle are listed on Yelp. Also,
Luca showed that changes in Yelp ratings are associated
with changes in revenues [16]. These studies indicate the
potential of Yelp data as representative of the true population. However, to the best of our knowledge, none of
this research analyzed Yelp as a source of business failure
information. Therefore, to test the representativeness of
Yelp data as a source of business failure information, we
compared the number of closed businesses collected from
Yelp to the number of bankrupted businesses as reported by
bankruptcy filings of the Northern Illinois (which includes
Chicago) open court records collected by the bankruptcy
data project at Harvard 2 . Both the data from Yelp and
court bankruptcy filings are between January 2006 and July
2012. Figure 1 shows the plot of the two normalized time
series data . As shown in the plot, there is a strong correlation between the number of closed businesses computed from
Yelp and the number of bankrupted businesses (correlation
coeff. = 0.7164). Also, the two datasets have a similar trend
as shown in Figure 1.
Figure 1: Normalized Yelp Closed Businesses versus Normalized Bankrupted Businesses between Jan
2006 and July 2012
4.
ECONOMETRIC FRAMEWORK
In this section, we develop the econometric model. To
motivate the need for an econometric model, consider the
task of deciding whether running a daily deal increases the
risk of business failure. We can model this as a regression
problem of the form
f ailurei = β ∗ dailydeali + ei
(1)
where ”f ailurei ” is whether business ”i” fails, , ”dailydeali ”
indicates whether the business ran a daily deal and ei is the
error term and captures all the factors that are not included
in the model. These factors are referred to as the ”unobserved heterogeneity”. The unobserved heterogeneity can
include factors such as location specific risk, category risk
.. .etc. It can also include factors that are correlated with
the business decision to run a daily deal such as whether
the business is struggling. Consider for example the class of
struggling businesses that use daily deals as a last resort to
attract more customers. In expectation, a struggling business will have a higher risk of failure. However, if we were
to perform a regression analysis using Equation 1, we will
over estimate the impact of daily deals (β) because a large
number of the struggling businesses that used daily deals
failed. While in fact, these businesses did not fail because
of daily deals, they failed because of internal problems that
happened to be correlated with the decision to offer a daily
2
http://bdp.law.harvard.edu
deal. In this case, we have an omitted variable bias [10]. The
omitted variable bias will only arise however, if the omitted
variable (which in our case is whether the business was struggling), is correlated with one of the regressors ( the decision
to use daily deal).
The problem can also be viewed from the lens of endogeneity. A perquisite to any regression model is that the regressors are exogenous, [10] i.e., the regressor variable comes
from outside of the model and cannot be explained by any of
the variables of the model. However, as we have seen in the
above example, the regressor dailydeal was correlated with
the errors. Consequently, the variable dailydeal is not exogenous as it it can be explained in part by the errors. The
omitted variable bias is a form of endogeneity that results
in a biased estimator.
If we wanted to estimate the impact of daily deals on
the business survival, we will need to find an instrument
[10]. An instrument is a variable that is correlated with the
endogenous variable (dailydeal) but is not correlated with
error (struggling business). For example, we could use the
size of the daily deals providers sales force as an instrument
since it is likely to be correlated with decision to adopt daily
deals but unlikely to be correlated with the error term. A
model with an instrumental variable is estimated using 2
stage least squares (2sls). In future work, we will address
the use of instrumental variables approach.
However, in this paper, we don’t directly attempt to model
the impact of the daily deal on business survival. Instead,
we focus on assessing the correlation between the unobserved
factors that make a business offer a daily deal and the unobserved factors that influence failure. In that vein, we consider two different econometric models. In section 4.1, we
use a bivariate probit model to check whether the errors in
predicting if the business decision to run a daily deal are
correlated with the errors in predicting the business failure.
Since the errors model the unobserved heterogeneity, the
bivariate model will help identify whether the unobserved
factors that make a business run a daily deal are also correlated with the unobserved factors that contribute to the
business failure. In section 4.2, we relax the conditions that
the dependent variables (run a daily deal and failure) are
binary and use the framework of seemingly unrelated regression ” (SURE) to answer the same question addressed
by the bivariate profit.
Our decision to use the SURE model is motivated by two
factors. First, the bivariate probit model makes the explicit
assumption that the error terms are jointly normal. While
this might sound like a reasonable assumption, we have no
way of checking its validity. On the other other hand, the
SURE model does not make any assumptions about the joint
distributions of the error, but rather uses the variance of
these distributions to derive its estimate. Therefore, the
use of two models provides us with a robustness test as in
[19]. The second reason we consider the SURE model is that
recent developments in econometrics [2] that used actual experimental data do not show any advantages of enforcing the
limited range of dependent variables.
An important question is whether a single model best explains the two dependent outcomes (daily deals adoption
and survival ) or whether we need two separate models. In
section 4.3, we describe a specification test based on the
likelihood of the data and show how they relate to parametric technique, that test the null hypothesis of no correlation
between the error terms.
4.1
Bivariate Probit
Assume that we have a random sample of N observations
where each observation is donated by i such that i = 1, ..., N .
In ordinary regression models, we typically observe only one
dependent variable for each observation Y = (Y1 , ..., YN ).
However, in the general case general case, we can observe
multiple dependent variables for each observation. Let Yji
denote the response of the ith observational unit for the j th
dependent variable. A typical situation is the case when we
observe 2 variables, such that Y1i and Y2i are two binary
dependent variables.
Traditional probit models can generally be described as latent variable models, in which we define a latent variable Y ∗
such that Y = 1(Y ∗ >0) . In this section, we consider the bivariate probit model. The bivariate probit model belongs to
the generalized class that is usually used to estimate several
correlated binary variables jointly. These often represent
two interrelated decisions, for example, to adopt two different, but related, policy initiatives. In the bivariate probit
model, we have two separate probit models with correlated
error terms. Specifically, we have two binary dependent variables for each ith observational unit : Yji = Y1i , Y2i such that
j = 1, 2. Therefore, we have the following model
Y1i∗ = X1i β1 + ε1i
(2)
Y2i∗
(3)
= X2i β2 + ε2i
use maximum likelihood estimation to estimate the parameters. The likelihood function L is defined as:
Y
L =
P (Y1i = 1, Y2i = 1)Y1i Y2i
i∈N
P (Y1i = 0, Y2i = 1)(1−Y1i )Y2i
P (Y1i = 1, Y2i = 0)Y1i (1−Y2i )
P (Y1i = 0, Y2i = 0)(1−Y1i)(1−Y2i )
Substituting the latent variables Y1∗ and Y2∗ in the Probability functions and taking the logarithm gives the log likelihood function LL:
X
LL =
Y1i Y2i ln P (ε1i > −X1i β1 , ε2i > −X2i β2 )
i∈N
+(1 − Y1i )Y2i ln P (ε1i < −X1i β1 , ε2i > −X2i β2 )
+Y1i (1 − Y2i ) ln P (ε1i > −X1i β1 , ε2i < −X2i β2 )
+(1 − Y1i )(1 − Y2i ) ln P (ε1i < −X1i β1 , ε2i < −X2i β2 )
(9)
After rearranging the terms, the log-likelihood function becomes:
X
LL = =
Y1i Y2i ln Φ(X1i β1 , X2i β2 , ρ)
i∈N
+(1 − Y1i )Y2 ln Φ(−X1i β1 , X2i β2 , −ρ)
where Y1i∗ and Y2i∗ are the latent variables and they are related to the dependent variables by the following equation
∗ >0)
Y1i = 1(Y1i
(4)
∗ >0)
Y2i = 1(Y2i
(5)
The vectors X1i denotes the [N1 × 1] vector of exogenous
regressor for dependent variable ”1” . X2i denotes the [N2 ×
1] vector of exogenous regressor for dependent variable ”2”
. In the bivariate probit model, the main assumption is
that error terms ε1i , ε2i are independent across observations
“i” but may have cross-equation correlations. Therefore, we
have E[εji εjk |x] = 0 ∀i 6= k. In addition, the error terms
are drawn from a bivariate normal distribution [10]
0
1 ρ
ε1i
|X ∼ N
,
(6)
ε2i
0
ρ 1
Where ρ is a correlation parameter denoting the extent to
which the two ε’s covary. The conditional expectation for
the bivariate normal distribution is given by
E(ε2i |ε1i > z) = ρ
φ1 (z)
Φ1 (−z)
(7)
where Φ1 , φ1 are the univariate normal cumulative distribution and density functions respectively.Equation 7 is the
Inverse Mills ratio and has been used extensively in econometrics , [13] [10]. Equation 7 shows that the in case of bivariate probit, the errors are not independent, for example
the error in estimating the probability of offering a Groupon
deal corresponds in expectation to a large error in estimating the probability of business failures. The errors typically
correspond to unobserved variables and by implication the
Inverse Mills ratio indicates that these two variables move in
synchronization. Fitting the bivariate probit model involves
estimating the values of the parameters β1 , β2 , and ρ. We
(8)
+Y1i (1 − Y2i ) ln Φ(X1i β1 , −X2i β2 , −ρ)
+(1 − Y1i )(1 − Y2i ) ln Φ(−X1i β1 , −X2i β2 , ρ)
(10)
Note that Φ is the cumulative distribution function of the
multivariate normal distribution|bivariate normal distribution and φ is the corresponding density function. Y1i and
Y2i in the log-likelihood function are observed variables being equal to one or zero.
4.2
Seemingly Unrelated Regression
In the previous section, we enforced the constraint that
the dependent variable have a limited range limited dependent variables. The bivariate probit frameworks explicitly
uses bivariate normal distribution to model the joint distribution of the error terms. To test whether this assumption
is indeed valid, we relax these constraints and use the framework of seemingly unrelated regression (SURE) proposed by
Zellner [21]. Furthermore, there is mounting evidence in the
econometrics literature [2] that argues in favor of using Ordinary least squares (OLS) even when the dependent variables
are binary.
Y1i = X1i β1 + ε1i
(11)
Y2i = X2i β2 + ε2i
(12)
As before, the assumption of the model is that error terms
ε1i , ε2i are independent across observations ”i” but may have
cross-equation correlations. Therefore we have E[εji εjk |x] =
0 ∀i 6= k and
ε1i
ε2i
|X ∼ N
0
0
,
σ11 2
ρσ11 σ22
ρσ11 σ22
σ22 2
(13)
X11
y11
y12 X12
.. ..
. .
y1N X1N
y21 = 0
y22 0
. .
.. ..
0
y2N
0
ε11
ε12
0
..
..
.
.
0
β1 + ε1N
X21
β
ε
2
21
ε22
X22
.
..
.
.
.
X2N
ε2N
(14)
Or in a more compact notation Y = Xβ + ε,
E[ε|X] =
0, Var[ε|X] = Ω.
The SURE regression differs from the OLS in that the
covariance matrix is not spherical i.e., Var[ε|X] 6= σ 2 In
where IN is the identity matrix. In ordinary least squares,
the Best Linear unbiased estimator for the parameters β
is given by βbOLS = (X 0 X)−1 X 0 y. Once we have a nonspherical covariance matrix OLS is not efficient. To overcome this restriction, we use feasible generalized least squares
which is a two stage estimator. In the first stage, we run
ordinary least squares estimation assuming that the two
equations are independent. The residuals from the OLS
are used to estimate the elements of the covariance matrix
σ̂ij = N1 εˆi T εˆj . In the second stage, we run weighted least
squares using the previously estimated covariance matrix.
Feasible generalized least squares method estimates β by
minimizing the squared Mahalanobis length of the residual
vector β̂F GLS = arg min (Y − Xb)0 Ω−1 (Y − Xb) (note the
b
in case of ordinary least squares Ω is diagonal and therefore
β̂OLS = arg min (Y − Xb)0 (Y − Xb) ). The explicit form of
b
the estimator is given by
β̂F GLS = (X 0 Ω−1 X)−1 X 0 Ω−1 Y.
(15)
To test whether the two equations are best modeled using
a SURE or can be modeled using ordinary least squares,
it suffices to test whether the errors ε1t , ε2t are correlated.
The Breusch-Pagan test [4] which is widely used in detecting
heteroskedasticity can be applied.
4.3
Specification Testing
We use the Akaike Information Criteria (AIC) [10] to test
whether the data is best described by two separate models
or a single joint model. In the case of the bivariate probit
model, the joint model will have one additional parameter
(the correlation coefficient). We therefore compare the log
likelihood of the joint models in Equation 10 to the log likelihood of the separate equations and test whether the difference is greater than “1.0”. In the case of SURE model,
we replace the log likelihood with the the sum of squared
residuals [10]. For the parametric approach, we employ the
Cramer-Rao bound to compute parameter’s mean and variance from the maximum likelihood estimator [10]. We use
a “t-test” to test whether the parameters including the correlation coefficient ρ are different from “0”. In the SURE
setting, we use the Breusch -Pagan framework to test the
correlation coefficient.
5.
EXPERIMENTS AND RESULTS
In this section, we present the experimental setup and results of several regression models both in a univariate framework and a multiple equation framework. In the univariate
framework, we use a probit model to analyze the significance
of the different factors that may impact the business survival. Similarly, we use a probit model to analyze the significance of the different factors that influence the business
decision to make a daily deal in Groupon. In the multiple
equations framework, we use bivariate probit and seemingly
unrelated regression frameworks to jointly model the business survival and the business daily deal decision. Here, our
goal is to specifically investigate whether the unobserved
factors that make the business offer a daily deal are correlated with the unobserved factors that impact the business
survival. We conduct several statistical tests to test if the
data is best described by a joint model and whether the correlation between the unobserved factors is significant. Our
results indicate that joint models fit the data better than
single univariate models and there is strong significant correlation for the two business categories restaurants and spas.
5.1
Sampling
When modeling business survival and Groupon decision
we need to take into account the relative frequency of both
events. On one hand, the probability of a business failure in
any given year is fairly low (in order of 8%). On the other
hand, daily deals is a relatively new phenomena which is
still being evaluated, the number of businesses that leverage a daily deal are fairly low compared to the businesses
that don’t offer daily deals. In that vein, we are attempting
to model two rare events : daily deal decision and business
failure. Statistical models tend to underestimate the probability of rare events[10]. Since the vast majority of the businesses will be non-groupon and non-closed, the model will
assign a large negative constant to the two equations that
model the daily deal and failure models. The large constants
will make the error terms small and can impact the model’s
ability to detect the correlation between the error terms in
the two equations. To address this problem, we first apply
the Bivariate Probit and the SURE to the entire data set.
We then restrict the models such that there is no constant
terms. We find that in the first case (unrestricted full sample), a positive but not statistically significant correlation
between the error terms. In case of the full sample but noconstant , we find that there is a strong positive correlation
between the error terms. This confirms our intuition about
the inability of the full data model to capture the two rare
events. Therefore, we randomly sampled from the population of ”nonGrouponBusiness” to account for the sparsity
problems and selection biases that can be caused by data
collection.
5.2
Univariate Probit Models
A probit model is a type of regression used to model a
binary dependent variable. Here, we define two binary dependent variables. First, we define the variable isClosed to
represent whether the business has failed (isClosed=1) or the
business is still operating (isClosed=0). Second, we define
the variable isGroupon to represent whether the business has
made at least one daily deal (isGroupon=1) at Groupon or
the business did not make any deals (isGroupon=0). We develop a business survival model and Groupon decision model
to analyze the factors that influence isClosed and isGroupon
respectively.
Business Survival Model.
Symbol
isClosed
Variable
isClosed
isGroupon
isGroupon
f zrisk
f prisk
gzrisk
gprisk
rate
nreview
price
Fail Ziprisk
Fail Pricerisk
Groupon Ziprisk
Groupon Pricerisk
Rating
Reviews
Price Category
Description
A binary outcome indicating whether the business failed (isClosed =1) or operating
(isClosed = 0)
A binary outcome indicating whether the business made a deal (isGroupon =1) or
no deals (isGroupon = 0)
The percentage of failed businesses in the same zip code
The percentage of failed businesses in the same price category
The percentage of businesses that made Groupon deals in the same zip code
The percentage of businesses that made Groupon deals in the same price category
Average Yelp Rating
Number of Yelp Reviews
{1,2,3,4} From cheap to most expensive
Table 3: Description of Variables
Table 4: Yelp Population Data
Table 5: Yelp Population Data
Dependent Variable: isClosed
AUC = 0.674
Variable
Coefficient
(Std. Err.)
Fail Ziprisk
6.786∗∗
(0.317)
Fail Pricerisk
7.039∗∗
(0.380)
Rating × Reviews -0.003∗∗
(0.000)
Reviews
0.008∗∗
(0.002)
Rating
-0.061∗∗
(0.010)
Intercept
-2.287∗∗
(0.060)
Significance levels : † : 10%
∗ : 5%
∗∗ : 1%
Dependent Variable: isGroupon
AUC = 0.894
Variable
Coefficient
(Std. Err.)
Groupon Ziprisk
4.521∗∗
(0.136)
Groupon Pricerisk 10.771∗∗
(0.888)
Rating × Reviews -0.001∗∗
(0.000)
Reviews
0.006∗∗
(0.001)
Rating
0.057∗∗
(0.021)
Intercept
-2.848∗∗
(0.093)
Significance levels : † : 10%
∗ : 5%
∗∗ : 1%
We model the business survival variable isClosed as a probit
function of a number of business variables we collected from
Yelp (refer to Table 3 for notations):
isClosed = f (rate×nreview, rate, nreview, f zrisk, f prisk)
(16)
We tried a number of other specifications and selected the
specification of Equation 16 based on the AIC. Table 4 shows
the results of the model trained on Yelp population data we
collected from Chicago. Although the model is simple, it
fits the data well (AUC=0.674) and all the variables are
statistically significant (p − value < 0.01).
We can gain further insight by examining the marginal
contributions of each of the factors to the probability of survival. Therefore, we make the following observations from
the results:
• When the average Yelp rating is higher, the risk of
failure gets lower. Higher rated business tend to be
more successful.
• When the number of reviews is higher, the risk of failure increases. A business with high number of reviews
has been around for a longer time and its risk increases
with time.
• When the average Yelp rating weighted by the number of reviews rate × nreview increases, the risk of
failure gets lower. This is consistent with the theory
of Bayesian learning presented in [16]. This is because
the average Yelp rating weighted by the number of
reviews gives more precise information compared to
rating or number of reviews only.
• The business location makes a difference, some zip
codes are riskier than others. This is consistent with
previous work on restaurant failure by zip code.
• The price risk computed for the business price category
is also significant similar to the zip code risk.
Groupon Adoption Model.
Similar to the survival model, we model the Groupon decision variable isGroupon as a probit function of a number of
business variables from Yelp (as in Table 3):
isGroupon = f (rate×nreview, rate, nreview, gzrisk, gprisk)
(17)
We selected the specification of Equation 17 based on the
AIC . Table 5 shows the results of the model trained on
Yelp population data from Chicago. The model has a good
accuracy of AUC = 0.894, that shows the ability of the
model to predict whether a business will decide to make
a daily deal based on some business parameters collected
from Yelp. Also, the regressors are statistically significant
(p − value < 0.01). We can gain further insight by examining the marginal contributions of each of the factors to the
probability of offering a daily deal. Therefore, we make the
following observations from the results:
• When the average Yelp rating is higher, the probability
of daily deal increases. We conjecture that this is due
in part to how daily deal sales force selects the business
to approach.
• When the number of reviews is higher, the probability
of daily deal increases. We conjecture that this is also
due in part to how daily deal sales force selects the
business to approach.
• When the average Yelp rating weighted by the number
of reviews rate × nreview increases, the probability of
daily deal gets lower. This is an indication of a more
successful business that would not need to make a deal
at Groupon, especially that Groupon takes 50% of the
deal revenue [3]. This results is also consistent with
the theory of Bayesian learning presented in [16].
• The business location in terms of zip code makes a
difference, some zip codes are more likely to offer daily
deals. The higher the number of businesses that make
a deal in the same zip code, the higher the chance of a
business to adopt a deal. This is part due to completive
pressure and in part due to how daily deal sales team
targets geographical areas.
• The business price category is also significant as we
mentioned before.
5.3
Joint Model
We jointly modeled the Groupon and failure models using
Equations 17 and 16 using a bivariate probit model. Table 6
shows that the data is best modeled by a bivariate probit
since the log likelihood of the bivariate model differs from
that of the two separate equations by more than ”1.0”. In addition, Tables 7 and 8 show that for the two largest daily deal
categories, the correlation between the unobserved factors
”ρ” is positive and significant (0.281 in Restaurants/Bars,
0.24 in Beauty/Spas). As a robustness test, we used the
SURE model to compute the correlation between the unobserved errors and tested its significance using the BreuschPagan test (correlation=0.054 in Restaurants/Bars, 0.060 in
Beauty/Spas). It should be noted that the bivariate probit
model operates in the space of latent variables that can be
in range [−∞, 0] when the dependent variable is ”0” and in
the range of [0, ∞] when the dependent variable is ”1”. The
residuals are therefore computed in that space and can assume large values. On the other hand, the SURE model
operates directly in the space of dependent variable that assume range between ”0” and ”1”. Therefore the residuals are
smaller in case of SURE. This explains why the correlation
coefficient in case of bivariate probit differs from that in case
of SURE model.
Previous work has shown that restaurants with their high
marginal cost, low fixed cost and inability to schedule the arrival of daily deal customers [11] are not well suited to daily
deals. On the other hand, Spas with their low marginal cost,
high fixed cost and ability to schedule daily deal customers
are better suited for daily deals. Conversely, a daily deal is
a more desperate measure for a restaurant than a Spa. This
is validated by our results, if a restaurant offers a daily deal
with out having any strong reason to do so, its probability of
failing increases more than a Spa that had some motivation
for offering a daily deal. Similar to what we did in the univariate case, we analyzed the contributions of factors, and
we have the following observations:
• When the ratings weighted by the number of reviews
rate × nreview increases, the risk of failure gets lower.
Also, the probability of offering a daily deal gets lower.
• Unlike the univariate case, the ratings rate is not a significant factor. However, the weighted ratings are significant. This is consistent with the theory of Bayesian
learning presented in [16].
• As in the univariate case, when the number of reviews
is higher, the risk of failure increases.
Restaurants and Bars
Model Name
No. Params
isGroupon+isClosed
10
Bivariate Probit
11
Beauty and Spas
Model Name
No. Params
isGroupon+isClosed
10
Bivariate Probit
11
AIC
1976.46
1968.33
AIC
1029.97
1028.89
Table 6: AIC for univariate versus bivarite probit
We have also tried other business categories, for example,
Health & Medical, Active Life,..etc. However, there was a
lack of significance of correlation between the unobserved
factors for these categories. We conjecture the lack of significance is because we did not have enough samples in those
categories. In future work, we aim to extend our study to
daily deal data from Yipit [1] as well as collecting data from
other sources such as city department of revenue to gain a
better access to businesses that are not represented well by
our current data.
6.
CONCLUSION AND FUTURE WORK
In this work, we studied whether daily deal adoption signals additional information about the business. For the two
largest daily deal categories restaurants and spas, we found
that the unobserved factors that contribute to a business
decision to offer a daily are positively correlated with unobserved factors that contribute to a business failure. Restaurants had a higher correlation while spas had a lower correlation. These results indicate that daily deal provide a
strong signal of business survival for restaurants and to a
lesser extent for spas. Our results also show that social media sites such as Yelp provide a rich set of information that
can be used to model business. In particular, we found that
consistent with Bayesian learning theory, the rating of business weighted by the number of reviews provides a statistically significant predictor of business failure.Ceteris paribus,
a business with a high number of positive reviews has higher
odds of survival.
In future, we plan to extend our work in multiple directions. The first direction is to consider other business categories. In this work, we were limited to the daily deal data
from [5]. Second, we plan to consider daily deal providers
other than Groupon. In the case of Groupon, we modeled
the marketers decision to use daily deal as a binary decision.
We also modeled the business failure as a binary decision.
This allowed us to leverage the framework of bivariate probit
to jointly model daily deal and business failure. In the more
general setting, the marketer can choose to offer a daily deal
through a number of providers or not to offer a deal. In
the general case, the marketer’s choice is best modeled as a
multinomial. We plan to investigate techniques from Multilevel Multiprocess Models (MLMP) [15]. The data needed
to undertake these two directions can be obtained from daily
deal aggregators such as Yipit [1]
We also plan to address the dynamics of Groupon adoption and business failure. In particular, we plan to test
whether daily deal adoption is stationary, whether business
failures are stationary and whether the two time series are
co-integrated [12]. This will allow us to investigate whether
there is a long term equilibrium between daily deal adoption
and business failure.
Table 7: Restaurants and Bars: Bivariate Probit
Table 8: Beauty and Spas: Bivariate Probit
Bivariate Probit
Coefficient
(Std. Err.)
Equation 1 : isGroupon
Groupon Pricerisk 13.978∗∗
(2.122)
Groupon Ziprisk
5.715∗∗
(0.535)
Rating × Reviews -0.002∗∗
(0.001)
Reviews
0.010∗∗
(0.002)
Rating
0.180∗
(0.073)
Intercept
-3.385∗∗
(0.293)
Equation 2 : isClosed
Fail Ziprisk
8.791∗∗
(1.021)
Fail Pricerisk
12.928∗∗
(3.541)
Rating × Reviews -0.002∗
(0.001)
Reviews
0.007†
(0.004)
Rating
-0.013
(0.070)
Intercept
-4.202∗∗
(0.495)
Equation 3 : Joint
athrho
0.288∗∗
(0.09161)
rho
0.281
(.08440)
SURE: Breusch-Pagan test of independence
Correlation
0.0543∗∗
Significance levels : † : 10%
∗ : 5%
∗∗ : 1%
Bivariate Probit
Coefficient
(Std. Err.)
Equation 1 : isGroupon
Groupon Pricerisk 17.203∗∗
(3.326)
Groupon Ziprisk
4.460∗∗
(0.373)
Rating × Reviews -0.004∗∗
(0.001)
Reviews
0.024∗∗
(0.006)
Rating
-0.087
(0.059)
Intercept
-2.613∗∗
(0.288)
Equation 2 : isClosed
Fail Ziprisk
5.199∗∗
(1.556)
Fail Pricerisk
8.906∗
(4.538)
Rating × Reviews -0.007∗
(0.003)
Reviews
0.026∗∗
(0.010)
Rating
0.103
(0.090)
Intercept
-4.201∗∗
(0.719)
Equation 3 : Joint
athrho
0.25†
(0.1433)
rho
0.246
(0.1346)
SURE: Breusch-Pagan test of independence
Correlation
0.060∗∗
Significance levels : † : 10%
∗ : 5%
∗∗ : 1%
Variable
Last but not least, we plan to investigate the causal impact
of Groupon on metrics other than business failure. We plan
to investigate the Groupon impact on sales, number of orders
and order size. This will help us better understand how
the daily deals impact revenue: a) do they impact revenue
through a change in the number of orders. or b) do daily
deals have a stronger impact on order size or c) do daily deals
have an impact on both number and size of orders. We will
also look at using instruments [10] to get an estimate of the
casual impact of Groupon. We are investigating Groupon
sales force as a potential instrument.
7.
REFERENCES
[1] http://www.yipit.com/.
[2] J. D. Angrist and J.-S. Pischke. Mostly Harmless
Econometrics: An Empiricists Companion. Princeton
University Press, 2008.
[3] A. Arabshahi. Undressing groupon: An analysis of the
groupon business model, 2010.
[4] T. Breusch and A. Pagan. Simple test for
heteroscedasticity and random coefficient variation.
Econometrica, 47(5), 1979.
[5] J. Byers, M. Mitzenmacher, and G. Zervas. Daily
deals: Prediction, social diffusion, and reputational
ramifications. In Proceedings of the fifth ACM
international conference on Web search and data
mining, pages 543–552. ACM, 2012.
[6] J. W. Byers, M. Mitzenmacher, and G. Zervas. The
groupon effect on yelp ratings: a root cause analysis.
In Proceedings of the 13th ACM Conference on
Electronic Commerce, EC ’12, pages 248–265, New
York, NY, USA, 2012. ACM.
[7] U. Dholakia. How effective are groupon promotions for
businesses? 2010.
[8] U. Dholakia. How businesses fare with daily deals: A
Variable
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
multi-site analysis of groupon, livingsocial, opentable,
travelzoo, and buywithme promotions. 2011.
B. Edelman, S. Jaffe, and S. Kominers. To groupon or
not to groupon: The profitability of deep discounts.
2010.
W. H. Green. Econometric Analysis. Pearson, 6
edition, 2008.
S. Gupta, T. Keinungham, R. Weaver, and
L. Williams. Are daily deals good for merchants?
J. D. Hamilton. Time Series Analysis. Princeton
University Press, 1994.
J. Heckman. Sample selection bias as a specification
error. Econometrica, 47(1):153–162, 1979.
G. Z. John Byers, Michael Mitzenmacher. Daily deals:
Prediction, social diffusion, and reputational
ramifications. In 5th ACM International Conference
on Web Search and Data Mining (WSDM), 5th ACM
International Conference on Web Search and Data
Mining (WSDM).
L. Lillard and L. Waite. Til death do us part: marital
disruption and mortality. The American Journal of
Sociology, 100, 1995.
M. Luca. Reviews, reputation, and revenue: The case
of yelp.com. Technical Report 12-016, Harvard
Business School Working Paper, 2011.
J. McCrary. Manipulation of the running variable in
the regression discontinuity design. Journal of
Econometrics, 2008.
R. Pindyck and D. Rubinfeld. Microeconomics.
Prentice Hall, 2008.
K. Tucker and A. Goldfarb. Online display
advertising: Targeting and intrusiveness. Marketing
Science, 30(3):389–404, 2011.
M. Ye, T. Sandholm, C. Wang, C. Aperjis, and B. A.
Huberman. Collective attention and the dynamics of
group deals. In Proceedings of the 21st international
conference companion on World Wide Web, WWW
’12 Companion, pages 1205–1212, New York, NY,
USA, 2012. ACM.
[21] A. Zelner. An efficient method of estimating seemingly
unrelated regression equations and tests for
aggregation bias. Journal of the American Statistical
Association, 57:348–368, 1962.
[22] G. Zervas. The evolving quality of groupon deals
(draft). 2012.
| 5cs.CE
|
1
ComSens: Exploiting Pilot Diversity for Pervasive Integration of
Communication and Sensing in MIMO-TDD-Frameworks
Mohammadreza Mousaei, Mojtaba Soltanalian, and Besma Smida, Senior Member, IEEE
arXiv:1709.07407v1 [cs.IT] 5 Sep 2017
Department of Electrical and Computer Engineering, University of Illinois at Chicago (UIC), Chicago, IL, USA
Email: {mmousa3, msol, smida} @uic.edu
Abstract—In this paper, we propose a fully-integrated radar and
communication system – named ComSens. We utilize two different
pilot sequences (one for uplink and one for downlink) with the
condition that they must be uncorrelated to each other. Within
such a framework, the signal received from end-user and the backscattered signal from the desired objects have uncorrelated pilots.
Thus, the base-station is able to distinguish data signal from user
and back-scattered signal from object. We assume a time division
duplex (TDD) framework. The pilot sequences are designed for
MIMO channels. We evaluate channel MSE as a figure of merit for
communication system. We also show that the designed pilots are
uncorrelated for a range of time lags. Moreover, designed uplink
pilot has negligable autocorrelation for a range of time lags leading
to an impulse-like autocorrelation for radar sensing.
I. I NTRODUCTION
Due to the increasing demand in wireless communication
services, achieving higher data rates and more reliable transmissions have become a fundamental goal [1]–[4]. Given the
ever-increasing demand for both high-speed data services and
accurate remote sensing capabilities, modern wireless systems
will increasingly require more efficient strategies for use of the
available frequency spectrum [5]–[9]. In particular, the coexistence of communication and radar systems has recently attracted
a significant research interest [10]–[14]. For example, different
schemes for coexisting communication and radar systems has
been proposed; see e.g. [15]–[21] and the references therein.
While integrating radar and communication operation in one
system has been considered in the literature, such efforts are
typically centered around incorporating communication as a
secondary operation alongside a primary radar operation. The
research in [15] exploits the main lobe of the beam for radar
purposes, and the sidelobes (which are of no significance to
the radar pulse compression) for data transmission purposes.
The research works [16], [17] approach the same problem by
devising similar methods to allow comparably low data rates
into an already existing radar system.
In this work, we propose an integrated system of communication and sensing (which we call ComSens) that relies on the
communication pilot overhead— thus paving the way for pilot
design and exploiting pilot diversity to achieve a satsifactory
performance in both communication and radar tasks. Note that:
Pilot (or training) based channel estimation is very common
[22], [23]. Accurate knowledge of channel state information
(CSI) is important for wireless communication systems [24],
[25]. Most modern wireless systems acquire the CSI with the
assistance of pilot signals (a.k.a. training sequences) that are
inserted within the transmit signals periodically [26], [27]. In
such scenarios, the transmitter sends training sequences – known
to the receiver – enabling the receiver to perform channel
estimation on the basis of the received training symbols.
Communication devices are more ubiquitous than radar systems [28], [29]. We note that incorporating the communication
signals in the primary radar probing waveforms may not be an
efficient fusion of communication and radar systems. In fact,
the communication task must play a primary rule not only
because of the pervasive usage of comuunication devices, but
also the fact that the communication systems typically require
a larger capacity of conveying information than radar systems.
Additionally, considering the communication operation as the
primary lays the ground for making the radar systems ubiquitous
(for example having radar capability on cellphones).
A. Contributions
The key departure from prior works on integrated radarcommunication systems is that we (a) incorporate a radar system
in an already existing communication system, particularly by (b)
using the novel idea of designing different training signals such
that sensing and communication can co-exist. (c) incorporating
such a system in TTD framework in a MIMO system and
designing the pilot sequences.
B. Notation
We use bold lowercase letters for vectors/sequences and bold
uppercase letters for matrices. (·)T , (·)∗ and (·)H denote the
vector/matrix transpose, the complex conjugate, and the Hermitian transpose, respectively. kxkn or the ln -norm of the vector
1
P
x is defined as ( k |x(k)|n ) n where {x(k)} are the entries of
x. The Frobenius norm of a matrix X (denoted by kXkF ) with
P
1
2 2
entries {X(k, l)} is equal to
. Finally, R and
|X(k,
l)|
k,l
C represent the set of real and complex numbers, respectively.
II. F USION OF C OMMUNICATION AND R ADAR O PERATIONS
In this section, we describe the problem settings and explain
our proposed scheme in more details.
A. The Proposed Integration Scheme
The ComSens framework operates by exploiting the two-way
communication between the base-station and end-users. Before
2
Fig. 1. The ComSens communication and sensing integration framework.
discussing the issue of designing the pilot sequences, we will first
address how the base-station and end-user exchange messages
and sense the environment at the same time and over the same
frequency. We consider a model with M -user multiple-accessbroadcast channel (MABC) as depicted in Fig. 1. Note that such
a system model, in which several end-users wish to exchange
messages with a central node, or base-station, is a model that
captures the behavior of current and future cellular networks. We
assume half-duplex end-user nodes that may transmit or receive
at a given time, on a given frequency, but not both, leading to the
need to describe protocols, or which nodes transmit when. We
consider time division duplex (TDD) two-way system as duplex
scheme and for multiple access both FDMA and TDMA can
be used. For each user, time is devided into forward channel
and reverse channel (as in TDD scheme). During the former
time, Base-station transmits the packet and during the latter user
transmits the packet. The base-station (BS) sends a packet s
to the end-user U. The end-user can extract its own message
after channel estimation using the downlink pilot – labeled X.
Contrary to most of the current works on integrated radar and
communication systems, the data transmission proposed here
is similar to the conventional half-duplex transmission. This
guarantees a high-data rate to efficiently accommodate downlink
traffic. At the same time, the packet s is reflected from objects in
the neighborhood. The base-station observes the echo of its own
transmit signal, and detects the presence of objects and their
distance and relative velocity. With ComSens, the base-station
jointly estimates the radar return and extract the uplink message
from end-user U after channel estimation using uplink pilot –
labeled Y . The principal constraint in the performance of radar
sensing is the simultaneous reception of the radar echo and uplink
packet. Therefore the main goal of this work is to design the
uplink and downlink pilot sequences. We design the two pilot
sequences to be uncorrelated to each other so that they can be
distinguished from each other at the base-station. After separation
of two pilots, the base-station uses the packet with uplink pilot
for communication purposes and the reflected downlink packet
for sensing.
Remark: From the above discussion it must be clear that
ComSens uplink communications may be subject to interference
from the radar echo. Note that the echo signal is received at
the base-station with high attenuation due to the two-way link
(from the base-station to the object and from the object to the
base-station) and the absorption at the object so its impact on
the uplink communication is negligible.
B. Time and Range Analysis
1. Limitations:
Consider one TDD frame for an end-user U (as it is shown
in Fig. 1). At the forward channel time, base-station transmits
the packet. End-user receives the packet at tmod + tU
delay where
tmod is the modulation and transmission time and tU delay is the
propagation time between end-user and the base-station. Packet
is processed at the end-user in tpr time. Then, end-user transmits
the packet in the reverse channel time and base-station receives
it at the time t1 = tmod + 2tU
delay + tpr . On the other hand,
transmitted packet from the base-station is also received at the
Ob
object at the time tmod + tOb
delay where tdelay is the propagation
time between object and the base-station. The packet is then
back-scattered from the object and received at the base-station
at the time t2 = tmod + 2tOb
delay . We design downlink pilot
and uplink pilot to be uncorrelated to each other for k time
lags. Therefore, if two received signals (from user and object)
have arrival time difference (t2 − t1 ) of at most k, they are
distinguishable from each other. On the other hand, if t2 −t1 > k,
the radar signal cannot be recognized and it will be considered as
weak interference for communication system. Consequently, our
proposed integrated radar system will perform when t2 − t1 ≤ k.
Substituting t1 and t2 we have:
Ob
tU
delay − tdelay ≤
tpr + k
2
(1)
3
where
tOb,U
delay =
dOb,U
νTs
(2)
and dU and dOb are respectively the distance of user and the
object from the base-station, Ts is symbol time in our system
and ν is the speed of electromagnetic wave in the space. Using
Eq. (1, 2) we have
dOb ≤ dU +
νT (tpr + k)
2
(3)
2. Practical Scenario: Communication cell towers have a
range between 35km to 72km. We consider our user to be
(as a medium distance) at the distance dU = 25km of the
base-station. Assume that the symbol time Ts = 25µs and
processing time tpr = Ts where speed of electromagnetic wave
is ν = 3 × 108 , assuming we design our pilots to be uncorrelated
for k = 4. Such a system would have a radar range of 43.75km
(dOb ≤ 43.75km).
C. Channel Model
We consider the same settings as in [30] and [31]. More
precisely, we consider a narrowband block fading point-to-point
MIMO channel with nT transmit and nR receive antennas.
Assume that P ∈ CB×nT be a matrix whose rows are the
pilot sequence at each transmitter antenna. At the training phase,
channel can be described as
Y = HP T + N
Fig. 2. The radar operational range of ComSens.
where M SE = tr[θ]. Now let
Q,
R
eR
P
H
e
RP
e RP
eH
M +P
III. P ILOT S EQUENCE D ESIGN
In this section, we design the pilot coefficients gathered in the
matrix P , in order to produce an accurate estimate of the channel
H–while simultaneously satisfying a set of radar performance
criteria. For an accurate channel estimation, one may resort
to a minimization of the channel mean-squared error (MSE),
expressed as [30], [31]
h
−1 i
M SE = tr R−1 + (P ⊗ I nR )H M −1 (P ⊗ I nR )
. (5)
e , P ⊗ In ∈ CBnR ×nT nR , and note that using the matrix
Let P
R
inversion lemma we have
−1
e H M −1 P
e
θ ,
R−1 + P
(6)
−1
eH M + P
e RP
eH
e R,
= R − RP
P
(7)
∈ C(B+nT )nR ×(B+nT )nR ,
(8)
(4)
where Y ∈ CnR ×B is the received sequence, H ∈ CnR ×nT
is the MIMO channel when H(i, j) denotes the MIMO channel gain between ith transmitter and j th receiver and N ∈
CnR ×B is the noise matrix. We assume Gaussian noise i.e.
vec(N ) ∼ CN (0, M ) where M ∈ CBnR ×BnR denotes noise
covariance matrix. We also assume vec(H) ∼ CN (0, R) where
R ∈ CnT nR ×nT nR denotes channel covariance matrix.
!
T
U , (I nT nR 0nT nR ×BnR ) ∈ C
(B+nT )nR ×nT nR
,
(9)
and observe that [32],
U H Q−1 U = θ −1 .
(10)
In light of the above, the authors in [31] propose a cyclic
optimization approach to minimizing the MSE in (5): Consider
an auxiliary variable V ∈ CnT nR ×BnR such that
h
i
F (V , P ) := tr V H QV .
(11)
The minimizer V of (11) can be obtained as [33, p. 354]
!
I nT nR
V∗=
−1
e RP
eH
eR
− M +P
P
(12)
By substituting (12) in (11), one can verify that
F (V ∗ , P ) = tr [θ] = M SE.
(13)
Therefore, in order to optimize the MSE we can use a cyclic
optimization of (11) with respect to V and P . In particular, it
was shown in [31] that the optimization of (11) with respect to
P can be cast at each (cyclic) iteration as:
min
P h+1 ∈Ω
(h)
P (h+1) − P Σ
2
,
2
(14)
4
(h)
where P Σ is constructed from P (h) at each iteration (see [31]
for details). For the two-part pilot employed in ComSens, define:
P DL
P UL
:= X
:= Y
(15)
(16)
where X ∈ CB×nT is the downlink pilot contributing at both
radar and communication modes and Y ∈ CB×nR is the uplink
pilot which contributes only in communication mode. Thus, (14)
becomes
2
2
2
min ||X − X Σ ||2 + ||Y − Y Σ ||2 ,
Xi
(17)
Yi
where the constraint set Ω is yet to be defined. As indicated
earlier, X and Y should have low correlation with each other
and X should have an impulse-like autocorrelation. We describe
the pilot constraints in three categories:
1) Both pilot sequences should have fixed transmit powers
given by
||xq ||22 ≤ p,
||y l ||22 ≤ p,
1 ≤ q ≤ nT
1 ≤ l ≤ nR
(18)
(19)
where xq and y l are column vectors of X and Y and p
is the power upper-bound.
2) To resolve ambiguity between radar reflections and
communication signals, pilot sequences (and their time
lags up to k lags) should be uncorrelated to each other;
i.e their cross correlation must be zero or very small at
least for a number of time lags (forming a zero correlation
zone [34], [35]):
X T J i Y ' 0nT ×nR ,
0 ≤ i ≤ k,
(20)
where J k ∈ C B×B is a shift matrix that shifts a matrix
by k time lags. Clearly J 0 is identity matrix.
3) Radar pilot sequence should be impulse-like; i.e. its
auto-correlation must be zero or very small at least for a
number of time lags:
X T J i X ' 0nT ×nT ,
0 ≤ i ≤ k.
(21)
Consequently, one can solve the following optimization problem to design our pilot sequences:
2
2
min
||X − X Σ ||2 + ||Y − Y Σ ||2
s.t.
||xq ||22 ≤ p, 1 ≤ q ≤ nT ;
||y l ||22 ≤ p, 1 ≤ l ≤ nR ;
xTq J i y l ≤ , 1 ≤ i ≤ k;
X,Y ∈Ω
xTq J i xq ≤ ,
(22)
1 ≤ i ≤ k;
where is a very small number (in this paper we use 10−5 ) to
achieve equality constraints. In order to tackle (22) we can use
cyclic optimization[36]. We define:
(23)
Then one can perform a cyclic procedure to minimize G(X, Y )
as follows: We start with an initial value Y = Y 0 . Then
we comupte X i by tackling minimization problem in Eq. (24)
and Y i by tackling minimization problem in Eq. (25). More
precisely:
s.t.
X,Y ∈Ω
2
G(X, Y ) := ||X − X Σ ||2 + ||Y − Y Σ ||2
s.t.
=
arg min G(X, Y i−1 )
X
||xq ||22 ≤ p,
xTq J m y i−1
= 0,
l
T
xq J m xq ≤ ,
(24)
1 ≤ q ≤ nT ;
1 ≤ m ≤ k;
1 ≤ m ≤ k;
= arg min G(X i , Y )
Y
||y l ||22 ≤ p,
(xiq )T J m y l
(25)
1 ≤ l ≤ nT ;
≤ 0, 1 ≤ m ≤ k;
where 1 ≤ q ≤ nT and 1 ≤ l ≤ nR . Note that since now the
second constraint in both (24) and (25) are affine constraints,
we replaced them with equality. Eq. (25) is now a convex
optimization problem and solvable using convex optimization.
However, the third constraint in (24) is not convex. We can
rewrite Eq. (24) in form:
Xi
s.t.
=
arg min G(X, Y i−1 )
X
||xq ||22 ≤ p,
xTq J m y i−1
=
l
T
T
xq (J m + J m
(26)
0,
1 ≤ q ≤ nT ;
1 ≤ m ≤ k;
+ 2I m )xq ≤ 2p,
1 ≤ m ≤ k;
the third constraint in (26) is now in quadratic convex form
since (J Tm + J m + 2I m ) is a symmetric positive semi-definite
matrix. Note that the optimization problem is still the same (since
xTq J m xq ≤ and J Tm = J −m then xTq J Tm xq ≤ also holds
and from the first constraint xTq I m xq ≤ p). Now we can follow
the steps of the algorithm below to design the pilot sequence.
Algorithm 1 Cyclic Algorithm For Constrained Pilot Sequence
Design
Step 0: Initialize P DL and P U L using a random matrix in Ω.
Step 1: Compute the minimizer V of (11) using (12).
Step 2: Update the current design of X and Y by solving
cyclic optimization problem (26) and (25) µ times (or until
convergence).
Step 3: Repeat steps 1 and 2 until a stop criterion is satisfied,
e.g. M SE (m+1) − M SE (m) < η for some given η > 0, where
m denotes the outer loop iteration.
IV. S IMULATION R ESULTS
In this section, we evaluate the performance of the communication method with respect to the channel MSE metric and
for the radar mode we illustrate the cross and auto-correlation
between two pilot sequences.
5
102
0
Autocorrelation
10
10-2
10-4
10-6
-8
10
-8
-6
-4
-2
0
2
4
6
8
Time lags
A. Simulation Settings
We used the exponential model to generate covariance matrices. This model is particularly appropriate whenever a control
over correlation is required. For a covariance matrix C, we let
[C]k,l = ρl−k for k ≤ l, and [C]k,l = [C]∗l,k for l < k, with
|ρ| < 1 denoting the correlation coefficient. Furthermore, we
assume that both the channel matrix R and the noise matrix M
follow the Kronecker model; i.e for covariance matrix R defined
as R = (RTT ⊗ RR ) we suppose ρrt = 0.9e−jθrt and ρrr =
0.65e−jθrr to construct RT and RR (at the transmit side and the
receive side, respectively) using exponential model. Also, for covariance matrix of noise M defined as M = (MTT ⊗MR ) where
MR = RR , we let ρmt = 0.8e−jθmt to construct MT at the
transmitter side.The phase arguments (θrt , θrr , θmt ) appearing
above were chosen randomly as (0.8349π, 0.4289π, 0.5361π).
We also normalize R and M such that tr{R} = 1 and
tr{M} = 1, and define the pilot sequence-to-noise ratio (SNR)
as SNR , γ, and γ = kP k2F denotes the total training energy.
We consider γ = BnT , and set the stop threshold of the iteration
loop in Algorithm I as η = 10−5 .
B. Channel MSE Metric
We show the performance of the suggested approach for
communication purposes using MSE as the figure of merit. We
consider a 4 × 4 MIMO channel with B = 8. The results are
shown in Fig. 3. For each power, we have used the proposed
method 50 times ,using different initializations, and have reported
the average of the obtained MSE values. It can be observed from
Fig. 3 that the proposed method performs better in each iteration
until it converges to the optimal MSE.
C. Radar Pilot Sequence Specifications
To ensure radar part of the system performs properly, our pilot
sequence for radar part should have very small auto-correlation
for at least a range of time lags so that this pilot sequence have an
Fig. 4. Autocorrelation of radar pilot signal (xT
l J i xl ) where 1 ≤ l ≤ 8, and
each l denotes a transmit antenna so we have totally 4 autocorrelation plots in
this figure, and also −8 ≤ i ≤ 8 denote time lags
10-5
Cross-correlation
Fig. 3. MSE comparison of different schemes for a 4 × 4 MIMO channel where
B = 8. In order to demonstrate the improvement of MSE values through the
iterations, values of MSE at each iterations is shown.
10-10
10-15
10-20
-8
-6
-4
-2
0
2
4
6
8
Time lags
Fig. 5. Cross-correlation of radar and communication pilot signals (xT
q J i yl )
where 1 ≤ q ≤ 8, and each q denotes different transmit antenna, 1 ≤ l ≤ 8, and
each l denotes different receive antenna so we have totally 16 cross-correlation
plots in this figure, and also −8 ≤ i ≤ 8 denote time lags
impulse like shape. In Fig. 4 this auto-correlation is shown. For
each lag, auto-correlation level is shown in dB. Fig. 4 shows that
auto-correlation levels for time lags 2-8 are almost zero compared
to autocorrelation for the first lag. Which gives us the impulselike correlation for the pilot sequence contributing in sensing
mode.
D. Correlation of pilots
The key factor for our system to distinguish between radar
signal and communication signal is that two pilot sequences
should be uncorrelated with each other for a number of time
lags. Fig. 5 shows cross-correlation between two pilot signals
for our simulations in dB. As it is obvious from simulations
results, correlation between these two signals are really small so
they can be assumed uncorrelated.
6
V. C ONCLUSION
The idea of designing pilot sequences for a communication
system to be able to operate also in an integrated radar mode
has been proposed and the protocol and limitations has been
explained. We evaluate the channel MSE for communication
system and show that pilot sequences are uncorrelated, and also
one of the sequences have impulse-like correlation (suitable for
radar sensing). The proposed system can perform as a radar
and communication system. Considering the communication
operation as the primary also lays the ground for making the
radar systems ubiquitous.
R EFERENCES
[1] A. Sani and A. Vosoughi, “Bandwidth and power constrained distributed
vector estimation in wireless sensor networks,” in Military Communications
Conference, MILCOM 2015-2015 IEEE. IEEE, 2015, pp. 1164–1169.
[2] S. Hajizadeh and G. A. Hodtani, “Three-receiver broadcast channels with
side information,” in Information Theory Proceedings (ISIT), 2012 IEEE
International Symposium on. IEEE, 2012, pp. 393–397.
[3] T. X. Tran, M.-P. Hosseini, and D. Pompili, “Mobile edge computing:
Recent efforts and five key research directions,” MMTC CommunicationsFrontiers, vol. 12, no. 4, pp. 29–34, 2017.
[4] M. Karimzadeh Kiskani, B. Azimdoost, and H. Sadjadpour, “Effect of
social groups on the capacity of wireless networks,” IEEE Transactions
on Wireless Communications, vol. 15, no. 1, pp. 3–13, 2016.
[5] M.-P. Hosseini, M. R. Nazem-Zadeh, D. Pompili, K. Jafari, K. Elisevich,
and H. Soltanian-Zadeh, “Support vector machine with nonlinear-kernel
optimization for lateralization of epileptogenic hippocampus in mr images,”
in Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual
International Conference of the IEEE. IEEE, 2014, pp. 1047–1050.
[6] V. Naghshin, M. C. Reed, and N. Aboutorab, “Coverage analysis of
packet multi-tier networks with asynchronous slots,” IEEE Transactions
on Communications, vol. 65, no. 1, pp. 200–215, Jan 2017.
[7] A. Sani and A. Vosoughi, “Distributed vector estimation for power-and
bandwidth-constrained wireless sensor networks,” IEEE Transactions on
Signal Processing, vol. 64, no. 15, pp. 3879–3894, 2016.
[8] S. Vahidian, S. Assa, and S. Hatamnia, “Relay selection for securityconstrained cooperative communication in the presence of eavesdropper’s
overhearing and interference,” IEEE Wireless Communications Letters,
vol. 4, no. 6, pp. 577–580, Dec 2015.
[9] M.-P. Hosseini, T. Tran, D. Pompili, K. Elisevich, and H. Soltanian-Zadeh,
“Deep learning with edge computing for localization of epileptogenicity using multimodal rs-fmri and eeg big data,” in IEEE International Conference
on Autonomic Computing (ICAC). IEEE, 2017.
[10] M. Yazdani, J. K. Lee, J. Mautz, E. Arvas, and K. Sun, “Shortcut waves
in a radially uniaxial sphere,” in 2015 IEEE International Symposium on
Antennas and Propagation USNC/URSI National Radio Science Meeting,
July 2015, pp. 1112–1113.
[11] V. Naghshin, A. M. Rabiei, N. C. Beaulieu, and B. Maham, “Accurate
statistical analysis of a single interference in random networks with
uniformly distributed nodes,” IEEE Communications Letters, vol. 18, no. 2,
pp. 197–200, February 2014.
[12] M. K. Kiskani and H. Sadjadpour, “Throughput analysis of decentralized
coded content caching in cellular networks,” IEEE Transactions on Wireless
Communications, vol. 16, no. 1, pp. 663–672, 2017.
[13] M. H. Moghaddam, M. J. Azizipour, S. Vahidian, and B. Smida, “A
framework for super-resolution of scalable video via sparse reconstruction
of residual frames,” arXiv preprint arXiv:1707.09926, 2017.
[14] M. Yazdani, J. Mautz, L. Murphy, and E. Arvas, “High-frequency scattering
from radially uniaxial dielectric sphere,” IEEE Antennas and Wireless
Propagation Letters, vol. 14, pp. 1577–1581, 2015.
[15] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “A dual function radar-communications system using sidelobe control and waveform
diversity,” in 2015 IEEE Radar Conference (RadarCon), May 2015, pp.
1260–1263.
[16] S. D. Blunt, P. Yatham, and J. Stiles, “Intrapulse radar-embedded communications,” IEEE Transactions on Aerospace and Electronic Systems, vol. 46,
no. 3, pp. 1185–1200, July 2010.
[17] J. Euzire, R. Guinvarc’h, M. Lesturgie, B. Uguen, and R. Gillard, “Dual
function radar communication time-modulated array,” in 2014 International
Radar Conference, Oct 2014, pp. 1–4.
[18] Y. L. Sit, C. Sturm, L. Reichardt, T. Zwick, and W. Wiesbeck, “The
OFDM joint radar-communication system: An overview,” in Proc. Int. Conf.
Advances in Satellite and Space Communications (SPACOMM 2011), 2011,
pp. 69–74.
[19] C. Sturm, T. Zwick, and W. Wiesbeck, “An OFDM system concept
for joint radar and communications operations,” in Vehicular Technology
Conference, 2009. VTC Spring 2009. IEEE 69th, April 2009, pp. 1–5.
[20] Y. L. Sit, L. Reichardt, C. Sturm, and T. Zwick, “Extension of the OFDM
joint radar-communication system for a multipath, multiuser scenario,” in
2011 IEEE RadarCon (RADAR), May 2011, pp. 718–723.
[21] C. Sturm and W. Wiesbeck, “Waveform design and signal processing
aspects for fusion of wireless communications and radar sensing,” Proceedings of the IEEE, vol. 99, no. 7, pp. 1236–1259, July 2011.
[22] M. Mousaei and B. Smida, “Optimizing pilot overhead for ultra-reliable
short-packet transmission,” in 2017 IEEE International Conference on
Communications (ICC), May 2017, pp. 1–5.
[23] B. Hassibi and B. M. Hochwald, “How much training is needed in multipleantenna wireless links?” IEEE Transactions on Information Theory, vol. 49,
no. 4, pp. 951–963, 2003.
[24] S. Hatamnia, S. Vahidian, S. Aı̈ssa, B. Champagne, and M. AhmadianAttari, “Network-coded two-way relaying in spectrum-sharing systems
with quality-of-service requirements,” IEEE Transactions on Vehicular
Technology, vol. 66, no. 2, pp. 1299–1312, 2017.
[25] M. Karimzadeh Kiskani and H. R. Sadjadpour, “Capacity of cellular
networks with femtocache,” in Computer Communications Workshops (INFOCOM WKSHPS), 2016 IEEE Conference on. IEEE, 2016.
[26] S. Vahidian, E. Soleimani-Nasab, S. Assa, and M. Ahmadian-Attari, “Bidirectional af relaying with underlay spectrum sharing in cognitive radio
networks,” IEEE Transactions on Vehicular Technology, vol. 66, no. 3, pp.
2367–2381, March 2017.
[27] M. K. Kiskani and H. R. Sadjadpour, “Multihop caching-aided coded multicasting for the next generation of cellular networks,” IEEE Transactions
on Vehicular Technology, vol. 66, no. 3, pp. 2576–2585, 2017.
[28] V. Naghshin and M. C. Reed, “On capacity and association area characterization in small cell-based multi-tier networks,” IEEE Wireless Communications Letters, vol. 4, no. 5, pp. 505–508, Oct 2015.
[29] S. Hataminia, S. Vahidian, M. Mohammadi, and M. Ahmadian-Attari,
“Performance analysis of two-way decode-and-forward relaying in the
presence of co-channel interferences,” IET Communications, vol. 8, no. 18,
pp. 3349–3356, 2014.
[30] N. Shariati, J. Wang, and M. Bengtsson, “Robust training sequence design
for correlated MIMO channel estimation,” IEEE Transactions on Signal
Processing, vol. 62, no. 1, pp. 107–120, Jan. 2014.
[31] M. Soltanalian, M. M. Naghsh, N. Shariati, P. Stoica, and B. Hassibi,
“Training signal design for correlated massive mimo channel estimation,”
IEEE Transactions on Wireless Communications, vol. 16, no. 2, pp. 1135–
1143, Feb 2017.
[32] K. B. Petersen, M. S. Pedersen et al., The matrix cookbook. Technical
University of Denmark, 2008.
[33] P. Stoica and R. Moses, Spectral Analysis of Signals. New Jersey:
Prentice Hall, 2005. [Online]. Available: http://user.it.uu.se/∼ps/ref.html
[34] P. Z. Fan, N. Suehiro, N. Kuroyanagi, and X. M. Deng, “Class of binary
sequences with zero correlation zone,” Electronics Letters, vol. 35, no. 10,
pp. 777–779, May 1999.
[35] H. Torii, M. Nakamura, and N. Suehiro, “A new class of zero-correlation
zone sequences,” IEEE Transactions on Information Theory, vol. 50, no. 3,
pp. 559–565, March 2004.
[36] P. Stoica and Y. Selen, “Cyclic minimizers, majorization techniques, and the
expectation-maximization algorithm: a refresher,” IEEE Signal Processing
Magazine, vol. 21, no. 1, pp. 112–114, Jan 2004.
| 7cs.IT
|
Which Training Methods for GANs do actually Converge?
Lars Mescheder 1 Andreas Geiger 1 2 Sebastian Nowozin 3
arXiv:1801.04406v2 [cs.LG] 19 Feb 2018
Abstract
Recent work has shown local convergence of
GAN training for absolutely continuous data and
generator distributions. In this paper, we show
that the requirement of absolute continuity is necessary: we describe a simple yet prototypical
counterexample showing that in the more realistic case of distributions that are not absolutely
continuous, unregularized GAN training is not
always convergent. Furthermore, we discuss regularization strategies that were recently proposed
to stabilize GAN training. Our analysis shows
that GAN training with instance noise or zerocentered gradient penalties converges. On the
other hand, we show that Wasserstein-GANs and
WGAN-GP with a finite number of discriminator
updates per generator update do not always converge to the equilibrium point. We discuss these
results, leading us to a new explanation for the
stability problems of GAN training. Based on our
analysis, we extend our convergence results to
more general GANs and prove local convergence
for simplified gradient penalties even if the generator and data distribution lie on lower dimensional
manifolds. We find these penalties to work well in
practice and use them to learn a generative image
model of all 1000 Imagenet classes in a single
GAN with little hyperparameter tuning.
1. Introduction
Generative Adversarial Networks (GANs) (Goodfellow
et al., 2014) are powerful latent variable models that can be
used to learn complex real-world distributions. Especially
for images, GANs have emerged as one of the dominant
approaches for generating new realistically looking samples
after the model has been trained on some dataset.
However, while very powerful, GANs can be hard to train
and in practice it is often observed that gradient descent
1
MPI Tübingen, Germany 2 ETH Zürich, Switzerland
Microsoft Research, Cambridge, UK. Correspondence to: Lars
Mescheder <[email protected]>.
3
Method
unregularized (Goodfellow et al., 2014)
WGAN (Arjovsky et al., 2017)
WGAN-GP (Gulrajani et al., 2017)
DRAGAN (Kodali et al., 2017)
Instance noise (Sønderby et al., 2016)
ConOpt (Mescheder et al., 2017)
Gradient penalties (Roth et al., 2017)
Gradient penalty on real data only
Gradient penalty on fake data only
Local
convergence
(a.c. case)
!
%
%
!
!
!
!
!
!
Local
convergence
(general case)
%
%
%
%
!
!
!
!
!
Table 1. Convergence properties of different GAN training algorithms for general GAN-architectures. Here, we distinguish between the case where both the data and generator distributions are
absolute continuous (a.c.) and the general case where they may lie
on lower dimensional manifolds.
based GAN optimization does not lead to convergence. As
a result, a lot of recent research has focused on finding
better training algorithms (Arjovsky et al., 2017; Gulrajani
et al., 2017; Kodali et al., 2017; Sønderby et al., 2016; Roth
et al., 2017) for GANs as well as gaining better theoretically
understanding of their training dynamics (Arjovsky et al.,
2017; Arjovsky & Bottou, 2017; Mescheder et al., 2017;
Nagarajan & Kolter, 2017; Heusel et al., 2017).
Despite practical advances, the training dynamics of GANs
are still not completely understood. Recently, Mescheder
et al. (2017) and Nagarajan & Kolter (2017) showed that
local convergence and stability properties of GAN training can be analyzed by examining the eigenvalues of the
Jacobian of the the associated gradient vector field: if the
Jacobian has only eigenvalues with negative real-part at the
equilibrium point, GAN training converges locally for small
enough learning rates. On the other hand, if the Jacobian has
eigenvalues on the imaginary axis, it is generally not locally
convergent. Moreover, Mescheder et al. (2017) showed that
if there are eigenvalues close but not on the imaginary axis,
the training algorithm might require intractably small learning rates to achieve convergence. While Mescheder et al.
(2017) observe eigenvalues close to the imaginary axis in
practice, this observation does not answer the question if
eigenvalues close to the imaginary axis are a general phenomenon and if yes, whether they are indeed the root cause
for the training instabilities that people observe in practice.
Which Training Methods for GANs do actually Converge?
A partial answer to this question was given by Nagarajan
& Kolter (2017), who showed that for absolutely continuous data and generator distributions1 all eigenvalues of the
Jacobian have negative real-part. As a result, GANs are locally convergent for small enough learning rates in this case.
However, the assumption of absolute continuity is not true
for common use cases of GANs, where both distributions
lie on lower dimensional manifolds (Sønderby et al., 2016;
Arjovsky & Bottou, 2017).
In this paper we show that this assumption is indeed necessary: by considering a simple yet prototypical example
of GAN training we analytically show that (unregularized)
GAN training is not always locally convergent. We also
discuss how recent techniques for stabilizing GAN training affect local convergence on our example problem. Our
findings show that neither Wasserstein GANs (WGANs) (Arjovsky et al., 2017) nor Wasserstein GANs with Gradient
Penalty (WGAN-GP) (Gulrajani et al., 2017) nor DRAGAN
(Kodali et al., 2017) converge on this simple example for a
fixed number of discriminator updates per generator update.
On the other hand, we show that instance noise (Sønderby
et al., 2016; Arjovsky & Bottou, 2017), zero-centered gradient penalties (Roth et al., 2017) and consensus optimization
(Mescheder et al., 2017) lead to local convergence.
Based on our analysis, we give a new explanation for the
instabilities commonly observed when training GANs based
on discriminator gradients orthogonal to the tangent space
of the data manifold. We also introduce simplified gradient
penalties for which we prove local convergence. We find
that these gradient penalties work well in practice, allowing
us to learn a generative image model of all 1000 Imagenet
classes in a single GAN.
In summary, our contributions are as follows:
• We identify a simple yet prototypical counterexample
showing that (unregularized) gradient descent based
GAN optimization is not always locally convergent
• We discuss if and how recently introduced regularization techniques stabilize training
• We introduce simplified gradient penalties and prove
local convergence for the regularized GAN training
dynamics
All proofs can be found in the supplementary material.
1
Nagarajan & Kolter (2017) also proved local convergence for
a slightly more general family of probability distributions where
the support of the generator is equal to the support of the true
data distribution near the equilibrium point. Alternatively, they
show that their results also hold when the discriminator satisfies
certain (strong) smoothness conditions. However, the conditions
are usually hard to satisfy in practice without prior knowledge
about the support of the true data distribution.
2. Instabilities in GAN training
2.1. Background
GANs are defined by a min-max two-player game between
a discriminative network Dψ (x) and generative network
Gθ (z). While the discriminator tries to distinguish between
real data point and data points produced by the generator,
the generator tries to fool the discriminator. It can be shown
(Goodfellow et al., 2014) that if both the generator and
discriminator are powerful enough to approximate any realvalued function, the unique Nash-equilibrium of this two
player game is given by a generator that produces the true
data distribution and a discriminator which is 0 everywhere
on the data distribution.
Following the notation of Nagarajan & Kolter (2017), the
training objective for the two players can be described by
an objective function of the form
L(θ, ψ) = Ep(z) [f (Dψ (Gθ (z)))]
+ EpD (x) [f (−Dψ (x))] (1)
for some real-valued function f . The common choice
f (t) = − log(1 + exp(−t)) leads to the loss function considered in the original GAN paper (Goodfellow et al., 2014).
For technical reasons we assume that f is continuously differentiable and satisfies f 0 (t) 6= 0 for all t ∈ R.
The goal of the generator is to minimize this loss whereas
the discriminator tries to maximize it. Our goal when training GANs is to find a Nash-equilibrium, i.e. a parameter
assignment (θ∗ , ψ ∗ ) where neither the discriminator nor the
generator can improve their utilities.
GANs are usually trained using Simultaneous or Alternating
Gradient Descent (SimGD and AltGD). Both algorithms
can be described as fixed point algorithms (Mescheder et al.,
2017) that apply some operator Fh (θ, ψ) to the parameter
values (θ, ψ) of the generator and discriminator, respectively.
For example, simultaneous gradient descent corresponds
to the operator Fh (θ, ψ) = I + h v(θ, ψ), where v(θ, ψ)
denotes the gradient vector field
−∇θ L(θ, ψ)
v(θ, ψ) :=
.
(2)
∇ψ L(θ, ψ)
Similarly, alternating gradient descent can be described by
an operator Fh = F2,h ◦ F1,h where F1,h and F2,h perform
an update for the generator and discriminator, respectively.
Recently, it was shown (Mescheder et al., 2017) that local
convergence of GAN training near an equilibrium point
(θ∗ , ψ ∗ ) can be analyzed by looking at the spectrum of the
Jacobian Fh0 (θ∗ , ψ ∗ ) at the equilibrium: if Fh0 (θ∗ , ψ ∗ ) has
eigenvalues with absolute value bigger than 1 , the training
algorithm will generally not converge to (θ∗ , ψ ∗ ). On the
other hand, if all eigenvalues have absolute value smaller
Which Training Methods for GANs do actually Converge?
y
y
Definition 2.1. In the Dirac-GAN, the true (univariate) data
distribution pD is given by pD = δ0 and the generator is
given by pθ = δθ . The discriminator is given given by a
linear function: Dψ (x) = ψ · x.
Dψ (x)
p D = δ0
pθ = δθ
pD = δ0
p θ = δθ
Dψ (x)
x
x
z
(a) t = t0
(b) t = t1
Figure 1. Visualization of the counterexample showing that in the
general case, gradient descent GAN optimization is not convergent:
(a) In the beginning, the discriminator pushes the generator towards
the true data distribution and the discriminator’s slope increases.
(b) When the generator reaches the target distribution, the slope of
the discriminator is largest, pushing the generator away from the
target distribution. This results in oscillatory training dynamics
that never converge.
than 1, the training algorithm will converge to (θ∗ , ψ ∗ ) with
linear rate O(|λmax |k ) where λmax is the eigenvalue of
F 0 (θ∗ , ψ ∗ ) with the biggest absolute value. If all eigenvalues F 0 (θ∗ , ψ ∗ ) are on the unit circle, the algorithm can be
convergent, divergent or neither, but if it is convergent it
will generally converge with a sublinear rate. A similar result (Khalil, 1996) also holds for the (idealized) continuous
system
θ̇(t)
−∇ψ L(θ, ψ)
=
(3)
∇θ L(θ, ψ)
ψ̇(t)
which corresponds to training the GAN with infinitely small
learning rate: if all eigenvalues of the Jacobian v 0 (θ∗ , ψ ∗ )
at a stationary point (θ∗ , ψ ∗ ) have negative real-part, the
continuous system converges locally to (θ∗ , ψ ∗ ) with linear convergence rate. On the other hand, if v 0 (θ∗ , ψ ∗ ) has
eigenvalues with positive real-part, the continuous system
is not locally convergent. If all eigenvalues have zero realpart, it can be convergent, divergent or neither, but if it is
convergent, it will generally converge with a sublinear rate.
For simultaneous gradient descent linear convergence can
be achieved if and only if all eigenvalues of the Jacobian
of the gradient vector field v(θ, ψ) have negative real part
(Mescheder et al., 2017). This situation was also considered
by Nagarajan & Kolter (2017) who examined the asymptotic
case of step sizes h that go to 0 and proved local convergence
for absolutely continuous generator and data distributions
under certain regularity assumptions.
2.2. The Dirac-GAN
Simple experiments, simple theorems are the building
blocks that help us understand more complicated systems.
Ali Rahimi - Test of Time Award speech, NIPS 2017
In this section, we describe a simple yet prototypical counterexample which shows that in the general case, unregularized GAN training is neither locally nor globally convergent.
Note that in the Dirac-GAN, both the generator and the
discriminator have exactly one parameter. This situation
is visualized in Figure 1. In this setup, the GAN training
objective (1) is given by
L(θ, ψ) = f (ψθ) + f (0)
(4)
While using linear discriminators might appear restrictive,
the class of linear discriminators is in fact as powerful as
the class of all real-valued functions for this example: when
we use f (t) = − log(1 + exp(−t)) and we take the supremum over ψ in (4), we obtain (up to scalar and additive
constants) the Jensen-Shannon divergence between pθ and
pD . The same holds true for the Wasserstein-divergence,
when we use f (t) = t and put a Lipschitz constraint on the
discriminator (see Section 3.1).
We show that the training dynamics of GANs lead to divergent behavior in this simple setup.
Lemma 2.2. The unique equilibrium point of the training
objective in (4) is given by θ = ψ = 0. Moreover, the
Jacobian of the gradient vector field at the equilibrium point
has the two eigenvalues ±f 0 (0) i which are both on the
imaginary axis.
We now take a closer look at the training dynamics produced
by various algorithms for training the Dirac-GAN. First, we
consider the (idealized) continuous system in (3): while
Lemma 2.2 shows that the continuous system is generally
not linearly convergent to the equilibrium point, it could
in principle converge with a sublinear convergence rate.
However, this is not the case as the next lemma shows:
Lemma 2.3. The integral curves of the gradient vector field
v(θ, ψ) do not converge to the Nash-equilibrium. More
specifically, every integral curve (θ(t), ψ(t)) of the gradient
vector field v(θ, ψ) satisfies θ(t)2 + ψ(t)2 = const for all
t ∈ [0, ∞).
Note that our results do not contradict the results of Nagarajan & Kolter (2017) and Heusel et al. (2017): our example
violates Assumption IV in Nagarajan & Kolter (2017) that
the support of the generator distribution is equal to the support of the true data distribution near the equilibrium. It also
violates the assumption in Heusel et al. (2017) that the optimal discriminator parameter vector is a continuous function
of the current generator parameters2 . In fact, unless θ = 0,
2
This assumption is usually even violated by WassersteinGANs, as the optimal discriminator parameter vector as a function
of the current generator parameters can have discontinuities near
the Nash-equilibrium. See Section 3.1 for details.
Which Training Methods for GANs do actually Converge?
2.3. Where do instabilities come from?
Our simple example shows that naive gradient based GAN
optimization does not always converge to the equilibrium
point. To get a better understanding of what can go wrong
for more complicated GANs, it is instructive to analyze
these instabilities in depth for this simple example problem.
(a) SimGD
(b) AltGD
Figure 2. Training behavior of the Dirac-GAN. The starting iterate
is marked in red.
there is not even an optimal discriminator parameter vector for the Dirac-GAN. Indeed, we find that two-time scale
updates as suggested by Heusel et al. (2017) do not help convergence towards the Nash-equilibrium (see Figure 16 in the
supplementary material). However, our example seems to
be a prototypical situation for (unregularized) GAN training
which usually deals with distributions that are concentrated
on lower dimensional manifolds (Arjovsky & Bottou, 2017).
We now take a closer look at the discretized system.
Lemma 2.4. For simultaneous gradient descent, the Jacobian of the update operator Fh (θ, ψ) has
p eigenvalues
λ1/2 = 1 ± hf 0 (0)i with absolute values 1 + h2 f 0 (0)2
at the Nash-equilibrium. Independently of the learning rate,
simultaneous gradient descent is therefore not stable near
the equilibrium. Even stronger, for every initial condition
and learning rate h > 0, the norm of the iterates (θk , ψk )
obtained by simultaneous gradient descent is monotonically
increasing.
The behavior of simultaneous gradient descent on our example problem is visualized in Figure 2a.
Similarly, for alternating gradient descent we have
Lemma 2.5. For alternating gradient descent with ng generator and nd discriminator updates, the Jacobian of the
update operator Fh (θ, ψ) has eigenvalues
λ1/2
α2
=1−
±
2
s
α2
1−
2
2
− 1.
(5)
√
with α := ng nd hf 0 (0). For α ≤ 2, all eigenvalues are
hence on the unit circle. Moreover for α > 2, there are
eigenvalues outside the unit circle.
Even though Lemma 2.5 shows that alternating gradient
descent does not converge linearly to the Nash-equilibrium,
it could in principle converge with a sublinear convergence
rate. However, this is very unlikely because – as Lemma 2.3
shows – even the continuous system does not converge. Indeed, we empirically found that alternating gradient descent
oscillates in stable cycles around the equilibrium and shows
no sign of convergence (Figure 2b).
To understand the instabilities, we have to take a closer
look at the oscillatory behavior that GANs exhibit both for
the Dirac-GAN and for more complex systems. An intuitive explanation for the oscillations is given in Figure 1:
when the generator is far from the true data distribution,
the discriminator pushes the generator towards the true data
distribution. At the same time, the discriminator becomes
more certain, which increases the discriminator’s slope (Figure 1a). Now, when the generator reaches the target distribution (Figure 1b), the slope of the discriminator is largest,
pushing the generator away from the target distribution. As
a result, the generator moves away again from the true data
distribution and the discriminator has to change its slope
from positive to negative. After a while, we end up with a
similar situation as in the beginning of training, only on the
other side of the true data distribution. This process repeats
indefinitely and does not converge.
Another way to look at this is to consider the local behavior
of the training algorithm near the Nash-equilibrium. Indeed,
near the Nash-equilibrium, there is nothing that pushes the
discriminator towards having zero slope on the true data
distribution. Even if the generator is initialized exactly on
the target distribution, there is no incentive for the discriminator to move to the equilibrium discriminator. As a result,
training is unstable near the equilibrium point.
This phenomenon of discriminator gradients orthogonal to
the data distribution can also arise for more complex examples: as long as the data distribution is concentrated on a
low dimensional manifold and the class of discriminators
is big enough, there is no incentive for the discriminator to
produce zero gradients orthogonal to the tangent space of
the data manifold and hence converge to the equilibrium
discriminator. Even if the generator produces exactly the
true data distribution, there is no incentive for the discriminator to produce zero gradients orthogonal to the tangent
space. When this happens, the discriminator does not provide useful gradients for the generator orthogonal to the data
distribution and the generator does not converge.
Note that these instabilities can only arise if the true data
distribution is concentrated on a lower dimensional manifold. Indeed, Nagarajan & Kolter (2017) showed that under some suitable assumptions - gradient descent based
GAN optimization is locally convergent for absolutely continuous distributions. Unfortunately, this assumption may
not be satisfied for data distributions like natural images to
Which Training Methods for GANs do actually Converge?
posed in the original GAN paper (Goodfellow et al., 2014)
leads to convergence of the continuous system, albeit with
an extremely slow convergence rate. A more detailed discussion and an analysis of Consensus optimization (Mescheder
et al., 2017) can be found in the supplementary material.
(b) Non-saturating GAN
3.1. Wasserstein GAN
(c) WGAN (nd = 5)
(d) WGAN-GP (nd = 5)
The two-player GAN game can be interpreted as minimizing
a probabilistic divergence between the true data distribution
and the distribution produced by the generator (Nowozin
et al., 2016; Goodfellow et al., 2014). This divergence is
obtained by considering the best-response strategy for the
discriminator, resulting in an objective function that only
contains the generator parameters. Many recent regularization techniques for GANs are based on the observation
(Arjovsky & Bottou, 2017) that this divergence may be discontinuous with respect to the parameters of the generator
or may even take on infinite values if the support of the data
distribution and the generator distribution do not match.
(e) Consensus optimization
(f) Instance noise
(g) Gradient penalty
(h) Gradient penalty (CR)
(a) Standard GAN
Figure 3. Convergence properties of different GAN training algorithms using alternating gradient descent with recommended
number of discriminator updates per generator update (nd = 1
if not noted otherwise). The shaded area in Figure 3c visualizes
the set of forbidden values for the discriminator parameter ψ. The
starting iterate is marked in red.
which GANs are commonly applied (Arjovsky & Bottou,
2017). Moreover, even if the data distribution is absolutely
continuous but concentrated along some lower dimensional
manifold, the eigenvalues of the Jacobian of the gradient
vector field will be very close to the imaginary axis, resulting in a highly ill-conditioned problem. This was observed
by Mescheder et al. (2017) who examined the spectrum
of the Jacobian for a data distribution given by a circular
mixture of Gaussians with small variance.
To make the divergence continuous with respect to the parameters of the generator, Wasserstein GANs (WGANs)
Arjovsky et al. (2017) replace the Jensen-Shannon divergence used in the original derivation of GANs (Goodfellow
et al., 2014) with the Wasserstein-divergence. As a result,
Arjovsky et al. (2017) propose to use f (t) = t and restrict
the class of discriminators to Lipschitz continuous functions
with Lipschitz constant equal to some g0 > 0. While a
WGAN converges if the discriminator is always trained until convergence, in practice WGANs are usually trained by
running only a fixed finite number of discriminator updates
per generator update. However, near the Nash-equilibrium
the optimal discriminator parameters can have a discontinuity as a function of the current generator parameters: in
the Dirac-GAN, the optimal discriminator has to move from
ψ = −1 to ψ = 1 when θ changes signs. As the gradients
get smaller near the equilibrium point, the gradient updates
do not lead to convergence for the discriminator. Overall,
the training dynamics are again determined by the Jacobian
of the gradient vector field near the Nash-equilibrium:
Lemma 3.1. WGANs trained with simultaneous or alternating gradient descent with a fixed number of discriminator
updates per generator update and a fixed learning rate
h > 0 do generally not converge to the Nash equilibrium
for the Dirac-GAN.
3. Regularization strategies
As we have seen in Section 2, unregularized GAN training
does not always converge to the Nash-equilibrium. In this
section, we discuss how several regularization techniques
that have recently been proposed, influence convergence of
the Dirac-GAN.
Interestingly, we also found that the non-saturating loss pro-
The training behavior of the WGAN is visualized in Figure 3c. We stress that this analysis only holds if the discriminator is trained with a fixed number of discriminator updates
(as it is usually done in practice). More careful training that
ensures that the discriminator is kept exactly optimal or
two-timescale training (Heusel et al., 2017) might be able
to ensure convergence for the WGAN.
Which Training Methods for GANs do actually Converge?
y
Dψ (x)
x
(a) Example with instance noise
(b) Eigenvalues
Figure 4. Dirac-GAN with instance noise. While unregularized
GAN training is inherently unstable, instance noise can stabilize it:
(a) Near the Nash-equilibrium, the discriminator is pushed towards
the zero discriminator. (b) As we increase the noise level σ from
0 to σcritical , the real part of the eigenvalues at the equilibrium
point becomes negative and the absolute value of the imaginary
part becomes smaller. For noise levels bigger than σcritical all
eigenvalues are real-valued and GAN training hence behaves like
a normal optimization problem.
The convergence properties of WGANs were also considered by Nagarajan & Kolter (2017) who showed that even
for absolutely continuous densities and infinitesimal learning rates, WGANs are not always locally convergent.
We also found that WGAN-GP (Gulrajani et al., 2017) does
not converge for the Dirac-GAN (Figure 3d). Please see the
supplementary material for details.
3.2. Instance noise
A common technique to stabilize GANs is to add instance
noise (Sønderby et al., 2016; Arjovsky & Bottou, 2017), i.e.
independent Gaussian noise, to the data points. While the
original motivation was to make the probabilistic divergence
between data and generator distribution well-defined for distributions that do not have common support, this does not
clarify the effects of instance noise on the training algorithm
itself and its ability to find a Nash-equilibrium. Interestingly,
however, it was recently shown (Nagarajan & Kolter, 2017)
that in the case of absolutely continuous distributions, gradient descent based GAN optimization is - under suitable
assumptions - locally convergent.
Indeed, when we apply instance noise to the Dirac-GAN we
have:
Lemma 3.2. When using Gaussian instance noise with standard deviation σ, the eigenvalues of the Jacobian of the
gradient vector field are given by
p
λ1/2 = f 00 (0)σ 2 ± f 00 (0)2 σ 4 − f 0 (0)2 .
(6)
In particular, all eigenvalues of the Jacobian have negative
real-part at the Nash-equilibrium if f 00 (0) < 0 and σ > 0.
Hence, simultaneous and alternating gradient descent are
both locally convergent for small enough learning rates.
Interestingly, Lemma 3.2 shows that there is a critical noise
2
level given by σcritical
= |f 0 (0)|/|f 00 (0)|. If the noise level
is smaller than the critical noise level, the eigenvalues of
the Jacobian have non-zero imaginary part which results
in a rotational component in the gradient vector field near
the equilibrium point. If the noise level is larger than the
critical noise level, all eigenvalues of the Jacobian become
real-valued and the rotational component in the gradient
vector field disappears. The optimization problem is best
behaved when we select σ = σcritical : in this case we can
even achieve quadratic convergence for h = |f 0 (0)|−1 . The
effect of instance noise on the eigenvalues is visualized in
Figure 4b, which shows the traces of the two eigenvalues as
we increase σ from 0 to 2σcritical .
Figure 3f shows the training behavior of the GAN with
instance noise, showing that instance noise indeed creates a
strong radial component in the gradient vector field which
makes the training algorithm converge.
3.3. Zero-centered gradient penalties
Motivated by the success of instance noise to make the f divergence between two distributions well-defined, Roth
et al. (2017) derived a local approximation to instance noise
that results in a zero-centered3 gradient penalty for the discriminator.
In our simple example, a penalty on the squared norm of the
gradients of the discriminator (no matter where) results in
the regularizer
γ
R(ψ) = ψ 2 .
(7)
2
This regularizer does not include the weighting terms considered by Roth et al. (2017). However, the same analysis
can also be applied to the regularizer with the additional
weighting, yielding almost exactly the same results (see
Section D.2 of the supplementary material).
Lemma 3.3. The eigenvalues of the Jacobian of the gradient vector field for the gradient-regularized GAN at the
equilibrium point are given by
r
γ
γ2
λ1/2 = − ±
− f 0 (0)2 .
(8)
2
4
In particular, for γ > 0 all eigenvalues have negative real
part. Hence, simultaneous and alternating gradient descent
are both locally convergent for small enough learning rates.
As for instance noise, there is a critical regularization parameter γcritical = 2|f 0 (0)| that results in a locally rotation
free vector field. A visualization of the training behavior of
the Dirac-GAN with gradient penalty is shown in Figure 3g.
Figure 3h illustrates the training behavior of the GAN with
gradient penalty and critical regularization (CR). In particular, we see that near the Nash-equilibrium the vector field
3
In contrast to the gradient regularizers used in WGAN-GP
and DRAGAN which are not zero-centered.
Which Training Methods for GANs do actually Converge?
does not have a rotational component anymore and hence
behaves like a normal optimization problem.
4. General convergence results
In Section 3 we analyzed the convergence properties of
various regularization strategies on the Dirac-GAN. In this
section, we consider general GAN problems. First, we introduce two simplified versions of the zero-centered gradient
penalty proposed by Roth et al. (2017). We then show that
these gradient penalties allow us to extend the convergence
proof by Nagarajan & Kolter (2017) to the case where the
generator and data distribution do not locally have the same
support.4 As a result, our convergence proof for the regularized training dynamics also holds for the more realistic case
where both the generator and data distributions may lie on
lower dimensional manifolds.
4.1. Simplified gradient penalties
Our analysis suggests that the main effect of the zerocentered gradient penalties proposed by Roth et al. (2017)
on local stability is to penalize the discriminator for deviating from the Nash-equilibrium. The simplest way to achieve
this is to penalize the gradient on real data alone: when the
generator distribution produces the true data distribution
and the discriminator is equal to 0 on the data manifold, the
gradient penalty ensures that the discriminator cannot create
a non-zero gradient orthogonal to the data manifold without
suffering a loss in the GAN game.
This leads to the following regularization term:
γ
(9)
R1 (ψ) := EpD (x) k∇Dψ (x)k2 .
2
Note that this regularizer is a simplified version of to the
regularizer derived by Roth et al. (2017). However, our
regularizer does not contain the additional weighting terms
and penalizes the discriminator gradients only on the true
data distribution.
We also consider a similar regularization term given by
γ
(10)
R2 (θ, ψ) := Epθ (x) k∇Dψ (x)k2
2
where we penalize the discriminator gradients on the current
generator distribution instead of the true data distribution.
Note that on the Dirac-GAN from Section 2, both regularizers reduce to the gradient penalty from Section 3.3 whose
behavior is visualized in Figure 3g and Figure 3h.
4.2. Convergence
In this section we present convergence results for the regularized GAN-training dynamics for both regularization terms
4
Assumption IV in Nagarajan & Kolter (2017)
R1 (ψ) and R2 (ψ) under some suitable assumptions.5
Let (θ∗ , ψ ∗ ) denote an equilibrium point of the regularized
training dynamics. In our convergence analysis, we consider
the realizable case, i.e. we assume that there are generator
parameters that make the generator produce the true data
distribution:
Assumption I. We have pθ∗ = pD and Dψ∗ (x) = 0 in
some local neighborhood of supp pD .
Like Nagarajan & Kolter (2017), we assume that f satisfies
the following property:
Assumption II. We have f 0 (0) 6= 0 and f 00 (0) < 0.
An extension of our convergence proof for f (t) = t (as in
WGANs) can be found in the supplementary material.
The convergence proof is complicated by the fact that for
neural networks, there generally is not a single equilibrium
point (θ∗ , ψ ∗ ), but a submanifold of equivalent equilibria
corresponding to different parameterizations of the same
function. We therefore define the reparameterization manifolds MG and MD . To this end, let
h(ψ) := EpD (x) |Dψ (x)|2 + k∇x Dψ (x)k2 .
(11)
The reparameterization manifolds are then defined as
MG := {θ | pθ = pD }
MD := {ψ | h(ψ) = 0}. (12)
To prove local convergence, we have to assume some regularity properties for MG and MD near the equilibrium
point. To state these assumptions, we need
g(θ) := Epθ (x) [∇ψ Dψ (x)|ψ=ψ∗ ] .
(13)
Assumption III. There are -balls B (θ∗ ) and B (ψ ∗ )
around θ∗ and ψ ∗ so that MG ∩ B (θ∗ ) and MD ∩ B (ψ ∗ )
define C 1 - manifolds. Moreover, the following holds:
(i) if v ∈ Rn is not in the tangent space of MD at ψ ∗ ,
then ∂v2 h(ψ ∗ ) 6= 0.
(ii) if w ∈ Rm is not in the tangent space of MG at θ∗ ,
then ∂w g(θ∗ ) 6= 0.
While formally similar, the two conditions in Assumption III
have very different meanings: the first condition is a simple
regularity property that means that the geometry of MD
can be locally described by the second derivative of h. The
second condition implies that the discriminator is strong
enough so that it can detect any deviation from the equilibrium generator distribution. Indeed, this is the only point
where we assume that the class of representable discriminators is sufficiently expressive (and excludes, for example,
the trivial case Dψ = 0 for all ψ).
5
Our results also hold for any convex combination of R1 and
R2 and the regularizer with the additional weighting terms derived
by Roth et al. (2017). See the supplementary material for details.
Which Training Methods for GANs do actually Converge?
We are now ready to state our main convergence result. To
this end, consider the regularized gradient vector field
−∇θ L(θ, ψ)
ṽi (θ, ψ) :=
.
(14)
∇ψ L(θ, ψ) − ∇ψ Ri (θ, ψ)
Theorem 4.1. Assume Assumption I, II and III hold for
(θ∗ , ψ ∗ ). For small enough learning rates, simultaneous
and alternating gradient descent for ṽ1 and ṽ2 are both
convergent to MG × MD in a neighborhood of (θ∗ , ψ ∗ ).
Moreover, the rate of convergence is at least linear.
Theorem 4.1 shows that GAN training with our gradient
penalties is convergent when initialized sufficiently close
the equilibrium point. While this does not show that the
method is globally convergent, it at least shows that near the
equilibrium the method is well-behaved.
4.3. Stable equilibria for unregularized GAN training
As we have seen in Section 2, unregularized GAN training
does not always converge to the Nash-equilibrium. However,
this does not rule out the existence of stable equilibria for
every GAN architecture. In Section E of the supplementary
material, we identify two forms of stable equilibria that may
exist for unregularized GAN training (Energy Solutions and
Full-Rank Solutions). However, such solutions require much
more expressive discriminators and it is not yet clear under
what conditions such solutions exist for high dimensional
data distributions.
5. Experiments
2D-Problems Measuring convergence for GANs is hard
for high dimensional problems, because we lack an evaluation metric that can reliably detect non-convergent behavior.
We therefore first examine the behavior of the different regularizers on simple 2D examples where we can assess convergence using an estimate of the Wasserstein-1-distance.
To this end, we run 5 different training algorithms on 4 different 2D-examples for 6 different GAN architectures. For
each method, we try both stochastic gradient descent and
RMS-Prop with 4 different learning rates. For the R1 , R2
and WGAN-GP regularizers we try 3 different regularization parameters. For each method and architecture, we pick
the best hyperparameter setting. We train all method for 50k
iterations. Please see the supplementary material for details.
The results are shown in Figure 5. We see that the R1 and
R2 regularizers perform similarly and they achieve slightly
better results than unregularized training or training with
WGAN-GP. In the supplementary material we show that the
R1 and R2 regularizers find solutions where the discriminator is 0 in a neighborhood of the true data distribution,
whereas unregularized training and WGAN-GP converge
(a) 2D Gaussian
(b) Line segment
(c) Circle
(d) Four line segments
Figure 5. Wasserstein-1-distance to true data distribution for 4 different 2D-data-distributions, 6 different architectures (small bars)
and 5 different training methods. Here, we abbreviate WGANGP with 1 and 5 discriminator update(s) per generator update as
WGP-1 and WGP-5.
to energy solutions which we define in Section E.1 of the
supplementary material.
Imagenet To test how well the gradient penalties from
Section 4.1 perform on more complicated tasks, we train
a convolutional GAN consisting of ResNet-architectures
(He et al., 2016b) with 55 layers for both the generator and
discriminator on the ILSVRC dataset (Russakovsky et al.,
2015). While we find that unregularized GAN training
quickly leads to mode-collapse on this architecture, our simple R1 -regularizer enables stable training. Some samples
from the model after 35 epochs of training and more details
on the experimental setup can be found in the supplementary
material.
6. Conclusion
In this paper, we analyzed the stability of GAN training on
a simple yet prototypical example. Due to the simplicity of
the example, we were able to analyze the convergence properties of the training dynamics analytically and we showed
that (unregularized) gradient based GAN optimization is
not always locally convergent. Our findings also show that
WGANs and WGAN-GP do not always lead to local convergence whereas instance noise and zero-centered gradient
penalties do. Based on our analysis, we extended our results
to more general GANs and we proved local convergence for
simplified zero-centered gradient penalties under suitable
assumptions. In the future, we would like to extend our
theory to the non-realizable case and examine the effect of
finite sampling sizes on the GAN training dynamics.
Which Training Methods for GANs do actually Converge?
References
Abadi, Martín, Barham, Paul, Chen, Jianmin, Chen, Zhifeng,
Davis, Andy, Dean, Jeffrey, Devin, Matthieu, Ghemawat,
Sanjay, Irving, Geoffrey, Isard, Michael, Kudlur, Manjunath, Levenberg, Josh, Monga, Rajat, Moore, Sherry,
Murray, Derek Gordon, Steiner, Benoit, Tucker, Paul A.,
Vasudevan, Vijay, Warden, Pete, Wicke, Martin, Yu, Yuan,
and Zheng, Xiaoqiang. Tensorflow: A system for largescale machine learning. In 12th USENIX Symposium on
Operating Systems Design and Implementation, OSDI
2016, Savannah, GA, USA, November 2-4, 2016., pp.
265–283, 2016.
Arjovsky, Martín and Bottou, Léon. Towards principled
methods for training generative adversarial networks.
CoRR, abs/1701.04862, 2017.
Arjovsky, Martín, Chintala, Soumith, and Bottou, Léon.
Wasserstein GAN. CoRR, abs/1701.07875, 2017.
Barratt, Shane and Sharma, Rishi. A note on the inception
score. arXiv preprint arXiv:1801.01973, 2018.
Berthelot, David, Schumm, Tom, and Metz, Luke. BEGAN:
boundary equilibrium generative adversarial networks.
CoRR, abs/1703.10717, 2017.
Bertsekas, Dimitri P. Nonlinear programming. Athena
scientific Belmont, 1999.
Goodfellow, Ian J., Pouget-Abadie, Jean, Mirza, Mehdi,
Xu, Bing, Warde-Farley, David, Ozair, Sherjil, Courville,
Aaron C., and Bengio, Yoshua. Generative adversarial
nets. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal,
Quebec, Canada, pp. 2672–2680, 2014.
Gulrajani, Ishaan, Ahmed, Faruk, Arjovsky, Martín, Dumoulin, Vincent, and Courville, Aaron C. Improved
training of wasserstein gans. In Advances in Neural Information Processing Systems 30: Annual Conference
on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp. 5769–5779,
2017.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Identity mappings in deep residual networks. In
Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14,
2016, Proceedings, Part IV, pp. 630–645, 2016a. doi:
10.1007/978-3-319-46493-0_38.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 770–778, 2016b.
Heusel, Martin, Ramsauer, Hubert, Unterthiner, Thomas,
Nessler, Bernhard, and Hochreiter, Sepp. Gans trained
by a two time-scale update rule converge to a local nash
equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long
Beach, CA, USA, pp. 6629–6640, 2017.
Khalil, Hassan K. Nonlinear systems. Prentice-Hall, New
Jersey, 2(5):5–1, 1996.
Kodali, Naveen, Abernethy, Jacob D., Hays, James, and
Kira, Zsolt. How to train your DRAGAN. CoRR,
abs/1705.07215, 2017.
Krizhevsky, Alex and Hinton, Geoffrey. Learning multiple
layers of features from tiny images. 2009.
Liu, Ziwei, Luo, Ping, Wang, Xiaogang, and Tang, Xiaoou.
Deep learning face attributes in the wild. In Proceedings
of International Conference on Computer Vision (ICCV),
2015.
Mescheder, Lars M., Nowozin, Sebastian, and Geiger, Andreas. The numerics of gans. In Advances in Neural
Information Processing Systems 30: Annual Conference
on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp. 1823–1833,
2017.
Nagarajan, Vaishnavh and Kolter, J. Zico. Gradient descent GAN optimization is locally stable. In Advances
in Neural Information Processing Systems 30: Annual
Conference on Neural Information Processing Systems
2017, 4-9 December 2017, Long Beach, CA, USA, pp.
5591–5600, 2017.
Nowozin, Sebastian, Cseke, Botond, and Tomioka, Ryota.
f-gan: Training generative neural samplers using variational divergence minimization. In Advances in Neural
Information Processing Systems 29: Annual Conference
on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 271–279, 2016.
Paszke, Adam, Gross, Sam, Chintala, Soumith, Chanan,
Gregory, Yang, Edward, DeVito, Zachary, Lin, Zeming,
Desmaison, Alban, Antiga, Luca, and Lerer, Adam. Automatic differentiation in pytorch. 2017.
Radford, Alec, Metz, Luke, and Chintala, Soumith. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint
arXiv:1511.06434, 2015.
Roth, Kevin, Lucchi, Aurélien, Nowozin, Sebastian, and
Hofmann, Thomas. Stabilizing training of generative
adversarial networks through regularization. In Advances
in Neural Information Processing Systems 30: Annual
Which Training Methods for GANs do actually Converge?
Conference on Neural Information Processing Systems
2017, 4-9 December 2017, Long Beach, CA, USA, pp.
2015–2025, 2017.
Russakovsky, Olga, Deng, Jia, Su, Hao, Krause, Jonathan,
Satheesh, Sanjeev, Ma, Sean, Huang, Zhiheng, Karpathy, Andrej, Khosla, Aditya, Bernstein, Michael, Berg,
Alexander C., and Fei-Fei, Li. ImageNet Large Scale
Visual Recognition Challenge. International Journal of
Computer Vision (IJCV), 115(3):211–252, 2015. doi:
10.1007/s11263-015-0816-y.
Salimans, Tim, Goodfellow, Ian, Zaremba, Wojciech, Cheung, Vicki, Radford, Alec, Chen, Xi, and Chen, Xi.
Improved techniques for training gans. In Lee, D. D.,
Sugiyama, M., Luxburg, U. V., Guyon, I., and Garnett, R.
(eds.), Advances in Neural Information Processing Systems 29, pp. 2234–2242. Curran Associates, Inc., 2016a.
Salimans, Tim, Goodfellow, Ian J., Zaremba, Wojciech,
Cheung, Vicki, Radford, Alec, and Chen, Xi. Improved
techniques for training gans. In Advances in Neural Information Processing Systems 29: Annual Conference on
Neural Information Processing Systems 2016, December
5-10, 2016, Barcelona, Spain, pp. 2226–2234, 2016b.
Sønderby, Casper Kaae, Caballero, Jose, Theis, Lucas, Shi,
Wenzhe, and Huszár, Ferenc. Amortised MAP inference
for image super-resolution. CoRR, abs/1610.04490, 2016.
Tieleman, Tijmen and Hinton, Geoffrey. Lecture 6.5rmsprop: Divide the gradient by a running average of
its recent magnitude, 2012.
Zhao, Junbo Jake, Mathieu, Michaël, and LeCun, Yann.
Energy-based generative adversarial network. CoRR,
abs/1609.03126, 2016.
Which Training Methods for GANs do actually Converge?
Supplementary material
A. Preliminaries
In this section we first summarize some results from the
theory of discrete dynamical systems. We also prove a discrete version of a basic convergence theorem for continuous
dynamical systems from Nagarajan & Kolter (2017) which
allows us to make statements about training algorithms for
GANs for finite learning rates. Afterwards, we summarize
some results from Mescheder et al. (2017) about the convergence properties of simultaneous and alternating gradient
descent. Moreover, we state some eigenvalue bounds that
were derived by Nagarajan & Kolter (2017) which we need
to prove Theorem 4.1 on the convergence of the regularized
GAN training dynamics.
A.1. Discrete dynamical systems
In this section, we recall some basic definitions from the theory of discrete nonlinear dynamical systems. For a similar
description of the theory of continuous nonlinear dynamical
systems see for example Khalil (1996) and Nagarajan &
Kolter (2017).
In this paper, we consider continuously differentiable operators F : Ω → Ω acting on an open set Ω ⊂ Rn . A fixed
point of F is a point x̄ ∈ Ω such that F (x̄) = x̄. We are
interested in stability and convergence of the fixed point iteration F (k) (x) near the fixed point. To this end, we first have
to define what we mean by stability and local convergence:
Definition A.1. Let x̄ ∈ Ω be a fixed point of a continuously
differentiable operator F : Ω → Ω. We call x̄
• stable if for every > 0 there is δ > 0 such that
kx − x̄k < δ implies kF (k) (x) − x̄k < for all k ∈ N.
• asymptotically stable if it is stable and there is δ > 0
such that kx − x̄k < δ implies that F (k) (x) converges
to x̄
• exponentially stable if there is λ ∈ [0, 1), δ > 0 and
C > 0 such that kx − x̄k < δ implies
kF (k) (x) − x̄k < Ckx − x̄kλk
responding algorithm linearly convergent. Moreover, if x̄
is exponentially stable, we call the infimum of all λ so that
(15) holds for some C > 0 the convergence rate of the fixed
point iteration.
As it turns out, local convergence of fixed point iterations
can be analyzed by examining the spectrum of the Jacobian
of the fixed point operator. We have the following central
Theorem:
Theorem A.2. Let F : Ω → Ω be a C 1 -mapping on an
open subset Ω of Rn and x̄ ∈ Ω be a fixed point of F .
Assume that the absolute values of the eigenvalues of the
Jacobian F 0 (x̄) are all smaller than 1. Then the fixed point
iteration F (k) (x) is locally convergent to x̄. Moreover, the
rate of convergence is at least linear with convergence rate
|λmax | where λmax denotes the eigenvalue of F 0 (x̄) with
the largest absolute value.
Proof. See Bertsekas (1999), Proposition 4.4.1.
For the proof of Theorem 4.1 in Section D, we need a generalization of Theorem A.2 that takes into account submanifolds of fixed points. The next theorem is a discrete version
of Theorem A.4 from Nagarajan & Kolter (2017) and we
prove it in a similar way:
Theorem A.3. Let F (α, γ) define a C 1 -mapping that maps
some domain Ω to itself. Assume that there is a local neighborhood U of 0 such that F (0, γ) = (0, γ) for γ ∈ U . Moreover, assume that all eigenvalues of J := ∇α F (α, 0) |α=0
have absolute value smaller than 1. Then the fixed point iteration defined by F is locally convergent to M := {(0, γ) |
γ ∈ U } with linear convergence rate in a neighborhood of
(0, 0). Moreover, the convergence rate is |λmax | with λmax
the eigenvalue of J with largest absolute value.
Proof. In the following, we write F (α, γ)
=
(F1 (α, γ), F2 (α, γ)), so that the fixed point iteration
can be written as
(15)
for all k ∈ N.
If x̄ is asymptotically stable fixed point of F , we call the
algorithm obtained by iteratively applying F locally convergent to x̄. If x̄ is exponentially stable, we call the cor-
αk+1 = F1 (αk , γk ) γk+1 = F2 (αk , γk ).
(16)
We first examine the behavior of F1 near (0, 0). To this end,
we develop F1 into a Taylor-Series
F1 (α, γ) = Jα + g1 (α, γ)
(17)
Which Training Methods for GANs do actually Converge?
We first show that for any c > 0 we have kg1 (α, γ)k ≤
ckαk sufficiently close to (0, 0): because F1 (0, γ) = 0 for
all γ close to 0, g1 (α, γ) must be of the form g1 (α, γ) =
h1 (α, γ)α with h1 (0, 0) = 0. This shows that for any c > 0
there is indeed an open neighborhood V of (0, 0) so that
|g1 (α, γ)| ≤ ckαk for all (α, γ) ∈ V .
According to Bertsekas (1999), Proposition A 15, we can
select for every > 0 a norm k · kQ on Rn such that
kJαkQ < (|λmax | + )kαkQ
(18)
for α ∈ Rn where |λmax | denotes the eigenvalue of J with
the largest absolute value.
Hence, for (α, γ) ∈ V ,
kF1 (α, γ)kQ ≤ kJαkQ + kg1 (α, γ)kQ
< (|λmax | + + c)kαkQ
(19)
Because we can make c + as small as we want, this shows
that kαk k ≤ Cλk kα0 k for some C > 0 and λ ∈ [0, 1), if
α0 and all γl for l = 0, . . . , k − 1 are sufficiently close to 0 .
We therefore have to show that the iterates γk stay in a given
local neighborhood of 0, i.e. kγk k ≤ d for some d > 0,
when α0 and γ0 are initialized sufficiently close to 0.
To show this, we develop F2 into a Taylor-series around 0:
F2 (α, γ) = γ + g2 (α, γ).
(20)
Again, we see that g2 must be of the form g2 (α, γ) =
h2 (α, γ)α, showing that kg2 (α, γ)k ≤ c0 kαkQ for some
fixed constant c0 > 0 (note that in general h2 (0, 0) 6= 0).
We therefore have
where L(θ, ψ) is the GAN training objective defined in (1).
Similarly, alternating gradient descent can be described by
an update operator of the form Fh = F2,h ◦ F1,h where F1,h
and F2,h are given by
θ − h∇θ L(θ, ψ)
F1,h (θ, ψ) =
(23)
ψ
θ
F2,h (θ, ψ) =
.
(24)
ψ + h∇ψ L(θ, ψ)
Moreover, we defined the gradient vector field
−∇θ L(θ, ψ)
v(θ, ψ) =
.
∇ψ L(θ, ψ)
(25)
To understand convergence of simultaneous and alternating
gradient descent, we have to understand when the Jacobian
of the corresponding update operator has only eigenvalues
with absolute value smaller than 1.
Lemma A.4. The eigenvalues of the Jacobian of the update
operator for simultaneous gradient descent are given by
λ = 1 + hµ with µ the eigenvalues of v 0 (θ∗ , ψ ∗ ). Assume
that v 0 (θ∗ , ψ ∗ ) has only eigenvalues with negative real part.
The eigenvalues of the Jacobian of the update operator Fh
for simultaneous gradient descent are then all in the unit
circle if and only if
h<
1
| Re(λ)|
2
1+
Im(λ)
Re(λ)
2
(26)
for all eigenvalues λ of v 0 (θ∗ , ψ ∗ ).
Proof. For simultaneous gradient descent we have
kγk − γ0 k =
k−1
X
kg2 (αl , γl )k ≤
l=0
≤
k−1
X
l=0
k−1
X
c0 kαl kQ
Fh (θ, ψ) = (θ, ψ) + hv(θ, ψ)
l=0
Cc0
Cc λ kα0 kQ ≤
kα0 kQ
1−λ
0 l
(21)
(27)
and hence Fh0 (θ∗ , ψ ∗ ) = I + hv 0 (θ∗ , ψ ∗ ). Therefore the
eigenvalues are given by λ = 1 + hµ with µ the eigenvalues
of v 0 (θ∗ , ψ ∗ ).
1−λ
Hence, if we initialize α0 within kα0 kQ ≤ 2CC
0 d and
d
γ0 within kγ0 k ≤ 2 , we have kγk k ≤ d for all k ∈ N,
concluding the proof.
To see when |λ| < 1, we write µ = −a + ib with a, b ∈ R
and a > 0. Then
A.2. Simultaneous and Alternating Gradient Descent
which is smaller than 1 if and only if
In this section, we recall some results by Mescheder et al.
(2017) about the convergence properties of simultaneous
and alternating gradient descent as algorithms for training
generative adversarial networks.
Recall that simultaneous gradient descent can be described
by an update operator of the form
θ − h∇θ L(θ, ψ)
Fh (θ, ψ) =
(22)
ψ + h∇ψ L(θ, ψ)
|λ|2 = (1 − ha)2 + h2 b2
h<
2a
.
a2 + b2
(28)
(29)
Dividing both the numerator and denominator by a2 shows
the assertion.
Lemma A.5. Assume that v 0 (θ∗ , ψ ∗ ) has only eigenvalues with negative real part. For h > 0 small enough, the
eigenvalues of the Jacobian of the update operator Fh for
alternating gradient descent are then all in the unit circle.
Which Training Methods for GANs do actually Converge?
Proof. The Jacobian of the update operator Fh = Fh,2 ◦Fh,1
at an equilibrium is
0
0
Fh0 (θ∗ , ψ ∗ ) = Fh,2
(θ∗ , ψ ∗ ) · Fh,1
(θ∗ , ψ ∗ ).
(30)
However, we have
0
Fh,i
(θ∗ , ψ ∗ ) = I + hvi0 (θ∗ , ψ ∗ )
(31)
for i ∈ {1, 2} where
−∇θ L(θ, ψ)
0
0
v2 (θ, ψ) =
∇ψ L(θ, ψ)
v1 (θ, ψ) =
(32)
(33)
denote the components of the gradient vector field. Hence
Fh0 (θ∗ , ψ ∗ ) = I + h(v10 (θ∗ , ψ ∗ ) + v20 (θ∗ , ψ ∗ ))
+ h2 v20 (θ∗ , ψ ∗ )v10 (θ∗ , ψ ∗ )
= I + hv 0 (θ∗ , ψ ∗ ) + O(h2 ). (34)
The eigenvalues λ of Fh0 (θ∗ , ψ ∗ ) are hence given by λ = 1+
hµ + O(h2 ) where µ denote the eigenvalues of v 0 (θ∗ , ψ ∗ ).
In particular, if the real part of µ is negative and h > 0 small
enough, 1 + hµ will be in the unit circle. Because the error
is in O(h2 ), this also holds for λ for h small enough.
In the proof of Theorem 4.1 we will use local coordinates,
i.e. a diffeomorphism φ that maps a local neighborhood of
(θ∗ , ψ ∗ ) to an open subset of Rn+m . The vector field v and
the update operator F then have the following representation
in the local coordinates:
Fhφ (α) := φ ◦ Fh ◦ φ−1 (α)
(35)
v φ (α) = φ0 (θ, ψ) · (v ◦ φ(α))
(36)
While in local coordinates, the simple relationships between Fhφ (α) and v φ (α) needed to prove Lemma A.4 and
Lemma A.5 do not hold anymore, the spectrum can be described in the same way:
Remark A.6. Assume (θ∗ , ψ ∗ ) is a fixed point of Fh and a
stationary point of v. Let α∗ = φ(θ∗ , ψ ∗ ). Then
(Fhφ )0 (α∗ ) = φ0 (θ∗ , ψ ∗ )Fh0 (θ∗ , ψ ∗ )φ0 (θ∗ , ψ ∗ )−1
(37)
(vhφ )0 (α∗ )
(38)
0
∗
= φ (θ , ψ
∗
)vh0 (θ∗ , ψ ∗ )φ0 (θ∗ , ψ ∗ )−1 .
Hence, (Fhφ )0 (α∗ ) and Fh0 (θ∗ , ψ ∗ ) have the same spectrum.
The same also holds for (v φ )0 (α∗ ) and v 0 (θ∗ , ψ ∗ ).
Proof. This follows from the chain and product rules by
using the fact that Fh (θ∗ , ψ ∗ ) = (θ∗ , ψ ∗ ) and v(θ∗ , ψ ∗ ) =
0.
As we will see in the proof of Theorem 4.1, this allows us to
apply Theorem A.3 to situations where the stationary points
lie on a lower dimensional manifold instead of a space of
the form {0}k × Rn+m−k .
A.3. Eigenvalue bounds
When analyzing the convergence properties of GANs, we
have to analyze the spectrum of real-valued matrices of the
from
0 −B T
(39)
B −Q
with Q symmetric positive definite. To this end, we need
the following important theorem from Nagarajan & Kolter
(2017) which gives explicit bounds on the real part of the
eigenvalues:
Theorem A.7. Assume J ∈ R(n+m)×(n+m) is of the following form:
0 −B T
J=
(40)
B −Q
where Q ∈ Rm×m is a symmetric positive definite matrix
and B ∈ Rm×n has full column rank. Then all eigenvalues
λ of J satisfy Re(λ) < 0. More precisely
• if Im(λ) = 0
Re(λ) ≤ −
λmin (Q)λmin (B T B)
(41)
λmax (Q)λmin (Q) + λmin (B T B)
• if Im(λ) 6= 0
Re(λ) ≤ −
λmin (Q)
2
(42)
Proof. See Nagarajan & Kolter (2017), Lemma G.2.
In Section E.1, we need a generalization of Theorem A.7.
Using almost exactly the same proof as for Theorem A.7,
we obtain
Theorem A.8. Assume J ∈ R(n+m)×(n+m) is of the following form:
−P −B T
J=
(43)
B
−Q
where P ∈ Rn×n is a symmetric positive semi-definite
matrix, Q ∈ Rm×m is a symmetric positive definite matrix
and B ∈ Rm×n has full column rank. Then all eigenvalues
λ of J satisfy Re(λ) < 0.
Proof. Let v T = (aT , bT ) denote some eigenvector of
J with corresponding eigenvalues λ = λr + iλi , where
λr , λi ∈ R. Then
λr =
1 T
v̄ (J + J T )v = −āT P a − b̄T Qb.
2
(44)
Because both P and Q are positive semi-definite, we have
λr ≤ 0. Because Q is positive definite, it suffices to show
that b 6= 0 to prove λr < 0.
Assume that b = 0. Because v is an eigenvector of J, we
have Ba − Qb = λb and therefore Ba = 0. Because B
Which Training Methods for GANs do actually Converge?
has full-column rank, this shows a = 0 and hence v = 0.
However, this contradicts the fact that v is an eigenvector of
J. All in all, this show that b 6= 0 and thus λr ≤ −b̄T Qb <
0 as required.
Proof. The loss in (4) can be rewritten as
L(θ, ψ) = f (θψ) + const
(51)
It is easy to check that the gradient vector field is given by
For applying Theorems A.2, we have to show that the Jacobian of the update operator Fh only has eigenvalues with
absolute value smaller than 1. For simultaneous and alternating gradient descent this can be achieved (Lemma A.4
and A.5), if the Jacobian of the gradient vector field v only
has eigenvalues with negative real-part. While this condition suffices to prove convergence for small learning rates,
Mescheder et al. (2017) showed that simultaneous and alternating gradient descent might still require intractably small
learning rates if the imaginary part of the eigenvalues is
large. However, in our case we have the following simple
bound on the imaginary part of the eigenvalues:
Lemma A.9. Let
−P −B T
J=
(45)
B
−Q
where P ∈ Rn×n and Q ∈ Rm×m are symmetric. All
eigenvalues λ of J then satisfy
q
| Im(λ)| ≤ λmax (B T B).
(46)
Note that this bound is independent from P and Q.
Proof. Assume v, kvk = 1, is an eigenvector of J with
eigenvalue λ. Then
Im(λ) = v̄ T Ja v.
with Ja :=
inequality
1
2i (J
(47)
− J T ). Hence, by the Cauchy-Schwarz
| Im(λ)| ≤ kvkkJa vk = kJa vk.
(48)
But, if v T = (aT , bT ),
kJa vk2 = bBB T b + aB T Ba ≤ λmax (B T B).
v(θ, ψ) =
−f 0 (θψ)ψ
f 0 (θψ)θ
.
(52)
Because L(θ, 0) = L(0, ψ) = const for all θ, ψ ∈ R,
(θ, ψ) = (0, 0) is indeed a Nash-equilibrium for the game
defined by (51). Because we assume f 0 (t) 6= 0 for all t ∈ R,
we have v(θ, ψ) = 0 if and only if (θ, ψ) = (0, 0), showing
that (0, 0) is indeed the unique Nash-equilibrium.
Moreover, the Jacobian v 0 (θ, ψ) of v is given by
−f 00 (θψ)ψ 2
0
f (θψ) + f 00 (θψ)θψ
−f 0 (θψ) − f 00 (θψ)θψ
. (53)
f 00 (θψ)θ2
Evaluating it at the Nash equilibrium θ = ψ = 0, we obtain
v 0 (0, 0) =
0
−f 0 (0)
0
f (0)
0
(54)
which has the eigenvalues ±f 0 (0)i.
Lemma 2.3. The integral curves of the gradient vector field
v(θ, ψ) do not converge to the Nash-equilibrium. More
specifically, every integral curve (θ(t), ψ(t)) of the gradient
vector field v(θ, ψ) satisfies θ(t)2 + ψ(t)2 = const for all
t ∈ [0, ∞).
Proof. Let R(θ, ψ) := 12 (θ2 + ψ 2 ). Then
d
R(θ(t), ψ(t))
dt
= θ(t)v1 (θ(t), ψ(t)) + ψ(t)v2 (θ(t), ψ(t)) = 0, (55)
(49)
showing that R(θ, ψ) is indeed constant for all t ∈ [0, ∞).
This shows
| Im(λ)| ≤
q
λmax (B T B).
(50)
B. Proofs for the Dirac-GAN
This section contains the proofs for our results from Section 2 and Section 3 on the properties of the Dirac-GAN.
Lemma 2.2. The unique equilibrium point of the training
objective in (4) is given by θ = ψ = 0. Moreover, the
Jacobian of the gradient vector field at the equilibrium point
has the two eigenvalues ±f 0 (0) i which are both on the
imaginary axis.
Lemma 2.4. For simultaneous gradient descent, the Jacobian of the update operator Fh (θ, ψ) has
p eigenvalues
λ1/2 = 1 ± hf 0 (0)i with absolute values 1 + h2 f 0 (0)2
at the Nash-equilibrium. Independently of the learning rate,
simultaneous gradient descent is therefore not stable near
the equilibrium. Even stronger, for every initial condition
and learning rate h > 0, the norm of the iterates (θk , ψk )
obtained by simultaneous gradient descent is monotonically
increasing.
Proof. The first part is a direct consequence of Lemma A.4.
Which Training Methods for GANs do actually Converge?
To see the the norms of the iterates (θk , ψk ) is monotonically
increasing, we calculate
2
2
θk+1
+ ψk+1
= (θk − hf 0 (θk ψk )ψk )2 + (ψk + hf 0 (θk ψk )θk )2
= θk2 + ψk2 + h2 f 0 (θk ψk )2 (θk2 + ψk2 )
≥
θk2
+
ψk2 .
(56)
Lemma 2.5. For alternating gradient descent with ng generator and nd discriminator updates, the Jacobian of the
update operator Fh (θ, ψ) has eigenvalues
s
2
α2
α2
λ1/2 = 1 −
±
1−
− 1.
(5)
2
2
√
with α := ng nd hf 0 (0). For α ≤ 2, all eigenvalues are
hence on the unit circle. Moreover for α > 2, there are
eigenvalues outside the unit circle.
Proof. The update operators for alternating gradient descent
are given by
θ − hf 0 (θψ)ψ
F1 (θ, ψ) =
(57)
ψ
θ
F2 (θ, ψ) =
.
(58)
ψ + hf 0 (θψ)θ
Hence, the Jacobians of these operators at 0 are given by
1 −hf 0 (0)
0
F1 (0, 0) =
(59)
0
1
1
0
F20 (0, 0) =
.
(60)
hf 0 (0) 1
As a result, the Jacobian of the combined update operator is
n
(F2nd ◦ F1 g )0 (0, 0) = F20 (0, 0)nd · F10 (0, 0)ng
1
−ng hf 0 (0)
=
. (61)
nd hf 0 (0) −ng nd h2 f 0 (0)2 + 1
An easy calculation shows that the eigenvalues of this matrix
are
s
2
α2
α2
λ1/2 = 1 −
±
1−
−1
(62)
2
2
√
with α = ng nd hf 0 (0) which are on the unit circle if and
only if α ≤ 2.
Lemma 3.1. WGANs trained with simultaneous or alternating gradient descent with a fixed number of discriminator
updates per generator update and a fixed learning rate
h > 0 do generally not converge to the Nash equilibrium
for the Dirac-GAN.
Proof. First, consider simultaneous gradient descent. Assume that the iterates (θk , ψk ) converge towards the equilibrium point (0, 0). Note that (θk+1 , ψk+1 ) 6= 0 if (θk , ψk ) 6=
0. We can therefore assume without loss of generality that
(θk , ψk ) 6= 0 for all k ∈ N.
Because limk→∞ ψk = 0, there exists k0 such that for all
k ≥ k0 we have |ψk | < 1. For k ≥ k0 we therefore have
θk+1
1 −h
θk
=
.
(63)
ψk+1
h 1
ψk
For k ≥ k0 , the iterates are therefore given by
θk
θk0
1 −h
= Ak−k0
with A =
. (64)
ψk
ψk0
h 1
However, the eigenvalues of A are given by λ1/2 = 1 ±
√
hi which both have absolute value 1 + h2 > 1. This
contradicts the assumption that (θk , ψk ) converges to (0, 0).
A similar argument also hold for alternating gradient descent.
In this case, A is given by
n
n
1 0 d 1 −h g
1
−hng
=
. (65)
h 1
0 1
hnd 1 − h2 ng nd
The eigenvalues of A as in (65) are given by
s
2
h2 ng nh
h2 ng nh
1−
±
1−
− 1.
2
2
(66)
At least one of these eigenvalues has absolute value greater
or equal to 1. Note that for almost all initial conditions
(θ0 , ψ0 ), the the inner product between the eigenvector corresponding to the eigenvalue with modulus bigger than 1
will be nonzero for all k ∈ N. Since the recursion in (63)
is linear, this contradicts the fact that (θk , ψk ) → (0, 0),
showing that alternating gradient descent generally does not
converge to the Nash-equilibrium either.
Lemma 3.2. When using Gaussian instance noise with standard deviation σ, the eigenvalues of the Jacobian of the
gradient vector field are given by
p
λ1/2 = f 00 (0)σ 2 ± f 00 (0)2 σ 4 − f 0 (0)2 .
(6)
In particular, all eigenvalues of the Jacobian have negative
real-part at the Nash-equilibrium if f 00 (0) < 0 and σ > 0.
Hence, simultaneous and alternating gradient descent are
both locally convergent for small enough learning rates.
Proof. When using instance noise, the GAN training objective (1) is given by
h
i
Eθ̃∼N (θ,σ2 ) f (θ̃ψ) + Ex∼N (0,σ2 ) [f (−xψ)] . (67)
Which Training Methods for GANs do actually Converge?
The corresponding gradient vector field is hence given by
−ψf 0 (θ̃ψ)
ṽ(θ, ψ) = Eθ̃,x
.
(68)
θ̃f 0 (θ̃ψ) − xf 0 (−xψ)
The Jacobian ṽ 0 (θ, ψ) is therefore
−f 00 (θ̃ψ)ψ 2
−f 0 (θ̃ψ) − f 00 (θ̃ψ)θ̃ψ
Eθ̃,x
f 0 (θ̃ψ) + f 00 (θ̃ψ)θ̃ψ f 00 (θ̃ψ)θ̃2 + x2 f (−xψ)
(69)
Evaluating it at θ = ψ = 0 yields
0
−f 0 (0)
(70)
ṽ 0 (0, 0) =
f 0 (0) 2f 00 (0)σ 2
whose eigenvalues are given by
p
λ1/2 = f 00 (0)σ 2 ± f 00 (0)2 σ 4 − f 0 (0)2 .
(71)
In particular, for γ > 0 all eigenvalues have negative real
part. Hence, simultaneous and alternating gradient descent
are both locally convergent for small enough learning rates.
Proof. The regularized gradient vector field becomes
−f 0 (θψ)ψ
ṽ(θ, ψ) =
.
(72)
f 0 (θψ)θ − γψ
The Jacobian ṽ 0 (θ, ψ) is therefore given by
−f 00 (θψ)ψ 2
−f 0 (θψ) − f 00 (θψ)θψ
. (73)
f 0 (θψ) + f 00 (θψ)θψ
f 00 (θψ)θ2 − γ
whose eigenvalues are given by
r
γ
γ2
λ1/2 = − ±
− f 0 (0)2 .
2
4
Especially in the beginning of training, the discriminator
can reject samples produced by the generator with high
confidence (Goodfellow et al., 2014). When this happens,
the loss for the generator may saturate so that the generator
receives almost no gradient information anymore.
To circumvent this problem Goodfellow et al. (2014) introduced a nonsaturating objective for the generator. In
nonsaturating GANs, the generator objective is replaced
with6
max Epθ (x) f (−Dψ (x)).
(76)
θ
In our example, this is maxθ f (−ψθ).
Lemma 3.3. The eigenvalues of the Jacobian of the gradient vector field for the gradient-regularized GAN at the
equilibrium point are given by
r
γ
γ2
− f 0 (0)2 .
(8)
λ1/2 = − ±
2
4
Evaluating it at θ = ψ = 0 yields
0
−f 0 (0)
0
ṽ (0, 0) =
f 0 (0)
−γ
C.1. Nonsaturating GAN
(74)
While the nonsaturating generator objective was originally
motivated by global stability considerations, we investigate
its effect on local convergence. A linear analysis similar to
normal GANs yields
Lemma C.1. The unique Nash-equilibrium for the nonsaturating GAN on the example problem is given by θ = ψ = 0.
The eigenvalues of the Jacobian of the gradient vector field
at the equilibrium are ±f 0 (0)i which are both on the imaginary axis.
Proof. The gradient vector field for the nonsaturating GAN
is given by
0
−f (−θψ)ψ
v(θ, ψ) =
.
(77)
f 0 (θψ)θ
As in the proof of Lemma 2.2, we see that (ψ, θ) = (0, 0)
defines the unique Nash-equilibrium for the nonsaturating
GAN.
Moreover, the Jacobian v 0 (θ, ψ) is
f 00 (−θψ)ψ 2
−f 0 (−θψ) + f 00 (−θψ)θψ
.
f 0 (θψ) + f 00 (θψ)θψ
f 00 (θψ)θ2
(78)
At θ = ψ = 0 we therefore have
0
−f 0 (0)
v 0 (0, 0) =
.
(79)
f 0 (0)
0
with eigenvalues λ1/2 = ±f 0 (0)i.
(75)
C. Other regularization strategies
In this section we discuss further regularization techniques
for GANs on our example problem that were omitted in the
main text due to space constraints.
Lemma C.1 implies that simultaneous gradient descent is
not locally convergent for a nonsaturating GAN and any
learning rate h > 0, because the eigenvalues of the Jacobian
of the corresponding update operator Fh all have absolute
value larger than 1 (Lemma A.4). While Lemma C.1 also
rules out linear convergence towards the Nash-equilibrium
in the continuous case (i.e. for h → 0), the continuous
6
Goodfellow et al. (2014) used f (t) = − log(1 + exp(−t)).
Which Training Methods for GANs do actually Converge?
training dynamics could in principle still converge with a
sublinear convergence rate. Indeed, we find this to be the
case for the Dirac-GAN. We have
The training dynamics are
convergent with logarith hence
mic convergence rate O √1t .
Lemma C.2. For every integral curve of the gradient vector
field of the nonsaturating Dirac-GAN we have
For a more formal proof, first note that w is nonincreasing
by the first part of the proof. Moreover, for every > 0
there is δ > 0 such that for w < δ:
d
(θ(t)2 + ψ(t)2 ) = 2 [f 0 (θψ) − f 0 (−θψ)] θψ.
dt
(80)
For concave f this is nonpositive. Moreover, for f 00 (0) <
0, the continuous training dynamics of the nonsaturating
Dirac-GAN converge with logarithmic convergence rate.
Proof. The gradient vector field for the nonsaturating DiracGAN is given by
0
−f (−θψ)ψ
v(θ, ψ) =
.
(81)
f 0 (θψ)θ
Hence, we have
d
(θ(t)2 + ψ(t)2 ) = v1 (θ, ψ)θ + v2 (θ, ψ)ψ
dt
= 2θψ [f 0 (θψ) − f 0 (−θψ)] . (82)
For concave f , we have
f 0 (θψ) − f 0 (−θψ)
≤0
2θψ
and hence
d
(θ(t)2 + ψ(t)2 ) ≤ 0.
dt
f 0 (0) − ≤ φ̇ ≤ f 0 (0) +
00
2
ẇ ≤ (f (0) sin (2φ) + )w .
k∈Z
for t in a union of intervals QT ⊆ [0, T ] with total length
at least βbαT c with some constants α, β > 0 which are
independent of T .
For these t ∈ QT we have sin2 (2φ(t)) ≥
f 00 (0) < 0, this shows
1 00
f (0) + w(t)2
ẇ(t) ≤
2
2
2
(83)
w(t) ≤
(84)
(85)
4
ẇ = 4f (0)θ ψ + O(|θψ| )
= f 00 (0)w2 sin2 (2φ) + O(|w|4 ).
(86)
(87)
When we ignore higher order terms, we can solve this system explicitly7 for φ and w:
φ(t) = f 0 (0)(t − t0 )
w(t) =
7
(89)
2
−f 00 (0)t
+
f 00 (0)
4f 0 (0)
sin(4f 0 (0)(t − t0 )) + c
1
2.
Because
(94)
for t ∈ QT and small enough. Solving the right hand
formally yields
To intuitively understand why the continuous system converges with logarithmic convergence rate, note that near the
equilibrium point
have in polar coordi√ we asymptotically
√
nates (θ, ψ) = ( w cos(φ), w sin(φ)):
00
(92)
This implies that for every time interval [0, T ], φ(t) is in
[ π
π 3π
π
+k ,
+k
(93)
8
2 8
2
1
−( 21 f 00 (0)
(90)
For solving the ODE we use the separation of variablestechnique and the identity
Z
sin(2ax)
2 sin2 (ax)dx = x −
.
(88)
2a
+ )t + c
.
(95)
As w(t) is nonincreasing for t ∈
/ QT and the total length of
QT is at least βbαT c this shows that
Now assume that f 0 (0) 6= 0 and f 00 (0) < 0.
φ̇ = f 0 (0) + O(|w|1/2 )
(91)
2
w(T ) ≤
−( 12 f 00 (0)
1
.
+ )βbαT c + c
(96)
The training dynamics
hence converge with logarithmic
convergence rate O √1t .
Note that the standard choice f (t) = − log(1 + exp(−t))
is concave and satisfies f 00 (0) = − 14 < 0. Lemma C.1 is
hence applicable and shows that the GAN training dynamics for the standard choice of f converge with logarithmic
convergence rate in the continuous case. The training behavior of the nonsaturating GAN on our example problem
is visualized in Figure 3b.
C.2. Wasserstein GAN-GP
In practice, it can be hard to enforce the Lipschitz-constraint
for WGANs. A practical solution to this problem was given
by Gulrajani et al. (2017), who derived a simple gradient
penalty with a similar effect as the Lipschitz-constraint.
The resulting training objective is commonly referred to as
WGAN-GP.
Which Training Methods for GANs do actually Converge?
Similarly to WGANs, we find that WGAN-GP does not converge for the Dirac-GAN. A similar analysis also applies to
the DRAGAN-regularizer proposed in (Kodali et al., 2017).
The regularizer proposed by Gulrajani et al. (2017) is given
by
γ
2
R(ψ) = Ex̂ (k∇x Dψ (x̂)k − g0 )
(97)
2
where x̂ is sampled uniformly on the line segment between
two random points x1 ∼ pθ (x1 ), x2 ∼ pD (x2 ).
For the Dirac-GAN, it simplifies to
R(ψ) =
γ
(|ψ| − g0 )2
2
The corresponding gradient vector field is given by
−ψ
ṽ(θ, ψ) =
.
θ − sign(ψ)γ(|ψ| − g0 )
(98)
(99)
Lemma C.3. WGAN-GP trained with simultaneous or alternating gradient descent with a fixed number of generator
and discriminator updates and a fixed learning rate h > 0
does not converge locally to the Nash equilibrium for the
Dirac-GAN.
Proof. First, consider simultaneous gradient descent. Assume that the iterates (θk , ψk ) converge towards the equilibrium point (0, 0). For almost all initial conditions8 we have
(θk , ψk ) 6= (0, 0) for all k ∈ N. This implies
(100)
and hence limk→∞ |ψk+1 − ψk | = h|g0 | 6= 0, showing
that (θk , ψk ) is not a Cauchy sequence. This contradicts the
assumption that (θk , ψk ) converges to the equilibrium point
(0, 0).
A similar argument also holds for alternating gradient descent.
The training behavior of WGAN-GP on our example problem is visualized in Figure 3d.
As for WGANs, we stress that this analysis only holds if the
discriminator is trained with a fixed number of discriminator
updates per generator update. Again, more careful training
that ensures that the discriminator is kept exactly optimal or
two-timescale training (Heusel et al., 2017) might be able
to ensure convergence for WGAN-GP.
8
Depending on γ, h and g0 modulo a set of measure 0.
Consensus optimization (Mescheder et al., 2017) is an algorithm that attempts to solve the problem of eigenvalues with
zero real-part by introducing a regularization term that explicitly moves the eigenvalues to the left. The regularization
term in consensus optimization is given by
γ
kv(θ, ψ)k2
2
γ
= (k∇θ L(θ, ψ)k2 + k∇ψ L(θ, ψ)k2 ). (101)
2
R(θ, ψ) =
As was proved by Mescheder et al. (2017), consensus optimization converges locally for small learning rates h > 0
provided that the Jacobian v 0 (θ∗ , ψ ∗ ) is invertible.9
Indeed, for the Dirac-GAN we have
Note that the gradient vector field has a discontinuity at
the equilibrium point, as the gradient vector field takes on
values with norm bigger than some fixed constant in every
neighborhood of the equilibrium point. As a result, we have
|ψk+1 − ψk | = h|θk − γψk − sign(ψk )g0 |
C.3. Consensus optimization
Lemma C.4. The eigenvalues of the Jacobian of the gradient vector field for consensus optimization at the equilibrium
point are given by
λ1/2 = −γf 0 (0)2 ± if 0 (0)
(102)
In particular, all eigenvalues have a negative real part
−γf 0 (0)2 . Hence, simultaneous and alternating gradient
descent are both locally convergent using consensus optimization for small enough learning rates.
Proof. As was shown by Mescheder et al. (2017), the Jacobian of the modified vector field ṽ at the equilibrium point
is
ṽ 0 (0, 0) = v 0 (0, 0) − γv 0 (0, 0)| v 0 (0, 0).
In our case, this is
−γf 0 (0)2
f 0 (0)
−f 0 (0)
−γf 0 (0)2 .
(103)
(104)
A simple calculation shows that the eigenvalues of ṽ 0 (0, 0)
are given by
λ1/2 = −γf 0 (0)2 ± if 0 (0).
(105)
This concludes the proof.
A visualization of consensus optimization on the DiracGAN is given in Figure 3e.
Unfortunately, consensus optimization has the drawback
that it can introduce new spurious points of attraction to the
GAN training dynamics. While this is usually not a problem
9
Mescheder et al. (2017) considered only the case of isolated
equilibrium points. However, by applying Theorem A.3, it is
straightforward to generalize their proof to the case where we are
confronted with a submanifold of equivalent equilibrium points.
Which Training Methods for GANs do actually Converge?
for simple examples, it can be a problem for more complex
ones like deep neural networks.
A similar regularization term as in consensus optimization
was also independently proposed by Nagarajan & Kolter
(2017). However, Nagarajan & Kolter (2017) proposed to
only regularize the component ∇ψ L(θ, ψ) of the gradient
vector field corresponding to the discriminator parameters.
Moreover, the regularization term is only added to the generator objective to give the generator more foresight. It can
be shown (Nagarajan & Kolter, 2017) that this simplified
regularization term can in certain situations also make the
training dynamics locally convergent, but might be better
behaved at stationary points of the GAN training dynamics
that do not correspond to a local Nash-equilibrium. Indeed, a more detailed analysis shows that this simplified
regularization term behaves similarly to instance noise and
gradient penalties (which we discussed in Section 3.2 and
Section 3.3) on the Dirac-GAN.
D. General convergence results
In this section, we prove Theorem 4.1. To this end, we
extend the convergence proof by Nagarajan & Kolter (2017)
to our setting. We show that by introducing the gradient
penalty terms Ri (θ, ψ), we can get rid of the assumption that
the generator and data distributions locally have the same
support. As we have seen, this makes the theory applicable
to more realistic cases, where both the generator and data
distributions typically lie on lower dimensional manifolds.
D.1. Convergence proof
To prove Theorem 4.1, we first need to understand the local
structure of the gradient vector field v(θ, ψ). Recall that the
gradient vector field v(θ, ψ) is defined as
−∇θ L(θ, ψ)
v(θ, ψ) :=
(106)
∇ψ L(θ, ψ)
with
L(θ, ψ) = Ep(z) [f (Dψ (Gθ (z)))]
+ EpD (x) [f (−Dψ (x))] . (107)
Lemma D.1. The gradient of L(θ, ψ) with respect to θ is
given by
∇θ L(θ, ψ) = Ep(z) f 0 (Dψ (Gθ (z)) [∇θ Gθ (z)]T
· ∇x Dψ (Gθ (z)) . (108)
Similarly, the gradient of L(θ, ψ) with respect to ψ is given
by
∇ψ L(θ, ψ) = Epθ (x) [f 0 (Dψ (x))∇ψ Dψ (x)]
− EpD (x) [f 0 (−Dψ (x))∇ψ Dψ (x)] . (109)
Proof. This is just the chain rule.
Lemma D.2. Assume that (θ∗ , ψ ∗ ) satisfies Assumption I.
The Jacobian of the gradient vector field v(θ, ψ) at (θ∗ , ψ ∗ )
is then
T
0
−KDG
v 0 (θ∗ , ψ ∗ ) =
.
(110)
KDG KDD
The terms KDD and KDG are given by
KDD = 2f 00 (0) EpD (x) [∇ψ Dψ∗ (x)∇ψ Dψ∗ (x)T ] (111)
KDG = f 0 (0)∇θ Epθ (x) [∇ψ Dψ∗ (x)] |θ=θ∗
(112)
Proof. First note that by the definition of v(θ, ψ) in (106),
the Jacobian v 0 (θ∗ , ψ ∗ ) of v(θ, ψ) is given by
−∇2θ L(θ∗ , ψ ∗ ) −∇2θ,ψ L(θ∗ , ψ ∗ )
.
(113)
∇2θ,ψ L(θ∗ , ψ ∗ )
∇2ψ L(θ∗ , ψ ∗ )
By Assumption I, Dψ∗ (x) = 0 in some neighborhood
of supp pD . Hence, we also have ∇x Dψ∗ (x) = 0 and
∇2x Dψ∗ (x) = 0 for x ∈ supp pD . By taking the derivative of (108) with respect to θ and using ∇x Dψ∗ (x) =
0 and ∇2x Dψ∗ (x) = 0 for x ∈ supp pD we see that
∇2θ L(θ∗ , ψ ∗ ) = 0.
To show (111) and (112), simply take the derivative of
(109) with respect to θ and ψ and evaluate at it at (θ, ψ) =
(θ∗ , ψ ∗ ).
We now take a closer look at the regularized vector field.
Recall that we consider the two regularization terms
γ
(114)
R1 (θ, ψ) := EpD (x) k∇x Dψ (x)k2
2
γ
R2 (θ, ψ) := Epθ (x) k∇x Dψ (x)k2 .
(115)
2
As discussed in Section 4.1, the regularization is only applied to the discriminator. The regularized vector field is
hence given by
−∇θ L(θ, ψ)
ṽ(θ, ψ) :=
.
(116)
∇ψ L(θ, ψ) − ∇ψ Ri (θ, ψ)
Lemma D.3. The gradient ∇ψ Ri (θ, ψ) of the regularization terms Ri , i ∈ {1, 2}, with respect to ψ are
∇ψ R1 (θ, ψ) = γ EpD(x) [∇ψ,x Dψ (x)∇x Dψ (x)] (117)
∇ψ R2 (θ, ψ) = γ Epθ (x) [∇ψ,x Dψ (x)∇x Dψ (x)] . (118)
Proof. These equations can be derived by take the derivative
of (114) and (115) with respect to ψ.
Lemma D.4. The second derivatives ∇2ψ Ri (θ∗ , ψ ∗ ) of the
regularization terms Ri , i ∈ {1, 2}, with respect to ψ at
(θ∗ , ψ ∗ ) are both given by
LDD := γ EpD (x) [∇ψ,x Dψ∗ (x)∇ψ,x Dψ∗ (x)T ] . (119)
Which Training Methods for GANs do actually Converge?
satisfy
By Assumption III, this can only hold if v is in the tangent
space of MD at ψ ∗ .
Proof. ∇2ψ Ri (θ∗ , ψ ∗ ), i ∈ {1, 2}, can be computed by taking the derivative of (117) and (118) with respect to ψ and
using the fact that ∇x Dψ∗ (x) = 0 in a neighborhood of
supp pD .
Lemma D.6. Assume that Assumption III holds. If w 6= 0
is not in the tangent space of MG at θ∗ , then KDG w 6= 0.
Moreover,
both
regularization
∇θ,ψ Ri (θ∗ , ψ ∗ ) = 0.
terms
Moreover, we clearly have ∇θ,ψ R1 (θ∗ , ψ ∗ ) = 0, because
R1 does not depend on θ. To see that ∇θ,ψ R2 (θ∗ , ψ ∗ ) = 0,
take the derivative of (118) with respect to θ and use the
fact that ∇x Dψ∗ (x) = 0 and ∇2x Dψ∗ (x) = 0 for x ∈
supp pD .
As a result, the Jacobian ṽ 0 (θ∗ , ψ ∗ ) of the regularized gradient vector field at the equilibrium point is given by
T
0
−KDG
0 ∗
∗
ṽ (θ , ψ ) =
.
(120)
KDG KDD − LDD
For brevity, we define MDD := KDD − LDD .
To prove Theorem 4.1, we have to show that ṽ 0 (θ∗ , ψ ∗ ) is
well behaved when restricting it to the space orthogonal to
the tangent space of MG × MD at (θ∗ , ψ ∗ ):
Lemma D.5. Assume that Assumptions II and III hold.
If v 6= 0 is not in the tangent space of MD at ψ ∗ , then
v̄ T MDD v < 0.
Proof. By Lemma D.2, we have
h
i
2
v T KDD v = 2f 00 (0) EpD (x) (∇ψ Dψ∗ (x)T v)
(121)
and by Lemma D.4
h
i
2
v T LDD v = γ EpD (x) k∇x,ψ Dψ∗ (x)vk .
(122)
By Assumption II, we have f 00 (0) < 0. Hence, v T MDD v ≤
0 and v T MDD v = 0 implies
∇ψ Dψ∗ (x)T v = 0 and
∇x,ψ Dψ∗ (x)v = 0
(123)
for all x ∈ supp pD .
Proof. By Lemma D.2, we have
KDG w = f 0 (0) ∇θ Epθ (x) [∇ψ Dψ∗ (x)] |θ=θ∗ w
= f 0 (0)∂w g(θ). (127)
for
g(θ) := Epθ (x) [∇ψ Dψ∗ (x)] .
(128)
By Assumption III, this implies KDG w 6= 0 if w is not in
the tangent space of MG at θ∗ .
We are now ready to prove Theorem 4.1:
Theorem 4.1. Assume Assumption I, II and III hold for
(θ∗ , ψ ∗ ). For small enough learning rates, simultaneous
and alternating gradient descent for ṽ1 and ṽ2 are both
convergent to MG × MD in a neighborhood of (θ∗ , ψ ∗ ).
Moreover, the rate of convergence is at least linear.
Proof. First note that by Lemma D.1 and Lemma D.3
v(θ, ψ) = 0 for all points (θ, ψ) ∈ MG × MD , because
Dψ (x) = 0 and ∇x Dψ (x) = 0 for all x ∈ supp pD and
ψ ∈ MD . Hence, MG × MD consists only of equilibrium
points of the regularized gradient vector fields.
Let Tθ∗ MG and Tψ∗ MD denote the tangent spaces of MG
and MD at θ∗ and ψ ∗ .
We now want to show that both simultaneous and alternating
gradient descent are locally convergent to MG × MD for
the regularized gradient vector field ṽ(θ, ψ). To this end,
we want to apply Theorem A.3. By choosing local coordinates θ(α, γG ) and ψ(β, γD ) for MG and MD and using
Remark A.6, we can assume without loss of generality that
θ∗ = 0, ψ ∗ = 0 as well as
MG = Tθ∗ MG = {0}k × Rn−k
Let
l
h(ψ) := EpD (x) |Dψ (x)|2 + k∇x Dψ (x)k2 .
(124)
Using the fact that Dψ (x) = 0 and ∇x Dψ (x) = 0 for
x ∈ supp pD , we see that the Hessian of h(ψ) at ψ ∗ is
∇2ψ h(ψ ∗ ) = 2 EpD (x) ∇ψ Dψ (x)∇ψ Dψ (x)T
+ ∇ψ,x Dψ (x)∇ψ,x Dψ (x)T (125)
The second directional derivate ∂v2 h(ψ) is therefore
∂v2 h(ψ) = 2 EpD (x) |∇ψ Dψ (x)T v|2
+ k∇x,ψ Dψ (x)vk2 = 0. (126)
MD = Tψ∗ MD = {0} × R
m−l
(129)
.
(130)
This allows us to write10 ṽ(θ, ψ) = ṽ(α, γG , β, γD )
In order to apply Theorem A.3, we have to show that
∇(α,β) ṽ(θ∗ , ψ ∗ ) only has eigenvalues with negative realpart.
By Lemma D.2, ∇(α,β) ṽ(θ∗ , ψ ∗ ) is of the form
0
K̃DG
T
−K̃DG
K̃DD − L̃DD
(131)
10
By abuse of notation, we simply write θ = (α, γG ) and
ψ = (β, γD ).
Which Training Methods for GANs do actually Converge?
where K̃DD , K̃DG and L̃DD denote the submatrices of
KDD , KDG and LDD corresponding to the (α, β) coordinates.
We now show that M̃DD := K̃DD − L̃DD is negative definite and K̃DG has full column rank.
To this end, first note that
ṽ T M̃DD ṽ = v T MDD v
(132)
with v T := (ṽ T , 0). Note that v ∈
/ Tψ∗ MD for ṽ 6= 0.
Hence, by Lemma D.5 we have that ṽ T M̃DD ṽ < 0 if ṽ 6= 0.
As a result, we see that M̃DD is symmetric negative definite.
Similarly, for wT := (w̃T , 0), the components of KDG w
corresponding to the β-coordinates are given by K̃DG w̃.
Again, we have w ∈
/ Tθ∗ MG for w̃ 6= 0. Hence, by
Lemma D.6 we have that KDG w 6= 0 if w̃ 6= 0. Because the
components of KDG w corresponding to the γD coordinates
are 0, this shows that K̃DG w̃ 6= 0. K̃DG therefore has full
column rank.
Theorem A.7 now implies that all eigenvalues of
∇(α,β) ṽ(θ∗ , ψ ∗ ) have negative real part. By Lemma A.4,
Lemma A.5 and Theorem A.3, simultaneous and alternating gradient descent are therefore both convergent to
MG × MD near (θ∗ , ψ ∗ ) for small enough learning rates.
Moreover, the rate of convergence is at least linear.
D.2. Extensions
In the proof of Theorem 4.1 we have assumed that f 00 (0) <
0. This excludes the function f (t) = t which is used in
Wasserstein-GANs. We now show that our convergence
proof extends to the case where f (t) = t when we modify
Assumption III as little bit:
Remark D.7. When we replace h(ψ) with
h̃(ψ) := EpD (x) k∇x Dψ (x)k2
(133)
and MD with M̃D := {ψ | h̃(ψ) = 0} the results of
Theorem 4.1 still hold for f (t) = t.
implies that pθ = pD and hence, by Lemma D.1,
∇ψ L(θ, ψ) = Ex∼pD [∇ψ Dψ (x)]
− Ex∼pD [∇ψ Dψ (x)] = 0. (134)
To see why Lemma D.5 still holds, first note that for f (t) =
t, we have f 00 (0) = 0, so that by Lemma D.2 KDD = 0.
Hence,
v T MDD v = −v T LDD v.
(135)
We therefore have to show that v T LDD v 6= 0 if v is not in
the tangent space of MD .
However, we have seen in the proof of Lemma D.5 that
h
i
2
v T LDD v = γ EpD (x) k∇x,ψ Dψ∗ (x)vk .
(136)
Hence v T LDD v = 0 implies ∇x,ψ Dψ∗ (x)v = 0 for x ∈
supp pD and thus
∂v2 h(ψ) = 2 EpD (x) k∇x,ψ Dψ (x)vk2 = 0. (137)
By Assumption III, this can only be the case if v is in the
tangent space of MD . This concludes the proof.
In Section D.1, we showed that both regularizers R1 and R2
from Section 4.1 make the GAN training dynamics locally
convergent. A similar, but slightly more complex regularizer
was also proposed by Roth et al. (2017) who tried to find a
computationally efficient approximation to instance noise.
The regularizer proposed by Roth et al. (2017) is given by
a linear combination of R1 and R2 where the weighting is
adaptively chosen depending on the logits of Dψ (x) of the
current discriminator at a data point x:
RRoth (θ, ψ) = Epθ (x) (1 − σ(Dψ (x)))2 k∇x Dψ (x)k2
+ EpD (x) σ(Dψ (x)))2 k∇x Dψ (x)k2 (138)
Indeed, we can show that our convergence proof extends
to this regularizer (and a slightly more general class of
regularizers):
Remark D.8. When we replace the regularization terms R1
and R2 with
Proof. Almost everything in the proof of Theorem 4.1 still
holds for these modified assumptions. The only thing that
we have to show is that MG × MD still consists only of
equilibrium points and that Lemma D.5 still holds in this
setting.
To see the former, note that by Lemma D.1 we still have
∇θ L(θ, ψ) = 0 for (θ, ψ) ∈ MG × MD , because we have
∇x Dψ (x) = 0 for ψ ∈ MD and x ∈ supp pD . On the
other hand, for f (t) = t we also have ∇ψ L(θ, ψ) = 0
if θ ∈ MG , because for θ ∈ MG the definition of MG
R3 (θ, ψ) = Epθ (x) w1 (Dψ (x))k∇x Dψ (x)k2
+ EpD (x) w2 (Dψ (x))k∇x Dψ (x)k2 (139)
so that w1 (0) > 0 and w2 (0) > 0, the results of Theorem 4.1 still hold.
Proof. Again, we have to show that MG × MD still consists only of equilibrium points and that Lemma D.5 still
holds in this setting.
Which Training Methods for GANs do actually Converge?
However, by using ∇x Dψ (x) = 0 for x ∈ supp pD and
ψ ∈ MD , it is easy to see that ∇ψ R3 (θ, ψ) = 0 for all
(θ, ψ) ∈ MG × MD , which implies that MG × MD still
consists only of equilibrium points.
To formulate Assumption III0 , we need
To see why Lemma D.5 still holds in this setting, note that
(after a little bit of algebra) we still have ∇θ,ψ R3 (θ∗ , ψ ∗ ) =
0 and
1
∇2ψ R3 (θ∗ , ψ ∗ ) = (w1 (0) + w2 (0))LDD .
(140)
γ
Assumption III0 . There are -balls B (θ∗ ) and B (ψ ∗ )
around θ∗ and ψ ∗ so that MG ∩ B (θ∗ ) and MD ∩ B (ψ ∗ )
define C 1 - manifolds. Moreover, the following holds:
The proof of Lemma D.5 therefore still applies in this setting.
E. Stable equilibria for unregularized GAN
training
In Section 2, we have seen that unregularized GAN training
is not always locally convergent to the equilibrium point.
Moreover, in Section 4, we have shown that zero-centered
gradient penalties make general GANs locally convergent
under some suitable assumptions.
While our results demonstrate that we cannot expect unregularized GAN training to lead to local convergence for
general GAN architectures, there can be situations where
unregularized GAN training has stable equilibria. Such equilibria usually require additional assumptions on the class of
representable discriminators.
In this section, we identify two types of stable equilibria.
For the first class of stable equilibria, which we call Energy
Solutions, the equilibrium discriminator forms an energy
function for the true data distributions and might be a partial
explanation for the success of autoencoder-based discriminators (Zhao et al., 2016; Berthelot et al., 2017). For the
second class, which we call Full-rank solutions, the discriminator learns a representation of the data distribution with
certain properties and might be a partial explanation for the
success of batch-normalization for training GANs (Radford
et al., 2015).
E.1. Energy Solutions
For technical reasons, we assume that supp pD defines a
C 1 -manifold in this section.
Energy solutions are solutions where the discriminator forms
a potential function for the true data distribution. Such
solutions (θ∗ , ψ ∗ ) satisfy the following property:
Assumption I0 . We have pθ∗ = pD , Dψ∗ (x) = 0,
∇x Dψ∗ (x) = 0 and v T ∇2x Dψ∗ (x)v > 0 for all x ∈
supp pD and v not in the tangent space of supp pD at x.
We also need a modified version of Assumption III which
ensures certain regularity properties of the reparameterization manifolds MG and MD near the equilibrium (θ∗ , ψ ∗ ).
g̃(ψ) := ∇θ Epθ (x) [Dψ (x)]
θ=θ ∗
.
(141)
(i) if v is not in the tangent space of MD at ψ ∗ , then
∂v g̃(ψ ∗ ) 6= 0.
(ii) if w is not in the tangent space of MG at θ∗ , then
there is x ∈ supp pD so that ∇θ Gθ∗ (x)w is not in the
tangent space of supp pD at x ∈ supp pD .
The first part of Assumption III0 implies that the generator gradients become nonzero whenever the discriminator
moves away from an equilibrium discriminator. The second
part of Assumption III0 means that every time the generator
leaves the equilibrium, it pushes some data point aways
from supp pD , i.e. the generator is not simply redistributing
mass on supp pD .
In Theorem E.2 we show that energy solutions lead to local
convergence of the unregularized GAN training dynamics.
For the proof, we first need a generalization of Lemma D.2:
Lemma E.1. Assume that (θ∗ , ψ ∗ ) satisfies Assumption I0 .
The Jacobian of the gradient vector field v(θ, ψ) at (θ∗ , ψ ∗ )
is then given by
T
KGG −KDG
0 ∗
∗
v (θ , ψ ) =
.
(142)
KDG KDD
The terms KDD and KDG are given by
KGG = − f 0 (0) Ep(z) [∇θ Gθ∗ (z)]T
∇2x Dψ∗ (Gθ∗ (z))∇θ Gθ∗ (z)
(143)
KDD = 2f 00 (0) EpD (x) [∇ψ Dψ∗ (x)∇ψ Dψ∗ (x)T ] (144)
T
KDG = f 0 (0) ∇θ Epθ (x) [∇ψ Dψ∗ (x)] |θ=θ∗
(145)
Proof. Almost all parts of the proof of Lemma D.2 are
still valid. The only thing that remains to show is that
∇2θ L(θ∗ , ψ ∗ ) = −KGG . To see this, just take the derivative
of (108) with respect to θ and use the fact that ∇x Dψ (x) =
0 for x ∈ supp pD .
We are now ready to formulate our convergence result for
energy solutions:
Theorem E.2. Assume Assumption I0 , II and III0 hold for
(θ∗ , ψ ∗ ). Moreover, assume that f 0 (0) > 0. For small
enough learning rates, simultaneous and alternating gradient descent for the (unregularized) gradient vector field
v are both convergent to MG × MD in a neighborhood
of (θ∗ , ψ ∗ ). Moreover, the rate of convergence is at least
linear.
Which Training Methods for GANs do actually Converge?
Proof (Sketch). The proof is similar to the proof of Theorem 4.1.
First, note that MG × MD still only consists of equilibrium points. Next, we introduce local coordinates and show
that for v not in the tangent space of MG at θ∗ , we have
v T KGG v < 0. This can be shown using Lemma E.1, Assumption I0 and the second part of Assumption III0 .
Assumption I00 . We have pθ∗ = pD and ψ1∗ = 0.
We again consider reparameterization manifolds, which we
define as follows in this section:
MG := {θ | pθ = pD }
Moreover, let
Moreover, we need to show that for w not in the tangent
T w 6= 0. This can be
space of MD at ψ ∗ , we have KDG
shown by applying the first part of Assumption III0 .
The rest of the proof is the same as the proof of Theorem 4.1,
except that we have to apply Theorem A.8 instead of Theorem A.7.
Note that energy solutions are only possible, if the discriminator is able to satisfy Assumption III0 . This is not the
case for the Dirac-GAN from Section 2. However, if we
use a quadratic discriminator instead, there are also energy
solutions to the unregularized GAN training dynamics for
the Dirac-GAN. To see this, we can parameterize Dψ (x) as
Dψ (x) := ψ1 x2 + ψ2 x.
(146)
It is easy to check that the Dirac-GAN with a discriminator as in (146) indeed has energy solutions: every (θ, ψ)
with θ = 0 and ψ2 = 0 defines an equilibrium point of
the Dirac-GAN and the GAN-training dynamics are locally
convergent near this point if ψ1 > 0. Note however, that
even though all equilbria with ψ1 > 0 are points of attraction for the continuous GAN training dynamics, they may
not be attractors for the discretized system when ψ1 is large
and the learning rate h is fixed. In general, the conditioning
of energy solutions depends on the condition numbers of
the Hessians ∇2x Dψ∗ (x) at all x ∈ supp pD . Indeed, the
presence of ill-conditioned energy solutions might be one
possible explanation why WGAN-GP often works well in
practice although it is not even locally convergent for the
Dirac-GAN.
E.2. Full-Rank Solutions
In practice, Dψ (x) is usually implemented by a deep neural
network. Such discriminators can be described by functions
of the form
Dψ (x) = ψ1T ηψ2 (x)
(147)
with a vector-valued C 1 -functions ηψ2 and ψ = (ψ1 , ψ2 ).
ηψ2 can be regarded as a feature-representation of the data
point x.
We now state several assumptions that lead to local convergence in this situation.
The first assumption can be seen as a variant of Assumption I
adapted to this specific situation:
M0D := {ψ | ψ1 = 0}. (148)
g(θ) = Epθ (x) ηψ2∗ (x) .
(149)
Assumption III now becomes:
Assumption III00 . There is an -ball B (θ∗ ) around θ∗ so
that MG defines a C 1 - manifold11 . Moreover, the following
holds:
(i) The matrix EpD (x) ηψ2∗ (x)ηψ2∗ (x)T has full rank.
(ii) if w is not in the tangent space of MG at θ∗ , then
∂w g(θ∗ ) 6= 0.
We call a function ηψ2∗ that satisfies the first part of Assumption III00 a full-rank representation of pD . Moreover,
if ηψ2∗ satisfies the second part of Assumption III00 , we call
ηψ2∗ a complete representations, because the second part
of Assumption III00 implies that every deviation from the
Nash-equilibrium pθ∗ = pD is detectable using ηψ2∗ .
In practice, complete full-rank representations might only
exist if the class of discriminators is very powerful or the
class of generators is limited. Especially the second part of
Assumption III00 might be hard to satisfy in practice. Moreover, finding such representations might be much harder
than finding equilibria for the regularized GAN-training
dynamics from Section 4.
Nonetheless, we have the following convergence result for
GANs that allow for complete full-rank representations:
Theorem E.3. Assume Assumption I0 , Assumption II and
III0 hold for (θ∗ , ψ ∗ ). For small enough learning rates,
simultaneous and alternating gradient descent for the (unregularized) gradient vector field v are both convergent to
MG × M0D in a neighborhood of (θ∗ , ψ ∗ ). Moreover, the
rate of convergence is at least linear.
Proof (Sketch). The proof is again similar to the proof of
Theorem 4.1. We again introduce local coordinates and
show that for w not in the tangent space of M0D at ψ ∗ , we
have wT KDD w < 0. To see this, note that w must have a
nonzero ψ1 component if it is not in the tangent space of
M0D at ψ ∗ . However, using (111), we see that the submatrix
of KDD corresponding to the ψ1 coordinates is given by
K̃DD = 2f 00 (0) EpD (x) ηψ2∗ (x)ηψ2∗ (x)T .
(150)
11
Note that M0D is a C 1 -manifold by definition in this setup.
Which Training Methods for GANs do actually Converge?
This matrix is negative definite by Assumption II and the
first part of Assumption III00 .
Moreover, by applying (112), we see that the component
of KDG w, w ∈ Rn , corresponding to the ψ1 coordinates is
given by
∂w g(θ∗ ) = f 0 (0)∇θ Epθ (x) ηψ2∗ (x) θ=θ∗ w.
(151)
Using the second part of Assumption III00 , we therefore see
that for w not in the tangent space of MG at θ∗ , we have
KDG w 6= 0.
The rest of the proof is the same as the proof of Theorem 4.1.
For the Dirac-GAN from Section 2, we can obtain a complete full-rank representation, when we parameterize the
discriminator Dψ as Dψ (x) = ψ exp(x), i.e. if we set
ψ1 := ψ and ηψ2 (x) := exp(x). It is easy to check that ηψ2
indeed defines a complete full-rank representation and that
the Dirac-GAN is locally convergent to (θ∗ , ψ ∗ ) = (0, 0)
for this parameterization of Dψ (x).
F. Experiments
In this section, we describe additional experiments and give
more details on our experimental setup. If not noted otherwise, we always use the nonsaturating GAN-objective introduced by Goodfellow et al. (2014) for training the generator.
For WGAN-GP we use the generator and discriminator objectives introduced by Gulrajani et al. (2017).
2D-Problems For the 2D-problems, we run unregularized
GAN training, R1 -regularized and R2 -regularized GAN
training as well WGAN-GP with 1 and 5 discriminator
update per generator update. We run each method on 4
different 2D-examples for 6 different GAN architectures.
The 4 data-distributions are visualized in Figure 7. All 6
GAN architectures consist of 4-layer fully connected neural
networks for both the generator and discriminator, where
we select the number of hidden units from {8, 16, 32} and
use select either leaky RELUs (i.e. ϕ(t) = max(t, 0.2t)) or
Tanh-activation functions.
For each method, we try both Stochastic Gradient Descent (SGD) and RMS-Prop with 4 different learning
rates: for SGD, we select the learning rate from {5 ·
10−3 , 10−2 , 2 · 10−2 , 5 · 10−2 }. For RMSProp, we select
it from {5 · 10−5 , 10−4 , 2 · 10−4 , 5 · 10−4 }. For the R1 -,
R2 - and WGAN-GP-regularizers we try the regularization
parameters γ = 1, γ = 3 and γ = 10. For each method
and architecture, we pick the hyperparameter setting which
achieves the lowest Wasserstein-1-distance to the true data
distribution. We train all methods for 50k iterations and
we report the Wasserstein-1-distance averaged over the last
10k iterations. We estimate the Wasserstein-1-distance using the Python Optimal Transport package12 by drawing
2048 samples from both the generator and the true data
distributions.
The results are shown in Figure 5. The best solution found
by each method on the “Circle”-distribution is shown in Figure 8. We see that the R1 - and R2 -regularizers converge to
solutions for which the discriminator is 0 in a neighborhood
of the true data distribution. On the other hand, unregularized training and WGAN-GP converge to energy solutions
where the discriminator forms a potential for the true data
distribution. Please see Section E.1 for details.
CIFAR-10 To test our theory on real-world tasks, we train
a DC-GAN architecture (Radford et al., 2015) with 3 convolutional layers and no batch-normalization on the CIFAR-10
dataset (Krizhevsky & Hinton, 2009). We apply different
regularization strategies to stabilize the training. To compare
the different regularization strategies, we measure the inception score (Salimans et al., 2016a) over Wall-clock-time.
We implemented the network in the Tensorflow framework
(Abadi et al., 2016). For all regularization techniques, we
use the RMSProp optimizer (Tieleman & Hinton, 2012)
with α = 0.9 and a learning rate of 10−4 .
For the R1 and R2 regularizers from Section 4.1 we use a
regularization parameter of γ = 10. For the WGAN-GP
regularizer we also use a regularization parameter of γ = 10
as suggested by Gulrajani et al. (2017). We train all methods
using 1 discriminator update per generator update except
for WGAN-GP, for which we try both 1 and 5 discriminator
updates
The inception score (Salimans et al., 2016b) over time for
the different regularization strategies is shown in Figure 6.
As predicted by our theory, we see that the R1 and R2 regularizers from Section 4.1 lead to stable training whereas
unregularized GAN training is not stable. We also see that
WGAN-GP with 1 or 5 discriminator updates per generator
update lead to similar final inception scores on this architecture. The good behavior of WGAN-GP is surprising considering the fact that it does not even converge locally for the
Dirac-GAN. One possible explanation is that WGAN-GP
oscillates in narrow circles around the equilibrium which
might be enough to produce images of sufficiently high
quality. Another possible explanation is that WGAN-GP
converges to an energy or a full-rank solution (Section E)
for this example.
celebA We also train deep convolutional neural networks
on the celebA dataset (Liu et al., 2015). We use a similar
training setup as for the CIFAR-10 experiment, but we use
a slightly different architecture (Table 3). In particular, we
12
http://pot.readthedocs.io
Which Training Methods for GANs do actually Converge?
use a constant number of filters in each layer and we use
additional preactivation ResNet-blocks (He et al., 2016b;a).
We use 5 × 5 kernels for all convolutional and transposed
convolutional layers, except for the ResNet-blocks, in which
we use 3 × 3 kernels. We also find it helpful to multiply the
output of the ResNet-blocks with 0.1. We implemented the
network in the Tensorflow framework and use the RMSProp
optimizer with α = 0.9 and a learning rate of 10−4 . As a
regularization term, we use the R1 -regularizer with γ = 10.
For the latent code z, we use a 256 dimensional Gaussian
distribution.
We find that the R1 successfully stabilizes training of this
architecture. Some random samples can be seen in Figure 9.
Imagenet For the Imagenet experiment, we use ResNetarchitectures both for the generator and discriminator, both
having 55 layers in total. Both the generator and discriminator are conditioned on the labels of the input data. The
architectures for the generator and discriminator that we use
for the Imagenet experiment are shown in Table 4. As in the
celebA-experiment, we use preactivation ResNet-blocks and
Leaky RELU-nonlinearities everywhere. We also multiply
the output of the ResNet blocks with 0.1. For the generator,
we sample a latent variable z from a 256-dimensional uniform distribution on [−1, 1]256 and concatenate it with a 256
dimensional embedding of the labels. The resulting 512dimensional vector is then fed into the first fully connected
layer of the generator. The discriminator takes as input an
image and outputs a 1000 dimensional vector. Depending
on the label of the input, we select the corresponding index
in this vector and use it as the logits for the GAN-objective.
We implemented the network in the Pytorch framework
(Paszke et al., 2017) and use the RMSProp optimizer with
α = 0.99, = 10−5 and an initial learning rate of 10−4 .
We use a batch size of 128 and we trained the networks on
4 GeForce GTx 1080 Ti GPUs for 35 epochs. Every 10
epochs, we anneal the learning rate by a factor of 2.
We find that while training this GAN without any regularization quickly leads to mode collapse, using the R1 regularizers from Section 4.1 leads to stable training.13
Some random (unconditional) samples can be seen in Figure 10. Moreover, Figure 11 and Figure 12 show conditional
samples for some selected Imagenet classes. While not completely photorealistic, we find that our model can produce
convincing samples from all 1000 Imagenet classes.
Figure 6. Inception score over time for various regularization strategies. While the inception score can be problematic for evaluating
probabilistic models (Barratt & Sharma, 2018), it still gives a
rough idea about the convergence and stability properties of different training methods.
Layer
Fully Connected
Reshape
TransposedConv2D
TransposedConv2D
TransposedConv2D
output size
filter
256 · 4 · 4
256 × 4 × 4
128 × 8 × 8
64 × 16 × 16
3 × 32 × 32
256 → 256 · 4 · 4
256 → 128
128 → 64
64 → 3
(a) Generator architecture
Layer
Conv2D
Conv2D
Conv2D
Reshape
Fully Connected
output size
filter
64 × 16 × 16
128 × 8 × 8
256 × 4 × 4
256 · 4 · 4
256 · 4 · 4
3 → 64
64 → 128
128 → 256
256 · 4 · 4 → 1
(b) Discriminator architecture
Table 2. Architectures for CIFAR-10-experiment.
13
Under PyTorch, we sometimes still experienced some instabilities in the beginning of training, but these instabilities only
occurred during the first 1000 iterations and could be fixed by
restarting the training with a new random initialization. We did
not experience any instabilities under Tensorflow.
Which Training Methods for GANs do actually Converge?
(a) 2D-Gaussian
(b) Line
(c) Circle
(d) Four lines
Figure 7. The four 2D-data distributions on which we test the different algorithms.
(a) unregularized
(b) R1
(c) R2
(d) WGAN-GP-1
(e) WGAN-GP-5
Figure 8. Best solutions found by the different algorithms for learning a circle. The blue points are samples from the true data distribution,
the orange points are samples from the generator distribution. The colored areas visualize the gradient magnitude of the equilibrium
discriminator. We find that while the R1 and R2 regularizers converge to equilibrium discriminators that are 0 in a neighborhood of the
true data distribution, unregularized training and WGAN-GP converge to energy solutions (Section E.1).
(a) Training distribution
(b) Random samples
Figure 9. Results on the celebA dataset (Liu et al., 2015) (64 × 64) for a DC-GAN (Radford et al., 2015) based architecture with additional
residual connections (He et al., 2016b). For both the generator and the discriminator, we do not use batch normalization and use a constant
number of filters in each layer.
Which Training Methods for GANs do actually Converge?
(a) Training distribution
(b) Random samples
Figure 10. Unconditional results on the Imagenet dataset (Russakovsky et al., 2015) with resolution 128 × 128. The inception score is
18.5 ± 0.4.
Which Training Methods for GANs do actually Converge?
(a) tench
(b) papillon
(c) weevil
(d) admiral
(e) lighthouse
(f) bell cote
(g) castle
(h) dam
(i) dock
Figure 11. Class conditional results on the Imagenet dataset.
Which Training Methods for GANs do actually Converge?
(a) home theater
(b) jack-o’-lantern
(c) passenger car
(d) police van
(e) rugby ball
(f) ski
(g) trifle
(h) pizza
(i) valley
Figure 12. Class conditional results on the Imagenet dataset.
Which Training Methods for GANs do actually Converge?
Layer
Layer
output size
filter
Fully Connected
Reshape
TransposedConv2D
128 · 4 · 4
128 × 4 × 4
128 × 8 × 8
256 → 128 · 4 · 4
128 → 128
Resnet-Block (3x)
TransposedConv2D
128 × 8 × 8
128 × 16 × 16
128 → 64 → 128
128 → 128
Resnet-Block (3x)
TransposedConv2D
128 × 16 × 16
128 × 32 × 32
128 → 64 → 128
128 → 128
Resnet-Block (3x)
TransposedConv2D
128 × 32 × 32
3 × 64 × 64
128 → 64 → 128
64 → 3
(a) Generator architecture
Layer
output size
filter
Conv2D
128 × 32 × 32
3 → 128
Resnet-Block (2x)
Conv2D
128 × 32 × 32
128 × 16 × 16
128 → 64 → 128
128 → 128
Resnet-Block (2x)
Conv2D
128 × 16 × 16
128 × 8 × 8
Resnet-Block (2x)
Conv2D
Resnet-Block (2x)
Reshape
Fully Connected
output size
filter
Fully Connected
Reshape
1024 · 4 · 4
1024 × 4 × 4
512 → 1024 · 4 · 4
-
Resnet-Block (4x)
NN-Upsampling
1024 × 4 × 4
1024 × 8 × 8
1024 → 512 → 1024
-
Conv2D
Resnet-Block (4x)
NN-Upsampling
1024 × 8 × 8
1024 × 8 × 8
1024 × 16 × 16
1024 → 1024
1024 → 512 → 1024
-
Conv2D
Resnet-Block (4x)
NN-Upsampling
512 × 16 × 16
512 × 16 × 16
512 × 32 × 32
1024 → 512
512 → 256 → 512
-
Conv2D
Resnet-Block (4x)
NN-Upsampling
256 × 32 × 32
256 × 32 × 32
256 × 64 × 64
512 → 256
256 → 128 → 256
-
Conv2D
Resnet-Block (4x)
NN-Upsampling
128 × 64 × 64
128 × 64 × 64
128 × 128 × 128
256 → 128
128 → 64 → 128
-
Conv2D
Resnet-Block (4x)
Conv2D
64 × 128 × 128
64 × 128 × 128
3 × 128 × 128
128 → 64
64 → 32 → 64
16 → 3
(a) Generator architecture
Layer
output size
filter
Conv2D
64 × 128 × 128
3 → 64
128 → 64 → 128
64 → 128
Resnet-Block (4x)
Conv2D
64 × 128 × 128
128 × 64 × 64
64 → 32 → 64
64 → 128
128 × 8 × 8
128 × 4 × 4
128 → 64 → 128
128 → 128
Resnet-Block (4x)
Conv2D
128 × 64 × 64
256 × 32 × 32
128 → 64 → 128
128 → 256
128 × 4 × 4
128 · 4 · 4
128 · 4 · 4
128 → 64 → 128
128 · 4 · 4 → 1
Resnet-Block (4x)
Conv2D
256 × 32 × 32
512 × 16 × 16
256 → 128 → 256
256 → 512
(b) Discriminator architecture
Resnet-Block (4x)
Conv2D
512 × 16 × 16
1024 × 8 × 8
512 → 256 → 512
512 → 1024
Table 3. Architectures for celebA-experiment.
Resnet-Block (4x)
Conv2D
1024 × 8 × 8
1024 × 4 × 4
1024 → 512 → 1024
1024 → 1024
Resnet-Block (4x)
Reshape
1024 × 4 × 4
1024 · 4 · 4
1024 → 512 → 1024
-
1024 · 4 · 4
1024 · 4 · 4 → 1000
Fully Connected
(b) Discriminator architecture
Table 4. Architectures for Imagenet-experiment.
Which Training Methods for GANs do actually Converge?
(a) Standard GAN
(b) Non-saturating GAN
(c) WGAN
(d) WGAN-GP
(e) Consensus optimization
(f) Instance noise
(g) Gradient penalty
(h) Gradient penalty (CR)
Figure 13. Convergence properties of different GAN training algorithms using simultaneous gradient descent. The shaded area in
Figure 13c visualizes the set of forbidden values for the discriminator parameter ψ. The starting iterate is marked in red.
(a) Standard GAN
(b) Non-saturating GAN
(c) WGAN
(d) WGAN-GP
(e) Consensus optimization
(f) Instance noise
(g) Gradient penalty
(h) Gradient penalty (CR)
Figure 14. Convergence properties of different GAN training algorithms using alternating gradient descent with 1 discriminator
update per generator update The shaded area in Figure 14c visualizes the set of forbidden values for the discriminator parameter ψ.
The starting iterate is marked in red.
Which Training Methods for GANs do actually Converge?
(a) Standard GAN
(b) Non-saturating GAN
(c) WGAN
(d) WGAN-GP
(e) Consensus optimization
(f) Instance noise
(g) Gradient penalty
(h) Gradient penalty (CR)
Figure 15. Convergence properties of different GAN training algorithms using alternating gradient descent with 5 discriminator
updates per generator update. The shaded area in Figure 15c visualizes the set of forbidden values for the discriminator parameter
ψ. The starting iterate is marked in red.
Which Training Methods for GANs do actually Converge?
(a) SimGD
(b) AltGD (nd = 1)
(c) AltGD (nd = 5)
Figure 16. Convergence properties of our GAN using two time-scale training as proposed in (Heusel et al., 2017). On our example
problem, we do not see any sign of convergence when training with two time-scales. The starting iterate is marked in red.
| 2cs.AI
|
1
A Tractable Framework for the Analysis of Dense
Heterogeneous Cellular Networks
Serkan Ak, Hazer Inaltekin, Member, IEEE,
arXiv:1610.05617v3 [cs.IT] 26 Mar 2018
H. Vincent Poor, Fellow, IEEE
Abstract
This paper investigates the downlink performance of dense K-tier heterogeneous cellular networks (HCNs)
under general settings. First, Gaussian approximation bounds for the standardized aggregate wireless interference (AWI) in dense K-tier HCNs are obtained for when base stations (BSs) in each tier are distributed over
the plane according to a spatial and general Poisson point process. The Kolmogorov-Smirnov (KS) distance is
used to measure deviations of the distribution of the standardized AWI from the standard normal distribution.
An explicit and analytical expression bounding the KS distance between these two distributions is obtained
as a function of a broad range of network parameters such as per-tier transmission power levels, per-tier
BS intensity, BS locations, general fading statistics, and general bounded path-loss models. Bounds achieve
a good statistical match between the standardized AWI distribution and its normal approximation even for
moderately dense HCNs. Second, various spatial performance metrics of interest such as outage capacity,
ergodic capacity and area spectral efficiency in the downlink of K-tier HCNs for general signal propagation
models are investigated by making use of the derived distribution approximation results. Considering two
specific BS association policies, it is shown that the derived performance bounds track the actual performance
metrics reasonably well for a wide range of BS intensities, with the gap among them becoming negligibly small
for denser HCN deployments. Finally, both analytical and numerical results on the area spectral efficiency reveal
a non-linear growth trend with diminishing returns of HCN performance. Hence, the SIR invariance property
does not hold under bounded path-loss models, which is a critical finding from the HCN design perspective.
In particular, it points out a critical BS density beyond which the HCN performance starts to decline due to
excessive wireless interference.
S. Ak is with the Wireless Networking and Communications Group (WNCG), The University of Texas at Austin, Austin, TX 78701,
USA (e-mail: [email protected]). H. Inaltekin is with the Department of Electrical and Electronic Engineering at the University of
Melbourne, VIC 3010, Australia (e-mail: [email protected]). H. V. Poor is with the Department of Electrical Engineering,
Princeton University, Princeton, NJ 08544, USA (e-mail: [email protected]).
This research was supported in part by the European Union Research Executive Agency Marie Curie FP7-Reintegration-Grants
under Grant PCIG10-GA-2011-303713, in part by the Scientific and Technological Research Council of Turkey (TUBITAK) under
Grant 115E162, and in part by the U.S. National Science Foundation under Grants CNS-1702808 and ECCS-1647198. This work was
presented in part at the IEEE International Symposium on Information Theory (ISIT), Barcelona, Spain, July 2016 [1], [2].
2
Index Terms
Heterogeneous Cellular Networks, 5G, Downlink Interference, Gaussian Approximation, Outage Capacity,
Poisson Point Processes.
I. I NTRODUCTION
A. Background and Motivation
Fifth generation (5G) wireless networks are conceived as highly heterogeneous consisting of multipletiers of network elements with much denser deployments, e.g., hundreds of transmitters per unit area
and more advanced communication protocols to deal with excessive data demand from mobile users
[3]–[11]. Modeling and analyzing performance of such multi-tier heterogeneous cellular networks
(HCNs) using spatial point processes have recently gained increasing popularity [12]–[16]. Poisson
cluster processes have also recently been used for the same purpose in [17], [18]. In particular, it is
shown in [12] that using a Poisson point process (PPP) model even for macro cell BS locations provides
us with an approximation as good as the one provided by the conventional grid based model [19] for
the actual network performance. The PPP based location model leads to a lower performance bound,
whilst the grid model resulting in an upper performance bound, with almost the same deviations from
the experimental data [12]. The appropriateness of the random spatial models for network modeling
and analysis is expected to increase even further with the more irregular topology of HCNs.
Along with PPP based modeling for analytical tractability, there is an increasing tendency towards
network densification in order to improve data rates in HCNs [5]–[9]. From an engineering point of
view, the significant capacity improvements are expected to emerge automatically in ultra-dense HCNs
due to shorter connection distances arising from the deployment of hundreds (or, even thousands) of
small cell base-stations (BSs) per square kilometer. Contrary to this intuitive expectation, an important
observation in some of these previous studies such as [6]–[8] is that there is a certain level of network
densification to achieve the best data rate performance. It is also concluded in these papers that this
critical level of network densification, depending on network modeling assumptions, is inherently
related to a variety of networks parameters such as association policy, fading distribution, path-loss
model, and even the antenna heights and elevation angles.
Our results in this paper provide new insights into the performance gains that can be achieved through
network densification in HCNs. Specifically, we prove that network interference resembles a Gaussian
distributed random variable with a growing mean in dense HCNs under any given bounded path-loss
model. Utilizing the discovered statistical structure of network interference across HCN performance
3
analysis, it is also shown that network interference renders the advantages of network densification
moot after some certain BS density level. Roughly speaking, this finding implies that it is the bounded
nature of any path-loss model that will restrict the benefits of network densification strategies in
next-generation wireless communication systems. Last but not least, the discovered Gaussian structure
of network interference is also important from the network analysis and planning perspective as it
facilitates the performance analysis of dense HCNs greatly by eliminating the need for complicated
numerical integral computations for inverting complex-valued characteristic functions and/or Laplace
transforms, as illustrated later in the paper through bounding various HCN capacity metrics.
More generally, a major design issue for PPP based HCN models, dense or not, is characterization
and mitigation of aggregate wireless interference (AWI), e.g., see [20] and [21]. Although the statistical
characterization of AWI is possible for the special case of Rayleigh fading and the classical unbounded
path-loss model to compute various performance metrics such as outage probability in closed form
[13]–[15], computation of the exact AWI distribution is a very challenging task that usually does
not result in closed form expressions for general signal propagation models [12], [22]. In addition to
network planning insights mentioned above, another important contribution of the current paper is to
generalize those previously known results and techniques to study the HCN performance under more
general and heterogeneous communication scenarios that take into account the general bounded pathloss models, general fading distributions and spatial distributions of BSs including non-homogeneous
PPPs. This is achieved by leveraging our methodological analysis approximating the standardized AWI
distribution as a normal distribution.
B. Main Contributions
The main contributions of the current paper are listed below.
1) Development of a New Analytical Framework for the HCN Performance Analysis: This paper
develops a new analytical framework for the HCN performance analysis. The developed framework
results in performance figures arbitrarily close to the actual network performance in the dense network
limit. Specifically, a thorough mathematical analysis of the AWI distribution under general network
settings is often a prohibitively hard task. To overcome this difficulty, we introduce a simple and
effective methodological approach for examining the statistical structure of AWI in the downlink of a
dense K-tier HCN, wherein the network tiers are differentiated from each other in terms of transmission
power levels, spatial BS distributions and RF signal propagation characteristics. The proposed approach
provides us with a principled framework to analyze the downlink performance of spatially dense Ktier HCNs, where BS locations in each tier follow a PPP with possibly non-homogeneous spatial BS
4
intensity. The signal power attenuation due to path-loss is modeled through a general bounded pathloss function decaying to zero with distance, which can vary from one tier to another. Fading and
shadowing are also accounted for in the employed signal propagation model without assuming any
specific distribution functions for these other random wireless channel dynamics.
2) Derivation of Gaussian Approximation Bounds for the Distribution of AWI in HCNs: This paper
derives tight Gaussian approximation bounds for the distribution of AWI for the downlink HCN
communications. More specifically, measuring the distance between the standardized downlink AWI
and normal distributions by means of the Kolmogorov-Smirnov (KS) distance, we obtain an analytical
expression for deviations between them, which paves the way for illuminating statistical behavior of
the AWI and in turn for designing an efficient K-tier HCN. Briefly, the stated distance consists of
two parts: (i) a scaling coefficient and (ii) a multiplicative positive function c(x) with x ∈ R being
the point at which we want to estimate the value of the standardized AWI distribution.
The scaling coefficient has two important properties. First, it is related to the skewness characteristic
of the individual interference terms constituting the standardized AWI, which mainly depends on
various network parameters at each tier such as transmission powers, BS distribution and signal
propagation characteristics. The second property is its monotonically decaying nature to zero with
denser deployments of BSs per tier. The function c(x) is uniformly bounded by a small constant and
approaches zero for large absolute values of x at a rate |x|−3 , which makes the derived bounds on the
tails of the standardized downlink AWI distribution tight even for sparsely deployed HCNs.
3) Derivation of the Bounds on Various HCN Performance Metrics Under General Settings:
Utilizing the derived Gaussian approximation results across HCN performance analysis, this paper
produces analytical expressions for upper and lower bounds on performance metrics of interest in K-tier
HCNs under general network settings and general PPPs. In particular, we focus on the downlink data
rates in a dense K-tier HCN under two different BS association policies. We obtain tight performance
bounds on the downlink outage capacity, ergodic capacity and area spectral efficiency (ASE) in HCNs
under both association policies for general signal propagation models. The derived bounds approximate
the actual HCN performance metrics accurately for a wide range of BS intensities in each tier. The
gap between bounds becomes negligibly small as the BSs are more densely deployed.
4) Network Design Insights: This paper also produces new network design insights for K-tier
HCNs. In particular, based on the derived analytical expressions, we show that the outage probability
increases if the BSs in a K-tier HCN are more densely deployed. This finding implies that the celebrated
signal-to-interference ratio (SIR) invariance property [12], [13] does not hold anymore if a bounded
path-loss model is used to characterize large-scale wireless propagation losses in the downlink of an
5
HCN. Hence, there is a fundamental capacity limit beyond which the data rate performance of an HCN
cannot be increased indefinitely by adding extra infrastructure. As a result, we must either mitigate
interference more efficiently or find the optimum BS intensity per tier maximizing delivered data rates
to mobile users in order to reap the benefits from network densification.
Our simulations corroborate these findings by indicating that an increase in the network intensity
induces degradations in outage capacity and ergodic capacity, each of which scales at least as fast as
, where λ = [λ1 , . . . , λK ]⊤ is the BS intensity vector. Further, the ASE performance of a
Θ kλk−1
2
K-tier HCN exhibits a non-linear growth trend with diminishing returns and a plateauing behavior
after some BS density level. From an HCN design perspective, this result further supports the necessity
of setting BS intensities at each tier appropriately for the proper and cost-effective delivery of data
services to mobile users. Last but not least, the general path-loss model we consider in this paper
covers the stretched exponential path-loss models and bounded multi-slope path-loss models [8], [23],
[24]. Hence, our results are also appropriate for characterizing the performance of emerging millimeter
wave (mmWave) communications in the dense network limit. Overall, the proposed approach can be
extended to other HCN settings and performance metrics, with the potential of providing new network
design insights and shedding light on HCN performance beyond specific selections of the path-loss
model and the fading distribution. To this end, some potential extensions of the proposed approach
are further discussed in Section VII.
C. Paper Organization and Notation
The rest of the paper is organized as follows. In Section II, we present a survey of relevant previous
work by comparing and contrasting existing results with those we obtain in the current paper. In
Section III, we introduce the system model. In Section IV, we present our methodological approach
for approximating the AWI distribution in K-tier HCNs as a Gaussian distribution and illustrate
numerical examples of these analytical findings. In Section V, we introduce BS association policies
for performance analysis and derive our bounds tracking the performance metrics, especially outage
characteristics of mobile users, under these association policies. We illustrate simulation results on the
derived performance bounds in Section VI. We provide a discussion on the scope, generalizations and
limitations of the derived results in Section VII. Finally, Section VIII concludes the paper.
We use boldface letters, upper-case letters and calligraphic letters to denote vector quantities, random
variables and sets, respectively. | · | notation is used to measure the magnitudes of scalar quantities,
whereas k · k2 notation is used to measure the Euclidean norms of vector quantities. 1{·} is used to
denote the indicator function. Expected value and variance of a random variable X are denoted by
6
E [X] and Var (X), respectively. As is standard, when we write f (t) = O (g (t)), f (t) = Ω (g (t)) and
(t)
f (t) = o (g (t)) as t → t0 for two positive functions f (t) and g (t), we mean lim supt→t0 fg(t)
< ∞,
(t)
(t)
> 0 and limt→t0 fg(t)
= 0, respectively. f (t) is said to be Θ (g (t)) as t → t0 if f (t) =
lim inf t→t0 fg(t)
O (g (t)) and f (t) = Ω (g (t)) as t → t0 .
II. R ELATED WORK
The early work in the literature focusing on the design and analysis of wireless networks by means
of stochastic geometry based models includes [25]–[29]. These papers considered traditional singletier macro cell deployments and obtained various approximations on the distribution of AWI using
characteristic functions [25], LePage series [26], Edgeworth expansion [27], geometrical considerations
[28] and skewed stable distributions [29]. In [25], a closed form expression was also obtained for the
AWI distribution under the assumption of no fading and unbounded power-law decaying path-loss
function when the path-loss exponent is 4. More recently, generalized shot-gun models are considered
for one-, two- and three-dimensional wireless networks to derive semi-analytical expressions for the
downlink coverage probability for arbitrary fading and general path-loss models in [30], optimum
downlink coverage for Poisson cellular networks subject to transmit power, BS density and transmit
power density constraints is derived in [31], and Berry-Esseen types of bounds were obtained in
[32]–[34], but again by considering only single-tier wireless networks. In [35] and [36], the authors
used a Gaussian signal approximation technique to calculate bit error rate and symbol error rate as
performance metrics for cellular networks.
The current paper differs from the above previous work in several important aspects. In particular,
this paper extends the previous known results approximating the AWI distribution for macro cell
deployments to more heterogeneous and complex wireless communication environments when compared to [25]–[34]. Functional dependencies among different tiers to approximate the AWI distribution
in the downlink of a K-tier HCN are clearly identified. While the authors in [29] showed that the
AWI distribution can be modeled as a skewed stable distribution with unity skew parameter under the
unbounded path-loss model, our approximation technique uses general bounded path-loss models and
demonstrates that the skewness of the standardized AWI distribution decreases to zero as the K-tier
HCN under consideration gets denser. We discover that the behavior of AWI changes from being
heavy-tailed to an exponentially decaying light-tailed one for bounded path-loss models decaying to
zero.
This paper differs from [30]–[33] by focusing on a multi-tier network scenario with distinct and
general network parameters in each tier, which enables us to discover further insights into the behavior
7
of AWI in dense HCNs via Lemmas 1 and 2 in Section IV. The authors in [30] focus only on single-tier
networks, and their equivalence results are between one-, two- and three-dimensional Poisson wireless
networks. As a result, the results presented in [30] are not applicable to our case to first reduce the
multi-tier network to a single-tier network, and then use the Gaussian approximation results presented
in our previous work [33], [34]. In comparison with the preliminary results presented in [34], the
association policies (APs) studied for K-tier HCNs in Section V are much richer in terms of the
network parameters that they include, which provides additional insights into various capacity metrics
for dense K-tier HCNs.
The papers [35] and [36] differ from the current paper by considering diversity at the bit and
symbol level through bit-error-rate and/or symbol-error-rate calculations based on an equivalencein-distribution (EiD) technique. The introduced EiD technique requires characteristic function (CF)
calculations for the interference signal, while our Gaussian approximation approach does not require
any such CF (and/or Laplace transform) calculations due the employed Berry-Esseen technique.
Further, we only consider the statistics of the AWI power in this paper, which is the mainstream
approach taken in many previous studies, i.e., see [22]. Amplitude statistics of the AWI considered in
[35] and [36] are strongly related to the type of signaling scheme employed, which is not within the
scope of the current paper.
The related work also includes the papers that use stochastic geometry to model and analyze HCNs
such as [37]–[43]. To start with, the papers [37] and [38] calculate the statistics of the signal-tointerference-plus-noise ratio (SINR) in the downlink of HCNs by utilizing moment generating functions
(MGF). In [39], the authors investigated a gamma distribution approximation for the distribution of
AWI clogging a fixed-size cell with a guard zone and a dominant interferer. In [40], the author derived
the downlink SINR distribution for K-tier HCNs by assuming the classical unbounded path-loss model,
Rayleigh faded wireless links and the nearest base-station (BS) association rule. In [41], they considered
vector broadcast channels operating according to opportunistic beamforming in a K-tier HCN setting,
and obtained tight approximations for beam outage probabilities and ergodic aggregate data rates for
Rayleigh faded propagation environments with homogeneous PPP distributions for the locations of
access points.
In [42], the authors extended the previous results in [30] for four different association policies
in HCNs (i.e., max-SINR, nearest-BS, maximum received instantaneous power and maximum biased
received power association models) for an unbounded path-loss model with specific functional form
(having varying path-loss exponents from tier-to-tier) and with arbitrary fading when the BS locations
in each tier are given by homogeneous PPPs. They obtained semi-analytical expressions for the
8
downlink performance of HCNs involving complex-valued integrals. In [43], the authors used the
factorial moments for the SINR process to show that a generic K-tier HCN can be transformed into a
stochastically equivalent single-tier network when the BSs in each tier are distributed over the plane
according to homogeneous PPPs and the path-loss model is given by the unbounded inverse power-law
function.
The current paper differs from [37] and [38] by utilizing general and tier-dependent bounded
path-loss models as well as using general fading distributions. In comparison with wireless network
performance results obtained for homogeneous PPPs, we show that incorporating the skewness of
the standardized AWI via the Berry-Esseen theorem makes the Gaussian approximation analytically
more useful and numerically more tractable to accurately approximate the AWI distribution even for
non-homogeneous PPPs and general fading models. When compared with the results reported in [39]–
[41], our network set-up is much richer, allowing non-homogeneous PPPs for BS locations and general
signal propagation models including fading and shadowing.
This paper differs from [42] in two important aspects. Firstly, our network model is not restricted to
homogeneous PPPs and the unbounded path-loss model. Secondly, our Gaussian approximation technique does not necessitate an inversion of complex-valued characteristic functions. More specifically,
our model generalizes the results in [42] to non-homogeneous PPPs with arbitrary mean-measure and
to bounded path-loss models with arbitrary functional forms by only requiring computation of integrals
with respect to the generalized fading distributions on the real line to obtain tight performance bounds
on the downlink of K-tier HCNs. The approach presented in [43] for mapping a K-tier HCN to a
stochastically equivalent single-tier planar wireless network is promising, but the key propagation invariance property disappears for general bounded path-loss models. Hence, such stochastic equivalence
results are not directly applicable to the K-tier HCN setup with general bounded path-loss models and
non-homogeneous PPPs studied in this paper. More generally, we observe that the SINR invariance
property does not hold in our case, which reinforces the similar observations obtained for stretched
exponential path-loss models and multi-slope path-loss models in [8], [24], respectively.
III. S YSTEM M ODEL
In this section, we will introduce the details of the studied downlink model in a K-tier cellular topology, the details of the spatial processes determining BS locations, the signal propagation characteristics
and the association policy under which the network performance of a K-tier HCN is determined.
9
A. The Downlink Model in a K-Tier Cellular Topology
We consider an overlay K-tier HCN in which the BSs in all tiers are fully-loaded (i.e., no empty
queues) and have access to the same communication resources both in time and frequency. The BSs
in different tiers are differentiated mainly on the basis of their transmission powers, with Pk > 0
being the transmission power of a tier-k BS for k = 1, . . . , K. As is standard in stochastic geometric
modeling, it is assumed that BSs are distributed over the plane according to a general PPP with
differing spatial density among the tiers. Further, the signal propagation characteristics (including both
large-scale path-loss and small-scale fading) also vary from one tier to another. The details of BS
location processes and signal propagation are elaborated below.
(o)
(o)
We place a test user at an arbitrary point x(o) = x1 , x2
∈ R2 and consider signals coming
from all BSs in all tiers as the downlink AWI experienced by this test user. Since we focus on the
downlink analysis, we assume that the uplink and downlink do not share any common communication
resources. Therefore, the uplink interference can be ignored for the analysis of downlink AWI. This
setting is general enough to illuminate the effects of various network parameters such as transmission
powers and BS intensity in each tier on the distribution of the AWI seen by the test user.
B. BS Location Processes
The BS locations in tier-k, k = 1, . . . , K, independently form a spatial planar PPP Φk . Λk represents
the mean measure (alternatively called the intensity measure or spatial density) of the kth tier BSs.
We do not assume any specific functional form for Λk and hence do not restrict our attention only
to homogeneous PPPs. For each (Borel) subset S of R2 , Λk (S) gives us the average number of BSs
lying in S. We will assume that Λk is locally finite i.e., Λk (S) < ∞ for all bounded subsets S of R2 ,
and Λk (R2 ) = ∞, i.e., there is an infinite population of tier-k BSs scattered all around in R2 . For the
whole HCN, the aggregate BS location process, which is the superposition of all individual position
S
processes, is denoted by Φ = K
k=1 Φk .
For mathematical convenience, we also express Φk as a discrete sum of Dirac measures as Φk (S) =
P
(k)
∈ S ⊆ R2 , and zero otherwise. The level of AWI at
j≥1 δX (k) (S), where δX (k) (S) = 1 if Xj
j
j
x(o) from tier-k BSs depends critically on the distances between the points of Φk and x(o) . It is wellP
known from the theory of Poisson processes that the transformed process j≥1 δT X (k) is still Poisson
j
(on the positive real line) with mean measure given by Λk ◦ T −1 , where T (x) = x − x(o) 2 =
r
2
2
(o)
(o)
x1 − x1
+ x2 − x2
and T −1 (S) = {x ∈ R2 : T (x) ∈ S} for all S ⊆ R [44]. We will
R
assume that Λk ◦ T −1 has a density in the form Λk ◦ T −1 (S) = λk S µk (t)dt. Here, λk is a modeling
10
parameter pertaining to the kth tier, which can be interpreted as the BS intensity parameter, that will
enable us to control the average number of tier-k BSs whose distances from x(o) belong to S and
interfere with the signal reception at the test user.
In order to understand the model further, it is insightful to consider a homogeneous PPP with density
λk for tier-k BSs. In this case, Λk (S) is given by Λk (S) = λk · area (S) and the specific functional
form of µk (t) becomes equal to µk (t) = 2πt1{t≥0} , which can be easily obtained by transforming the
points of the homogeneous PPP over R2 to R by using the Euclidean metric as a mapping between
R2 and R. This is the PPP model for BS locations frequently considered in most previous papers, e.g.,
see [12]–[14]. In our model above, on the other hand, we allow the possibility of Λk (S) to depend
on the set S ⊆ R2 through an arbitrary functional form, which is not necessarily translation invariant,
as long as Λk constitutes a locally finite measure over the subsets of R2 . As a result, the functional
form of µk (t) can assume any shape from the set of functions over R, rather than being restricted to
the specific form µk (t) = 2πt1{t≥0} . In this regard, λk just functions as a tunable network parameter
in order for us to obtain a dense HCN deployment. An alternative approach to obtain a dense HCN
deployment is to vary the functional form of µk (t) itself. However, this approach does not result in a
single parameter representation to plot the downlink AWI distributions and HCN performance metrics.
The analytical results we derive in this paper hold for any functional form of µk (t) satisfying the
above mild conditions on Λk .
C. Signal Propagation Model
We model the large scale signal attenuation for tier-k, k = 1, . . . , K, by a bounded monotone nonincreasing path-loss function Gk : [0, ∞) 7→ [0, ∞). Gk asymptotically decays to zero at least as fast
as t−αk for some path-loss exponent αk > 2. To ensure the finiteness of AWI at the test user, we
require the relationship µk (t) = O (tαk −1−ǫ ) as t → ∞ to hold for some ǫ > 0.
The fading (power) coefficient for the wireless link between a BS located at point X ∈ Φ and
the test user is denoted by HX .1 The fading coefficients {HX }X∈Φ form a collection of independent
random variables (also independent of Φ), with those belonging to the same tier, say tier-k, having a
common probability distribution with density qk (h), h ≥ 0. The first, second and third order moments
(k)
(k)
(k)
of fading coefficients are assumed to be finite, and are denoted by mH , mH 2 and mH 3 , respectively,
1
For simplicity, we only assign a single fading coefficient to each BS. In reality, it is expected that the channels between a BS and all
potential receivers (intended or unintended) experience different (and possibly independent) fading processes. Our simplified notation
does not cause any ambiguity here since we focus on the outage and rate performance of the test user at a given arbitrary position in
R2 in the remainder of the paper.
11
for tier-k. We note that this signal propagation model is general enough that HX ’s could also be
thought to incorporate shadow fading effects due to blocking of signals by large obstacles existing
in the communication environment, although we do not model such random factors explicitly and
separately in this paper. Further, the model can also be extended to the load-aware analysis of HCNs,
multi-antenna communications and power-controlled BSs, as elaborated in Section VII of the paper.
D. Association Policy, Interference Power and Performance Measures
Association policy is a key mechanism that determines the outage and rate performance experienced
by the test user as it regulates the useful signal power as well as the interference power at the test
user. Hence, we first formally define it to facilitate the upcoming discussion.
K
K
2
Definition 1: An association policy A : Ω × R∞
+ × R+ × R+ 7→ R is a mapping that takes a
BS configuration ϕ ∈ Ω (i.e., a countable point measure), fading coefficients {Hx }x∈ϕ , transmission
K
power levels {Pk }K
k=1 and biasing coefficients {βk }k=1 as an input and determines the BS location to
which the test user is associated as an output.
For the HCN model explained above, the output of A is a random point X ⋆ = (X1⋆ , X2⋆ ) ∈ Φ
since the BS locations and fading coefficients are random elements. Biasing coefficients are important
design parameters to offload data from bigger cells to the smaller ones. Two other important random
quantities related to X ⋆ are the tier index A⋆ to which X ⋆ belongs and the distance between
(o)
(o)
∈ R2 , which is denoted by R⋆ = X ⋆ − x(o) 2 =
X ⋆ and the test user x(o) = x1 , x2
r
2
2
(o)
(o)
X1⋆ − x1
+ X2⋆ − x2
. Using these definitions along with considering all the signal impairments due to fading and path-loss, the total interference power at the test user is written as
X
Iλ =
PX HX GX (T (X)) ,
(1)
X∈Φ\{X ⋆ }
where λ = [λ1 , . . . , λK ]⊤ , and it is understood that PX = Pk and GX = Gk if X ∈ Φk . This
parametrization of AWI is chosen to emphasize the dependence of its distribution on the BS intensity
parameter λk of each tier.
SINR is the main performance determinant for the HCN model in question. Given an association
policy A, the SINR level experienced by the test user is equal to
SINRA =
PA⋆ HX ⋆ GA⋆ (T (X ⋆ ))
,
1
N0 + PG
Iλ
where N0 is the constant background noise power and PG ≥ 1 is a processing gain constant that
signifies the interference reduction capability, if possible, of the test user. We also let SNRk =
Pk
N0
to
12
denote the signal-to-noise ratio (SNR) for tier-k. Next, we define the main performance metrics used
to measure the HCN outage and rate performance.
Definition 2: For a target bit rate τ , τ -outage probability is equal to
Pr (τ -outage) = Pr {log (1 + SINRA ) < τ } .
Similarly, for a target outage probability γ, the outage capacity achieved by the test user under the
association policy A is equal to
Co (γ) = sup {τ ≥ 0 : Pr (τ -outage) ≤ γ} ,
which is the maximum data rate supported with outage probability not exceeding γ.2
Unlike the outage capacity in which (the instantaneous) SINRA is assumed to be an unknown constant
for the duration of channel coherence time, ergodic capacity is the average of (the instantaneous)
capacity that can be achieved by averaging over a large number of coherence time intervals leading
to Definition 3 below.
Definition 3: The ergodic capacity achieved by the test user under the association policy A is equal
to
Cerg = E[log (1 + SINRA )].
Another system performance metric we analyze in this paper is the area spectral efficiency (ASE).
In contrast to the outage capacity and ergodic capacity, the ASE metric captures the collective network
performance. For the sake of simplicity, we will study the ASE for when BSs in each tier are
homogeneously distributed. Its formal definition is given below.
Definition 4: Consider a K-tier HCN in which the BSs in tier-k are distributed according to
a homogeneous PPP with intensity λk for k = 1, . . . , K. Then, the ASE, measured in terms of
Nats/Sec/Hz/Area, achieved by the network is equal to
ASE (λ, γ) =
K
X
k=1
λk (1 − γk )Co (k, γk ) ,
where λ = [λ1 , . . . , λK ]⊤ is the vector of BS intensities, γ = [γ1 , . . . , γK ]⊤ is the vector of target
outage probabilities with γk , k = 1, . . . , K, being the target outage probability for tier-k BSs, and
Co (k, γk ) is the conditional outage capacity in tier-k, conditioned on the event that a mobile user is
connected to a tier-k BS.
2
The focus is more on the outage probability, rather than the outage capacity, in most previous work such as [12]–[14]. The outage
probability and outage capacity are related performance metrics for HCN analysis, with an extra step required to calculate the outage
capacity. The main reason for us to focus on the outage capacity in this paper is to have the same units with the ergodic capacity metric.
13
Since ASE measures the collective network performance, rather than the one observed at a point, it is
assumed that all BSs in the network serve a mobile user. Hence, the above definition with conditional
outage capacities makes sense. In the next section, we will first present our methodological approach
establishing the Gaussian approximation bounds to measure the proximity of the AWI distribution
to the normal distribution. These approximation results will be leveraged in Section V to obtain the
outage and rate performance of HCNs.
IV. G AUSSIAN A PPROXIMATION
FOR THE
AWI D ISTRIBUTION
In this section, we will establish the Gaussian approximation bounds under different spatial distribution assumptions for the standardized AWI distribution in the downlink of an HCN.3 Then, we will
present numerical examples validating our theoretical work.
A. Analytical Results
In this part, we present our theorems providing explicit upper and lower bounds on the AWI
distribution, which the test user experiences at a specific location in a K-tier HCN. These bounds
will clearly show the functional dependence between the downlink AWI distribution and a broad
range of network parameters such as transmission power levels, BS distribution over the plane and
signal propagation characteristics in each tier. We will also specialize these approximation results to
the commonly used homogeneous PPPs at the end of this part.
Theorem 1: For all x ∈ R,
where Ξ =
PK
k=1 PK
)
Iλ − E [Iλ ]
≤ x − Ψ(x) ≤ Ξ · c(x),
Pr p
Var [Iλ ]
(k) R
λk Pk3 m 3 0∞ G3k (t)µk (t)dt
31.935
H
3 , c(x) = min 0.4785, 1+|x|3 and Ψ(x) =
R
k=1
(
(k)
H2
λk Pk2 m
∞
0
G2k (t)µk (t)dt
2
which is the standard normal cumulative distribution function (CDF).
(2)
√1
2π
Rx
t2
e− 2 dt,
−∞
Proof: We only give a sketch of the proof. The complete details can be found in [45]. There are
several critical steps involved in the proof of this theorem. First, we show that the Laplace transform
LIλ (s) = E [exp (−sIλ )] for Iλ exists for all s ≥ 0 under our modeling assumptions in Section
(k)
(k)
III. Second, we introduce a sequence of auxiliary i.i.d. random variables U1,n , . . . , U Λ ,n for each
⌈ n,k ⌉
λk µk (t)
n ∈ N and for each tier k ∈ {1, . . . , K}, with a common PDF fk (t) = Λn,k 1{0≤t≤n} for tier-k,
3
We do not assume any interference protection or cancellation to derive the results in this section. In Section V, we show that the
results in this section can be modified in a straightforward manner to incorporate the effects of association policies on the K-tier HCN
performance.
14
Rn
where Λn,k = λk 0 µk (t) dt and ⌈.⌉ is the smallest integer greater than or equal to its argument. We
n
o∞
P
P⌈Λn,k ⌉ (k) (k)
(k)
(k)
(k)
is an i.i.d. collection
H
G
U
and
H
define In = K
I
,
where
I
=
P
n
k
k
i=1
i
i,n
i
k=1 n
i=1
of random variables with the common probability density function qk (h) for k = 1, . . . , K. We prove
that In converges in distribution to Iλ by showing that LIn (s) converges to LIλ (s) pointwise as n
tends to infinity. Third, we obtain a bound on the deviations of the distribution of the standardized
In −E[In ]
, from a standard normal distribution by using the form of the Berryversion of In , i.e., √
Var(In )
Esseen theorem given in [46]. Finally, by combining the results proven in the second and third steps,
we conclude the proof.
Measuring the KS distance, Theorem 1 provides us with an explicit expression for the deviations
between the standardized AWI and normal distributions. Several important remarks about this result
are in order. First, the standardized AWI can take negative values due to the centering operation,
which makes the deviations from the normal distribution bounded for negative values of x, as desired
in Theorem 1. Second, since a bounded path-loss model is used in each tier, the decay rate of the tails
of the AWI distribution also depends on the fading distribution parameters in each tier as indicated by
our bound in Theorem 1. A similar phenomenon was also observed in [22], [47]. We refer interested
readers to [22], [47] for a thorough comparison between bounded and unbound path-loss models.
Third, the scaling coefficient Ξ appearing in Theorem 1 is linked to the main network parameters such
as transmission power levels, distribution of BSs over the plane and signal propagation characteristics.
Starting with the BS intensity parameters λk , k = 1, . . . , K, we observe that the rate of growth of
the expression appearing in the denominator of Ξ is half an order larger than that of the expression
appearing in the numerator of Ξ as a function of λk . This observation implies that the derived Gaussian
approximation becomes tighter for denser deployments of HCNs. A formal statement of this result is
given in Lemma 1.
Fourth, the same functional form for c(x) was also obtained in papers [33], [34]. The reason is the
fact that the same form of the Berry-Esseen theorem from [46] is applicable to both single-tier and
multi-tier networks, but not the stochastic equivalence between a single-tier network and multi-tier
HCNs [42], [43]. Fifth, an approach alternative to our Gaussian approximation method in Theorem 1
would be to use characteristic functions and/or Laplace transforms in an inversion formula to derive
performance metrics for K-tier HCNs. However, this is not tractable for the HCN setup with nonhomogeneous PPPs, general path-loss models and general fading distributions considered in this paper.
For example, an important feature of our Gaussian approximation result in Theorem 1 is that it only
depends on the fading distributions through their second and third moments. On the other hand,
15
fading distributions appear in a more convoluted manner in the method of characteristic functions
and/or Laplace transforms that necessitates the computation of integrals with respect to both the mean
measure of the underlying BS processes and the per-tier fading distributions. Finally, the skewness of
the standardized AWI is related to its third moment, whose absolute value is upper bounded by Ξ in
Theorem 1, and hence a decreasing function of each BS intensity.
Lemma 1: The scaling coefficient Ξ appearing in the Gaussian approximation result in Theorem 1
is bounded above by Ξ ≤ √ δ for some finite positive constant δ.
kλk2
R
(k) R ∞
3 (k) ∞ 3
Proof: Let ak = Pk mH 3 0 Gk (t)µk (t)dt and bk = Pk2 mH 2 0 G2k (t)µk (t)dt. Then,
PK
ak λk
kλk2 kak2
Ξ = k=1
3 ≤
32
PK
PK
2
k=1 λk bk
k=1 λk bk
due to the Cauchy-Schwarz inequality. Further, we can lower-bound the sum in the denominator above
as
K
X
k=1
λk bk
! 23
≥
min bk
1≤k≤K
K
X
k=1
! 32
|λk |
3
≥ ǫ (kλk2 ) 2 ,
where the last inequality follows from the equivalence of all the norms in finite dimensional vector
spaces. Combining these two inequalities, we conclude the proof.
In the case of ultra-dense HCNs, the asymptotic behavior of the scaling
coefficient
Ξ (and the related
skewness behavior) is formally described by Lemma 1, i.e., Ξ = O
√1
kλk2
as kλk2 goes to infinity.
Further, the effect of an association policy on δ in Lemma 1 can be seen more clearly by considering
specific association policies (i.e., see Lemma 3 or Lemma 4). In particular, the association policies
considered in this paper modify the lower limits of the integrals in the numerator and denominator of
Ξ. The numerical value of the ratio of these integrals are usually small numbers a little greater than
one, i.e., see [33] for one specific example. Moreover, the existence of the multiplying function c(x)
makes our bounds further sharpened, and hence the effect of a specific association policies on our
Gaussian approximation stays limited.
Following a similar approach above, we can also see that changing transmission powers is not
as effective as changing BS intensity parameters to improve the Gaussian approximation bound
in Theorem 1. This is expected since the power levels are assumed to be deterministic (i.e., no
power control is exercised) and therefore they do not really add to the randomness coming from
the underlying spatial BS distribution over the plane and the path-loss plus fading characteristics
modulating transmitted signals. Another important observation we have in regards to the combined
effect of the selection of transmission powers per tier and the moments of fading processes in each
16
tier on the Gaussian approximation result in Theorem 1 is that our approximation bounds benefit from
the fading distributions with restricted dynamic ranges and the alignment of received AWI powers due
to fading and path-loss components. This observation is made rigorous through the following lemma.
R∞
R∞
(k)
Lemma 2: Let ak = λk 0 G3k (t)µk (t)dt, bk = λk 0 G2k (t)µk (t)dt and ck = Pk2 mH 2 . Then, the
scaling coefficient Ξ appearing in the Gaussian approximation result in Theorem 1 is bounded below
by
Ξ≥
1
kck2 kbk2
23 X
K
3
ak ck2 ,
k=1
with equality achieved if fading processes in all tiers are deterministic and the vectors b = [b1 , . . . , bK ]⊤
and c = [c1 , . . . , cK ]⊤ are parallel.
Proof: Using ak , bk and ck introduced above, we can write a lower bound for Ξ as
PK
PK
3 (k)
3 (k)
k=1 ak Pk mH 3
k=1 ak Pk mH 3
Ξ=
P
32
23 =
3
PK
K
ck
2
b
b
c
(kck
)
2
k=1 k kck2
k=1 k k
32 X
K
1
(k)
≥
ak Pk3 mH 3 .
kck2 kbk2
k=1
3
(k)
(k)
(k) 2
Using Jensen’s inequality, we also have mH 3 ≥ mH 2 . Using this lower bound on mH 3 in the
23 P
3
K
1
2
above expression, we finally have Ξ ≥ kck2 kbk2
k=1 ak ck .
In addition to the above fundamental properties of the scaling coefficient Ξ, it is also worthwhile
to mention that the Gaussian approximation bound derived in Theorem 1 is a combination of two
different types of Berry-Esseen bounds embedded in the function c(x). One of these bounds is a
uniform bound that helps us to estimate the standardized AWI distribution uniformly as
)
(
Iλ − E [Iλ ]
≤ x − Ψ(x) ≤ Ξ · 0.4785
Pr p
Var [Iλ ]
for all x ∈ R. On the other hand, the other one is a non-uniform bound that helps us to estimate the
tails of the standardized AWI distribution as
)
(
31.935
Iλ − E [Iλ ]
≤ x − Ψ(x) ≤ Ξ ·
Pr p
1 + |x|3
Var [Iλ ]
and decays to zero as a third order inverse power law.
Up to now, we considered general PPPs for the distribution of BSs in each tier. One simplifying
assumption in the literature is to assume that PPPs determining the locations of BSs are homogeneous.
In this case, µk (t) for all tiers is given by µk (t) = 2πt1{t≥0} , where 1{·} is the indicator function.
Using this expression for µk (t) in Theorem 1, we obtain the following approximation result for the
17
distribution of AWI when all BSs are homogeneously distributed over the plane according to a PPP
with differing BS intensity parameters λk from tier to tier.
Theorem 2: Assume that Φk is a homogeneous PPP with a mean measure given Λk (S) = λk ·area (S).
Then, for all x ∈ R,
where Ξ =
√1
2π
PK
(
k=1 PK
)
Iλ − E [Iλ ]
Pr p
≤ x − Ψ(x) ≤ Ξ · c(x),
Var [Iλ ]
(k) R
λk Pk3 m 3 0∞ G3k (t)tdt
31.935
H
3 , c(x) = min 0.4785, 1+|x|3 and Ψ(x) =
R
k=1
(k)
H2
λk Pk2 m
∞
0
G2k (t)tdt
2
which is the standard normal CDF.
(3)
√1
2π
Rx
Proof: The proof follows from Theorem 1 by replacing µk (t) with 2πt1{t≥0} .
When all network parameters are assumed to be the same, i.e., the same transmission power levels,
fading distributions and BS distributions for all tiers, the HCN in question collapses to a single tier
network. In this case, the Gaussian approximation result is given below.
(k)
Corollary 1: Assume Pk = P , µk (t) = 2πt1{t≥0} , Gk (t) = G (t), λk = λ, mH 2 = mH 2 and
(k)
mH 3 = mH 3 for all k = 1, . . . , K. Then, for all x ∈ R, we have
)
(
Iλ − E [Iλ ]
≤ x − Ψ(x) ≤ Ξ · c(x),
Pr p
Var [Iλ ]
where Ξ =
√1 √ 1
2π Kλ
mH 3
R∞
G3 (t)tdt
3 , and c(x) and Ψ(x) are as given in Theorem 1.
(mH 2 ) ( 0 G2 (t)tdt) 2
We note that this is the same result obtained in [33] as a special case of the network model studied
3
2
0
R∞
in this paper.
B. Numerical Verification of the Gaussian Approximation Results
In this part, we will illustrate the analytical Gaussian approximation results derived for the standardized AWI distribution in Section IV-A for a specific three-tier HCN scenario. To this end, we will use a
1
bounded power-law path-loss model G (t) = 1+t
α and the recently proposed stretched exponential path
loss model G (t) = exp −αtβ for dense cellular networks [8].4 The path-loss models are assumed
to be the same for all tiers with various values of the parameters α, β > 0, and hence we do not
index the path-loss models with the subscript k below. Similar conclusions continue to hold for other
path-loss models.
The BSs in each tier are distributed over the plane according to a PPP, with BS intensity parameters
given by λ1 = 0.1κ, λ2 = κ and λ3 = 5κ. Here, κ is our unit-less control parameter to control
4
With a slight abuse of notation, we adopted the same notation from [8] in order the represent the stretched exponential path-loss
model. Hence, β parameter in this path-loss model should not confused with the biasing coefficients introduced in Section III.
t2
e− 2 dt,
−∞
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
CDF
CDF
18
0.5
0.4
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
0
-6
6
-5
-4
-3
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.4
0.3
0.2
0.1
-4
-3
-2
-1
0
1
2
3
4
5
0
-6
6
2
3
4
5
6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
Standardized Interference Power
1
1
0.8
0.8
0.6
0.6
CDF
CDF
1
0.1
-5
Standardized Interference Power
0.4
0.2
0.4
0.2
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
0
-6
-5
-4
Standardized Interference Power
Fig. 1.
0
0.4
0.2
0
-6
-1
0.5
0.3
0
-6
-2
Standardized Interference Power
CDF
CDF
Standardized Interference Power
-3
-2
-1
0
1
2
3
4
5
6
Standardized Interference Power
Gaussian approximation bounds for the standardized AWI CDFs (top and middle figures). Comparison of the simulated
standardized AWI CDFs with the standard normal CDF (bottom figures). Parameter selection for the stretched exponential distribution
is taken from [8]. Rayleigh fading with unit mean power is assumed.
the average number of BSs interfering with the signal reception at the test user. The test user is
assumed to be located at the origin without loss of any generality. The random fading coefficients
in all tiers are assumed to be independent and identically distributed random variables, drawn from
a Rayleigh distribution with unit mean power gain. Our results are qualitatively the same for other
fading distributions such as Nakagami and Rician fading distributions. The transmission power levels
are set as P1 = 4P2 = 16P3 , where P2 is assumed to be unity.
We start with the case of homogeneously distributed BSs over the plane in Fig. 1. The nonhomogeneous case is analyzed by introducing no-BS zones in Fig. 2. We only plot the CDF of
AWI in Fig. 1 since Theorem 1 measures the deviations between the CDFs of standardized AWI and
a standard normal random variable. In the top and middle figures in Fig. 1, we present the upper and
lower bounds for the deviations between the standardized AWI and normal distributions, i.e., we plot
19
the expressions Ψ(x)+Ξ·c(x) and Ψ(x)−Ξ·c(x) appearing in Theorem 1, for the two path-loss models
with a variety of κ values. Two different regimes are apparent in thesefigures. For the
moderate values
at which we want to estimate the CDF of standardized AWI, i.e., Pr
Iλ −E[Iλ ]
√
≤x
Var[Iλ ]
with moderate x
values, our uniform Berry-Esseen bound, which is Ξ · 0.4785, provides better estimations for the AWI
distribution. On the other hand, for absolute
values larger
than 4.0359 at which we want to estimate
Iλ −E[Iλ ]
≤ x with |x| larger than 4.0359, our non-uniform
the CDF of standardized AWI, i.e., Pr √
Var[Iλ ]
31.935
Berry-Esseen bound, which is Ξ · 1+|x|
3 , is tighter. These figures also clearly demonstrate the effect of
the BS intensity parameters λk on our Gaussian approximation bounds. As suggested by Lemma 1, the
KS distance between the standardized AWI and normal distributions approaches zero at a rate at least
as fast as √ 1
. Further, even if all BS intensity parameters are fixed, the distance between the upper
and lower bounds in Theorem 1 disappears at a rate O |x|−3 as |x| → ∞ due to the non-uniform
kλk2
bound. Lastly, our bounds appear to be tighter for the stretched exponential path-loss model. This is
mainly because of the path-loss model dependent constants appearing in Theorem 1.
When we compare the top left-hand side and right-hand side figures for the bounded power-law
path-loss model in Fig. 1, we observe a better approximation behavior for smaller values of the
path-loss exponent α. This is again because of the path-loss model dependent constants appearing
in Theorem 1. For this particular choice of the path-loss model and BS distribution over the plane,
our approximation results benefit from small values of the path-loss exponent, although the difference
between them becomes negligible for moderate to high values of κ.
We also performed Monte-Carlo simulations to compare simulated standardized AWI distributions
with the normal distribution for 104 random BS configurations. For these simulations, we only used
the more conservative bounded power-law path-loss model. The results are expected to be even better
for the stretched exponential path-loss model. Briefly, the bottom figures in Fig. 1 provide further
numerical evidence for the Gaussian approximation of AWI in HCNs. Surprisingly, there is a good
match between the simulated standardized AWI distribution and the standard normal CDF even for
sparsely populated HCNs, i.e., κ = 1, which is better than what is predicted by our upper and lower
Gaussian approximation bounds.
Finally, in addition to the numerical verification of the Gaussian approximation bounds in Theorem
1 for homogeneous PPPs, we also performed a similar numerical analysis for HCNs with nonhomogeneous BS deployments in Fig. 2. In this figure, we provide numerical results illustrating the
derived Gaussian approximation bounds for two different types of non-homogeneous spatial distributions for the locations of BSs. The non-homogeneous BS deployment is obtained by allowing the
1
1
0.8
0.8
0.6
0.6
CDF
CDF
20
0.4
0.2
0
-6
0.4
0.2
-5
-4
-3
-2
-1
0
1
2
3
4
5
0
-6
6
-5
-4
-3
1
1
0.8
0.8
0.6
0.6
0.4
0.2
0
-6
0
1
2
3
4
5
6
0.2
-5
-4
-3
-2
-1
0
1
2
3
4
5
0
-6
6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
Standardized Interference Power
1
1
0.8
0.8
0.6
0.6
CDF
CDF
-1
0.4
Standardized Interference Power
0.4
0.2
0
-6
-2
Standardized Interference Power
CDF
CDF
Standardized Interference Power
0.4
0.2
-5
-4
-3
-2
-1
0
1
2
Standardized Interference Power
3
4
5
6
0
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
Standardized Interference Power
Fig. 2. Gaussian approximation bounds for the standardized AWI CDFs for non-homogeneous BS deployments (top and middle figures).
Comparison of the simulated standardized AWI CDFs with the standard normal CDF for non-homogeneous BS deployments (bottom
figures). Parameter selection for the stretched exponential distribution is taken from [8]. Rayleigh fading with unit mean power is
assumed.
possibility of BSs not being deployed in some certain regions of the network domain, possibly due to
some potential environmental limitations. More specifically, we either consider a guard zone of radius
5 unit distances around the origin or an annulus with two different values for inner radius, i.e., 2 and
10 unit distances, and two different values for outer radius, i.e., 20 and 50 unit distances. The BSs in
all tiers are assumed to be homogeneously distributed over the plane according to a PPP with κ = 50,
without allowing any BS either inside the guard zone or inside the annulus region.
As Fig. 2 indicates, our Gaussian approximation bounds continue to hold in the considered nonhomogeneous BS deployment scenarios when the AWI power is measured at the origin. In particular,
the stretched exponential path-loss model leads to tighter bounds than the bounded power-law pathloss model does as in the case of homogeneously distributed BSs. Further, different BS deployment
21
scenarios seem to have little impact on the bounds for the stretched exponential path-loss model. For
the bounded power-law path-loss model, on the other hand, the Gaussian approximation bounds for
the AWI distribution becomes tighter than those for the case of homogeneously deployed BSs when
there is a guard zone around the origin in which no BSs are allowed to lie. Considering the annulus
region in which no BSs are allowed to lie for the bounded power-law path-loss model, the Gaussian
approximation bounds for the AWI distribution are as good as those obtained for the homogeneously
deployed HCN scenario with little impact of the annulus radii on the bounds for the steeper path-loss
function. Lastly, the simulation results presented for the bounded power-law path-loss model in the
bottom figures in Fig. 2 corroborate the Gaussian approximation bounds illustrated in the upper figures.
In the following section, we will make use of the analytical findings in this section to obtain various
performance metrics in K-tier HCNs under general settings. In the light of these metrics, we will also
gain insights into the impact of network densification on the outage and rate performance of HCNs
under specific association policies including both biased and non-biased BS selection strategies.
V. O UTAGE
AND
R ATE P ERFORMANCE
OF
K- TIER HCN S
In this section, we will derive the performance bounds on the HCN capacity metrics, i.e., outage
capacity, ergodic capacity and ASE, under two specific association policies: (i) a generic association
policy and (ii) biased average received signal strength (BARSS) association policy. However, it should
be noted that the analytical approach developed below is general enough for any association policy
that preserves the Poisson distribution property for BS locations given the information of X ⋆ . The
validity and utility of our analytical results will be numerically illustrated by simulations in Section
VI
A. Generic Association Policy
We start our discussion with the generic association policy. The generic association policy is the
policy under which the test user is connected to a BS in tier-k at a (deterministic) distance r > 0,
and the locations of the rest of the (interfering) BSs in each tier form a non-homogeneous PPP over
R
R2 \B x(o) , di with mean measure satisfying the functional form Λi ◦ T −1 (S) = λi S µi (t)dt given
this connection information for i = 1, . . . , K, where B x(o) , di is the planar ball centered at x(o)
with radius di ≥ 0. B x(o) , di can be thought to signify an exclusion region around the test user due
to operation of the HCN network protocol stack.
The study of the generic association policy, which may seem a little artificial at first sight, will set
the stage for us to analyze both outage and ergodic capacity performance of the BARSS association
22
policy later in this section. The following lemma establishes the Gaussian approximation bounds for
the distribution of (standardized) aggregate interference Iλ at the test user under the generic association
policy by specializing Theorem 1 to this case.
Lemma 3: Under the generic association policy described above,
)
(
Iλ − E [Iλ ]
≤ x − Ψ(x) ≤ Ξ · c(x)
Pr p
Var [Iλ ]
R x − t2
31.935
1
√
2
for all x ∈ R, where Ψ(x) = 2π −∞ e dt which is the standard normal CDF, c(x) = min 0.4785, 1+|x|3 ,
(i) R
PK
λi Pi3 m 3 d∞ G3i (t)µi (t)dt
H
i
and Ξ = i=1 P
3 .
R
K
i=1
(i)
H2
λi Pi2 m
∞
di
G2i (t)µi (t)dt
2
Since the outage and ergodic capacity metrics given in Definition 2 and Definition 3 heavily depend
on the level of AWI at the test user, the above Gaussian approximation bound plays a key role to
obtain performance upper and lower bounds on the outage and rate performance of HCNs. Below, we
start with the outage capacity metric.
Theorem 3: Let ζk (h, τ, r) =
Pk
hGk (r)
−SNR−1
k
eτ −1
√
PG−E[Iλ ]
Var[Iλ ]
. Then, Pr (τ -outage) under the generic asso-
ciation policy is bounded above and below as
1 − E Vk+ (Hk , τ, r) ≤ Pr (τ -outage) ≤ 1 − E Vk− (Hk , τ, r) ,
where Hk is a generic random variable with PDF qk , and the functions Vk+ and Vk− are given as
Vk+ (h, τ, r) = min {1, Ψ (ζk (h, τ, r)) + Ξ · c (ζk (h, τ, r))} 1
h≥
SNR−1 (eτ −1)
k
Gk (r)
Vk− (h, τ, r) = max {0, Ψ (ζk (h, τ, r)) − Ξ · c (ζk (h, τ, r))} 1
h≥
SNR−1 (eτ −1)
k
Gk (r)
.
(4)
and
(5)
Proof: Please see Appendix A.
Using the bounds on Pr (τ -outage), we can bound Co (γ) for the generic association policy as below.
Theorem 4: Co (γ) under the generic association policy is bounded above and below as
and
Co (γ) ≤ sup τ ≥ 0 : 1 − E Vk+ (Hk , τ, r) ≤ γ
Co (γ) ≥ sup τ ≥ 0 : 1 − E Vk− (Hk , τ, r) ≤ γ .
Proof: The proof follows from that the upper (lower) bound on the outage probability crosses the
target outage probability γ earlier (later) than Pr (τ -outage) as τ increases.
An important high level perspective on the detrimental effects of the network interference on the
HCN outage performance can be obtained if we study the outage capacity bounds given in Theorem
23
4 as a function of λ, e.g., for homogeneous PPPs. At each fading state Hk = h, it can be shown
as kλk2
that the outage capacity scales with the BS intensity parameters according to Θ kλk−1
2
grows to infinity. This observation is different from the scale-invariance property of SINR statistics
with BS intensity observed in some previous work such as [13], [14], [30], [42]. The main reason is
that the increase in Iλ with denser HCN deployments cannot be counterbalanced by an increase in the
received power levels with bounded path-loss models. That is, the benefits of shorter communication
distances are eliminated by an increase in network interference levels in dense HCNs when a bounded
path-loss model is used to characterize large-scale wireless propagation losses. From an HCN design
perspective, this result implies that it is imperative to set BS intensities at each tier appropriately for
the proper delivery of data services with minimum required QoS to the end users.
Unlike the outage capacity metric, the ergodic capacity is more suitable for delay insensitive data
traffic and obtained by averaging instantaneous data rates over long time intervals. In the following
theorem, the bounds on the ergodic capacity are given, where no channel state information (CSI) is
assumed at the transmitter side.
Theorem 5: Cerg under the generic association policy is bounded above and below as
Z ∞
Z ∞
−
E Vk (Hk , τ, r) dτ ≤ Cerg ≤
E Vk+ (Hk , τ, r) dτ .
0
0
Proof: Please see Appendix B.
B. BARSS Association Policy
Now, we study the HCN outage and rate performance under the BARSS association policy, in which
the test user associates itself to the BS X ⋆ given by
X ⋆ = arg max βX PX GX (kXk2 ) ,
X∈Φ
where it is understood that βX = βk if X ∈ Φk . Consider the event Ek (r) that A⋆ = k and R⋆ = r,
i.e., Ek (r) is the event that the test user is associated with a tier-k BS at a distance r under the
BARSS association policy. Then, the locations of BSs in tier-i form a non-homogeneous PPP over
(k)
(k)
−1 βk Pk
2
(o)
R \B x , Qi (r) given the event Ek (r) for i = 1, . . . , K, where Qi (r) = Gi
Gk (r) , and
β i Pi
5
G−1
i (y) = inf {x ≥ 0 : Gi (x) = y} if y ∈ [0, Gi (0)] and zero otherwise. This observation puts us back
into the generic association policy framework, and the derivation of the bounds for the conditional
outage probability/capacity and ergodic capacity on the conditioned event Ek (r) proceeds as before.
5
Here, for the simplicity of analysis, we assume that path-loss functions are continuous.
24
Averaging over the event Ek (r), we obtain bounds on the unconditional outage probability and capacity
metrics. To this end, we need the following lemmas.
Lemma 4: Under the BARSS association policy described above, for all x ∈ R,
)
(
Iλ − E [Iλ ]
≤ x Ek (r) − Ψ(x) ≤ Ξk (r) · c(x),
Pr p
Var [Iλ ]
where Ξk (r) =
3.
PK
(i)
H3
λi Pi3 m
i=1 P
K
R∞
(k)
Q
(r)
i
2 (i)
i=1 λi Pi mH 2
G3i (t)µi (t)dt
R∞
G2 (t)µi (t)dt
(k)
Q
(r) i
i
3
2
, and c(x) and Ψ(x) are as defined in Lemma
We note that this is almost the same result as in Lemma 3, except for a small change in the definition
of the constant Ξ to show its dependence on the conditioned event Ek (r). In order to achieve averaging
over the event Ek (r), we need to know the connection probability to a tier-k BS and the conditional
PDF of the connection distance given that the test user is associated with a tier-k BS. To this end, we
first obtain the connection probability to a tier-k BS, which we denote by p⋆k , Pr {A⋆ = k}, in the
next lemma for non-homogeneous PPPs.
Lemma 5: The probability that the test user is associated with a tier-k BS is
Z ∞Y
K
(k)
⋆
pk =
fRk (u)du,
exp −Λi B x(o) , Qi (u)
0
(6)
i=1
i6=k
(k)
β k Pk
G
(u)
, βk is the biasing factor for tier-k BSs and Λk (·) is the mean
where Qi (u) = G−1
k
β i Pi
i
(k)
(k)
measure of Φk , B x(o) , Qi (u) is the ball in R2 centered at x(o) with radius Qi (u), and fRk (u) =
(o)
e−Λk (B(x ,u)) d Λ B x(o) , u is the nearest tier-k BS distance distribution for the test user.
du
k
Proof: Please see Appendix C.
Below, we focus on an important special case in which BS locations follow a homogeneous PPP
in each tier, which leads to Lemma 6 below.
Lemma 6: Let a0 = 0, aK+1 = +∞ and ai =
β i Pi
G (0)
β k Pk i
enumeration of ai ’s in descending order, i.e., aπ(i) ≥ aπ(i+1)
for i = 0, . . . , K. Then, p⋆k is given by
K Z rj
X
⋆
pk = 2πλk
u exp −π
j=1
rj−1
λk u 2 +
for i ∈ {1, . . . , K} \ {k}. Let π(i) be an
a
for i = 0, . . . , K − 1. Let ri = G−1
π(i)
k
j−1
X
i=1
2
(k)
λπ(i) Qπ(i) (u)
!!
du.
(7)
Proof: Please see Appendix D.
Several important remarks are in order regarding Lemma 6. The integration in (7) is with respect
to the nearest tier-k BS distance distribution to which the test user is associated. Hence, the BSs in
some tiers are inactive with regard to contributing to the association probability for different ranges of
25
the nearest distance from Φk to the origin, which is why we divide the integration limits into disjoint
intervals from rj−1 to rj for j = 1, . . . , K. This behavior is different than that observed in [14], which
is again a manifestation of the bounded nature of the path-loss model. In the next lemma, we derive
the PDF of R⋆ given A⋆ = k for a non-homogeneous PPP in each tier.
Lemma 7: The PDF of R⋆ given A⋆ = k is
K
1 Y
(k)
exp −Λi B x(o) , Qi (u)
fRk (u),
fk (u) = ⋆
pk i=1
(8)
i6=k
where fRk (u) is as given in Lemma 5.
Proof: Please see Appendix E.
In the next lemma, we again consider the important special case of homogeneous PPPs.
Lemma 8: Let a0 = 0, aK+1 = +∞ and ai =
β i Pi
G (0)
β k Pk i
enumeration of ai ’s in descending order, i.e., aπ(i) ≥ aπ(i+1)
for i ∈ {1, . . . , K} \ {k}. Let π(i) be an
aπ(i)
for i = 0, . . . , K − 1. Let ri = G−1
k
for i = 0, . . . , K. Then, the conditional PDF fk (u) of R⋆ given A⋆ = k is given as
K
2πλk X
fk (u) = ⋆
u exp −π
pk j=1
λk u 2 +
j−1
X
i=1
2
(k)
λπ(i) Qπ(i) (u)
!!
1{u∈[rj−1 ,rj )} .
(9)
Proof: Please see Appendix F.
The conditional connection PDF fk (u) given in (9) can be significantly simplified for small numbers
of tiers. A reduced expression for one particular but important case of a two-tier HCN is given by the
following corollary.
β 1 P1
Corollary 2: Assume K = 2, β1 P1 G1 (0) ≤ β2 P2 G2 (0) and u⋆ = G−1
G
(0)
. Then,
2
β 2 P2 1
2
2πλ1
(1)
2
1{u≥0}
f1 (u) = ⋆ u exp −π λ1 u + λ2 Q2 (u)
p1
and
2
2πλ2
2πλ2
(2)
2
2
1{u≥u⋆ } .
f2 (u) = ⋆ u exp −πλ2 u 1{u<u⋆ } + ⋆ u exp −π λ2 u + λ1 Q1 (u)
p2
p2
Using these preliminary results, the performance bounds on the outage probability, outage capacity,
ergodic capacity and ASE under the BARSS association policy are given in theorems below.
Theorem 6: Let Vbk± (h, τ, r) be defined as in (4) and (5), respectively, by replacing Ξ with Ξk (r)
given in Lemma 4. Then, Pr {τ -outage} under the BARSS association policy is bounded below and
above as
Pr {τ -outage} ≥ 1 −
K
X
k=1
p⋆k
Z
∞
0
h
i
+
b
fk (r)E Vk (Hk , τ, r) dr
26
and
K
X
Pr {τ -outage} ≤ 1 −
Z
p⋆k
k=1
∞
0
h
i
−
b
fk (r)E Vk (Hk , τ, r) dr.
Proof: The proof follows from calculating these bounds for Pr {τ -outage | Ek (r)} using Theorem
3, and then averaging them by using (6) and (8).
Theorem 7: Co (γ) under the BARSS association policy is bounded above and below as
(
)
Z ∞
K
h
i
X
Co (γ) ≤ sup τ ≥ 0 : 1 −
p⋆k
fk (r)E Vbk+ (Hk , τ, r) dr ≤ γ
0
k=1
and
(
Co (γ) ≥ sup τ ≥ 0 : 1 −
K
X
p⋆k
Z
∞
0
k=1
h
i
fk (r)E Vbk− (Hk , τ, r) dr ≤ γ
)
.
Proof: The proof follows from that the upper (lower) bound on the outage probability crosses the
target outage probability γ earlier (later) than Pr (τ -outage) as τ increases.
Similar to Theorem 5, the performance bounds for achievable ergodic capacity are given in the next
theorem.
Theorem 8: The ergodic capacity achievable under the BARSS association policy is bounded above
and below as
Cerg ≤
XK
p⋆k
Cerg ≥
XK
p⋆k
and
k=1
k=1
Z
∞
0
Z
0
∞
Z
∞
0
Z
0
∞
E[Vbk+ (H, τ, r)] dτ fk (r) dr
E[Vbk− (H, τ, r)] dτ fk (r) dr,
where Vbk+ (Hk , τ, r) and Vbk− (Hk , τ, r) are defined as in Theorem 6.
PK ⋆ R ∞
Proof: The ergodic capacity can be written as Cerg =
k=1 pk 0 Cerg (k, r)fk (r)dr, where
Cerg (k, r) is the conditional ergodic capacity given that the test user is connected to a BS at a
distance r and located in tier-k under the BARSS association policy. Then, similar to Appendix B,
the proof follows from calculating these bounds for 1 − Pr {τ -outage | Ek (r)} using Theorem 3, and
then averaging them by using (6) and (8) for non-homogeneous PPPs, or by using (7) and (9) for
homogeneous PPPs, respectively.
Available spectrum for wireless networks is often very limited. Hence, it is of prime importance to
investigate the ASE of a multi-tier HCN defined as the sum of the achievable bit rates per second per
hertz per unit area. The formal definition is given in Definition 4. In the next theorem, we provide upper
and lower bounds for the ASE metric in a K-tier HCN, which requires the calculation of conditional
outage capacity in each tier. For this calculation, we allow the possibility of having different target
outage probabilities for different tiers in order to provide a flexibility of setting a balance between
27
link reliability and rate. The target outage probability for tier-k, k = 1, . . . , K, will be represented
by γk ∈ (0, 1) in Theorem 9. We only consider homogeneous PPPs in Theorem 9 to avoid technical
complexities arising in the case of non-homogeneous PPPs. Further, since the ASE metric characterizes
the collective network performance, rather than the one observed at a specific point as above, it is
assumed that all BSs are serving a user.
Theorem 9: The ASE under the BARSS association policy is bounded above and below as
XK
ASE (λ, γ) ≤
λk (1 − γk ) Co+ (k, γk )
k=1
and
ASE (λ, γ) ≥
∆
where Co+ (k, γk ) = sup τ ≥ 0 : ρ+
k ≤ γk
XK
k=1
λk (1 − γk ) Co− (k, γk ),
∆
and Co− (k, γk ) = sup τ ≥ 0 : ρ−
k ≤ γk
are conditional
∆
T
outage capacity upper and lower bounds in tier-k, λ = [λ1 , . . . , λK ] , γ = [γ1 , . . . , γK ]T , and ρ+
k =
h
i
h
i
R∞
R
∆
∞
b−
1 − 0 fk (r) E Vbk+ (Hk , τ, r) dr and ρ−
k = 1 − 0 fk (r) E Vk (Hk , τ, r) dr are conditional outage
probabilities given that a mobile user is connected to a tier-k BS, Vb + (Hk , τ, r) and Vb − (Hk , τ, r) are
k
k
given as in Theorem 6 specialized to the homogeneous PPPs, and fk (r) is given in (9).
Proof: The proof easily follows from Theorems 6 and 7.
VI. S IMULATIONS
AND
N UMERICAL R ESULTS
In this part, we present our simulation results illustrating the upper and lower bounds on the HCN
capacity metrics derived in Section V. We will focus only on the homogeneous PPPs in this section.
The main reason for this choice is to be able to relate the obtained HCN performance metrics to
the conclusions drawn in previous papers since homogeneous PPPs are the de facto BS deployment
model in almost all previous work. Further, without having detailed data for the BS deployment,
homogeneous PPPs are the most appropriate and widely accepted model for illustrating the derived
analytical results.
In particular, we will investigate Co (γ), Cerg and ASE (λ, γ) under the BARSS association policy. N0
is set to zero and all fading coefficients are independently drawn from Nakagami-m distribution with
unit mean power gain and m = 5. The path-loss function is conservatively taken to be G(t) =
1
1+tα
for
all tiers. Considering the Gaussian approximation bounds illustrated in Fig. 1, the stretched exponential
path-loss model is expected to lead to tighter performance bounds. The transmission powers are set
as P1 = 10P2 = 50P3 , while we set BS locations according to homogeneous PPPs with intensities
given as λ1 = 0.1κ, λ2 = κ and λ3 = 5κ. Here, κ is our (unit-less) control parameter to control
the average number of BSs per unit area. For the 2-tier scenario, only {Pk , λk }2k=1 are considered.
28
1
Outage Capacity and Bounds
(Nats Per Second Per Hertz)
1.2
Outage Capacity and Bounds
(Nats Per Second Per Hertz)
Upper Bound
Lower Bound
Simulated Rate
Heuristic Rate
1.4
1
0.8
α = 3.3
0.6
0.4
α = 2.7
0.2
1
3
6
9
12
HCN Intensity Factor, κ
15
Upper Bound
Lower Bound
Simulated Rate
Heuristic Rate
0.9
0.8
0.7
0.6
α = 3.3
0.5
0.4
0.3
α = 2.7
0.2
18
0.1
1
3
6
9
12
HCN Intensity Factor, κ
15
18
Fig. 3. Upper and lower bounds on Co (γ) for 2- and 3-tier HCNs in the left-hand side and right-hand side figures, respectively.
The target outage probability is 0.15 for Fig. 3 and PG is set to 25 for both figures. For the sake of
simplicity, we assume that network layer queues at BSs are fully-loaded, an extension of which to the
lightly loaded case is elaborated in Section VII and will be considered as a future work [48].6
We plot the bounds in Theorem 7 on Co (γ) for 2-tier and 3-tier HCNs as a function of κ in Fig. 3.
Two different values of α are used. As this figure shows, both upper and lower bounds approximate
Co (γ) within 0.06 Nats/Sec/Hz for α = 2.7 and within 0.15 Nats/Sec/Hz for α = 3.3 in the 2-tier
scenario. They are tighter for the 3-tier scenario due to denser HCN deployment. The heuristic rate
curve, which is the arithmetic average of the upper and lower bounds, almost perfectly track Co (γ)
for all cases considered in Fig. 3.
An interesting observation is the monotonically decreasing nature of Co (γ) with κ. This is in
-type scaling behavior of outage capacity in Section
accordance with the discussion on the Θ kλk−1
2
V. Hence, we cannot improve the downlink data rates indefinitely in an HCN by adding more BS
infrastructure. We must either mitigate interference more efficiently or find the optimum BS intensity
per tier maximizing delivered data rates per unit area.
We plot Cerg and the corresponding bounds given in Theorem 8 for the 2-tier HCN scenario as a
function of κ in Fig. 4. We set the path-loss exponent α to 3. For moderate values of κ, we observe
that while our upper bound approximates Cerg within one Nats/Sec/Hz, the lower bound is closer than
0.3 Nats/Sec/Hz. Our bounds become very close to the simulated rate for moderate to high values of
6
Our parameter selections in this section are deliberately different than those in Section IV in order to illustrate that our results
continue to hold broadly for any set of network parameter combinations.
Ergodic Capacity
(Nats Per Second Per Hertz)
3
Upper Bound
Simulated Capacity
Lower Bound
2.5
2
1.5
1
0.5
1
3
6
9
12
15
HCN Intensity Factor, κ
18
21
Area Spectral Efficiency
(Nats per Second Per Hertz Per Unit Area)
29
1.6
1.4
1.2
1
0.8
Simulated ASE
Heuristic ASE
Upper Bound
Lower Bound
Macro Contrb.
Micro Contrb.
0.6
0.4
0.2
0
1
3
6
9
12
Intensity Factor, κ
15
18
Fig. 4. Change of ergodic capacity for various values of κ.
Fig. 5. Change of area spectral efficiency for various values
λ1 = 0.1κ, λ2 = κ. α, β1 and β2 are set to 3, 1 and 1,
of κ. λ1 = 0.1, λ2 = κ, βratio = β2 /β1 = 3 and α = 3.
respectively.
κ, and especially the gaps among the bounds and the Cerg become negligibly small for high values of
κ.
When we compare the bounds on outage and ergodic capacities depicted in Figs. 3 and 4, an
important observation we have is the upper bound for the ergodic capacity being looser than the one
for the outage capacity in sparse HCNs. The main reason for this behavior is the nature of Gaussian
approximation bounds appearing in Theorem 1. In particular, the outage probability, and hence the
outage capacity, depends on the tail of the AWI distribution. Due to the existence of the function
c(x) in the Gaussian approximation bounds in Theorem 1, the derived bounds on the tails of the
standardized downlink AWI distribution are tight even for the sparsely deployed HCNs. The ergodic
capacity calculations, on the other hand, require an averaging over the entire range of AWI values
and this range includes the interference values for which our bounds are not as tight as for higher
interference values. This leads to the observed behavior of obtained bounds being a little looser for the
ergodic capacity, especially the upper bound, when compared to those obtained for outage capacity.
Finally, Fig. 5 depicts the changes of ASE for a 2-tier HCN scenario as a function of κ, where BS
intensities are set as λ1 = 0.1 and λ2 = κ. Note that we fix λ1 at 0.1 to see how ASE changes by
adding more unplanned infrastructure to the network as λ2 grows. Biasing coefficient for tier-1, i.e.,
β1 , is assumed to be unity while the biasing value β2 for tier-2 is set to 3. Target outage probabilities
in both tiers are assumed to be identical, i.e., γk = γ, and taken as 0.15 for maintaining the same link
reliability in each tier.
30
We firstly observe that our upper and lower bounds again accurately characterize the simulated ASE
curve. Further, the heuristic ASE curve, which is taken to be the arithmetic average of the upper and
lower bounds, almost perfectly tracks the simulated ASE. We note three different forms of behavior
of ASE as a function of κ. For sparse to moderate values of κ, i.e., between 1 and 6, ASE increases
rapidly due to efficient utilization of the spectrum which is made available by tier-2 BSs. For κ from
moderate to high values, i.e., between 6 and 12, improvement in ASE starts to slow down due to
the pressure of growing AWI. For κ values higher than 12, ASE almost stops improving due to the
overwhelming growth in AWI. Especially, the contribution of macro BSs to the overall ASE is much
more negatively affected than that of micro BSs by the elevated AWI levels in the dense HCN regime,
as illustrated by the macro and micro contribution curves. These observations point out that significant
ASE gains can be achievable through biasing along with applying an efficient interference management
and suppression in hyper-dense multi-tier HCNs. Otherwise, the network density should be carefully
adjusted to maintain target ASE.
VII. D ISCUSSION : S COPE
OF THE
PAPER , E XTENSIONS
AND
L IMITATIONS
OF THE
P ROPOSED
A PPROACH
Our main aim in this paper is to provide tight Gaussian approximation bounds for AWI in HCNs
and to illustrate the utility of these bounds for the system-level HCN performance analysis. Hence, we
consider a simple but practically relevant system model in Section III for the sake of avoiding more
involved analysis in Section IV and Section V. In this section, on the other hand, we discuss some
notable extensions of our baseline model and the required changes for these extensions.
A. Load-Aware Modeling and Analysis
Our network model in Section III assumes that all the BSs are fully-loaded and have access to the
same communication resources both in time and frequency. However, in reality, small cells have fewer
users with lighter traffic loads [48]. This observation implies that the fully-loaded network scenario
can only provide lower bounds for the actual network performance. Hence, an important extension of
our results in this paper lies in the design and analysis of load-aware HCNs [48]–[50].
In particular, the load-aware analysis of HCNs depends primarily on the thinning of the underlying
BS location processes, which can be categorized into two classes of (i) uncorrelated thinning [48],
[49] and (ii) correlated thinning [50]. For the case of uncorrelated thinning, it is important not to turn
off the serving BS and the analysis of the total received power at a typical point plays an important
role to discover the outage probability in load-aware HCNs [48]. In this regard, our main Gaussian
31
approximation result in Theorem 1 can be directly applied to the conditionally thinned BS location
processes, without turning off the serving BS, in order to obtain the outage probability and other
related performance measures. Our framework is flexible enough to allow the consideration of both
location independent thinning [48] and location dependent thinning [49] in a similar fashion.
For the case of correlated thinning, on the other hand, the thinning probability depends on the
configuration of all BSs [50]. Therefore, it is not clear if the thinned BS location processes will still
maintain their Poisson property. If they do, Theorem 1 can still be applied to approximate the AWI
distribution. Otherwise, our approach in this paper cannot be used since the Poisson assumption is
critical for the proof of Theorem 1.
B. Multiple-Input Multiple-Output Communications
In addition to network densification, another communication strategy that features prominently in
IEEE 802.11e WiMAX and 3GPP LTE-A standards is multiple-input multiple-output (MIMO) communications [35], [41], [51]–[53]. Multiple antenna techniques offer both multiplexing and diversity
gains in order to improve spectral efficiency [54]. Two important modifications are needed in order to
accommodate MIMO techniques in our baseline model in Section III. The first is the consideration of
beamforming, which will bring additional degrees-of-freedom for communications over single-input
single-output (SISO) systems. The second is the inter-antenna interference.
More specifically, assuming that a BS belonging to tier-k has Mk transmission antennas and serves
Nk single-antenna mobile users,7 the transmitted signal from a BS located at X ∈ Φk can be
represented as
sX =
p
Pk
Nk
X
bX,i sX,i ,
i=1
Nk
k
where {bX,i }N
i=1 and {sX,i }i=1 are the sets of beamforming vectors and data symbols used by the BS
X ∈ Φk to communicate with Nk mobile users simultaneously. Then, a test user o located at x(o) and
connected to the BS X ⋆ ∈ Φk will receive the following baseband signal
yo =
p
|
Pk Gk (T
(X ⋆ ))h⊤
X ⋆ bX ⋆ ,o sX ⋆ ,o
{z
intended signal
+
X
X∈Φ\{X ⋆ }
7
|
p
}
Nk
X
p
⋆
h⊤
+ Pk Gk (T (X ))
X ⋆ bX ⋆ ,i sX ⋆ ,i
i=1,i6=o
|
PX GX (T (X))h⊤
X sX ,
{z
inter-cell interference
Usually, Nk is taken to be smaller than Mk .
}
{z
inter-antenna interference
}
32
where hX represents the amplitude fading coefficients for the channel from a BS located at X ∈ Φ
to the test user. As it is clear from this baseband signal representation, the inter-antenna interference
is not related to interfering BS locations and can be handled directly by the standard techniques from
probability theory. Further, if zero-forcing beamforming is employed, it can be eliminated from yo .
The inter-cell interference, on the other hand, depends on the locations of interfering BSs. However,
its structure is the same with the one studied in Section III and Section IV, except the new terms
appearing in the form of h⊤
X sX . Hence, the distribution of the interference power can be approximated
exactly as in Theorem 1 once the second and third moments for h⊤
X sX
2
are determined. Due to the
linearity of expectation operator, this is usually a straightforward task.
Lastly, we only considered single-antenna mobile users above for simplicity. Similar arguments with
receiver beamforming not depending on the interfering BS configuration hold when mobile users have
multiple antennas. If the receiver beamforming vectors depend on the locations of intefering BSs, then
this will induce a correlation between the BS location processes and the random variables modulating
the path-loss coefficients. In this case, our Gaussian approximation results cannot be applied directly.
C. Power Control
Another important and pretty much standard technique in wireless communications is the control of
transmission powers through a power control mechanism [55]–[57]. In particular, transmission powers
need to be allocated to each link for data communications, but this allocation has knock-on effects
on other links in the network due to the broadcast nature of the shared wireless medium. Hence, an
efficient power control mechanism is critical for achieving high data rates in HCNs. In this paper,
for the simplicity of exposition, we do not explicitly consider any power control mechanism in our
baseline model in Section III, which is also the case in most previous papers [8], [12]–[15], [48]–[51],
[53]. However, it is straightforward to incorporate the effect of any pairwise power control mechanism
between a transmitter and its receiver on our Gaussian approximation result in Theorem 1.
To see this more clearly, assume that PX (RX , UX ) is the power control mechanism used by the
BS located at X ∈ Φ, where RX and UX are the distance and the (power) fading coefficient between
this BS and its receiver. Then, the total interference power at the test user can be written as
X
Iλ =
PX (RX , UX ) HX GX (T (X)) ,
X∈Φ\{X ⋆ }
which is the same interference power expression in (1), except the fact that the transmission powers now
depend on connection distances between BSs and their intended receivers as well as the corresponding
fading gains. We note that the functional form for PX (RX , UX ) can be assumed to be the same for
33
all BSs belonging to the same tier. As a result, using PX (RX , UX ) HX as marks of BSs, observing
(k)
(k)
that these marks are independent of other BS locations in Φ \ {X} and replacing Pk2mH 2 and Pk3 mH 3
with the second and third moments of these marks, we can obtain a version of Theorem 1 suitable
for characterizing AWI distribution in power controlled HCNs. We note that if the employed power
control mechanism in a network tier also depends on the configuration of BSs in the network, then
the marks are not independent of the BS locations anymore and our Gaussian approximation results
cannot be applied directly in this case.
D. Uplink Communications and Other Network Technologies
Last but not least, we conclude this section of the paper with a brief discussion on the applicability
of our results for the HCN uplink communications and other network technologies. Although our main
focus in this paper is on the downlink communications in an HCN setting [13], the same analysis
directly extends to the HCN uplink communications [58]–[60].
To see this, first consider the simplest case of mobile terminals with identical transmission technology without any power control. This is exactly the case of single-tier cellular networks with the
understanding of interference signals emanating from mobile users rather than BSs. Hence, the structure
of the AWI power at any test BS remains the same as in (1) and Theorem 1 can be directly applied to
approximate uplink AWI statistics. If the transmission technologies of mobile users are, on the other
hand, dissimilar and allow them to use different but constant transmission power levels, perhaps due
to different manufacturer specifications, then this case can be treated exactly as in the HCN downlink
communications after a classification of transmission technologies into multiple groups. Finally, if a
pairwise power control mechanism is employed in the uplink communications, then we can use selected
power levels and fading coefficients as marks associated with mobile users and still use Theorem 1
as explained above to characterize the uplink AWI statistics in HCNs.
In addition to uplink communications, the analytical framework developed in this paper can find
further applications in other emerging network technologies such as Internet-of-Things (IoT) and
machine-type communications [61], [62]. Alongside mobile-edge computing and fog computing technologies [63], [64], it is expected that billions of smart devices will be accessing the same spectrum for
Internet connectivity in near future [65]. LTE for machine-type communications (LTE-M), extended
coverage GSM (EC-GSM) and narrowband IoT (NB-IoT) are the recent 3GPP study items in order
to accommodate billions of IoT devices in the same spectrum via the cellular infrastructure [66]–
[68]. In addition, LoRa and SigFox have been proposed for the unlicensed spectrum [69]. Due to
large numbers of IoT devices, we expect that the Gaussian approximation regime discovered in this
34
paper will continue to hold, possibly much more tightly, in both licensed and unlicensed bands for
machine-type communications. Since our network model is built upon general assumptions for pathloss, fading and transmitter location processes, the specific nature of transmitters is irrelevant and
Theorem 1 can still be applied to obtain Gaussian approximation bounds for the emerging machinetype communications technology once the device locations are characterized through an appropriate
PPP. Further details are left as a future work.
VIII. C ONCLUSIONS
AND
F UTURE WORK
This paper has examined various performance metrics of interest for the downlink in dense K-tier
HCNs under general settings including non-homogeneous PPPs by introducing a principled methodology. To this end, we have first investigated the Gaussian approximation for the downlink AWI
distribution in dense K-tier HCNs. Analytical bounds measuring the Kolmogorov-Smirnov distance
between the AWI distribution and Gaussian distribution have been obtained. The derived bounds have
also been illustrated numerically through simulations of particular three-tier HCN scenarios. A good
statistical fit between the simulated (centralized and normalized) AWI distribution and the standard
normal distribution has been observed even for moderate values of BS intensities.
Secondly, we have examined the downlink capacity metrics of dense K-tier HCNs under general
signal propagation models, allowing for the use of general bounded path-loss functions, arbitrary
fading distributions and general PPPs. Tight upper and lower bounds on the outage capacity, ergodic
capacity and area spectral efficiency have been obtained for two specific association policies - the
generic association policy and the BARSS association policy. The validity of our analytical results has
also been confirmed by simulations. An important design insight for dense K-tier HCNs arising from
this study is that the celebrated SIR invariance property does not hold anymore if a bounded path-loss
model is used to characterize large-scale wireless propagation losses in the downlink of a K-tier HCN.
Hence, it is imperative to either mitigate interference more efficiently or find the optimum BS intensity
per tier maximizing delivered data rates to mobile users in order to reap the benefits from network
densification.
The proposed approach can be extended to load-aware analysis of HCNs, MIMO communications,
power-controlled transmissions, uplink communications, emerging telecommunications technologies
with densely deployed transmitters and other association policies. It has the potential of understanding
the dense HCN performance and design beyond specific selections of the path-loss model and the
fading distribution. Utilizing these results, the future plans of the authors include the development of
novel spectrum management techniques for dense K-tier HCNs, an investigation of multi-slope path-
35
loss models for emerging millimeter wave communications under general settings, and an investigation
of provably near-optimum network control mechanisms such as optimum hybrid-access and cell-range
expansion.
A PPENDIX A
T HE P ROOF
OF
T HEOREM 3
In this appendix, we will provide the proof for Theorem 3 establishing the outage capacity bounds
for the generic association policy. Given that the test user is associated to a BS at a distance r in
tier-k, we can express the τ -outage probability as
Pr (τ -outage) = Pr {log (1 + SINRA ) < τ }
Z ∞
hGk (r)
−1
= 1 − SNR−1 (eτ −1) Pr Iλ ≤ Pk
− SNRk PG qk (h)dh,
τ −1
k
e
G (r)
k
where the last equality follows from the fact that Iλ is a positive random variable, and we have
(eτ −1)SNR−1
−1
k (r)
k
−
SNR
. By using Lemma 3 and the natural
PG
<
0
if
and
only
if
h
<
Pk hG
τ
k
e −1
Gk (r)
bounds 0 and 1 on the probability, we can upper and lower bound Pr (τ -outage) as
Z ∞
Pr (τ -outage) ≤ 1 − SNR−1 (eτ −1) max {0, Ψ (ζk (h, τ, r)) − Ξ · c (ζk (h, τ, r))} qk (h)dh
"
k
Gk (r)
= 1 − E max {0, Ψ (ζk (Hk , τ, r)) − Ξ · c (ζk (Hk , τ, r))} 1
= 1 − E Vk− (Hk , τ, r)
and
Pr (τ -outage) ≥ 1 −
Z
∞
SNR−1 (eτ −1)
k
Gk (r)
Hk ≥
SNR−1 (eτ −1)
k
Gk (r)
#
min {1, Ψ (ζk (h, τ, r)) + Ξ · c (ζk (h, τ, r))} qk (h)dh
"
= 1 − E min {1, Ψ (ζk (Hk , τ, r)) + Ξ · c (ζk (Hk , τ, r))} 1
= 1 − E Vk+ (Hk , τ, r) ,
Hk ≥
SNR−1 (eτ −1)
k
Gk (r)
#
where Ξ and c(x) are as given in Lemma 3, Ψ(x) is the standard normal CDF and 1{·} is the indicator
function.
A PPENDIX B
T HE P ROOF
OF
T HEOREM 5
The achievable ergodic capacity can be simply written as
Z ∞
(a)
Cerg = ESINRA [log (1 + SINRA )] =
1 − Pr (τ -outage)dτ,
0
36
where (a) follows from the fact that log (1 + SINRA ) is a positive random variable and hence its
R∞
expectation is equal to 0 Pr {log (1 + SINRA ) > τ } dτ . Using this observation, we can easily obtain
R∞
the upper and lower bounds on the ergodic capacity as Cerg ≤ 0 E Vk+ (Hk , τ, r) dτ and Cerg ≥
R∞ −
E
V
(H
,
τ,
r)
dτ by using the steps similar to those in Appendix A.
k
k
0
A PPENDIX C
T HE P ROOF
OF
L EMMA 5
In this appendix, we will derive the connection probability of the test user to a serving BS in
∆
tier-k under the BARSS association policy, which is denoted by p⋆k =Pr {A⋆ = k}. Let Ri be the
nearest distance from Φi to the test user for i = 1, . . . , K. Then, utilizing the structure of the BARSS
association policy, this probability can be written as
Z ∞Y
K
o
n
⋆ (a)
pk =
Pr βi Pi Gi (Ri ) ≤ βk Pk Gk (u) Rk = u fRk (u)du
0
(b)
=
Z
0
i=1
i6=k
K
∞Y
i=1
i6=k
Pr {βi Pi Gi (Ri ) ≤ βk Pk Gk (u)} fRk (u)du,
(10)
where the identity (a) follows from the conditional independence of the events
{βi Pi Gi (Ri ) ≤ βk Pk Gk (Rk )} for i ∈ {1, . . . , K} \ {k}
given any particular realization of Rk , and the identity (b) follows from the independence of the nearest
BS distances from different tiers. Each probability term in (10) can be written as
n
o
(k)
(k)
Pr {βi Pi Gi (Ri ) ≤ βk Pk Gk (u)} = Pr Ri ≥ Qi (u) = exp −Λi B X (0) , Qi (u)
,
(11)
where the last equality follows from the nearest BS distance distribution for Ri . Using (11), we can
R ∞ QK
(k)
write p⋆k as p⋆k = 0
i=1 exp
−Λi B X (0) , Qi (u)
fRk (u)du.
i6=k
A PPENDIX D
T HE P ROOF
OF
L EMMA 6
(k)
with
For the case of spatial homogeneous PPPs, we just need to replace Λi B X (0) , Qi (u)
2
(k)
πλi Qi (u) . Without loss of generality, we assume that the location of the test user is at the origin,
i.e., X (0) = (0, 0). Thus, using Lemma 5, we can write p⋆k as
Z ∞Y
K
2
(k)
⋆
pk =
fRk (u)du.
exp −πλi Qi (u)
0
i=1
i6=k
(12)
37
We note that exponent in (12) can be written as a summation
PK
i=1,i6=k
λi
(k)
Qi (u)
2
terms inside the summation may not be active for some particular values of u if Gk (u) ≥
Recalling the definition of ai ,
β i Pi
G (0),
β k Pk i
we observe that the condition Gk (u) ≥
, and some
β i Pi
G (0).
β k Pk i
β i Pi
G (0)
β k Pk i
holds
if and only if u ≤ G−1
k (ai ). Introducing a0 = 0 and aK+1 = +∞ to have the integration limits
from 0 to ∞, and enumerating
ai ’s in descending
order for i 6= k, we finally arrive the desired
2
j−1
K R
P
P
rj
(k)
λπ(i) Qπ(i) (u)
fRk (u) du, where π(i) is an enumeration of ai ’s
result p⋆k =
exp −π
rj−1
i=1
j=1
aπ(i) for i = 0, . . . , K.
in descending order, i.e., aπ(i) ≥ aπ(i+1) for i = 0, . . . , K − 1 and ri = G−1
k
A PPENDIX E
T HE P ROOF
OF
L EMMA 7
In this appendix, we will derive the conditional PDF of the connection distance R⋆ given the event
{A⋆ = k}. To this end, we will first calculate the conditional CDF of R⋆ given {A⋆ = k}, which will
be denoted by FR⋆ |{A⋆ =k} (r). Let Ri be the nearest distance from Φi to the test user for i = 1, . . . , K.
Then,
1
Pr {R⋆ ≤ r and A⋆ = k}
p⋆k
Z r
K
\
1
Pr
{βi Pi Gi (Ri ) ≤ βk Pk Gk (r)} Rk = u fRk (u)du.
= ⋆
pk 0
i=1
FR⋆ |{A⋆ =k} (r) =
(13)
i6=k
Using the conditional independence of the events {βi Pi Gi (Ri ) ≤ βk Pk Gk (Rk )} for i ∈ {1, . . . , K} \ {k}
for any given particular realization of Rk and the independence of the nearest BS distances from
different tiers, we can further simplify (13) as
Z r
K
\
1
Pr
{βi Pi Gi (Ri ) ≤ βk Pk Gk (u)} Rk = u fRk (u)du
FR⋆ |{A⋆ =k}(r) = ⋆
pk 0
i=1
i6=k
Z K
1 rY
(k)
= ⋆
exp −Λi B x(o) , Qi (u)
fRk (u)du.
pk 0 i=1
(14)
i6=k
We obtain the conditional PDF of R⋆ given A⋆ = k by differentiating (14) with respect to r. This
leads to
fk (r) =
K
1 Y
(k)
(o)
exp
−Λ
B
x
,
Q
(r)
fRk (r),
i
i
p⋆k i=1
i6=k
where fRk (r) = e−Λk (B(x
(o) ,r
)) d Λ B x(o) , r is the nearest tier-k BS distance distribution.
dr k
38
A PPENDIX F
T HE P ROOF OF L EMMA 8
2
(k)
(k)
(o)
Similar to Appendix D, the mean measure Λi B x , Qi (r)
reduces to πλi Qi (r) when
only homogeneous PPPs are considered. Without loss of generality, the location of the test user is
assumed to be at the origin, i.e., x(o) = (0, 0). Thus, we can rewrite fk (r) as
K
X
2
1
(k)
λi Qi (r) fRk (r) for r ≥ 0.
fk (r) = ⋆ exp −π
pk
i=1
(15)
i6=k
We observe that the summation term appearing in (15) is exactly the same one appeared in (12).
Hence, the same enumeration step can be carried out to arrive at the final result
!
j−1
K
2
X
1 X
(k)
fk (r) = ⋆
λπ(i) Qπ(i) (r)
fRk (r)1{r∈[rj−1 ,rj )} ,
exp −π
pk j=1
i=1
where a0 = 0, aK+1 = +∞, ai =
β i Pi
G (0)
β k Pk i
in descending order, i.e., aπ(i) ≥ aπ(i+1)
for i ∈ {1, . . . , K} \ {k}, π(i) is an enumeration of ai ’s
a
for i = 0, . . . , K.
for i = 0, . . . , K − 1 and ri = G−1
π(i)
k
R EFERENCES
[1] S. Ak, H. Inaltekin and H. V. Poor, “Gaussian approximation for the downlink interference in heterogeneous cellular networks,”
in Proc. IEEE International Symposium on Information Theory (ISIT), Barcelona, Spain, Jul. 2016.
[2] S. Ak, H. Inaltekin and H. V. Poor, “Downlink outage performance of heterogeneous cellular networks,” in Proc. IEEE International
Symposium on Information Theory (ISIT), Barcelona, Spain, Jul. 2016.
[3] J. Andrews, H. Claussen, M. Dohler, S. Rangan and M. Reed, “Femtocells: Past, present, and future,” IEEE J. Sel. Areas Commun.,
vol. 30, no. 3, pp. 497-?508, Apr. 2012.
[4] A. Ghosh, N. Mangalvedhe, R. Ratasuk, B. Mondal, M. Cudak, E. Visotsky, T. A. Thomas, J. G. Andrews, P. Xia, H. S. Jo, H.
S. Dhillon and T. D. Novlan, “Heterogeneous cellular networks: From theory to practice,” IEEE Trans. Commun. Mag., vol. 50,
no. 6, pp. 54–64, June 2012.
[5] I. Hwang, B. Song and S. S. Soliman, “A holistic view on hyper-dense heterogeneous and small cell networks,” IEEE Commun.
Mag., vol. 51, no. 6, pp. 20–27, June 2013.
[6] V. M. Nguyen, and M. Kountouris, “Performance limits of network densification,” IEEE J. Sel. Areas Commun., vol. 35, no.6, pp.
1294-1308, Mar. 2017.
[7] M. Ding, D. Lopez-Perez and G. Mao, “Ultra-dense networks: Is there a limit to spatial spectrum reuse?,” arXiv:1704.00399
[cs.NI], Apr. 2017. [Online]. Available: https://arxiv.org/abs/1704.00399.
[8] A. AlAmmouri, J. G. Andrews and F. Baccelli, “SINR and throughput of dense cellular networks with stretched exponential path
loss,” submitted to IEEE Trans. Wireless Commun.. [Online]. Available: https://arxiv.org/abs/1703.08246.
[9] D. Lopez-Perez, M. Ding, H. Claussen and A. H. Jafari, “Towards 1 Gbps/UE in cellular systems: Understanding ultra-dense small
cell deployments,” IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp. 2078–2101, Nov. 2015.
[10] O. Galinina, A. Pyattaev, S. Andreevy, M. Dohler and Y. Koucheryavy, “5G multi-RAT LTE-WiFi ultra-dense small cells:
Performance dynamics, architecture, and trends,” IEEE J. Sel. Areas Commun., vol. 33, no. 6, pp. 1224–1240, May 2015.
39
[11] D. Liu, L. Wang, Y. Chen, M. Elkashlan, K.-K. Wong, R. Schober and L. Hanzo, “User association in 5G networks: A survey
and an outlook,” IEEE Commun. Surveys Tuts., vol. 18, no. 2, pp. 1018–144, second quarter, 2016.
[12] J. G. Andrews, F. Baccelli and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans. on
Commun., vol. 59, no. 11, pp. 3122–3134, Nov. 2011.
[13] H. S. Dhillon, R. K. Ganti, F. Baccelli and J. G. Andrews, “Modeling and analysis of K-tier downlink heterogeneous cellular
networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 550–560, Apr. 2012.
[14] H.-S. Jo, Y. J. Sang, P. Xia and J. G. Andrews, “Heterogeneous cellular networks with flexible cell association: A comprehensive
downlink SINR analysis,” IEEE Trans. Wireless Commun., vol. 11, no. 10, pp. 3484–3495, Oct. 2012.
[15] T. Samarasinghe, H. Inaltekin and J. S. Evans, “On the outage capacity of opportunistic beamforming with random user locations,”
IEEE Trans. Commun., vol. 62, no. 8, pp. 3015-3026, Aug. 2014.
[16] H. ElSawy, E. Hossain and M. Haenggi, “Stochastic geometry for modeling, analysis, and design of multi-tier and cognitive
cellular wireless networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 3, pp. 996–1019, June 2013.
[17] C. Saha, M. Afshang, and H. S. Dhillon, “3GPP-inspired HetNet model using Poisson cluster process: Sum-product functionals
and downlink coverage,” May 2017, [Online]. Available: http://arxiv.org/abs/1705.01699
[18] M. Afshang, and H. S. Dhillon, “Poisson cluster process based analysis of HetNets with correlated user and base station locations,”
Apr. 2017. [Online]. Available: http://arxiv.org/abs/1612.07285
[19] V. H. Macdonald, “Advanced mobile phone service: The cellular concept,” Bell Syst. Tech. J., vol. 58, no. 1, pp.15–41, Jan. 1979.
[20] S. Dep, P. Monogioudis, J. Miernik and J. Seymour, “Algorithms for enhanced inter-cell interference coordination (eICIC) in LTE
HetNets,” IEEE/ACM Trans. Networking, vol. 22, no. 1, pp. 137–150, Feb. 2014.
[21] E. Hossain, M. Rasti, H. Tabassum and A. Abdelnasser, “Evolution toward 5G multi-tier cellular wireless networks: An interference
management perspective,” IEEE Wireless Commun., vol. 21, no. 3, pp.118–127, Jun. 2014.
[22] M. Haenggi and R. K. Ganti, “Interference in large wireless networks,” Foundations and Trends in Networking, vol. 3, no. 2, pp.
127-248, 2009.
[23] S. Deng, M. K. Samimi and T. S. Rappaport, “28GHz and 73GHz millimeter-wave indoor propagation measurements and path
loss models,” in Proc. 2015 ICC Workshop. IEEE, pp. 1244-1250, Jun. 2015.
[24] X. Zhang and J. G. Andrews, “Downlink cellular network analysis with multi-slope path loss models,” IEEE Trans. Commun.,
vol. 63, no. 5, pp. 1881-1894, May 2015.
[25] E. S. Sousa and J. A. Silvester, “Optimum transmission ranges in a direct-sequence spread-spectrum multihop packet radio network,”
IEEE J. Sel. Areas Commun., vol. 8, no. 5, pp. 762-771, Jun 1990.
[26] J. Ilow and D. Hatzinakos, “Analytic alpha-stable noise modeling in a Poisson field of interferers or scatterers,” IEEE Trans. Signal
Process., vol. 46, no. 6, pp. 1601-1611, June 1998.
[27] C. C. Chan and S. V. Hanly, “Calculating the outage probability in a CDMA network with spatial Poisson traffic,” IEEE Trans.
Veh. Technol., vol. 50, no. 1, pp. 183-204, Jan 2001.
[28] S. P. Weber and J. G. Andrews, “Bounds on the SIR distribution for a class of channel models in ad hoc networks,” in Proc. 49th
IEEE Global Telecommunications Conference, San Francisco, CA, Nov. 2006.
[29] M. Z. Win, P. C. Pinto and L. A. Shepp, “A mathematical theory of network interference and its applications,” Proc. IEEE, vol.
97, no. 2, pp. 205-230, Feb. 2009.
[30] P. Madhusudhanan, J. G. Restrepo, Y. Liu, T. X. Brown and K. R. Baker, “Downlink performance analysis for a generalized
shotgun cellular system,” IEEE Trans. Wireless Commun., vol. 13, no. 12, pp. 6684-6696, Dec. 2014.
[31] T. Samarasinghe, H. Inaltekin and J. S. Evans, “On optimal downlink coverage in Poisson cellular networks with power density
constraints,” IEEE Trans. Commun., vol. 62, no. 4, pp. 1382-1392, Apr. 2014.
[32] M. Aljuaid and H. Yanikomeroglu, “Investigating the Gaussian convergence of the distribution of the aggregate interference power
in large wireless networks,” IEEE Trans. Veh. Technol., vol. 59, no. 9, pp. 4418-4424, Nov. 2010.
40
[33] H. Inaltekin, “Gaussian approximation for the wireless multi-access interference distribution,” IEEE Trans. Signal Process., vol.
60, no. 11, pp. 6114-6120, Nov. 2012.
[34] H. Inaltekin, “Gaussian approximation for the wireless multi-access interference distribution and its applications,” Technical Report,
Antalya Int. Univ., 2012. [Online]. Available: http://arxiv.org/abs/1204.3812.
[35] M. Di Renzo and W. Lu, “Stochastic geometry modeling and performance evaluation of MIMO cellular networks using the
equivalent-in-distribution (EiD)-based approach,” IEEE Trans. Commun., vol. 63, no. 3, pp. 977996, March 2015.
[36] L. H. Afify, H. ElSawy, T. Al-Naffouri, and M.-S. Alouini, “Error performance analysis in K-tier uplink cellular networks using
a stochastic geometric approach,” in Proc. the 4th IEEE International Workshop on Small Cell and 5G Networks (SmallNets),
London, UK, Jun. 2015.
[37] M. D. Renzo, A. Guidotti and G. E. Corazza, “Average rate of downlink heterogeneous cellular networks over generalized fading
channels: A stochastic geometry approach,” IEEE Trans. Commun., vol. 61, no. 7, pp. 3050?-3071, Jul. 2013.
[38] A. Shojaeifard, K. A. Hamdi, E. Alsusa, D. K. C. So and J. Tang, “Exact SINR statistics in the presence of heterogeneous
interferers,” IEEE Trans. Inf. Theory, vol. 61, no. 12, pp. 6759-6773, Dec. 2015.
[39] R. W. Heath, M. Kountouris and T. Bai, “Modeling heterogeneous network interference with using Poisson point processes,” IEEE
Trans. Signal Process., vol. 61, no. 16, Aug. 2013.
[40] S. Mukherjee, “Distribution of downlink SINR in heterogeneous cellular networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3,
pp. 575-585, Apr. 2012.
[41] T. Samarasinghe, H. Inaltekin and J. S. Evans, “Modeling and analysis of opportunistic beamforming for Poisson wireless networks,”
IEEE Trans. Wireless Commun., vol. 15, no. 5, pp. 3732-3745, May 2016.
[42] P. Madhusudhanan, J. G. Restrepo, Y. Liu and T. X. Brown, “Analysis of downlink connectivity models in a heterogeneous cellular
network via stochastic geometry,” IEEE Trans. on Wireless Commun., vol. 15, no. 6, pp. 3895-3907, Jun. 2016.
[43] B. Blaszczyszyn and H. P. Keeler, “Studying the SINR process of the typical user in Poisson networks by using its factorial
moment measures,” IEEE Trans. Inf. Theory, vol. 61, pp. 6774-6794, Dec. 2015.
[44] J. F. C. Kingman, Poisson Processes, Clarendon Press, Oxford, 1993.
[45] S. Ak, H. Inaltekin and H. V. Poor, “A tractable framework for the analysis of general multi-tier heterogeneous cellular networks,”
Technical Report, Princeton University, May 2017. [Online]. Available: https://arxiv.org/abs/1610.05617.
[46] L. H. Y. Chen and Q. M. Shao, “Stein’s method for normal approximation,” IMS Lecture Notes Series, National University of
Singapore, vol. 4, pp. 1-59, Feb. 2005.
[47] H. Inaltekin, M. Chiang, H. V. Poor, and S. B. Wicker, “On unbounded path-loss models: effect of singularity on wireless network
performance,” IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1078-1092, Sep. 2009.
[48] H. Dhillon, R. Ganti and J. G. Andrews, “Load-aware modeling and analysis of heterogeneous cellular networks,” IEEE Trans.
Wireless Commun., vol. 12, no. 4, pp. 1666–1677, Apr. 2013.
[49] M. D. Renzo, W. Lu and P. Guan, “The intensity matching approach: A tractable stochastic geometry approximation to system-level
analysis of cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 9, pp. 5963–5983, Sep. 2016.
[50] A. Shojaeifard, K. A. Hamdi, E. Alsusa, D. K. C. So and J. Tang, “A unified model for the design and analysis of spatially-correlated
load-aware HetNets,” IEEE Trans. Commun., vol. 62, no. 11, pp. 1–16, Nov. 2014.
[51] H. S. Dhillon, M. Kountouris and J. G. Andrews, “Downlink MIMO HetNets: Modeling, ordering results and performance analysis,”
IEEE Trans. Wireless Commun., vol. 12, no. 19, pp. 5208–5222, Oct. 2013.
[52] A. Gupta and R. K. Jha, “A survey of 5G network: Architecture and emerging technologies,” IEEE Access, vol. 3, pp. 1206-1232,
Aug. 2015.
[53] A. Shojaeifard, K. A. Hamdi, E. Alsusa, D. K. C. So, J. Tang and K.-K Wong, “Design, modeling, and performance analysis of
multi-antenna heterogenous cellular networks,” IEEE Trans. Commun., vol. 64, no. 7, pp. 3104–3118, Jul. 2016.
[54] D. Tse and P. Viswanath., Fundamentals of Wireless Communication, Cambridge University Press, New York, NY, 2005.
41
[55] M. Chiang, P. Hande, T. Lian and C. W. Tan, “Power control in wireless cellular networks,” Foundations and Trends in Networking,
vol. 2, no. 4, pp. 381–553, Jun. 2008.
[56] H. Inaltekin and S. V. Hanly, “Optimality of binary power control for the single cell uplink,” IEEE Trans. Inf. Theory, vol. 58,
no. 10, pp. 6484–6498, Oct. 2012.
[57] H. ElSawy, A. S.-Salem, M.-S. Alouini and M. Z. Win, “Modeling and analysis of cellular networks using stochastic geometry:
A tutorial,” IEEE Commun. Surveys Tuts., vol. 19, no. 1, pp. 167–203, first quarter, 2017.
[58] T. D. Novlan, H. S. Dhillon and J. G. Andrews, “Analytical modeling of uplink cellular networks,” IEEE Trans. Wireless Commun.,
vol. 12, no. 6, pp. 2669-2679, Jun. 2013.
[59] Z. Z.-Yazdi and S. Jalali, “Outage analysis of uplink two-tier networks,” IEEE Trans. Commun., vol. 62, no. 9, pp. 3351-3362,
Sep. 2014.
[60] M. D. Renzo and P. Guan, “Stochastic geometry modeling and system-level analysis of uplink heterogeneous cellular networks
with multi-antenna base stations,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2453–2476, Jun. 2016.
[61] A. A.-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari and M. Ayyash, “Internet of things: A survey on enabling technologies,
protocols, and applications,” IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp. 2347–2376, fourth quarter, 2015.
[62] Y. Sarikaya, H. Inaltekin, T. Alpcan and J. S. Evans, “Stability and dynamic control of underlay mobile edge networks,” IEEE
Trans. Mobile Comput., to appear.
[63] M. Chiang and T. Zhang, “Fog and IoT: An overview of research opportunities,” IEEE Internet Things J., vol. 3, no. 6, pp.
854–864, Jun. 2016.
[64] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta and D. Sabella, “On multi-access edge computing: A survey of the emerging
5G network edge cloud architecture and orchestration,” IEEE Commun. Surveys Tuts., vol. 19, no. 3, third quarter, 2017.
[65] D. Lund, C. MacGillivray, V. Turner and M. Morales, “Worldwide and regional Internet of things (IoT) 2014 − 2020 forecast: A
virtuous circle of proven value and demand,” Market Analysis, International Data Cooperation, May 2014. [Online]. Available:
https://www.business.att.com/content/article/IoT-worldwide regional 2014-2020-forecast.pdf
[66] Ericsson,
“Cellular
networks
for
massive
IoT,”
Ericsson
White
Paper,
Jane.
2016.
[Online].
Available:
https://www.ericsson.com/assets/local/publications/white-papers/wp iot.pdf
[67] Nokia,
“LTE
evolution
for
IoT
connectivity,”
Nokia
White
Paper,
2017.
[Online].
Available:
https://resources.ext.nokia.com/asset/200178
[68] M. Shirvanimoghaddam, M. Dohler and S. J. Johnson, “Massive non-orthogonal multiple access for cellular IoT: Potentials and
limitations,” IEEE Commun. Mag., vol. 55, no. 9, pp. 55-61, Sep. 2017.
[69] U. Raza, P. Kulkarni and M. Sooriyabandara, “Low power wide area networks: An overview,” IEEE Commun. Surveys Tuts., vol.
19, no. 2, pp. 855–873, second quarter, 2017.
| 7cs.IT
|
arXiv:1705.01715v2 [stat.ME] 26 May 2017
Statistical inference using differentially private
bi-degree sequences and synthetic directed graphs
Ting Yan∗
Central China Normal University
Abstract
Although a lot of approaches are developed to release network data with a
rigourous differentially privacy guarantee, how to valid statistical inference (in particular for asymptotic analysis) using released data in many network models is still
unknown or not properly explored. Even in the simple p0 model that is an exponential family model with the bi-degree sequence as the sufficient statistic, how to
estimate the model parameter using the noisy bi-degree sequence is still not addressed. In this paper, we release the bi-degree sequence of a directed graph using a
discrete Laplace mechanism and propose an efficient algorithm for finding a maximum likelihood estimate of the bi-degree sequence, which is equivalent to projecting
the noisy bi-degree sequence on the set of all graphical bi-degree sequences with the
L1 distance. Along the way, our algorithm also outputs a synthetic directed graph.
By using the outputs, we present a differentially private estimator of the parameter in the p0 model. We show that the estimator is asymptotically consistent and
normally distributed, whose rate of convergence is the same as as the non private
estimator. Numerical studies demonstrate our theoretical findings.
Key words: Differentially private; Maximum likelihood estimator; Network data;
Noisy bi-degree sequence; p0 model.
Mathematics Subject Classification: 62F12, 91D30.
1
Introduction
As more and more network data (of all kinds, but especially social ones) are made publicly
available, the data privacy becomes an important issue in network data analysis since
they may contain sensitive information about individuals and their relationships (e.g.,
sexual relationships, email exchanges, financial transactions). The demand to preserve
privacy has leaded to a rapid development on algorithms to release network data or their
aggregate network statistics safely [e.g., Lu and Miklau (2014); Task and Clifton (2012);
Wang et al. (2006); Zhou et al. (2008)]. At the same time, it brings challenge for statistical
Department of Statistics, Central China Normal University, Wuhan, 430079, China.
[email protected].
∗
1
Email:
inference since the original data are not observed and inference must be based on the noisy
data. In addition, the structure of network data are non-standard. As a result, how to
estimate model parameters and analyze their asymptotic properties using noisy data in
many network models are still unknown or have not been properly explored.
The degree sequence of an undirected graph or the bi-degree sequence of a directed
graph preliminarily summarizes the information contained in network data and reflect the
heterogeneity of nodes to participate in network connection. One popular way to model
the degree sequences is putting them as the sufficient statistics for exponential family
distributions on graphs such as the p1 model for directed cases [Holland and Leinhardt
(1981)] and its special version–β-model for undirected case [Chatterjee et al. (2011)].
The theoretical properties of such random graph models including the consistency and
asymptotic normality of the maximum likelihood estimator (MLE) have been established;
see Chatterjee et al. (2011), Rinaldo et al. (2013), Perry and Wolfe (2012), Olhede and
Wolfe (2012), Yan and Xu (2013), Sadeghi and Rinaldo (2014) in the undirected case
and Yan et al. (2016) in the directed case. It is noted that such network models are
non-standard since the number of parameters increases as the number of nodes grows.
In this paper, we focus on the directed graphs. The bi-degrees may carry confidential
and sensitive information. For instances, the out-degree of one individual reveals how
many people are infected by him/her in sexually transmitted disease networks and such
information is sensitive. In this case, it is required to limit disclosure of the bi-degrees.
Differential Privacy (DP) proposed by Dwork et al. (2006) gives a rigorous definition
on data privacy and has been widely used. Within network data, DP is divided into
Edge Differential Privacy (EDP) (e.g., Nissim et al. (2007)) and Node Differential Privacy (e.g., Kasiviswanathan et al. (2013)), designed to limit disclosure of edge and node
(along with its edges) information, respectively. The aim of this paper is to design an efficient algorithm to release the bi-degree sequence under EDP and perform valid statistical
inference.
There have been efforts to release the degrees of undirected graphs in a differentially
private manner recently. Hay et al. (2009) describe an efficient algorithm to release the
degree partition. Karwa and Slavković (2016) use a discrete Laplace mechanism to release
the degree sequence and prove that a differentially private estimator of the parameters
in the β-model is asymptotically consistent and normally distributed. Moreover, they
also characterize conditions that lead to a quadratic time algorithm to check for the
existence of MLE of the β-model and construct an algorithm to output a synthetic graph.
However, it is still open for statistic inference based on the noisy bi-degree sequence. In
this paper, we release the bi-degree sequence of directed graphs by using a discrete Laplace
mechanism as well with a rigorous EDP guarantee. If we directly use the released bisequence for inference, it ignores the noise addition process and may lead to inconsistent,
or sometimes even non-existent parameter estimates. To address this problem, we propose
an effective algorithm for finding an MLE of the bi-degree sequence, which is equivalent
to projecting the noisy bi-degree sequence on the set of all graphical bi-degree sequences
with the L1 distance. Along the way, it also output a synthetic directed graph. We
characterize the relationship between the privacy parameter and the error of the MLE of
the bi-degree sequence. By using the output, we present a differentially private estimator
of the parameter in the p0 model, which is an exponential family model with the bi-
2
degree sequence as the sufficient statistic. We show that the estimator is asymptotically
consistent and normally distributed, whose rate of convergence is the same as as the non
private estimator.
For the rest of the paper, we proceed as follows. In Section 2, we introduce some notations and the background on the p0 model, the characterization of the Fisher information
matrix and the EDP. In Section 3, we release the bi-degree sequence using a discrete
Laplace mechanism and generate synthetic directed graphs. In Section 4, we present the
consistency and asymptotic normality of a differentially private estimator of the parameter in the p0 model. In Section 5, we carry out the simulation studies to evaluate the
theoretical results and analyze two real network data. We make the summary and further
discussion in Section 6. All proofs are regelated into Section 7.
2
2.1
Preliminaries
The bi-degree sequence and the p0 model
Let Gn be a simple directed graph on n ≥ 2 nodes labeled by “1, . . . , n”. Here, “simple”
means that there are no multiple edges and no self-loops in Gn . Let A = (ai,j ) be the
adjacency matrix of Gn , where aij is an indictor of the directed edge from head node i to
tail node j. If there exists a directed edge from i to j, then aP
i,j = 1; otherwise ai,j = 0.
Since Gn is loopless, let ai,i = 0 for convenience. Let d+
=
i
j6=i ai,j be the out-degree
+
+
+ ⊤
of node i andPd = (d1 , . . . , dn ) be the out-degree sequence of the graph Gn . Similarly,
−
−
− ⊤
define d−
i =
j6=i aj,i as the in-degree of vertex i and d = (d1 , . . . , dn ) as the in-degree
−
+ −
sequence. The pair d = (d+ , d− ) or {(d+
1 , d1 ), . . . , (dn , dn )} are the bi-degree sequence.
When the expectation of a bi-degree sequence is given, the probability distribution on
graph Gn to admit the maximum entropy belongs to the exponential family distribution
with the bi-degree sequence as the sufficient statistic according to the maximum entropy
principle [Wu (1997)]. Hereafter, we refer to this model as the p0 model. The subscript
“0” means a simpler model than the p1 model that considers the dyad reciprocity effect
[Holland and Leinhardt (1981)]. It is also a directed version of the β-model [Chatterjee et
al. (2011)]. Such models are useful to re-construct the networks when the node degrees are
the only information available. Moreover, they can be served as null models for hypothesis
testing [Zhang and Chen (2013)]. The p0 model can be represented as:
P(Gn ) =
X
X
1
exp(
αi d +
βj d−
i +
j ),
c(α, β)
i
j
(1)
where c(α, β) is a normalizing constant, α = (α1 , . . . , αn )⊤ and β = (β1 , . . . , βn )⊤ . The
outgoingness parameter αi characterizes how attractive the node is and the incomingness
parameter βi illustrate the extent to which the node is attracted to others. Since an outedge from vertex i pointing to j is the in-edge of j coming from i, it is immediate that the
sum of out-degrees is equal to that of in-degrees. If one transforms (α, β) to (α − c, β + c),
the probability distribution does not change. For the sake of the identification of model
parameters, we set βn = 0 as in Yan et al. (2016). This model can be formulated by
an array of mutually independent Bernoulli random variables ai,j , 1 ≤ i 6= j ≤ n with
3
probabilities:
eαi +βj
P(ai,j = 1) =
.
1 + eαi +βj
P
The normalizing constant c(α, β) is i6=j log(1 + eαi +βj ) and the likelihood equations are
Pn
eαi +βk
i = 1, . . . , n,
d+
=
i
k=1,k6=i 1+eαi +βk ,
P
α
+β
j
n
e k
d−
j =
k=1,k6=j 1+eαk +βj , j = 1, . . . , n − 1.
(2)
The fixed point iteration algorithm by Chatterjee et al. (2011) or the minorizationmaximization algorithm by Hunter (2004) can be used to solve the above system of equations. Yan et al. (2016) prove the consistency and asymptotic normality of the MLE in
the p0 model when n goes to infinity.
2.2
Characterization of the Fisher information matrix
For a subset C ⊂ Rn , let C 0 and C denote the interior and closure of C, respectively. For
a vector x = (x1 , . . . , xn )⊤ ∈ Rn , denote by kxk∞ = max1≤i≤n |xi |, the ℓ∞ -norm of x. For
an n × n matrix J = (Ji,j ), let kJk∞ denote the matrix norm induced by the ℓ∞ -norm on
vectors in Rn , i.e.
n
X
kJxk∞
kJk∞ = max
= max
|Ji,j |.
1≤i≤n
x6=0 kxk∞
j=1
In order to characterize the Fisher information matrix in the p0 model, we introduce
a class of matrices. Given two positive numbers m and M with M ≥ m > 0, we say the
(2n − 1) × (2n − 1) matrix V = (vi,j ) belongs to the class Ln (m, M) if the following holds:
P2n−1
P2n−1
vn,j ,
vi,j ≤ M, i = 1, . . . , n − 1; vn,n = j=n+1
m ≤ vi,i − j=n+1
vi,j = 0, i, j = 1, . . . , n, i 6= j,
vi,j = 0, i, j = n + 1, . . . , 2n − 1, i 6= j,
m ≤ vi,j = vj,i ≤ M, i = 1, . . . , n, j = n + 1, . . . , 2n − 1, j 6= n + i,
vi,n+i = vn+i,i = 0, i = 1, . . . , n − 1,
P
P
vi,i = nk=1 vk,i = nk=1 vi,k , i = n + 1, . . . , 2n − 1.
(3)
Clearly, if V ∈ Ln (m, M), then V is a (2n − 1) × (2n − 1) diagonally dominant, symmetric
nonnegative matrix.
P2n−1
P2n−1
Define v2n,i = vi,2n := vi,i − j=1;j6
i=1 v2n,i .
=i vi,j for i = 1, . . . , 2n − 1 and v2n,2n =
Then m P
≤ v2n,i ≤ MPfor i = 1, . . . , n − 1, v2n,i = 0 for i = n, n + 1, . . . , 2n − 1 and
v2n,2n = ni=1 vi,2n = ni=1 v2n,i . Yan et al. (2016) propose to approximate the inverse of
4
V , V −1 , by the matrix S = (si,j ), which is defined as
δi,j
1
+ v2n,2n
, i, j = 1, . . . , n,
vi,i
1
−
,
i = 1, . . . , n, j = n + 1, . . . , 2n − 1,
v2n,2n
si,j =
1
i = n + 1, . . . , 2n − 1, j = 1, . . . , n,
− v2n,2n ,
δ
i,j + 1 , i, j = n + 1, . . . , 2n − 1,
vi,i
(4)
v2n,2n
where δi,j = 1 when i = j and δi,j = 0 when i 6= j. To quantify the accuracy of this
approximation, we define another norm k · k for a matrix A = (ai,j ) by kAk := maxi,j |ai,j |.
The upper bound of the approximation error is given below.
Proposition 1 (Proposition 1 in Yan et al. (2016)). If V ∈ Ln (m, M) with M/m = o(n),
then for large enough n,
c1 M 2
kV −1 − Sk ≤ 3
.
m (n − 1)2
where c1 is a constant that does not depend on M, m and n.
Note that if M and m are bounded constants, then the upper bound of the above
approximation error is on the order of n−2 .
2.3
Edge differential privacy
Differential privacy (DP) proposed by Dwork et al. (2006) is a random data releasing
mechanism that defines a condition probability distribution on the output for a given
input with a restriction. Intuitively, it requires that the distribution of the output is the
same whether or not an individual’s data appear in the dataset. From a theoretical point,
any test statistic has no power for testing whether an individual is in the datasets or not;
see Wasserman and Zhou (2010) for a rigourous proof.
Within network data, DP is divided into node DP and Edge DP (EDP) depending on
the definition of the graph neighbor. If two graphs are called neighbors if they differ in
exactly one edge, then DP is EDP. Analogously, we can define node differential privacy
by letting graphs be neighbors if one can be obtained from the other by removing a
node and its adjacent edges. Intuitively, EDP protects edges not to be detected, whereas
node differential privacy protects nodes together with their adjacent edges (that is, all
information pertaining to nodes). Similar to Karwa and Slavković (2016), we use EDP
here. According to Theorem 2.4 in Wasserman and Zhou (2010), one can show that EDP
makes it nearly impossible to test the presence or absence of an edge in the graph, thus
providing protection. Let δ(G, G′ ) be the number of edges on which G and G′ differ.
The formal definition is as follows. Some notations are quoted from Karwa and Slavković
(2016).
Definition 2.1 (Edge differential privacy). Let ǫ > 0. A randomized mechanism (or a
family of conditional probability distribution) Q(·|G) is ǫ-edge differentially private if
sup
sup log
G,G′ ∈G,δ(G,G′ )=1 S∈S
5
Q(S|G)
≤ǫ
Q(S|G′ )
where G is the set of all directed graphs of interest on n nodes and S is the set of all
possible outputs (or the range of Q).
The privacy parameter ǫ is chosen by the person administering the privacy policy and
is public, which controls the trade-off between privacy and accuracy. Small value of ǫ
means more privacy protection. The most common mechanism to release the output of
any function f under differential privacy is the Laplace Mechanism [e.g., Dwork et al.
(2006)] which adds Laplace noise proportional to the global sensitivity of f .
Definition 2.2. (Global Sensitivity). Let f : G → Z k . The global sensitivity of f is
defined as
GS(f ) = max
kf (G) − f (G′ )k1
′
δ(G,G )=1
where k · k1 is the L1 norm.
A common algorithm to achieve the EDP is to add discrete Laplace noise to the input
if the outputs are the network statistics. Its EDP property is given in Lemma 1. The proof
of Lemma 1 is similar to that of Lemma 1 in Karwa and Slavković (2016) for undirected
cases and we omit it.
Lemma 1. (Discrete Laplace Mechanism). Let f : G → Z k . Let Z1 , . . . , Zk be independent and identically distributed discrete Laplace random variables with pmf defined as
follows:
1 − λ |z|
λ , z ∈ Z, λ ∈ (0, 1).
P(Z = z) =
1+λ
Then the algorithm which on input G outputs f (G) + (Z1 , . . . , Zk ) is ǫ-edge differentially
private, where ǫ = −GS(f ) log λ.
One nice property of differential privacy is that any function of a differentially private
mechanism is also differentially private.
Lemma 2 (Dwork et al. (2006); Wasserman and Zhou (2010)). Let f be an output of
an ǫ-differentially private mechanism and g be any function. Then g(f (G)) is also ǫdifferentially private.
By Lemma 2, any post-processing done on the noisy bi-degree sequences obtained as
an output of a differentially private mechanism is also differentially private.
3
Releasing bi-degree sequence and synthetic graph
In this section, we release the bi-degree sequence under EDP by using a discrete Laplace
mechanism in Lemma 1. Then we “de-noise” noisy bi-sequence by using maximum likelihood estimation. Since each of these steps uses only the output of a differentially private
algorithm, by Lemma 2, the generated synthetic graphs of Algorithm 2 are also differentially private.
6
3.1
Releasing the bi-degree sequence
We release the bi-degree sequence d = (d+ , d− ) under EDP by using a discrete Laplace
mechanism in Lemma 1. Note that f (Gn ) = (d+ , d− ). If we add or remove a directed
edge i → j in Gn , then the out-degree of the head node and the in-degree of the tail
node associated with the changed edge increase or decrease 1 each. Therefore, the global
sensitivity for the bi-degree sequence is 2. Algorithm 1 describes the process of releasing
d and return a differentially private bi-sequence.
Algorithm 1
Input: The bi-degree sequence d of a directed Gn and privacy parameter ǫ
Output: Differentially private answer to the bi-degree sequence of Gn
1: Let d = (d+ , d− ) be the bi-degree sequence of Gn
2: for i = 1 → n do
−
3:
Generate two independent e+
i and ei from discrete Laplace with λ = exp(−ǫ/2);
+
−
−
−
4:
Let zi+ = d+
i + ei and zi = di + ei
5: end for
6: return z = (z + , z − )
3.2
The MLE of bi-degrees and synthetic directed graphs
The output of Algorithm 1 may not be the graphical bi-degree sequence. There have been
several characterizations for the bi-degree sequence [e.g., Fulkerson (1960); Kleitman and
Wang (1973); Majcher (1985)]. A necessary condition for graphical bi-degree sequences is
that the sum of in-degrees is equal to that of out-degrees. To check what are the chances
that this condition holds, we carry out some simulations. We use the p0 model to generate
the random graphs and record their bi-degree sequences. Then use Algorithm 1 to output
the bi-sequence z. We set αi , βP
n = 100. We repeat 10, 000 simulations
i ∼ U(0, 1)
P and
+
−
and record the frequency that i zi = i zi holds. The simulation results show that
this condition holds with at most 1%. Therefore, the output generally can not be directly
used for inference.
To valid inference, we need to denoise the bi-sequence z. Note that the number of
−
parameters to be estimated (d+
i , di , i = 1, . . . , n) is equal to the number of observations
(zi+ , zi− , i = 1, . . . , n), and the parameter space is discrete and very large–the convex hull
of the parameter set is full dimensional for n ≥ 4. Note the parameter λ in the noise
addition process of Algorithm 1 is known. Let Bn be the set of all possible bi-degrees
sequence of graph Gn . To estimate d, we treat d as the fixed unknown parameter in Bn .
The likelihood on observation z with the parameter d is
L(d|z) = c(λ) exp{−(
n
X
i=1
|zi+
−
d+
i |
+
n
X
i=1
|zi− − d−
i |)}.
The MLE dˆ lying in the set Bn of d is
dˆ = arg min (kz + − d+ k1 + kz − − d− k1 ).
d∈Bn
7
(5)
Computing an MLE of d is equivalent to finding a L1 “projection” of z on Bn . Note that
ˆ Along
the optimization is not unique. We propose Algorithm 2 to produce the MLE d.
ˆ
the way, it also output a directed graph with d as its bi-degree sequence. The correctness
of Algorithm 2 is given in Theorem 1, whose proof is in Section 7.1.
Algorithm 2
Input: A bi-sequence of integers z = (z + , z − )
Output: A directed graph Gn on n vertices with bi-degree sequence dˆ
1: Let Gn be the empty graph on n vertices
2: Let S = {1, . . . , n} \ {i : zi+ ≤ 0}
3: while |S| > 0 do
4:
T = {1, . . . , n} \ {i : zi− ≤ 0}
5:
Let zi+∗ = maxi∈S zi+ and i∗ = min{i ∈ S : zi+ = zi+∗ }.
6:
Let T = T \ {i∗ } and pos = |T |.
7:
Let hi∗ = min(zi+∗ , pos).
8:
Let I =indices of hi∗ highest values in z − (T ) where z − (T ) is the sequence z −
restricted to the index set T
9:
Add a directed edge from i∗ to k in Gn for each k ∈ I
10:
Let zi− = zi− − 1 for all i ∈ I and S = S \ {i∗ }
11: end while
Theorem 1 (MLEs of bi-degrees). Let z = (z + , z − ) be a bi-sequence of integers obtained
from Algorithm 1. The bi-degree sequence of graph Gn produced by Algorithm 2 is an MLE
of d.
The directed Havel-Hakimi algorithm proposed by Erdós et, al. (2010) plays a key role
in the proof of Theorem 1. It is a “certifying” algorithm that produces a certificate that a
bi-degree sequence is graphical, i.e., if the input to the algorithm is graphical, it outputs
a directed graph that realizes it. We prove Theorem 1 by converting such a certifying
algorithm into Algorithm 2 that performs L1 “projection” on the set Bn .
Remark 1. In step 8 of Algorithm 2, if some in-degrees of z − (T ) are equal, we arrange
them by the decreasing order of their corresponding out-degrees. Assume that the order
is zi−1 ≥ · · · ≥ zi−k . Then we select their top hi∗ values. This rule applies hereafter and we
will not emphasize it.
The next lemma characterizes the error between dˆ and d in terms of the privacy
parameter ǫn .
Lemma 3. Let Gn be a directed graph and d be its bi-degree sequence. Let z be the output
of Algorithm 1 and dˆ be the MLE of d obtained from z using Algorithm 2 with the private
parameter ǫn . When ǫn (c + 1) ≥ 4 log n, we have
P(kdˆ − dk∞ > c) ≤
4
,
n
where for two bi-sequence a = (a+ , a− ) and b = (b+ , b− ), ka − bk∞ is defined as
ka − bk∞ = max{ka+ − b+ k∞ , ka− − b− k∞ }
8
(6)
Lemma 3 characterizes the relationship between the privacy parameter and the error.
As expected, the privacy parameter ǫn is smaller, the error between the original bi-degree
and its MLE dˆ becomes larger. For any fixed τ ∈ (0, 1/2), if ǫn = Ω(n−(1/2−τ ) ), then
kdˆ − dk∞ = Op (n(1/2−τ ) log n).
(7)
The above equation will be used in next section.
4
Asymptotic properties of a DP estimator
In this section, we propose a DP estimator of the parameter in the p0 model (1) using the
output dˆ of Algorithm 2. We defined a DP estimating equation as follows:
Pn
eαi +βk
=
dˆ+
i = 1, . . . , n,
i
k=1,k6=i 1+eαi +βk ,
P
α
+β
i
n
e k
dˆ−
i =
k=1,k6=i 1+eαk +βi , i = 1, . . . , n − 1.
(8)
Let θ = (α1 , . . . , αn , β1 , . . . , βn−1)⊤ . The solution θb to the above equations is defined as
the DP estimator of θ, where θb = (α̂1 , . . . , α̂n , β̂1 , . . . , β̂n−1 )⊤ and β̂n = 0. The system of
equations (8) is precisely the equation (2) by replacing dˆ with d. By Lemma 2, θb is DP
since dˆ satisfies DP. The existence and consistency of θb is state blow.
Theorem 2. Assume that θ∗ ∈ R2n−1 with kθ∗ k∞ ≤ τ log n, where 0 < τ < 1/24 is a
constant, and that A ∼ Pθ∗ , where Pθ∗ denotes the probability distribution (1) on A under
the parameter θ∗ . If ǫn = Ω(n−(1/2−ρ) ) for a fixed ρ ∈ (0, 1/2), then as n goes to infinity,
with probability approaching one, the DP estimator θb exists and satisfies
1/2 8kθ ∗ k∞
(log
n)
e
∗
kθb − θ k∞ = Op
= op (1).
n1/2
Further, if θb exists, it is unique.
We briefly describe its proof here and the details are regelated to Section 7.3. Define
a system of functions:
Pn
eαi +βk
−
= dˆ+
i
k=1;k6=i 1+eαi +βk ,
P
−
eαk +βj
Fn+j (θ) = dˆj − nk=1;k6=j 1+e
αk +βj ,
Fi(θ)
F (θ)
i = 1, . . . , n,
j = 1, . . . , n,
(9)
= (F1(θ), . . . , F2n−1(θ))⊤.
Note the solution to the equation F (θ) = 0 is precisely the DP estimator. We construct
a Newton iterative sequence. If the initial value is chosen as the true value θ∗ , then it
is left to obtain the upper bound of error between the initial point and the solution to
show the consistency. By calculations, it can be shown that the Jacobin matrix of F (θ)
belongs to the matrix class Ln (m, M) defined at (3). It guarantees that the Newton
iterative sequence converges with a geometric convergence of rate under some additional
mild conditions as demonstrated in Theorem 5.
9
b
Next, we present asymptotic normality of θ.
Theorem 3. Assume that A ∼ Pθ∗ and ǫn = Ω(n−(1/2−ρ) ) for a fixed ρ ∈ (0, 1/2). If
kθ∗ k∞ ≤ τ log n, where τ ∈ (0, 1/44) is a constant, then for any fixed k ≥ 1, as n → ∞,
the vector consisting of the first k elements of (θb − θ∗ ) is asymptotically multivariate
normal with mean 0 and covariance matrix given by the upper left k × k block of S defined
at (4).
−
+ −
⊤
We give a brief description for the proof of Theorem 3 here. Let g = (d+
1 , . . . , dn , d1 , . . . , dn−1 )
ˆ+ ˆ−
ˆ− ⊤
and ĝ = (dˆ+
1 , . . . , dn , d1 , . . . , dn−1 ) . We use V to denote the Fisher information matrix
of θ in the p0 model. If we apply Taylor’s expansion to each component of ĝ − Eg, then
the second order term in the expansion is V (θb − θ). Since V −1 does not have a closed
form, we work with S defined at (4) to approximate it and Proposition 1 gives an upper
bound on the approximation error. Then we present θb − θ as the sum of S(ĝ − Eg) and a
remainder. The proof is achieved by establishing the asymptotic normality of S(ĝ − Eg)
and showing the remainder is negligible.
Remark 2. By Theorem 3, for any fixed i, as n → ∞, the convergence rate of θ̂i is
∗
1/2
1/vi,i . Since (n − 1)e−2kθ k∞ /4 ≤ vi,i ≤ (n − 1)/4, the rate of convergence is between
∗
O(n−1/2 ekθ k∞ ) and O(n−1/2 ), which is the same as the non private estimate.
5
5.1
Numerical studies
Simulation
In this section, we evaluate the asymptotic results through numerical simulations. Similar
to Yan et al. (2016), the settings of parameter values take a linear form. Specifically, we
∗
set αi+1
= (n − 1 − i)L/(n − 1) for i = 0, . . . , n − 1. We considered four different values
for L, L = 0, log(log n), (log n)1/2 and log n, respectively. For the parameter values of β,
let βi∗ = αi∗ , i = 1, . . . , n − 1 for simplicity and βn∗ = 0 by default. We set ǫ = 2 for the
privacy parameter.
Note that by Theorem 3, ξˆi,j = [α̂i − α̂j −(αi∗ −αj∗ )]/(1/v̂i,i + 1/v̂j,j )1/2 , ζ̂i,j = (α̂i + β̂j −
αi∗ − βj∗ )/(1/v̂i,i + 1/v̂n+j,n+j )1/2 , and η̂i,j = [β̂i − β̂j − (βi∗ − βj∗ )]/(1/v̂n+i,n+i + 1/v̂n+j,n+j )1/2
are all asymptotically distributed as standard normal random variables, where v̂i,i is the
b Therefore, we assess the asymptotic normality
estimate of vi,i by replacing θ∗ with θ.
of ξˆi,j , ζ̂i,j and η̂i,j using the quantile-quantile (QQ) plot. Further, we also record the
coverage probability of the 95% confidence interval, the length of the confidence interval,
and the frequency that the DP estimate does not exist. The results for ξˆi,j , ζ̂i,j and η̂i,j
are similar, thus only the results of ξˆi,j are reported. Each simulation is repeated 10, 000
times.
We consider two values for n, n = 100 and 200 and find that the QQ-plots for them
are similar. Therefore, we only show the QQ-plots when n = 200 in Figure 1 to save
space. In this figure, the horizontal and vertical axes are the theoretical and empirical
quantiles, respectively, and the straight lines correspond to the reference line y = x. When
L = 0, the empirical quantiles agree well with the theoretical ones while there are notable
deviations for pair (n − 1, n) when L ≥ log(log n). When L = log n, the private estimate
10
−3
−1
1
3
−3
−1
1
3
−1 0 1 2 3
−3
−1
1
3
−3
−1
1
3
−3
−1
1
3
−1
1
3
−3
−1
1
3
−3
−1
1
3
−3
−3
−1 0 1 2 3
−3
−1 0 1 2 3
−3
−1 0 1 2 3
−3
−3
−1 0 1 2 3
3
−1 0 1 2 3
1
−3
−1 0 1 2 3
−3
−1 0 1 2 3
−1
−1 0 1 2 3
−3
−3
L = log(log(n))
L = (log(n))1/2
(i,j) = (n−1, n)
(i,j) = (n/2, n/2+1)
−3
L=0
(i,j) = (1,2)
1/2
Figure 1: The QQ plots of v̂ii (θ̂i − θi ).
did not exist in all repetitions (see Table 1, thus the corresponding QQ plot could not be
shown).
Table 1 reports the coverage probability of the 95% confidence interval for αi − αj ,
the length of the confidence interval, and the frequency that the MLE did not exist.
As expected, the length of the confidence interval increases as L increases and decreases
as n increases, which qualitatively agree with the theory. When L ≤ log(log(n)), most
of simulated coverage probabilities are close the targeted level. When L = (log n)1/2
(Conditions in Theorem 3 no longer hold), the DP estimate failed to exist with a positive
probabilty; when L = log n, the MLE did not exist in any of the repetitions.
5.2
Real data analysis
1. We analyze the Children’s Friendship data [Anderson et al. (1999)], downloaded from
http://moreno.ss.uci.edu/data.html. This is a directed network data set about children’s friendships in elementary schools. The original data were collected by Parker and
Asher (1993) and contain 881 children in 36 classrooms in the third, fourth and fifth
grades in five US public elementary schools. Anderson et al. (1999) revisited this data
11
Table 1: The reported values are the coverage frequency (×100%) for αi − αj for a pair
(i, j) / the length of the confidence interval / the frequency (×100%) that the DP estimate
did not exist.
n
(i, j)
100
(1,2)
(50,51)
(99,100)
200
(1,2)
(100,101)
(199,200)
L=0
L = log(log n)
L = (log(n))1/2
L = log(n)
92.98/0.57/2.41 93.21/0.10/2.44 97.23/1.46/66.46 NA/NA/100
93.48/0.57/2.41 93.87/0.76/2.44 92.07/0.93/66.46 NA/NA/100
92.45/0.57/2.41 85.78/0.63/2.44 83.75/0.68/66.46 NA/NA/100
95.15/0.40/0
95.23/0.40/0
95.24/0.40/0
94.08/0.75/0
94.10/0.55/0
87.68/0.45/0
96.62/1.10/18.7
94.26/0.68/18.7
85.87/0.48/18.7
NA/NA/100
NA/NA/100
NA/NA/100
and construct the Children’s Friendship data by choosing three of the 36 classrooms, one
from each grade. Here, we only use the data set from the third grade for analysis, which
contains 22 nodes and 177 directed edges representing the friendships from i to j that
child i said j is his friend.
We evaluate how close the private estimator (α̂, β̂) is to the MLE (α̃, β̃) fitted in the p0
model by using the original bi-degree sequence. The private estimator (α̂, β̂) were fitted
from the output of Algorithm 2. We chose ǫ equal to 1, 2 and 3 as done in Karwa and
Slavković (2016) and repeated to release the bi-degree sequence using Algorithm 1 1, 000
times for each ǫ. Then we computed the average private estimate and the upper (97.5th )
in blue color and the lower (2.5th ) quantiles in orange color of the estimates conditional on
the event that the private estimate exists. The frequencies that the private estimate fails
to exist are 93.8%, 58.9% and 39.4% for ǫ = 1, 2, 3, respectively. The results are shown
in Figure 2 with the estimates of α (β) on the vertical axis and out-degree (in-degree)
on the horizontal axis. The black point indicates α̃ or β̃ and the red point indicates the
mean value of α̂ or β̂). Also plotted the upper (97.5th ) and the lower (2.5th ) quantiles of
the estimates. The results show that the mean estimate is very close to the MLE and the
MLE lies within the 95% confidence interval. Moreover, as expected, as ǫ increases, the
length of confidence interval becomes smaller.
2. We analyze Lazega’s Law Firm data [Lazega (2001)], also download from http://
moreno.ss.uci.edu/data.html. This data set comes from a network study of corporate
law partnership that was carried out in a Northeastern US corporate law firm in New
England during 1988–1991. Lazega (2001) gave a description network analyses of this
data set. This data set are non-symmetric and includes three types of measurements of
networks among the 71 attorneys of this firm–coworker, advice and friendship. We use
the cowork data set for analysis. The cowork relationship from attorney i to j means that
i said j had worked with himself in the past year. In this data set, node 8 is isolated and
we removed it before analysis. The left data have 70 lawyers and 756 directed edges.
12
ε=2
4
6
8
10
14
−4
−3
−2
−1
0
1
2
−4
−3
−2
−1
0
1
2
1
0
−1
−4
−3
−2
α estimate
ε=3
2
ε=1
MLE
Mean Estimate
4
6
8
10
14
4
6
4
6
8
10
14
4
6
8
10
14
−2
−1
0
1
2
3
4
3
2
1
−2
−1
0
1
0
−2
−1
β estimate
2
3
4
Upper
Lower
4
out−degree
4
6
8
10
14
8
10
14
in−degree
Figure 2: The differentially private estimate (α̂, β̂) with the MLE for the children data
set (third grade). The plots show the median and the upper (97.5th ) and the lower (2.5th )
quantiles.
Similar to the analysis of Children’s Friendship data, we also chose ǫ equal to 1, 2
and 3 and repeated the simulation 1, 000 times for each ǫ. Then compute the average
private estimate and the upper (97.5th ) and the lower (2.5th ) quantiles of the estimates.
The frequencies that the private estimate fails to exist are 98.0%, 62.1% and 46.1% for
ǫ = 1, 2, 3, respectively. The results are shown in Figure 3. From this figure, we can see
that the mean value of α̂ or β̂ deviates from the MLE a little. But the MLE still lies in
the 95% confidence interval. On the other hand, as ǫ increases, the length of confidence
interval becomes smaller.
6
Summary and discussion
We have presented a discrete Laplace mechanism to release the bi-degree sequences of a
directed graph with an EDP guarantee. The noisy bi-degree sequence generally cannot be
directly used for inference since it is not be graphical with a large probability according
to the simulations. We propose an efficient maximum likelihood algorithm for estimating
the original bi-degree sequence, in which it also outputs a directed graph with the MLE
of the original bi-degree sequence as its bi-degree sequence. We also characterize the
error between the original bi-degrees and its MLEs. By using the MLE of the bi-degree
sequence, we propose a differential private estimator of the parameter in the p0 model and
show that it is asymptotical consistent and normally distributed. From utility perspective,
13
5
10
15
20
25
−5
−4
−3
−2
−1
−2
−3
−5
−4
−3
−5
−4
α estimate
ε=3
−1
ε=2
−2
−1
ε=1
MLE
Mean Estimate
5
10
15
20
25
5
10
15
20
25
5
10
15
20
25
5
10
15
20
25
3
2
1
−1
0
1
−1
0
1
−1
0
β estimate
2
3
Upper
Lower
2
3
out−degree
5
10
15
20
25
in−degree
Figure 3: The differentially private estimate (α̂, β̂) with the MLE for the Lazega’s cowork
data set.
it is no cost for the asymptotic analysis.
There are open questions for further study. Although the general necessary and sufficient condition for the existence of the MLE in exponential families has been derived
[Barndorff-Nielsen (1973)], it is not possible to check whether the condition hold in many
problems. In the p0 model, this condition becomes that the bi-degree sequence d lies in
the relative interior of convex hull of Bn . As pointed out by Rinaldo et al. (2013) in
their supplementary material, the combinatorial complexity of the convex hull of Bn is
quite high and increases very rapidly with at least an exponential order on n. Unlike the
undirected graph, the facets of this convex hull is still unknown. To solve the estimating
equation (8), we use the iterative scaling algorithm propose by Holland and Leinhardt
(1981) to compute the solution. This algorithm does not directly produce the DP estimate θ̂. Instead, it output the DP estimate of the probability p̂i,j with edge ai,j occurs
by replacing the parameter αi and βj in pij with their DP estimates. We say the MLE
does not exist if some pij s are zeros or ones. Like the undirected case in which Karwa and
Slavković (2016) devise a quadratic time algorithm to detect the boundary points of the
convex hull of degrees, it is of interest to investigate whether there are similar results in
directed cases. Moreover, Karwa et al. (2014) and Karwa et al. (2017) propose methods to
release network data in a wide class of exponential random graph models (ERGMs) in a
differentially private manner and use Markov chain Monte Carlo techniques to fit ERGMs
to the produced synthetic networks. However, there lacks of theoretical investigations on
the properties of the estimators such as quantifying the error of the estimator between in
14
the original network data and in noisy network data.
7
7.1
proofs
Proofs for Theorem 1
In this subsection, we show that Algorithm 2 finds a solution to the optimization problem
(5). We convert a directed Havel-Kakimi algorithm [Havel (1955); Hakimi (1962)] proposed by Erdós et, al. (2010) into Algorithm 2 to achieve the proof. Similar to addressing
the problem of undirected cases in Karwa and Slavković (2016), there are also two main
steps here. First, we reduce the global optimization to a local optimization by ignoring
the indices with negative entries in z + and z − and restricting to bi-degree sequences with
their out-degrees and in-degrees are point-wise bounded by z + and z − , respectively. Second, we use the so called k-out-star graphs to select the optimal directions by noticing
that every point in Bn can be written as a sum of special bi-degree sequences of k-out-star
graphs.
Erdós et, al. (2010) introduce the notation of normal order. We say that the bidegree sequence is in normal order if the entries satisfy the following properties: for each
−
−
−
+
+
−
−
i = 1, . . . , n−2, we either have d−
i > di+1 or di = di+1 and di ≥ di+1 . We use d(1) , . . . , d(n)
to denote the normal order. Note that we made no ordering assumption about node n.
The following theorem characterizes the bi-degree sequence.
Theorem 4 (Theorem 2 in Erdós et, al. (2010)). Assume that the bi-degree sequence
−
+
(d+ , d− ) (with d+
j + dj > 0, j ∈ [1, n]) is in normal order and dn > 0 (that is the
out-degree of the last vertex is positive). Then (d+ , d− ) is bi-graphical if and only if the
bi-degree sequence
(
(
+
+
d−
d
,
k
=
6
n
−
k − 1, k ≤ dn
k
,
b
=
b+
=
k
k
d−
k > d+
0, k = n
n
k,
−
with zero elements removed (those j for which d+
j = dj = 0) is bi-graphical.
By Theorem 4, we can use a recursive method to check whether a bi-sequence of
integers is in Bn . To speed up the recursive process, at each step, we choose the node
with the largest out-degree as the node “n” and arrange the left nodes in normal order,
although the node “n” is chosen arbitrarily. At step 1, we choose the node with the largest
out-degree as the node “n” and remove d−
n connections from vn to nodes with largest indegrees. Then remove the nodes that have lost both their in- and out- degrees in the
process. Repeat this step until all out-degrees become zero. Since the sum of out-degrees
is equal to that of in-degrees, all in-degrees also become zero when all out-degrees become
zero. At the end of the procedure if we are left with a bi-sequence of 0’s, the original
bi-sequence is in Bn . Note that each node in this process is picked at most once. The
number of repeats is at most n. So the algorithm is fast and efficient. This method can be
modified to characterize every point inside Bn . Before we present this characterization,
we will define k-out-star sequences.
15
Definition 7.1. A k-out-star graph centered at node i is any graph where there are only
k out-edges from i pointing to k other nodes in the directed graph Gn . Node i is called
the center and the k nodes to which it points are called leaf nodes. A bi-sequence dk(i) =
(d+k(i) , d−k(i) ) is said to be a k-out-star sequence centered at node i if there exists a graph
Gn such that Gn is a k-out-star graph with its center at i and bi-degree sequence dk(i) .
Similarly, we can define a k-in-star graph bk(i) with i as its center and k leaf nodes pointing
to i.
In a k-out-star sequence, the number of out-degrees equal to k is 1 and that of indegrees equal to 1 is k. When k < n, the k-out star graph have isolated nodes. Due to
that the exact ordering of the leaf nodes have not specified, dk(i) is actually a set of bidegree sequences. For example, ({3, 0, 0, 0, 0}, {0, 1, 1, 1}) and ({3, 0, 0, 0, 0}, {3, 1, 0, 1, 1})
are both 3-out-star sequences centered at node 1, we represent all such sequences (with
abuse of notation) by d1(3) . When the ordering of the leaf nodes is to be fixed, we specify
it explicitly. Let Kn be the set of all k-out-star bi-degree sequences on n nodes. Lemma
4 characterizes all bi-degree sequences in terms of k-out-star degree sequences.
Lemma 4. Every bi-degree sequence d can
written as a sum of n k-out-star sequences,
Pbe
n
each centered at a distinct node i.e., d = i=1 g ki(li ) where g ki(li ) ∈ Kn .
Proof. Let d = (d+ , d− ) be any bi-degree sequence of the graph Gn . Consider repeated
applications of Theorem 4 to d. Note that at the end of each application, some nodes may
loss their out-degrees and in-degrees. In this case, we still work with bi-sequences of the
same length 2n and append 0s in the appropriate locations. Specifically, let r i = (r i+ , r i− )
be the bi-sequence obtained at the end of each application, call it the ith residual bisequence. Its construction is described below.
At the initial step, r 1 = d. At step i, r i+1 is obtained from r i by subtracting a k-outstar sequence, i.e., r i+1 = r i − g ki(li ) . g ki (li ) is the bi-degree sequence of a k-out-star graph
G′n centered at node li where li is the index of the ith largest element of r i+ and k = rlii .
The leaf nodes of G′n are the nodes with k largest elements in {r1i− , . . . , rni− } \ {rli−
}. If
i
there are leaf nodes with the same in-degrees, arrange them into the decreasing order of
their out-degrees.
Since at each step, one out-degree becomes zero, this procedure terminates after at
most n steps. Thus it generates at most n residual sequences. Moreover, as d is graphical,
rn is the 0 sequence. Finally, r i+1 −P
r i = g ki(li ) for i = 2, . . . , n and r 1 = d, r n = 0.
Adding these inequalities, we get d = i g ki(li ) . Since each g ki (li ) is a k-out-star sequence,
g ki(li ) ∈ Kn .
Definition 7.2. (Directed Havel-Hakimi decomposition). The Directed Havel-Hakimi decomposition of a bi-degree sequence d is defined as the set of k-out-star sequences obtained
after the application of Theorem 4 and is denoted by H(d) = {g 1, . . . , g n } where g i = g ki (li ) .
Lemma 4 shows that every bi-degree sequence can be written as a sum of k-out-star
sequences, thus every bi-degree sequence has a Havel-Hakimi decomposition. In many
proofs below, we reduce a bi-degree sequence by a k-out-star sequence. The proposition
below gives a condition when the resulting sequence is always graphical.
16
Proposition 2. Let d be a bi-degree sequence.
′
k(i)
(1)Let k ≤ d+
, such that d − d′ is also
i . Then there exists a k-out-star sequence d in g
graphical.
′
k(i)
(2)Let k ≤ d−
, such that d − d′ is also
i . Then there exists a k-in-star sequence d in b
graphical.
Proof. (1) Since d is a graphical bi-degree sequence, it follows that node i points to d+
i
other nodes. Since k ≤ d+
i , it is possible to delete these k out-edges with the head node i
in the graph. Clearly, the bi-degree sequence of this graph is d − d′ .
(2) Since d is a graphical bi-degree sequence, it follows that d−
i nodes points to i. Since
−
k ≤ di , it is possible to delete these k in-edges with the tail node i in the graph. Clearly,
the bi-degree sequence of this graph is d − d′.
The next two propositions allow us to restrict the search for optimal bi-degree sequences in the set of bi-degree sequences, whose out-degrees are point-wise bounded by
z + after eliminating the negative coordinates of z + .
Proposition 3. Let z + = (z1+ , . . . , zn+ ) and z − = (z1− , . . . ,P
zn− ) be sequences
Pof integers.
+
−
+
+
Let I1 = {i : zi > 0} and I2 = {i : zi > 0}. Let fz (a) = i |zi − ai | + i |zi− − a−
i |.
Let d be any degree sequence such that f (d) = mina∈Bn fz (a) = d.
+ c
(1)If d+ (I1c ) > 0, then there exists a degree sequence d+
∗ such that d∗ (I ) = 0 and f (d) =
f (d∗).
− c
(2)If d− (I2c ) > 0, then there exists a degree sequence d−
∗ such that d∗ (I ) = 0 and f (d) =
f (d∗).
Proof. The proofs of the first part and second part are similar. We only give the proof of
c
the first part. If d+
i = 0, ∀i ∈ I1 , the proposition is true by letting d∗ = d. Hence assume
that there exists at least one i ∈ I c such that d+
i > 0. Let d∗ be the bi-degree sequence
d+
(i)
i
obtained from d by reducing it with a g
out-star sequence, as follows:
(
(
0,
k
=
i
d−
−
k − 1, k ∈ J
d+
=
,
d
=
,
∗k
∗k
+
c
dk , k 6= i
d−
k, k ∈ J
+
where J is the set of d+
i nodes to which the center node i points to. Here, J ⊂ {i : di > 0}
17
∗
and |J| = d+
i . By Proposition 2 (1), d∗ is graphical. Next let us show that f (d ) ≤ f (d).
X
X
f (d∗ ) =
|zi+ − d+
|
+
|zi− − d−
∗i
∗i |
i
=
X
j6=i
≤
=
X
j6=i
X
j6=i
≤
X
i
i
|zi+
−
d+
i |
+ |zi+ | +
+
|zi+ − d+
i | + |zi | +
X
i∈J
X
j∈J
|zi− − d−
i |+
+
+
|zi+ − d+
i | + |zi | + |di | +
|zi+ − d+
i |+
= f (d)
X
i
X
|zi− − d−
i + 1| +
|zi− − d−
i |
X
i
X
i∈J c
1+
i∈J
|zi− − d−
i |
|zi− − d−
i |
X
i∈J c
|zi− − d−
i |
But d is such that arg mina∈Bn fz (a) = d, hence f (d∗ ) = f (d). If there is more than one
j ∈ I1 such that dj > 0, we can redefine d∗ iteratively until there are no such j left.
+ −
Proposition
4.
P +
PLet−z = −(z , z ) be a bi-sequence of n nonnegative integers. Let fz (a) =
−
i |zi − ai | +
i |zi − ai |. Let d be any degree sequence such that f (d) = mina∈Bn fz (a).
+
(1)There exists a degree sequence d∗ such that d+
∗i ≤ zi , ∀ i and fz (d∗ ) = fz (d).
−
(2)There exists a degree sequence d∗ such that d∗i ≤ zi− , ∀ i and fz (d∗ ) = fz (d).
Proof. The proofs of parts (1) and (2) are similar and we only give the proof of part (1).
+
∗
If d+
i ≤ zi , ∀ i, the proposition is true by letting d = d. Hence assume that there exists
+
∗
at least one i such that d+
i > zi . Let d be defined as follows:
(
(
+
d−
z
,
k
=
i
−
k − 1, k ∈ I
k
,
,
d
=
d+
=
∗k
∗k
+
c
d−
dk , k 6= i
k, k ∈ I
+
∗
where I is the index set such that |I| = d+
i − zi . Clearly, by Proposition 2, d is a
bi-degree sequence because it is obtained by reducing d with a k-out-star sequence, where
+
k = zi+ − d+
i ≤ zi .
Next let us show that fz (d∗ ) ≤ fz (d).
X
X
fz (d∗ ) =
|zi+ − d+
|zi− − d−
∗i | +
∗i |
i
=
X
k6=i
≤
X
k6=i
i
|zk+
−
d+
k|
+ |zi+ − zi+ | +
X
k∈I
|zi− − d−
i + 1| +
+
+
|zk+ − d+
k | + |zi − zi | + |I| +
= fz (d).
X
k∈I
X
k∈I c
|zi− − d−
i |+
|zk− − d−
k|
X
k∈I c
|zk− − d−
k|
But d is such that f (d) = mina∈Bn fz (a), hence f (d∗ ) = f (d). If there is more than one i
+
such that d+
i > zi , we can redefine d∗ iteratively until there are no such i left.
18
The next lemma shows that we can always reduce the L1 distance of any bi-degree sequence d by replacing the k-out-star sequences in its directed Havel-Hakimi decomposition
with an appropriately chosen k-out-star sequences by solving a sequential optimization
problem. Let B≤z be the set of all possible bi-degrees sequences with their out-degrees
and in-degrees pointwise bounded by z + and z − , respectively. Let K≤z be the set of all
possible k-out-star sequences with their out-degrees and in-degrees pointwise bounded by
z + and z − , respectively.
Lemma 5. Let d be any bi-degree sequence in B≤z and let H(d) = {g ij }nj=1 be its directed
Havel-Hakimi decomposition where g ij is a k-out-star sequence centered at node ij . Let
xi1 , . . . , xin be the following k-out-star sequences defined recursively:
xi1 = arg
min
P
g∈K≤z + ,g+
ik+1
x
=
j6=1
g ij ∈B≤z +
fz (g),
fz (
argmin
g ∈ K≤z \ {xij }kj=1
Pn
ij
ij
j=1 x + g +
j=k+2 g ∈ B≤z
k
X
xij + g)
j=1
Pk
Let dk for k = 1, . . . , n be constructed sequentially by replacing the k-out-star sequence in
H(dk−1) centered at node ik by xik as follows:
1
i1
d =x +
X
ij
k
g ,
d =
k
X
ij
x +
j=1
j6=1
n
X
g ij .
j=k+1
Then, fz (dn ) ≤ fz (d) and each dk ∈ B≤z .
Proof. For two bi-sequences z and a, let kz − ak1 = kz + − a+ k1 + kz − − a− k1 . Then we
have
k
fz (d ) − fz (d
k+1
) = kz −
ik+1
= x
= fz (
k
X
j=1
−g
k
X
j=1
≥ 0,
ij
x −
ik+1
n
X
j=k+1
ij
g k1 − kz −
= kz −
k
X
j=1
xij + g ik+1 ) − fz (
ij
x −g
k
X
k+1
X
j=1
ik+1
ij
x −
n
X
j=k+2
k1 − kz −
g ij k 1
k
X
j=1
xij − xik+1 k1
xij + xik+1 )
j=1
where the second equality due to that each bi-sequence is pointwise bounded by z. Adding
these inequalities for k = 0 to k = n−1, we get fz (d0 )−fz (dn ) ≥ 0, as required. Moreover,
each dk is clearly a bi-degree sequence, as dk is obtained from dk+1 by replacing a k-out-star
sequence from its directed Havel-Hakimi decomposition.
Proposition 5. Given a non negative bi-sequence z, the element in the set K≤z that
19
solves the following optimization problem
min kz − gk1
g∈K≤z
is the k-out-star sequence of the following graph G∗ : Let i∗ = {i : zi+∗ = maxi zi+ }, and
k = zi∗ . Let I be the index set of k largest elements of z − excluding i∗ . In G∗ , add an
out-edge from i∗ to i for all i ∈ I.
Proof. Any k-out-star sequence can be selected by selecting a node c as center and connecting k out-edges from it to k other tail nodes. Thus, if E = {j: there exists an out-edge
from c to j }, then the objective function that we need to minimize is
X
X
|zi− − 1| + |zc+ − k| +
(|zi+ | + |zi− |).
i∈E
i∈E c \{c}
The result follows by noticing that the optimal k-out-star sequence can be selected by
first selecting the star center c and then selecting E. Clearly, the optimal center is the
node with highest “demand”, i.e., dc = di∗ = maxi zi+ . Next, connecting this node to di∗
nodes with highest “demand” gives the optimal k-out-star sequence.
We are now ready to present the proof of Theorem 1.
Proof of Theorem 1. Let d∗ be the optimal degree sequence. Let I1 = {zi : zi+ ≤ 0} and
I2 = {zi : zi− ≤ 0}. By Proposition 3, we can set d+
∗ (I1 ) = 0 and d∗ (I2 ) = 0. This is
done by Steps 2 and 4 of Algorithm 2. By Proposition 4, we reduce a global optimization
problem into a local optimization problem by restricting the bi-degree sequences bounded
point-wise by z. As a result, we only need to find the optimum over the set B≤z .
By Lemma 5, we can construct the optimal bi-degree sequence over B≤z by starting
with any bi-degree sequence d0 and replacing it by k-out-star sequences defined in Lemma
5. Since 0 is also a bi-degree sequence, we set d0 = 0. This is doneP
in Step 1. Then, using
n
the notation in Lemma 5, the optimal bi-degree sequence is d = nj=1 xij , where
ik+1
x
=
argmin
fz (
{xij }kj=1
g ∈ K≤z \
Pk
ij
j=1 x + g ∈ B≤z
k
X
xij + g)
j=1
Next
that Steps 3 to 10 of Algorithm 2 construct xij iteratively. Let z k = z −
Pk show
ij
j=1 x , then
fz k (g)
xik+1 =
argmin
g ∈ K≤z k \ {xij }kj=1
g ∈ B≤z k
Thus, each xik+1 can be found using the result in Proposition 5. Note that to enforce the
condition g ∈ K≤z k \ {xij }kj=1 , we need to exclude the nodes with non-positive in-degrees
from consideration. This is done in Step 4. Step 5 select i∗ (i.e., ik+1 ). Steps 7 and 8
decide the optimal set of in-neighborhoods of the center node i∗ according to Proposition
5. Note that step 7 is needed to make sure that the out-degree is not larger than the
20
number of nodes available to connect to. Finally, Steps 5 to 9 construct the optimal
bi-degree sequence xij = xi∗ and add the directed edges from i∗ pointing to nodes in I to
Gn .
7.2
Proof of Lemma 3
+
+
+
ˆ+
Proof. Note that dˆ+
i ≤ zi if zi ≥ 0 and di = 0 if zi < 0. Thus, we have
+
+
+
+
max |dˆ+
i − di | ≤ max |zi − di | = max |ei |.
i
i
i
−
−
Similarly, we also have maxi |dˆ−
i − di | ≤ maxi |ei |. Let e1 , . . . , en be independent and
p|e| , 0 < p < 1 and
identically distributed random variables with pmf P (e1 = e) = 1−p
1+p
e ∈ Z. By Proposition E in supplementary material of Karwa and Slavković (2016), we
have
2p[c]+1 n
P(max |ei | > c) = 1 − (1 −
) ,
i
1+p
Therefore, we have
+
−
ˆ−
P(max{max |d̂+
i − di |, max |di − di |} ≥ c)
i
≤
=
i
+
−
P (max{max |ei |, max |ei |}
i
i
−ǫn (c+1)/2
2e
)2n .
1 − (1 −
1 + e−ǫn /2
≥ c)
Note that when ǫn (c + 1) > 2 log 2, e−ǫn (c+1)/2 < 1/2. Here, ǫn (c + 1) ≥ 4 log n. Since
the function f (x) = 1 − (1 − x)n is an increasing function on x when x ∈ (0, 1), we have
1 − (1 −
2e−ǫn (c+1)/2 2n
) ≤ 1 − (1 − 2e−ǫn (c+1)/2 )2n .
1 + e−ǫn /2
On the other hand, (1 − x)n ≥ 1 − nx when x ∈ (0, 1). So, we have
1 − (1 − 2e−ǫn (c+1)/2 )2n ≤ 1 − (1 − 2n × 2e−ǫn (c+1)/2 ) = 4ne−ǫn (c+1)/2 .
When ǫn (c + 1) ≥ 4 log n, we have
4n
+
−
ˆ−
P(max{max |d̂+
→ 0.
i − di |, max |di − di |} ≥ c) ≤
i
i
n2
7.3
Proofs for Theorem 2
The technique to prove the consistency of the DP estimator here is partly similar to that
used in Yan et al. (2016). The main ideas for the proofs are constructing a Newton
iterative sequence that converges to the DP estimate and obtaining the convergence rate
of the sequence. For an ad hoc system of equations, Yan et al. (2016) establish a geometric
21
convergence of rate for the Newton iterative sequence.
Theorem 5 (Theorem 7 in Yan et al. (2016)). For a bi-sequence (d, b) with d = (d1 , . . . , dn )
and b = (b1 , . . . , bn ), define a system of equations:
Fi (θ) = di −
Fn+j (θ) = bj −
n
X
f (αi + βk ), i = 1, . . . , n,
k=1,k6=i
n
X
k=1,k6=j
f (αk + βj ), j = 1, . . . , n − 1,
F (θ) = (F1 (θ), . . . , Fn (θ), Fn+1 (θ), . . . , F2n−1 (θ))⊤ ,
(10)
where f (·) is a continuous function with the third derivative. Let D ⊂ R2n−1 be a convex
set and assume for any x, y, v ∈ D, we have
k[F ′ (x) − F ′ (x)]vk∞ ≤ K1 kx − yk∞ kvk∞ ,
max kFi′ (x) − Fi′ (y)k∞ ≤ K2 kx − yk∞ ,
i=1,...,2n−1
(11)
(12)
where F ′ (θ) is the Jacobin matrix of F on θ and Fi′ (θ) is the gradient function of Fi
on θ. Consider θ(0) ∈ D with Ω(θ(0) , 2r) ⊂ D, where r = k[F ′ (θ(0) )]−1 F (θ(0) )k∞ . For
any θ ∈ Ω(θ(0) , 2r), we assume that F ′ (θ) ∈ Ln (m, M) or −F ′ (θ) ∈ Ln (m, M). For
k = 1, 2, . . ., define the Newton iterates θ(k+1) = θ(k) − [F ′ (θ(k) )]−1 F (θ(k) ). Let
ρ=
c1 (2n − 1)M 2 K1
K2
+
.
3
2
2m n
(n − 1)m
(13)
If ρr < 1/2, then θ(k) ∈ Ω(θ(0) , 2r), k = 1, 2, . . ., are well-defined and satisfy
kθ(k+1) − θ(0) k∞ ≤ r/(1 − ρr).
(14)
Further, limk→∞ θ(k) exists and the limiting point is precisely the solution of F (θ) = 0 in
the range of θ ∈ Ω(θ(0) , 2r).
To apply the above theorem, it is natural to consider a system of DP equations in (8).
Then we need to verify the conditions (11) and (12) and to compute (13). Lemma 2 in
Yan et al. (2016) gives that under what assumption that these conditions hold, restated
below.
Lemma 6 (Lemma 2 in Yan et al. (2016)). Take D = R2n−1 and θ(0) = θ∗ in Theorem
5. Let f (αi + βj ) = eαi +βj /(1 + eαi +βj ) and F (θ) defined in (10) for a bi-sequence (d, b).
Assume that
p
kF (θ)k∞ ≤ c1 (n − 1) log(n − 1),
(15)
hold, where c1 is a constant. Then equations (11) and (12) hold with the numbers K1 , K2
and r in Theorem 5 as
K1 = n − 1, K2 =
n−1
(log n)1/2
6kθ ∗ k∞
2kθ ∗ k∞
, r≤
c
e
+
c
e
,
2
3
2
n1/2
22
where c2 and c3 are constants.
Next, we will verify that −F ′ (θ) ∈ Ln (m, M) with F (θ) defined at (9) and equation
(15) holds. The Jacobian matrix F ′ (θ) of F (θ) can be calculated as follows. For i =
1, . . . , n,
n
X
eαi +βk
∂Fi
∂Fi
= 0, l = 1, . . . , n, l 6= i;
=−
,
∂αl
∂αi
(1 + eαi +βk )2
k=1;k6=i
∂Fi
eαi +βj
∂Fi
=−
,
j
=
1,
.
.
.
,
n
−
1,
j
=
6
i;
=0
∂βj
(1 + eαi +βj )2
∂βi
and for j = 1, . . . , n − 1,
eαl +βj
∂Fn+j
∂Fn+j
=−
, l = 1, . . . , n, l 6= j;
= 0,
α
+β
2
j
l
∂αl
(1 + e
)
∂αj
n
X
∂Fn+j
eαk +βj
∂Fn+j
=−
= 0, l = 1, . . . , n − 1.
,
α
+β
2
j
k
∂βj
(1 + e
)
∂βl
k=1;k6=j
First, note that since the Jacobian is diagonally dominant with nonzero diagonals, it is
positive definite, implying that the DP equations (8) have a unique solution. Second, it is
easy to see that −F ′ (θ) ∈ Ln (m, M), thus Proposition 1 and Theorem 5 can be applied.
To verify equation (15), we depend on one lemma in Yan et al. (2016).
Lemma 7 (Lemma 3 in Yan et al. (2016)). With probability at least 1 − 4n/(n − 1)2 , for
a bi-degree sequence d = (d+ , d− ), we have
p
+
−
−
max{max |d+
(n − 1) log(n − 1).
i − E(di )|, max |dj − E(dj )|} ≤
i
j
In view of Lemma 7, it is left to bound kdˆ− Edk∞ in order to obtain (15), where k · k∞
is defined at (6). By Lemma 3 and (7), if ǫn = Ω(n−(1/2−ρ) ) for a fixed ρ ∈ (0, 1/2), then
we have
kdˆ − Edk∞ ≤ kdˆ − dk∞ + kd − Edk∞ = Op {(n log n)1/2 }.
(16)
We are now ready to present the proof of Theorem 2.
Proof of Theorem 2. Assume that condition (15) holds. In the Newton’s iterates, we
choose the initial value θ(0) as θ∗ . If θ ∈ Ω(θ∗ , 2r), then −F ′ (θ) ∈ Ln (m, M) with
∗
1
e2(kθ k∞ +2r)
M = , m=
.
4
(1 + e2(kθ∗ k∞ +2r) )2
(17)
If kθ∗ k∞ ≤ τ log n with the constant τ satisfying 0 < τ < 1/16, then as n → ∞,
∗
n−1/2 (log n)1/2 e8kθ k ≤ n−1/2+8τ (log n)1/2 → 0. By Lemma 6 and condition (15), for
23
sufficiently small r,
c1 (2n − 1)M 2 (n − 1)
(n − 1)
ρr ≤
+
2m3 n2
2m(n − 1)
1/2
(log n)
6kθ ∗ k∞
2kθ ∗ k∞
×
c
e
+
c
e
11
12
1/2
n
∗
∗
(log n)1/2 e8kθ k∞
(log n)1/2 e12kθ k∞
+O
.
≤ O
n1/2
n1/2
Therefore, if kθ∗ k∞ ≤ τ log n, then ρr → 0 as n → ∞. Consequently, by Theorem 5,
b then it satisfies
limn→∞ θb(n) exists. Denote the limit as θ,
∗
(log n)1/2 e8kθ k∞
∗
b
kθ − θ k∞ ≤ 2r = O
= o(1).
n1/2
By (16), condition (15) holds with probability approaching one if we set ǫn = n−(1/2−ρ)
for a fixed ρ ∈ (0, 1/2), thus the above inequality also holds with probability approaching
one. The uniqueness of the MLE due to that F ′ (θ) is positive definite.
7.4
Proofs for Theorem 3
The proof of the asymptotic normality of the DP estimator is partly similar to that of the
non-private case in Yan et al. (2016). Wherein Yan et al. work with the original bi-degree
ˆ The proof proceeds in three main steps.
sequence d, here we do with its DP estimator d.
First, we show that the first k elements of dˆ − Ed is asymptotical normality. Second, we
apply Taylor’s expansion to the system of equations in (8) and obtain the expression of
θ̂ − θ, where the main item is V −1 (dˆ− Ed). Third, we work with the approximate inverse
S given in (4), instead of V −1 , to bound the remainder. For sake of clarity of exposition,
we reproduce some results in Yan et al. (2016).
∗
Proposition 6 (Proposition 2 in Yan et al. (2016)). Assume that A ∼ Pθ∗ . If ekθ k∞ =
o(n1/2 ), then for any fixed k ≥ 1, as n → ∞, the vector consisting of the first k elements of
S{g − E(g)} is asymptotically multivariate normal with mean zero and covariance matrix
given by the upper left k × k block of S.
Lemma 8 (Lemma 8 Yan et al. (2016)). Let R = V −1 − S and U = Cov[R(g − Eg)].
Then
∗
(1 + e2kθ k∞ )4
−1
.
(18)
kUk ≤ kV − Sk + 4kθ∗ k∞
(n − 1)2
4e
Lemma 9. If kθ∗ k∞ ≤ τ log n and τ < 1/40, then for any i,
θbi − θi∗ = [V −1 (ĝ − Eg)]i + op (n−1/2 ).
(19)
Proof. The proof is very similar to the proof of Lemma 9 in Yan et al. (2016). It only
ˆ
requires verification of the fact that all the steps hold by replacing d with d.
The asymptotic normality of ĝ − Eg is stated in the following proposition.
24
∗
Proposition 7. Assume that A ∼ Pθ∗ and ǫn = Ω(n−(1/2−ρ) for a fixed ρ. If ekθ k∞ =
o(n1/2 ), then for any fixed k ≥ 1, as n → ∞, the vector consisting of the first k elements
of S(ĝ − Eg) is asymptotically multivariate normal with mean zero and covariance matrix
given by the upper left k × k block of S.
Proof. Recall that
eαi +βj
vij =
, i 6= j,
(1 + eαi +βj )2
vii =
n
X
vij .
j6=i,j=1
Since ex /(1 + ex )2 is an increasing function on x when x ≥ 0 and a decreasing function
when x ≤ 0, we have
∗
n−1
(n − 1)e2kθ k∞
≤
v
≤
, i = 1, . . . , 2n.
ii
∗
(1 + e2kθ k∞ )2
4
By Lemma 3 and (7), we have
kĝ − gk∞ = Op (n1/2−ρ log n),
So, we have
ĝi − gi gi − Egi
ĝi − Egi
=
+
vii
vii
vii
∗
1/2−ρ
n
log ne2kθ k∞
gi − Egi
= Op (
)+
n
vii
g
−
Eg
i
i
= op (n−1/2 ) +
vii
Consequently, Proposition 7 immediately follows Proposition 6.
Proof of Theorem 3. By Lemma 9 and noting that V −1 = S + R, we have
Since kĝ − gk∞
(θb − θ)i = [S(ĝ − Eg)]i + [R{ĝ − E(g)}]i + op (n−1/2 ).
√
= Op ( log n), we have
[R{ĝ − g}]i = Op (n
M2 p
log n) = op (n−1/2 ),
m3 n2
where M and m are given in (17). Combing Lemma 8, we have
[R(ĝ − Eg)]i = [R(ĝ − g)]i + [R(g − Eg)]i = op (n−1/2 ).
Therefore, Theorem 3 follows directly from Proposition 7.
25
Acknowledgements
Yan is partially supported by by the National Natural Science Foundation of China (No.
11401239).
References
Anderson, C. J., Wasserman S. and Crouch B. (1999). A p∗ primer: logit models for social
networks. Social Networks, 21, 37–66.
Barndorff-Nielsen, O. (1973). Exponential Families and Conditioning. Ph. D. thesis, Univ.
of Copenhagen.
Chatterjee, S., Diaconis, P., and Sly, A. (2011). Random graphs with a given degree
sequence. Annals of Applied Probability, 21, 1400–1435.
Dwork, C., Mcsherry, F., Nissim, K. and Smith, A. (2006). Calibrating noise to sensitivity
in private data analysis. Proceedings of the 3rd Theory of Cryptography Conference,
265–284.
Erdós, P. L., Péter L., Miklós, I., and Toroczkai, Z. (2010) A simple Havel-Hakimi
type algorithm to realize graphical degree sequences of directed graphs. The Electronic
Journal of Combinatorics, 17, Research Paper R66.
Fulkerson D. R. (1960). Zero-one matrices with zero trace. Pacific J. Math., 10, 831–836.
Hakimi, S. L. (1962). On realizability of a set of integers as degrees of the vertices of a
linear graph. I. Journal of the Society for Industrial and Applied Mathematics, 496–506.
Havel, V. (1955). A remark on the existence of finite graphs. Casopis Pest. Mat., 80,
477–480.
Hay, M., Li, C., Miklau, G. and Jensen, D. (2009). Accurate estimation of the degree
distribution of private networks. In Data Mining, 2009. ICDM09. Ninth IEEE International Conference on 169C178. IEEE.
Holland, P. W. and Leinhardt, S. (1981). An exponential family of probability distributions for directed graphs (with discussion). Journal of the American Statistical Association, 76, 33–65.
Hunter, D. R. (2004). MM algorithms for generalized BradleyCTerry models. The Annals
of Statistics , 32, 384–406.
Kasiviswanathan S.P., Nissim K., Raskhodnikova S., Smith A. (2013). Analyzing Graphs
with Node Differential Privacy. In: Sahai A. (eds) Theory of Cryptography. Lecture
Notes in Computer Science, vol 7785. Springer, Berlin, Heidelberg.
Karwa, V. and Slavković, A. (2016). Inference using noisy degrees-Differentially private
beta model and synthetic graphs. The Annals of Statistics, 44, 87–112.
Karwa V., Slavković A. and Krivitsky P. (2014). Differentially private exponential random
graph models, Available at https://arxiv.org/abs/1409.4696.
Karwa V., Krivitsky P. and Slavković A. (2017). Sharing social network data: differentially
private estimation of exponential family random-graph models. Journal of the Royal
Statistical Society: Series C (Applied Statistics), 66, 481–500.
Kleitman D. and Wang D. (1973). Algorithms for constructing graphs and digraphs with
given valences and factors. Discrete Math., 6, 79–88.
Lazega E. (2001). The Collegial Phenomenon: The Social Mechanisms of Cooperation
Among Peers in a Corporate Law Partnership. Oxford University Press. Oxford.
Lu W. and Miklau G. (2014). Exponential random graph estimation under differential pri26
vacy. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge
discovery and data mining (KDD ’14), ACM, New York, NY, USA, 921–930.
Majcher Z. (1985). Matrices representable by directed graphs. Archivum Mathematicum,
4, 205–218.
Nissim K., Raskhodnikova S. and Smith A. (2007). Smooth sensitivity and sampling in
private data analysis. In Proceedings of the thirty-ninth annual ACM Symposium on
Theory of Computing, 75–84. ACM.
Olhede, S. C. and Wolfe, P. J. (2012). Degree-based network models. Available at http://
arxiv.org/abs/1211.6537.
Parker, J. G. and Asher, S. R. (1993). Friendship and friendship quality in middle childhood: Links with peer group acceptance and feelings of loneliness and social dissatisfaction. Developmental Psychology, 29, 611–621.
Perry, P. O. and Wolfe, P. J. (2012). Null models for network data. Available at http://
arxiv.org/abs/1201.5871.
Rinaldo, A., Petrović, S., and Fienberg, S. E. (2013). Maximum likelihood estimation in
the β-model. The Annals of Statistics, 41, 1085–1110.
Sadeghi, K. and Rinaldo, A. (2014). Statistical models for degree distributions of networks, NIPS 2014 Workshop “From Graphs to Rich Data”. Available at http://arxiv.
org/abs/1411.3825.
Sampson, S. F. (1968). A novitiate in a period of change: An experimental and case study
of social relationships. PhD thesis, Cornell University, September.
Task C. and Clifton C. (2012). A guide to differential privacy theory in social network
analysis. 2012 IEEE/ACM International Conference on Advances in Social Networks
Analysis and Mining, 411–417
Wang D., Liau C. and Hsu T. (2006). Privacy protection in social network data disclosure
based on granular computing. 2006 IEEE International Conference on Fuzzy Systems,
997–1003
Wasserman, L. and Zhou, S. (2010). A statistical framework for differential privacy. J.
Amer. Statist. Assoc. 105, 375–389.
Wu, N. (1997). The maximum entropy method. New York, Springer.
Yan, T., Leng, C. and Zhu, J. (2016). Asymptotics in directed exponential random graph
models with an increasing bi-degree sequence. The Annals of Statistics, 44, 31–57.
Yan, T. and Xu, J. (2013). A central limit theorem in the β-model for undirected random
graphs with a diverging number of vertices. Biometrika, 100, 519–524.
Zhang, J. and Chen, Y. (2013). Sampling for Conditional Inference on Network Data.
Journal of the American Statistical Association, 108, 1295–1307.
Zhou B., Pei J. and Luk W. (2008). A brief survey on anonymization techniques for privacy
preserving publishing of social network data. ACM SIGKDD Explorations Newsletter
archive, 10, 12–22
27
| 10math.ST
|
Human Pose Regression by Combining Indirect Part Detection and Contextual
Information
Diogo C. Luvizon
Hedi Tabia
ETIS Lab., UMR 8051, Université Paris Seine,
Université Cergy-Pontoise, ENSEA, CNRS.
David Picard
arXiv:1710.02322v1 [cs.CV] 6 Oct 2017
{diogo.luvizon, hedi.tabia, picard}@ensea.fr
Abstract
Networks (CNN) [42] alongside consistent computational
power increase. This can be seen as the shift from classical
approaches [32, 25] to deep architectures. In many recent
works from different domains, CNN based methods have
overcome classical approaches by a large margin [18, 37].
A key benefit from CNN is that the full pipeline is differentiable, allowing end-to-end learning. In the context of
human pose estimation, the first methods using deep neural networks tried to do regression directly by learning a
non-linear mapping function from RGB images to joint coordinates [42]. By contrast, the majority of the methods
in the state of the art tackle pose estimation as a detection
problem by predicting heat maps that corresponds to joint
locations [27, 5]. In such methods, the ground truth is artificially generated from joint positions, generally as a 2D
Gaussian distribution centered on the joint location, while
the context information is implicitly learned by the hidden
convolutional layers.
Despite achieving state-of-the-art accuracy on 2D pose
estimation, detection based approaches have some limitations. For example, such methods relies on additional steps
to convert heat maps to joint positions, usually by applying
the argmax function, which is not differentiable, breaking
the learning chain on neural networks. Additionally, the
precision of predicted keypoints is proportional to that of
the heat map resolution, which leads the top ranked methods [13, 27] to high memory consumption and high computational requirements.
Moreover, regression based methods are conceptually
more adapted to 2D and 3D scenarios and can be used indistinctly on both cases [36, 38]. However, the regression
function map is sub-optimally learned, resulting in lower
results if compared to detection based approaches. In this
paper, we aim at solving this problem by bridging the gap
between detection and regression based methods. We propose to replace the argmax function, used to convert heat
maps into joint locations, by what we call the Soft-argmax
function, which keeps the properties of specialized part detectors while being fully differentiable. With this solution,
In this paper, we propose an end-to-end trainable regression approach for human pose estimation from still images.
We use the proposed Soft-argmax function to convert feature maps directly to joint coordinates, resulting in a fully
differentiable framework. Our method is able to learn heat
maps representations indirectly, without additional steps of
artificial ground truth generation. Consequently, contextual information can be included to the pose predictions
in a seamless way. We evaluated our method on two very
challenging datasets, the Leeds Sports Poses (LSP) and the
MPII Human Pose datasets, reaching the best performance
among all the existing regression methods and comparable
results to the state-of-the-art detection based approaches.
1. Introduction
Human pose estimation from still images is a hard task
since the human body is strongly articulated, some parts
may not be visible due to occlusions or low quality images,
and the visual appearance of body parts can change significantly from one pose to another. Classical methods use
keypoint detectors to extract local information, which are
combined to build pictorial structures [16]. To handle difficult cases of occlusion or partial visualization, contextual
information is usually needed to provide visual cues that
can be extracted from a broad region around the part location [15] or by interaction among detected parts [45]. In
general, pose estimation can be seen from two different perspectives, namely as a correlated part detection problem or
as a regression problem. Detection based approaches commonly try to detect keypoints individually, which are aggregated in post-processing stages to form one pose prediction.
In contrast, methods based on regression use a function to
map directly input images to body joint positions.
In the last few years, pose estimation have gained attention with the breakthrough of deep Convolutional Neural
1
we are able to optimize our model from end-to-end using regression losses, i.e., from input RGB images to final (x, y)
body joint coordinates.
The contributions of our work are the following: first,
we present a human pose regression approach from still
images based on the Soft-argmax function, resulting in an
end-to-end trainable method which does not require artificial heat maps generation for training. Second, the proposed
method can be trained using an insightful regression loss
function, which is directly linked to the error distance between predicted and ground truth joint positions. Third, in
the proposed architecture, contextual information is directly
accessible and is easily aggregated to the final predictions.
Finally, the accuracy reached by our method surpasses that
of regression methods ans is close to that of state-of-theart detection methods, despite using a much smaller network. Some examples of our regressed poses are shown in
Fig. 1. Additionally, we provide our implementation of the
proposed method in Python using the open source Keras library [10].1
The rest of this paper is divided as follows. In the next
section, we present a review of the most relevant related
work. The proposed method is presented in Section 3. In
Section 4, we show the experimental evaluations, followed
by our conclusions in Section 5.
2. Related work
(a) Input image
Several approaches for human pose estimation have been
presented for both 2D [14, 26] and 3D [22] scenarios, as
well as for video sequences [29]. Among classical methods, Pictorial Structures [2] and poselet-based features [31]
have been widely used in the past. In this section, due to
limited space, we focus on CNN based methods that are
more related to our work. We briefly refer to the most recent works, splitting them as regression based and detection
based approaches.
Regression based approaches. Some methods tackled
pose estimation as a keypoint regression problem. One of
the first regression approaches was proposed by Toshev and
Szegedy [42] as a holistic solution based on cascade regression for body part detection, where individual joint positions are recursively improved, taking the full frame as input. Pfister et al. [30] proposed the Temporal Pose ConvNet
to track upper body parts, and Carreira et al. [6] proposed
the Iterative Error Feedback by injecting the prediction error
back to the input space, improving estimations recursively.
Recently, Sun et al. [38] proposed a structured bone based
representation for human pose, which is statistically less
variant than absolute joint positions and can be indistinctly
used for both 2D and 3D representations. However, the
(b) Part-based maps
(c) Pose
Figure 1: Test samples from the Leeds Sports Poses (LSP)
dataset. Input image (a), the predicted part-based maps encoded as RGB image for visualizasion (b), and the regressed
pose (c). Corresponding human limbs have the same colors
in all images. This figure is better seen in color.
method requires converting pose data to the relative bone
based format. Moreover, those results are all outperformed
by detection based methods.
Detection based approaches. Pischulin et al. [33] proposed DeepCut, a graph cutting algorithm that relies on
body parts detected by DeepPose [42]. This method has
been improved in [21] by replacing the previous CNN by
a deep Residual Network (ResNet) [19], resulting in very
competitive accuracy results, specially on multi-person detection. More recent methods have shown significant improvements on accuracy by using fully convolutional models to generate belief maps (or heat maps) for joint probabilities [34, 3, 5, 27, 13]. For example, Bulat et al. [5]
proposed a two-stages CNN for coarse and fine heat map regression using pre-trained models, and Gkioxari et al. [17]
presented a structured prediction method, where the prediction of each joint depends on the intermediate feature
maps and the distribution probability of the previously pre-
1 The Python source code will be publicly available after acceptance at
https://github.com/dluvizon/pose-regression.
2
dicted joints. Following the tendency of deeper models
with residual connections, Newell et al. [27] proposed a
stacked hourglass network with convolutions in multi-level
features, allowing reevaluation of previous estimations due
to a stacked block architecture with many intermediate supervisions. The part-based learning process can benefit
from intermediate supervision because it acts as constraints
on the lower level layers. As a result, the feature maps on
higher levels tend to be cleaner. Recently, Chu et al. [13]
extended the stacked hourglass network by increasing the
network complexity and using a CRF based attention map
alongside intermediate supervisions to reinforce the task of
learning structural information. These results provide to our
knowledge state-of-the-art performance.
All the previous methods that are based on detection
need additional steps on training to produce artificial ground
truth from joint positions, which represent an additional
processing stage and additional hype-parameters, since the
ground truth heat maps have to be defined by hand. On
evaluation, the inverse operation is required, i.e., heat maps
have to be converted to joint positions, generally using the
argmax function. Consequently, in order to achieve good
precision, predicted heat maps need reasonable spacial resolution, which increases quadratically the computational cost
and memory usage. In order to provide an alternative to
heat maps based approaches, we present our framework in
the following section.
maps that are output by Block-B. Similarly to recent approaches [27, 13], on each prediction block we produce one
estimation that is used as intermediate supervision, providing better accuracy and more stability to the learning process. As a convention, we use the generic term “heat map”
to refer both to part-based and contextual feature maps,
since these feature maps converge to heat maps like representations.
The proposed CNN model is partially based on
Inception-v4 [39] and on the Stacked Hourglass [27] networks. We also inspired our model on the “extreme” inception (Xception) [9] network, which relies on the premise
that convolutions can be separated into spatial convolutions
(individual for each channel) followed by a 1 × 1 convolution for cross-channel projection, resulting in significant reduction on the number of parameters and on computations.
This idea is called depthwise separable convolution (SepConv) and an optimized implementation is available on the
TensorFlow framework.
In our network, the “Stem” is based on Inception-v4’s
stem followed by a SepConv layer in parallel with a shortcut layer, as presented in Fig. 3a. For Block-A, we use
a similar architecture as the Stacked Hourglass replacing
all the residual blocks by a residual separable convolution
(Res-SepConv), as depicted in Fig. 3b. Additionally, our
approach increased the results from [27] with only three feature map resolutions, from 32 × 32 to 8 × 8, instead of the
original five resolutions, from 64 × 64 to 4 × 4.
At each prediction stage, Block-B is used to transform
input feature maps into Md part-based detection maps (Hd )
and Mc context maps (Hc ), resulting in M = Md +Mc heat
maps. For the specific problem of pose estimation, Md corresponds to the number of joints NJ , and Mc = Nc NJ ,
where Nc is the number of context maps for each joint.
The produced heat maps are projected back to the feature
space and reintroduced to the network flow by a 1 × 1
convolution. Similar techniques have been used by many
previous works [5, 27, 13], resulting in significant gain of
performance. From the generated heat maps, our method
computes the predicted joint locations and joint probabilities in the regression block, which has no trainable parameters. The architecture of Block-B and the regression stage
is shown in Fig. 4.
3. Proposed method
The proposed approach is an end-to-end trainable network which takes as input RGB images and outputs two
vectors: the probability pn of joint n being in the image
and the regressed joint coordinates yn = (xn , yn ), where
n = {1, 2, . . . , NJ } is the index of each joint and NJ is
the number of joints. In what follows, we first present the
global architecture of our method, and then detail its most
important parts.
3.1. Network architecture
An overview of the proposed method is presented in
Fig. 2. Our approach is based on a convolutional neural network essentially composed of three parts: one entry
flow (Stem), Block-A and Block-B. The role of the stem
is to provide basic feature extraction, while Block-A provides refined features and Block-B provides body-part and
contextual activation maps. One sequence of Block-A and
Block-B is used to build one prediction block, which output is used as intermediate supervision during training. The
full network is composed by the stem and a sequence of
K prediction blocks. The final prediction is the output of
the K th prediction block. To predict the pose at each prediction block, we aggregate the 2D coordinates generated
by applying Soft-argmax to the part-based and contextual
3.2. Proposed regression method
As presented in Section 2, traditional regression based
methods use fully connected layers on feature maps and
learn the regression mapping. However, this approach usually gives sub-optimal solutions. While state-of-the-art
methods are overwhelmingly based on part detection, approaches based on regression have the advantages of providing directly the pose prediction as joint coordinates without
additional steps or post-processing. In order to provide an
3
CNN
Input
frame
Stem
6
25
+
Block-A1
Block-B1
...
+
+
+
Block-AK
Block-BK
3
×
Feature
maps
6×
25
Context
maps
Part-based
maps
...
...
SAM
Aggregation
Prediction
block 1
32×32×Mc
...
Hc,1
Hd,1
SAM
32×32×Md
32×32×576
Hd,K
Pose +
probability
Supervision
SAM
Hc,K
Aggregation
Prediction
block K
Pose +
probability
SAM
Supervision
Figure 2: Overview of the proposed approach for pose regression.
Output (32×32×576)
+
Shortcut
SepConv
Conv
(576,3×3)
(576,1×1)
Output (32×32×Nf)
Inception-v4
"Stem"
Output (W×H×Nfout)
+
Shortcut
SepConv
Conv
(Nfout,5×5)
(Nfout,1×1)
Input (256×256×3)
Input (W×H×Nfin)
(a)
(b)
Supervision
Conv
(Nf,1×1)
Predicted
pose
Joint
regression
Heat
maps
Figure 3: In the proposed network architecture, the Stem
(a) is based on Inception-v4’s stem [39] followed by a separable convolution in parallel to a shortcut connection. In
(b), we present the residual separable convolution (ResSepConv), used to replace the residual block in the Stacked
Hourglass [27] model. If Nf in is equal to Nf out , the shortcut convolution is replaced by the identity mapping.
Joint
probability
SepConv
(Nf,5×5)
Supervision
Figure 4: Network architecture of Block-B and an overview
of the regression stage. The input is projected into M heat
maps (Md + Mc ) which are then used for pose regression.
ensures each heat maps is normalized. Then, we define the
Soft-argmax as follows:
Ψd (h) =
k=1
l=1
ehk,l
,
W X
H
X
Wi,j,d Φ(hi,j ),
(2)
i=1 j=1
Let us define the Softmax operation on a single heat map
h ∈ RW ×H as:
Φ(hi,j ) = PW
Regression
stage
Input (32×32×Nf)
Soft-argmax layer
ehi,j
PH
MaxPooling
+ Activation
Conv
(M,1×1)
Block-B
alternative to detection based methods, we propose an efficient and fully differentiable way to convert heat maps directly to (x , y) coordinates, which we call Soft-argmax. Additionally, the Soft-argmax operation can be implemented as
a CNN layer, as detailed in the next section.
3.2.1
+
+
where d is a given component x or y, and W is a W ×H ×2
weight matrix corresponding to the coordinates (x , y). The
matrix W can be expressed by its components Wx and
Wy , which are 2D discrete normalized ramps, defined as
follows:
(1)
where hi,j is the value of heat map h at location (i , j ), and
W × H is the heat map size. Contrary to the more common
cross-channel softmax, we use here a spatial softmax that
Wi,j,x =
4
i
j
, Wi,j,y = .
W
H
(3)
Finally, given a heat map h, the regressed location of the
predicted joint is given by
y = (Ψx (h), Ψy (h))T .
specialized part-based heat maps and context heat maps,
which are defined as Hd = [hd1 , . . . , hdNJ ] and Hc =
[hc1,1 , . . . , hcNc ,Nj ], respectively. Additionally, we define
the joint probability related to each context map as pci,n ,
where i = {1, . . . , Nc } and n = {1, . . . , Nj }.
Finally, the nth joint position from detection and contextual information aggregated is given by:
(4)
This Soft-argmax operation can be seen as a weighted
average of points distributed on an uniform grid, with the
weights being equal to the corresponding heat map.
In order to integrate the Soft-argmax layer into a deep
network, we need its derivative with respect to h:
PNc
yn =
PW PH
ehi,j ( k=1 l=1 ehk,l − ehi,j )
∂Ψd (hi,j )
.
= Wi,j,d
PW PH
∂hi,j
( k=1 l=1 ehk,l )2
(5)
The Soft-argmax function can thus be integrated in a trainable framework by using back propagation and the chain
rule on equation (5). Moreover, from equation (5), we can
see that the gradient is exponentially increasing for higher
values, resulting in very discriminative response at the joint
position.
The implementation of Soft-argmax can be easily done
with recent frameworks, such as TensorFlow, just by concatenating a spatial softmax followed by one convolutional
layer with 2 filters of size W × H , with fixed parameters
according to equation (3).
Unlike traditional argmax, Soft-argmax provides subpixel accuracy, allowing good precision even with very low
resolution. Moreover, the Soft-argmax operation allows
to learn very discriminative heat maps directly from the
(x , y) joint coordinates without explicitly computing artificial ground truth. Samples of heat maps learned by our
approach are shown in Fig. 5.
3.2.2
c
c
i=1 pi,n yi,n
,
+ (1 − α) P
Nc
c
i=1 pi,n
(6)
where ynd = Soft-argmax(hdn ) is the predicted location from
c
= Soft-argmax(hci,n ) and
the nth part based heat map, yi,n
c
pi,n are respectively the location and the probability for the
ith context heat map for joint n, and α is a hyper-parameter.
From equation (6) we can see that the final prediction is
a combination of one specialized prediction and Nc contextual predictions pondered by their probabilities. The
contextual weighted contribution brings flexibility, allowing
specific filters to be more responsive to particular patterns.
This aggregation scheme within the learning stage is only
possible because we have the joint probability and position
directly available inside the network in a differentiable way.
4. Experiments
We evaluate the proposed method on the very challenging MPII Human Pose [1] and Leeds Sports Poses
(LSP) [23] datasets. The MPII dataset contains 25K images collected from YouTube videos, including around 28K
annotated poses for training and 15K poses for testing. The
annotated poses have 16 body joints, some of them are not
present and others are occluded but can be predicted by the
context. The LSP dataset is composed by 2000 annotated
poses with up to 14 joint locations. The images were gethered from Flickr with sports people. The details about training the model and achieved accuracy results are given as
follows.
Joint probability
Additionally to the joint locations, we estimate the joint
probability pn , which corresponds to the probability of the
nth joint being present in the image. The estimated joint
probability is given by the sigmoid activation on the global
max-pooling from heat map hn . Despite giving an additional piece of information, the joint probability does not
depends on additional parameters and is computationally
negligible, compared to the cost of convolutional layers.
3.2.3
αynd
4.1. Training
The proposed network was trained simultaneously on
joints regression and joint probabilities. For joints regression, we use the elastic net loss function (L1 + L2):
Detection and context aggregation
Even if the correlation between some joints can be learned
in the hidden convolutional layers, the joint regression approach is designed to locate body parts individually, resulting in low flexibility to learn from the context. For example, the same filters that give high response to images of
a clean head, also must react positively to a hat or a pair
of sunglasses. In order to provide multi-source information to the final prediction, we include in our framework
Ly =
NJ
1 X
kyn − ŷn k1 + kyn − ŷn k22 ,
NJ n=1
(7)
where yn and ŷn are respectively the ground truth and the
predicted nth joint coordinates. In this case, we use directly the joint coordinates normalized to the interval [0, 1],
where the top-left image corner corresponds to (0, 0), and
the bottom-right image corner corresponds to (1, 1).
5
Table 1: Results on LSP test samples using the PCK measure at 0.2 with OC annotations.
For joint probability estimation, we use the binary cross
entropy loss function on the joint probability p:
Lp =
Method
NJ
1 X
[(pn − 1) log (1 − p̂n ) − pn log p̂n ], (8)
NJ n=1
Kiefel and Gehler [24]
Ramakrishna et al. [35]
Pishchulin et al. [32]
Ouyang et al. [28]
Chen and Yuille [7]
Yang et al. [44]
Chu et al. [12]
Pishchulin et al. [33]
where pn and p̂n are respectively the ground truth and the
predicted joint probability.
We optimize the network using back propagation and the
RMSProp optimizer, with batch size of 16 samples. For the
MPII dataset, we train the network for 120 epochs. The
learning rate begins at 10−3 and decreases by a factor of
0.4 when accuracy on validation plateaus. We use the same
validation split as proposed in [40]. On the LSP dataset, we
start from the model trained on MPII and fine-tuned it for
more 70 epochs, beginning with learning rate 2 · 10−5 and
using the same decrease procedure. The full training of our
network takes three days on the relatively outdated NVIDIA
GPU Tesla K20 with 5GB of memory.
Data augmentation. We used standard data augmentation on both MPII and LSP datasets. Input RGB images were cropped and centered on the main subject with
a squared bounding box, keeping the people scale (when
provided), then resized to 256 × 256 pixels. We perform
random rotations (±40◦ ) and random rescaling from 0.7 to
1.3 on MPII and from 0.85 to 1.25 on LSP to make the
model more robust to image changes.
Parameters setup. The network model is defined according to Fig. 2 and composed of eight prediction blocks
(K = 8). We trained the network to regress 16 joints with
2 context maps for each joint (Nj = 16, Nc = 2). In the
aggregation stage, we use α = 0.8.
Our method
Head Sho. Elb. Wri.
Detection based methods
83.5 73.7 55.9 36.2
84.9 77.8 61.4 47.2
87.5 77.6 61.4 47.6
86.5 78.2 61.7 49.3
91.5 84.7 70.3 63.2
90.6 89.1 80.3 73.5
93.7 87.2 78.2 73.8
97.4 92.0 83.8 79.0
Regression based method
97.4 93.8 86.8 82.3
Hip
Knee
Ank.
PCK
73.7
73.6
79.0
76.9
82.7
85.5
88.2
93.1
70.5
69.1
75.2
70.0
78.1
82.8
83.0
88.3
66.9
68.8
68.4
67.6
72.0
68.8
80.9
83.7
65.8
69.0
71.0
70.0
77.5
81.5
83.6
88.2
93.7
90.9
88.3
90.5
Table 2: Results on LSP test samples using the PCP measure with OC annotations.
Method
Kiefel and Gehler [24]
Pishchulin et al. [31]
Ramakrishna et al. [35]
Ouyang et al. [28]
Pishchulin et al. [32]
Chen and Yuille [7]
Yang et al. [44]
Chu et al. [12]
Pishchulin et al. [33]
Our method
Torso
Upper Lower Upper
leg
leg
arm
Detection based methods
84.3
74.5
67.6
54.1
87.4
75.7
68.0
54.4
88.1
79.0
73.6
62.8
88.6
77.8
71.9
61.9
88.7
78.9
73.2
61.8
92.7
82.9
77.0
69.2
96.5
88.7
81.7
78.8
95.4
87.6
83.2
76.9
96.0
91.0
83.5
82.8
Regression based method
98.2
93.8
89.8
85.8
Forearm
Head
PCP
28.3
33.7
39.5
45.4
45.0
55.4
66.7
65.2
71.8
78.3
77.4
80.4
84.3
85.1
87.8
83.1
89.6
96.2
61.2
62.8
67.8
68.7
69.2
75.0
81.1
81.1
85.0
75.5
96.0
88.4
Table 3: Results on LSP test samples using the PCK measure at 0.2 with PC annotations.
Method
Pishchulin et al. [32]
Chen and Yuille [7]
Fan et al. [15]
Tompson et al. [41]
Yang et al. [44]
Rafi et al. [34]
Yu et al. [46]
Belag. and Ziss. [4]
Lifshitz et al. [26]
Pishchulin et al. [33]
Insafutdinov et al. [21]
Wei et al. [43]
Bulat and Tzimi. [5]
Chu et al. [13]
4.2. Results
LSP dataset. We evaluate our method on the LSP
dataset using two metrics, the “Percentage of Correct Parts”
(PCP) and the “Probability of Correct Keypoint” (PCK)
measures, as well as two different evaluation protocols,
“Observer-Centric“ (OC) and “Person-Centric“ (PC), resulting in four different evaluation settings. Our results
compared to the state-of-the-art results on the LSP dataset
using OC annotations are present in Table 1 (PCK measure)
and Table 2 (PCP measure). In both cases, we overcome
the best scores by a significant margin, specially with respect to the lower leg and the ankles, on which we increase
the results of Pishchulin et al. [33] by 6.3% and 4.6%, respectively. To the best of our knowledge, we are the sole
regression method to report results using this evaluation settings.
Using the PC annotations on LSP, we achieved the best
results among regression based approaches and the second
general score, as presented in Tables 3 and 4. On the PCK
measure, we outperform the results reported by Carreira et
Carreira et al. [6]
Our method
Head Sho. Elb. Wri.
Detection based methods
87.2 56.7 46.7 38.0
91.8 78.2 71.8 65.5
92.4 75.2 65.3 64.0
90.6 79.2 67.9 63.4
90.6 78.1 73.8 68.8
95.8 86.2 79.3 75.0
87.2 88.2 82.4 76.3
95.2 89.0 81.5 77.0
96.8 89.0 82.7 79.1
97.0 91.0 83.8 78.1
97.4 92.7 87.5 84.4
97.8 92.5 87.0 83.9
97.2 92.1 88.1 85.2
98.1 93.7 89.3 86.9
Regression based methods
90.5 81.8 65.8 59.8
97.5 93.3 87.6 84.6
Hip
Knee
Ank.
PCK
61.0
73.3
75.7
69.5
74.8
86.6
91.4
83.7
90.9
91.0
91.5
91.5
92.2
93.4
57.5
70.2
68.3
71.0
69.9
83.8
85.8
87.0
86.0
86.7
89.9
90.8
91.4
94.0
52.7
63.4
70.4
64.2
58.9
79.8
78.7
82.8
82.5
82.0
87.2
89.9
88.7
92.5
57.1
73.4
73.0
72.3
73.6
83.8
84.3
85.2
86.7
87.1
90.1
90.5
90.7
92.6
81.6
92.8
70.6
92.0
62.0
90.0
73.1
91.1
al. [6] (CVPR 2016), which is the only regression method
reported on this setup, by 18.0%.
MPII dataset. On the MPII dataset, we evaluate our
method using the “Single person” challenge [1]. The scores
were computed by the providers of the dataset, since the test
labels are not publicly available. As shown in Table 5, we
reached a test score of 91.2%, which is only 0.7% lower
then the best result using detection based method, and 4.8%
6
Table 4: Results on LSP test samples using the PCP measure with PC annotations.
Method
Torso
Pishchulin et al. [32]
Tompson et al. [41]
Fan et al. [15]
Chen and Yuille [7]
Yang et al. [44]
Rafi et al. [34]
Belag. and Ziss. [4]
Yu et al. [46]
Lifshitz et al. [26]
Pishchulin et al. [33]
Insafutdinov et al. [21]
Wei et al. [43]
Bulat and Tzimi. [5]
Chu et al. [13]
Carreira et al. [6]
Our method
Upper Lower Upper
leg
leg
arm
Detection based methods
88.7
63.6
58.4
46.0
90.3
70.4
61.1
63.0
95.4
77.7
69.8
62.8
96.0
77.2
72.2
69.7
95.6
78.5
71.8
72.2
97.6
87.3
80.2
76.8
96.0
86.7
82.2
79.4
98.0
93.1
88.1
82.9
97.3
88.8
84.4
80.6
97.0
88.8
82.0
82.4
97.0
90.6
86.9
86.1
98.0
92.2
89.1
85.8
97.7
92.4
89.3
86.7
98.4
95.0
92.8
88.5
Regression based methods
95.3
81.8
73.3
66.7
98.2
93.6
91.0
86.6
Forearm
Head
PCP
35.2
51.2
49.1
58.1
61.8
66.2
69.4
72.6
71.4
71.8
79.5
77.9
79.7
81.2
85.1
83.7
86.6
85.6
83.9
93.3
89.4
83.0
94.8
95.8
95.4
95.0
95.2
95.7
58.0
66.6
70.1
73.6
74.8
81.2
82.1
85.4
84.3
84.3
87.8
88.3
88.9
90.9
51.0
78.2
84.4
96.8
72.5
89.4
higher than the second score using regression.
Taking into account the competitiveness of the MPII Human Pose challenge2 , our score represents a very significant improvement over regression based approaches and
a promising result compared to detection based methods.
Moreover, our method is much simpler than the stacked
hourglass network from Newell et al. [27] or its extension
from Chu et al. [13]. Due to limited memory resources, we
were not able to train these two models in our hardware.
Despite that, we reach comparable results with a model that
fits in much smaller GPUs.
Figure 5: Indirectly learned part-based heat maps from our
method. All the joints encoded to RGB are shown in the first
image (top-left corner) and the final pose is shown in the last
image (bottom-right corner). On each column, the intermediate images correspond to the predicted heat maps before
(left) and after (right) the Softmax normalization. The presented heat maps correspond to right ankle, right hip, right
wrist, right shoulder, upper neck, head top, left knee, and
left wrist.
4.3. Discussion
As suggested in Section 3.2.1, the proposed Soft-argmax
function acts as a constrain on the regression approach,
driving the network to learn part-based detectors indirectly.
This effect provides the flexibility of regression based methods, which can be easily integrated to provide 2D pose estimation to other applications such as 3D pose estimation or
action recognition, while preserving the performance of detection based methods. Some examples of part-based maps
indirectly learned by our method are show in Fig. 5. As we
can see, the responses are very well localized on the true
location of the joints without explicitly requiring so.
Additionally to the part-based maps, the contextual maps
give extra information to refine the predicted pose. In some
cases, the contextual maps provide strong responses to regions around the joint location. In such cases, the aggregation scheme is able to refine the predicted joint position.
On the other hand, if the contextual map response is weak,
the context reflects in very few changes on the pose. Some
examples of predicted poses and visual contributions from
2 MPII
contextual aggregation are shown in Fig. 6. The contextual
maps are able to increase the precision of the predictions
by providing complementary information, as we can see for
the right elbows of the poses in Fig. 6.
5. Conclusion
In this work, we presented a new regression method for
human pose estimation from still images. The method is
based on the Soft-argmax operation, a differentiable operation that can be integrated in a deep convolutional network to learn part-based detection maps indirectly, resulting in a significant improvement over the state-of-the-art
scores from regression methods and very competitive re-
Leader Board: http://human-pose.mpi-inf.mpg.de
7
Table 5: Comparison results with state-of-the-art methods on the MPII dataset on testing, using PCKh measure with threshold
as 0.5 of the head segment length. Detection based methods are shown on top and regression based methods on bottom.
Method
Pishchulin et al. [32]
Tompson et al. [41]
Tompson et al. [40]
Hu and Ramanan [20]
Pishchulin et al. [33]
Lifshitz et al. [26]
Gkioxary et al. [17]
Rafi et al. [34]
Belagiannis and Zisserman [4]
Insafutdinov et al. [21]
Wei et al. [43]
Bulat and Tzimiropoulos [5]
Newell et al. [27]
Chu et al. [13]
Chou et al. [11]
Chen et al. [8]
Rogez et al. [36]
Carreira et al. [6]
Sun et al. [38]
Our method
(a)
(b)
Head Shoulder Elbow Wrist
Detection based methods
74.3
49.0
40.8
34.1
95.8
90.3
80.5
74.3
96.1
91.9
83.9
77.8
95.0
91.6
83.0
76.6
94.1
90.2
83.4
77.3
97.8
93.3
85.7
80.4
96.2
93.1
86.7
82.1
97.2
93.9
86.4
81.3
97.7
95.0
88.2
83.0
96.8
95.2
89.3
84.4
97.8
95.0
88.7
84.0
97.9
95.1
89.9
85.3
98.2
96.3
91.2
87.1
98.5
96.3
91.9
88.1
98.2
96.8
92.2
88.0
98.1
96.5
92.5
88.5
Regression based methods
–
–
–
–
95.7
91.7
81.7
72.4
97.5
94.3
87.0
81.2
98.1
96.6
92.0
87.5
(c)
(d)
Hip
Knee
Ankle
Total
36.5
77.6
80.9
81.9
82.6
85.3
85.2
86.8
87.9
88.4
88.4
89.4
90.1
90.6
91.3
90.2
34.4
69.7
72.3
74.5
75.7
76.6
81.4
80.6
82.6
83.4
82.8
85.7
87.4
88.0
89.1
89.6
35.2
62.8
64.8
69.5
68.6
70.2
74.1
73.4
78.4
78.0
79.4
81.7
83.6
85.0
84.9
86.0
44.1
79.6
82.0
82.4
82.4
85.0
86.1
86.3
88.1
88.5
88.5
89.7
90.9
91.5
91.8
91.9
–
82.8
86.5
90.6
–
73.2
78.5
88.0
–
66.4
75.4
82.7
74.2
81.3
86.4
91.2
(e)
(f)
Figure 6: Samples of context maps aggregated to refine predicted pose. Input image (a), part-based detection maps (b),
predicted pose without context (c), two different context maps (d) and (e), and the final pose with aggregated predictions (f).
8
sults compared to detection based approaches. Additionally, we demonstrate that contextual information can be
seamless integrated into our framework by using additional
context maps and joint probabilities. As a future work, other
methods could be build up to our approach to provide 3D
pose estimation or human action recognition from pose in a
fully differentiable way.
[14] M. Dantone, J. Gall, C. Leistner, and L. V. Gool. Human
Pose Estimation Using Body Parts Dependent Joint Regressors. In Computer Vision and Pattern Recognition (CVPR),
pages 3041–3048, June 2013.
[15] X. Fan, K. Zheng, Y. Lin, and S. Wang. Combining local
appearance and holistic view: Dual-source deep neural networks for human pose estimation. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), June
2015.
[16] P. F. Felzenszwalb and D. P. Huttenlocher. Pictorial structures for object recognition. International Journal of Computer Vision, 61(1):55–79, 2005.
[17] G. Gkioxari, A. Toshev, and N. Jaitly. Chained Predictions
Using Convolutional Neural Networks. European Conference on Computer Vision (ECCV), 2016.
[18] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), June 2016.
[19] K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning
for Image Recognition. In IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2016.
[20] P. Hu and D. Ramanan. Bottom-up and top-down reasoning with convolutional latent-variable models. In 2016 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR). 2016.
[21] E. Insafutdinov, L. Pishchulin, B. Andres, M. Andriluka,
and B. Schiele. DeeperCut: A Deeper, Stronger, and Faster
Multi-Person Pose Estimation Model. In European Conference on Computer Vision (ECCV), May 2016.
[22] C. Ionescu, F. Li, and C. Sminchisescu. Latent structured
models for human pose estimation. In International Conference on Computer Vision (ICCV), pages 2220–2227, Nov
2011.
[23] S. Johnson and M. Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In Proceedings of the British Machine Vision Conference, 2010.
[24] M. Kiefel and P. V. Gehler. Human Pose Estimation with
Fields of Parts, pages 331–346. Springer International Publishing, Cham, 2014.
[25] L. Ladicky, P. H. S. Torr, and A. Zisserman. Human pose estimation using a joint pixel-wise and part-wise formulation.
In Computer Vision and Pattern Recognition (CVPR), 2013.
[26] I. Lifshitz, E. Fetaya, and S. Ullman. Human Pose Estimation Using Deep Consensus Voting, pages 246–260. Springer
International Publishing, Cham, 2016.
[27] A. Newell, K. Yang, and J. Deng. Stacked Hourglass Networks for Human Pose Estimation. European Conference on
Computer Vision (ECCV), pages 483–499, 2016.
[28] W. Ouyang, X. Chu, and X. Wang. Multi-source deep learning for human pose estimation. In Computer Vision and Pattern Recognition (CVPR), pages 2337–2344, June 2014.
[29] T. Pfister, J. Charles, and A. Zisserman. Flowing convnets
for human pose estimation in videos. In International Conference on Computer Vision (ICCV), 2015.
[30] T. Pfister, K. Simonyan, J. Charles, and A. Zisserman. Deep
convolutional neural networks for efficient pose estimation
in gesture videos. In Asian Conference on Computer Vision
(ACCV), 2014.
Acknowledgment
This work was partially supported by the Brazilian National Council for Scientific and Technological Development (CNPq) – Grant 233342/2014-1.
References
[1] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2D
Human Pose Estimation: New Benchmark and State of the
Art Analysis. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), June 2014.
[2] M. Andriluka, S. Roth, and B. Schiele. Pictorial structures
revisited: People detection and articulated pose estimation.
In Computer Vision and Pattern Recognition (CVPR), pages
1014–1021, June 2009.
[3] V. Belagiannis, C. Rupprecht, G. Carneiro, and N. Navab.
Robust optimization for deep regression. In International
Conference on Computer Vision (ICCV), pages 2830–2838,
Dec 2015.
[4] V. Belagiannis and A. Zisserman. Recurrent human pose
estimation. CoRR, abs/1605.02914, 2016.
[5] A. Bulat and G. Tzimiropoulos. Human pose estimation via
Convolutional Part Heatmap Regression. In European Conference on Computer Vision (ECCV), pages 717–732, 2016.
[6] J. Carreira, P. Agrawal, K. Fragkiadaki, and J. Malik. Human pose estimation with iterative error feedback. In 2016
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733–4742, June 2016.
[7] X. Chen and A. Yuille. Articulated pose estimation by a
graphical model with image dependent pairwise relations. In
Advances in Neural Information Processing Systems (NIPS),
2014.
[8] Y. Chen, C. Shen, X. Wei, L. Liu, and J. Yang. Adversarial
posenet: A structure-aware convolutional network for human
pose estimation. CoRR, abs/1705.00389, 2017.
[9] F. Chollet. Xception: Deep learning with depthwise separable convolutions. CoRR, abs/1610.02357, 2016.
[10] F. Chollet et al. Keras. https://github.com/
fchollet/keras, 2015.
[11] C. Chou, J. Chien, and H. Chen. Self adversarial training for
human pose estimation. CoRR, abs/1707.02439, 2017.
[12] X. Chu, W. Ouyang, H. Li, and X. Wang. Structured feature
learning for pose estimation. In Computer Vision and Pattern
Recognition (CVPR), 2016.
[13] X. Chu, W. Yang, W. Ouyang, C. Ma, A. L. Yuille, and
X. Wang. Multi-context attention for human pose estimation. arXiv preprint arXiv:1702.07432, 2017.
9
[46] X. Yu, F. Zhou, and M. Chandraker. Deep Deformation
Network for Object Landmark Localization, pages 52–70.
Springer International Publishing, Cham, 2016.
[31] L. Pishchulin, M. Andriluka, P. Gehler, and B. Schiele. Poselet Conditioned Pictorial Structures. In Computer Vision and
Pattern Recognition (CVPR), pages 588–595, June 2013.
[32] L. Pishchulin, M. Andriluka, P. V. Gehler, and B. Schiele.
Strong appearance and expressive spatial models for human
pose estimation. In International Conference on Computer
Vision (ICCV), pages 3487–3494, 2013.
[33] L. Pishchulin, E. Insafutdinov, S. Tang, B. Andres, M. Andriluka, P. Gehler, and B. Schiele. DeepCut: Joint Subset
Partition and Labeling for Multi Person Pose Estimation. In
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.
[34] U. Rafi, I. Kostrikov, J. Gall, and B. Leibe. An efficient
convolutional network for human pose estimation. In BMVC,
volume 1, page 2, 2016.
[35] V. Ramakrishna, D. Munoz, M. Hebert, A. J. Bagnell, and
Y. Sheikh. Pose Machines: Articulated Pose Estimation via
Inference Machines. In European Conference on Computer
Vision (ECCV), 2014.
[36] G. Rogez, P. Weinzaepfel, and C. Schmid. LCR-Net:
Localization-Classification-Regression for Human Pose. In
Conference on Computer Vision and Pattern Recognition
(CVPR), June 2017.
[37] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. In International
Conference on Learning Representations, 2015.
[38] X. Sun, J. Shang, S. Liang, and Y. Wei. Compositional
human pose regression. arXiv preprint arXiv:1702.07432,
2017.
[39] C. Szegedy, S. Ioffe, and V. Vanhoucke. Inception-v4,
inception-resnet and the impact of residual connections on
learning. CoRR, abs/1602.07261, 2016.
[40] J. Tompson, R. Goroshin, A. Jain, Y. LeCun, and C. Bregler.
Efficient object localization using Convolutional Networks.
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 648–656, June 2015.
[41] J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler. Joint
training of a convolutional network and a graphical model
for human pose estimation. In Z. Ghahramani, M. Welling,
C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27,
pages 1799–1807. Curran Associates, Inc., 2014.
[42] A. Toshev and C. Szegedy. DeepPose: Human Pose Estimation via Deep Neural Networks. In Computer Vision and
Pattern Recognition (CVPR), pages 1653–1660, 2014.
[43] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. In IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2016.
[44] W. Yang, W. Ouyang, H. Li, and X. Wang. End-to-end learning of deformable mixture of parts and deep convolutional
neural networks for human pose estimation. In Computer
Vision and Pattern Recognition (CVPR), 2016.
[45] Y. Yang, S. Baker, A. Kannan, and D. Ramanan. Recognizing proxemics in personal photos. In 2012 IEEE Conference
on Computer Vision and Pattern Recognition, pages 3522–
3529, June 2012.
10
| 1cs.CV
|
1
Stability, convergence and Hopf bifurcation
analyses of the classical car-following model
Gopal Krishna Kamath∗ , Krishna Jagannathan and Gaurav Raina
Department of Electrical Engineering, Indian Institute of Technology Madras, Chennai 600 036, India
arXiv:1607.08779v5 [cs.SY] 29 Mar 2018
Email: {ee12d033, krishnaj, gaurav}@ee.iitm.ac.in
Abstract
Reaction delays play an important role in determining the qualitative dynamical properties of a platoon of vehicles
traversing a straight road. In this paper, we investigate the impact of delayed feedback on the dynamics of the
Classical Car-Following Model (CCFM). Specifically, we analyze the CCFM in no delay, small delay and arbitrary
delay regimes. First, we derive a sufficient condition for local stability of the CCFM in no-delay and small-delay
regimes using. Next, we derive the necessary and sufficient condition for local stability of the CCFM for an arbitrary
delay. We then demonstrate that the transition of traffic flow from the locally stable to the unstable regime occurs
via a Hopf bifurcation, thus resulting in limit cycles in system dynamics. Physically, these limit cycles manifest as
back-propagating congestion waves on highways.
In the context of human-driven vehicles, our work provides phenomenological insight into the impact of reaction
delays on the emergence and evolution of traffic congestion. In the context of self-driven vehicles, our work has
the potential to provide design guidelines for control algorithms running in self-driven cars to avoid undesirable
phenomena. Specifically, designing control algorithms that avoid jerky vehicular movements is essential. Hence, we
derive the necessary and sufficient condition for non-oscillatory convergence of the CCFM. Next, we characterize the
rate of convergence of the CCFM, and bring forth the interplay between local stability, non-oscillatory convergence
and the rate of convergence of the CCFM.
Further, to better understand the oscillations in the system dynamics, we characterize the type of the Hopf
bifurcation and the asymptotic orbital stability of the limit cycles using Poincaré normal forms and the center manifold
theory. The analysis is complemented with stability charts, bifurcation diagrams and MATLAB simulations.
Index Terms
Transportation networks, car-following models, time delays, stability, convergence, Hopf bifurcation.
I. I NTRODUCTION
Intelligent transportation systems constitute a substantial theme of discussion on futuristic smart cities. A prospective solution to increase resource utilization is to use self-driven vehicles, which may also mitigate traffic congestion [30, Section 5.2], [36]. To that end, it is imperative to design stable control algorithms for these vehicles. Since
∗
Corresponding author
This is an extension of our preliminary work that appeared in Proceedings of the 53rd Annual Allerton Conference on Communication,
Control and Computing, pp. 538-545, 2015. DOI: 10.1109/ALLERTON.2015.7447051
March 30, 2018
DRAFT
2
a good design process requires an in-depth understanding of vehicular dynamics, a class of dynamical models –
known as car-following models – have been developed and studied [3]–[5], [8], [10], [25], [40].
An important consideration in the study of car-following models is the delay in the dynamical variables. Delays
arise due to various factors such as sensing, mechanical motions, communication and signal processing. These
delays are known to have a variety of effects on the properties of a dynamical system [17]. Specifically, delays can
readily lead to oscillations and instability [31], [40].
In this paper, we investigate the impact of delayed feedback on the qualitative dynamical properties of a platoon
of vehicles driving on a straight road. Specifically, we focus on analyzing the effect of delayed feedback on the
Classical Car-Following Model (CCFM). In the specific context of human-driven vehicles, the dominant sources
of delay are the physiological delay and the mechanical delay [31]. In contrast, self-driven vehicles tend to have
smaller reaction delays than their human-driven counterparts, and typically occur due to the delays in sensing,
computation and actuation [1]. Hence, we analyze local stability of the CCFM in three regimes – no delay, small
delay and arbitrary delay.
In addition to stability, non-oscillatory convergence and rate of convergence constitute two properties of practical
interest, which we also explore for the case of the CCFM. Such conditions could aid in ensuring smooth traffic flow
by avoiding jerky vehicular motion, thereby improving ride quality. The theoretical analyses could offer suggestions
for design guidelines.
In the context of human-driven vehicles, our investigation into the impact of reaction delay enhances phenomenological insights into the emergence and evolution of traffic congestion. For example, a peculiar phenomenon known
as a ‘phantom jam’ – the emergence of a back-propagating congestion wave in motorway traffic, seemingly out
of nowhere – has been observed in the real world [4], [5]. Previous studies [4], [5] have shown that a change in
driver’s sensitivity (for instance, a sudden deceleration) can lead to such oscillatory behaviour. In this paper, we
show that similar oscillations could also result from an increase in the driver’s reaction delay. More generally, our
study leads to an important observation that the transition of traffic flow from stability to instability could take
place due to a variation in many combinations of model parameters. In order to capture this complex dependence
on various parameters, we introduce an exogenous, non-dimensional parameter in our dynamical model, set to unity
on the stability boundary. We then analyze the system behavior as this exogenous parameter pushes the system
across the stability boundary, and show that limit cycles emerge due to a Hopf bifurcation.
The impact of the reaction delay is perhaps even more important in the context of self-driven vehicles. Self-driven
vehicles are envisioned to have reduced reaction delays as compared to a human driver. As a result, self-driven
vehicles facilitate smaller equilibrium separation between consecutive vehicles [30, Section 5.2]. This, in turn,
improves resource utilization without compromising safety [36]. In contrast to the case of human-driven vehicles,
the parameters in the control algorithm – known as upper longitudinal control algorithm [30, Section 5.2] – for selfdriven vehicles need to be tuned appropriately. To that end, our analyses and findings highlight the quantitative impact
of delayed feedback on the design of control algorithms for self-driven vehicles. In particular, the combination of
stability and convergence analyses may help in the design of various aspects of longitudinal control algorithms [30,
Section 5.2]. We complement our theoretical analyses using stability charts, bifurcation diagrams and MATLAB
March 30, 2018
DRAFT
3
simulations.
A. Related work on car-following models
The work by Chandler et al. [26] as well as the one by Herman et al. [28] constitute two of the earliest known
investigations on stability of car-following models. The CCFM was proposed in [3], although the main objective
therein was to understand the resulting macroscopic behavior. Several related models, and their modifications,
have been investigated in [6], [26] and [28]. For a recent exposition of linear stability analysis as applied to carfollowing models, see [27]. The aforementioned investigations mainly use transform techniques to derive conditions
for stability.
In contrast, [40] and some of the references therein consider the issue of stability from a dynamical systems
perspective. Specifically, [40] studies some stability properties of the CCFM. However, the aforementioned works
do not consider the delay in the the self-velocity term. To make the model more realistic, we accounted for this
delay in our previous work [8]. Therein, we studied a particular case of the CCFM called the Reduced Classical
Car-Following Model (RCCFM), and showed that it loses local stability via a Hopf bifurcation. This paper extends
the results presented in [8] to the CCFM, and also derives conditions that may ensure good ride quality, in addition
to characterizing the time taken by a platoon to reach its equilibrium. To the best of our knowledge, ours is the
first work to characterize such a metric. Further, we show that oscillations in state variables are a manifestation of
limit cycles, and not centers as asserted in [40]. For a recent review on stability analyses as applied to car-following
models, see [18]. For an exposition on the use of time-delayed equations for traffic-flow modeling, see [32].
Note that several dynamical models have also been studied in the Physics literature beginning with the Optimal
Velocity Model (OVM) [25]. In fact, it is known that some of these models lose local stability via a Hopf bifurcation
as well [9], [10], [14]. The OVM has also been studied as a Fillipov system by interpreting negative inter-vehicular
distance as an overtaking maneuver [22]. Further, macroscopic traffic jams resulting due to the OVM have been
studied using the Korteweg-de Vries equation; see [23] and references therein for details. However, this body of
literature assumes the vehicles to be traveling on a single-lane circular loop, thus mathematically yielding periodic
boundary conditions. In contrast, the CCFM and related models differ at a fundamental level by assuming the
vehicular motion on a single-lane straight road. Thus, we do not attempt to compare our results with those derived
for the OVM and related models.
From a vehicular dynamics perspective, most upper longitudinal controllers in the literature assume the lower
controller’s dynamics to be well-modeled by a first-order control system, in order to capture the delay lag [30, Section
5.3]. The upper longitudinal controllers are then designed to maintain either constant velocity, spacing or time gap;
for details, see [29] and the references therein. Specifically, Rajamani et al. [29] prove that synchronization with
the lead vehicle is possible by using information only from the vehicle directly ahead. This reduces implementation
complexity, and does not mandate vehicles to be installed with communication devices.
However, in the context of autonomous vehicles, communication systems are required to exchange various system
states required for the control algorithm. This information is used either for distributed control [29] or coordinated
March 30, 2018
DRAFT
4
control [41]. Formation and platoon stabilities have also been studied considering information flow among the
vehicles [33], [37]. For an extensive review, see [19].
In contrast to stabilizing platoons of autonomous vehicles (our scenario), it has been shown that well-placed,
communicating autonomous vehicles may be used to stabilize platoons of human-driven vehicles as well [11]. More
generally, the platooning problem has been studied as a consensus problem with delays [24]. Such an approach
aids the design of coupling protocols between interacting agents (in this context, vehicles). In contrast, we provide
design guidelines to appropriately choose protocol parameters, given a coupling protocol (the CCFM).
B. Our contributions
Our contributions can be summarized as follows.
(1) We make the CCFM more realistic by accounting for the delay in the self-velocity term.
(2) We show that, in the absence of reaction delays, the CCFM is locally stable for all parameter values of practical
interest. When the delays are rather small, we derive a sufficient condition for local stability of the CCFM
using a linearization of the time variable.
(3) We derive the necessary and sufficient condition for the local stability of the CCFM for an arbitrary delay.
We then show that, upon violation of this condition, the CCFM loses local stability via a Hopf bifurcation.
Indeed, this helps us understand that the oscillations emerge as a consequence of limit cycles, and centers as
asserted in the literature.
(4) In the case of human-driven vehicles, our work enhances phenomenological insights into the emergence and
evolution of traffic congestion. For example, the notion of Hopf bifurcation provides a mathematical framework
to offer a possible explanation for the observed ‘phantom jams.’
(5) We derive the necessary and sufficient condition for non-oscillatory convergence of the CCFM. This is useful in
the context of a transportation network since oscillations lead to jerky vehicular movements, thereby degrading
ride quality and possibly causing collisions.
(6) We characterize the rate of convergence of the CCFM, thereby gaining insight into the time required for
the platoon to attain the desired equilibrium, when perturbed. Such perturbations occur, for instance, when a
vehicle departs from a platoon.
(7) We highlight the three-way trade-off between local stability, non-oscillatory convergence and the rate of
convergence. Considering this trade-off, we suggest some guidelines to appropriately choose parameters for
the upper longitudinal control algorithm in self-driven vehicles.
(8) We characterize the type of Hopf bifurcation and the asymptotic orbital stability of the emergent limit cycles
using Poincaré normal forms and the center manifold theory.
(9) We corroborate the analytical results with the aid of stability charts, numerical computations and simulations
conducted using MATLAB.
The remainder of this paper is organised as follows. In Section II, we introduce the CCFM. In Sections III, IV
and V, we characterize the stable region for the CCFM in no-delay, small-delay and arbitrary-delay regimes
respectively. We understand the stable region by characterizing the region of non-oscillatory convergence of the
March 30, 2018
DRAFT
5
CCFM in Section VI, and the rate of convergence of the CCFM in Section VII. In Section VIII, we present the
local Hopf bifurcation analysis for the CCFM. In Section IX, we present the simulation results before concluding
in Section X.
II. M ODELS
We begin this section with an overview of the setting of our work. We then briefly explain the CCFM.
A. The setting
We study a platoon of N + 1 ‘ideal’ (i.e., zero length) vehicles traversing an infinitely long, single-lane road
without overtaking. The lead vehicle is indexed 0, its follower 1, and so forth. Each vehicle updates its acceleration
based on a combination of its position, velocity and acceleration and those of the vehicle directly ahead. Let
xi (t), ẋi (t) and ẍi (t) denote the position, velocity and acceleration of the ith vehicle respectively, at time t. The
acceleration and velocity profiles of the lead vehicle are assumed to be known. In particular, we restrict ourselves
to leader profiles that converge, in finite time, to ẍ0 = 0 and 0 < ẋ0 < ∞; that is, there exists a finite T0 such
that ẍ0 (t) = 0, ẋ0 (t) = ẋ0 > 0, ∀ t ≥ T0 . We use the terms “driver” and “vehicle” interchangeably throughout.
Further, we use SI units throughout.
B. The Classical Car-Following Model (CCFM)
A key feature of the CCFM is that the acceleration of each vehicle is dependent on three quantities: (i) its own
velocity, (ii) velocity relative to the vehicle directly ahead, and (iii) distance to the vehicle directly ahead. The
exact dependence has been modeled in the literature as [3]
m
ẍi (t) = αi
(ẋi (t)) (ẋi−1 (t − τ ) − ẋi (t − τ ))
(xi−1 (t − τ ) − xi (t − τ ))
l
,
(1)
for i ∈ {1, 2, · · · , N }. Here, αi > 0 represents the ith driver’s sensitivity coefficient, for each i ∈ {1, 2, · · · , N }.
Also, m ∈ [−2, 2] and l ∈ R+ are model parameters that contribute to the non-linearity. Note that the reaction delay
m
is neglected in the self-velocity term (ẋi (t)) . While self velocity might be available almost immediately, it takes
some non-negligible time to execute the required control action. Also, from an analytical viewpoint, ignoring delays
(in general) may generate inaccurate results. Thus, we account for the delay in the self-velocity term. Further, to
make the model more realistic, we assume heterogeneity in reaction delays of different vehicles.
It is apparent from (1) that the state variable xi (t) becomes unbounded as t → ∞ for each i. Therefore, similar
to [40], we transform the model in (1) using yi (t) + bi = xi−1 (t) − xi (t) and vi (t) = ẏi (t) = ẋi−1 (t) − ẋi (t) for
i ∈ {1, 2, · · · , N }. Here, bi denotes the desired equilibrium separation for the ith pair, yi (t) + bi represents the
separation between vehicles i − 1 and i at time t, and vi (t) corresponds to the relative velocity of the ith vehicle
with respect to the (i − 1)th vehicle at time t. The transformed model is thus obtained as
v̇i (t) = βi−1 (t − τi−1 )vi−1 (t − τi−1 ) − βi (t − τi )vi (t − τi ),
ẏi (t) = vi (t),
March 30, 2018
(2)
DRAFT
6
for i ∈ {1, 2, · · · , N }. Here,
βi (t) = αi
(ẋ0 (t) − v0 (t) − · · · − vi (t))
l
(yi (t) + bi )
m
.
Note that y0 , v0 , α0 and τ0 are dummy variables introduced for notational brevity, all of which are set to zero. We
emphasize that y0 and v0 are not state variables.
Note that yi (t) + bi , and not yi (t), represents the headway at time t. In fact, yi (t) represents the variation of
the headway about its equilibrium bi . Thus, yi (t) may become negative. However, the model breaks down when
yi (t) + bi becomes zero for l > 0 [40]. Also, the CCFM possesses an inherent “repulsion” property, which may
be illustrated as follows. Suppose that the vehicle indexed i approaches the vehicle indexed i − 1 at a relatively
higher velocity. When the distance becomes very small (mathematically, < 1 meter), the ith vehicle decelerates
rather rapidly. This can be inferred from (1). This helps avoid collision (hence the term “repulsion”), thus ensuring
yi (t) + bi > 0.
Since equations of the form (2) are hard to analyze, we obtain sufficient conditions for their stability by analyzing
them in the neighborhood of their equilibria. To that end, note that vi∗ = 0, yi∗ = 0 i = 1, 2, · · · , N is an equilibrium
for system (2). Linearizing (2) about this equilibrium, we obtain
∗
v̇i (t) = βi−1
vi−1 (t − τi−1 ) − βi∗ vi (t − τi ),
(3)
ẏi (t) = vi (t),
for i ∈ {1, 2, · · · , N }. Here, βi∗ = αi (ẋ0 )m /(bi )l denotes the equilibrium coefficient for the ith vehicle.
Notice from (3) that the evolution of vi (t), in the vicinity of its equilibrium, is not affected by the evolution
of yi (t). Further, yi (t) can be obtained by integrating vi (t). Thus, we drop the variables {yi (t)}N
i=1 when dealing
with the linearized system. This yields
∗
v̇i (t) = βi−1
vi−1 (t − τi−1 ) − βi∗ vi (t − τi ).
(4)
In the remainder of this paper, we study system (4) to deduce various conditions for the CCFM. It may be noted
that (4) is similar in form to the linearized RCCFM [8, Equation (3)]. However, the equilibrium coefficient βi∗ now
accounts for the non-linearity parameter l ∈ R+ .
III. T HE NO - DELAY REGIME
In this section, we consider the idealistic case of drivers that can react instantaneously to stimuli. This results in
zero reactions delays, and hence the linear model described by system (4) boils down to the following system of
Ordinary Differential Equations (ODEs):
∗
v̇i (t) = βi−1
vi−1 (t) − βi∗ vi (t),
(5)
for i ∈ {1, 2, · · · , N }. This can be succinctly written in matrix form as follows:
V̇(t) = AV(t),
March 30, 2018
(6)
DRAFT
7
where V(t) = [v1 (t) v2 (t) · · · vN (t)]T ∈ RN , and A ∈ RN ×N . The matrix A, known as the dynamics matrix [21,
Section 2.2], is a lower-triangular matrix, given by:
−βi∗ , i = j,
Aij = βj∗ ,
i = j + 1,
0,
elsewhere.
To characterize the stability of system (5), we require the eigenvalues of the dynamics matrix corresponding to
system (6) to be negative [15, Theorem 5.1.1]. Since A is a lower-triangular matrix, the characteristic polynomial
is given by the product of the diagonal elements of the matrix (λI − A) [16, Lemma 6.9.1]. Therefore, we have
f (λ) = det(λI − A) =
N
Y
(λ + βi∗ ) = 0.
(7)
i=1
Therefore, eigenvalues corresponding to system (5) are located at −βi∗ , i ∈ {1, 2, · · · , N }. Note that, from physical
constraints, αi > 0 and bi > 0 ∀i. This ensures βi∗ > 0 ∀i, for all physically relevant systems. Hence, the
corresponding eigenvalues will lie in the open left-half of the Argand plane, thereby ensuring the stability of
system (4) for all physically relevant values of the parameters.
IV. T HE SMALL - DELAY REGIME
In this section, we analyze system (2) in the small-delay regime. A way to obtain insights for small delays is
to conduct a linearization on time. Thus, we obtain a system of ODEs, which serves as an approximation to the
original infinite-dimensional system (4), for small delays. We derive the criterion for this system of ODEs to be
stable, thereby emphasizing the design trade-off inherent among various system parameters and the reaction delay.
We begin by applying the Taylor series approximation to the time-delayed state variables thus: vi (t − τi ) ≈
vi (t) − τi v̇i (t). Using this approximation for terms in (4), and re-arranging the resulting equations, we obtain
v̇i (t) +
∗
∗
βi−1
βi−1
τi−1
βi∗
v̇i−1 (t) =
vi−1 (t) −
vi (t),
∗
∗
1 − βi τi
1 − βi τi
1 − βi∗ τi
(8)
for i ∈ {1, 2, · · · , N }. This can be succinctly written in matrix form as
B V̇(t) = As V(t),
where V(t) = [v1 (t) v2 (t) · · · vN (t)]T ∈ RN . The matrix As
βi∗
− 1−β
∗ ∗,
i τi
∗
βj
Asij =
1−βi∗ τi∗ ,
0,
(9)
is as defined
i = j,
i = j + 1,
elsewhere,
and B is given by
Bij =
March 30, 2018
1,
i = j,
−β ∗ τj
j
1−βi∗ τi , i = j + 1,
0,
elsewhere.
DRAFT
8
Note that B is a lower-triangular matrix with unit diagonal entries. Hence, it is invertible, and the inverse is also
a lower-triangular matrix having unit diagonal elements. Further, since A is a lower-triangular matrix as well, the
dynamics matrix corresponding to system (8), i.e., Ã = B −1 As , is a lower-triangular matrix since it is the product
of two lower-triangular matrices [13, Section 1.4]. Further, due to the said structures, the diagonal elements of Ã
are given by
Ãii =
βi∗
, i ≥ 1.
1 − βi∗ τi
(10)
Therefore, the characteristic polynomial corresponding to system (8) is the product of the diagonal entries of the
matrix (λI − Ã) [16, Lemma 6.9.1]. That is,
f (λ) = det (λI − Ã) =
N
Y
λ + Ãii = 0.
(11)
i=1
This shows that the eigenvalues of system (8) are located at −Ãii , i ∈ {1, 2, · · · , N }. Hence, for system (8) to be
stable, the diagonal entries of its dynamics matrix à have to be positive. From (10), this is satisfied if and only if
βi∗ τi < 1, i ∈ {1, 2, · · · , N }.
(12)
Hence, the above equation represents the necessary and sufficient condition for stability of the time-linearized
system (8). Further, as noted in Section II, (12) is a sufficient condition for the local stability of the CCFM,
described by system (2).
V. T HE H OPF BIFURCATION
Having studied system (2) in the no-delay and the small-delay regimes, in this section, we focus on the arbitrarydelay regime. We derive the necessary and sufficient condition for the local stability of system (2), and show that
the corresponding traffic flow transits from the locally stable to the unstable regime via a Hopf bifurcation [2].
A. Transversality condition
Hopf bifurcation is a phenomenon wherein a system undergoes a stability switch due to a pair of conjugate
eigenvalues crossing the imaginary axis in the Argand plane [17, Chapter 11, Theorem 1.1]. Mathematically, a
Hopf bifurcation analysis is a rigorous way of proving the emergence of limit cycles in non-linear dynamical
systems.
In order to ascertain whether the CCFM undergoes a stability loss via a Hopf bifurcation, we follow [12]
and introduce an exogenous, non-dimensional parameter κ > 0. A general system of delay differential equations
ẋ(t) = f (x(t), x(t−τ1 ), · · · , x(t−τn )) is modified to ẋ(t) = κf (x(t), x(t−τ1 ), · · · , x(t−τn )) with the introduction
of the exogenous parameter. In the specific case of the CCFM, introducing κ in (2) results in
v̇i (t) = κβi−1 (t − τi−1 )vi−1 (t − τi−1 ) − κβi (t − τi )vi (t − τi ),
ẏi (t) = κvi (t),
March 30, 2018
(13)
DRAFT
9
for each i ∈ {1, 2, · · · , N }. We linearize this about the all-zero equilibrium, and drop yi ’s, to obtain
∗
v̇i (t) = κβi−1
vi−1 (t − τi−1 ) − κβi∗ vi (t − τi ),
(14)
for i ∈ {1, 2, · · · , N }. The characteristic equation pertaining to (14) is [8, Equation (15)]
λ + κβi∗ e−λτi = 0.
(15)
It is well known that for (14) to be stable, all the roots of (15) must lie in the open-left half of the Argand
plane [15, Theorem 5.1.1]. Hence, to analyze the local stability of system (13), we search for a conjugate pair of
eigenvalues of (15) that crosses the imaginary axis in the Argand plane, thereby pushing the system into an unstable
√
regime. To that end, we substitute λ = jω, with j = −1, in (15) to obtain
κβi∗ cos(ωτi ) = 0, and ω − κβi∗ sin(ωτi ) = 0.
The first equality implies ωτi = (2n + 1) π2 for n = 0, 1, 2, · · · . Using this, the second equality then results in κβi∗
= ω for n = 0, 2, 4, · · · . Therefore, when a conjugate pair of eigenvalues is on the imaginary axis in the Argand
plane, we have
π
, n = 0, 1, 2, · · · ,
2τi
π
= (2n + 1) ∗ , n = 0, 2, 4, · · · ,
2βi τi
ω0 = (2n + 1)
κcr
(16)
(17)
where κcr is the critical value of κ at ω = ω0 .
To show that system (13) undergoes a Hopf bifurcation at κ = κcr for each n ∈ {0, 2, 4, · · · }, we need to prove
the transversality condition of the Hopf spectrum. That is, we must show that [17, Chapter 11, Theorem 1.1]
dλ
6= 0
(18)
Re
dκ κ=κcr
holds for each n ∈ {0, 2, 4, · · · }. Therefore, we differentiate (15) with respect to κ. Algebraic manipulations then
yield
dλ
Re
dκ
=
κ=κcr
2βi∗ τi2 ω02
> 0,
(2n + 1)(1 + τi2 ω02 )π
(19)
for n ∈ {0, 2, 4, · · · }. This implies that system (13) undergoes a Hopf bifurcation at κ = κcr for each n ∈
{0, 2, 4, · · · }. Hence, κ < κcr when n = 0 is the necessary and sufficient condition for system (13) to be locally
stable.
First, we remark that κ = κcr is the equation of the stability boundary, also known as the Hopf boundary. Once
we obtain the expression for κcr , we tune the system parameters such that the non-dimensional parameter is unity
on the stability boundary, i.e., we set 2βi∗ τi = π to make κcr unity. Next, note that the system loses stability when
the very first conjugate pair of eigenvalues, corresponding to n = 0 in (17), crosses the imaginary axis. Further
increase in κ cannot restore system stability – indeed, the derivative in (19) is positive for each n ∈ {0, 2, 4, · · · }.
That is, an increase in κ results in the eigenvalues moving to the right in the Argand plane, making it impossible
to regain stability. Lastly, it is clear from (17) that αi and τi are inversely related on the Hopf boundary, i.e., when
κcr = 1. Hence, we set αi τi = c, a real constant, in order to study the trade-off between the leader’s profile ẋ0 ,
March 30, 2018
DRAFT
4
5
−2
3
2
1
Leader’s velocity, ẋ0
l=2
l=1
l=0
Unstable region
l=0
l=1
l=2
Stable region
Unstable region
0
0.2
0.4
0.6
Stable region
0
Leader’s velocity, ẋ0
0.8
10
−1.5
−1
−0.5
Non-linearity parameter, m
0
0.8
1.2
1.6
Non-linearity parameter, m
(a)
2
(b)
Fig. 1: The local stability regions for the CCFM are depicted, with variation in non-linearity parameter l; a visual
representation of (20) with c = 1 and bi > 1. (a) is for m < 0, whereas (b) is for m > 0. Pictorially, we restrict
m ∈ [0.8, 2] for clarity of visual representation. As l increases, the CCFM becomes resilient to instability since
bi > 1.
and the non-linearity parameters l and m. The resulting necessary and sufficient condition for the local stability of
system (2) is
(ẋ0 )m
π
< .
(bi )l
2c
(20)
Notice that we recover the necessary and sufficient condition for the local stability of the RCCFM [8, Equation
(20)] if (i) the non-linearity parameter l is set to zero, or (ii) the equilibrium headway bi is set to unity. For these
cases, the inference drawn in [8] holds: When m > 0, slow lead vehicles stabilize the system, and for m < 0,
fast lead vehicles are required to ensure system stability. From Fig.s 1a and 1b, notice that the above inference
holds for l > 0 as well. However, note that the non-linearity parameter l affects the resilience of the CCFM to
instability. Specifically, if the equilibrium headway bi > 1, then the locally stable region expands with an increase
in l. However, when bi < 1, the locally stable region shrinks with an increase in l.
B. Discussion
A few comments are in order.
(1) The foregoing analysis serves to clarify that the oscillations in state variables are a manifestation of limit
cycles (isolated closed orbits in phase space) that emerge due to a Hopf bifurcation, and not centers (family of
concentric closed orbits) as asserted in [40]. Further, as pointed out in the Introduction, these emergent limit
cycles physically manifest themselves as a back-propagating congestion wave, known as a ‘phantom jam.’
Therefore, the foregoing analysis offers a possible explanation of a commonly-observed phenomenon.
(2) Note that the non-dimensional parameter κ introduced in Section V-A is not a system parameter; it is an
exogenous mathematical entity to aid the analysis. Its usefulness is at the edge of the stability boundary,
March 30, 2018
DRAFT
11
wherein it is used to push the system into the unstable regime in a controlled manner, as described in the
analysis.
(3) It is well known in the control literature that a suitable variation in gain parameter can destabilize a system [30,
Section 3.7]. Hence, to ensure that the bifurcation phenomenon is not an artifact of the exogenous parameter,
it is required to verify that the transversality condition of the Hopf spectrum is satisfied for at least one system
parameter beforehand. For the case of the CCFM, following the derivation in Section V-A, it is easy to prove
that the CCFM could undergo a Hopf bifurcation due to an appropriate variation in any of αi , τi , ẋ0 , l or m.
(4) Note that the D-partitioning and its “dual” τ -decomposition approaches [38, Section 3.3] are used extensively
in the literature to study local stability properties of delay-differential equations. While the former assumes the
delay to be fixed and independent of other parameters, the latter allows only the delay to be varied. In contrast,
our approach allows stability analysis to be conducted by a continuous variation of any parameter (including the
exogenous parameter). Additionally, the use of an exogenous parameter as the bifurcation parameter captures
any inter-dependence among model parameters, and generally simplifies the resulting algebra. Further, note
that the bifurcation approach helps understand how local stability is lost and also approximates the trajectory
of the CCFM in the vicinity of the equilibrium using non-linear terms (up to third order in most cases) – key
additions in comparison to other widely-used approaches. This helps deduce the stability of the emergent limit
cycles. The said analysis for the CCFM can be found in Section VIII.
(5) Substituting n = 0 in (17), and letting κ = 1 on the stability boundary, the necessary and sufficient condition
for the local stability of system (2) becomes
βi∗ τi <
π
.
2
(21)
Note that when τi = 0, (21) is trivially satisfied. This, in turn, implies that the CCFM is stable for all parameter
values, in the absence of reaction delays as seen in Section III. However, as the delay increases, (21) will be
violated, thus resulting in loss of local stability of the CCFM. This, in turn, validates our claim that delays
play an important role in determining the qualitative behavior of the CCFM.
(6) Note that (21) coincides with the necessary and sufficient condition derived in [40, Section 3.1]. In fact,
the characteristic equation of the form (15) (with κ = 1) arise in several applications including population
dynamics [20], engineering [17], consensus dynamics [34] and vehicular dynamics [40]. In general, such
equations have been analyzed using both time-domain [15], [20] and spectral-domain methods [38], [40].
However, to the best of our knowledge, none of these works apply the method used in this paper. Further,
note that the evolution equations are non-linear time-delay equations. Hence, the analysis goes beyond that of
a linear time-delay system; see Section VIII for details.
VI. N ON - OSCILLATORY CONVERGENCE
In this section, we characterize the region of non-oscillatory convergence. Mathematically, this amounts to ensuring
that the eigenvalues corresponding to system (4) are negative real numbers. Qualitatively, non-oscillatory convergence
March 30, 2018
DRAFT
12
avoids jerky vehicular motion since relative velocities and headways constitute dynamical variables. Such results
could help ensure the smooth flow of traffic, and hence improve the ride quality.
In the above spirit, following [35], we derive the necessary and sufficient condition for non-oscillatory convergence
of the CCFM. The characteristic equation pertaining to system (4), after dropping the subscript ‘i’ for convenience,
is f (λ) = λ + β ∗ e−λτ = 0 [8, Equation (8)]. Substituting λ = −σ − jω and simplifying, we obtain
σ = β ∗ eστ cos(ωτ ), and ω = β ∗ eστ sin(ωτ ).
(22)
These, in turn, yield tan(ωτ ) = ω/σ. To ensure that ω = 0 is the only solution of this equation, the necessary and
sufficient condition is στ ≥ 1. Re-writing (22), we have
sin(ωτ )
= 1.
β ∗ τ eστ
ωτ
In the limit ω → 0, the term within the brackets represents sinc(0) = 1. Moreover, the exponential term is bounded
by e since στ ≥ 1. Hence, the boundary of non-oscillatory convergence is β ∗ τ e = 1, and the corresponding
necessary and sufficient condition for non-oscillatory convergence is
β∗τ ≤
1
.
e
(23)
Notice that the region in the parameter space described by (23) is a strict subset of the region described by (21).
Therefore, from these two equations, we can summarize the conditions for the local stability of the CCFM as
follows.
(1) If β ∗ τ ∈ [0, π/2), the system is locally stable.
(2) Additionally, if β ∗ τ ∈ [0, 1/e], the system converges asymptotically to the equilibrium in a non-oscillatory
fashion.
(3) Contrarily, if β ∗ τ ∈ (1/e, π/2), the state variable oscillates about the equilibrium, converging asymptotically.
Note that, despite differing in the method of derivation, (23) agrees with the condition for non-oscillatory condition
derived in [40, Section 3].
VII. R ATE OF CONVERGENCE
Rate of convergence is an important performance metric that dictates the time a dynamical system takes to attain
the desired equilibrium, when perturbed. In the context of a transportation network, it is related to the time required
to attain the uniform traffic flow, once the traffic flow is perturbed (by events such as the departure of a vehicle
from the platoon). Following [7], we characterize the rate of convergence for the CCFM.
The characteristic equation pertaining to system (4), with the subscript ‘i’ dropped for ease of exposition, is
f (λ) = λ + β ∗ e−λτ = 0 [8, Equation (8)]. In time domain, this corresponds to a system ẋ(t) = −β ∗ x(t − τ ),
March 30, 2018
DRAFT
15
13
l = 0.8
0
Rate of convergence
5
10
l=1
0 τ1
l = 1.2
τ∗
τ2
0.5
Reaction delay, τ
Fig. 2: Variation in the rate of convergence of the CCFM as the reaction delay is increased, for l ∈ {0.8, 1, 1.2}.
where x is an arbitrarily chosen dynamical variable. The rate of convergence of such a system is the reciprocal of
the smallest among σ1 , σ2 and σ3 , where these quantities are obtained by solving the equations [7, Theorem 2]
στ = 1,
στ e−στ = β ∗ τ,
m
m
e− tan(m) = β ∗ τ, m = στ tan(m),
sin(m)
respectively. The rate of convergence is maximum at τ ∗ = 1/(β ∗ e). For τ < τ ∗ , the rate of convergence increases,
whereas it decreases for τ > τ ∗ [7].
We solve the above equations using MATLAB to illustrate the variations in the rate of convergence for the
CCFM, as the reaction delay is varied. To that end, we consider a tagged vehicle following a lead vehicle with
an equilibrium velocity of 10. The tagged vehicle has a sensitivity coefficient of α = 0.7 and tries to maintain an
equilibrium headway of 20. We fix m = 2, and consider l ∈ {0.8, 1, 1.2}.
The rate of convergence for this system is plotted in Fig. 2. It can be seen that the rate of convergence increases
with τ for τ < τ ∗ , and decreases when the reaction delay is varied beyond τ ∗ . Also, note that the condition
for the maximum rate of convergence coincides with the boundary for non-oscillatory convergence of the CCFM,
β ∗ τ ∗ e = 1. Hence, it would be optimal to choose parameters satisfying this equation. The said figure portrays τ ∗
only for the l = 1 case.
However, in practice, system parameters may vary. This will result in a shift of the operating point of the CCFM,
and may result in a trade-off between the rate of convergence and non-oscillatory convergence of the CCFM. Notice
from Fig. 2 that, for a given value of non-linearity parameter l, the rate of convergence is not symmetric about τ ∗ .
In the vicinity of τ ∗ , if the operating point of the CCFM shifts to the left of τ ∗ , the system retains its non-oscillatory
behavior and the rate of convergence reduces drastically. On the other hand, if the operating point of the CCFM
shifts to the right in the vicinity of τ ∗ , the system converges to the equilibrium in an oscillatory fashion, but the
reduction in the rate of convergence is not as drastic. However, if the reaction delay increases considerably beyond
March 30, 2018
DRAFT
14
τ ∗ , then not only does the system exhibit oscillatory convergence, it may also converge to the equilibrium very
slowly. This is portrayed in Fig. 2 using τ1 = τ ∗ /3 and τ2 = 3τ ∗ , for l = 1. Clearly, the rate of convergence at τ2
is much lesser than that at τ1 .
Finally, notice from Fig. 2 that, an increase in l leads to a decrease in the rate of convergence. However, as
discussed in Section V-A, an increase in l makes the system relatively resilient to instability since b1 = 20 > 1.
Thus, there is a three-way trade-off involving the system’s resilience to instability, rate of convergence and nonoscillatory convergence. Note that Fig. 2 brings forth this trade-off for a fixed set of parameters. However, the same
is true for other parameter values as well.
Note that the characteristic equation captures the closed-loop pairwise interaction in the platoon. To characterize
e
th
the time taken by a platoon to reach an equilibrium (denoted by TCCF
M ), we first define the time taken by the i
pair of vehicles in the platoon following the standard control-theoretic notion of “settling time.” That is, by tei (),
we denote the minimum time taken by the time-domain trajectory of the ith pair to enter, and subsequently remain
within, the -band around its equilibrium. For simplicity, we drop the explicit dependence on . Then, the platoon
dynamics is said to converge to the uniform flow when the dynamics of each pair has settled inside the -band of
its respective equilibrium. Therefore, we have
e
TCCF
M =
max
i=1,2,...,N
tei .
(24)
e
Here, given > 0, tei is computed for the pair that has the least rate of convergence. This, in turn, yields TCCF
M.
Note that the convergence of the CCFM is asymptotic, i.e., the system does not (strictly) converge to the equilibrium
in finite time. Hence, we make use of the settling time concept.
VIII. H OPF BIFURCATION ANALYSIS
In the previous sections, we have characterized the stable region for the CCFM, and studied two of its most
important properties; namely, non-oscillatory convergence and the rate of convergence. We have also proved, by
means of the transversality condition of the Hopf spectrum (19), that system (2) loses stability via a Hopf bifurcation.
In this section, we study the CCFM when it is pushed just beyond the stable region. We characterize the type of the
bifurcation and the asymptotic orbital stability of the emergent limit cycles, following closely the style of analysis
presented in [2], by using Poincaré normal forms and the center manifold theory.
We begin by denoting the non-linear part of the RHS of (13) as fi . That is, for i ∈ {1, 2, · · · , N },
fi , κβi−1 (t − τi−1 )vi−1 (t − τi−1 ) − κβi (t − τi )vi (t − τi ).
(25)
Let µ = κ − κcr . Observe that the system undergoes a Hopf bifurcation at µ = 0, where κ = κcr . Henceforth, we
consider µ as the bifurcation parameter. An incremental change in κ from κcr to κcr + µ, where µ > 0, pushes the
system in to its unstable regime. We now provide a concise step-by-step overview of the detailed local bifurcation
analysis, before delving into the technical details.
Step 1: Using Taylor series expansion, we segregate the RHS of (25) into linear and non-linear parts. We then
cast this into the standard form of an Operator Differential Equation (OpDE).
March 30, 2018
DRAFT
15
Step 2: At the critical value of the bifurcation parameter, i.e., at µ = 0, the system has exactly one pair of
purely imaginary eigenvalues with non-zero angular velocity, as given by (16). The linear eigenspace spanned by
the corresponding eigenvectors is called the critical eigenspace. The center manifold theorem [2] guarantees the
existence of a locally invariant 2-dimensional manifold that is a tangent to the critical eigenspace at the equilibrium
of the system.
Step 3: Next, we project the system onto its critical eigenspace as well as its complement, at the critical value
of the bifurcation parameter. This helps describe the dynamics of the system on the center manifold, with the aid
of an ODE in a single complex variable.
Step 4: Finally, using Poincaré normal forms, we evaluate the Lyapunov coefficient and the Floquet exponent,
which characterize the type of the Hopf bifurcation and the asymptotic orbital stability of the emergent limit cycles
respectively.
We begin the analysis by expanding (25) about the all-zero equilibrium using Taylor’s series, to obtain
m ∗ 2
m
2
β v
(−τi−1 ) + βi∗ vi,t
(−τi )
ẋ0 i−1 (i−1),t
ẋ0
i−2
i−1
m ∗ X
m ∗ X
− βi−1
vn,t (−τi−1 )v(i−1),t (−τi−1 ) + βi−1
vn,t (−τi )vi,t (−τi )
ẋ0
ẋ0
n=1
n=1
∗
v̇i (t) = − κβi∗ vi,t (−τi ) + κβi−1
v(i−1),t (−τi−1 ) −
−
+
l
bi−1
∗
βi−1
v(i−1),t (−τi−1 )y(i−1),t (−τi−1 ) +
l ∗
β vi,t (−τi )yi,t (−τi )
bi i
i−1 i−1
m(m − 1) ∗ 3
m(m − 1) ∗ 3
m(m − 1) ∗ X X
vi,t (−τi )vn,t (−τi )vk,t (−τi )
β
v
(−τ
)
−
β
v
(−τ
)
−
β
i−1
i
2(ẋ0 )2 i−1 i−1
2(ẋ0 )2 i i
2(ẋ0 )2 i n=1
k=1
+
i−2 X
i−2
X
m(m − 1) ∗
β
2(ẋ0 )2 i−1 n=1
k=1
v(i−1),t (−τi−1 )vn,t (−τi−1 )vk,t (−τi−1 ) −
i−1
2m(m − 1) ∗ X 2
β
vi,t (−τi )vn,t (−τi )
i
3(ẋ0 )2
n=1
i−2
X
+
i−2
X
2m(m − 1) ∗
lm
∗
∗
β
v
(−τ
)v
(−τ
)
−
β
vi,t (−τi )vn,t (−τi )yi,t (−τi )
i−1
n,t
i−1
i−1
(i−1),t
3(ẋ0 )2
3(bi )(ẋ0 ) i n=1
n=1
+
i−1
X
lm
lm
∗
2
βi−1
v(i−1),t (−τi−1 )vn,t (−τi−1 )y(i−1),t (−τi−1 ) −
βi∗ vi,t
(−τi )yi,t (−τi )
3(bi−1 )(ẋ0 )
3(
ẋ
)(b
)
0
i
n=1
+
lm
β ∗ v2
(−τi−1 )y(i−1),t (−τi−1 )
3(ẋ0 )(bi−1 ) i−1 (i−1),t
(26)
ẏi (t) =κvi (t),
where we use the shorthand vi,t (−τi ) to represent vi (t − τi ).
In the following, we use C k (A; B) to denote the linear space of all functions from A to B which are k times
differentiable, with each derivative being continuous. Also, we use C to denote C 0 , for convenience.
With the concatenated state S(t), note that (2) is of the form:
dS(t)
= Lµ St (θ) + F(St (θ), µ),
dt
(27)
where t > 0, µ ∈ R, and where for τ = max τi > 0,
i
St (θ) = S(t + θ), S : [−τ, 0] −→ R2N , θ ∈ [−τ, 0].
March 30, 2018
DRAFT
16
Here, Lµ : C [−τ, 0]; R2N −→ R2N is a one-parameter family of continuous, bounded linear functionals, whereas
the operator F : C [−τ, 0]; R2N −→ R2N is an aggregation of the non-linear terms. Further, we assume that
F(St , µ) is analytic, and that F and Lµ depend analytically on the bifurcation parameter µ, for small |µ|. The
objective now is to cast (27) in the standard form of an OpDE:
dSt
= A(µ)St + RSt ,
dt
(28)
since the dependence here is on St alone rather than both St and S(t). To that end, we begin by transforming the
linear problem dS(t)/dt = Lµ St (θ). We note that, by the Riesz representation theorem [39, Theorem 6.19], there
exists a 2N × 2N matrix-valued measure η(·, µ) : B C [−τ, 0]; R2N −→ R2N ×2N , wherein each component of
η(·) has bounded variation, and for all φ ∈ C [−τ, 0]; R2N , we have
Lµ φ =
Z0
dη(θ, µ)φ(θ).
(29)
−τ
In particular,
Lµ St =
Z0
dη(θ, µ)S(t + θ).
−τ
Motivated by the linearized system (4), we define
Ã
dη =
κIN ×N
where
0N ×N
0N ×N
dθ,
−κβi∗ δ(θ + τi ), i = j,
(Ã)ij = κβi∗ δ(θ + τi ),
i = j + 1, j ≥ 1,
0,
otherwise.
For instance, when N = 2,
−κβ1∗ δ(θ + τ1 )
0
0 0
κβ1∗ δ(θ + τ1 ) −κβ2∗ δ(θ + τ2 ) 0 0
dθ.
dη =
κ
0
0 0
0
κ
0 0
For φ ∈ C 1 [−τ, 0]; C2N , we define
A(µ)φ(θ) =
dφ(θ) ,
θ ∈ [−τ, 0),
dθ
R0
−τ
dη(s, µ)φ(s) ≡ Lµ ,
(30)
θ = 0,
and
Rφ(θ) =
March 30, 2018
0,
F(φ, µ),
θ ∈ [−τ, 0),
θ = 0.
DRAFT
17
With the above definitions, we observe that dSt /dθ ≡ dSt /dt. Hence, we have successfully cast (27) in the form
of (28). To obtain the required coefficients, it is sufficient to evaluate various expressions for µ = 0, which we
use henceforth. We start by finding the eigenvector of the operator A(0) with eigenvalue λ(0) = jω0 . That is, we
want an 2N × 1 vector (to be denoted by q(θ)) with the property that A(0)q(θ) = jω0 q(θ). We assume the form:
q(θ) = [1 φ1 φ2 · · · φ2N −1 ]T ejω0 θ , and solve the eigenvalue equations. That is, we need to solve
−κβ1∗ e−jω0 τ1
κβ ∗ e−jω0 τ1 − κβ ∗ e−jω0 τ2
1
1
2
..
φ1
.
∗
∗ −jω0 τN
φ2
κβN −1 e−jω0 τN −1 − κβN
e
= jω0
.. ,
.
κΘ
φ2N −2
κφ1 Θ
..
φ2N −1
.
κφN −1 Θ
where Θ = j(e−jω0 τ − 1)/ω0 . This, in turn, necessitates the following assumption: −κβ1∗ e−jω0 τ1 = jω0 φ0 . Then,
for i ∈ {1, 2, · · · N − 1}, and k ∈ {N, N + 1, · · · 2N − 1},
φi =
κβi∗ e−jω0 τi φi−1
κΘφN −k
, and φk =
,
∗
−jω
τ
0
i+1
jω0 + κβi+1 e
jω0
where we set φ0 = 1 for notational brevity.
We define the adjoint operator as follows:
A∗ (0)φ(θ) =
− dφ(θ) ,
dθ
R0 T
dη (s, 0)φ(−s),
θ ∈ (0, τ ],
θ = 0,
−τ
where dη T is the transpose of dη. We note that the domains of A and A∗ are C 1 [−τ, 0]; C2N and C 1 [0, τ ]; C2N
respectively. Therefore, if jω0 is an eigenvalue of A, then −jω0 is an eigenvalue of A∗ . Hence, to find the eigenvector
of A∗ (0) corresponding to −jω0 , we assume the form: p(θ) = B[ψ2N −1 ψ2N −2 ψ2N −3 · · · 1]T ejω0 θ , and solve
A∗ (0)p(θ) = −jω0 p(θ). Simplifying this, we obtain
−κβ1∗ ejω0 τ1 ψ2N −1 + κβ1∗ ejω0 τ1 ψ2N −2 + κψN −1 Θ̃
−κβ ∗ ejω0 τ2 ψ
∗ jω0 τ2
ψ2N −3 + κψN −2 Θ̃
ψ2N −1
2N −2 + κβ2 e
2
..
ψ2N −2
.
∗ jω0 τN
ψ
−κβN e
ψN + κψ0 Θ̃
2N −3
= −jω0
.. .
.
c
c
ψ1
..
1
.
c
March 30, 2018
DRAFT
18
Here, we set ψ0 = 1 for notational brevity and c is some constant which can be shown to be −jω0 /B. Then, for
i ∈ {1, 2, · · · N − 1}, and k ∈ {N + 1, N + 2, · · · 2N − 1}, we have
ψN =
κΘ̃ψ0
∗ ejω0 τN − jω
κβN
0
, ψi = −jω0 , and ψi =
∗
jω0 τ2N −i
κβ2N
+ κΘ̃ψN −i
−i e
.
∗
jω
τ
0
2N
−i
κβ2N −i e
− jω0
The normalization condition for Hopf bifurcation requires that hp, qi = 1, thus yielding an expression for B.
For any q ∈ C [−τ, 0]; C2N and p ∈ C [0, τ ]; C2N , the inner product is defined as
Z0
hp, qi , p̄ · q −
Zθ
p̄T (ζ − θ)dηq(ζ) dζ,
(31)
θ=−τ ζ=0
where the overbar represents the complex conjugate and the “ · ” represents the regular dot product. The value of
B such that the inner product between the eigenvectors of A and A∗ is unity can be shown to be
B=
1
,
ζ1 + ζ2 + ζ3 + ζ4
where
ζ1 = Θ
∗
N
−1
X
κψN φ̄i , ζ2 =
i=0
N
−1
X
i=1
κβi∗ τi ejω0 τi φi−1 (ψ2N −i − ψ2N −i−1 ),
∗
φN −1 ψN τN ejω0 τN , and ζ4 =
ζ3 = −κβN
2N
−1
X
ψ2N −1−i φ̄i .
i=0
In the above, we define φ0 = ψ0 = 0 for notational brevity.
For St , a solution of (28) at µ = 0, we define
z(t) = hp(θ), St i, and w(t, θ) = St (θ) − 2Real(z(t)q(θ)).
Then, on the center manifold C0 , we have w(t, θ) = w(z(t), z̄(t), θ), where
w(z(t), z̄(t), θ) = w20 (θ)
z2
z̄ 2
+ w02 (θ) + w11 (θ)z z̄ + · · · .
2
2
(32)
Effectively, z and z̄ are the local coordinates for C0 in C in the directions of p and p̄ respectively. We note that
w is real if St is real, and we deal only with real solutions. The existence of the center manifold C0 enables the
reduction of (28) to an ODE in a single complex variable on C0 . At µ = 0, the said ODE can be described as
ż(t) = hp, ASt + RSt i ,
= jω0 z(t) + p̄(0).F (w(z, z̄, θ) + 2Real(z(t)q(θ))) ,
= jω0 z(t) + p̄(0).F0 (z, z̄).
(33)
This is written in abbreviated form as
ż(t) = jω0 z(t) + g(z, z̄).
(34)
The objective now is to expand g in powers of z and z̄. However, this requires wij (θ)’s from (32). Once these are
evaluated, the ODE (33) for z would be explicit (as given by (34)), where g can be expanded in terms of z and z̄
as
g(z, z̄) = p̄(0).F0 (z, z̄) = g20
March 30, 2018
z2
z̄ 2
z 2 z̄
+ g02 + g11 z z̄ + g21
+ ··· .
2
2
2
(35)
DRAFT
19
Next, we write ẇ = Ṡt − żq − z̄˙ q̄. Using (28) and (34), we then obtain the following ODE:
Aw − 2Real(p̄(0).F0 q(θ)),
θ ∈ [−τ, 0),
ẇ =
Aw − 2Real(p̄(0).F0 q(0)) + F0 , θ = 0.
This can be re-written using (32) as
ẇ = Aw + H(z, z̄, θ),
(36)
where H can be expanded as
H(z, z̄, θ) = H20 (θ)
z2
z̄ 2
z 2 z̄
+ H02 (θ) + H11 (θ)z z̄ + H21 (θ)
+ ··· .
2
2
2
(37)
˙ Using (32) and (34) to replace wz ż (and their
Near the origin, on the manifold C0 , we have ẇ = wz ż + wz̄ z̄.
conjugates, by their power series expansion) and equating with (36), we obtain the following operator equations:
(2jω0 − A)w20 (θ) = H20 (θ),
(38)
−Aw11 = H11 (θ),
(39)
−(2jω0 + A)w02 (θ) = H02 (θ).
(40)
We start by observing that
St (θ) = w20 (θ)
z̄ 2
z2
+ w02 (θ) + w11 (θ)z z̄ + zq(θ) + z̄ q̄(θ) + · · · .
2
2
From the Hopf bifurcation analysis [2], we know that the coefficients of z 2 , z̄ 2 , z 2 z̄, and z z̄ terms are used to
approximate the system dynamics. Hence, we only retain these terms in the expansions.
To obtain the effect of non-linearities, we substitute the aforementioned terms appropriately in the non-linear
terms of (26) and separate the terms as required. Therefore, for each i ∈ {1, 2, . . . , 2N }, we have the non-linearity
term to be
Fi = F20i
z2
z̄ 2
z 2 z̄
+ F02i + F11i z z̄ + F21i
,
2
2
2
(41)
where, for i ∈ {1, 2, · · · , N }, the coefficients are given by
m
l
m
l
∗
∗
∗
−2jω0 τi−1
+4
F20i =βi−1 w20(i−1) (−τi−1 ) − βi w20i (−τi ) − 4
+
βi−1 e
+
βi∗ e−2jω0 τi ,
ẋ0
bi−1
ẋ0
bi
m
m
∗
−2jω0 τi−1
−4
βi−1 (i − 2)e
+4
βi∗ (i − 1)e−2jω0 τi ,
ẋ0
ẋ0
m
l
m
l
∗
∗
∗
2jω0 τi−1
F02i =βi−1 w02(i−1) (−τi−1 ) − βi w02i (−τi ) − 4
+
βi−1 e
+4
+
βi∗ e2jω0 τi ,
ẋ0
bi−1
ẋ0
bi
m
m
∗
−4
βi−1
(i − 2)e2jω0 τi−1 + 4
βi∗ (i − 1)e2jω0 τi ,
ẋ0
ẋ0
m
l
m
l
∗
∗
F11i =βi−1
w11(i−1) (−τi−1 ) − βi∗ w11i (−τi ) − 2
+
βi−1
+2
+
βi∗ ,
ẋ0
bi−1
ẋ0
bi
m
l
∗
F21i = − 2
βi−1
+
w20(i−1) ejω0 τi−1 + 2w11(i−1) (−τi−1 )e−jω0 τi−1
ẋ0
bi−1
m
l
+2
+
βi∗ w20i ejω0 τi + 2w11i (−τi )e−jω0 τi
ẋ0
bi
March 30, 2018
DRAFT
20
−
+
m
ẋ0
m
ẋ0
∗
βi−1
i−2
X
(w20n (−τi−1 ) + w20(i−1) (−τi−1 ))ejω0 τi−1 + 2(w11n (−τi−1 ) + w11(i−1) (−τi−1 ))e−jω0 τi−1
n=1
βi∗
i−1
X
(w20n (−τi ) + w20i (−τi ))ejω0 τi + 2(w11n (−τi ) + w11i (−τi ))e−jω0 τi
n=1
2m(m − 1)(i − 2)
lm(i − 2)
m(m − 1) m(m − 1)(i − 2)2
lm
+
+
+
+ (2e
+
2(ẋ0 )2
(ẋ0 )2
3(ẋ0 )2
3(bi−1 )(ẋ0 ) 3(bi−1 )(ẋ0 )
2m(m − 1)(i − 1) lm(i − 1)
m(m − 1) m(m − 1)(i − 1)2
lm
−ω0 τi ∗
+
+
+
βi )
+
.
− (2e
2(ẋ0 )2
(ẋ0 )2
3(ẋ0 )2
3(bi )(ẋ0 )
3(bi )(ẋ0 )
−ω0 τi−1
∗
βi−1
)
We represent the vector of non-linearities used in (33) as F0 = [F1 F2 · · · FN ]T . Next, we compute g using
F0 as
g(z, z̄) = p̄(0).F0 = B̄
N
X
l=1
ψ̄N −l Fl .
(42)
Substituting (41) in (42), and comparing with (35), we obtain
gx = B̄
N
X
l=1
ψ̄N −l Fxl ,
(43)
where x ∈ {20, 02, 11, 21}. Using (43), the corresponding coefficients can be computed. However, computing g21
requires w20 (θ) and w11 (θ). Hence, we perform the requisite computation next. For θ ∈ [−τ, 0), H can be simplified
as
H(z, z̄, θ) = −Real (p̄(0).F0 q(θ)) ,
z̄ 2
z2
= − g20 + g02 + g11 z z̄ + · · · q(θ)
2
2
2
z̄
z2
− ḡ20 + ḡ02 + ḡ11 z z̄ + · · · q̄(θ),
2
2
which, when compared with (37), yields
H20 (θ) = −g20 q(θ) − ḡ20 q̄(θ),
(44)
H11 (θ) = −g11 q(θ) − ḡ11 q̄(θ).
(45)
From (30), (38) and (39), we obtain the following ODEs:
ẇ20 (θ) = 2jω0 w20 (θ) + g20 q(θ) + ḡ02 q̄(θ),
(46)
ẇ11 (θ) = g11 q(θ) + ḡ11 q̄(θ).
(47)
Solving (46) and (47), we obtain
ḡ02
g20
q(0)ejω0 θ −
q̄(0)e−jω0 θ + e e2jωθ ,
jω0
3jω0
g11
ḡ11
w11 (θ) =
q(0)ejω0 θ −
q̄(0)e−jω0 θ + f,
jω0
jω0
w20 (θ) = −
(48)
(49)
for some vectors e and f, to be determined.
March 30, 2018
DRAFT
21
To that end, we begin by defining the following vector: F̃20 , [F201 F202 · · · F20N ]T . Equating (38) and (44),
and simplifying, yields the operator equation: 2jω0 e − A e e2jω0 θ = F̃20 . On simplification, we obtain
2jω0 + κβ1∗ e1
2jω0 + κβ2∗ e2 − κβ1∗ e1
..
.
∗
∗
2jω0 + κβN
eN − κβN −1 eN −1 = F̃20 .
−κτ e1 + 2jω0 eN +1
.
..
−κτ eN + 2jω0 e2N
This, in turn, yields
ei =
∗
F20(N +i) + κτ ei
F20i − κβi−1
ei−1
, and, eN +i =
,
2jω0 + κβi∗
2jω0
(50)
for i ∈ {1, 2, · · · , N }. Here, we set e0 = 0 for notational brevity.
Next, equating (39) and (45), and simplifying, we obtain the operator equation Af = −F̃11 , with F̃11 ,
[F111 F112 · · · F11N ]T . To solve this, we make the assumption that, for i ∈ {1, 2, · · · , N − 1},
∗
F20i τ + κβi−1
fi−1 + βi∗ F11(N +i) = 0
Therefore, on simplification, the above-mentioned operator equation yields
−κβ1∗ f1
κβ1∗ f1 − κβ2∗ f2
..
.
∗
∗
κβN −1 fN −1 − κβN fN = −F̃11 .
κτ f1
.
..
κτ fN
On solving this, we obtain for i ∈ {1, 2, · · · , N },
fi =
∗
F11i + κβi−1
fi−1
, and, fN +i = c,
∗
κβi
(51)
where c is an arbitrary constant, which we set to zero for simplicity. We also set f0 = 0 for notational brevity.
Substituting for e and f from (50) and (51) in (48) and (49) respectively, we obtain w20 (θ) and w11 (θ). This, in
turn, facilitates the computation of g21 . We can then compute
1
g21
j
2
2
g20 g11 − 2|g11 | − |g02 | +
,
c1 (0) =
2ω0
3
2
0
dλ
Re[c1 (0)]
α (0) = Re
, µ2 = −
, and β2 = 2Re[c1 (0)].
dκ κ=κcr
α0 (0)
Here, c1 (0) is known as the Lyapunov coefficient and β2 is the Floquet exponent. It is known from [2] that these
quantities are useful since
March 30, 2018
DRAFT
Amplitude (relative velocity)
−0.6
0
0.6
22
l = 0.95
l=1
l = 1.05
κcr = 1
Bifurcation parameter, κ
1.05
Fig. 3: Bifurcation diagram: Variation in the amplitude of relative velocity of the CCFM as the non-dimensional
parameter κ is varied, for l ∈ {0.95, 1, 1.05}.
(i) If µ2 > 0, then the bifurcation is supercritical, whereas if µ2 < 0, then the bifurcation is subcritical.
(ii) If β2 > 0, then the limit cycle is asymptotically orbitally unstable, whereas if β2 < 0, then the limit cycle is
asymptotically orbitally stable.
We now present numerically-constructed bifurcation diagrams to gain some insight into the effect of various
parameters on the amplitude of the limit cycle.
Bifurcation diagrams
We next present bifurcation diagrams, numerically constructed using the scientific computation software MATLAB. We implement a discrete version of system (13) with update time Ts = 0.01 s. We then vary the nondimensional parameter κ in the range [1, 1.05], and record the amplitude of the relative velocity in steady state.
The resulting plot of the envelope of the relative velocity as a function of the non-dimensional parameter is called
a bifurcation diagram.
For illustration, we consider a single vehicle following a lead vehicle whose equilibrium velocity is 10. For the
follower vehicle, we initialize the parameters as follows. α1 = 0.7, b1 = 20 and m = 2. We set the reaction delay
τ1 = τcr ≈ 0.45, to ensure κcr = 1. Next, we vary the non-dimensional parameter in the vicinity of unity, and
record the resulting amplitude of the relative velocity for l = 0.95, 1, 1.05. The resulting bifurcation diagram is
portrayed in Fig. 3.
It can be inferred from Fig. 3 that, there is no monotonicity in the amplitude of relative velocity with an increase
in the non-linearity parameter l. This is unlike the result presented in [8], wherein monotonicity of the amplitude
of relative velocity with an increase in m was shown numerically, for l = 0 (the RCCFM). While Fig. 3 was
constructed for m = 2, extensive computations reveal a lack of monotonicity in the amplitude of limit cycle with
an increase in l.
March 30, 2018
DRAFT
23
−10
v(t), y(t) ( × 10−3)
0
10
v3(t)
y3(t)
60
65
Time (in seconds)
70
5
τ 2 = 3τ
∗
2
4
Time (in seconds)
τ1 = τ ∗/3
6
(a)
τ 2 = 3τ ∗
−5
0
τ1 = τ ∗/3
Headway, y(t) (× 10−3)
0
Relative velocity, v(t) (× 10−3)
−10
0
10
Fig. 4: Simulations: Emergence of limit cycles in the headway and the relative velocity, as predicted by the analysis.
0
2
4
Time (in seconds)
6
(b)
Fig. 5: Simulations: Non-oscillatory and oscillatory solutions of the CCFM. (a) portrays the relative velocity
solutions, whereas (b) shows the headway solutions. These serve to validate our analytical insight.
IX. S IMULATIONS
We now present the simulation results for the CCFM, that serve to corroborate our analytical findings. We make
use of the scientific computation software MATLAB to implement a discrete version of system (2) with update
time Ts = 0.01 s, thus simulating the CCFM.
We initialize the parameters with the following values. N = 4, α1 = 0.5, α2 = 0.6, α3 = 0.7, α4 = 0.8,
τ1 = 0.5, τ2 = 0.4, τ3 = τcr ≈ 0.45 and τ4 = 0.3. The leader’s velocity profile is considered to be 10(1 − e−10t ),
thus ensuring an equilibrium velocity of 10. Further, we fix m = 2, l = 1 and desired headways bi = 20 ∀i. Fig. 4
shows the emergence of limit cycles in the state variables of the third vehicle, as predicted by our analysis. Also
notice the phase shift between the relative velocity and headway solutions, as a consequence of obtaining the latter
by integrating the former.
March 30, 2018
DRAFT
24
Next, we validate that the CCFM does indeed exhibit non-oscillatory convergence to the all-zero equilibrium, as
predicted by (23). We also compare the rate of convergence when the reaction delay satisfies τ < τ ∗ and τ > τ ∗ , as
discussed in Section VII. We make use of the same parameter values as above, except for reaction delays; these are
identically set to τ1 = 1/(3eβ ∗ ) and τ2 = 3/(eβ ∗ ). Fig.s 5a and 5b portray the solutions for relative velocity and
headway respectively. To ensure comparison with Fig. 2, these plots correspond to the third vehicle, i.e., α = 0.7
s−1 and l = 1. Notice that τ1 and τ2 , which are not in the vicinity of τ ∗ corresponding to the third vehicle, as
seen from Fig. 2. Hence, the solutions corresponding to τ1 attain their equilibria much faster than those pertaining
to τ2 .
X. C ONCLUDING R EMARKS
In this paper, we highlighted the importance of delayed feedback in determining the qualitative dynamical
properties of a platoon of vehicles driving on a straight road. Specifically, we analyzed the Classical Car-Following
Model (CCFM) in three regimes – no delay, small delay and arbitrary delay. Control-theoretic analyses helped
us derive conditions for its local stability. In particular, the analysis for small-delay regime yielded a sufficient
condition for the local stability of the CCFM, whereas we obtained the necessary and sufficient condition for the
local stability of the CCFM in the arbitrary-delay regime.
We then proved that the CCFM undergoes a loss of stability via a Hopf bifurcation. Mathematically, this
result proves the emergence of limit cycles, which physically manifests as a back-propagating congestion wave.
Even though the parameters are not strictly controllable in the case of human drivers, our work enhances the
phenomenological insights into ‘phantom jams.’ Our analyses made use of an exogenous, non-dimensional parameter
that served to handle the complex relation which could exist among the various model parameters.
We then derived the necessary and sufficient condition for non-oscillatory convergence of the CCFM. Designing
control algorithms that conform to this condition ensures that jerky vehicular motions are avoided, thus guaranteeing
smooth traffic flow and improving ride quality. Next, we characterized the rate of convergence of the CCFM, and
highlighted the three-way trade-off between local stability, non-oscillatory convergence and the rate of convergence.
Finally, we characterized the type of Hopf bifurcation and the asymptotic orbital stability of the limit cycles,
which emerge when the stability conditions are just violated, using Poincaré normal forms and the center manifold
theory. The analyses were complemented by stability charts, numerically constructed bifurcation diagrams and
MATLAB simulations. These serve to highlight the impact of various model parameters on system stability as well
as the relative velocity amplitude of the emergent limit cycles.
Avenues for further research
There are numerous avenues that merit further investigation. In the context of the CCFM, we have addressed
the issue of pairwise stability of vehicles in this work. However, string stability of a platoon of vehicles running
the CCFM remains to be studied. Also, from a practical standpoint, the parameters of the CCFM may vary, for
varied reasons. Hence, it becomes imperative that the longitudinal control algorithm be robust to such parameter
variations, and to unmodeled vehicular dynamics.
March 30, 2018
DRAFT
25
ACKNOWLEDGEMENTS
This work is undertaken as a part of an Information Technology Research Academy (ITRA), Media Lab Asia,
project titled “De-congesting India’s transportation networks.” The authors are also thankful to Debayani Ghosh
and Sreelakshmi Manjunath for many helpful discussions.
R EFERENCES
[1] A. Kesting and M. Treiber, “How reaction time, update time, and adaptation time influence the stability of traffic flow,” Computer-Aided
Civil and Infrastructure Engineering, vol. 23, pp. 125-137, 2008.
[2] B.D. Hassard, N.D. Kazarinoff and Y.-H. Wan, “Theory and Applications of Hopf Bifurcation.,” Cambridge University Press, 1981.
[3] D.C. Gazis, R. Herman and R.W. Rothery, “Nonlinear follow-the-leader models of traffic flow,” Operations Research, vol. 9, pp. 545-567,
1961.
[4] D. Chowdhury, L. Santen and A. Schadschneider, “Statistical physics of vehicular traffic and some related systems,” Physical Reports,
vol. 329, pp. 199-329, 2000.
[5] D. Helbing, “Traffic and related self-driven many-particle systems,” Reviews of Modern Physics, vol. 73, pp. 1067-1141, 2001.
[6] E.A. Unwin and L. Duckstein, “Stability of reciprocal-spacing type car-following models,” Transportation Science, vol. 1, pp. 95-108,
1967.
[7] F. Brauer, “Decay rates for solutions of a class of differential-difference equations,” SIAM Journal on Mathematical Analysis, vol. 10, pp.
783-788, 1979.
[8] G.K. Kamath, K. Jagannathan and G. Raina, “Car-following models with delayed feedback: local stability and Hopf bifurcation,” in
Proceedings of the 53rd Annual Allerton Conference on Communication, Control and Computing, 2015.
[9] G. Orosz, B. Krauskopf and R.E. Wilson, “Bifurcations and multiple jams in a car-following model with reaction-time delay,” Physica D,
vol. 211, pp. 277-293, 2005.
[10] G. Orosz and G. Stépán, “Subcritical Hopf bifurcations in a car-following model with reaction-time delay,” Proceedings of the Royal
Society A, vol. 642, pp. 2643-2670, 2006.
[11] G. Orosz, “Connected cruise control: modelling, delay effects, and nonlinear behaviour,” Vehicle System Dynamics, vol. 54, pp. 1147-1176,
2016.
[12] G. Raina, “Local bifurcation analysis of some dual congestion control algorithms,” IEEE Transactions on Automatic Control, vol. 50, pp.
1135-1146, 2005.
[13] G. Strang, “Linear Algebra and Its Applications,” Cengage Learning, Fourth Edition, 2006.
[14] I. Gasser, G. Sirito and B. Werner, “Bifurcation analysis of a class of ‘car following’ traffic models”, Physica D, vol. 197, pp. 222-241,
2004.
[15] I. Györi and G. Ladas, “Oscillation Theory of Delay Differential Equations With Applications,” Clarendon Press, 1991.
[16] I.N. Herstein, “Topics in Algebra,” John Wiley & Sons, 1975.
[17] J.K. Hale and S.M.V. Lunel, “Introduction to Functional Differential Equations,” Springer-Verlag, 2011.
[18] J. Sun, Z. Zheng and J. Sun, “Stability analysis methods and their applicability to car-following models in conventional and connected
environments,” Transportation Research Part B, vol. 109, pp. 212-237, 2018.
[19] K.C. Dey, L. Yan, X. Wang, Y. Wang, H. Shen, M. Chowdhury, L. Yu, C. Qiu and V. Soundararaj, “A review of communication, driver
characteristics, and controls aspects of cooperative adaptive cruise control (CACC),” IEEE Transactions on Intelligent Transportation
Systems, vol. 17, pp. 491-509, 2016.
[20] K. Gopalsamy, “Stability and Oscillations in Delay Differential Equations of Population Dynamics,” Kluwer Academic Publishers, 1992.
[21] K.J. Åström and R.M. Murray, “Feedback Systems: An Introduction for Scientists and Engineers,” Princeton University Press, 2008.
[22] L. Bur̆ic̆ and V. Janovský, “On pattern formations in a class of traffic models,” Physica D, vol. 237, pp. 28-49, 2008.
[23] L. Hattam, “KdV cnoidal waves in a traffic flow model with periodic boundaries,” Physica D, vol. 348, pp. 44-53, 2017.
[24] M. di Bernardo, A. Salvi and S. Santini, “Distributed consensus strategy for platooning of vehicles in the presence of time-varying
heterogeneous communication delays,” IEEE Transactions on Intelligent Transportation Systems, vol. 16, pp. 102-112, 2015.
[25] M. Bando, K. Hasebe, K. Nakanishi and A. Nakayama, “Analysis of optimal velocity model with explicit delay,” Physical Review E, vol.
58, pp. 5429-5435, 1998.
March 30, 2018
DRAFT
26
[26] R.E. Chandler, R. Herman and E.W. Montroll, “Traffic dynamics: studies in car following,” Operations Research, vol. 6, pp. 165-184,
1958.
[27] R.E. Wilson and J.A. Ward, “Car-following models: fifty years of linear stability analysis - a mathematical perspective,” Transportation
Planning and Technology, vol. 34, pp. 3-18, 2011.
[28] R. Herman, E.W. Montroll, R.B. Potts and R.W. Rothery, “Traffic dynamics: analysis of stability in car following,” Operations Research,
vol. 7, pp. 86-106, 1959.
[29] R. Rajamani and C. Zhu, “Semi-autonomous adaptive cruise control systems,” IEEE Transactions on Vehicular Technology, vol. 51, pp.
1186-1192, 2002.
[30] R. Rajamani, “Vehicle Dynamics and Control,” Springer, Second Edition, 2012.
[31] R. Sipahi and S.I. Niculescu, “Analytical stability study of a deterministic car following model under multiple delay interactions,” in
Proceedings of Mechanical and Industrial Engineering Faculty Publications, 2006.
[32] R. Sipahi and S.I. Niculescu, “Deterministic Time-Delayed Traffic Flow Models: A Survey,” in Complex Time-Delay System: Theory and
Applications. F.M. Atay: Springer, 2010, pp. 297-322.
[33] R.U. Chavan, M. Belur, D. Chakraborty and D. Manjunath, “On the stability and formations in ad hoc multilane vehicular traffic,” in
Proceedings of the 7th International Conference on Communication Systems and Networks (COMSNETS), 2015.
[34] R. Olfati-Saber and R.M. Murray, “Consensus problems in networks of agents with switching topologies and time-delays,” IEEE
Transactions on Automatic Control, vol. 49, pp. 1520-1533, 2004.
[35] S. Deb and R. Srikant, “Global stability of congestion controllers for the Internet,” IEEE Transactions on Automatic Control, vol. 48, pp.
1055-1060, 2003.
[36] S. Greengard, “Smart transportation networks drive gains,” Communications of the ACM, vol. 58, pp. 25-27, 2015.
[37] T.H. Summers, C. Yu, S. Dasgupta and B.D.O. Anderson, “Control of minimally persistent leader-remote-follower and coleader formations
in the plane,” IEEE Transactions on Automatic Control, vol. 56, pp. 2778-2792 , 2011.
[38] W. Michiels and S-I. Niculescu, “Stability and Stabilization of Time-Delay Systems: An Eigenvalue-Based Approach,” Advances in Design
and Control, SIAM, 2007.
[39] W. Rudin, “Real & Complex Analysis,” Tata McGraw Hill Publications, Third Edition, 1987.
[40] X. Zhang and D.F. Jarrett, “Stability analysis of the classical car-following model,” Transportation Research Part B, vol. 31, pp. 441-462,
1997.
[41] Z. Qu, J. Wang and R.A. Hull, “Cooperative control of dynamical systems with application to autonomous vehicles,” IEEE Transactions
on Automatic Control, vol. 53, pp. 894-911, 2008.
March 30, 2018
DRAFT
| 3cs.SY
|
Online Article Ranking as a Constrained, Dynamic,
Multi-Objective Optimization Problem
Jeya Balaji Balasubramanian
Akshay Soni
Intelligent Systems Program,
University of Pittsburgh, Pittsburgh
[email protected]
Yahoo! Research, Sunnyvale
[email protected]
Nikolay Laptev
arXiv:1705.05765v1 [cs.AI] 16 May 2017
Yashar Mehdad
Airbnb, San Francisco
[email protected]
Abstract
The content ranking problem in a social news website,
is typically a function that maximizes a scalar metric of interest like dwell-time. However, like in most
real-world applications we are interested in more than
one metric—for instance simultaneously maximizing
click-through rate, monetization metrics, dwell-time—
and also satisfy the traffic requirements promised to different publishers. All this needs to be done on online
data and under the settings where the objective function
and the constraints can dynamically change; this could
happen if for instance new publishers are added, some
contracts are adjusted, or if some contracts are over.
In this paper, we formulate this problem as a constrained, dynamic, multi-objective optimization problem. We propose a novel framework that extends a
successful genetic optimization algorithm, NSGA-II, to
solve this online, data-driven problem. We design the
modules of NSGA-II to suit our problem. We evaluate
optimization performance using Hypervolume and introduce a confidence interval metric for assessing the
practicality of a solution. We demonstrate the application of this framework on a real-world Article Ranking
problem. We observe that we make considerable improvements in both time and performance over a bruteforce baseline technique that is currently in production.
1
Introduction
Ranking a list of content based on a query (e.g., search results, or news articles where the query is the user) has been
studied in depth. An example of this is article ranking in a
news website that is typically aimed at maximizing a single
scalar objective like dwell-time or click-through-rate (CTR).
These are known as single-objective optimization (SOO)
problems. However, most real-world content ranking platforms tries to optimize multiple objectives like dwell-time,
CTR, daily active users (DAU), and monetization metrics
simultaneously. Note that some of these objectives may be
conflicting in the sense that if one of them increases, others
may be forced to decrease; for instance, generally if one tries
to increase the dwell-time, then CTR drops. This generates
the need for formulating this as a multi-objective optimization (MOO) problem.
Copyright c 2017, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
University of California, Los Angeles
[email protected]
To make it further complicated, the article ranking problems are generally subject to constraints such as traffic
shaping requirements—these are partnership contracts between content aggregators and third party content providers
to acquire content in return for a target traffic or revenue
promises. The traffic, or revenue commitment, is documented in a contract and then manually tracked and reported
by the partnership lead. This is usually done by pulling the
data from various tools and sending a spreadsheet to the partners. This is a highly laborious process that is not scalable
while also not providing the level of information granularity
needed. One major challenge of many content aggregators is
the control over achieving these targets; the current model is
based on after-the-fact reporting rather than controlled targeting. The gap between the commitment and actual target
achievement is often large and can range from 10% to 50%
under achievement.
Finally, the objectives of interest are generally dynamic in
nature i.e., they may change with time. For example, CTR
is known to show significant temporal variations with timeof-day, day-of-week effect, and CTR decay due to repeated
exposure to the same articles (Agarwal, Chen, and Elango
2009). This requires us to optimize target traffic with respect
to the dynamic behavior of these objectives.
1.1
Problem description
An objective variable, y, is any random variable of interest that we want to optimize. An objective space, Y , is a
vector of all objective variables that we want to simultaneously optimize. We represent the objective space as an
m-dimensional vector, ~y = (y1 , y2 , . . . , ym ) with m different objectives. For example in the news ranking problem,
we may want to simultaneously optimize on total clicks,
and dwell-time. A design variable, x, is any other random
variable that can affect the value of the objective variables.
This is usually a variable we can easily measure and control. The design space, X, is a vector of all design variables that can affect our objectives. We represent the design
space as an n-dimensional vector, ~x = (x1 , x2 , . . . , xn ),
with n candidate design variables. Examples of design variables in ranking can include the amount of user interest
(views, likes, comments), page presentation, publisher reputation, etc. An objective function is a mapping function from
the design space to the objective space, f : X → Y . A
multi-objective function can be represented as a vector of m
scalar objective functions, one for each objective variable,
F (~x) = (f1 (~x), f2 (~x), . . . , fm (~x)).
We formulate the generic MOO problem as the following
minimization problem,
min F (~x) = [f1 (~x), f2 (~x), . . . , fm (~x)]
~
x∈S
We start by first discussing the current ranking approach for
a popular News platform and then discuss its shortcomings
and challenges. Then we propose our MOO based solution
to this problem.
2.1
The objective space can be subject to a set of constraints
specified by the set S. The constraints can be represented
in terms of equality constraints, h(·), and/or inequality constraints, g(·). For example, in traffic shaping, these constraints can be a requirement for total clicks to be greater
than some target value for a partner, θclicks . This constraint
can be re-written as: fClicks (~x) − θClicks ≥ 0 or g(~x) ≥ 0,
where g(~x) = fClicks (~x) − θClicks .
The image of the feasible set under F (~x) for all possible
values of ~x under constraints S is called the attained set.
This is represented as C = {~y ∈ Rm : ~y = F (~x), ~x ∈ S}.
Typically, a single-objective optimization returns a single
globally optimal solution. In contrast, a solution to a MOO
problem is a set of solutions, where each solution is said to
be Pareto optimal.
Definition 1. Pareto optimal: A solution ~x∗ ∈ X to a
minimization problem is said to be Pareto optimal iff there
does not exist another solution ~x ∈ X, such that, F (~x) ≤
F (~x∗ ), and fi (~x) < fi (~x∗ ) for at least one of the objectives.
(Marler and Arora 2004)
A Pareto set (or Pareto front) is a set of all Pareto optimal
solutions. This set is composed of extreme solutions (design
variable assignments) which is optimal for one of the m objectives, while it is sub-optimal on others. All other solutions
represent a range of trade-offs between the different objectives. In practice, it is not feasible to compute all solutions
in the attained set C, therefore we require approximate solutions with some theoretical guarantees.
There are two important properties, or theoretical guarantees, of a good solution set that we are interested in 1) Convergence: each solution should be as close to the global optimum as possible; and 2) Diversity: the solution set should
be uniformly spread over the Pareto front. Convergence ensures optimal solutions, and diversity ensures that we cover
a wide variety of trade-offs between the multiple objectives.
Let variable t represent time. The generalized dynamic
MOO can be formulated using Equation (1) with a time parameter
min F (~x, t) = [f1 (~x, t), f2 (~x, t), . . . , fm (~x, t)]
Approach
(1)
where, S = {~x ∈ Rn : h(~x) = 0, g(~x) ≥ 0}
~
x∈St
2
(2)
where, St = {~x ∈ Rn : h(~x, t) = 0, g(~x, t) ≥ 0}
Our contributions in this work include the development
of a novel framework to perform constrained, MOO and
extend it to optimize on dynamic objectives. To the best
of our knowledge, this is the first real-world application of
MOO in an online data-driven problem. We propose modules in our framework more suited to the online problem.
This presented us with new challenges including practicality of the solutions. We address those concerns with a new
metric called Confidence Uncertainty.
Article Ranking
In this section we describe a simple Article Score as a linear additive function of user activities on each article (design
variables). Higher user activities tends to correspond to better user-engagement. For each news article, i, we compute
the article score as
Scorei = α·Freshness+β· Views+γ ·Likes+φ·Comments
(3)
where F reshness indicates how recent the article is; newer
articles tend to be more interesting on a social news website, number of V iews, number of Likes, and number of
Comments are user activity signals on the news article. α,
β, γ, and φ are the corresponding parameters of this scoring model. These parameters (or weights) are currently estimated using an optimization procedure called Grid Search.
Once the values of the parameters are assigned, each article
evaluates to an article score. This score is sorted to obtain a
ranked list of articles.
The Grid Search uses a brute force search approach to uniformly sample the values of the parameters by finding the
values of α, β, γ, and φ that maximizes a particular scalar
metric, like dwell-time. The data for this optimization task
is generated on a small fraction of randomly chosen users,
called the exploration bucket. User activities on articles and
the user engagement metrics, like dwell-time, are recorded
for each article for different choices of the four parameters
we are trying to learn. The design assignment that generates the highest dwell-time is considered to be optimal and
is assigned to generate the scores used for the general (out
of exploration bucket) user. The model is periodically updated to adapt to the changing dynamics in user behavior
and traffic obligations. The reader can generalize this application to any optimization task which involves parameter
(design variables) tuning of a function that evaluates to one
or more metric (objectives) in an online optimization task.
While the current approach is simple and works reasonably well in practice, it involves searching over the entire
space of parameters in a brute-force manner. The exploration
bucket users are subjected to low-quality experience while
we try to optimize these parameters. Searching over the entire space of possible parameters takes time and generally
hits on user experience as well. We formulate this parameter searching problem as an optimization problem, and do a
more principled and adaptive search of the parameter space.
Based on the parts of the parameter space we have explored
so far, our approach makes a decision on where to explore
next and which areas need not be explored any further.
2.2
The MOO framework
We first define the concept of non-dominance that is closely
related to the concept of Pareto optimality.
Definition 2. Non-Dominated and Dominated points: A vector of objective functions, F (~x∗ ) ∈ C, is non-dominated iff
there does not exist another vector, F (~x) ∈ C, such that,
F (~x) ≤ F (~x∗ ) with fi (~x) < fi (~x∗ ) for at least one of the
objectives. Otherwise, F (~x∗ ) is said to be dominated. (Marler and Arora 2004)
The subtle difference between the concept of Pareto optimality (Definition 1) and non-dominance (Definition 2) is
that Pareto optimality is defined in terms of the design space,
while non-dominance is defined in terms of the objective
space. For all practical purposes, this distinction is not important to us.
(Deb et al. 2002) propose a fast and efficient algorithm for
non-dominated sorting (NDS) to sort solutions in the objective space and identify the set of non-dominated solutions.
They use this NDS algorithm in their multi-objective evolutionary algorithm, NSGA-II, to identify the Pareto set in the
solutions available in each generation. The computational
complexity of this sorting approach is O(mK 2 ), where m
is the number of objectives, and K is size of the input to the
NDS algorithm. In NSGA-II, this size is fixed to be equal to
the user input parameter, population size.
Grid Search (Baseline model) Grid Search is the current
single-metric optimization technique used for our application. It is a simple brute force search over all possible values of the design variables. For each assignment of the design variables, it’s objective evaluation is computed using
the objective function. This objective function is a machine
learning model, usually Gradient Boosted Decision Trees
(GBDT) learned from the training data.
We extend this algorithm to handle multiple objectives.
We do so by using the non-domination heuristic. We use
NDS to identify the Pareto set from a set of different design
variable assignments. We also extend the current approach
to handle constraints by simply checking if each evaluation
meets the constraints, before we sort it using NDS.
Algorithm 1: Grid Search (Baseline)
Input: Training data (D), Design constraints (S),
Objective functions (F ), Increment percentage
(inc)
Output: Approximated Pareto set, Q
1 Initialize an empty list, f inal ← ∅.
2 Initialize a design vector p with dimension n assigning
the minimum value for each design variable seen in
the training data D.
3 Let Π be a list of all possible permutations of values of
p, where each variable takes discrete value increments,
inc% inclusively between the minimum and
maximum value of the design variable seen in D.
4 foreach π ∈ Π do
5
if F (π) satisfies constraint S then
6
Add π to f inal
7
end
8 end
9 Q ← Pareto-set in Non-Dominated-Sort(final)
10 return Q
Algorithm 1 describes our extension of the existing simple
brute force algorithm, the Grid Search, to handle multiple
objectives with constraints. The feasible solutions are sorted
using NDS. The algorithm takes as input the training data, D
(from exploration bucket), Design constraints, S (eg., traffic
shaping requirements), Objective functions, F (eg., GBDT
model to predict Dwell time from an assignment of design
variables.), and an increment, inc to define the granularity
of the different values explored for each design variables.
The time complexity of NDS is O(mK 2 ) for m objectives and K solutions. In the worst case, Grid Search returns
v d solutions, where v is the number of different values generated from a granularity of inc. The overall complexity is
O(mv d ). It can be clearly seen that it scales very poorly with
the number of design variables, d.
DO-NSGA-II In GA terminology, a collection of candidate solutions, ~x, is called a population. Typically, GAs initialize with random solutions as the initial population. GAs
then use two operators to generate new solutions from existing ones: crossover and mutation. An iteration of generating
new solutions from existing ones, using crossover and mutation operators, is called a generation. In the crossover operator two (or more) solutions, called parents, are selected
from the current population, their solutions are combined together to produce new solutions, called children. Crossover
ensures that elements of good solutions make it to the next
generation promoting convergence. The mutation operator
introduces random changes to a solution. The mutation operator creates diversity in the solution set and also helps escape local optima. The crossover and mutation operator is
sometimes preceded by the selection operator that identifies
parents from the population, which are promising candidates
to approach convergence towards global optima, and passes
them to the crossover and mutation operator. Each of the
three operators: selection, crossover, and mutation, can be
considered as separate modules in the GA.
We extend NSGA-II (Deb et al. 2002) (Non-dominated
Sorting genetic Algorithm) from which we used the NDS
algorithm to sort candidate solutions and identify the Pareto
set in Grid Search. Other than the fast and efficient sorting algorithm NSGA-II is a popular choice in MOO literature. Another advantage of NSGA-II is that there exists
a dynamic version to handle a dynamic objective function
(Deb, Karthik, and others 2007) which we can build on for
our application. The authors test their approach with a numerical function and not in a data-driven online setting. We
extend the NSGA-II framework to handle dynamic online
data. We call this extension- Dynamic, Online NSGA-II or
DO-NSGA-II.
Fitness function: The fitness function of a GA is the
heuristic used by the algorithm to evaluate the merit of a solution. DO-NSGA-II uses Pareto-dominance ranking as the
fitness function. We define this heuristic in Definition 2. A
set of solutions in a population, set A, in the GA is ranked
better (lower) than another set of solutions, set B, if each
individual solution in set A are non-dominated by each individual solution in set B.
Diversity: An important advantage of using GA is the
ability to control diversity of the population. DO-NSGA-II
uses crowding distance to prefer diverse solutions. A front
is a set of solutions in a population with the same Paretodominance rank. For each solution within a front, for each
objective, i, the solutions within a front are first sorted with
respect to that objective. The first and last solutions in the
sorted list are given a distance score of infinity (to protect extreme solutions). For all the solutions in between, the score
is computed by taking the normalized difference between the
objective values of the solutions, before and after the current
f (~
x
)−fj (~
xi−1 )
solution in the sorted list, Distance(~xi ) ← j fi+1
.
max −f min
j
j
For a detailed review of different fitness functions and diversity metrics used in GA literature, refer to (Konak, Coit, and
Smith 2006).
Selection operator: We use constraint dominance binary
selection (Deb et al. 2002) as our selection operator. This
module handles the constraints in DO-NSGA-II and promotes diversity. It first randomly chooses two solutions from
the current population. If one of the two is feasible, the feasible solution is selected to the next generation. If both are
unfeasible, the solution with the smaller constraint violation
is selected. If both solutions are feasible, the solution with
the better Pareto rank is selected. If they belong to the same
front, then the solution with the larger crowding distance is
selected. The choice of this algorithm is particularly attractive to our problem where the algorithm adapts to dynamic
changes in the constraints (changing traffic shaping targets).
Crossover operator: We use Simulated Binary Crossover
(SBX) (Deb and Beyer 2001) algorithm that is widely used
in literature. We introduce two parameters to DO-NSGA-II
that SBX depends upon: probability of crossover, Pc , and
the spread factor of the polynomial distribution ηc . Smaller
values of ηc return children that are very different from the
parents, while larger values return children similar to the
parents.
Mutation operator: We use Highly-Disruptive Polynomial
Mutation (Hamdan 2012b) algorithm, shown to be effective
at escaping the local optima on standard synthetic problems.
In principle it is similar to SBX, and introduces similar parameters. the probability of mutation, Pm , and the spread
factor for a polynomial distribution ηm .
The purpose of the mutation operator in our application is
to handle dynamics in the objective. Ideally we do not want
a big loss in performance to a dynamic change of objective. We motivate our implementation using Hypermutation
(Cobb 1990). Here, we increase Pm when a change is detected. We detect a change in DO-NSGA-II by re-evaluating
all the existing solutions, if the objective value has changed
from the previous generation, Pm increases.
The time-complexity of DO-NSGA-II remains O(mK 2 )
for m objectives and K solutions.
3
3.1
Experiments
Methodology
We evaluate DO-NSGA-II and Grid Search using three experiments. For each of the experiment, we analyze an online
dataset from a real-world problem of article ranking in a social news website as described in the Approach section. In
Algorithm 2: DO-NSGA-II
Input: Design constraints (S), Objective functions (F ),
Population size (K), Number of Generations
(E), SBX parameters (Pc , ηc ), Mutation
parameters (Pm , ηm ), Hypermutation period,
(epoch)
Output: Approximated Pareto set, Q
1 Randomly initialize K solutions, G0 . Sort the solutions,
Non-Dominated-Sort(G0 ).
2 children ← Crossover(ηc , G0 ) with a probability
Pc and Mutation(ηm , G0 ) with a probability Pm .
3 for t = 1 to E do
4
if objective change detected from Gt−1 then
5
Increase Pm for epoch generations.
6
Gt ← Gt−1 ∪ children
7
Sort the solutions, Non-Dominated-Sort(Gt )
8
Gt ← Choose top K solutions from Gt with the
best Pareto rank. In case of a tie, choose solution
with better crowding distance.
9
children ← Crossover(ηc , G0 ) with a
probability Pc and Mutation(ηm , G0 ) with a
probability Pm .
10 Q ← Pareto-set in Non-Dominated-Sort(GE )
11 return Q
the first experiment, we study the multi-objective optimization aspect of our framework. In the second experiment, we
study the constraints handling capabilities of DO-NSGAII and Grid Search. In the third and final experiment, we
examine the dynamics handling capabilities of our framework. Grid Search cannot handle dynamic behavior without
re-evaluating from the scratch.
For our experiments, we set the number of generations
in DO-NSGA-II to 500 and 500 population size. The SBX
parameters are Pc = 0.9, and ηc = 15, which are standard values used in literature. For mutation parameters, we
set Pm = 1/n and ηm = 1, as recommended by (Hamdan
2012a). We use our implemented hypermutation to increase
Pm = 1 when a dynamic change in objective is detected.
3.2
Evaluation
In the introduction section, we mentioned that the two important properties of a good approximated Pareto set were
convergence and diversity. Hypervolume (Hv.) metric is the
hypervolume (scalar value) occupied by some reference
point and the approximated Pareto set treated as a minimization problem. This metric is commonly accepted since
a large value indicates both better convergence and better diversity. For a solution set with k solutions and m objectives,
we compute the hypervolume using an O(k m−2 log k) time
and linear space complexity implementation as described
in (Fonseca, Paquete, and López-Ibánez 2006). We perform
min-max scaling of the objective space before computing
the hypervolume to avoid bias in the metric from different
scales of the dimensions of the objective space. Due to the
min-max scaling, the worst result in the minimization prob-
lem cannot be greater than (1.0, 1.0). So, we use a reference
point value of [2.0, 2.0].
We introduce a new metric called the Confidence Uncertainty (CU), which is the volume (scalar value) occupied by
the 90% confidence intervals of each of the m objectives.
We estimate the 90% confidence interval of each solution,
across each objective, using quantile regression. We return
CU as the product, of m objectives, from the min-max normalized difference between the upper and lower bounds of
the 90% confidence interval.
3.3
Datasets
Static data For the article ranking problem, we collect
data over a week on a small fraction of randomly chosen
users visiting the social news website. Each instance is a
unique article on the website. There are 28922 instances
in the data. We collect user-behavior data on these articles.
There are 4 design variables: article freshness, total user
views, likes, and comments. The 2 objective variables of our
interest are the total article clicks, and dwell-time (in milliseconds).
Dynamic data The dynamics of our objectives are a function of time. The number of clicks or dwell time, any article
receives depends upon the time of day. For example, it correlates to the work routine of a large subset of the users. Data
was collected at 4 different time intervals of the day: 1) Time
step 1, [00:00, 6:00] hours (50588 instances), 2) Time step
2, [06:00, 12:00] hours (51284 instances), 3) Time step 3,
[12:00, 18:00] hours (72019 instances), and 4) Time step 4,
[18:00, 00:00] hours (72705 instances). We choose 4 time
steps spanning across a day to demonstrate the capabilities
of DO-NSGA-II in handling dynamics. In practice, data can
be collected over finer time-steps depending upon the application.
Article freshness is computed as the difference (in hours)
between the article publish time of each article to the article
with the latest publish time (smaller value indicates newer
article). The other design variables (views, likes and comments) are count data. For each of the 4 design variables and
2 objectives (xi for instance i), we compute derived features
by taking their log(xi + 1 × 10−5 ). We add 1 × 10−5 to
the count data as a smoothing factor to naturally handle the
minimum value of zero when we take the log of the feature.
3.4
Modeling objectives
We split the training data into 80% training set, 10% test
set, and 10% validation set. We model the objectives using
Gradient Boosted Decision Trees (GBDT) in the scikit-learn
python package (Buitinck et al. 2013). Optimal parameters
for the GBDT are found using hyperopt (Bergstra, Yamins,
and Cox 2013) while optimizing on the coefficient of determination (R2 ) on the test set. The model performance on the
validation data are reported as follows.
Static data The Clicks model has an R2 of 0.5323 and a
mean squared error (MSE) of 5.3213. The Dwell time model
has an R2 of 0.56879 and MSE of 4.6169.
Dynamic data In time steps 1 through 4, the four Clicks
models have R2 values of (0.5223, 0.4701, 0.4629, 0.5059),
and MSE values of (5.8485, 6.4737, 6.2731, 5.9325).
The four Dwell time models have R2 values of
(0.5453, 0.5092, 0.5426, 0.5278), and MSE values of
(4.8183, 5.0986, 4.8743, 4.9541)
3.5
Experiment 1: Optimization performance
In this experiment, we compare the quality of the Pareto
curve returned by DO-NSGA-II and compare to Grid
Search, the baseline algorithm. Figure 1 shows the Pareto
curves we obtain from the two algorithms (shown as a maximization problem). We can immediately see that the solutions from DO-NSGA-II are better converged and spread
than Grid Search.
Figure 1: Comparing Pareto curve obtained by Grid Search
(green triangles) and DO-NSGA-II (red circles)
Total solutions
Hv
Average Hv
Grid Search
DO-NSGA-II
45
3.7161
3.6352
699
3.7905
3.7256
Table 1: Experimental results
The results in Table 1 shows that Grid Search returns
fewer solutions than DO-NSGA-II. The Figure 1 shows few
data points because a lot of design variable assignments
evaluate to a small number of different objective values. So,
they are stacked in the figure. DO-NSGA-II achieves a much
better overall Hypervolume. DO-NSGA-II also achieves a
much better Average Hypervolume, the average Hypervolume occupied by each individual solution in the approximated Pareto set. This measure shows that any arbitrary solution returned by DO-NSGA-II is likely to have a better
Hypervolume than one from Grid Search.
3.6
Experiment 2: Constrained MOO
In this experiment, we observe the constraint handling capabilities of DO-NSGA-II and Grid Search. We set a target
constraint on the Clicks objective, which could represent the
number of clicks required in a traffic shaping problem for an
article from a specific publisher to meet the contract requirements from that publisher. We test two scenarios 1) Possible
constraint, shows a reasonable number of clicks we need our
article ranking model to achieve; 2) Impossible constraint,
tests a constraint which would be unreasonable to achieve in
light of observed training data.
a smooth transition between the different pareto fronts, allowing for reasonable solution at any given generation in an
online setting.
References
(a) Possible constraint
(b) Impossible constraint
Figure 2: DO-NSGA-II and Grid Search under constraints
on Clicks objective.
Figure 2a shows the result under achievable constraint
of log10 (Clicks) > 6.25. We observe that DO-NSGA-II,
with its selection operator, adapts the search towards achieving the constraint. Grid Search cannot adapt and returns the
same solution set as in Experiment 1. Fewer solutions meet
the constraints. Figure 2a shows the result under an impossible to achieve constraint of log10 (Clicks) > 6.75. DONSGA-II still offers the best possible solution among the
solutions that do not meet the constraints. For this reason,
we chose this way of handling the constraints, organically,
during search.
3.7
Experiment 3: Dynamic optimization
This experiment studies the dynamics handling capability of
DO-NSGA-II. Grid Search cannot handle dynamics in the
objective.
(a) Pareto front over 4 time
steps.
(b) Change in Hypervolume
due to change in Pareto front.
Figure 3: DO-NSGA-II adapting under dynamics of different time steps.
Figure 3a shows the changing pareto front through the
four different time steps. The approximated pareto set in the
different time steps is represented with different shapes and
colors. Figure 3b shows how the hypervolume changes as the
algorithm adapts to the changing objective. We mark generations (0, 500, 1000, 1500) where the different time step
models transition. Our design of hypermutation allows for
[Agarwal, Chen, and Elango 2009] Agarwal, D.; Chen, B.C.; and Elango, P. 2009. Spatio-temporal models for estimating click-through rate. In Proceedings of the 18th international conference on World wide web, 21–30. ACM.
[Bergstra, Yamins, and Cox 2013] Bergstra, J.; Yamins, D.;
and Cox, D. D. 2013. Making a science of model search:
Hyperparameter optimization in hundreds of dimensions for
vision architectures. ICML (1) 28:115–123.
[Buitinck et al. 2013] Buitinck, L.; Louppe, G.; Blondel, M.;
Pedregosa, F.; Mueller, A.; Grisel, O.; Niculae, V.; Prettenhofer, P.; Gramfort, A.; Grobler, J.; Layton, R.; VanderPlas,
J.; Joly, A.; Holt, B.; and Varoquaux, G. 2013. API design
for machine learning software: experiences from the scikitlearn project. In ECML PKDD Workshop: Languages for
Data Mining and Machine Learning, 108–122.
[Cobb 1990] Cobb, H. G. 1990. An investigation into the
use of hypermutation as an adaptive operator in genetic algorithms having continuous, time-dependent nonstationary
environments. Technical report, DTIC Document.
[Deb and Beyer 2001] Deb, K., and Beyer, H.-G. 2001. Selfadaptive genetic algorithms with simulated binary crossover.
Evolutionary Computation 9(2):197–221.
[Deb et al. 2002] Deb, K.; Pratap, A.; Agarwal, S.; and Meyarivan, T. 2002. A fast and elitist multiobjective genetic
algorithm: Nsga-ii. IEEE transactions on evolutionary computation 6(2):182–197.
[Deb, Karthik, and others 2007] Deb, K.; Karthik, S.; et al.
2007. Dynamic multi-objective optimization and decisionmaking using modified nsga-ii: a case study on hydrothermal power scheduling. In International Conference
on Evolutionary Multi-Criterion Optimization, 803–817.
Springer.
[Fonseca, Paquete, and López-Ibánez 2006] Fonseca, C. M.;
Paquete, L.; and López-Ibánez, M. 2006. An improved
dimension-sweep algorithm for the hypervolume indicator.
In 2006 IEEE International Conference on Evolutionary
Computation, 1157–1163. IEEE.
[Hamdan 2012a] Hamdan, M. 2012a. The distribution index in polynomial mutation for evolutionary multiobjective
optimisation algorithms: An experimental study. In International Conference on Electronics Computer Technology,
Kanyakumari, India.
[Hamdan 2012b] Hamdan, M. 2012b. On the disruptionlevel of polynomial mutation for evolutionary multiobjective optimisation algorithms. Computing and Informatics 29(5):783–800.
[Konak, Coit, and Smith 2006] Konak, A.; Coit, D. W.; and
Smith, A. E. 2006. Multi-objective optimization using genetic algorithms: A tutorial. Reliability Engineering & System Safety 91(9):992–1007.
[Marler and Arora 2004] Marler, R. T., and Arora, J. S. 2004.
Survey of multi-objective optimization methods for engineering. Structural and multidisciplinary optimization
26(6):369–395.
| 2cs.AI
|
Minimax Optimal Estimators for Additive Scalar
Functionals of Discrete Distributions
Kazuto Fukuchi
arXiv:1701.06381v3 [cs.IT] 7 Dec 2017
1
∗1,2
and Jun Sakuma
†1,3,4
Department of Computer Science, Graduate School of System and
Information Engineering, University of Tsukuba
2
Japan Society for Promotion of Science
3
JST CREST
4
RIKEN Center for Advanced Intelligence Project
December 8, 2017
Abstract
In this paper,
P we consider estimators for an additive functional of φ, which is defined as
θ(P ; φ) = ki=1 φ(pi ), from n i.i.d. random samples drawn from a discrete distribution
P = (p1 , ..., pk ) with alphabet size k. We propose a minimax optimal estimator for the
estimation problem of the additive functional. We reveal that the minimax optimal rate
is characterized by the divergence speed of the fourth derivative of φ if the divergence
speed is high. As a result, we show there is no consistent estimator if the divergence
speed of the fourth derivative of φ is larger than p−4 . Furthermore, if the divergence
speed of the fourth derivative of φ is p4−α for α ∈ (0, 1), the minimax optimal rate is
2−2α
2
obtained within a universal multiplicative constant as (n lnkn)2α + k n .
1
Introduction
Let P be a probability measure with alphabet size k, and X be a discrete random variable
drawn from P . Without loss of generality, we can assume that the domain of P is [k], where
we denote [m] = {1, ..., m} for a positive integer m. We use a vector representation of P ;
P = (p1 , ..., pk ) where pi = P {X = i}. Let φ be a mapping from [0, 1] to R+ . Given a set of
i.i.d. samples Sn = {X1 , ..., Xn } from P , we deal with the problem of estimating an additive
functional of φ. The additive functional θ of φ is defined as
θ(P ; φ) =
k
X
φ(pi ).
i=1
We simplify this notation to θ(P ; φ) = θ(P ). Most entropy-like criteria can be formed in
terms of θ. For instance, when φ(p) = −p ln p, θ is Shannon entropy. For a positive real α,
letting φ(p) = pα , ln(θ(P ))/(1 − α) becomes Rényi entropy. More generally, letting φ = f
where f is a concave function, θ becomes f -entropies (Akaike, 1998).
Techniques for the estimation of the entropy-like criteria have been considered in various
fields, including physics (Lake and Moorman, 2011), neuroscience (Nemenman et al., 2004),
and security (Gu et al., 2005). In machine learning, methods that involve entropy estimation
∗ [email protected]
† [email protected]
1
were introduced for decision-trees (Quinlan, 1986), feature selection (Peng et al., 2005), and
clustering (Dhillon et al., 2003). For example, the decision-tree learning algorithms, i.e.,
ID3, C4.5, and C5.0 construct a decision tree in which the criteria for the tree splitting are
defined based on Shannon entropy (Quinlan, 1986). Similarly, information theoretic feature
selection algorithms evaluate the relevance between the features and the target using the
entropy (Peng et al., 2005).
The goal of this study is to derive the minimax optimal estimator of θ given a function
φ. For the precise definition of the minimax optimality, we introduce P
the minimax risk.
n
A sufficient statistic of P is a histogram N = (N1 , ..., Nk ), where Nj = i=1 1{Xi =j} and
k
N ∼ Multinomial(n, P ). The estimator of θ is defined as a function θ̂ : [n] → R. Then, the
quadratic minimax risk is defined as
2
R∗ (n, k; φ) = inf sup E θ̂(N ) − θ(P )
,
θ̂ P ∈Mk
where Mk is the set of all probability measures on [k], and the infimum is taken over
all estimators θ̂. With this definition of the minimax risk, an estimator θ̂ is minimax
(rate-)optimal if there exists a positive constant C such that
2
sup E θ̂(N ) − θ(P )
≤ CR∗ (n, k; φ).
P ∈Mk
A natural estimator of θ is the plugin or the maximum likelihood estimator, in which the
estimated value is obtained by substituting the empirical mean of the probabilities P into
θ. However, the estimator has a large bias for large k. Indeed, the plugin estimators for
φ(p) = −p ln p and φ(p) = pα have been shown to be suboptimal in the large-k regime in
recent studies (Wu and Yang, 2016; Jiao et al., 2015; Acharya et al., 2015).
Recent studies investigated the minimax optimal estimators for φ(p) = −p ln p and φ(p) = pα
in the large-k regime (Wu and Yang, 2016; Jiao et al., 2015; Acharya et al., 2015). However,
the results of these studies were only derived for these φ. Jiao et al. (2015) suggested that
the estimator is easily extendable to the general additive functional, although they did not
prove the minimax optimality.
In this paper, we propose a minimax optimal estimator for the estimation problem of the
additive functional θ for general φ under certain conditions on the smoothness. Our estimator achieves the minimax optimal rate even in the large-k regime for φ ∈ C 4 [0, 1] such
that φ(4) (p) is finite for p ∈ (0, 1], where C 4 [0, 1] denotes a class of four times differentiable
functions from [0, 1] to R. For such φ, we reveal a property of φ which can substantially
influence the minimax optimal rate.
Related work. The simplest way to estimate θ is to use the so-called plugin estimator or
the maximum likelihood estimator, in which the empirical probabilities are substituted into
θ as P . Letting P̃ = (p̂1 , ..., p̂k ) and p̂i = Ni /n, the plugin estimator is defined as
θplugin (N ) = θ(P̃ ).
The plugin estimator is asymptotically consistent under weak assumptions for fixed k (Antos
and Kontoyiannis, 2001). However, this is not true for the large-k regime. Indeed, Jiao et al.
(2015) and Wu and Yang (2016) derived a lower bound for the quadratic risk for the plugin
estimator of φ(p) = p ln(1/p) and φ(p) = pα . In the case of Shannon entropy, the lower
bound is given as
2
h
i
ln2 k
k
2
+
,
sup E (θplugin (N ) − θ(P )) ≥ C
n2
n
P ∈Mk
2
where C denotes a universal constant. The first term k 2 /n2 comes from the bias and it indicates that if k grows linearly with respect to n, the plugin estimator becomes inconsistent.
This means the plugin estimator is suboptimal in the large-k regime. Bias-correction methods, such as (Miller, 1955; Grassberger, 1988; Zahl, 1977), can be applied to the plugin estimator of φ(p) = −p ln p to reduce the bias whereas these bias-corrected estimators are still
suboptimal. The estimators based on Bayesian approaches in (Schürmann and Grassberger,
1996; Schober, 2013; Holste et al., 1998) are also suboptimal (Han et al., 2015).
Many researchers have studied estimators that can consistently estimate the additive functional with sublinear samples with respect to the alphabet size k to derive the optimal
estimator in the large-k regime. The existence of consistent estimators even with sublinear
samples were first revealed in Paninski (2004), but an explicit estimator was not provided.
Valiant and Valiant (2011a) introduced an estimator based on linear programming that
consistently estimates φ(p) = −p ln p with sublinear samples. However, the estimator of
(Valiant and Valiant, 2011a) has not been shown to achieve the minimax rate even in a
more detailed analysis in (Valiant and Valiant, 2011b). Recently, Acharya et al. (2015)
showed that the bias-corrected estimator of Rényi entropy achieves the minimax optimal
rate in regard to the sample complexity if α > 1 and α ∈ N, but they did not show the
minimax optimality for other α. Jiao et al. (2015) introduced a minimax optimal estimator
for φ(p) = −p ln p for any α ∈ (0, 3/2) in the large-k regime. Wu and Yang (2015) derived
a minimax optimal estimator for φ(p) = 1p>0 . For φ(p) = −p ln p, Jiao et al. (2015); Wu
and Yang (2016) independently introduced the minimax optimal estimators in the large-k
regime. In the case of Shannon entropy, the optimal rate was obtained as
ln2 k
k2
+
.
(n ln n)2
n
The first term indicates that the introduced estimator can consistently estimate Shannon
entropy if n ≥ Ck/ ln k.
The estimators introduced by Wu and Yang (2016); Jiao et al. (2015); Acharya et al. (2015)
are composed of two estimators: the bias-corrected plugin estimator and the best polynomial estimator. The bias-corrected plugin estimator is composed of the sum of the plugin
estimator and a bias-correction term which offsets the second-order approximation of the
bias as in (Miller, 1955). The best polynomial estimator is an unbiased estimator of the
polynomial that best approximates φ in terms of the uniform error. Specifically, the best
approximation for the polynomial of φ in an interval I ⊆ [0, 1] is the polynomial g that minimizes supx∈I |φ(x) − g(x)|. Jiao et al. (2015) suggested that this estimator can be extended
for the general additive functional θ. However, the minimax optimality of the estimator was
only proved for specific cases of φ, including φ(p) = −p ln p and φ(p) = pα . Thus, to prove
the minimax optimality for other φ, we need to individually analyze the minimax optimality for specific φ. Here, we aim to clarify which property of φ substantially influences the
minimax optimal rate when estimating the additive functional.
Besides, the optimal estimators for divergences with large alphabet size have been investigated in (Bu et al., 2016; Han et al., 2016; Jiao et al., 2016). The estimation problems of
divergences are much complicated than the additive function, while the similar techniques
were applied to derive the minimax optimality.
Our contributions. In this paper, we propose the minimax optimal estimator for θ(P ; φ).
We reveal that the divergence speed of the fourth derivative of φ plays an important role
in characterizing the minimax optimal rate. Informally, for β > 0, the meaning of “the
divergence speed of a function f (p) is p−β ” is that |f (p)| goes to infinity at the same speed
as p−β when p approaches 0. When the divergence speed of the fourth derivative of φ(p) is
p−β , the fourth derivative of φ diverges faster as β increases.
Our results are summarized in Figure 1. Figure 1 illustrates the relationship between the
divergence speed of the fourth derivative of φ and the minimax optimality of the estimation
3
Figure 1: Relationship between the divergence speed of the fourth derivative of φ and the
minimax optimality of the estimation problem of θ(P ; φ).
problem of θ(P ; φ). In Figure 1, the outermost rectangle represents the space of the four
times continuous differentiable functions C 4 [0, 1]. The innermost rectangle denotes the
subset class of C 4 [0, 1] such that the absolute value of its fourth derivative φ(4) (p) is finite
for any p ∈ (0, 1]. In this subclass of φ, the horizontal direction represents the divergence
speed of the fourth derivative of φ, in which a faster φ is on the left-hand side and a
slower φ is on the right-hand side. The φ with an explicit form and divergence speed is
denoted by a point in the rectangle. For example, the black circle denotes φ(p) = −p ln p
where the divergence speed of the fourth derivative of this φ is p−3 . Class B denotes a
set of any function φ such that the divergence speed of the fourth derivative is pα−4 where
α ∈ (0, 1). As already discussed, existing methods have achieved minimax optimality in
the large-k regime for specific φ, including φ(p) = −p ln p (black circle in Figure 1) and
φ(p) = pα (middle line in Figure 1 where the white circle denotes that there is no α > 0
such that the divergence speed is p−3 ).
We investigate the minimax optimality of the estimation problem of θ for φ in Class A and
Class B. Class A is a class of φ such that the divergence speed of the fourth derivative is faster
than p−4 . Class B is a class of φ such that the divergence speed of the fourth derivative is
pα−4 where α ∈ (0, 1). In Class A, we show that we cannot construct a consistent estimator
of θ for any φ in Class A (the leftmost hatched area in Figure 1, Proposition 1). In other
words, the minimax optimal rate is larger than constant order if the divergence speed of the
fourth derivative is faster than p−4 . Thus, there is no need to derive the minimax optimal
estimator in Class A.
Also, we derive the minimax optimal estimator for any φ in Class B (the middle hatched
area in Figure 1, Theorem 1). For example, φ(p) = pα (Réyni entropy case), φ(p) =
cos(cp)pα , and φ(p) = ecp pα for α ∈ (0, 1) include the coverage of our estimator, where c is
a universal constant. Intuitively, since the large derivative makes the estimation problem θ
more difficult, the minimax rate decreases if the derivative of φ diverges faster. Our minimax
optimal rate reflects this behavior. For φ in Class B, the minimax optimal rate is obtained
as
k2
k 2−2α
+
,
(n ln n)2α
n
4
where k & ln 3 n if α ∈ (0, 1/2]. We can clearly see that this rate decreases for larger α, i.e.,
a slower divergence speed.
4
Currently, the minimax optimality of φ in Class C is an open problem. However, we provide
a notable discussion in Section 3.
2
Preliminaries
Notations. We now introduce some additional notations. For any positive real sequences
{an } and {bn }, an & bn denotes that there exists a positive constant c such that an ≥ cbn .
Similarly, an . bn denotes that there exists a positive constant c such that an ≤ cbn .
Furthermore, an bn implies an & bn and an . bn . For an event E, we denote its
complement by E c . For two real numbers a and b, a ∨ b = max{a, b} and a ∧ b = min{a, b}.
For a function φ : R → R, we denote its i-th derivative as φ(i) .
Poisson sampling. We employ the Poisson sampling technique to derive upper and
lower bounds for the minimax risk. The Poisson sampling technique models the samples as independent Poisson distributions, while the original samples follow a multinomial
distribution. Specifically, the sufficient statistic for P in the Poisson sampling is a histogram Ñ = (Ñi , ..., Ñk ), where Ñ1 , ..., Ñk are independent random variables such that
Ñi ∼ Poi(npi ). The minimax risk for Poisson sampling is defined as follows:
2
R̃∗ (n, k; φ) = inf sup E θ̂(Ñ ) − θ(P )
.
{θ̂} P ∈Mk
The minimax risk of Poisson sampling well approximates that of the multinomial distribution. Indeed, Jiao et al. (2015) presented the following lemma.
Lemma 1 (Jiao et al. (2015)). The minimax risk under the Poisson model and the multinomial model are related via the following inequalities:
R̃∗ (2n, k; φ) − sup |θ(P )|e−n/4 ≤ R∗ (n, k; φ) ≤ 2R̃∗ (n/2, k; φ).
P ∈Mk
Lemma 1 states R∗ (n, k; φ) R̃∗ (n, k; φ), and thus we can derive the minimax rate of the
multinomial distribution from that of the Poisson sampling.
Best polynomial approximation. Acharya et al. (2015); Wu and Yang (2016); Jiao et al.
(2015) presented a technique of the best polynomial approximation for deriving the minimax
optimal estimators and their lower bounds for the risk. Let PL be the set of polynomials of
degree L. Given a function φ, a polynomial p, and an interval I ⊆ [0, 1], the uniform error
between φ and p on I is defined as
sup|φ(x) − p(x)|.
(1)
x∈I
The best polynomial approximation of φ by a degree-L polynomial with a uniform error is
achieved by the polynomial p ∈ PL that minimizes Eq (1). The error of the best polynomial
approximation is defined as
EL (φ, I) = inf sup|φ(x) − p(x)|.
p∈PL x∈I
The error rate with respect to the degree L has been studied since the 1960s (Timan et al.,
1965; Petrushev and Popov, 2011; Ditzian and Totik, 2012; Achieser, 2013). The polynomial
that achieves the best polynomial approximation can be obtained, for instance, by the Remez
algorithm (Remez, 1934) if I is bounded.
5
3
Main results
Suppose φ is four times continuously differentiable on (0, 1]1 . We reveal that the divergence
speed of the fourth derivative of φ plays an important role for the minimax optimality of the
estimation problem of the additive functional. Formally, the divergence speed is defined as
follows.
Definition 1 (divergence speed). For an integer m ≥ 1, let φ be an m times continuously
differentiable function on (0, 1]. For β > 0, the divergence speed of the mth derivative of φ
is p−β if there exist finite constants W > 0, cm , and c0m such that for all p ∈ (0, 1]
φ(m) (p) ≤ βm−1 W p−β + cm , and φ(m) (p) ≥ βm−1 W p−β + c0m ,
where βm =
Qm
i=1 (i
− m + β).
A larger β implies faster divergence. We analyze the minimax optimality for two cases:
the divergence speed of the fourth derivative of φ is i) larger than p−4 (Class A), and ii)
pα−4 (Class B), for α ∈ (0, 1).
Minimax optimality for Class A. We now demonstrate that we cannot construct a
consistent estimator for any n and k ≥ 3 if the divergence speed of φ is larger than p−4 .
Proposition 1. Let φ be a continuously differentiable function on (0, 1]. If there exists
finite constants W > 0 and c01 such that for p ∈ (0, 1]
φ(1) (p) ≥ W p−1 + c01 ,
then there is no consistent estimator, i.e., R∗ (n, k; φ) & 1.
The proof of Proposition 1 is given in Appendix D. From Lemma 15, the divergence speed of
the first derivative is p−1 if that of the fourth derivative is p−4 . Thus, if the divergence speed
of φ is greater than p−4 , we cannot construct an estimator that consistently estimates θ for
any probability measure P ∈ Mk . Consequently, there is no need to derive the minimax
optimal estimator in this case.
Minimax optimality for Class B. We derive the minimax optimal rate for φ in which the
divergence speed of its fourth derivative is pα−4 for α ∈ (0, 1). Thus, we make the following
assumption.
Assumption 1. Suppose φ is four times continuously differentiable on (0, 1]. For α ∈ (0, 1),
the divergence speed of the fourth derivative of φ is pα−4 .
Note that a set of φ satisfying Assumption 1 is Class B depicted in Figure 1. The divergence speed increases as α decreases. Under Assumption 1, we derive the minimax optimal
estimator of which the minimax rate is given by the following theorems.
4
1/α
Theorem 1. Under Assumption 1 with α ∈ (0, 1/2], if n & kln k and k & ln 3 n,
R∗ (n, k; φ)
k2
.
(n ln n)2α
Otherwise, there is no consistent estimator, i.e., R∗ (n, k; φ) & 1.
1/α
Theorem 2. Under Assumption 1 with α ∈ (1/2, 1), if n & kln k
R∗ (n, k; φ)
k2
k 2−2α
.
+
(n ln n)2α
n
Otherwise, there is no consistent estimator, i.e., R∗ (n, k; φ) & 1.
1 We
say that a function φ : [0, 1] → R+ is differentiable at 1 if limh→−0
6
φ(1+h)−φ(1)
h
exists.
Theorems 1 and 2 are proved by combining the results in Sections 6 and 7. The minimax
optimal rate in Theorems 1 and 2 are characterized by the parameter for the divergence
speed α from Assumption 1. From Theorems 1 and 2, we can conclude that the minimax
optimal rate decreases as the divergence speed increases.
The explicit estimator that achieves the optimal minimax rate shown in Theorems 1 and 2
are described in the next section.
Remark 1. Assumption 1 covers φ(p) = pα for α ∈ (0, 1), but does not for all existing works.
For φ(p) = −p ln(p) and φ(p) = pα with α ≥ 1, the divergence speed of these φ is lower
than pα−4 for α ∈ (0, 1). Indeed, the divergence speed of φ(p) = −p ln(p) and φ(p) = pα for
α ≥ 1 are p−3 and pα−4 , respectively. We can expect that the corresponding minimax rate
is characterized by the divergence speed even if the divergence speed is lower than pα−4 for
α ∈ (0, 1). The analysis of the minimax rate for lower divergence speeds remains an open
problem.
4
Estimator for θ
In this section, we describe our estimator for θ in detail. Our estimator is composed of
the bias-corrected plugin estimator and the best polynomial estimator. We first describe
the overall estimation procedure on the supposition that the bias-corrected plugin estimator
and the best polynomial estimator are black boxes. Then, we describe the bias-corrected
plugin estimator and the best polynomial estimator in detail.
For simplicity, we assume the samples are drawn from the Poisson sampling model, where
we first draw n0 ∼ Poi(2n), and then draw n0 i.i.d. samples Sn0 = {X1 , ..., Xn0 }. Given the
samples Sn0 , we first partition the samples into two sets. We use one set of the samples
to determine whether the bias-corrected plugin estimator or the best polynomial estimator
n0
should be employed, and the other set to estimate θ. Let {Bi }i=1 be i.i.d. random variables
drawn from the Bernoulli distribution with parameter 1/2, i.e., P{Bi = 0} = P{Bi = 1} =
1/2 for i = 1, ..., n0 . We partition (X1 , ..., Xn0 ) according to (B1 , ..., Bn0 ), and construct the
histograms Ñ and Ñ 0 , which are defined as
0
Ñi =
n
X
0
1Xj =i 1Bj =0 ,
Ñi0
=
j=1
n
X
1Xj =i 1Bj =1 , for i ∈ [n0 ].
j=1
Then, Ñ and Ñ 0 are independent histograms, and Ñi , Ñi0 ∼ Poi(npi ).
Given Ñ 0 , we determine whether the bias-corrected plugin estimator or the best polynomial
estimator should be employed for each alphabet. Let ∆n,k be a threshold depending on n
and k to determine which estimator is employed, which will be specified as in Theorem 5
on page 10. We apply the best polynomial estimator if Ñi0 < 2∆n,k , and otherwise, i.e.,
Ñi0 ≥ 2∆n,k , we apply the bias-corrected plugin estimator. Let φpoly and φplugin be the best
polynomial estimator and the bias-corrected plugin estimator for φ, respectively. Then, the
estimator of θ is written as
θ̃(Ñ ) =
k
X
1Ñ 0 ≥2∆n,k φplugin (Ñi ) + 1Ñ 0 <2∆n,k φpoly (Ñi ) .
i
i
i=1
Finally, we truncate θ̃ so that the final estimate is not outside of the domain of θ.
θ̂(Ñ ) =(θ̃(Ñ ) ∧ θsup ) ∨ θinf ,
where θinf = inf P ∈Mk θ(P ) and θsup = supP ∈Mk θ(P ). Next, we describe the details of the
best polynomial estimator φpoly and the bias-corrected plugin estimator φplugin .
7
Best polynomial estimator. The best polynomial estimator is an unbiased estimator
L
of the polynomial that provides the best approximation of φ. Let {am }m=0 be coefficients
of the polynomial that achieves the best approximation of φ by a degree-L polynomial
4∆
with range I = [0, nn,k ], where L is as specified in Theorem 5 on page 10. Then, the
approximation of φ by the polynomial at point pi is written as
L
X
φL (pi ) =
am pm
i .
(2)
m=0
From Eq (2), an unbiased estimator of φL can be derived from an unbiased estimator of pm
i .
For the random variable Ñi drawn from the Poisson distribution with mean parameter npi ,
i!
becomes (npi )m . Thus, (Ñnim)m
the expectation of the mth factorial moment (Ñi )m = (Ñ Ñ−m)!
i
is an unbiased estimator of pm
i . Substituting this into Eq (2) gives the unbiased estimator
of φL (pi ) as
L
X
am
(Ñi )m .
nj
m=0
φ̄poly (Ñi ) =
Next, we truncate φ̄poly so that it is not outside of the domain of φ(p). Let φinf, ∆n,k =
n
inf p∈[0, ∆n,k ] φ(p) and φsup, ∆n,k = supp∈[0, ∆n,k ] φ(p). Then, the best polynomial estimator
n
n
n
is defined as
φpoly (Ñi ) = (φ̄poly (Ñi ) ∧ φsup, ∆n,k ) ∨ φinf, ∆n,k .
n
n
Bias-corrected plugin estimator. In the bias-corrected plugin estimator, we apply the
bias correction of (Miller, 1955). Applying the second-order Taylor expansion to the bias of
the plugin estimator gives
"
!
#
!
!2
(2)
Ñi
Ñi
φ (pi ) Ñi
E φ
− φ(pi ) ≈Eφ(1) (pi )
− pi +
− pi
n
n
2
n
=
pi φ(2) (pi )
.
2n
(2)
Thus, we include − Ñi φ 2n(2Ñi /n) as a bias-correction term in the plugin estimator φ(Ñi /n),
which offsets the second-order approximation of the bias. However, we do not directly apply
the bias-corrected plugin estimator to estimate φ(pi ) for two reasons. First, the derivative
of φ is large near 0, which results in a large bias, and second, φ(p) for p > 1 is undefined
even though Ñi /n can exceed 1. Thus, we apply the bias-corrected plugin estimator to the
function φ̄ ∆n,k defined below instead of φ. Define
n
HL (p; φ, a, b)
L
L−m
X
X (−1)` (L + `)!
φ(m) (a)
(p − a)m (p − b)L+1
(a − b)−L−1−` (p − a)`
m!
`!L!
m=1
`=0
L−m
L
X φ(m) (a)
X L+1
p−a
=φ(a) +
(p − a)m
B`,L+`+1
,
m!
L+`+1
b−a
m=1
=φ(a) +
`=0
n
where Bν,n (x) = ν xν (1 − x)n−ν denotes the Bernstein basis polynomial. Then,
HL (p; φ, a, b) denotes a function that interpolates between φ(a) and φ(b) using Hermite inter(i)
polation. From generalized Hermite interpolation (Spitzbart, 1960), HL (a; φ, a, b) = φ(i) (a)
8
(i)
for i = 0, ..., L and HL (b; φ, a, b) = 0 for i = 1, ..., L. The function φ̄ ∆n,k is defined as
n
∆n,k ∆n,k
∆n,k
; φ,
,
H4
2n
n
2n
H4 p; φ, ∆n,k , ∆n,k
n
2n
φ̄ ∆n,k (p) =
n
H4 (p; φ, 1, 2)
H4 (2; φ, 1, 2)
φ(p)
From this definition, φ̄ ∆n,k = φ if p ∈ [
n
∆n,k
n , 1].
(1)
if p ≤
if
∆n,k
,
2n
∆n,k
∆n,k
<p<
,
2n
n
if 1 < p < 2,
if p ≥ 2,
otherwise .
From Hermite interpolation, the function
(4)
φ̄ ∆n,k is four times differentiable on R+ and φ̄ ∆n,k (p) = ... = φ̄ ∆n,k (p) = 0 for p ≤
n
n
∆n,k
2n
and
n
p ≥ 2. By introducing φ̄ ∆n,k , we can bound the fourth derivative of φ̄ ∆n,k using ∆n,k , and
n
n
this enables us to control the bias with the threshold parameter ∆n,k . Using φ̄ ∆n,k instead
n
of φ yields the bias-corrected plugin estimator
N̄i
Ñi (2)
N̄i
φplugin (Ñi ) = φ̄ ∆n,k
− 2 φ̄ ∆n,k
.
n
2n
n
n
n
5
(3)
Remark about Differentiability for Analysis
Why is the minimax rate characterized by the divergence speed of the fourth derivative? Indeed, most of the results can be obtained on a weaker assumption compared to Assumption 1
regarding differentiability, which is formally defined as follows.
Assumption 2. Suppose φ is two times continuously differentiable on (0, 1]. For α ∈ (0, 1),
the divergence speed of the second derivative of φ is pα−2 .
Assumption 2 only requires two times continuous differentiability, whereas Assumption 1
requires four times. Only the analysis of the bias-corrected plugin estimator requires Assumption 1 to achieve the minimax rate due to the bias-correction term in Eq (3). The
bias-correction term is formed as the plugin estimator of the second derivative of φ, and
its convergence rate is highly dependent on the smoothness of the second derivative. The
smoothness of the second derivative of φ is characterized by the fourth derivative of φ,
and thus Assumption 1 is required to derive the error bound of the bias-corrected plugin
estimator. Another bias-correction method might weaken the assumption as in Assumption 2.
6
Analysis of Lower Bound
In this section, we derive a lower bound for the minimax rate of θ. Under Assumption 2,
we can derive the lower bound of the minimax risk as in the following theorem.
Theorem 3. Under Assumption 2, for k ≥ 3, we have
R∗ (n, k; φ) &
k 2−2α
.
n
The lower bound is obtained by applying Le Cam’s two-point method (see (Tsybakov,
2009)). The details of the proof of Theorem 3 can be found in Appendix B. Next, we derive
another lower bound for the minimax rate.
9
Theorem 4. Under Assumption 2, if n &
k1/α
ln k ,
R∗ (n, k; φ) &
we have
k2
,
(n ln n)2α
4
where we need k & ln 3 n if α ∈ (0, 1/2].
The proof is accomplished in the same manner as (Wu and Yang, 2016, Proposition 3). The
details of the proof of Theorem 4 are also found in Appendix B. Combining Theorems 3
2−2α
2
and 4, we get the lower bounds in Theorems 1 and 2 as R∗ (n, k; φ) & (n lnk n)2α ∨ k n &
k2
(n ln n)2α
7
+
k2−2α
n .
Analysis of Upper Bound
Here, we derive the upper bound for the worst-case risk of the estimator.
Theorem 5. Suppose ∆n,k =√C2 ln n and L = bC1 ln nc where C1 and C2 are universal
constants such that 6C1 ln 2 + 4 C1 C2 (1 + ln 2) < 1 and C2 > 16. Under Assumption 1, the
worst-case risk of θ̂ is bounded above by
2
k2
k 2−2α
+
,
sup E θ̂(Ñ ) − θ(P )
.
2α
(n ln n)
n
P ∈Mk
4
where we need k & ln 3 n if α ∈ (0, 1/2].
To prove Theorem 5, we derive the bias and the variance of θ̂.
Lemma 2. Given P ∈ Mk , for 1 . ∆n,k ≤ n, the bias of θ̂ is bounded above by
k
h
i X
h
i
(e/4)∆n,k + Bias φplugin (Ñi ) − φ(pi ) 1npi >∆n,k
Bias θ̃(Ñ ) − θ(P ) .
i=1
!
h
i
−∆n,k /8
+ Bias φpoly (Ñi ) − φ(pi ) 1npi ≤4∆n,k + e
.
Lemma 3. Given P ∈ Mk , for 1 . ∆n,k ≤ n, the variance of θ̂ is bounded above by
k
h
i X
h
i
Var θ̃(Ñ ) − θ(P ) .
(e/4)∆n,k + Var φplugin (Ñi ) − φ(pi ) 1npi >∆n,k
i=1
h
i
+ Var φpoly (Ñi ) − φ(pi ) 1npi ≤4∆n,k + e−∆n,k /8
!
h
i
h
i2
+ Bias φplugin (Ñi ) − φ(pi ) + Bias φpoly (Ñi ) − φ(pi )
1∆n,k ≤pi ≤4∆n,k .
The proofs of Lemmas 2 and 3 are left to Appendix C. As proved in Lemmas 2 and 3,
the bounds on the bias and the variance of our estimator are obtained with the bias and
the variance of the plugin and the best polynomial estimators for each individual alphabet.
Thus, we next analyze the bias and the variance of the plugin and the best polynomial
estimators.
Analysis of the best polynomial estimator. The following lemmas provide the upper
bounds on the bias and the variance of the best polynomial estimator.
10
Lemma 4. Let Ñ ∼ Poi(np). Given an integer L and a positive real ∆, let φL (p) =
PL
m
be the optimal uniform approximation of φ by degree-L polynomials on [0, ∆],
m=0 am p
PL
and gL (Ñ ) = m=0 am (Ñ )m /nm be an unbiased estimator of φL (p). Under Assumption 2,
we have
h
i r
h
i ∆ α
.
Bias (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − φ(p) . Var gL (Ñ ) − φL (p) +
L2
Lemma 5. Let Ñ ∼ Poi(np). Given an integer L and a positive real ∆ & n1 , let φL (p) =
PL
m
be the optimal uniform approximation of φ by degree-L polynomials on [0, ∆],
m=0 am p
PL
and gL (Ñ ) = m=0 am (Ñ )m /nm be an unbiased estimator of φL (p). Assume Assumption 2.
If p ≤ ∆ and 2∆3 L ≤ n, we have
√
i ∆3 L64L (2e)2
Var (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − φ(p) .
n
h
∆nL
.
The proofs of Lemmas 4 and 5 can be found in Appendix C.
Analysis of the plugin estimator. The following lemmas provide the upper bounds for
the bias and the variance of the plugin estimator.
Lemma 6. Assume Assumption 1 and n1 . ∆ < p ≤ 1. Let Ñ ∼ Poi(np). Then, we have
!
!
#
"
Ñ (2) Ñ
1
p
Ñ
− 2 φ̄∆
− φ(p) . 2 2−α + 2 .
Bias φ̄∆
n
2n
n
n ∆
n
Lemma 7. Assume Assumption 1 and n1 . ∆ < p ≤ 1. Let Ñ ∼ Poi(np). Then, we have
"
!
!
#
Ñ (2) Ñ
pφ(2) (p)
p2α−1
1
p
Ñ
Var φ̄∆
− 2 φ̄∆
− φ(p) +
.
+ 4 4−2α + .
n
2n
n
2n
n
n ∆
n
The proofs of Lemmas 6 and 7 are left to Appendix C.
Proof for the Upper Bound. Combining Lemmas 2 to 7, we prove Theorem 5.
Proof of Theorem 5. Set L = bC1 ln nc and ∆n,k = C2 ln n where C1 and C2 are some
positive constants. Substituting Lemmas 4 to 7 into Lemmas 2 and 3 yields
h
i
Bias θ̂(Ñ ) − θ(P )
.
k
X
i=1
1
nC2 (ln 4−1)
pi
1
(ln n)2 n3C1 ln 2+2
+ α
+
+
n (ln n)2−α
n2
n2
√
C1 C2 (ln 2+1)
1
1
+
+ C /8
(n ln n)α
n 2
√
≤
k
nC2 (ln 4−1)
k
1
k(ln n)2 n3C1 ln 2+2
+ α
+
+
n (ln n)2−α
n2
n2
C1 C2 (ln 2+1)
+
11
!
k
k
+ C /8 ,
(n ln n)α
n 2
and
h
i
Var θ̂(Ñ ) − θ(P )
.
k
X
i=1
1
nC2 (ln 4−1)
p2α−1
pi
1
(ln n)4 n6C1 ln 2+4
+ 1pi ≥C2 ln n/n i
+
+ 2α
+
n
n (ln n)4−2α
n
n4
√
1
nC2 /8
.
+
1
pi
(ln n)2 n3C1 ln 2+2
+ 2+
α
2−α
n (ln n)
n
n2
C1 C2 (ln 2+1)
√
C1 C2 (ln 2+1)
+
1
+
(n ln n)α
k 2−2α
k
1
k
+
∨ 2α 1−2α + 2α
4−2α
n √ n ln
n
n n (ln n)
1
k
k
k(ln n)4 n6C1 ln 2+4 C1 C2 (ln 2+1)
+ C /8 + 2α
+ 4
+
4
4−2α
2
n
n (ln n)
n
n
k
nC2 (ln 4−1)
+
k(ln n)4 n6C1 ln 2+4
+
n4
√
C1 C2 (ln 2+1)
+
k
,
(n ln n)2α
where
we use Lemmas 17 and 18. For δ > 0, as long as C2 (ln 4 − 1) ≥ 2α + δ, 6C1 ln 2 +
√
4 C1 C2 (ln 2 + 1) ≤ 3 − 2α − δ, and C2 /8 ≥ 2α + δ, we have
h
i2 1
k2
k2
1
k2
Bias θ̂(Ñ ) − θ(P ) . 4 + 2α+δ +
. 4+
2α
n
n
(n ln n)
n
(n ln n)2α
h
i k 2−2α
k
k
k
∨ 2α 1−2α + 2α+δ +
Var θ̂(Ñ ) − θ(P ) .
n
(n ln n)2α
n ln
n n
k 2−2α
k
k
.
∨ 2α 1−2α +
n
n ln
n (n ln n)2α
(4)
(5)
There exist the constants C1 and C2 that satisfies these conditions, for example, C1 <
1/6 ln 2 and C2 > 16. Since θ̂(Ñ ), θ(P ) ∈ [θinf , θsup ], the bias-variance decomposition gives
2
2
sup E θ̂(Ñ ) − θ(P )
≤ sup E θ̃(Ñ ) − θ(P )
P ∈Mk
P ∈Mk
h
i2
h
i
≤ Bias θ̃(Ñ ) − θ(P )
+ Var θ̃(Ñ ) − θ(P ) .
(6)
Substituting Eqs (4) and (5) into Eq (6) yields
2 k 2−2α
k
k2
sup E θ̂(Ñ ) − θ(P )
.
∨ 2α 1−2α +
.
n
n ln
n (n ln n)2α
P ∈Mk
4
If α ∈ (0, 1/2] and k & ln 3 , the last term is dominated. If α ∈ (1/2, 1), the term
is dominated by
k
2−2α
n
k
n2α ln1−2α n
.
Acknowledgment
This work is supported by JST CREST and KAKENHI No. 16H02864.
References
Hirotogu Akaike. Information theory and an extension of the maximum likelihood principle.
In Selected Papers of Hirotugu Akaike, pages 199–213. Springer, 1998.
12
!2 !
Douglas E Lake and J Randall Moorman. Accurate estimation of entropy in very short. Am
J Physiol Heart Circ Physiol, 300:H319–H325, 2011.
Ilya Nemenman, William Bialek, and Rob de Ruyter van Steveninck. Entropy and information in neural spike trains: Progress on the sampling problem. Physical Review E, 69(5):
056111, 2004.
Yu Gu, Andrew McCallum, and Don Towsley. Detecting anomalies in network traffic using
maximum entropy estimation. In Proceedings of the 5th ACM SIGCOMM conference on
Internet Measurement, pages 32–32. USENIX Association, 2005.
J. Ross Quinlan. Induction of decision trees. Machine learning, 1(1):81–106, 1986.
Hanchuan Peng, Fuhui Long, and Chris Ding. Feature selection based on mutual information
criteria of max-dependency, max-relevance, and min-redundancy. IEEE Transactions on
pattern analysis and machine intelligence, 27(8):1226–1238, 2005.
Inderjit S Dhillon, Subramanyam Mallela, and Dharmendra S Modha. Information-theoretic
co-clustering. In Proceedings of the ninth ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 89–98. ACM, 2003.
Yihong Wu and Pengkun Yang. Minimax rates of entropy estimation on large alphabets
via best polynomial approximation. IEEE Transactions on Information Theory, 62(6):
3702–3720, 2016.
Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman. Minimax estimation of
functionals of discrete distributions. Information Theory, IEEE Transactions on, 61(5):
2835–2885, 2015.
Jayadev Acharya, Alon Orlitsky, Ananda Theertha Suresh, and Himanshu Tyagi. The
complexity of estimating rényi entropy. In Proceedings of the Twenty-Sixth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January
4-6, 2015, pages 1855–1869. SIAM, 2015. doi: 10.1137/1.9781611973730.124.
Andrs Antos and Ioannis Kontoyiannis. Convergence properties of functional estimates for
discrete distributions. Random Structures & Algorithms, 19(3-4):163–193, 2001. ISSN
1098-2418. doi: 10.1002/rsa.10019. URL http://dx.doi.org/10.1002/rsa.10019.
G. A. Miller. Note on the bias of information estimates, 1955.
Peter Grassberger. Finite sample corrections to entropy and dimension estimates. Physics
Letters A, 128(6):369–373, 1988.
Samuel Zahl. Jackknifing an index of diversity. Ecology, 58(4):907–913, 1977. ISSN 00129658,
19399170.
Thomas Schürmann and Peter Grassberger. Entropy estimation of symbol sequences. Chaos:
An Interdisciplinary Journal of Nonlinear Science, 6(3):414–427, 1996.
S. Schober. Some worst-case bounds for bayesian estimators of discrete distributions. In
2013 IEEE International Symposium on Information Theory, pages 2194–2198, July 2013.
doi: 10.1109/ISIT.2013.6620615.
D Holste, I Grosse, and H Herzel. Bayes’ estimators of generalized entropies. Journal of
Physics A: Mathematical and General, 31(11):2551, 1998.
Yanjun Han, Jiantao Jiao, and Tsachy Weissman. Does dirichlet prior smoothing solve
the shannon entropy estimation problem? CoRR, abs/1502.00327, 2015. URL http:
//arxiv.org/abs/1502.00327.
Liam Paninski. Estimating entropy on m bins given fewer than m samples. IEEE Transactions on Information Theory, 50(9):2200–2203, 2004.
13
Gregory Valiant and Paul Valiant. Estimating the unseen: an n/log(n)-sample estimator
for entropy and support size, shown optimal via new clts. In Lance Fortnow and Salil P.
Vadhan, editors, Proceedings of the 43rd ACM Symposium on Theory of Computing,
STOC 2011, San Jose, CA, USA, 6-8 June 2011, pages 685–694. ACM, 2011a. doi:
10.1145/1993636.1993727. URL http://doi.acm.org/10.1145/1993636.1993727.
Gregory Valiant and Paul Valiant. The power of linear estimators. In Rafail Ostrovsky,
editor, IEEE 52nd Annual Symposium on Foundations of Computer Science, FOCS 2011,
Palm Springs, CA, USA, October 22-25, 2011, pages 403–412. IEEE Computer Society,
2011b. doi: 10.1109/FOCS.2011.81. URL http://dx.doi.org/10.1109/FOCS.2011.81.
Y. Wu and P. Yang. Chebyshev polynomials, moment matching, and optimal estimation of
the unseen. ArXiv e-prints, April 2015.
Y. Bu, S. Zou, Y. Liang, and V. V. Veeravalli. Estimation of kl divergence between largealphabet distributions. In 2016 IEEE International Symposium on Information Theory
(ISIT), pages 1118–1122, July 2016. doi: 10.1109/ISIT.2016.7541473.
Y. Han, J. Jiao, and T. Weissman. Minimax rate-optimal estimation of kl divergence between
discrete distributions. In 2016 International Symposium on Information Theory and Its
Applications (ISITA), pages 256–260, Oct 2016.
J. Jiao, Y. Han, and T. Weissman. Minimax estimation of the l1 distance. In 2016 IEEE
International Symposium on Information Theory (ISIT), pages 750–754, July 2016. doi:
10.1109/ISIT.2016.7541399.
A-F Timan, J Berry, and J Cossar. Theory of approximation of functions of a real variable.
1965.
Penco Petrov Petrushev and Vasil Atanasov Popov. Rational approximation of real functions, volume 28. Cambridge University Press, 2011.
Zeev Ditzian and Vilmos Totik. Moduli of smoothness, volume 9. Springer Science &
Business Media, 2012.
Naum I Achieser. Theory of approximation. Courier Corporation, 2013.
Eugene Y Remez. Sur la détermination des polynômes d’approximation de degré donnée.
Comm. Soc. Math. Kharkov, 10:41–63, 1934.
A Spitzbart. A generalization of hermite’s interpolation formula. The American Mathematical Monthly, 67(1):42–46, 1960.
Alexandre B. Tsybakov. Introduction to Nonparametric Estimation. Springer series in
statistics. Springer, 2009. ISBN 978-0-387-79051-0. doi: 10.1007/b13794. URL http:
//dx.doi.org/10.1007/b13794.
Lucien M Le Cam. Asymptotic Methods in Statistical Theory. Springer-Verlag New York,
Inc., New York, NY, USA, 1986. ISBN 0-387-96307-3.
Oleg Lepski, Arkady Nemirovski, and Vladimir Spokoiny. On estimation of the lr norm of
a regression function. Probability theory and related fields, 113(2):221–253, 1999.
T Tony Cai, Mark G Low, et al. Testing composite hypotheses, hermite polynomials and
optimal estimation of a nonsmooth functional. The Annals of Statistics, 39(2):1012–1041,
2011.
14
A
Error Rate of Best Polynomial Approximation
Here, we analyze the upper bound and the lower bound of the best polynomial approximation
error EL (φ, [0, ∆]). The upper bound and the lower bound are derived as follows.
Lemma 8. Under Assumption 2, for ∆ ∈ (0, 1], we have
α
∆
EL (φ, [0, ∆]) .
.
L2
Lemma 9. Under Assumption 2, for ∆ ∈ (0, 1] there is a positive constant c such that
2 α
L
lim inf
EL (φ, [0, ∆]) > c.
L→∞
∆
Combining Lemmas 8 and 9, we can conclude E( φ, [0, ∆])
lemmas are given as follows.
∆ α
.
L2
The proofs of these
Proof of Lemma 8. Letting φ∆ (p) = φ(∆x2 ), we have EL (φ, [0, ∆]) = EL (φ∆ , [−1, 1]). We
utilize the Jackson’s inequality to upper bound the best polynomial approximation error EL
by using the modulus of continuity defined as
ω(f, δ) =
{|f (x) − f (y)| : |x − y| ≤ δ}.
sup
x,y∈[−1,1]
To derive the upper bound of EL , we divide into two cases: α ∈ (0, 1/2] and α ∈ (1/2, 1).
Case α ∈ (0, 1/2]. From the Jackson’s inequality (Achieser, 2013), there is a trigonometric
polynomial TL with degree-L such that
1
.
sup |f (x) − TL (x)| . sup
|f (x) − f (y)| : |x − y| ≤
L
x∈[0,2π]
x,y∈[0,2π]
By the definition of EL , we have
EL (f, [−1, 1]) = inf
sup |f (x) − g(x)|
g∈PL x∈[−1,1]
= inf
sup |f (cos(x)) − g(cos(x))|
g∈PL x∈[0,2π]
1
. sup
|f (cos(x)) − f (cos(y))| : |x − y| ≤
L
x,y∈[0,2π]
1
= sup
|f (x) − f (y)| : cos−1 (x) − cos−1 (y) ≤
L
x,y∈[−1,1]
1
1
≤ sup
|f (x) − f (y)| : |x − y| ≤
= ω f,
,
L
L
x,y∈[−1,1]
(7)
where we use the fact that cos−1 (x) − cos−1 (y) ≥ |x − y| for x, y ∈ [−1, 1] to derive the
last line. From Lemma 15 and the fact that pα−1 ≥ 1 for p ∈ (0, 1], we have φ(1) (p) ≤ (W +
|c1 |)pα−1 for p ∈ (0, 1]. From the absolute continuousness of φ on (0, 1], for x, y ∈ (−1, 1]
where x ≤ y we have
Z y
|φ∆ (x) − φ∆ (y)| ≤
2∆tφ(1) ∆t2 dt
x
Z y
≤2∆α (W + |c1 |)
t2α−1 dt
x
∆α (W + |c1 |) 2α
=
y − x2α
α
∆α (W + |c1 |)
≤
(y − x)2α ,
α
15
where the last line is obtained since xβ for β ∈ (0, 1] is β-Holder continuous. This is valid
for the case x = 0 since |φ∆ (0) − φ∆ (y)| = limx→0 |φ∆ (x) − φ∆ (y)|. Thus, we have
ω(φ∆ , δ) ≤
∆α (W + |c1 |) 2α
δ .
α
Substituting this into Eq (7), we have
∆α (W + |c1 |) 1
.
EL (φ∆ , [−1, 1]) .
α
L2α
∆
L2
α
.
Case α ∈ (1/2, 1). From the Jackson’s inequality (Achieser, 2013), there is a trigonometric
polynomial TL with degree-L such that
1
1
(1)
(1)
.
sup |f (x) − TL (x)| .
f (x) − f (y) : |x − y| ≤
sup
L x,y∈[0,2π]
L
x∈[0,2π]
In the similar manner of the case α ∈ (0, 1/2], we have
EL (φ∆ , [−1, 1]) = inf
sup |φ∆ (cos(x)) − g(cos(x))|
g∈PL x∈[0,2π]
1
(1) 1
.
. ω φ∆ ,
L
L
(8)
Since pα−2 ≥ 1 for p ∈ (0, 1] and Assumption 2, we have φ(2) (p) ≤ (α1 W + |c2 |)pα−2 for
p ∈ (0, 1]. From the absolute continuousness of φ(1) on (0, 1], for x, y ∈ (−1, 1] where x ≤ y
we have
Z y
(1)
(1)
φ∆ (x) − φ∆ (y) ≤
2∆φ(1) ∆t2 + 4∆2 t2 φ(2) ∆t2 dt
Zxy
≤
2∆α (W + |c1 |)t2α−2 + 4∆α (α1 W + |c2 |)t2α−2 dt
x
2(W + |c1 |) + 4(α1 W + |c2 |) 2α−1
y
− x2α−1
2α − 1
α 2(W + |c1 |) + 4(α1 W + |c2 |)
≤∆
(y − x)2α−1 .
2α − 1
=∆α
Also, we use the fact that xβ for β ∈ (0, 1] is β-Holder continuous. Thus, we have
2(W + |c1 |) + 4(α1 W + |c2 |) 2α−1
(1)
ω φ ∆ , δ ≤ ∆α
δ
.
2α − 1
Substituting this into Eq (8), we have
2(W + |c1 |) + 4(α1 W + |c2 |) 1
1
EL (φ∆ , [−1, 1]) . ∆α
.
L
2α − 1
L1−2α
∆
L2
α
.
Proof of Lemma 9. Let φ∆ (x) = φ ∆ x+1
. Then, we have EL (φ, [0, ∆]) = EL (φ∆ , [−1, 1]).
2
To derive the lower bound of EL (φ∆ , [−1, 1]), we introduce the second-order Ditzian-Totik
modulus of smoothness (Ditzian and Totik, 2012) defined as
x+y
x+y
,
ωϕ2 (f, t) = sup
f (x) + f (y) − 2f
: |x − y| ≤ 2tϕ
2
2
x,y∈[−1,1]
16
where ϕ(x) =
√
1 − x2 . Fix y = −1, for t > 0 we have
x+y
|x − y| ≤ 2tϕ
⇐⇒
2
r
(x − 1)2
⇐⇒
x + 1 ≤ 2t 1 −
4
(x − 1)2
(x + 1)2
≤1−
⇐⇒
2
4t
4
−2
2
2
t (x + 1) + (x − 1) − 4 ≤ 0 ⇐⇒
(t−2 + 1)x2 + 2(t−2 − 1)x + (t−2 + 1) − 4 ≤ 0 ⇐⇒
2
4
t−2 − 1
(t−2 − 1)2
+ 1 − −2
≤ 0 ⇐⇒
x + −2
− −2
t +1
t + 1 (t + 1)2
2
2
4
x + 1 − −2
⇐⇒
≤ −2
t +1
(t + 1)2
4
−1 ≤ x ≤ −1 + −2
.
t +1
Thus, we have
ωϕ2 (φ∆ , t)
4
x−1
: −1 ≤ x ≤ −1 + −2
≥ sup φ∆ (x) + φ∆ (−1) − 2φ∆
2
t +1
x
∆x
2
: 0 ≤ x ≤ −2
= sup φ(∆x) + φ(0) − 2φ
2
t +1
x
Application of the Taylor theorem gives
∆x
x
x
(1) ∆x
(1) ∆x
φ(∆x) + φ(0) − 2φ
= λφ
0−
x−
+ λφ
2
2
2
2
2
Z x
Z x2
∆2 φ(2) (∆t)(0 − t)dt +
∆2 φ(2) (∆t)(x − t)dt
−
x
2
0
x
2
Z
=
∆2 φ(2) (∆t)tdt +
Z
x
∆2 φ(2) (∆t)(x − t)dt.
x
2
0
Letting p0 = (α1 W/(α1 W ∨ −c02 ))1/(2−α) , φ(2) (p) ≥ α1 W pα−2 + c02 ≥ 0 for (0, p0 ]. From
continuousness of φ(2) , φ(2) (x) has same sign in x ∈ (0, p0 ]. Since t ≥ 0 for t ∈ [0, x2 ] and
x − t ≥ 0 for t ∈ [ x2 , x], we have for x ∈ (0, p0 ]
∆x
φ(∆x) + φ(0) − 2φ
2
!
!
Z x
Z
Z x
Z
2
≥∆α α1 W
0
x
tα−2 tdt +
tα−2 (x − t)dt
x
2
x
2
+ c02 ∆2
(x − t)dt
tdt +
0
x
2
α−1
xα
x
x
1 xα
c02 ∆2 x2
α−1
α
=∆ α1 W
+
−
x
+
−
x
+
α2α
1 − α 2α−1
α 2α
4
0
2−α
α
W
c
∆
1
=∆α xα W (2−α − 1) +
(21−α − 1) + 2
x2−α & ∆α xα .
α
4
α
Thus, we have for sufficiently small t
ωϕ2 (φ∆ , t)
&∆
α
2
t−2 + 1
17
α
& ∆α t2α .
(9)
PL
With the definition of ωϕ2 (f, t), we have the converse result L12 m=1 (m+1)Em (f, [−1, 1]) &
2
−1
0
ωϕ (f, L ) (Ditzian and Totik, 2012). Let L be an integer such that L0 = c` L where c` > 1.
Then, we have
EL (φ, [0, ∆])
0
≥
1
L0 − L
L
X
Em (φ, [0, ∆])
m=L+1
0
1
≥ 02
L
L
X
(m + 1)Em (φ, [0, ∆])
m=L+1
0
L
L
1 X
1
1 X
(m + 1)Em (φ, [0, ∆]) − 02 E0 (φ, [0, ∆]) − 02
(m + 1)Em (φ, [0, ∆]).
≥ 02
L m=0
L
L m=1
(10)
α
From Lemma 16, we have |φ(x) − φ(y)| ≤ W
α ∆ + |c1 |∆ for x, y ∈ [0, ∆]. Substituting it
and Eq (9) into Eq (10) and applying the converse result and Lemma 8 yields that there
are constants C > 0 and C 0 > 0 such that
α
L
C0 X
∆
|c1 |
W
α
∆
−
∆
(m
+
1)
−
L02 α
L02
L02 m=1
m2
α
L
∆
W
|c1 |
C0 X
(m + 1)
− 02 ∆α − 02 ∆ − 02
L α
L
L m=1
m2
EL (φ, [0, ∆]) ≥Cωϕ2 (φ∆ , L0−1 ) −
≥C
∆α
L02α
L
∆α
W
|c1 |
2C 0 ∆α X 1−2α
α
−
m
∆
−
∆
−
L02α
αc2` L2α
c2` L2α
L02 m=1
!
Z L
∆α
W
|c1 |
2C 0 ∆α
2−2α
α
1−2α
≥C 02α − 2 2α ∆ − 2 2α ∆ −
L
∨
x
dx
L
αc` L
c` L
L02
0
≥C
W
|c1 |
2C 0 ∆α
∆α
α
−
∆
−
∆
−
2α
c2α
αc2 L2α
c2` L2α
((2 − 2α) ∧ 1)c2` L2α
` L
α `
1
∆
W
|c1 |∆−α
2C 0
= 2α
C
−
−
−
.
c`
L2
αc2−2α
c2−2α
((2 − 2α) ∧ 1)c2−2α
`
`
`
≥C
Thus, by taking sufficiently large c` , there is c > 0 such that
2 α
L
lim sup
EL (φ, [0, ∆]) > c.
∆
L→∞
B
Proofs for Lower Bounds
To prove Theorem 3, the Le Cam’s two-point method (See, e.g., (Tsybakov, 2009)). The
consequent corollary of the Le Cam’s two-point method is as follows.
Corollary 1. For any two probability measures P, Q ∈ Mk , we have
R̃∗ (n, k; φ) ≥
1
2
(θ(P ) − θ(Q)) exp(−nDKL (P, Q)),
4
where DKL (P, Q) denotes the KL-divergence between P and Q.
We provide the proof of Theorem 3.
18
Proof of Theorem 3. For ∈ (0, 1/2). Define two probability measures on [k] as
1
1
1
P = ,
, ...,
,
2 2(k − 1)
2(k − 1)
1
1
1
Q = (1 + ),
(1 − ), ...,
(1 − ) .
2
2(k − 1)
2(k − 1)
Then, the KL-divergence between P and Q is obtained as
DKL (P, Q) = −
1
1
1
ln(1 + ) − ln(1 − ) = − ln 1 − 2 ≤ 2 .
2
2
2
Applying the Taylor theorem gives that there exist ξ1 ∈ [1/2, (1 + )/2] and ξ2 ∈ [(1 −
)/2(k − 1), 1/2(k − 1)] such that
θ(Q) − θ(P )
1
1
1
φ(2) (ξ1 ) 2
φ(2) (ξ2 ) 2
1
− φ(1)
+
+
.
= φ(1)
2
2
2
2(k − 1)
8
8(k − 1)
From the reverse triangle inequality, we have
|θ(Q) − θ(P )|
1
φ(2) (ξ1 ) 2
φ(2) (ξ2 ) 2
1
1
1
− φ(1)
+
+
≥ φ(1)
2
2(k − 1)
2
2
8
8(k − 1)
1
φ(2) (ξ1 ) 2
1
1
φ(2) (ξ2 ) 2
1
− φ(1)
−
−
.
≥ φ(1)
2
2(k − 1)
2
2
8
8(k − 1)
Combining Assumption 2, Lemma 15, and the fact that ξ1 ≥ 1/2 and ξ2 ≥ 1/4(k − 1) yields
1
≥W 21−α (k − 1)1−α + c01 ,
φ(1)
2(k − 1)
1
φ(1)
≤W 21−α + c1 ,
2
φ(2) (ξ1 ) ≤α1 W 22−α + c2 ,
φ(2) (ξ2 ) ≤α1 W 42−α (k − 1)2−α + c2 .
Consequently, we have
|θ(Q) − θ(P )| ≥ W 2−α (k − 1)1−α − 1 − α1 (2−1 + 21−α (k − 1)1−α )
− 2−1 (c1 − c01 ) − c2 (2−3 + 2−3 (k − 1)−1 )2 .
√
Set = 1/ n. Applying Corollary 1, we have
R̃∗ (n, k; φ)
≥
W 2 (k − 1)2−2α
1
α1
α1 21−α
√
√
1
−
−
−
2−2α n
(k − 1)1−α
2(k − 1)α n
n
2α−3 c2
c1 − c01
2α−3 c2
√
√
−
− 1−α
−
2
W (k − 1)1−α
W (k − 1)1−α n W (k − 1)2−α n
&
k 2−2α
.
n
From Lemma 1, this lower bound is valid for R∗ (n, k; φ).
19
!2
The proof of Theorem 4 is following the proof of (Wu and Yang, 2016). For ∈ (0, 1), define
the approximate probabilities by
(
)
k
X
k
k
Mk () = {pi }i=1 ∈ R+ :
pi ≤ 1 − .
i=1
With this definition, we define the minimax risk for Mk () as
R̃∗ (n, k, ; φ) = inf
2
E θ̂(Ñ ) − θ(P ) .
sup
(11)
θ̂ P ∈Mk ()
The minimax risk of Poisson sampling can be bounded below by Eq (11) as
Lemma 10. Under Assumption 2, for any k, n ∈ N and any < 1/3,
2
W 1−α
W2
1 ∗
k
+ |c1 | e−n/32 − 2 k 2−2α 2α − c21 2 .
R̃ (n/2, k; φ) ≥ R̃ (n, k, ; φ) − 4
3
α
α
∗
Proof of Lemma 10. This proof is following the same manner of the proof of (Wu and Yang,
2016, Lemma 1). Fix δ > 0. Let θ̂(·, n) be a near-minimax optimal estimator for fixed sample
size n, i.e.,
h
i
2
sup E (θ̂(N, n) − θ(P )) ≤ δ + R∗ (k, n; φ).
P ∈Mk
For an arbitrary approximate distribution P ∈ Mk (), we construct an estimator
θ̃(Ñ ) = θ̂(Ñ , n0 ),
where Ñi ∼ Poi(npi ) and n0 =
Lemma 17, we have
P
i
Ni . From the triangle inequality, Lemma 16 and
1
(θ̃(Ñ ) − θ(P ))2
3
1
≤
3
1
≤
3
≤
1
3
!
P
θ̃(Ñ ) − θ Pk
i=1
+ θ Pk
pi
i=1
!2
!
P
pi
− θ(P )
α
!
k
k
X
pi
W X
pi
θ̃(Ñ ) − θ Pk
+
− pi + |c1 |
− pi
Pk
Pk
α
i=1 pi
j=1 pj
j=1 pj
i=1
i=1
α
!
k
k
X
W X pi
P
+
pj − 1
θ̃(Ñ ) − θ Pk
Pk
α i=1
i=1 pi
j=1 pj j=1
P
+ |c1 |
k
X
i=1
k
X
pi
Pk
j=1
pj
!2
pj − 1
j=1
!α
!2
k
k
X
pi
W αX
pi
+ |c1 |
θ̃(Ñ ) − θ Pk
+
Pk
Pk
α
i=1 pi
j=1 pj
j=1 pj
i=1
i=1
!
!2
1
P
W 1−α α
≤
k
+ |c1 |
θ̃(Ñ ) − θ Pk
+
3
α
i=1 pi
! !2
P
W2
≤ θ̃(Ñ ) − θ Pk
+ 2 k 2−2α 2α + c21 2 .
α
i=1 pi
1
≤
3
P
!
20
!2
For the first term, we observe that Ñ ∼ Multinomial(m, PPpi ) conditioned on n0 = m.
Therefore, we have
!!2
!!2
∞
X
P
P
n0 = mP{n0 = m}
=
E θ̃(Ñ , m) − θ Pk
E θ̃(Ñ ) − θ Pk
p
p
m=0
i=1 i
i=1 i
≤
∞
X
R̃∗ (m, k; φ)P{n0 = m} + δ.
m=0
From Lemma 16 and Lemma 17, we have
R̃∗ (m, k; φ) ≤
sup
2
(θ(P ) − θ(P 0 ))
P,P 0 ∈Mk
k
k
X
W X
α
≤ sup
|pi − p0i | + |c1 |
|pi − p0i |
α i=1
P,P 0 ∈Mk
i=1
!2
k
k
X
W X α
pi + |c1 |
pi
≤4 sup
α i=1
P ∈Mk
i=1
2
W 1−α
k
+ |c1 | .
≤4
α
!2
P
Note P
that R̃∗ (m, k; φ) is a decreasing function with respect to m. Since n0 ∼ Poi(n i pi )
and | i pi − 1| ≤ ≤ 1/3, applying Chernoff bound yields P{n0 ≤ n/2} ≤ e−n/32 . Thus,
we have
!!2
P
E θ̃(Ñ ) − θ Pk
i=1 pi
2
X
W 1−α
∗
0
k
+ |c1 | P{n0 ≤ n/K} + δ
≤
R̃ (m, k; φ)P{n = m} + 4
α
m≥n/K
2
W 1−α
∗
≤R̃ (n/K, k; φ) + 4
k
+ |c1 | e−n/32 + δ.
α
The arbitrariness of δ gives the desired result.
The lower bound of R̃∗ (n, k, ; φ) is given by the following lemma.
0
0
Lemma 11. Let U and U 0 be random variables such that
√ U, U ∈ [0, λ] and E[U ] = E[U ] ≤ 1
0
and |E[θ(U ) − θ(U )]| ≥ d, where λ ≤ k. Let = 4λ/ k. Then
d2 7
64W 2 λ2α
64c21 λ2
∗
0
R̃ (n, k, ; φ) ≥
− kTV(E[Poi(nU/k)], E[Poi(nU /k)]) − 2 2α−1 2 −
.
16 8
α k
d
kd2
Proof of Lemma 11. The proof follows the same manner of the proof of (Wu and Yang,
2016, Lemma 2) expect Eq (12) below. Let β = E[U ] = E[U 0 ] ≤ 1. Define two random
vectors
0
U1
Uk
U1
U0
P =
, ...,
,1 − β ,P0 =
, ..., k , 1 − β ,
k
k
k
k
√
where Ui a nd Ui0 are independent copies of U and U 0 , respectively. Put = 4λ/ k. Define
21
the two events:
"
E=
X Ui
k
i
"
E0 =
X U0
i
k
i
#
− β ≤ , |θ(P ) − E[θ(P )]| ≤ d/4 ,
#
− β ≤ , |θ(P 0 ) − E[θ(P 0 )]| ≤ d/4 .
Applying Chebyshev’s inequality, the union bound, the triangle inequality and Lemma 16
gives
(
)
X Ui
c
PE ≤P
− β > + P{|θ(P ) − E[θ(P )]| > d/4}
k
i
P
Var[U ] 16 i Var[φ(Ui /k)]
+
≤
k2 P
d2
16 i E[(φ(Ui /k) − φ(β/k))2 ]
1
≤ +
16
d2
P
P
32 i E[W 2 (Ui − β)2α ] 32 i E[c21 (Ui − β)2 ]
1
≤ +
+
16
α2 k 2α d2
k 2 d2
2 2α
2 2
1
32W λ
32c1 λ
≤ + 2 2α−1 2 +
(12)
16 α k
d
kd2
By the same manner, we have
PE 0c ≤
32W 2 λ2α
32c21 λ2
1
+ 2 2α−1 2 +
.
16 α k
d
kd2
We define two priors on the set Mk (), the conditional distributions π = PU |E and π 0 =
PU 0 |E 0 . By the definition of events E, E 0 and triangle inequality, we obtain that under π, π 0 ,
|θ(P ) − θ(P 0 )| ≥
d
.
2
By triangle inequality, we have the total variation of observations under π, π 0 as
TV(PÑ |E , PÑ 0 |E 0 ) ≤TV(PÑ |E , PÑ ) + TV(PÑ , PÑ 0 ) + TV(PÑ 0 , PÑ 0 |E 0 )
=PE c + TV(PÑ , PÑ 0 ) + PE 0c
≤TV(PÑ , PÑ 0 ) +
1
64W 2 λ2α
64c21 λ2
+ 2 2α−1 2 +
.
8 α k
d
kd2
From the fact that total variation of product distribution can be upper bounded by the
summation of individual ones, we obtain
TV(PÑ , PÑ 0 ) ≤
k
X
TV(PÑi , PÑ 0 ) + TV(n(1 − β), n(1 − β))
i
i=1
=kTV(E[Poi(nU/k)], E[Poi(nU 0 /k)]).
Then, applying Le Cam’s lemma (Le Cam, 1986) yields that
d2 7
64W 2 λ2α
64c21 λ2
R̃∗ (n, k, ; φ) ≥
− kTV(E[Poi(nU/k)], E[Poi(nU 0 /k)]) − 2 2α−1 2 −
.
16 8
α k
d
kd2
To derive the upper bound of TV(E[Poi(nU/k)], E[Poi(nU 0 /k)]), we apply the following
lemma proved by Wu and Yang (2016).
22
Lemma 12 (Wu and Yang (2016, Lemma 3)). Let V and V 0 be random variables on [0, M ].
If E[V j ] = E[V 0j ], j = 1, ..., L and L > 2eM , then
0
TV(E[Poi(V )], E[Poi(V )]) ≤
2eM
L
L
.
Under the condition of Lemma 12, the following lemmas provides the lower bound of
d.
Lemma 13. For any given integer L > 0, there exists two probability measures ν0 and ν1
on [0, λ] such that
EX∼ν0 [X m ] = EX∼ν1 [X m ], for m = 0, ..., L,
EX∼ν0 [φ(X)] − EX∼ν1 [φ(X)] = 2EL (φ, [0, λ]).
Lemma 13. The proof is almost same as the proof of Jiao et al. (2015, Lemma 10). It follows
directly from a standard functional analysis argument proposed by Lepski et al. (1999). It
suffices to replace xα with φ(x) and [0, 1] with [0, λ] in the proof of (Cai et al., 2011, Lemma
1).
As proved Lemma 13, we can choose the probability measures of U and U 0 in Lemma 10
so that d in Lemma 10 becomes the uniform approximation error of the best polynomial
EL (φ, [0, λ]). The analysis of the lower bound on EL (φ, [0, λ]) can be found in Appendix A.
By using the lower bound (in Lemma 9), we prove Theorem 4 as follows.
Proof of Theorem 4. Set L = bC1 ln nc and λ = C2 lnnn where C1 and C2 are universal
constants such that 2eC2 ≤ C1 . Assembling Lemmas 9 and 11 to 13, we have M = C2 lnkn ,
α
where c > 0 is an universal constant. Also, we have
|E[φ(U ) − φ(U 0 )]| = d ≥ ck Lλ2
R̃∗ (n, k, ; φ)
!
bC1 ln nc
4α
4α
2eC2 ln n
64c21 bC1 ln nc λ2−2α
d2 7
64W 2 bC1 ln nc
−k
−
.
≥
−
16 8
kbC1 ln nc
c2 α2 k 2α+1
c2 k 3
If α ∈ (1/2, 1), it is sufficient to prove Theorem 4 when k & n1−1/2α ln n because of Theorem 3. Hence,
4α
64W 2 bC1 ln nc
=o(1)
c2 α2 k 2α+1
4α
64c21 bC1 ln nc λ2−2α
=o(1).
c2 k 3
(13)
(14)
4
If α ∈ (0, 1/2], we assume k & ln 3 n. Then, we get Eqs (13) and (14). Moreover, for
bC1 ln nc
2eC2 ln n
sufficiently large C1 , we get k kbC
= o(1).Thus, we have
1 ln nc
R̃∗ (n, k, ; φ) & d2 &
k2
.
(n ln n)2α
The second term in Lemma 10 is bounded above as
2
W 1−α
k2
4
k
+ |c1 | e−n/32 = o
.
α
(n ln n)2α
23
(15)
For α ∈ (0, 1), we get an upper bound on the fourth term in Lemma 10 as
c21 λ2−2α L4α 2
·d
k2
4α
c2 λ2−2α bC1 ln nc
≤ 1
· d2 = o(1) · d2 .
k2
c21 2 ≤
If α ∈ (1/2, 1), the third term in Lemma 10 is bounded above as
W 2 2−2α 2α W 2 L4α
k
≤ 2 2 3α · d2
α2
c α k
4α
W 2 bC1 ln nc
≤
· d2 = o(1) · d2 .
c2 α2 k 3α
Then, Eq (15) and Lemma 10 gives
R̃∗ (n, k; φ) &
k2
.
(n ln n)2α
4
If α ∈ (0, 1/2], we assume k ≥ c0 ln 3 n for an arbitrary constant c0 > 0, and we get
W 2 2−2α 2α W 2 C14α
k
≤ 2 2 0 3α · d2 .
α2
c α c
Hence, for sufficiently small c0 , Eq (15) and Lemma 10 yields
R̃∗ (n, k; φ) &
C
k2
.
(n ln n)2α
Proofs for Upper Bounds
We use the following helper lemma for proving Lemma 3.
Lemma 14 (Cai et al. (2011), Lemma 4). Suppose 1E is an indicator random variable
independent of X and Y , then
2
Var[X1E + Y 1E c ] = Var[X]PE + Var[Y ]PE c + (E[X] − E[Y ]) PEPE c .
Proof of Lemma 2. From the property of the absolute value, the bias is bounded above as
i
h
Bias θ̂(Ñ ) − θ(P )
≤
k
X
h
i
h
i
Bias 1Ñ 0 ≥2∆n,k φplugin (Ñi ) − φ(pi ) + Bias 1Ñ 0 <2∆n,k φpoly (Ñi ) − φ(pi )
.
i
i
i=1
Because of the independence between Ñ and Ñ 0 , we have
h
i
h
i n
o
Bias 1Ñ 0 ≥2∆n,k φplugin (Ñi ) − φ(pi ) =Bias φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k
i
h
i
h
i n
o
Bias 1Ñ 0 <2∆n,k φpoly (Ñi ) − φ(pi ) =Bias φpoly (Ñi ) − φ(pi ) P Ñi0 < 2∆n,k
i
24
For p ∈ [
∆n,k ∆n,k
2n , n ],
from Lemmas 15 and 16, we have
∆n,k ∆n,k
H4 p; φ,
,
− φ(pi )
n
2n
m 4−m
4 φ(m) ∆n,k
X
X 4+1
n
∆n,k
≤
p−
B`,4+`+1
m!
n
4+`+1
m=1
`=0
∆n,k
n
∆
− nn,k
p−
∆n,k
2n
+ φ
≤
4
φ(m)
X
∆n,k
n
m!
m=1
∆n,k
2n
m 4−m
`
X 4 + `
4+1
`
4+`+1
4+`+1
`
!
∆n,k
− φ(pi )
n
4+1
`=0
+
≤
4
φ(m)
X
∆n,k
n
m!
m=1
≤5
4
X
m=1
αm−1 W
m!
∆n,k
2n
α
∆n,k
n
W
+ |c1 |
α
m
W
+ |c1 |
α
m
∆n,k
W
−m
2
+ cm
+
+ |c1 |,
2n
α
(5 − m) +
where we use 0 ≤ Bν,n (x) ≤ Bν,n (ν/n) to get the third line. From the assumption ∆n,k ≤ n,
we have
4
X
αm−1 W
W
∆n,k ∆n,k
,
− φ(p) ≤ 5
+ cm +
+ |c1 |.
H4 p; φ,
m
n
2n
m!2
α
m=1
Also, for p ∈ [1, 2], we have
|H4 (p; φ, 1, 2) − φ(pi )|
≤
4−m
4
X 4+1
X
φ(m) (1)
p−1
(p − 1)m
B`,4+`+1
+ |φ(1) − φ(pi )|
m!
4+`+1
2−1
m=1
`=0
≤5
4
X
φ
m=1
≤5
4
X
(m)
(1)
W
+
+ |c1 |
m!
α
(αm−1 W + cm ) +
m=1
For p ∈ (
∆n,k
n , 1),
W
+ |c1 |.
α
we have by Lemma 16 that
|φ(p) − φ(pi )| ≤
W
+ |c1 |.
α
Consequently, we have for p ≥ 0
φ̄ ∆n,k (p) − φ(pi ) ≤ 5
n
4
X
(αm−1 W + cm ) +
m=1
25
W
+ |c1 | . 1.
α
(16)
For p ∈ (
∆n,k ∆n,k
2n , n ),
∆n,k ∆n,k
p
(2)
H
p; φ,
,
2n 4
n
2n
(m−i)∨0
4
2
p X (m) ∆n,k X 2
∆n,k
1
=
φ
p−
2n m=1
n
n
i ((m − i) ∨ 0)!
i=0
4−m
X
∆
p − nn,k
4+1
(2−i)
B`,4+`+1
∆n,k
4+`+1
− 2n
`=0
(m−i)∨0
=
!
2
4
p X (m) ∆n,k X 2
∆n,k
1
φ
p−
2n m=1
n
n
i ((m − i) ∨ 0)!
i=0
!
(2−i)∧`
4−m
∆
X
X
p − nn,k
(4 + 1)(4 + ` + 1)!
j 2−i
,
(−1)
B`−j,4+`−1+i
∆n,k
(4 + ` + 1)(4 + ` − 1 + i)! j=0
j
− 2n
`=0
where the last line is obtained by using the fact B(1)
ν,n (x) = n(Bν−1,n−1 (x) − Bν,n−1 (x)).
Again, the fact 0 ≤ Bν,n (x) ≤ Bν,n (ν/n) gives
p
∆n,k ∆n,k
(2)
H4 p; φ,
,
2n
n
2n
(m−i)∨0
4
2
p X (m) ∆n,k X 2
1
∆n,k
≤
φ
i ((m − i) ∨ 0)! 2n
2n m=1
n
i=0
(2−i)∧`
4−m
X X
2−i
(4 + 1)(4 + `)!
(` − j)`−j (4 − 1 + i + j)4−1+i+j
j
(` − j)!(4 − 1 + i + j)!
(4 + ` − 1 + i)4+`−1+i
j=0
`=0
4
X
(m−2)∨0
5−m
∆n,k
((m − 2) ∨ 0)! 2n
m−1
m !
20(5 − m) ∆n,k
20(4 + (4 − m)(5 − m)) ∆n,k
+
+
(m − 1)!
2n
2m!
2n
!
4
α−m
(m−1)∨1
1 X
∆n,k
5−m
∆n,k
≤
αm−1 W
+ cm
n m=1
n
((m − 2) ∨ 0)! 2n
m
m+1 !
20(4 + (4 − m)(5 − m)) ∆n,k
20(5 − m) ∆n,k
+
.
+
(m − 1)!
2n
2m!
2n
≤
∆n,k
p
φ(m)
2n m=1
n
From the assumption ∆n,k ≤ n, we have
p
∆n,k ∆n,k
(2)
H4 p; φ,
,
2n
n
2n
!
α−1
4
1 X
∆n,k
≤
αm−1 W
+ cm
n m=1
n
(5 − m)
20(5 − m)
20(4 + (4 − m)(5 − m))
+
+
.
2m−1 ((m − 2) ∨ 0)! 2m (m − 1)!
2m+2 m!
26
From the assumption, there is a universal constant c > 0 such that ∆n,k ≥ c. Thus, we have
p
∆n,k ∆n,k
(2)
H4 p; φ,
,
2n
n
2n
4
X
cm
cα−1
≤
αm−1 W α +
n
n
m=1
(5 − m)
20(5 − m)
20(4 + (4 − m)(5 − m))
+
+
.
2m−1 ((m − 2) ∨ 0)! 2m (m − 1)!
2m+2 m!
Also, for p ∈ (1, 2), we have
p
(2)
H (p; φ, 1, 2)
2n 4
!
4
1 X
5−m
20(5 − m) 20(4 + (4 − m)(5 − m))
≤
(αm−1 W + cm )
.
+
+
n m=1
((m − 2) ∨ 0)!
(m − 1)!
2m!
Thus, we have for p ≥ 0
p (2)
φ̄ ∆ (p)
2n n,k
n
4
X
cα−1
|cm |
αm−1 W α +
≤
n
n
m=1
1∨
.
(5 − m)
20(5 − m) 20(4 + (4 − m)(5 − m))
+
+
((m − 2) ∨ 0)!
(m − 1)!
2m!
1
.
nα
.
(17)
Combining Eqs (16) and (17) yields for any pi ∈ [0, 1]
h
i
Bias φplugin (Ñi ) − φ(pi ) . 1.
Then, we have
h
i n
o
Bias φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k
h
i n
o
= Bias φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k 1npi ≤∆n,k
h
i n
o
+ Bias φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k 1npi >∆n,k
n
o
h
i
.P Ñi0 ≥ 2∆n,k 1npi ≤∆n,k + Bias φplugin (Ñi ) − φ(pi ) 1npi >∆n,k .
n
o
The Chernoff bound for the Poisson distribution gives P Ñi0 ≥ 2∆n,k 1npi ≤∆n,k ≤
(e/4)∆n,k . Thus, we have
h
i n
o
Bias φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k
h
i
.(e/4)∆n,k + Bias φplugin (Ñi ) − φ(pi ) 1npi >∆n,k .
Similarly, we have by the final truncation of φpoly and Lemma 16 that
h
i
W
Bias φpoly (Ñi ) − φ(pi ) ≤ sup |φ(p) − φ(pi )| ≤
+ |c1 |.
α
p∈[0,1]
27
(18)
n
o
The Chernoff bound yields P Ñi0 < 2∆n,k ≤ e−∆n,k /8 for pi > 4∆n,k . Thus, we have
h
i n
o
Bias φpoly (Ñi ) − φ(pi ) P Ñi0 < 2∆n,k
h
i n
o
≤ Bias φpoly (Ñi ) − φ(pi ) P Ñi0 < 2∆n,k 1npi ≤4∆n,k
h
i n
o
+ Bias φpoly (Ñi ) − φ(pi ) P Ñi0 < 2∆n,k 1npi >4∆n,k
h
i
W
+ |c1 | e−∆n,k /8 .
≤Bias φpoly (Ñi ) − φ(pi ) 1npi ≤4∆n,k +
α
(19)
Combining Eqs (18) and (19) gives the desired result.
Proof of Lemma 3. Because of the independence of Ñ1 , .., Ñk , Ñ10 , ..., Ñk0 ,
Lemma 14 gives
h
i
Var θ̂(Ñ ) − θ(P )
#
" k
X
≤Var
1Ñ 0 ≥2∆n,k φplugin (Ñi ) − φ(pi ) + 1Ñ 0 <2∆n,k φpoly (Ñi ) − φ(pi )
applying
i
i
i=1
≤
k
X
i
h
Var 1Ñ 0 ≥2∆n,k φplugin (Ñi ) − φ(pi ) + 1Ñ 0 <2∆n,k φpoly (Ñi ) − φ(pi )
i
i
i=1
≤
k
X
i=1
h
i n
o
Var φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k
h
i n
o
+ Var φpoly (Ñi ) − φ(pi ) P Ñi0 < 2∆n,k
h
i
h
i2
+ E φplugin (Ñi ) − φ(pi ) − E φpoly (Ñi ) − φ(pi )
(20)
!
n
o n
o
0
0
P Ñi ≥ 2∆n,k P Ñi < 2∆n,k
.
We can derive upper bounds on the first two terms of Eq (20) in the same manner of Eqs (18)
and (19) as
h
i n
o
Var φplugin (Ñi ) − φ(pi ) P Ñi0 ≥ 2∆n,k
h
i
. (e/4)∆n,k + Var φplugin (Ñi ) − φ(pi ) 1npi >∆n,k ,
and
h
i n
o
Var φpoly (Ñi ) − φ(pi ) P Ñi0 < 2∆n,k
h
i
. Var φpoly (Ñi ) − φ(pi ) 1npi ≤4∆n,k + e−∆n,k /8 .
By the Chernoff bound, we have
n
o n
o
P Ñi0 ≥ 2∆n,k P Ñi0 < 2∆n,k
n
o n
o
=(1pi <∆n,k + 1pi >4∆n,k + 1∆n,k ≤pi ≤4∆n,k )P Ñi0 ≥ 2∆n,k P Ñi0 < 2∆n,k
≤(e/4)∆n,k + e−∆n,k /8 + 1∆n,k ≤pi ≤4∆n,k .
28
Thus, we have the upper bound of the last term of Eq (20) as
h
i
h
i2 n
o n
o
E φplugin (Ñi ) − φ(pi ) − E φpoly (Ñi ) − φ(pi )
P Ñi0 ≥ 2∆n,k P Ñi0 < 2∆n,k
h
i
h
i2
≤ Bias φplugin (Ñi ) − φ(pi ) + Bias φpoly (Ñi ) − φ(pi )
(e/4)∆n,k + e−∆n,k /8 + 1∆n,k ≤pi ≤4∆n,k
. (e/4)∆n,k + e−∆n,k /8
h
i
h
i2
+ Bias φplugin (Ñi ) − φ(pi ) + Bias φpoly (Ñi ) − φ(pi )
1∆n,k ≤pi ≤4∆n,k .
Next, we prove the upper bounds on the bias and the variance of the best polynomial
estimator as follows:
Proof of Lemma 4. Let φ0sup,∆ = φsup,∆ ∨ supp∈[0,∆] φL (p) and φ0inf,∆ = φinf,∆ ∧
inf p∈[0,∆] φL (p). By the triangle inequality and the fact that gL is an unbiased estimator of
φL , we have
h
i
Bias (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − φ(p)
h
i
≤ Bias (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − (gL (Ñ ) ∧ φ0sup,∆ ) ∨ φ0inf,∆
h
i
h
i
+ Bias (gL (Ñ ) ∧ φ0sup,∆ ) ∨ φ0inf,∆ − φL (p) + Bias gL (Ñ ) − φ(p) .
By Chebyshev alternating theorem (Petrushev and Popov, 2011), the first term is bounded
above as
h
i
Bias (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − (gL (Ñ ) ∧ φ0sup,∆ ) ∨ φ0inf,∆
≤(φ0sup,∆ − φsup,∆ ) ∨ (φinf,∆ − φ0inf,∆ ) ≤ EL (φ, [0, ∆]).
Also, the third term is bounded above as
h
i
Bias gL (Ñ ) − φ(p) =|φL (p) − φ(p)| ≤ EL (φ, [0, ∆]).
The error bound of
α EL (φ, [0, ∆]) is derived in Appendix A. From Lemma 8, we have
EL (φ, [0, ∆]) . L∆2 . The second term has upper bound as
h
Bias (gL (Ñ ) ∧
φ0sup,∆ )
∨
φ0inf,∆
i r h
i2
− φL (p) =
E (gL (Ñ ) ∧ φ0sup,∆ ) ∨ φ0inf,∆ − φL (p)
s
2
.
≤ E (gL (Ñ ) ∧ φ0sup,∆ ) ∨ φ0inf,∆ − φL (p)
2
Since φL (p) ∈ [φ0inf,∆ , φ0sup,∆ ] for p ∈ [0, ∆], we have (gL (Ñ ) ∧ φ0sup,∆ ) ∨ φ0inf,∆ − φL (p) ≤
2
gL (Ñ ) − φL (p) . Thus, we have
h
i r
h
i
0
0
Bias (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − φL (p) ≤ Var gL (Ñ ) − φL (p) .
29
Proof of Lemma 5. It is obviously that truncation does not increase the variance, i.e.,
h
i
h
i
Var (gL (Ñ ) ∧ φsup,∆ ) ∨ φinf,∆ − φ(p) ≤ Var gL (Ñ ) − φ(p) .
Letting φ∆ (p) = φ(∆x) and a0 , ..., aL be coefficients of the optimal uniform approximation
PL
m
of φ∆ by degree-L polynomials on [0, 1], we have m=0 ∆nmam (Ñ )m = gL (Ñ ). Then, since
the standard deviation of sum of random variables is at most the sum of individual standard
deviation, we have
h
i
Var gL (Ñ ) − φ(p) ≤
q
L
X
∆m |am |
Var(Ñ )m
nm
m=1
!2
.
From (Petrushev and Popov, 2011) and the fact from Lemma 16 that φ is bounded, there
is a positive constant C such that |am | ≤ C23L . From (Wu and Yang, 2016), Var(Ñ )m is
decreasing monotonously as m increases, and for X ∼ Poi(λ)
!
√
2 λm
(2e)
√
∨1 .
Var(X)m ≤ (λm)m
π λm
By the assumption of p ≤ ∆ and monotonous, Var(Ñ )m ≤ Var(X)m where X ∼ Poi(∆n).
Thus, we have
q
L
h
i
X
√
∆m 23L
Var gL (Ñ ) .
(∆nL)m (2e)2 ∆nL
m
n
m=1
!2
r
L
√
X
∆3m Lm 3L
∆nL
≤
2 (2e)
.
nm
m=1
From the assumption
∆3 L
n
!2
≤ 12 , we have
!2
√
∆3m Lm 3L
∆nL
2 (2e)
c
nm
m=1
!m !2
r
L
√
3L
X
∆
≤ 23L (2e) ∆nL
n
m=1
r
Z L r 3 !x !!2
√
∆3 L
∆ L
3L
∆nL
≤ 2 (2e)
+
dx
n
n
1
r
r
2
!L r
√
3L
3L
3L
∆
2
∆
∆
≤23L (2e) ∆nL
+
−
3
n
n
n
ln ∆nL
r
!L−1 2
r
√
3L
3L
∆
2
∆
1 −
=
23L (2e) ∆nL 1 +
n
ln 2
n
L
X
r
m
√
16∆3 L64L (2e)2 ∆nL
≤
n √
∆3 L64L (2e)2 ∆nL
.
.
n
30
The proofs of the upper bounds on the bias and the variance of the bias-corrected plugin
estimator are obtained as follows.
Proof of Lemma 6. Applying Taylor theorem yields
#
!
"
Ñ
− φ(p)
Bias φ̄∆
n
"
!2
!
Ñ − np φ(2) (p) Ñ
Ñ (2) Ñ
(1)
= E φ (p)
+
−p −
φ̄
n
2
n
2n ∆ n
φ(3) (p)
+
6
"
Ñ
−p
n
!3
+ R3
"
!#
!#
(3)
p
φ
(p)
Ñ
1
Ñ
Ñ
(2)
E pφ(2) (p) − φ̄∆
+
,
≤
+ E R3
; φ̄∆ , p
2n
n
n
6n2
n
!#
Ñ
; φ̄∆ , p
n
(21)
where we use the fact that for X ∼ Poi(λ), E[(X − λ)2 ] = λ, E[(X − λ)3 ] = λ, and
R3 (x; φ̄∆ , p) denotes the reminder term of the Taylor theorem. The first term of Eq (21) is
bounded above as
"
!#
1
Ñ (2) Ñ
(2)
E pφ (p) − φ̄∆
2n
n
n
"
!
!!#
1
Ñ
Ñ
(2) Ñ
(2)
(2)
=
E φ (p) p −
+
φ (p) − φ̄∆
2n
n
n
n
"
!
!#
Ñ φ(3) (p) Ñ
Ñ
Ñ (2)
1
=
E
− p + R1
; φ̄ , p
2n
n
n
n
n ∆
!#
"
p φ(3) (p)
Ñ
Ñ (2)
,
(22)
; φ̄ , p
+ E
R1
≤
2n2
2n2
n ∆
where the last line is obtained by using the fact that for X ∼ Poi(λ), E[X(X − λ)] = λ,
(2)
and R1 (x; φ̄∆ , p) denotes the reminder term of the Taylor theorem. From Lemma 15, the
second term of Eq (21) and the first term of Eq (22) are bounded above as
p φ(3) (p)
α2 W pα−2 + c3 p
1
p
≤
. 2 2−α + 2
(23)
2
6n
6n2
n ∆
n
p φ(3) (p)
α2 W pα−2 + c3 p
1
p
≤
. 2 2−α + 2 .
(24)
2
2
2n
2n
n ∆
n
h
h
i
i
(2)
Ñ
Ñ
The rest is to derive the upper bound on E R3 Ñ
and E 2n
.
2 R1
n ; φ̄∆ , p
n ; φ̄∆ , p
Let p̂ = Ñ
n . From the mean value theorem, letting a function G(x) be continuous on the
closed interval and differentiable with non-vanishing derivative on the open interval between
p and p̂, there exists ξ between p and p̂ such that
φ̄(4) (ξ)
G(p̂) − G(p)
R3 p̂; φ̄∆ , p = ∆
(p̂ − ξ)3
.
6
G(1) (ξ)
Define G(x) =
1
x2 (p̂
− x)4 . Then, there exists ξ such that
(4)
φ̄ (ξ)
ξ 3 (p̂ − p)4
R3 p̂; φ̄∆ , p = − ∆
(p̂ − ξ)3 2
12
p (ξ + p̂)(p̂ − ξ)3
(4)
=−
ξ 3 φ̄∆ (ξ)
(p̂ − p)4
12p2 (ξ + p̂)
31
(25)
Thus, we have
(4)
ξ 3 φ̄∆ (ξ)
E R3 p̂; φ̄∆ , p ≤E
(p̂ − p)4
12p2 (ξ + p̂)
i
h
1
4
2 (4)
(p̂
−
p)
≤
E
ξ
φ̄
(ξ)
∆
12p2
(4)
≤
supξ∈R+ ξ 2 φ̄∆ (ξ)
12p2
1
1
+
≤
4n2
12pn3
E (p̂ − p)4
(4)
sup ξ 2 φ̄∆ (ξ) ,
ξ∈R+
where we use the fact that for X ∼ Poi(λ), E[X 4 ] = 3λ2 + λ. For ξ ∈ ( ∆
2 , ∆), we have
∆
(4)
H
ξ; φ, , ∆
2
4
4
4−m
X
X
X 5
4
1
(4−i) ξ − ∆
=
φ(m) (∆)
(ξ − ∆)(m−i)∨0
B`,5+`
i ((m − i) ∨ 0)!
5+`
∆/2
m=1
i=0
`=0
4
4
X
X
4
1
=
φ(m) (∆)
(ξ − ∆)(m−i)∨0
i
((m
−
i)
∨
0)!
m=1
i=0
(4−i)∧`
4−m
X
X
4−i
5(5 + `)!
ξ−∆
,
(−1)j
B`−j,1+`+i
j
(5 + `)(1 + ` + i) j=0
∆/2
`=0
where we use B(1)
ν,n (x) = n(Bν−1,n−1 (x) − Bν,n−1 (x)). Since 0 ≤ Bν,n (x) ≤ Bν,n (ν/n) ≤ 1,
there is a universal constant c > 0 such that for any i = 0, ..., 4
4−m
X
`=0
(4−i)∧`
X
5(5 + `)!
4−i
ξ−∆
(−1)j
B`−j,1+`+i
≤ c.
(5 + `)(1 + ` + i) j=0
j
∆/2
Thus, we have from Lemma 15 that
∆
2
(4)
ξ H
ξ; φ, , ∆
2
4
4
X
X
4
c
(m−i)∨0
ξ 2 (ξ − ∆)
≤
φ(m) (∆)
i
((m
−
i)
∨
0)!
m=1
i=0
4
4
X
X
4
c
≤
αm−1 W ∆α−m + cm
∆(2+m−i)∨2
i
((m
−
i)
∨
0)!
m=1
i=0
4 X
4
X
4
c
=
αm−2 W ∆(2+α−i)∨(2+α−m) + cm ∆(2+m−i)∨2
i ((m − i) ∨ 0)!
m=1 i=0
.∆α−2 .
32
Similarly, for ξ ∈ (1, 2)
ξ 2 H (4) (ξ; φ, 2, 1)
4
X
4
X
4
c
(m−i)∨0
ξ 2 (ξ − 1)
i
((m
−
i)
∨
0)!
m=1
i=0
4
4
X
X
4
4c
≤
(αm−1 W + cm )
i
((m
−
i) ∨ 0)!
m=1
i=0
≤
φ
(m)
(1)
.1.
For ξ ∈ [∆, 1], we have from Lemma 15 that
ξ 2 φ(4) (ξ) ≤ α1 W ξ α−2 + c4 ξ 2 . ∆α−2 .
Since φ̄∆ (ξ) = 0 for ξ ∈ [0, ∆/2] and ξ ≥ 2 by the construction, we have
(4)
sup ξ 2 φ̄∆ (ξ) . ∆α−2 .
(26)
ξ∈R+
Thus, we have
E R3 p̂; φ̄∆ , p .
1
1
+ 3 3−α .
n2 ∆2−α
n ∆
(27)
Define G(x) = 12 ( xp̂ − 1)2 . Then, the mean value theorem stats that there exists ξ such that
(2)
R1 (p̂; φ̄∆ , p) =
(4)
ξ 2 ( pp̂ − 1)2
φ̄∆ (ξ)
(p̂ − ξ)
2
p̂( p̂ξ − 1)
(4)
=
φ̄∆ (ξ) ξ 3 (p̂ − p)2
.
2
p2 p̂
Thus, we have
(4)
φ̄
(ξ)
3
2
∆
p̂
ξ (p̂ − p)
(2)
E
R1 (p̂; φ̄∆ , p) ≤E
2n
4n
p2
(4)
≤
=
supξ∈R+ ξ 3 φ̄∆ (ξ)
4np2
E (p̂ − p)2
1
(4)
sup ξ 3 φ̄∆ (ξ) .
4n2 p ξ∈R+
In the similar manner of Eq (26), we have
(4)
sup ξ 3 φ̄∆ (ξ) . ∆α−1 .
ξ∈R+
Thus, we have
1
1
p̂
(2)
R1 (p̂; φ̄∆ , p) . 2 1−α ≤ 2 2−α .
E
2n
n p∆
n ∆
By the assumption ∆ & n1 , we have
and (28) gives the desired result.
1
n3 ∆3−α
.
33
1
n2 ∆2−α .
(28)
Assembling Eqs (23), (24), (27)
Proof of Lemma 7. From the property of the variance and the triangle inequality, we have
"
!
!
#
Ñ (2) Ñ
pφ(2) (p)
Ñ
Var φ̄∆
− 2 φ̄∆
− φ(p) +
n
2n
n
2n
!
!
!2
(2)
Ñ
Ñ
pφ
(p)
Ñ
(2)
≤E φ̄∆
− 2 φ̄∆
− φ(p) +
n
2n
n
2n
!
!2
!
!2
(2)
Ñ
Ñ
pφ
(p)
Ñ
(2)
.
φ̄
(29)
≤2E φ̄∆
− φ(p) + 2E
−
n
2n2 ∆ n
2n
Applying Taylor theorem to the first term of Eq (29) gives
!
Ñ
φ̄∆
− φ(p)
n
!
!2
!3
!
Ñ
φ(2) (p) Ñ
φ(3) (p) Ñ
Ñ
(1)
= φ (p)
−p +
−p +
− p + R3
; φ̄∆ , p ,
n
2
n
6
n
n
where R3 Ñ
;
φ̄
,
p
denotes the reminder term of the Taylor theorem. From the triangle
∆
n
inequality, we have
φ̄∆
Ñ
n
!2
!
− φ(p)
!2
!4
2 Ñ
2 Ñ
= 4 φ(1) (p)
− p + φ(2) (p)
−p
n
n
!6
!!2
2
φ(3) (p)
Ñ
Ñ
+
− p + 4 R3
; φ̄∆ , p
. (30)
9
n
n
The central moments for X ∼ Poi(λ) are given as E[(X − λ)2 ] = λ,E[(X − λ)4 ] = 3λ2 + λ,
and E[(X − λ)6 ] = 15λ3 + 25λ2 + λ. Lemma 15, the triangle inequality and the assumption
1
n & ∆, the expectation of the first three terms in Eq (30) have upper bounds as
!2
2 Ñ
8W 2 p2α−1 + 8c21 p
p2α−1
p
E4 φ(1) (p)
−p ≤
.
+ ,
n
n
n
n
!4
2 Ñ
3p2
p
E φ(2) (p)
− p ≤ 2α12 W 2 p2α−4 + c22
+
n
n2
n3
p2α−1
p2α−1
p
+
+ 2
2
3
2
n ∆
n ∆
n
p2α−1
p
.
+ 2,
n
n
.
and
φ(3) (p)
E
9
2
Ñ
−p
n
!6
≤ 2α22 W 2 p2α−6 + c23
15p3
25p2
p
+
+
n3
n4
n5
p2α−1
p2α−1
p
p2α−1
+
+
+ 3
n3 ∆2
n4 ∆ 3
n 5 ∆4
n
p2α−1
p
.
+ 3.
n
n
.
34
From Eq (25), there exists ξ between p and p̂ such that
!!2
Ñ
4E R3
; φ̄∆ , p
n
!2
3 (4)
ξ
φ̄
(ξ)
∆
=4E
(p̂ − p)4
12p2 (ξ + p̂)
(4)
supξ∈R+ ξ 2 φ̄∆ (ξ)
≤
36p4
≤
2
E (p̂ − p)8
490
1
105
119
+
+
+
36n4
36n5 p 36n6 p2
36n7 p
(4)
2
sup ξ 2 φ̄∆ (ξ) ,
ξ∈R+
where we use E[(X − λ)8 ] = 105λ4 + 490λ3 + 119λ2 + λ for X ∼ Poi(λ).
supξ∈R+
(4)
ξ 2 φ̄∆ (ξ)
2
.∆
2α−4
from Eq (26) and ∆ &
4E R3
1
n
Since
by the assumption, we have
!!2
Ñ
; φ̄∆ , p
n
1
1
1
1
+ 5 5−2α + 6 6−2α + 7 7−2α
n4 ∆4−2α
n ∆
n ∆
n ∆
1
. 4 4−2α .
n ∆
.
(2)
Letting g(p) = pφ̄∆ (p), application of the Taylor theorem to the second term of Eq (29)
yields
!
!
!
Ñ (2) Ñ
Ñ
Ñ
pφ(2) (p)
1
−
≤
(φ(2)(p) + pφ(3)(p))
− p + R1
; g, p .
φ̄
2n2 ∆ n
2n
2n
n
n
The triangle inequality and E[(X − λ)2 ] = λ for X ∼ Poi(λ) give
!
!2
(2)
Ñ
pφ
(p)
Ñ
(2)
E
−
φ̄
2n2 ∆ n
2n
!2
!!2
2
2
(φ(2)(p)) + (pφ(3)(p)) Ñ
1
Ñ
≤
E
− p + 2 E R1
; g, p
n2
n
2n
n
!!2
2
2
p(φ(2)(p)) + p(pφ(3)(p))
1
Ñ
.
=
+ 2 E R1
; g, p
n3
2n
n
Applying Lemma 15 gives
p(φ(2)(p))2 + p(pφ(3)(p))2
n3
1
≤ 3 2α12 W 2 p2α−3 + 2pc22 + 2α12 W 2 p2α−3 + 2p3 c23
n
p2α−1
p
. 3 2 + 3
n ∆
n
p2α−1
p
.
+ 3.
n
n
35
1
2
Let p̂ = Ñ
n and G(x) = x (p̂ − x) . Then, the mean value theorem gives that there exists ξ
between p and p̂ such that
"
2 #
h
i
G(p̂) − G(p)
2
(1)
E (R1 (p̂; g, p)) =E g (ξ)(p̂ − ξ)
G(1) (ξ)
"
2 #
ξ 2 (p̂ − p)2
(1)
=E g (ξ)
p(p̂ + ξ)
2
1
3
sup ξg (1) (ξ)
≤ 2+ 3
n
n p ξ∈R+
2
3
1
(3)
(4)
≤ 2+ 3
sup 2ξ φ̄∆ (ξ) + ξ 2 φ̄∆ (ξ)
n
n p ξ∈R+
!
2
2
1
3
(3)
2 (4)
.
≤ 2+ 3
2 sup ξ φ̄∆ (ξ) + 2 sup ξ φ̄∆ (ξ)
n
n p
ξ∈R+
ξ∈R+
In the similar manner of Eq (26), we have
(3)
sup ξ φ̄∆ (ξ)
ξ∈R+
2
(4)
. ∆2α−4 , and sup ξ 2 φ̄∆ (ξ)
2
. ∆2α−4 .
ξ∈R+
Thus, we have
!!2
1
Ñ
1
1
1
.
E R1
; g, p
+ 5 5−2α . 4 4−2α .
2
4
4−2α
2n
n
n ∆
n ∆
n ∆
Consequently, we get the bound of the variance as
p2α−1
1
p
+ 4 4−2α + .
n
n ∆
n
D
Proof of Proposition 1
Proof of Proposition 1. It is obviously that if the output domain of φ is unbounded, i.e.,
there is a point
p0 ∈ [0,
1] such that |φ(p)| → ∞ as p → p0 , there is no consistent estimator.
Letting p0 =
W
W ∨−c01
, φ(1) (p) has same sign in (0, p0 ]. Thus, for any p ∈ (0, p0 ], we have
Z
p
φ(1) (x)dx
|φ(p) − φ(p0 )| =
p
Z p00
=
φ(1) (x) dx
p
Z
p0
≥W
p−1 dx + c01 (p0 − p)
p
≥W ln(p0 /p) + c01 (p0 − p).
Since |φ(p) − φ(p0 )| → ∞ as p → 0, φ is unbounded and we gets the claim.
36
E
Additional Lemmas
Here, we introduce some additional lemmas and their proofs.
Lemma 15. For a non-integer α, let φ be a m times continuously differentiable function
on (0, 1] where m ≥ 1 + α. Suppose that there exist finite constants W > 0, cm and c0m such
that
φ(m) (p) ≤ αm−1 W pα−m + cm , and φ(m) (p) ≥ αm−1 W pα−m + c0m .
Then, there exists finite constants cm−1 and c0m−1 such that
φ(m−1) (p) ≤ αm−2 W pα−m+1 + cm−1 , and φ(m−1) (p) ≥ αm−2 W pα−m+1 + c0m−1 ,
Qi
where α0 = 1 and αi =
j=1 (j
Proof of Lemma 15. Let pm =
− α) for i = 1, ..., m.
αm−1 W
αm−1 W ∨−c0m
(m)
1/(m−α)
. Then, φ(m) (p) > 0 for p ∈ (0, pm ).
From continuousness of φ(m) , φ (p) has same sign in p ∈ (0, pm ], and thus we have either
φ(m) (p) ≥ αm−1 W pα−m +c0m or φ(m) (p) ≤ −αm−1 W pα−m −c0m in p ∈ (0, pm ]. Since φ(m−1)
is absolutely continuous on (0, 1], we have for any p ∈ (0, 1]
Z p
φ(m−1) (p) = φ(m−1) (pm ) +
φ(m) (x)dx.
pm
The absolute value of the second term has an upper bound as
Z p
Z p
αm−1 W xα−m + cm dx
φ(m) (x)dx ≤
pm
pm
α−m+1
≤ αm−2 W pm
− pα−m+1 + cm (p − pm )
≤αm−2 W pα−m+1 + αm−2 W pα−m+1
+ cm (pm − p)
m
≤αm−2 W pα−m+1 + αm−2 W pα−m+1
+ |cm |.
m
Also, we have a lower bound of the second term as
Z p∧pm
Z p
Z p
φ(m) (x)dx +
φ(m) (x)dx
φ(m) (x)dx =
pm
pm
p∧pm
p∧pm
Z p
α−m
αm−1 W pm
+ cm dx
αm−1 W xα−m + c0m dx −
p∧pm
pm
α−m+1
≥ αm−2 W pα−m+1
− (p ∧ pm )
+ c0m ((p ∧ pm ) − pm )
m
α−m
− αm−1 W pm
+ cm (p − (p ∧ pm ))
Z
≥
≥ αm−2 W (p ∧ pm )
α−m+1
− αm−2 W pα−m+1
− |c0m (pm− (p ∧ pm ))|
m
α−m
− αm−1 W pm + cm (p − (p ∧ pm ))
α−m+1
≥ αm−2 W pα−m+1 − αm−2 W pm
−|c0m |pm .
α−m
− αm−1 W pm + cm (1 − pm )
Applying the triangle inequality and the reverse triangle inequality gives
Z p
Z p
φ(m) (x)dx − φ(m−1) (pm ) ≤ φ(m−1) (p) ≤
φ(m) (x)dx + φ(m−1) (pm ) .
pm
pm
α−m+1
Thus, setting cm−1 = αm−2 W pα−m+1
+|cm |+ φ(m−1) (pm ) and c0m−1 = −αm−2 W pm
−
m
0
α−m
(m−1)
|cm |pm − (αm−1 W pm + cm )(1 − pm ) − φ
(pm ) yields the claim.
37
Lemma 16. Under Assumption 1 or Assumption 2, for any p, p0 ∈ [0, 1]
|φ(p) − φ(p0 )| ≤
W
α
|p − p0 | + |c1 (p − p0 )|.
α
Proof of Lemma 16. We can assume p0 ≤ p without loss of generality. The absolute continuously of φ gives
Z p
Z p
(1)
0
φ(1) (x) dx .
φ (x)dx ≤
|φ(p) − φ(p )| =
p0
p0
From Lemma 15, we have
Z
0
p
(W xα−1 + c1 )dx
|φ(p) − φ(p )| ≤
p0
W α
(p − p0α ) + c1 (p − p0 )
α
W
α
≤ |p − p0 | + |c1 (p − p0 )|,
α
=
where the last line is obtained since a function xα for α ∈ (0, 1) is α-Holder continuous.
This is valid for the case p0 = 0. Indeed,
|φ(p) − φ(0)| = lim
|φ(p) − φ(p0 )|
p0 →0
W
0 α
0
|p
−
p
|
+
|c
(p
−
p
)|
≤ lim
1
p0 →0
α
W
α
= |p − 0| + |c1 (p − 0)|.
α
Lemma 17. Given α ∈ [0, 1], supP ∈Mk
Pk
i=1
1−α
pα
.
i =k
Proof of Lemma 17. If α = 1, the claim is obviously
Pntrue. Thus, we assume α < 1. We introduce the Lagrange multiplier λ for a constraint i=1 pi = 1, and let the partial derivative
Pk
Pk
of i=1 pα
i + λ(1 −
i=1 pi ) with respect to pi be zero. Then, we have
αpα−1
− λ = 0.
i
(31)
Since pα−1 is a monotone function, the solution of Eq (31) is given as pi = (λ/α)1/(α−1) ,
Pk
i.e., the values of p1 , ..., pk are same. Thus, the function i=1 pα
i is maximized at pi = 1/k
Pk
for i = 1, ..., k. Substituting pi = 1/k into i=1 pα
gives
the
claim.
i
Pk
1
Lemma 18. Given α
<
0 and ∆
≤
supP ∈Mk :∀i,pi ≥∆ i=1 pα
=
i
k,
((1 − (k − 1)∆)α + (k − 1)∆α ) ≤ k∆α .
Proof. From the KarushKuhnTucker conditions, letting P ∗ = (p∗1 , ..., p∗k ) be a probability
vector that attains the supremum, there exist real values λ and δi ≥ 0 such that
(p∗i )α−1 − λ − δi = 0,
and p∗i = ∆ only if δi > 0. Thus, we have
p∗i = λ1/(α−1) or p∗i = ∆.
38
Hence,
sup
k
X
P ∈Mk :∀i,pi ≥∆ i=1
pα
i =
max
m=1,...,k−1
(m∆α + (k − m)(1 − m∆)α ).
Since ∆α ≥ (1−m∆)α for m = 1, ..., k−1, the maximum is attained at m = k−1. Moreover,
we have (1 − (k − 1)∆)α ≤ ∆α , and thus we get the claim.
39
| 7cs.IT
|
arXiv:1712.00193v1 [cs.CV] 1 Dec 2017
Improving Smiling Detection with Race and Gender Diversity
Hee Jung Ryu
Margaret Mitchell
Hartwig Adam
[email protected]
[email protected]
[email protected]
Google Inc.
Google Inc.
Google Inc.
Abstract
Recent progress in deep learning has been accompanied
by a growing concern for whether models are fair for users,
with equally good performance across different demographics [26, 17]. In computer vision research, such questions
are relevant to face detection and the related task of face
attribute detection, among others [15].
We measure race and gender inclusion in the context
of smiling detection, and introduce a method for improving smiling detection across demographic groups. Our
method introduces several modifications over existing detection methods, leveraging twofold transfer learning to
better model facial diversity. Results show that this technique improves accuracy against strong baselines for most
demographic groups as well as overall.
Our best-performing model defines a new state-of-theart for smiling detection, reaching 91% on the Faces of
the World dataset. The accompanying multi-head diversity classifier also defines a new state-of-the-art for gender
classification, reaching 93.87% on the Faces of the World
dataset. This research demonstrates the utility of modeling
race and gender to improve a face attribute detection task,
using a twofold transfer learning framework that allows for
privacy towards individuals in a target dataset.
Figure 1. End-to-end Architecture of our Smile Detector. Gray
and blue layers are part of the pretrained face recognition model,
modeled after FaceNet v8 [36]. Blue layers are transferred to
the diversity and smile classifiers. Green denotes the layer pretrained for gender classification and from which we transfer learn
our smile classifier. Yellow denotes the layer pretrained for race
classification and from which we transfer learn our smile classifier.
Pink denotes layers trained from scratch for the smile classification
task.
short hair and no makeup may be classified as male [31]. A
black face may not be recognized at all in a face tracking
task, while a face with a white mask will be [9].
In this paper, we examine the problem of unequal performance across different races and genders in the context
of smiling detection and the ChaLearn: Looking at People
challenge [15]. We document disparities between demographic subgroups, working with four race subgroups and
two genders, which serve as an initial set of subgroups, with
room to grow.
We detail a method (see Figure 1 for details) to improve
accuracy across the demographic subgroups and overall
while preserving the demographic privacy of unique individuals. Our approach utilizes transfer learning, using some
of the last hidden layers of a face recognition model to train
a race classifier and a gender classifier (as shown in Figure
4) on held-out public data, then combining the last hidden
layers from these classifiers into a third model trained for
1. Introduction
Detecting and recognizing people and faces has become
an increasingly feasible machine learning task due to the
rise in deep learning techniques and large people-focused
datasets [3, 20, 36, 27, 24, 19, 44]. The closely related
task of face attribute detection has seen similar gains, with
work on everything from age estimation [25] to arched
eyebrow detection [22] demonstrating significant strides in
what modern machine learning can do.
However, this rise has come at a cost: As modern and
public technology has incorporated modern computer vision techniques, users have noticed a troubling gap between how well some demographics are recognized compared with others [26]. For example, a woman’s face with
1
the task of smiling detection. This results in overall improved smiling detection accuracy across the board, including improvements for minority demographic subgroups. It
also reduces the errors captured in false negative rate (FNR)
and false positive rate (FPR) calculations, decreasing the average false rate between the two. Our contributions include:
1. We analyze the diversity across existing smiling detection datasets, focusing on the ChaLearn: Looking at
People challenge [15], and documenting demographic
differences.
2. We propose an approach for face attribute detection
that is sensitive to race and gender underrepresentation
in training data, race and gender disparities in system
predictions, and user privacy.
3. We demonstrate improved performance for the smiling
task, reaching state-of-the-art on Faces of the World.
4. We present quantitative and qualitative results on system performance across race and gender subgroups.
2. Ethical Considerations
Intent. The intent of this work is to demonstrate the utility of reasoning about demographics – in this case, race and
gender – in order to do better across those demographics, as
well as overall, on a downstream task – in this case, smiling detection. What this research does not do is motivate
race and gender identification as an end-goal, nor motivate
inferring race and gender for individuals without their consent. Use of ideas in this work towards those goals are not
supported by the authors. Although this work concerns race
and gender, it does not provide evidence or support for inferring race and gender directly.
Race and Gender categories. The categories that we
choose within race and gender are imperfect for a number
of reasons, including the fact that there is no gold-standard
for ‘race’ categories; race is often a collection of a variety of
demographic information; and it is debatable whether gender should be treated as discrete categories or a spectrum. A
deep-dive into the subgroup categories is intentionally absent from this work in order not to promote the idea that the
subgroup categories we use are a good set to use generally.
3. Related Work
Research in computer vision has included work on issues
that have direct social impact, such as security and privacy
[37, 30], however, research on the related issue of diversity
in computer vision is surprisingly lacking (but see, e.g., [9]
for details about some of the pressing issues).
However, there are several other strands of research relevant to the work in this paper: 1) Face Attribute detection, including smile detection and gender/race classifica-
tion from face images, 2) Dataset Bias, and 3) Machine
learning fairness.
Face Attribute Detection. The ChaLearn “Looking at
People” challenge from 2016 [15], piqued the interests and
accelerated the research progress in detecting face attributes
such as smile and gender. The challenge provides the Faces
of the World (FotW) dataset for training and validation.
[46] won the first place of the challenge, utilizing multitask learning (MTL) and fine-tuning on top of a model
trained for face recognition [27]. [28] later published an
out-performing result for the same task on FotW utilizing
transfer learning from a face recognition model [35] as well
as MTL. The authors explore frozen parameter values rather
than fine-tuning all the parameters. [42] reached a stateof-the-art performance on the smiling task on the CelebA
test dataset [22]. The authors employ both transfer learning
from a face recognition model similar to [36] and MTL in
addition to providing additional features as inputs besides
the features extracted from a face recognition model. The
authors utilize the geo-location and weather information inferred from a given face as two additional input features to
their MTL model.
This work explored the latent relationship between
smaller attributes of the face and attributes related to identity, such as gender, finding that performance on the smaller
attribute tasks could be improved by also inferring identity
attributes.
Several lines of research on face detection and recognition utilize skin tone as a proxy for race [18, 43, 32, 23],
however, skin tone is a variable visual feature within any
race group (and across lighting conditions), and has been
shown to be one of the least helpful features for distinguishing between races [34, 7, 45, 4, 39].
[42] explores inferring race from geo-location features
extracted from a face. [16] explores inferring racial population geo location distribution from car maker geo location
distribution. However, the application of such race classification models are limited and have not been adopted as tools
to measure and improve fairness performance of a computer
vision model. [38] has open-sourced the first race classification model with five races, including a subset of the US
Census and Hispanic.
Bias and Diversity [40] is one of the first papers to
present the dataset bias problem, documenting the issue in
popular object recognition datasets and exploring its implications and effects on computer vision models. [15] points
out the importance of the diversity of the source material
in a vision dataset, and postulates that analysis made on
skewed datasets cannot be representative enough for benchmarking progress. The authors provide Faces of the World
(FotW), collected with the aim of achieving a uniform distribution across different gender (female and male) and ethnic groups (Asian, Black, Hispanic, White).
Recent work has also applied computer vision techniques to bring into light gender bias present in movies
[2]. Using a tool they introduced called GDIQ, the research
team detects faces and measures the scene time by gender. The team discovered that actresses are seen 60% fewer
times than actors.
Fairness in Machine Learning. The area of machine
learning and fairness has recently surged, with newly uncovered theoretical guarantees and demonstrated results using modern deep learning techniques. [13] advocates for
“Fairness through Awareness”, demonstrating the importance of being aware of sensitive characteristics such as gender and race in order to work towards models that work well
regardless of demographics. Proposals for fairness have
included parity, such as Demographic Parity [13, 17, 5],
which requires that the proportion of individuals from each
demographic group matches for different decisions of the
algorithm. However, as pointed out by [13], this provides
no guarantee that the quality of the algorithmic decisions are
equally good for each group. Building from this idea, [17]
discusses measures for equality, such as Equality of Opportunity, which requires equal False Negative Rates across
subgroups, and Equality of Odds, which additionally requires equal False Positive Rates across subgroups.
[5] applies Equality of Opportunity within a multi-task
paradigm, where the main head predicts a task such as income bracket, and the auxiliary head predicts a sensitive
attribute: The gradient from that prediction is then negated
and passed back into the model, effectively minimizing accuracy on the sensitive attribute task while maximizing attribute on the main task. In this work, we take a somewhat
opposite approach, learning race and gender, and then transferring that learning into a new model.
4. Learning Diversity
4.1. Demographic Accuracy
At the core of this work lies the idea that faces look different across different races and genders [34, 6], and that
it is equally important to do well on each demographic
group. To the best of our knowledge, it has not been studied whether there exists a direct relationship between smile
detection and race and gender subgroup classification tasks.
If such a relationship holds, then smile detection might be
improved across races and genders by the model learning
about them. As we will demonstrate, learning race and gender does indeed improve performance.
A classifier trained to maximize the likelihood of a given
dataset will learn representations within the bounds of that
dataset [40]: If the dataset exhibits racial disparity, then
models trained on it will learn to encode similar disparities,
with poor performance on underrepresented races and better
performance on the more represented races in the dataset.
Figure 2. Our multihead diversity classifier’s gender head sets a
new state-of-the-art accuracy on FotW and achieves the state-ofthe-art accuracy on CelebA. UMIACS [28] fine-tuned their gender model on FotW train dataset before evaluating against FotW
validation dataset. For accurate performance comparison to their
model, we adopted their method in this analysis. The models compared are [46, 21, 41, 14, 28, 42, 29, 22, 47] from left to right in
the legend.
Poor performance may take the form of low accuracy on
some subgroups, and frequent false positives or false negatives on some subgroups. These disparities may be missed
when the training and validation datasets reflect a similarly
skewed demographic distributions.
To evaluate systems in terms of their performance across
subgroups, we report values for accuracy per subgroup. Inspired by recent work in fairness in machine learning (see
Section 3), we also introduce a metric that evaluates false
positive rates and false negatives rates in a way that is robust to label and subgroup imbalances, the average false
rate (AFR). Given:
FP = False Positives; FN = False Negatives
TP = True Positives; TN = True Negatives
FP
FPR = False Positive Rate = FP+TN
FN
FNR = False Negative Rate = FN+TP
The Average False Rate is equal to the weighted average
between the False Positive Rate and False Negative Rate:
AFR =
(FNRn + FPRp)
n+p
(1)
where n and p are weights on FNR and FPR. Weighting
provides the opportunity to capture different relative importance between the two kinds of error rates depending on
the task the metric is used on. AFR is somewhat comparable to Error Rate, however, balances positive and negative
labels so that there is no effect of label imbalance in the
error calculation. In this work, we weight FPR and FNR
equally (n = p = 1). When we calculate performance for
the full dataset, in addition to reporting accuracy, we report the mean of the Average False Rate across subgroups
to additionally address issues with both label and class im-
balance.
4.2. Twofold Transfer Learning
Transfer learning is useful in the case where there is a
feature space X and a marginal distribution over the feature
space P (XS ), where XS = x1 , x2 ...xn ∈ X , as well as a
marginal distribution over the related feature space P (XD ),
where XD ∈ X . In this setting, X is the space of all face
features, XS are face images for the smiling task, and XD
are face images for the demographic classification task, race
or gender.
Formally, given:
• A face domain FR for the task of race classification TR
with labels YR and images XR
• A face domain FG for the task of gender classification
TG with labels YG and images XG
Figure 3. To understand how inclusive our multi-head diversity
classifier is on gender classification, we measured its accuracy by
race subgroups against FotW validation set.
relatively equal across different race subgroups in the FotW
dataset, as reported in Figure 3.
4.4. Race and Gender Architectures
• The target face domain FS for the task of smiling detection TS with labels YS and images XS
We optimize for TR and TG by learning P (YR |XR ),
P (YG |XG ) and then include those learned representations
in the target domain FS when learning P (YS |XS ), hence
the name twofold transfer learning. This has the added
benefit of protecting the privacy of individuals in the smiling detection task, without directly inferring the gender and
race of those individuals.
4.3. Sampling
One concern in training classifiers is dataset class imbalance, where some classes or subgroups are more prevalent than others in the training data. This negatively impacts how well a trained classifier can generalize beyond
the frequent groups in the training data to groups that are in
the minority. Working with object recognition and CNNs,
[8] demonstrates that oversampling outperforms other popular sampling methods such as undersampling in most types
of class imbalance. However, when we oversampled more
than 10 folds on our face dataset with a high class imbalance in the context of race classification, we did not see any
performance improvements. In order to balance our dataset
across different classes for training, we instead list the eight
(race × gender) combinations and undersample our dataset
in a stratified sampling fashion, with up to about 25K face
images per race and gender intersection.
A few minority subgroups in our dataset could not even
meet the under-sample threshold. For these, we employ
an off-line oversampling method in order to make sure
each training batch contains faces across all race × gender. Our resulting diversity classifier sets a new state-ofthe-art gender classification accuracy on FotW dataset with
93.87% and reaches the current state-of-the-art accuracy on
the CelebA datset. See Figure 2 for more comparisons. Our
multihead diversity classifier’s gender head also performs
Figure 4. Multihead (gender and race) diversity classifier is used to
provide features that improve the task of smiling detection across
different races.
Figure 4 provides a high-level overview of the architecture used for learning in the demographic domains, FR and
FG . First, a face detector (similar to Picasa [10]) is run to
crop and align faces. Using these face crops, we train a face
recognition system modeled after the FaceNet v8 architecture [36]. Following [11], we extract features from the layer
called avgpool, which we refer to as avgpool-facenet in the
rest of this paper. We add fully connected layers and then
fine-tune the model for the race and gender estimation tasks.
To train the multihead (race and gender) diversity classifier, we use data collected from publicly available resources
of celebrities appearing in [1, 33, 38], and further curate the
data so that train and validation sets consist of disjoint sets
of identities.
Unlike the traditional approach to labeling race and gender, we obscure the race and gender subgroup names whenever possible1 in order to avoid introducing unconscious,
societal bias associated with existing labels. We call gender
Gender 1 and Gender 2, with an eye towards adding more
gender subgroups in future work. We call the races Subgroup 1, Subgroup 2, Subgroup 3, Subgroup 4, and so on,
and similarly anticipate using this naming scheme to extend
to further race subgroups.
1 FotW is annotated with ‘Female’, ‘Male’, and ‘Not Sure’; CelebA has
‘Male’ and ‘Non-Male’; we adopt these labels for direct comparisons with
previous work on these datasets.
avgpool-facenet
inception(4e)
S1
85.7
86.1
S2
88.9
90.4
S3
89.4
90.9
S4
92.0
92.4
Other
88.4
87.6
Total
89.04
89.82
Table 1. Accuracy (in %) by different input features according to
the layers detailed for FaceNet v8 [36], Subgroups (S) 1-4. (Train:
FotW, Eval: FotW, 1FC(4096))
Figure 5. Example subgroups used in race classification.
Gender Classification. Our model’s performance on
gender classification in the FotW [15] and CelebA [22]
datasets are 93.87% and 99% respectively, setting the new
state-of-the-art gender classification accuracy. Figure 2 illustrates the model’s performance compared to other existing models. Additional experiments demonstrate that gender classification works surprisingly well with only the top
part of the face, also known as the “attention region” (eyes
and nose) [34]. The trained model achieves accuracy that
only slightly underperforms the experiments with full faces.
Race Classification. We explore four race subgroups for
which we were able to scrape over 100,000 web images of
famous identities listed in [1, 33, 38]. This allows us to train
a reasonably performing race head of a multihead (race and
gender) diversity classifier, however, this is a limited set of
all the races that might be considered.
With the initial classifier, we can demonstrate the potential usefulness of learning race when detecting face attributes. We train our model to reach 98% or greater AUC
across the four race subgroups.
4.5. Smile Detector Architecture
Figure 1 provides a high-level overview of the architecture used for learning to detect smiles in the smiling domain
FG , building from the learned representations FR and FG .
Figure 6 is a close-up on the part of the model architecture
trained from scratch and outputs smiling detection results.
down to total 3,072 bytes (= FC1 1,024B + FC2 1,024B + FC3
1,024B). Details of all architectures explored are presented
in Table 2.
Learning Rate. Learning rate plays an important role in
our small neural net for it to avoid getting stuck in a local
minima before having enough exposure to learn from and
fine-tune on its teacher model(s). It is also not ideal for the
model to overfit to the training dataset due to the learning
rate being under a certain threshold. In our experiments, we
adjusted the initial learning rate through brute force search,
with an aim to reduce the oscillation of the continuous training accuracy- and loss- curves. We found that initial learning rates between the range of 0.005 and 0.001 (inclusive)
work the best for our experiments, with an early stopping
between 60K and 100K training steps when trained with a
batch size of 64.
Number of Parameters Trained. 1024×1024+4096×
1024 × 2 + 1024 × 3 × 2 = 944,332 weight parameters
are trained from scratch. Comparison experiments with the
same number of parameters but without the transfer learning demonstrate that the transfer learning adds additional
benefit to the final model.
Optimization and Loss. We use softmax along with
a cross-entropy loss for smiling, following the state-ofthe-art smiling model [28]. For optimization, we use
‘AdaGrad’[12].
(a) 4096
(b) 4096, 4096
(c) 1024, 1024
(d) MaxPool, Conv, 2048, 512
(e) MaxPool7x7x1(V), 1024, 1024
(f) AvgPool7x7x1(V), 1024, 1024
Figure 6. Smile Detector Architecture (Close-up).
Input Features and Twofold Transfer Learning. We
transfer learn our final smile detection model with three sets
of features extracted from two different pretrained models:
a face recognition model and a multihead (race/gender) diversity classifier.
Bottleneck Layers. We use AvgPool of dimension 7x7
filter and stride 1 with ‘valid’ padding and smaller FCs as
our bottleneck layers to generalize our rich input features of
total size 58,368 bytes (= S 50,176B + R 4,096B + G 4,096B)
Accuracy
89.82%
89.72%
89.85%
90.01%
90.47%
90.57%
Table 2. Architecture experiment (Train: FotW, Eval: FotW, Input:
Inception(4e))
5. Smile Detection Datasets
5.1. Looking at People, Smile and Gender Challenge
The ChaLearn Looking at People2 Smile and Gender
Challenge [15] requires classifying images according to
2 http://chalearnlap.cvc.uab.es/challenge/13/
track/20/description/
C ELEBA F OT W
Total
Train
Valid
6171
3086
Train 162687
Valid 19863
Test
19955
Estimated Race
S1
S2
S3
Gender
S4 Other
8% 48% 11% 15% 19%
8% 53% 11% 17% 11%
7% 75% 7% 8%
7% 76% 6% 6%
9% 69% 9% 9%
4%
6%
6%
F
6. Experiments
M Other
52% 46% 2%
55% 44% 1%
0* 42% 58%
0* 42% 58%
0* 39% 61%
Table 3. Estimated Race and Groundtruth Gender Distribution across Datasets. Duplicated images in CelebA are counted
only once in this dataset analysis. *FotW comes with ‘Female’ (F)
and ‘Male’ (M) labels. CelebA comes with ‘Male’ (M) and ‘NonMale’ (Other).
6.1. Analyzing Inclusion
Figure 7 and Table 3 detail the racial inclusion in the
Faces of the World and CelebA datasets used in the LAP
challenge. For Faces of the World, the race model estimates
a high variance among the subgroups, with the plurality of
the data for Subgroup 2. Subgroup 1 is the smallest of the
subgroups, with only 523 training data instances. This is
contrary to the reported racial diversity of FotW in [15],
and may be due to the fact that these are estimated labels
from our model, as well as due to differences in how different parties will categorize racial subgroups. As we will
demonstrate, agreement on race categories is not needed to
improve performance.
For CelebA, there is less variance across subgroups, with
Subgroup 2 being the biggest.
6.2. Learning Smiling from Race and Gender
Figure 7. Diversity in FotW validation set measured by image
count.
gender and smiling. Models may be trained on the provided
FotW and CelebA datasets [22], as well as pretraining from
other sources.
In the Faces of the World (FotW) dataset, each image
displays a single face labeled with smile, and some with
gender. The train set has 2,234 smile and 3,937 no smile,
and the validation set has 1,969 smile and 1,117 no smile.
See Table 3 for more details across genders and races. We
estimate that Subgroup 2 makes up about half the data, with
the rest as minorities.
The Large-scale CelebFaces Attributes (CelebA) data
contains 162,770 images with duplicates and 162,687
unique images in the train set; 19,867 images with duplicates and 19,863 unique images in the validation set; and
19,962 images with duplicates and 19,863 unique images in
the test set. Each image is labeled with smiling/not-smiling
and male/not-male. We estimate Subgroup 2 to be the vast
majority here, with relatively little data for other race subgroups. See Table 3 for more details.
To improve our models for the smiling task, we first examine the effect of different layers as input to the model.
We compare avgpool-facenet and inception(4e) in Table 1,
where the ‘Other’ label is applied when the confidence score
is < 0.5 for all subgroups.
Inception(4e) yields the best results across all racial subgroups, boosting the overall accuracy by +0.78% and the
accuracy of the minority Subgroup 1 by +0.41% on a simple fully connected layer of output size 4096. In the rest
of the experiments, we use features extracted from the inception(4e) layer as input to our smile detection system.
Considering how avgpool-facenet helps with our multihead
(gender and race) diversity classifier and how the layer
comes after various bottleneck layers, we conjecture that
avgpool-facenet has a richer set of features, more representative of facial identities than inception(4e). Meanwhile, inception(4e) may have a richer set of features representative
of facial muscle movements like smiling. This layer may
encode other features (possibly lighting or head poses) irrelevant to smiling such that we see an accuracy improvement of +0.72% after introducing a bottleneck layer similar
to avgpool-facenet into our system (See (c) and (f) in Table
2 for details).
We then compare MaxPool, Conv, AvgPool with two
fully connected (FC) layers, comparing different FC layers with the inception(4e) layer. Moving from one fully
connected layer to two does not improve performance, and
changing the size of the layer does not lead to significant
gains. However, adding a pooling layer leads to additional
gains. In Table 2, (d) was inspired by [28] and (f), by [42].
Since (f) with AvgPool worked the best, we experiment with
another popular pooling layer MaxPool. The final baseline architecture that we adopt is (f) in Table 2, achieving
90.47% accuracy on the smiling detection task. From here
Train: FotW
Eval: FotW
Subgroup 1
Subgroup 2
Subgroup 3
Subgroup 4
Other
ACC .% AFR%
ACC .% AFR%
ACC .% AFR%
ACC .% AFR%
ACC .% AFR%
ACC .%
Total
M AFR%
SMILING
+G
+R
+R +G
88.5
88.5
88.5
88.5
12.2
12.1
12.1
12.1
90.9
91.2
91.6
91.4
10.7
10.1
9.7
9.9
90.0
90.9
90.6
90.3
12.2
10.9
11.3
11.8
94.3
94.1
94.5
94.5
6.2
6.1
5.9
5.9
88.2
88.2
88.2
88.2
18.4
18.4
18.4
18.4
90.57
90.76
90.96
90.86
12.0
11.5
11.5
11.6
on 1FC(2048)
on 1FC(3072)
on 2FCs(930,4608)
88.1
87.7
87.7
12.5
13.0
12.8
91.3
91.4
90.8
10.2
10.3
10.1
90.3
89.8
91.2
11.8
12.4
10.2
94.5
94.1
94.3
5.9
6.5
5.8
88.2
87.9
88.1
18.7
19.6
18.4
90.80
90.67
90.57
11.8
12.3
11.5
Table 4. Accuracy (ACC .), Average False Rate (AFR), and mean AFR (M AFR) by race subgroup for smiling detection with proposed
systems, on FotW dataset. +G denotes transferred representations from the gender head of the diversity classifier, +R from the race head
of the diversity classifier, and +R+G from both. The last three FC rows represent models with the same number of parameters as the +G,
+R, +R +G models, but without the transferred representations.
Train: FotW
Eval: FotW
Gender 1
ACC .% AFR%
Gender 2
ACC .% AFR%
SMILING
90.2
90.3
90.7
90.4
91.0
91.3
91.3
91.4
+G
+R
+R +G
10.6
10.4
10.0
10.5
12.5
11.6
11.8
11.5
Table 5. Accuracy (Acc.) and Average False Rate (AFR) by gender
for smiling detection with proposed systems, FotW dataset.
on, we call this baseline model S or SMILING (See Table 4).
As discussed earlier in Section 3, studies show a latent
relationship between physical appearance of lips and sexual and/or racial identity. From here, we postulate that the
appearance of smiling may likely differ by race and gender
and affects a smiling model’s performance. However, we
are often faced with face dataset lack of sexual and racial
diversity. To mitigate this issue, we train a multi-head diversity classifier that classifies gender and race (detailed in
Section 4.4) and augment our baseline model S with additional input features transfer learned from each head (race
and gender) of the diversity classifier (architecture found at
Figure 4). From here, we denote input features extracted
from GSubGFC in Figure 4, G and those extracted from
RSubGFC in Figure 4, R. We experiment with a few combinations of the input features on each of the datasets, included in the results below.
6.3. Evaluating
First, the face detector is run on the provided CelebA or
FotW thumbnails. If it fails to align the face (this happens
for 0.65% of CelebA test set faces and 12.57% of FotW validation set faces), the provided CelebA or FotW face crops
and alignments are used.
In addition to accuracy for smiling detection, we measure the Average False Rate (AFR) across race subgroups
(see Section 4.1). Intuitively, this measures how much the
model fails to predict smiling across smiling instances for
Figure 8. Our smile classifier sets a new state-of-the-art accuracy
on FotW. On CelebA, it is only behind by [42]. The models compared are [46, 21, 41, 14, 28, 42, 22, 47] from left to right in the
legend.
each race subgroup, and how much the model over-predicts
smiling across non-smiling instances for each race subgroup.
Results are shown in Table 4. We achieve an accuracy
higher than that of the previous state-of-the-art smile classifier [28], using both race and gender SMILING + R + G, and
set a new state-of-the-art performance of 90.96% with the
model that includes a learned race representation, SMIL ING + R .
To test whether the transfer learning has transferred representations that help for this task, we compare to a models with the same number of parameters but without the
transferred learning. See rows 1FC(2048), 1FC(3072),
2FCs(930,4608) in the lower half of Table 4. These correspond to 1 fully-connected layer with 2048 dimensions,
3072 dimensions, and 2 fully-connected layers with 930 and
4608 dimensions. We find that the models with the transfer
achieve higher overall accuracy, although, interestingly, the
same number of parameters lead to reduced AFR in some
cases, and the highest accuracy for Subgroup 3, outperform-
Figure 9. Smile Detector Performance by Gender. For both genders, AUCs are equally 0.96.
ing the next best performance achieved by adding gender.
The reason for this is not entirely clear, and we look towards ethical approaches for a deeper dive into differences
between subgroups in future work.
On FotW, our smile model SMILING + R + G reaches an
accuracy 1.62% (absolute) higher than the 2016 ChaLearn
Looking at People Challenge winner SIAT MMLAB [46],
and 0.13% (absolute) higher than the published state-of-theart UMIACS [28] on the FotW validation dataset. Both of
the models used transfer learning from a face recognition
model and MTL. SIAT MMLAB exploited fine-tuning the
entire face recognition model with a couple of iterations
of additional pre-training/fine-tuning using CelebA dataset
and then finally fine-tuning using FotW. UMIACS takes
a similar approach to ours in that the model is transfer
learned from features extracted from a multiple mid-/lowlayer of a face recognition model. They also pretrained with
CelebA and fine-tuned on FotW. Our best performing model
is trained on FotW train dataset on top of a face recognition
model, skipping the intermediate step of fine-tuning CelebA
dataset. Our model outperforms other models published
[21, 41, 14] with an increased performance gap as large as
11.66% absolute. Figure 8 presents an accuracy comparison
graph.
We also plot ROCs with respect to smiling by gender
in Figure 9. AUCs of our model per gender are equally
0.96, demonstrating Equality of AUC similar to the fairness
metrics of [17].
7. Discussion
This paper is one of the first to demonstrate that building an inclusive vision model and building a state-of-the-art
model can harmoniously go hand-in-hand. Moreover, this
vision diversity research result gives us more insight into
understanding our vision models, the interaction between
our vision models, and visually distinctive features on our
faces. We present a more in-depth analysis of race/gender
intersectionality in the Appendix.
Transferring race knowledge increases the accuracy for
subgroups 2, 3, and 4, and improves total accuracy to the
state of the art. It also decreases the average false rate for
the four race subgroups and two gender subgroups, as well
as overall. Notably, learning race improves performance
across both genders.
Transferring gender knowledge similarly leads to accuracy and false rate improvements, although it is not as impactful as race. The addition of gender features increases
the accuracy of the gender groups as well as race Subgroups
2 and 3, in addition to improving accuracy overall. The only
point in the evaluation where transferred knowledge hurts
performance is from the addition of gender for Subgroup 4,
where accuracy is less than that of the baseline SMILING.
Adding both race and gender does not significantly improve over adding race alone for all race subgroups and
overall, however, it does improve performance on Gender
2 beyond improvements provided by race or gender alone.
The limited improvements provided by the model + R + G in
Tables 4 and 5 may be due in part to the fact that with this
model comes with more parameters, which may be more
difficult to tune with the relatively small amount of data provided by the FotW dataset.
We did not see any accuracy change for subgroups with a
dataset size much smaller than a hypothetical threshold. For
example, we did not see any performance changes to race
Subgroup 1 (denoted as SubG1). We attribute this to the
data for that subgroup being too small compared to the rest,
as shown in Figure 7. This is similar to the class imbalance
problem discussed in Section 4.3. From this, we make a soft
conclusion that our methodology’s effectiveness is limited
when the size of a minority demographic subgroup dataset
does not meet some minimum threshold. This gives a strong
motivation for collecting a more diverse dataset in order to
achieve an even greater model performance boost.
Despite our diversity classifier’s state-of-the-art performance on gender classification (see Figure 3 and 2), the
model S+G fine-tuned on the gender-head of the diversity
classifier shows some accuracy drop in Subgroup 4 in Table
4. Looking at Figure 7, we see a good gender parity in the
dataset. This may explain why the additional two-gender
input features do not help much: Two genders are already
well represented in the training data.
It is worth noting that Subgroup 4 has the highest accuracy compared to the other subgroups, and is the only
subgroup that results in a decrease in accuracy when gender
representations are added. Perhaps the information transferred for gender is already better captured in the baseline
smiling model for this subgroup, although it’s not clear why
the model performs particularly well on this group. Also
noteworthy is the fact that Subgroup 3 achieves the best
performance when two fully connected layers are added –
8. Conclusion
Figure 10. Baseline vs. Models with Extra Input Features. Left
column shows examples the baseline model got wrong that the
transfer learned models corrected. Right column shows examples
the baseline model got right, and the transfer learned models got
wrong.
without the transferred knowledge. We note these differences for future work that dives into details of each race
subgroup.
Throughout our experiments, the transferred knowledge
does not impact the ‘Other’ category, which is a diverse set
of people from many different backgrounds. This is interesting because while we do see a positive effect of transfer
learning for specific race and gender subgroups, this knowledge does not transfer to the other potential subgroups outside of the four represented. Simply adding more parameters in 2 fully connected layers decreases the average false
rate very slightly, but also decreases accuracy significantly.
We think this could be attributed to the fact that our diversity
classifier’s race head was not trained to capture and extract
representative features of the Other group here.
Finally, Figure 10 shows examples of the images in
which additional gender and/or race feature inputs made
differences. The faces in the examples tend to have a thin
smile, a neutral face, or a mouth open. With the additional input features, we see more correct smile detections
in black-and-white photos and in faces turned away from
the camera.
In the future, we intend to explore methods to boost
the lowest demographic accuracy and overall accuracy even
when the disparity between demographic subgroups, like
gender, is as small as that of this dataset. In addition, we
also see a possibility that the performance boost by gender
features could have been better measured if we measured
the performances by the estimated gender slices rather than
the groundtruth gender slices.
We acknowledge that our efforts taking an unbiased look
at diversity and fairness in our models has its limitations. In
future work, we aim to add more race and gender groups,
and obscure specific race and gender groundtruth labels
from the start. For example, by using unsupervised or
weakly supervised learning, such as clustering, we may
be able to create coherent groups that pattern along demographic lines, without the need for explicit labels.
We have demonstrated the problem of race and gender differences in smiling detection, and have proposed
a method for improving accuracy across races and genders. The model utilizes twofold transfer learning for better
representations of the race and gender subgroups that are
present in faces, and reaches a new state-of-the-art accuracy
for smiling detection on the Faces of the World dataset. In
this process, we also reach a new state-of-the-art accuracy
on gender classification using sampling to balance demographic labels and classes (subgroups).
When we use only FotW and transferred racial representations, the smiling detection model is improved to a new
state-of-the-art, with 90.96% accuracy. This is 1.62% absolute improvement over the previous ChaLearn winner [46],
and 0.13% absolute improvement over the previous state of
the art [28]. By incorporating the learned representations of
both gender and race, Gender 2 further improves.
In future work, we plan to address the fact that our gender and race subgroup categories are limited and do not represent the real world to its entirety.
9. Acknowledgements
Special thanks to Blaise Agüera y Arcas and Corinna
Cortes for their supports and sponsorship. Thanks to Caroline Pantofaru, Sergio Guadarrama, Rahul Sukthankar, Florian Schroff, Marco Andreetto, David Karam, Krishna Somandepalli, Shrikanth Narayanan, and Moritz Hardt for
helpful conversations. Thanks to Divya Tyam, Andrew Zaldivar, and Ben Hutchinson for donating their face photos for
our research.
References
[1] Wikidata. http://www.wikidata.org/.
[2] H. Adam, I. Breckinridge-Jackson, R. Cooper, T. Guha,
I. Hall, C. wei Huang, N. Kumar, N. Malandrakis, M. Nasir,
A. Ramakrishna, K. Somandepalli, and Y. Zhu. The reel
truth: Women arent seen or heard: An automated analysis of gender representation in popular films. https:
//seejane.org/wp-content/uploads/gdiqreel-truth-women-arent-seen-or-heardautomated-analysis.pdf, 2016.
[3] B. Amos, B. Ludwiczuk, and M. Satyanarayanan. Openface:
A general-purpose face recognition library with mobile applications. Technical report, CMU-CS-16-118, CMU School
of Computer Science, 2016.
[4] G. Anzures, O. Pascalis, P. C. Quinn, A. M. Slater, and
K. Lee. Minimizing skin color differences does not eliminate the own- race recognition advantage in infants. Infancy,
page 640654, 2011.
[5] A. Beutel, J. Chen, Z. Zhao, and E. H. Chi. Data decisions
and theoretical implications when adversarially learning fair
representations. CoRR, abs/1707.00075, 2017.
[6] U. Bindal, P. G. Bindal, and N. A. bt Raml. Labial impressions: A tool for identification. Journal of Forensic Research, 2014.
[7] K. R. Brooks and O. S. Gwinn. No role for lightness in the
per- ception of black and white? simultaneous contrast affects per- ceived skin tone, but not perceived race. Perception, 39:11421145, 2010.
[8] M. Buda, A. Maki, and M. A. Mazurowski. A systematic
study of the class imbalance problem in convolutional neural
networks. CoRR, abs/1710.05381, 2017.
[9] J. Buolamwini.
How i’m fighting bias in algorithms.
https://www.ted.com/talks/joy_
buolamwini_how_i_m_fighting_bias_in_
algorithms#t-63664.
[10] D. Chen, S. Ren, Y. Wei, X. Cao, and J. Sun. Joint cascade
face detection and alignment. In ECCV, 2014.
[11] F. Cole, D. Belanger, D. Krishnan, A. Sarna, I. Mosseri, and
W. T. Freeman. Face synthesis from facial identity features.
CoRR, abs/1701.04851, 2017.
[12] J. C. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159,
2011.
[13] C. Dwork, M. Hardt, T. Pitassi, O. Reingold, and R. Zemel.
Fairness through awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, ITCS
’12, pages 214–226, New York, NY, USA, 2012. ACM.
[14] M. Ehrlich, T. J. Shields, T. Almaev, and M. R. Amer. Facial attributes classification using multi-task representation
learning. In The IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) Workshops, June 2016.
[15] S. Escalera, X. Baro, H. J. Escalante, and I. Guyon. Chalearn
looking at people: A review of events and resources. In Proceedings of the The 2017 International Joint Conference on
Neural Networks (IJCNN 2017). IEEE, 2017.
[16] T. Gebru, J. Krause, Y. Wang, D. Chen, J. Deng, E. A.
Lieberman, and L. Fei-Fei. Using deep learning and google
street view to estimate the demographic makeup of the US.
CoRR, abs/1702.06683, 2017.
[17] M. Hardt, E. Price, , and N. Srebro. Equality of opportunity in supervised learning. In D. D. Lee, M. Sugiyama,
U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances
in Neural Information Processing Systems 29, pages 3315–
3323. Curran Associates, Inc., 2016.
[18] H. Hill, V. Bruce, and S. Akamatsu. Perceiving the sex and
race of faces: The role of shape and color. In the Royal
Society B: Biological Sciences, volume 261, page 367373,
1995.
[19] G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller.
Labeled faces in the wild: A database for studying face
recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October
2007.
[20] I. Kemelmacher-Shlizerman, S. M. Seitz, D. Miller, and
E. Brossard. The megaface benchmark: 1 million faces for
recognition at scale. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 4873–
4882, 2016.
[21] C. Li, Q. Kang, and Q. Ge. Deepbe: Learning deep binary encoding for multi-label classification. In Computer
Vision and Pattern Recognition Workshops CVPRW, 2016
IEEE Conference on. IEEE.
[22] Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face
attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015.
[23] C. R. Malskies, E. Eibenberger, and E. Angelopoulou. The
recognition of ethnic groups based on histological skin properties. In the Vision, Modeling, and Visualization Workshop,
page 353360, 2011.
[24] A. Nech and I. Kemelmacher-Shlizerman. Level playing
field for million scale face recognition. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2017.
[25] Z. Niu, M. Zhou, L. Wang, X. Gao, and G. Hua. Ordinal regression with multiple output cnn for age estimation. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.
[26] E. O. of the President.
Big data: A report on algorithmic systems, opportunity, and civil rights.
https://obamawhitehouse.archives.gov/
sites/default/files/microsites/ostp/
2016_0504_data_discrimination.pdf.
[27] O. M. Parkhi, A. Vedaldi, and A. Zisserman. Deep face
recognition. In British Machine Vision Conference, 2015.
[28] C. D. C. Rajeev Ranjan, Swami Sankaranarayanan and
R. Chellappa. An all-in-one convolutional neural network
for face analysis. In 2017 IEEE 12th International Conference on Automatic Face & Gesture Recognition. IEEE.
[29] R. Ranjan, V. M. Patel, and R. Chellappa. Hyperface: A deep
multi-task learning framework for face detection, landmark
localization, pose estimation, and gender recognition. CoRR,
abs/1603.01249, 2016.
[30] N. Raval, A. Machanavajjhala, and L. P. Cox. Protecting
visual secrets using adversarial nets. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops, Honolulu, HI, USA, July 21-26,
2017, pages 1329–1332, 2017.
[31] K. Rodden. Is that a boy or a girl? exploring a neural networks construction of gender. https://medium.com/
\spacefactor\@m{}kerryrodden/is-that-aboy-or-a-girl-cb93abbae6da.
[32] S. M. M. Roomi, S. Virasundarii, S. Selvamegala,
S. Jeevanand-ham, and D. Hariharasudhan. Race classification based on facial features. In NCVPRIPG, page 5457,
2011.
[33] R. Rothe, R. Timofte, and L. V. Gool. Deep expectation
of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision
(IJCV), July 2016.
[34] F. S, H. H, and H. ZG. Learning race from face: A survey.
IEEE Trans Pattern Anal Mach Intell., 36:2483–509, Dec
2014.
[35] S. Sankaranarayanan, A. Alavi, C. D. Castillo, and R. Chellappa. Triplet probabilistic embedding for face verification
and clustering. CoRR, abs/1604.05417, 2016.
[36] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR,
2015.
[37] J. Sokolic, Q. Qiu, M. R. D. Rodrigues, and G. Sapiro.
Learning to identify while failing to discriminate. In The
IEEE International Conference on Computer Vision (ICCV),
Oct 2017.
[38] K. Somandepalli. Prediction race from face for movie
data.
https://github.com/usc-sail/micarace-from-face/wiki, 2017.
[39] M. A. Strom, L. A. Zebrowitz, S. Zhang, P. M. Bronstad, and
H. K. Lee. Skin and bones: The contribution of skin tone and
facial structure to racial prototypicality ratings. PLoS ONE,
7:18, 2012.
[40] A. Torralba and A. A. Efros. Unbiased look at dataset bias.
In CVPR, 2011.
[41] M. Uřičář, R. Timofte, R. Rothe, J. Matas, and L. V. Gool.
Structured output SVM prediction of apparent age, gender
and smile from deep features. In Proceedings of IEEE conference on Computer Vision and Pattern Recognition Workshops, Las Vegas, USA, 2016.
[42] J. Wang, Y. Cheng, and R. S. Feris. Walk and learn: Facial
attribute representation learning from egocentric video and
contextual data. In CVPR, 2016.
[43] Y. Xie, K. Luu, and M. Savvides. A robust approach to facial ethnicity classification on large scale face databases. In
BTAS, page 143 149, 2012.
[44] D. Yi, Z. Lei, S. Liao, and S. Z. Li. Learning face representation from scratch. CoRR, abs/1411.7923, 2014.
[45] L. Yin, J. Loi, and J. Jia. Towards race-related face identification: Research on skin color transfer. In IEEE Int. Conf.
Autom. Face Gesture Recog, page 362368, 2004.
[46] K. Zhang, L. Tan, Z. Li, and Y. Qiao. Gender and smile classification using deep convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition Workshops, pages 34–38, 2016.
[47] N. Zhang, M. Paluri, M. Ranzato, T. Darrell, and L. D. Bourdev. PANDA: pose aligned networks for deep attribute modeling. CoRR, abs/1311.5591, 2013.
10. Appendix
Evaluation Analysis at Gender and Race Subgroup Intersections, Accuracy (See Figure 11). Adding Race
(+R) increases accuracy over the baseline in all evaluations.
Adding gender on top of race (+R +G) can lead to further
gains. In isolation, adding the gender representation (+G)
helps for two of the four tested subgroups. For subgroup 4,
accuracy is decreased.
Statistical Significance (See Table 6) Differences between the baseline S and the twofold transfer learned models S+R, S+G, and S+R+G are statistically significant.
Difference between S+R and S+R+G is not significant.
We set the significance level α with Bonferroni correction for α=.05. With correction for the 6 tests we run, the
S
S
S+G
S+R
S+R+G
p-value
5.7e−7
1.0e−4
5.7e−7
χ2
25.0
15.1
25.0
S+G
p-value
5.7e−7
3.9e−2 *
9.8e−4 *
χ2
25.0
2.0*
0.0*
S+R
p-value
1.0e−4
3.9e−2 *
0.38*
χ2
15.1
2.0*
1.0*
S+R+G
p-value χ2
4.0e−4
12.4
9.8e−4 * 0.0*
0.38*
1.0*
-
Table 6. McNemar Test. (*) indicates P-value from an exact binomial test and χ2 from McNemar’s test with continuity correction.
Bonferroni-corrected α is .008. At this level, S+G and S+R
are not significantly different, although they are significant
at the uncorrected .05 significance level. Results demonstrate that transfer learned race and gender representations
help for improving smiling detection across races, with race
leading to the highest overall accuracy.
Figure 11. Green rows indicate accuracy with an upward boost from the baseline, S, using the transfer learned representations. The
best-performing model is in bold.
| 1cs.CV
|
Enhanced Discrete Particle Swarm Optimization Path
Planning for UAV Vision-based Surface Inspection
Manh Duong Phunga , Cong Hoang Quacha , Tran Hiep Dinhb , Quang Hab,∗
a Vietnam
arXiv:1706.04399v1 [cs.RO] 14 Jun 2017
b University
National University, 144 Xuan Thuy, Cau Giay, Hanoi, Vietnam
of Technology Sydney, 15 Broadway, Ultimo NSW 2007, Australia
Abstract
In built infrastructure monitoring, an efficient path planning algorithm is essential for robotic inspection of large surfaces using computer vision. In this
work, we first formulate the inspection path planning problem as an extended
travelling salesman problem (TSP) in which both the coverage and obstacle
avoidance were taken into account. An enhanced discrete particle swarm optimisation (DPSO) algorithm is then proposed to solve the TSP, with performance
improvement by using deterministic initialisation, random mutation, and edge
exchange. Finally, we take advantage of parallel computing to implement the
DPSO in a GPU-based framework so that the computation time can be significantly reduced while keeping the hardware requirement unchanged. To show
the effectiveness of the proposed algorithm, experimental results are included
for datasets obtained from UAV inspection of an office building and a bridge.
Keywords: Path planning, infrastructure monitoring, bridge inspection,
vision-based inspection, particle swarm optimization, unmanned aerial vehicle
1. Introduction
For robotics inspection of built infrastructure, computer vision can be used
to detect most surface deficiencies such as cracking, spalling, rusting, distortion,
∗ Corresponding
author
Email addresses: [email protected] (Manh Duong Phung), [email protected]
(Cong Hoang Quach), [email protected] (Tran Hiep Dinh), [email protected]
(Quang Ha)
Preprint submitted to Arxiv
June 15, 2017
misalignment, and excessive movements. Over the last decade, much research
effort has been devoted to this theme with computer vision becoming an important component of modern Structural Health Monitoring (SHM) systems for
built infrastructure such as rust detection of steel bridges [1], crack detection of
concrete bridges [2, 3, 4], or bridge condition assessment [5]. In this regard, it is
promising to integrate a computer vision system into mobile inspection robots,
such as unmanned aerial vehicles (UAVs) [6, 7] or ubiquitous robots [8, 9], especially when dealing with large and hardly accessible structures like tunnels
[10]. For this purpose, an efficient inspection path planning (IPP) algorithm is
therefore of crucial importance.
In vision-based inspection path planning, it is required to find a trajectory
that is informative enough to collect data from different views of a given structure so that the inspection robot can carry out the data acquisition of the region
of interest. Depending on size of the inspecting region, the trajectory can be
planned for multiple robots to coordinately conduct the data collection [11]. To
be visibly processed at a later time, the data collected are often from a sensor
of the time-of-flight (optical, sonar or radar) or passive optical (CCD camera)
type. Since the computational time for IPP rapidly increases with the area of
the region of interest, an IPP algorithm should meet the following criteria:
(i) capability of viewing/covering every surface of the region of interest via
at least one planned viewpoint of the inspection sensor,
(ii) obstacle avoidance for the robot,
(iii) generation of an ”optimal” path under available conditions, and
(iv) effectiveness in terms of processing time (for online re-planning and large
structure inspection).
Studies on IPP, in general, can be categorised into three groups, namely
cell decomposition, sub-problem separation, and other methods. In cell decomposition, the target space is decomposed in sub-regions called cells. The cell
shape can be trapezoidal, square, cubic, or customised depending on critical
points of Morse functions, often with a uniform size [12, 13, 14]. An exhaustive
path connecting each cell is then computed for the coverage, using typically a
2
heuristic algorithm such as wavefront [15] or spiral spanning tree [16]. Methods based on cell decomposition yield good results in terms of coverage and
obstacle avoidance. As the path generated, however, may not be optimal, it
is worth seeking a better and more feasible alternative. In this context, the
IPP separation approach tackling the non-deterministic polynomial time (NP)hard problems can be divided into two, the art gallery problem that finds the
smallest set of viewpoints to cover the whole gallery, and the travelling salesman problem (TSP) that finds the shortest path to visit a set of given cities
[17, 18, 19, 20, 21, 22]. Each problem can be solved separately using known
methods such as the randomised, incremental algorithm for the art gallery problem [23, 24] and the chained Lin-Kernighan heuristics for the TSP [25]. Other
approaches have focused on sampling the configuration space [26], using submodular objective function [27], or employing genetic algorithms [28] but they
often require constraining the robot to certain dynamic models or end with
near-optimal solutions. The requirements remain not only a shorter path but
also collision-free.
In this paper, the IPP problem is addressed by first formulating it as an
extended TSP. The enhanced discrete particle swarm optimisation (DPSO) is
then employed to solve the IPP. Finally, parallel computing based on graphical
processing units (GPU) is deployed to obtain the real-time performance. The
contributions of our approach are three folds: (i) By formulating the IPP as
an extended TSP, both the coverage and obstacle avoidance are simultaneously
taken into account. In addition, constraints related to the kinematic and dynamic models of the robot are separated from the DPSO solution so that this
solution can be applied to a broad range of robots. (ii) Three techniques including deterministic initialisation, random mutation, and edge exchange have been
proposed to improve the accuracy of DPSO. (iii) Parallel computation has been
implemented to significantly improve the time performance of DPSO. By utilising GPU, the parallel implementation does not add additional requirements to
the hardware, i.e. the developed software can run on popular laptop computers.
The rest of this paper is structured as follows. Section 2 introduces the
3
steps to formulate the IPP as an extended TSP. Section 3 presents the proposed
DPSO and its deployment for solving the IPP. Section 4 provides experimental
results. Finally, a conclusion is drawn to end our paper.
2. Problem formulation
Our ultimate goal is to design a path planning system for an UAV used for
inspecting planar surfaces of largely built structures like buildings or bridges.
The sensor used for the inspection is a CCD camera attached to a controllable
gimbal. We suppose that the 3D model of the structure and the environment
are known prior to planning, for example, by using laser scanners. Here, the
IPP objective is to find the shortest path for the UAV’s navigation and taking
photos of the target surfaces so that the images captured can be later processed
to detect potential defects or damages. We first consider the IPP as an extended
TSP and then solve it using the developed DPSO. This section presents the
computation of viewpoint selection and point-to-point pathfinding, which are
fundamental to formulate the extended TSP problem.
2.1. Viewpoint selection
The viewpoint selection involves finding a set of camera configurations that
together cover the whole surfaces of interest. Let P be a finite set of geometric
primitives pi comprising the surfaces to be covered. Each geometric primitive
pi corresponds to a surface patch within the field of view of the camera. Let C
be the configuration space such that every feasible configuration cj ∈ C maps
to a subset of P . Each configuration ci corresponds to a position (xi , yi , zi ) and
an orientation (ϕi , θi , ψi ) of the camera. Given a finite set of configurations C,
the viewpoint selection problem on one hand calls generally for the minimum
number of configurations ci such that all elements pi ∈ P are covered. On the
other hand, from image sticking and defect detection, the following requirements
are added to the system: (i) image capturing moment is when the camera is
perpendicular to the inspected surface, (ii) sufficiently high resolution to distinguish the smallest feature, sf , and (iii) overlapping of images to a percentage op
4
specified by the sticking algorithm. It turns out that those requirements simSmallest feature
Field of view
Surface
Working distance
Lens
Sensor
Sensor size
Sensor size
(b)
(a)
Figure 1: Camera for inspection: (a) Camera setup in the field; (b) Relation between
parameters of the camera and the field.
plify our selection problem. The perpendicular requirement confines the camera
orientation to the normal of the inspected surface. The resolution requirement
suggests the computation of the field of view of the camera as:
af ov =
1
rc sf ,
2
(1)
where rc is the camera resolution (see Fig. 1). Taken the overlapping percentage
into account, the geometric primitive pi is then:
pi = (1 − op )af ov .
(2)
The working distance from the camera to the surface can also be computed as:
dk =
af ov f
,
ss
(3)
where f and ss are respectively the focal length and sensor size of the camera.
From (2) and (3), it is possible to determine configurations ci to cover the set
5
Inspection surfaces
Primitive geometry (pi)
Viewpoint
Figure 2: Generation of inspection viewpoints.
of primitives P , as illustrated in Fig. 2. Specifically, for each surface Pk ⊂ P , a
grid with the cell size of pi is first established to cover Pk . A working surface Pk∗ ,
parallel to Pk and distant dk from Pk , is then created. Projecting the center of
each cell of Pk to Pk∗ gives the position component of viewpoint ci . The normal
of Pk defines the orientation component of ci , which is supposed to be fully
controlled by the inspecting UAV so that it can be omitted in our computation.
2.2. Point-to-point pathfinding
Given the viewpoints, the shortest, obstacle-free path between every pair
of them need be found to form a graph for later processing. Without loss of
generality, different motion planning approaches such as roadmap, decoupling,
potential field and mathematical programming can be used here depending on
the UAV model and dynamic constraints [29, 30]. In this work, the hierarchical
decoupled approach is employed in which open- and closed-loop controllers operating at a variety of rate are linked together from top to bottom [29, 31, 32].
Since the majority of UAVs currently in production often already equipped with
an inner-loop tracking controller and a waypoint following system, this approach
can be simplified to a discrete search that produces a set of waypoints connecting
6
two viewpoints while avoiding obstacles. For this, the workspace is first divided
into a grid of voxels. Each voxel has the free or occupied status corresponding
to the presence or absence of an object in that voxel. In order to consider the
UAV as a particle moving without collision between voxels, all the free voxels
in a sphere of a radius equal to the largest dimension of the UAV are marked
as occupied. Thus, the A* algorithm [33] can be used to find the shortest path
between viewpoints. In each step, the cost to move from one voxel to another
surrounding neighbour is computed as:
L(α, β, γ) = a1 α2 + a2 β 2 + a3 γ 2 ,
(4)
where coordinates α, β, γ ∈ {−1, 0, 1} indicate the position of neighbor, and
coefficients a1 , a2 and a3 assign a particular weight to each direction. The total
cost to move from a voxel p to the viewpoint g at step n is given by:
f (p) =
n
X
Lk + kp − gk2 ,
(5)
k=1
where Lk is the motion cost at step k.
2.3. Modelling the IPP as a TSP
For given viewpoints and paths between them, a graph can be built to model
the IPP as an extended TSP. We define each viewpoint as a node, i, and the
path between two viewpoints as an edge, eij . The length, lij , of edge eij is the
cost to travel from node i to node j determined by (5). If the path between
node i and node j is blocked due to obstacles, a virtual path between them is
defined and a very large cost is assigned for the path. Denoting the set of all
nodes by V and the set of all edges by E, we restrict motion of the UAV to the
graph G = (V, E). The IPP task is now to find a tour, with a minimum cost,
that visits each node (viewpoint) exactly once, including the way back to the
initial node. Let T be the set of these nodes.
By associating a binary variable
1 if edge e ∈ E is in tour
ij
λij =
0 otherwise
7
(6)
with each edge in the graph, the IPP is then formulated as follows:
min
X
lij λij
(7)
λij = 2 ∀i ∈ V
(8)
eij ∈E
X
subject to
j∈V, i6=j
X
λij ≤ |T | − 1
∀T ⊂ V, T 6= ∅
(9)
i,j∈T, i6=j
λij ∈ {0, 1},
(10)
where |T | is the number of nodes in the tour. The objective function in (7)
defines the shortest tour. The constraint in (8) implies each node in the graph
has exactly one incoming edge and one outgoing edge, i.e., the tour passes
through each node once, while condition (9) ensures no sub-tours, i.e., the tour
returns to the original node after visiting all other nodes.
3. Enhanced Discrete Particle Swarm Optimization for Inspection
Path Planning
Particle swarm optimization (PSO), inspired by social behavior of bird flocking or fish schooling, is a population-based stochastic technique designed for
solving optimization problems [34]. In PSO, a finite set of particles is generated, each particle seeks the global optimum by moving and evolving through
generations. Initially, each particle is assigned to a random position and velocity. It then moves by updating its best previous position, Pk , and the best
position of the swarm, Gk . Let xk and vk be respectively the position and velocity of a particle at generation k. The position and velocity of that particle
in the next generation is given by:
vk+1 ← w.vk + ϕ1 r1 .(Pk − xk ) + ϕ2 r2 .(Gk − xk )
(11)
xk+1 ← xk + vk+1 ,
(12)
8
where w is the inertial coefficient, ϕ1 is the cognitive coefficient, ϕ2 is the social
coefficient, and r1 , r2 are random samples of a uniform distribution in the range
[0,1]. Equations (11) and (12) imply that the motion of a given particle is
a compromise between three possible choices including following its own way,
moving toward its best previous position, or toward the swarm’s best position.
The ratio between choices is determined by the coefficients w, ϕ1 , and ϕ2 .
3.1. DPSO approach to the IPP
Since the IPP defined in (7) – (10) is a discrete optimization problem, enhanced algorithms for discrete particle optimization (DPSO) will be developed
for our problem, motivated by [35]. For this, let us begin with an outline of our
approach to solve the IPP problem using DPSO with improvements in initialization, mutation, edge exchange and parallel implementation.
First, let define the position of particles as sequences of N + 1 nodes, all
distinct, except that the last node must be equal to the first one:
x = (n1 , n2 , ..., nN , nN +1 ), ni ∈ V, n1 = nN +1 ,
(13)
where N is the number of nodes, N = |V |. Since each sequence is a feasible
tour satisfying (8) and (9), to minimise the objective function (7) according to
(11) and (12), we need to define the velocity and numerical operators for the
particles’ motion.
From (12), it can be seen that a new position of a particle can be evolved
from the position of its current generation via the velocity operator, considered
here as a list of node transpositions:
v = ((ni,1 , nj,1 ), (ni,2 , nj,2 ), ..., (ni,kvk , nj,kvk )),
(14)
where ni , nj ∈ V and kvk is the length of the transposition list.
In DPSO, particle velocities and positions are updated by using the following
operations:
• The addition between a position x and a velocity v is found by applying
the first transposition of v to x, then the second one to the result, etc.
9
For example, with x = (1, 4, 2, 3, 5, 1) and v = ((1, 2), (2, 3)), by applying
the first transposition of v to x and keeping in mind the equality between
the first and last nodes, we obtain (2,4,1,3,5,2). Then applying the second
transposition of v to that result gives (3,4,1,2,5,3), which is the final result
of x + v.
• The subtraction between a position x2 and a position x1 is defined as the
velocity v, i.e., x2 − x1 = v, such that by applying v to x1 we obtain back
x2 .
• The addition between a velocity v1 and a velocity v2 is defined as a new
velocity, v1 ⊕ v2 = v, which contains the transpositions of v1 followed by
the transpositions of v2 .
• The multiplication between a real coefficient c with a velocity v is a new
velocity, c.v, defined as follows:
– For c = 0, c.v = ∅.
– For 0 < c ≤ 1, c.v = ((ni,1 , nj,1 ), (ni,2 , nj,2 ), ..., (ni,ckvk , nj,ckvk )).
– For c < 0 and c > 1, we omit these cases since they do not occur in
our DPSO.
3.2. Augmentations to the DPSO
In order to speed up the convergence and avoid being stuck in the local
minimum, we propose to enhance optimisation performance of the DPSO as
follows.
3.2.1. Deterministic initialization
The swarm in DPSO, having no prior knowledge of the searching space, is
initialized with its particles at random positions. This initialization works well
for a relatively small search space.
For large structure, the searching result depends, to a great extent, on the
initial positions of the particles. Therefore, in order to increase the probability
10
Figure 3: Initialization of particle using back-and-forth path.
of reaching the global optimum, we propose to exploit features of viewpoints
to generate several seeding particles to facilitate the evolution of the swarm
in the search space. In our application, viewpoints are generated based on
a grid decomposition. Consequently, a back-and-forth tour would generate a
near-optimal path, as shown in Fig. 3, if no obstacles occur. From this observation, positions are deterministically assigned for several particles during the
initialization process.
3.2.2. Random mutation
Similar to other evolutionary optimisation techniques such as the genetic
algorithm or ant colony system, the PSO performs both exploration and exploitation of the search space. Initially, particles are far from each other so they
explore different regions in the search space. After evolving through generations, the swarm converges and starts to make more exploitation. At this stage,
distances between particles will gradually reduce to the size termed ”swarm
collapse” [34], whereby many particles will become almost identical.
In order to avoid the collapse situation and keep the balance between exploration and exploitation, random mutations for particles are employed. After
11
1
3
2
1
4
2
3
4
Exchange edges
(2,6) and (3,7)
6
8
6
5
7
8
7
5
Figure 4: DPSO augmentation using edge exchange.
every i generations, identical particles are filtered. The remaining are then
sorted according to their cost values. Finally, only one-third of the smallest
particles are kept for the next generation. All others are disturbed, each in
different and randomly-chosen dimensions.
3.2.3. Edge exchange
The enhancement is based on the geometric feature for which crossing edges
can be exchanged to result in a shorter tour. Here, as 3D cross checking may
be difficult, a complete search similarly to the 2-opt algorithm is employed to
compare each valid combination of the swapping mechanism for edges [36]. In
this search, every possible exchange of edges is evaluated and the one with
the most improvement is chosen. Figure 4 illustrates the case when an edge
exchange between (2,6) and (3,7) to shorten the tour. Since this augmentation
is computational demanding, it should be used only when the random mutation
does not make any difference.
3.2.4. Parallel implementation on GPU
Owing to the rapidly increased performance with thousands of cores, a
graphics processing unit (GPU) can outperform the traditional CPUs for problems that are suitable for processed by SIMD (single instruction multiple data).
As our optimisation algorithms are also a SIMD-based, we can take this advantage to implement in parallel the proposed DPSO in GPUs to reduce computation time.
The diagram and pseudo code for parallel implementation are shown in Fig. 5
12
Initialization
Particle
Particle
Particle
Compute velocity
Particle
Random mutation
Parallel
Compute position
Update local best
Evaluate fitness
Edge exchange
Global best fitness
Figure 5: Parallel implementation of the DPSO on GPU.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
/* Host:
*/
Load w, ϕ1 , ϕ2 , swarm size to global memory;
Load graph of shortest paths and travelling costs to global memory;
Copy initalized particles from CPU to global memory;
/* see Fig.7 line 1-8 */
Set threads per block = swarm size;
Call kernels to evolve each particle in a seperate thread;
/* Device:
*/
Kernel move particle(*particles){
Get *particle corresponding to thread id;
Update position and f itness of particle;
/* see Fig.7 line 10-17 */
Update global best to global memory and synchronize threads;
}
Kernel random mutation(*particles){
Sort particles using thrust library;
Randomize 2/3 worst particles;
Update global best to global memory and synchronize threads;
}
Kernel edge exchange(*graph){
foreach i < (number of nodes − 2) do
foreach i < j < (number of nodes − 2) do
Swap nodes(i, j) and evaluate f itness;
end
end
Update global best to global memory and synchronize threads;
}
Figure 6: Pseudo code for parallel computation of DPSO on GPU.
13
and Fig. 6 respectively. After initialization, parameters of a particle such as
the velocity, position, and fitness are computed in parallel, each particle in a
different thread. At the end of each generation, the results are saved to the
global memory to update these particle parameters and then a new parallel
computation round starts.
In UAVs, parallel programs can be implemented in recent onboard computers
having good GPU capability and low power consumption such as Jetson TK1
with 192 CUDA Cores (Kepler GPU), 5 W [37]. The board can be configured as
either the main or supplemental board communicated with other components
via standard communications protocols like MAVLink. However, if the battery
power is highly limited as in some micro UAVs, an alternative solution is to
stream the sensory data to the ground control station (GCS) and utilise the
GPU of a laptop to conduct the path planning. The result is then uploaded to
the UAV via GCS for planning/re-planning and navigation.
3.3. Enhanced DPSO Pseudo Code
For vision-based inspection, to take into account obstacle avoidance of the
UAV, a selected combination of random and deterministic initialization for each
particle in the swarm is performed on a CPU while its evolutions, including
computation of updated particles’ velocity and position, random mutation and
edge exchange, are implemented in parallel on a GPU.
By making use of all advantages of the enhanced DPSO algorithm, the
pseudo code for our proposed algorithm incorporating the above-mentioned augmentations is shown in Fig. 7.
4. Experimental results
Experiments have been carried out on two real datasets recorded by laser
scanners mounted on a UAV for inspection of an office building and a concrete
bridge. The first dataset represents a floor of the building with a size of 25 m
× 12 m × 8 m. The second dataset represents a part of the bridge including
14
1
2
3
4
5
6
7
8
9
10
11
12
/* ------------------------ Computation on CPU ------------------------- */
/* Initialization:
*/
Set swarm parameter w, ϕ1 , ϕ2 , swarm size;
foreach particle in swarm do
Initialize particle’s position with 10% specific and 90% random;
Compute f itness value of each particle;
Set local best value of each particle to itself;
Set velocity of each particle to zero;
end
Set global best to the best fit particle;
/* ------------------------ Computation on GPU ------------------------- */
/* Evolutions:
*/
repeat
foreach particle in swarm do
Compute new velocity;
/* using Eq.11 */
Compute new position;
/* using Eq.12 */
Update f itness of new position;
if new f itness < local best then
local best = new f itness;
end
13
14
15
16
20
end
if current generation reaches collapsed cycle then
Sort all particles by f itness;
Randomize 2/3 worst particles;
21
end
22
Find the particle with the best f itness and update global best;
if global best not improved then
/* Edge exchange */
foreach particle in swarm do
Swap each pair of nodes and evaluate f itness;
Choose the swap with best fit;
17
18
19
23
24
25
26
end
27
28
29
30
/* Random mutation */
end
until max generation not reached and
global best not remaining unchanged for a pre-specified number of generations;
Figure 7: Pseudo code of the enhanced DPSO algorithm.
piers and surfaces with a size of 22 m × 10 m × 4.5 m. Figures 8a and 9a show
the datasets in point cloud representation. In order to apply the IPP algorithm
to the datasets, planar surfaces and boundaries need to extracted from them.
For this task, we have developed a software for automatic interpretation of
15
unordered point cloud data described in details in [38]. The software uses the
Random sample consensus (RANSAC) algorithm combined with data obtained
from an inertial measurement unit (IMU) to detect planar surfaces. The convex
hull algorithm is then employed to determine their boundaries. The remaining
point cloud is clusterized to obstacle objects by finding the nearest neighbour
in a 3D Kd-tree structure. Through the software, users are able to select the
surfaces they want to inspect, as shown in Fig. 8b and Fig. 9b, respectively.
In all experiments, coefficients w = 1, ϕ1 = 0.4, ϕ2 = 0.4 are chosen for the
DPSO. The number of particles is set to 100. The random mutation is executed
in every three generations and the edge exchange is carried out if the random
mutation does not improve the result. The parallel implementation is developed
based on the CUDA platform. The programs, including both serial and parallel
versions, are executed in a laptop computer with CoreTM i7 CPU and GeForce@
GTX 960M GPU.
4.1. Path Generation and DPSO Convergence
Figures 8c and 9c show the paths generated to inspect three selected surfaces
of each dataset. Figures 8d and 9d show the paths in the appearance of obstacles.
It can be seen that the back-and-forth pattern is dominant in those paths,
except essential changes when having obstacles or switching between surfaces.
Figures 8e and 8f present the front and side views of a zoom-in part of the
inspection path showing that obstacles were avoided. Figures 9e and 9f show
similar results for the bridge dataset.
Figure 10 shows the graphs of the fitness value as an objective function of
the generation number for the two inspection cases of a building and a bridge.
In each graph, the fitness represents the cost to traverse the inspection path.
From the dataset of the office building, the DPSO by solving the extended TSP
improves 22.2 % of the travelling cost and converges within 60 generations. For
the second dataset of the bridge, those numbers are 37.9 % and 80, respectively.
The difference is accounted for by the variation in size of the inspection surfaces
and the structural complexity of the environments. That is to say in terms of al16
(a)
(b)
(c)
(d)
(e)
(f)
Figure 8: Experiment with the dataset recording one floor of an office building: (a) Raw
data recorded by laser scanners ; (b) Detected planar surfaces and their boundaries; (c) Path
planning to inspect the surfaces of ceiling, left wall, and back wall; (d) Inspection path with
the appearance of obstacles; (e) Part of inspection path avoiding an obstacle (front view); (f)
Part of inspection path avoiding an obstacle (side view);
gorithms, care should be given when considering parameters for the exploration
(number of particles) and exploitation (number of generations).
4.2. Effect of the augmentations on the DPSO
Table 1 presents the effect of augmentations on the performance improvement over DPSO in percentage by applying our enhanced algorithm. Here, with
the dataset obtained from building inspection, the deterministic initialization
17
(a)
(b)
(c)
(d)
(e)
(f)
Figure 9: Experiment with the dataset recording a part of a bridge: (a) Raw data recorded
by laser scanners ; (b) Detected planar surfaces and their boundaries; (c) Path planning to
inspect the piers, top surface, and slope surface; (d) Inspection path with the appearance of
obstacles; (e) Part of inspection path avoiding obstacles at surface bottom; (f) Part of
inspection path avoiding obstacles at surface top;
significantly improves the processing time by 2.8 times and slightly improves
the travelling cost by 1.4 %. Notably, the computational efficiency in terms of
fast convergence actually comes from the improvement of evolving generations
of the swarm by means of initialization. On the other hand, it is not surprised
that the edge exchange introduces some enhancement on the travelling cost as
it uses brute force transpositions. Likewise, the parallel implementation introduces the most significant impact on the computation time thanks to the parallel
processing capability taking advantage of the SIMD feature of the DPSO.
To show consistency in the effectiveness of the proposed approach, we com18
5000
Fitness
Fitness
5500
3300
3200
3100
3000
2900
2800
2700
2600
2500
2400
4500
4000
3500
0
10
20
30
40
Generation
50
3000
60
0
20
40
Generation
(a)
60
80
(b)
Figure 10: DPSO Convergence from datasets of: (a) Office building; (b) Concrete bridge.
Table 1: Percent improvement of the DPSO by augmentations
Building dataset
Algorithm
Bridge dataset
Time
Travelling cost
Time
Travelling cost
(%)
(%)
(%)
(%)
Initialization
280
1.4
310
1.7
Random mutation
x
5.0
x
5.5
Edge exchange
x
13.8
x
15.7
Parallel on GPU
6570
x
6720
x
x: not applicable
pare our enhanced DPSO algorithm not only with the conventional DPSO but
also with an ant colony system (ACS), where the ACS is implemented as in
[39]. In the comparison, each algorithm was executed over 15 trials. Table 2
shows the results expressed in the average value and the standard deviation of
the processing time and the travelling cost. Compared with the ACS algorithm,
our enhanced DPSO for the bridge inspection dataset has shown on average an
improvement of 15% in the travelling cost and 87 times in the computation time.
Owing to a significant improvement in processing time, the enhanced DPSO can
be applied for real-time automated inspection.
19
Table 2: Comparison between the enhanced DPSO, DPSO and ACS algorithms.
Building dataset
Algorithm
Bridge dataset
Time (s)
Travelling cost
Time (s)
Travelling cost
Enhanced DPSO
32.9±1.2
2490.2±38.9
41.6±1.5
3358.5±59.7
DPSO
2253.8±25.2
2998.1±44.3
2928.4±33.8 4130.7±65.8
ACS
2560.1±16.3
2763.6±56.8
3617.2±23.4 3862.3±87.4
5. Conclusion
In this paper, we have presented an enhanced discrete particle optimisation (DPSO) algorithm for solving the inspection path planning (IPP) problem
that is formulated as an extended travelling salesman problem (TSP) considering simultaneously the coverage and obstacle avoidance. By augmenting with
deterministic initialization, random mutation, edge exchange and parallel implementation on GPU, the proposed DPSO can greatly improve its performance
in both time and travelling cost. The validity and effectiveness of the proposed
technique are verified in successful experiments with two real-world datasets
collected by UAV inspection of an office building and a concrete bridge. In a
future work, the algorithm will be extended for inspection of non-planar surfaces and incorporation of online re-planning strategies to deal with inspection
of built infrastructure of an irregular shape.
Acknowledgments
The first author would like to acknowledge an Endeavours Research Fellowship (ERF-PDR-142403-2015) provided by the Australian Government. This
work is supported by the University of Technology Sydney Data Arena Research
Exhibit Grant 2016 and Vietnam National University Grant QG.16.29.
20
References
References
[1] K.-W. Liao, Y.-T. Lee, Detection of rust defects on steel bridge coatings via
digital image recognition, Automation in Construction 71 (Part 2) (2016)
294 – 306. doi:10.1016/j.autcon.2016.08.008.
[2] C. M. Yeum, S. J. Dyke, Vision-based automated crack detection for bridge
inspection, Comp.-Aided Civil and Infrastruct. Engineering 30 (10) (2015)
759–770. doi:10.1111/mice.12141.
[3] G. Li, S. He, Y. Ju, K. Du, Long-distance precision inspection method for
bridge cracks with image processing, Automation in Construction 41 (2014)
83 – 95. doi:10.1016/j.autcon.2013.10.021.
[4] R. Adhikari, O. Moselhi, A. Bagchi, Image-based retrieval of concrete crack
properties for bridge inspection, Automation in Construction 39 (2014) 180
– 194. doi:10.1016/j.autcon.2013.06.011.
[5] R. Zaurin, R. Catbas, Integration of computer imaging and sensor data for
structural health monitoring of bridges, Smart Materials and Structures
19 (1). doi:10.1088/0964-1726/19/1/015019.
[6] A. A. Woods, H. M. La, Q. Ha, A novel extended potential field controller
for use on aerial robots, in: Proceedings of the 12th IEEE International
Conference on Automation Science and Engineering (CASE), 2016, pp.
286–291. doi:10.1109/COASE.2016.7743420.
[7] A. Ellenberg, A. Kontsos, F. Moon, I. Bartoli, Bridge deck delamination
identification from unmanned aerial vehicle infrared imagery, Automation
in Construction 72 (Part 2) (2016) 155 – 165. doi:10.1016/j.autcon.
2016.08.024.
[8] T. Bock, The future of construction automation: Technological disruption
and the upcoming ubiquity of robotics, Automation in Construction 59
(2015) 113 – 121. doi:10.1016/j.autcon.2015.07.022.
21
[9] Y. Yu, N. Kwok, Q. Ha, Color tracking for multiple robot control using a
system-on-programmable-chip, Automation in Construction 20 (2011) 669
– 676. doi:10.1016/j.autcon.2011.04.013.
[10] R. Montero, J. Victores, S. Martnez, A. Jardn, C. Balaguer, Past, present
and future of robotic tunnel inspection, Automation in Construction 59
(2015) 99 – 112. doi:10.1016/j.autcon.2015.02.003.
[11] B. Zhang, W. Liu, Z. Mao, J. Liu, L. Shen, Cooperative and geometric
learning algorithm (CGLA) for path planning of UAVs with limited information, Automatica 50 (3) (2014) 809 – 820. doi:10.1016/j.automatica.
2013.12.035.
[12] H. Choset, Coverage for robotics - a survey of recent results, Annals of
Mathematics and Artificial Intelligence 31 (2001) 113–126. doi:10.1023/
A:1016639210559.
[13] H. Choset, K. Lynch, S. Hutchinson, G. Kantor, W. Burgard, L. Kavraki,
S. Thrun (Eds.), Principles of Robot Motion: Theory, Algorithms, and
Implementation, The MIT Press, 2005.
[14] E. Acar, H. Choset, A. Rizzi, P. Atkar, D. Hull, Morse decompositions for
coverage tasks, International Journal of Robotics Research 21 (4) (2002)
331–344. doi:10.1177/027836402320556359.
[15] V. Shivashankar, R. Jain, U. Kuter, D. Nau, Real-time planning for covering an initially-unknown spatial environment, in: Proceedings of the
Twenty-Fourth International Florida Artificial Intelligence Research Society Conference, 2011, pp. 63–68, isbn: 978-1-57735-501-4.
[16] Y. Gabriely, E. Rimon, Spiral-stc: an on-line coverage algorithm of grid
environments by a mobile robot, in: Proceedings of the IEEE International
Conference in Robotics and Automation (ICRA), Vol. 1, 2002, pp. 954–960.
doi:10.1109/ROBOT.2002.1013479.
22
[17] B. Englot, F. Hover, Three-dimensional coverage planning for an underwater inspection robot, International Journal of Robotics Research 32 (9-10)
(2013) 1048–1073. doi:10.1177/0278364913490046.
[18] G. A. Hollinger, B. Englot, F. S. Hover, U. Mitra, G. S. Sukhatme, Active planning for underwater inspection and the benefit of adaptivity,
The International Journal of Robotics Research 32 (1) (2013) 3 – 18.
doi:10.1177/0278364912467485.
[19] P. Janousek, J. Faigl, Speeding up coverage queries in 3d multi-goal
path planning, in: Proceedings of the IEEE International Conference on
Robotics and Automation (ICRA), 2013, pp. 5082–5087. doi:10.1109/
ICRA.2013.6631303.
[20] P. Wang, K. Gupta, R. Krishnamurti, Some complexity results for metric view planning problem with traveling cost and visibility range, IEEE
Transactions on Automation Science and Engineering 8 (3) (2011) 654 –
659. doi:10.1109/TASE.2011.2123888.
[21] P. S. Blaer, P. K. Allen, View planning and automated data acquisition
for three-dimensional modeling of complex sites, Journal of Field Robotics
26 (11-12) (2009) 865–891. doi:10.1002/rob.20318.
[22] M. Saha, T. Roughgarden, J.-C. Latombe, G. Snchez-Ante, Planning tours of robotic arms among partitioned goals, The International
Journal of Robotics Research 25 (3) (2006) 207–223.
doi:10.1177/
0278364906061705.
[23] T. Danner, L. E. Kavraki, Randomized planning for short inspection paths,
in: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2000, pp. 971–976. doi:10.1109/ROBOT.2000.844726.
[24] B. Englot, F. Hover, Planning complex inspection tasks using redundant
roadmaps, in: H. I. Christensen, O. Khatib (Eds.), Robotics Research,
23
Vol. 100 of Springer Tracts in Advanced Robotics, Springer International
Publishing, 2016, pp. 327–343. doi:10.1007/978-3-319-29363-9_19.
[25] D. Applegate, W. Cook, A. Rowe, Chained lin-kernighan for large traveling
salesman problems, INFORMS Journal on Computing 15 (1) (2003) 82–92.
doi:10.1287/ijoc.15.1.82.15157.
[26] G. Papadopoulos, H. Kurniawat, N. Patrikalakis, Asymptotically optimal
inspection planning using systems with differential constraints, in: Proceedings of the IEEE International Conference in Robotics and Automation
(ICRA), 2013, pp. 4126–4133. doi:10.1109/ICRA.2013.6631159.
[27] G. Hollinger, B. Englot, F. Hover, U. Mitra, G. Sukhatme, Uncertaintydriven view planning for underwater inspection, in: Proceedings of the
IEEE International Conference in Robotics and Automation (ICRA), 2012,
pp. 4884–4891. doi:10.1109/ICRA.2012.6224726.
[28] P. Jimenez, B. Shirinzadeh, A. Nicholson, G. Alici, Optimal area covering
using genetic algorithms, in: Proceedings of the IEEE/ASME International
Conference on Advanced Intelligent Mechatronics, 2007, pp. 1–5. doi:
10.1109/AIM.2007.4412480.
[29] C. Goerzen, Z. Kong, B. Mettler, A survey of motion planning algorithms
from the perspective of autonomous uav guidance, Journal of Intelligent
and Robotic Systems 57 (1) (2009) 65. doi:10.1007/s10846-009-9383-1.
[30] N. Dadkhah, B. Mettler, Survey of motion planning literature in the
presence of uncertainty: Considerations for uav guidance, Journal of
Intelligent & Robotic Systems 65 (1) (2012) 233–246.
doi:10.1007/
s10846-011-9642-9.
[31] S. Scherer, S. Singh, L. Chamberlain, M. Elgersma, Flying fast and
low among obstacles: Methodology and experiments., The International
Journal of Robotics Research 27 (5) (2008) 549–574.
0278364908090949.
24
doi:10.1177/
[32] D. Jung, P. Tsiotras, On-line path generation for unmanned aerial vehicles
using b-spline path templates, Journal of Guidance, Control, and Dynamics
36 (6) (2013) 1642–1653. doi:10.2514/1.60780.
[33] P. Hart, N. Nilsson, B. Raphael, A formal basis for the heuristic determination of minimum cost paths, IEEE Transactions on Systems Science and
Cybernetics 4 (2) (1968) 100–107. doi:10.1109/TSSC.1968.300136.
[34] J. Kennedy, R. Eberhart, Y. Shi (Eds.), Swarm Intelligence, Morgan Kaufmann, 2001.
[35] M. Clerc, Discrete particle swarm optimization, illustrated by the traveling salesman problem, in: G. Onwubolu, B. Babu (Eds.), New Optimization Techniques in Engineering, Vol. 141 of Studies in Fuzziness and
Soft Computing, Springer Berlin Heidelberg, 2004, pp. 219–239. doi:
10.1007/978-3-540-39930-8_8.
[36] G. Gutin, A. Punnen (Eds.), The Traveling Salesman Problem and Its
Variations, Springer US, 2007. doi:10.1007/b101971.
[37] Y. Ukidave, D. Kaeli, U. Gupta, K. Keville, Performance of the nvidia jetson tk1 in hpc, in: 2015 IEEE International Conference on Cluster Computing (CLUSTER), 2015, pp. 533–534. doi:10.1109/CLUSTER.2015.147.
[38] M. Phung, C. Quach, D. Chu, N. Nguyen, T. Dinh, Q. Ha, Automatic interpretation of unordered point cloud data for UAV navigation in construction,
in: Proceedings of the 14th International Conference on Control, Automation, Robotics and Vision (ICARCV), 2016. doi:10.1109/ICARCV.2016.
7838683.
[39] M. Dorigo, L. M. Gambardella, Ant colony system: a cooperative learning
approach to the traveling salesman problem, IEEE Trans. Evolutionary
Computation 1 (1) (1997) 53–66. doi:10.1109/4235.585892.
25
| 2cs.AI
|
Control design and analysis of a stochastic network control system
Mohammad Soltani1 , Abhyudai Singh2
arXiv:1704.00236v1 [cs.SY] 1 Apr 2017
Abstract
A Network Control System (NCS) consists of control components that interact with the plant over a shared network. The
system dynamics of a NCS could be subject to noise arising from randomness in the times at which the data is transmitted
over the network, corruption of the transmitted data by the communication network, and external disturbances that might affect
the plant. A question of interest is to understand how the statistics of the data transmission times affects the system dynamics,
and under what conditions the system is stable. Another related issue is designing a controller that meets desired performance
specifications (e.g., a specific mean and variance of the system state). Here, we consider a minimal NCS that consists of a
plant and a controller, and it is subject to random transmission times, channel corruption and external disturbances. We derive
exact dynamics of the first two moments of the system, and use them to derive the stability conditions of the system. We
further design a control law that steers the system to a desired mean and variance. Finally, we demonstrate our results using
different examples, and show that under some specific conditions, randomness in the data transmission times can even reduce
the variability contributed from disturbance.
I. INTRODUCTION
Advent of modern information and communication technologies has provided the ability to design control systems whose
elements are located at remote distances. In such Network Control Systems (NCSs), different components of a feedback loop
communicate through a shared network [1]–[3]. The dynamics of a NCS is affected by noise arising from different sources.
For example, different components of a NCS are typically designed to be facultative so as to reduce the energy consumption
and use the shared network resources efficiently [4], [5]. Since the availability of the network at a given time depends on
whether or not other components of the NCS are using it, the times at which a specific device is able to communicate is
inherently random. The transmitted data can further be corrupted by the noise in the communication channel. Moreover, the
plant dynamics could be subject to external disturbances. Collectively, these noise sources hamper the overall performance of
the system. Previous works on NCS have dealt with designing control strategies to improve system performance. In particular,
for NCS with limited resources, different strategies are desired that reduce the number of times a control component needs
to communicate over the network. Examples of such strategies include event-triggered control, where transmission happens
when the state of the system meets a certain condition [6]–[8]; and self-triggered control, where the state of the system at the
time of a transmission is used to determine the next transmission time [9]–[11]. Further, motivated by specific requirements
of these systems, new network protocols are designed [12].
Here, we analyze the effect of various noise sources on stability of the system, and also design a control law that can drive
the system to a desired state. We consider a NCS whose states are modeled as Stochastic Differential Equations (SDEs).
These SDEs naturally capture the disturbances in the system. For example, if the disturbance is state independent (dependent)
then the corresponding drift and diffusion terms of the SDEs could be taken as state independent (dependent). To capture
the randomness in the transmission times, we model them using a renewal process. We provide the necessary and sufficient
conditions on stability of first- and second-order moments of the states of NCS under consideration. Further, we quantify
the contribution of each noise source to the states of the system. We show that the exact results combined with stability
conditions can be used to design controllers to meet desired performance criterion. Finally ,we demonstrate our method via
different examples. An interesting observation is that while rare and randomly transmission times are expected to increase
noise in a system, these might even reduce the noise in the system in some parameter regimes.
The rest of this paper is organized as following: In Section II network control systems are formulated. Section III provides
exact moments of these systems, and in Section IV and V examples of such systems are investigated. Finally conclusion
and direction of future works are presented in Section VI.
II. S TOCHASTIC CONTROL OF LINEAR SYSTEMS
Let the states of the system x ∈ Rn×1 evolve according to the following stochastic differential equation
dx = (â + Ax(t) + Bu(t)) dt + (C + Dx(t)1n )dwn ,
m×1
n×1
n×n
(1)
n×m
n×n
where u(t) ∈ R
denotes the controller; â ∈ R
is a constant vector; A ∈ R
, B ∈ R
, C ∈ R
and
D ∈ Rn×n are constant matrices; and 1n is a 1 × n unit matrix. In addition, wn is an n-dimensional Weiner process that
satisfies the following
hdwn i = 0, hdwn dw>
(2)
n i = In dt.
1 M.
Soltani is with Department of Electrical and Computer Engineering, University of Delaware, Newark, DE USA 19716. [email protected]
Singh is with the Department of Electrical and Computer Engineering, Biomedical Engineering, Mathematical Sciences, Center for Bioinformatics
and Computational Biology, University of Delaware, Newark, DE USA 19716. [email protected]
2 A.
𝐱
𝐮(𝒕+
s)
Controller
𝐮
Hold
Plant
𝐮 ⟼ 𝑘𝐱 + 𝜼
𝝉⟼0
𝐱
𝒕𝑠
𝑑𝝉 = 𝑑𝑡,
𝑑𝐱 = 𝑎 + 𝐴𝐱 + 𝐵𝐮 𝑑𝑡
𝒕𝑠
+(𝐶 + 𝐷𝐱𝕝𝑛 )𝑑𝒘𝑛 ,
𝑠 = {1,2, … }
𝑑𝐮/𝑑𝑡 = 0
𝑠 = {1,2, … }
Transmission times
Fig. 1. Model schematic of a stochastic NCS model. Left: The controller is far from the plant hence the feedback loop is connected through a network.
In between transmission times, plant uses the previous control law which is maintained in hold. Any time that connection occurs, the hold reads the new
control law which is calculated based on the current values of the states of the system. Right: Mathematical representation of network control system.
Resets here are the times in which connection occurs. Any time that a transmission occurs, new control law is applied to the plant. However due to
measurements error, an extra term η is added to the system.
Here In is an n × n Identity matrix, and the symbol h i denotes the expected value. While the first part in the right-hand
side of (1) determines dynamics of the plant, the second part represent the contribution of disturbance to the system. The
state-independent disturbance is modeled through C, and Dx(t)1n represents state-dependent disturbance.
The controller connects to the plant at random times ts , s ∈ {1, 2, . . .} as shown in Fig. 1. We assume that the time
intervals between transmission times
T s ≡ ts − ts−1
(3)
are independent and identically distributed (iid) random variables which follow a generic, continuous probability density
function f with support over non-negative real line. Whenever the controller communicates with the plant, the control law
reset as
−
m×n
u(t+
,
(4)
s ) 7→ Kx(ts ) + η, K ∈ R
+
where t−
s and ts denote the time just before and after a transmission, respectively. The matrix K here consists of controller
gains and the term η denotes the noise of the communication channel. We assume that η is a vector of zero-mean noise
terms and hηη > i = Σ ∈ Rm×m , where Σ is a diagonal matrix. The control law is assumed to remain constant in between
the consecutive transmission events, i.e.,
du
= 0.
(5)
dt
In order to obtain a mathematically tractable model, we introduce a timer τ that measures the time since the last
transmission. The timer increases with time in between transmissions, and resets to zero whenever the transmissions occur.
Let the probability that a transmission occurs in an infinitesimal time interval, (t, t + dt], be h(τ )dt. Then, we have that
h(τ ) ≡
1−
f (τ )
Rτ
.
f (y)dy
y=0
(6)
Alternatively, the duration between events T s follows a probability density function f given by
T s ∼ f (τ ) = h(τ )e−
Rτ
0
h(y)dy
(7)
[13], [14], and the timer follows the following steady-state probability density function
τ ∼ p(τ ) =
1 − R τ h(y)dy
e 0
,
hT s i
(8)
where hT s i is the mean time interval in between transmissions [15]. Modeling the timing of transmissions through the timer
enables us to investigate statistical moments of such systems as described in the next section.
III. S TATISTICAL MOMENTS OF NCS
We start our analysis by introducing a new vector that contains both states and controller y = [x u]> ∈ Rn+m×1 .
Dynamics of this vector is obtained from (1) and (5) as
dy = (ây + Ay y(t)) dt + (Cy + Dy y(t)1n+m )dwn+m ,
where
â ≡
â
0
, Ay ≡
A
0
B
0
, Cy ≡
C
0
0
0
, Dy ≡
D
0
0
0
(9)
, dwn+m ≡
dwn
0
.
(10)
Furthermore, at any time when the controller transmits a new control law, the states of y change as
In 0
+
−
,
hy(ts )i = Jy(ts ), J ≡
K 0
(11)
−
where we used the fact that the states of the system will not change during the events thus hx(t+
s )i = x(ts ).
A. Mean of NCS
In Appendix A, we show that the steady-state mean of vector y is bounded, if and only if the expected value
Z ∞
eAy T s =
f (τ )eAy τ dτ
(12)
0
exists and all of the eigenvalues of the matrix Jy eAy T s are inside the unit cycle, i.e.
eig Jy heAy T s i < 1.
(13)
In this limit, the steady-state mean of vector hxi = limt→∞ hxi is
+
*
Z
Z Ts
−1
Ay τ
Ay T s
−Ay r
−Ay τ
Ay T s
hyi = e
In+m − Jy e
Jy e
e
ây dr + e
*
e
Ay T s
Z
+
Ts
e
−Ay r
âdr
Z
∞
=
0
Z
Ay τ
f (τ ) e
0
e
ây dr ,
(14)
τ
e
−Ay r
ây dr dτ
(15)
0
is obtained by taking expected value with respect to T s , and
Z
Z ∞
Z τ
Aτ
Aτ
Aτ
−Ar
e
=
p(τ )e dτ,
e
e
âdr =
0
−Ay r
0
0
where the vector
τ
0
∞
p(τ ) e
Aτ
0
τ
Z
e
−Ar
âdr dτ
(16)
0
are derived by taking expected value with respect to τ . In the next part, we provide an approach to derive the second-order
moments of the system.
B. Second-order moments of NCS
Our strategy is to transform the second-order moments to a similar form as in (9) and (11). To that end, we introduce a
new vector
µ = [x u vec(xx> ) vec(xu> ) vec(uu> )],
(17)
where vec(M ) denotes the vector transformation of a matrix M . Vectorization is transforming a matrix into a column vector
by putting all the columns of the matrix into a vector subsequently. In appendix B, we show that dynamics of µ in between
the events is given by
µ̇ = âµ + Aµ µ,
(18)
where
â
0
>
vec
CC
âµ ≡
0
0
, Aµ ≡
A
0
M1
0
0
B
0
0
Im ⊗ â
0
0
0
M2
0
0
0
0
M3
Im ⊗ A
0
0
0
0
Im ⊗ B
0
,
(19)
M1 ≡In ⊗ â + â ⊗ In + D ⊗ C 1n + C > 1n ⊗ D,
M2 ≡In ⊗ A + A ⊗ In + nD ⊗ D,
M3 ≡In ⊗ B + B ⊗ In ,
and ⊗ denotes the Kronecker product. Furthermore, at the time of transmission, the states of the vector µ change as (see
Appendix C)
0
I 0
0
0 0
0 K
0
0
0 0
−
0 .
0
0
I
0
0
,
R
≡
(20)
hµ(t+
)i
=
J
µ(t
)
+
R
,
J
≡
µ
µ
µ
µ
s
s
0 0 K ⊗ In 0 0
0
0 0 K ⊗K 0 0
Σ
The deterministic dynamics in (18) and stochastic resets in (20) are similar to those in (9) and (11). Therefore, with a
similar analysis as in Appendix A, the steady-state mean of vector µ is bounded, if and only if all of the eigenvalues of the
matrix Jµ eAµ T s are inside the unit cycle. In this limit hµi = limt→∞ hµi is given by
+
*
!
Z Ts
Z τ
Aµ τ
Aµ T s −1
Aµ T s
−Aµ r
−Aµ τ
−Aµ r
e
e
hµi = e
In+m+n2 +m2 +nm − Jµ e
Jµ e
âµ dr + Rµ + e
âµ dr .
0
0
(21)
The last term in the above expression can be computed by taking expected value with respect to T s
+ Z
*
Z τ
Z Ts
∞
−Aµ r
Aµ T s
e−Aµ r âµ dr dτ. Likewise, weget
f (τ ) eAµ τ
e
âdr =
e
eAµ τ =
Z
∞
p(τ )eAµ τ dτ,
0
eAµ τ
(22)
0
0
0
Z
τ
e−Aµ r âµ dr
0
Z
=
0
∞
Z
p(τ ) eAµ τ
τ
e−Aµ r âµ dr dτ
(23)
0
by taking expected value with respect to τ . Mean of µ contains all the second-order moments of the vector x.
C. Control design from steady-state moments
We can design the controller for having desired steady-state mean in (14) by choosing K. This equation combined with
stability condition tells whether the desired mean is possible or not, i.e., as long as the eigenvalues of the matrix Jµ eAµ T s
are inside the unit cycle, we can design the matrix of control gains K to have the desired steady-state mean of x and finite
second-order moments.
Moreover, by having exact solutions of the second-order moments, we can also design a controller that minimizes the
steady-state variance of the system. We can set up this problem as an optimization problem
>
MinimizeK vec(hxx> i) − vec(hxi hxi ) ,
(24a)
Subject to hxi = Constant.
(24b)
Variance due to
disturbance
Variance due to noise in
channel
Mean of transmission time intervals 𝑻𝒔
Normalized variance of x
Variance of x
It may not always be feasible to obtain a control law that meets all desired specifications. This happens when the degrees
of freedom in the controller is less than number of mean and variance terms. We illustrate this by an example in the next
section wherein one can design the controller to achieve to a specific mean, but then controller is unable to change the
variance to a desired level.
Variance due to noise in
channel
Variance due to
disturbance
Noise in transmission time intervals 𝐶𝑉%&
Fig. 2. Making transmission times more random can reduce variance in x. Left) Surprisingly by increasing the mean time interval in between the
transmissions, noise in x contributed from disturbance can reduce. On the other hand increasing T s will increase the noise contributed from noisy channel
drastically. It is because the added noise by channel remains in the system for a longer time before getting corrected by a new control. Right) Interestingly
noisy transmission times also can reduce the variance contributed from disturbance. Hence when noise in channel is small, randomly distributed control
sequence can be used to reduce variance in x. For this plot we used gamma distributed time intervals and variance of x is normalized to its value at the
hT 2 i−hT i2
s
s
beginning of the plot. Noise in the inter transmission times is quantified by coefficient of variation squared CVT2 =
. The parameters are
hT s i2
selected as â = 1, a = −1, c = 0.45, b = 0.5, k = 0.5, and σ = 1. 95% confidence intervals are obtained by running 1000 numerical simulations.
IV. A SCALAR NETWORK CONTROL SYSTEM
Suppose that the state of a system is governed via a one dimensional SDE as
dx = (â + ax(t) + bu(t)) dt + c dw,
(25)
−
u(t+
s ) 7→ kx(ts ) + η,
(26)
and the control law at the time of resets is
where η is a zero mean noise term with variance σ 2 . Based on (9), the vector y = [x u]> is governed via
dy = (ây + Ay y(t)) dt + Cy dw2 ,
where
c 0
â =
, Ay =
, Cy =
.
(28)
0 0
1 0
Further, at the time of connection the states reset as (11) with J =
. Hence, based on (14), the mean of the state
−k 0
x in steady-state is
â
hxi = −
.
(29)
a + bk
Interestingly, the mean of x in steady-state is independent of T s . This implies that as long as the system is stable, having
rare transmissions of the data does not change the mean of the states.
â
0
a b
0 0
(27)
A. Second-order moment
In order to derive the second-order moments, we define
µ = [hxi hui hx2 i hxui hu2 i].
For this µ, â and Aµ in (19) are defined as
aµ =
â
0
c2
0
0
, Aµ =
a
0
2â
2b
0
(30)
0 0
0 0
0 0
.
a b
0 0
b 0
0 0
0 2a
â 0
0 0
(31)
Furthermore, at any time when the transmission occurs over the network, the states of the system reset to
hµ+ (ts )i = Jµ µ(ts ) + Rµ ,
where
Jµ =
1
k
0
0
0
0
0
0
0
0
0
0
1
k
k2
0
0
0
0
0
0
0
0
0
0
, Rµ =
(32)
0
0
0
0
σ2
.
Using these matrices, the non-zero eigenvalues of Jµ eAµ T s are given by
(
)
a eaT s + b eaT s k − bk a e2aT s (a + bk)2 + bk(bk − 2a eaT s (a + bk))
Aµ T s
eig Jµ he
i =
,
.
a
a2
(33)
(34)
If these eigenvalues are inside the unit circle, then mean and variance of the state x is given by (29) and
Fluctuations contributed from noisy channel
z
var(x) =
hx2 i
2
− hxi = σ
2
2 b (ahT s i
}|
{
− eaT s + 1) a2 ( eaT s + 1)hT s i + 2bk( eaT s (ahT s i − 1) + 1)
a3 hT s i2 (a + bk)(a(heaT s i + 1) + b(heaT s i − 1)k)
Fluctuations contributed from randomness in dynamics of the system
z
+ c2
−a4 (
eaT s
+ 1)hT s
i2
+
a2 bkhT
}|
{
eaT s hT s i + e2aT s − 1) + 2b2 k2 (−ahT s i + eaT s − 1)( eaT s (ahT s i − 1) + 1)
.
2a3 hT i2 (a + bk)(a heaT s i + 1) + b(heaT s i − 1)k)
(35)
s i(−2a
Interestingly, in some parameter regimes rare control of system (i.e., transmitting the control law after long times) can
result in lower variance of the system. To see this, suppose that dynamics of the system are noisy and, i.e., c is large and
fluctuations due to disturbance are dominant. In this limit, rare control of the system can reduce the noise as shown in Fig.
2. Note that the variance of x in this limit is less than
c2
(36)
a + bk
which is the variance for the case in which controller and plant are connected all the time. This means that rare transmissions
on the network not only saves resources such as bandwidth and energy, but also reduces the variance in x. Moreover, in
Fig. 2 we also illustrated the effect of noise in transmission times on variance of x. As expected, increasing noise in timing
of transmissions increases fluctuations that arises from noise in channel. However, again these noisy transmission times can
be used to reduce the effect of fluctuations contributed from disturbance.
It is important to point that the aforementioned scenario just occurs in specific parameter regimes. In most of the cases,
noise increases by increasing mean time intervals and randomness in timing of transmissions.
−
B. Control design
For this system, the non-zero eigenvalues of the matrix Jy eAT s are given by (34). Therefore, in order to have finite
mean of x, it is necessary and sufficient that
a e2aT s (a + bk)2 + bk(bk − 2a eaT s (a + bk))
a eaT s + b eaT s k − bk
< 1,
< 1,
(37)
a
a2
where | | denotes the absolute value. In addition, we can design the controller for having desired mean of the state by
choosing k as
â + ahxi
k=−
.
(38)
bhxi
It should be noted that the designed value of k should satisfy (37). Since k is a scalar and we just have one degree of
freedom in designing controller, we can only achieve the desired mean level.
C. The limit of a = 0
As an interesting and easy to follow limit, consider that a = 0. In this limit, the non-zero eigenvalue of Jy eAy T s is
1 + bkhT s i. Thus, in order to have a finite mean of x, the sign of k should be selected opposite to that of b, mean of inter
transmission times should be finite, and |1 + bkhT s i| < 1. In this limit, mean of x in steady state is
â
.
(39)
bk
Furthermore, to have finite second-order moment, in addition to a finite mean, the following eigenvalue also should be
inside the unit circle
b2 k 2 hT s i2 CVT2 + (bkhT s i + 1)2 .
(40)
hxi = −
It follows that in this limit, noisy transmission times may result in instability of the system. If this eigenvalue is inside the
unit circle then the variance of x is
Fluctuations contributed from noisy channel
z
var(x) = σ
2
2 −b(CVT
Fluctuations contributed from disturbance
}|
{ z
}|
{
2
3
2
2
+ 1)hT s i(3bkhT s i + 4) + b2 hT 3s i/hT s i
2 (bkhT s i(3b(CVT + 1)khT s i + 8CVT ) − 8) − bkhT s i/hT s i
+c
,
4k(bkhT s i + 2)
8bk(bkhT s i + 2)
(41)
where hT 3s i denotes the third-order moment of the time interval in between transmissions. While for nonzero a, variance of
x depends on the entire distribution of T s ; for a = 0, it only depends on the first three moments of T s .
In order to further simplify these results, we assume that the time intervals are log-normally distributed (hT 3s i = hT s i3 (1+
CVT2 )) and noise in between time intervals is small, i.e. CVTi ≈ 0, i > 2. In this limit, variance of x simplifies to
Fluctuations contributed from noisy channel
Fluctuations contributed from disturbance
z
}|
}|
{ z
{
bCVT2 hT s i
bhT s i
CVT2 hT s i bkhT s i − 2
2
2
var(x) = σ −
−
+c
+
.
k(bkhT s i + 2)
2k
bkT + 2
4bk
(42)
Given the fact that for having a finite mean one of the parameters b and k should be negative, then clearly in this limit,
increasing mean transmission times or noise in transmission times increases variance of x. In the next section, we illustrate
our results via a two dimensional system.
D. A two states NCS example
Consider a NCS with two states x = [x1 x2 ]> that is governed by the following dynamics
dx = (â + Ax(t) + Bu) + Cdw2 ,
where
â =
a1
0
, A=
−γ1
a2
0
−γ2
1
0
, B=
0
1
(43)
√
, C=
a1
0
0
0
.
(44)
This example is motivated from biochemical reactors. Assume that x1 and x2 are levels of species 1 and 2, respectively.
Then, the system under consideration can be interpreted as having production of molecules of species 1 at a constant rate
a1 . The molecules species 2 are produced from species 1, i.e., its production rate is a2 x1 . The production of species x1 can
be assumed to be noisy as the dynamics corresponds to the Langevin approximation of this reaction [16]. Finally, molecules
of x1 and x2 degrade with rates γ1 and γ2 , respectively.
This biochemical reactor is controlled through a shared network. Any time that transmission happens, the control law
changes as
−k1 −k2
+
−
hu(ts )i = Kx(ts ), K =
.
(45)
0
−k3
There could be several ways to control this reactor. For example, a UV radiation that increases death rate of molecules can
be used [17], [18]. This control law is implemented by manipulating k1 and k3 . Another possibility is that the resources
needed to produce species 1 are controlled based on levels of species 2 through the parameter k3 . Such negative feedback
loops are common motifs in biological systems [19]–[21].
By introducing y = [x u]> ∈ Rn+m×1 , this system can be written in the form of (9). Hence the methods explained in
this paper can be applied, which results in
hx1 i =
a1 (γ2 + k3 )
γ2 + k3
hx1 i.
, hx2 i =
a2 k2 + γ1 (γ2 + k3 ) + γ2 k1 + k1 k3
a2
(46)
Note that we can select k1 , k2 , and k3 to have any desired mean level. Moreover, we have one degree of freedom (two
means and three parameters); hence, we can use this degree of freedom to minimize one of the terms in the vector of second
order moments. We do not show the second-order moments of x due to space constraints.
V. C ONCLUSION
We studied statistical moments of a network control system in which transmission times in between the controller and the
plant are randomly distributed. We derived exact solution of the mean and the second-order moments as well as the stability
conditions. We showed that these results can be used to design controllers that maintain mean of the states at desired levels.
We demonstrate our method on different examples. Surprisingly, as long as mean of the states of the plant is finite, mean of
the state of the system is independent of transmission times statistical moments. In addition, we observed that under specific
parameter regimes, rare transmissions not only save resources of the system, but may also reduces noise. Further, we showed
that for a fixed mean of transmission time intervals, noisy transmission times can be used to reduce the fluctuations in x.
Future work will extend the method explained here to network control systems in which more than just two parts are
connected via a network, i.e. a system with more than one set of random transmissions. Another avenue of research would
be the systems whose states switch between sub-systems, depending upon the values of the state. Finally, it would also be
interesting to extend this work to analyze network control systems with nonlinearities in their dynamics. While for nonlinear
systems exact solutions may not be available, moment closure methods can be used to obtain approximate analytical solutions
[22]–[25].
A PPENDIX
A. Proof of equation (14)
+
Using (9), the states of SHS right before sth event y(t−
s ) is related to y(ts−1 ) as
Z Ts
Z Ts
+
−
Ay Ts
−Ay r
A y Ts
y(ts ) =e
e
ây dr + e
y(ts−1 ) +
(Cy + Dy y(t)1n+m )dwn+m .
0
Thus the mean of the states after sth event is
*
hy(t+
s )i
(47)
0
=Jy
e
Ay T s
Z
+
Ts
e
0
−Ay r
âdr
+ Jy eAy T s
y(t+
s−1 ) .
(48)
Here we used the fact that the Ito integral
*Z
Ts
(Cy + Dy y(t)1n+m )dwn+m
+
=0
(49)
0
[26]. Hence from (48), the mean of the states right after an event in steady-state (s → ∞) exists if and only and if
eigenvalues of Jy eAy T s are inside the unite circle. In this limit the steady-state mean of the states right after an event
hy|τ =0 i can be written as
*
+
Z Ts
Ay T s −1
Ay T s
−Ay r
hy|τ =0 i = In+m − Jy e
Jy e
e
ây dr .
(50)
0
By using equation (50), the mean of the states in between events for any given τ is
*
+
Z Ts
Z
Ay τ
Ay T s −1
Ay T s
−Ay r
e
hy|τ =τ i =e
In+m − Jy e
J e
ây dr + eAy τ
0
τ
e−Ay r ây dr.
(51)
0
The mean of the states is obtained by unconditioning (51) with respect to τ using (8) as shown in (14).
B. Dynamics of second-order moments of NCS
We start by calculating dynamics of xx> in between random transmissions. Using (1), the time evolution of xx> is
obtained as
dxx>
> >
> >
= Axx> + xx> A> + Bux> + xu> B > + âx> + xâ> + CC > + C 1>
n x D + Dx1n C + nDxx D .
dt
By vectorization, we can rewrite (52) as
dvec xx>
= (In ⊗ A + A ⊗ In + nD ⊗ D) vec xx> + (In ⊗ B + B ⊗ In ) vec xu> + vec CC >
dt
+ (In ⊗ â + â ⊗ In + D ⊗ C 1n + C > 1n ⊗ D)x,
(52)
(53)
where we used the fact that for three matrices M , M 0 , and M 00
vec(M M 0 M 00 ) = (M 00> ⊗ M )vec(M 0 )
[27]. Since these dynamics depend on xu
>
we also add dynamics of xu
(54)
>
dxu>
=Axu> + Buu> + âu> .
dt
Again by using vectorization we have
dvec xu>
=(Im ⊗ A)vec xu> + (Im ⊗ B)vec uu> + (Im ⊗ â)u
dt
and finally we add dynamics of vec uu> by using (5)
dvec uu>
= 0.
dt
The combination of (53), (56), and (57) results in (18).
(55)
(56)
(57)
C. Second-order moments of NCS after transmissions
Based on (4)
> −
>
> +
> −
huu> (t+
s )i = Kxx (ts )K + Σ ⇒ vechuu (ts )i = K ⊗ Kvec(xx (ts )) + vec(Σ).
Further
hxu(t+
s )i
(58)
can be written as
> +
>
> +
> −
hxu> (t+
s )i = xx (ts )K ⇒ vechxu (ts )i = K ⊗ In vec(xu (ts )).
(59)
Overall we have
−
hµ(t+
s )i = Jµ µ(ts ) + Rµ .
(60)
where
Jµ ≡
In
0
0
0
0
0
K
0
0
0
0
0
In+m
K ⊗ In
K ⊗K
0
0
0
0
0
0
0
0
0
0
, Rµ ≡
â
0
0
0
vex (Σ)
.
(61)
Note that in deriving Jµ we used the fact that values of x before and after transmission are equal hence
+
> +
> +
− > −
hxx> (t+
s )i − hx(ts )ihx (ts )i = 0 ⇒ vechxx (ts )i = vec(x(ts )x (ts )).
(62)
R EFERENCES
[1] Hespanha JP (2005) A model for stochastic hybrid systems with application to communication networks. Nonlinear Analysis: Theory, Methods &
Applications 62: 1353–1383.
[2] Hespanha JP (2004) Stochastic hybrid systems: Applications to communication networks. In: Alur R, Pappas GJ, editors, Hybrid Systems: Computation
and Control, Springer-Verlag. pp. 387-401.
[3] Bohacek S, Hespanha JP, Lee J, Obraczka K (2003) A hybrid systems modeling framework for fast and accurate simulation of data communication
networks. In: Proc. of the ACM Int. Conf. on Measurements and Modeling of Computer Systems (SIGMETRICS). volume 31, pp. 58-69.
[4] Liberzon D, Nesic D (2007) Input-to-state stabilization of linear systems with quantized state measurements. IEEE Transactions on Automatic Control
52: 767-781.
[5] Gommans T, Antunes D, Donkers T, Tabuada P, Heemels M (2014) Self-triggered linear quadratic control. Automatica 50: 1279-1287.
[6] Tabuada P (2007) Event-triggered real-time scheduling of stabilizing control tasks. IEEE Transactions on Automatic Control 52: 1680-1685.
[7] Heemels WPMH, Johansson KH, Tabuada P (2012) An introduction to event-triggered and self-triggered control. In: 51st IEEE Conference on
Decision and Control (CDC). pp. 3270-3285.
[8] Antunes DJ, Khashooei BA (2016) Consistent event-triggered methods for linear quadratic control. In: 2016 IEEE 55th Conference on Decision and
Control (CDC). pp. 1358-1363.
[9] Anta A, Tabuada P (2010) To sample or not to sample: Self-triggered control for nonlinear systems. IEEE Transactions on Automatic Control 55:
2030-2042.
[10] Wang X, Lemmon MD (2009) Self-triggered feedback control systems with finite-gain calL2 stability. IEEE Transactions on Automatic Control 54:
452-467.
[11] Khashooei BA, Antunes DJ, Heemels WPMH (2017) Output-based event-triggered control with performance guarantees. IEEE Transactions on
Automatic Control : To appear.
[12] Dacic D, Nesic D (2008) Observer design for wired linear networked control systems using matrix inequalities. Automatica 44: 2840 - 2848.
[13] Ross SM (2010) Reliability theory. In: Introduction to Probability Models, Academic Press. Tenth edition, pp. 579 - 629.
[14] Evans M, Hastings N, Peacock B (2000) Statistical Distributions. Wiley, 3rd edition.
[15] Vargas-Garcia CA, Soltani M, Singh A (2016) Conditions for cell size homeostasis: A stochastic hybrid systems approach. IEEE Life Sciences Letters
.
[16] Gillespie DT (2000) The chemical Langevin equation. Journal of Chemical Physics 113: 297–306.
[17] Kobayashi H, Krn M, Araki M, Chung K, Gardner TS, et al. (2004) Programmable cells: Interfacing natural and engineered gene networks. Proceedings
of the National Academy of Sciences 101: 8414-8419.
[18] Szymaska P, Gritti N, Keegstra JM, Soltani M, Munsky B (2015) Using noise to control heterogeneity of isogenic populations in homogenous
environments. Physical Biology 12: 045003.
[19] Borri A, Palumbo P, Singh A (2016) The impact of negative feedback in metabolic noise propagation. IET Systems Biology : 179–186.
[20] Singh A (2011) Negative feedback through mRNA provides the best control of gene-expression noise. IEEE Transactions on Nanobioscience 10:
194–200.
[21] Soltani M, Vargas-Garcia CA, Kumar N, Kulkarni R, Singh A (2015) Approximate statistical dynamics of a genetic feedback circuit. Proc of the
2015 Amer Control Conference, Chicago, IL : 4424-4429.
[22] Singh A, Hespanha JP (2011) Approximate moment dynamics for chemically reacting systems. IEEE Transactions on Automatic Control 56: 414-418.
[23] Singh A, Hespanha JP (2010) Stochastic hybrid systems for studying biochemical processes. Philosophical Transactions of the Royal Society A 368:
4995-5011.
[24] Soltani M, Vargas-Garcia CA, Singh A (2015) Conditional moment closure schemes for studying stochastic dynamics of genetic circuits. IEEE
Transactions on Biomedical Systems and Circuits 9: 518-526.
[25] Singh A, Hespanha JP (2005) Models for multi-specie chemical reactions using polynomial stochastic hybrid systems. In: Proc. of the 44th IEEE
Conf. on Decision and Control, Seville, Spain. pp. 2969-2974.
[26] Øksendal B (2003) Stochastic differential equations. Springer.
[27] Macedo HD, Oliveira JN (2013) Typing linear algebra: A biproduct-oriented approach. Science of Computer Programming 78: 2160 - 2191.
| 3cs.SY
|
Biological Mechanisms for Learning: A Computational Model of Olfactory
Learning in the Manduca sexta Moth, with Applications to Neural Nets
arXiv:1802.02678v1 [q-bio.NC] 8 Feb 2018
Charles B. Delahunt∗, Jeffrey A. Riffell†, and J. Nathan Kutz‡
Abstract. The insect olfactory system, which includes the antennal lobe (AL), mushroom body (MB), and
ancillary structures, is a relatively simple neural system capable of learning. Its structural features,
which are widespread in biological neural systems, process olfactory stimuli through a cascade of
networks where large dimension shifts occur from stage to stage and where sparsity and randomness
play a critical role in coding. Learning is partly enabled by a neuromodulatory reward mechanism
of octopamine stimulation of the AL, whose increased activity induces rewiring of the MB through
Hebbian plasticity. Enforced sparsity in the MB focuses Hebbian growth on neurons that are the
most important for the representation of the learned odor. Based upon current biophysical knowledge, we have constructed an end-to-end computational model of the Manduca sexta moth olfactory
system which includes the interaction of the AL and MB under octopamine stimulation. Our model
is able to robustly learn new odors, and our simulations of integrate-and-fire neurons match the
statistical features of in vivo firing rate data. From a biological perspective, the model provides
a valuable tool for examining the role of neuromodulators, like octopamine, in learning, and gives
insight into critical interactions between sparsity, Hebbian growth, and stimulation during learning. Our simulations also inform predictions about structural details of the olfactory system that
are not currently well-characterized. From a machine learning perspective, the model yields bioinspired mechanisms that are potentially useful in constructing neural nets for rapid learning from
very few samples. These mechanisms include high-noise layers, sparse layers as noise filters, and a
biologically-plausible optimization method to train the network based on octopamine stimulation,
sparse layers, and Hebbian growth.
1. Introduction.
Learning is a vital function of biological neural networks, yet the underlying biomechanical
mechanisms responsible for robust and rapid learning are not well understood. The insect
olfactory network, and the moth’s olfactory network (MON) in particular (e.g. the Manduca
sexta moth), is a comparatively simple biological neural network capable of learning [1, 2], and
makes an ideal model organism for characterizing the mechanics of learning. It is amenable
to interrogation through experimental neural recordings of key, well-understood structural
components including the antenna lobe (AL) [3] and mushroom body (MB) [4]. In addition,
the AL-MB contain many structural motifs that are widespread in biological neural systems.
These motifs include: (i) the use of neuromodulators (octopamine and dopamine) in learning
[5], (ii) a cascading networks structure [6], (iii) large changes in dimensionality (i.e. numbers
of neurons) between networks [7], (iv) sparse encodings of data in high-dimensional networks
[8], (v) random connections [9], and (vi) the presence of noisy signals [10]. Bio-inspired design
principles suggest that each of the features has high value to the olfactory system. The
mechanism of octopamine/dopamine release during learning is of particular interest, since it
is not well-understood how this stimulation promotes the construction of new sparse codes in
the MB.
∗
Department of Electrical Engineering, University of Washington, Seattle, WA, USA. [email protected]
Department of Biology, University of Washington, Seattle, WA, USA. [email protected]
‡
Department of Applied Mathematics, University of Washington, Seattle, WA, USA. [email protected]
†
1
In this work, we build a computational model of the MON that is closely aligned with both
the known biophysics of the moth AL-MB and in vivo neural firing rate data, and that includes
the dynamics of octopamine stimulation. There exist computational neural network models
of the insect MB [11, 12, 13, 14, 15] inspired by but not closely tied to particular organisms
(indeed, [11] points out the advantages of this more general approach). Our goal here is to
analyze the interactions of AL, MB, and octopamine during associative learning. To this end,
we model the architecture and neural dynamics of the whole system: AL, MB, octopamine
stimulation, synaptic growth, odor projection onto the AL, and an extrinsic (readout) neuron
downstream. We model neurons as “integrate-and-fire” units [16]. In addition, we tether the
model closely to the known biophysics of the M. sexta moth. This allows us to calibrate the
model’s behavior to in vivo neural firing rate data from moth AL neurons, collected during
odor and octopamine stimulation.
We thus create a full, end-to-end neural network model (hereafter “Network Model”) that
demonstrates robust learning behavior while also tightly matching the structure and behavior
of a real biological system. This approach has three advantages: (i) we can meaningfully
compare Network Model simulation output to experimental data in order to tune model
parameters, (ii) simulation results can map back to the original biological system in order
to render meaningful insights, and (iii) each structural and dynamical feature of the system
can demonstrate its relevance to the task of learning in a neural net. In silico experiments
with the Network Model allow us to abstract out critical features in the moth’s toolkit, study
how these features enable learning in a neural network, and thus derive bio-inspired insight
into the mathematical framework that enables rapid and robust learning. Specifically, our
experiments elucidate mechanisms for fast learning from noisy data that rely on cascaded
networks, sparsity, and Hebbian plasticity.
These mechanisms have potential applications to engineered neural nets. NNs have
emerged as a dominant mathematical paradigm for characterizing neural processing and learning. This is not surprising given that they were inspired by the Nobel prize winning work of
Hubel and Wiesel on the primary visual cortex of cats [17]. These seminal experiments showed
that networks of neurons were organized in hierarchical layers of cells for processing visual
stimulus. The first mathematical model of a neural network, the Neocognitron in 1980 [18],
had many of the characteristic features of today’s deep neural networks (DNNs). However,
many of the biological motifs listed above are largely absent from engineered NNs, whose
principles and building blocks are biologically implausible even as DNNs have achieved great
success [19, 20]. For example, the AL-MB interaction with octopamine and Hebbian plasticity
operates in a fundamentally different manner than the backpropagation optimization used in
DNNs. We seek to characterize a set of biological elements, a “biological toolkit”, that can
be assembled into NNs that operate on fundamentally different principles than standard NNs
and that are capable of rapid and robust learning from very few training samples, an ability
common in biological NNs that remains challenging for today’s DNNs.
A brief summary of the AL-MB network: It is organized as a feed-forward cascade of five
distinct networks, as well as a reward mechanism [21, 22]. Roughly 30,000 noisy chemical
receptor neurons (RNs) detect odor and send signals to the Antenna Lobe (AL) [6]. The AL
acts as a pre-amplifier, providing gain control and sharpening of odor representations [23]. It
contains roughly 60 isolated units (glomeruli), each focused on a single odor stimuli feature
2
[21]. Glomeruli laterally inhibit each other, and project odor codes to the Mushroom Body
(MB). AL neurons are noisy [10]. The MB contains about 4000 Kenyon Cells (KCs). These
fire sparsely and encode odor signatures as memories [24, 4, 8]. MB sparsity is enforced by
global inhibition from the Lateral Horn [25]. Extrinsic Neurons (ENs), numbering ∼10’s, are
believed to be “readout neurons” that interpret the KC codes [26, 27]. In response to reward
(sugar at the proboscis), a large neuron sprays octopamine globally over the AL and MB,
causing generalized stimulation of neurons. Learning does not occur without this octopamine
input [28, 1]. The connections into the KCs (AL→KCs) and out of the KCs (KCs→ENs) are
plastic during learning [29, 6]. Figure 1.1 gives a system schematic (A) along with typical
firing rate (FR) timecourses (from simulation) for neurons in each network (B). More network
details are given in Methods.
2. Results.
We first show the calibration of our Network Model to in vivo data. We then describe neural
behaviors of the Network Model during learning and give results of learning experiments.
Finally, we give results of experiments on MB sparsity.
2.1. Calibration of Model.
The Network Model was calibrated to behave in a statistically similar way to three sets
of in vivo data measuring projection neuron (PN) firing rate (FR) activity in the AL (See
Methods for details): (i) PN spike counts with odor but without octopamine: 129 units
with FR>1 spike/sec, (ii) PN spike counts with odor and with octopamine: 180 units with
FR>1 spike/sec, and (iii) PN spike counts with odor, with and without octopamine: 52 units
with FR>1 spike/sec. Due to the limited number of experimental units, only qualitative
comparisons of the model and experiment could be made. Excessive tuning of the model
parameters would have served only to overfit the particular data, rather than matching true
PN behavior distributions or, more importantly, the general learning behavior of the moth.
Fig 2.1 shows the close match of typical Network Model PN statistics to in vivo PN statistics
based on mean (µ) and variance (σ) of spontaneous FRs and odor responses, both without
and with octopamine (details of metrics are given in Methods). Importantly, Fig 2.1 shows
significant octopamine-modulated increase in PN FR activity in the Network Model, consistent
with in vivo experiments involving octopamine stimulation.
There is limited experimental data measuring the FR activity of Kenyon cells (KC) in
the MB, and no data to our knowledge measuring KC in response to octopamine stimulation.
However, we note that the behavior of KCs during the application of octopamine to the AL,
either with or without odor, is not an artifact of parameter tuning. Rather, it follows from
the tuning the AL to match in vivo data. Specifically, PN FRs at baseline (with no odor or
octopamine), with odor alone, with octopamine alone, and with odor and octopamine, are
all determined by calibration of the model to in vivo data. KCs respond only to PNs and to
inhibition from the LH (See Fig 1.1). Calibrating the KC baseline response in the absence of
octopamine to in vivo data in [30] fixes the feed-forward connections from PNs. Assumed in
this model, due to lack of biophysical evidence, is that octopamine has no direct effect on KC
FRs. Thus KC behavior with octopamine is fully determined once the model is tuned to PN
data. This completes the calibration process of our model parameters. As Fig 2.1 shows, the
3
Figure 1.1. AL-MB overview. A: System schematic: Chemical sensors (RNs) excite a noisy pre-amp
network (AL), which feeds forward to a plastic sparse memory layer (MB), which excites readout (decision)
neurons (ENs). Green lines show excitatory connections, red lines show inhibitory connections (LH inhibition
of the MB is global). Light blue ovals show plastic synaptic connections into and out of the MB.
B: Neuron timecourse outputs from each network (typical simulation) with time axes aligned vertically. Timecourses are aligned horizontally with their regions-of-origin in the schematic. The AL timecourse shows all
responses within ± 2.5 std dev of mean spontaneous rate as medium blue. Responses outside this envelope are
yellow-red (excited) or dark blue (inhibited). MB responses are shown as binary (active/silent). Timecourse
events are as follows: (1) A period of no stimulus. All regions are silent. (2) Two odor stimuli are delivered,
3 stimulations each. AL, MB, and ENs display odor-specific responses. (3) A period of control octopamine, ie
without odor or Hebbian training. AL response is varied, MB and EN are silent. (4) The system is trained
(octopamine injected) on the first odor. All regions respond strongly. (5) A period of no stimulus. All regions
are silent, as in (1). (6) The stimuli are re-applied. The AL returns to its pre-trained activity since it is
not plastic. In contrast, the MB and EN are now more responsive to the trained odor, while response to the
untrained odor is unchanged. Green dotted line in the EN represents a hypothetical “action” threshold. The
moth has learned to respond to the trained odor.
model agrees well with in vivo experiment.
There are no bulk data, to our knowledge, measuring EN firing rates in response to odors
and/or octopamine. However, calibrating EN response is not necessary to demonstrate an
ability to learn. The key marker is post-training increase in EN response.
4
Figure 2.1. In vivo firing rate data and model calibration: Comparison of PN firing rate activity
from in vivo data and simulations. A: Histograms and CDFs of in vivo data and simulations. Col a: mean
spontaneous FRs µs . Col b: σs /µs of spontaneous FRs, a measure of noisiness of a PN. Col c: odor response,
measured as distance from µs in σs units. Distance > 2σs implies a strong activation/inhibition. Col d: odor
response during octopamine, in σs units distance from µs . Note that PN responses are broadened (i.e. more
PNs are strongly activated or inhibited). The dotted line in the CDF inset is the same as the CDF of the
odor response without octopamine, to show the broadening towards both extremes. Col e: change in mean
spontaneous FRs due to octopamine, measured in σs units distance from (non-octopamine) µs . Some PNs are
excited, some are inhibited.
B: Activity of PNs indexed by increasing spontaneous FR. Blue lines = mean spontaneous FRs µs (cf col 1).
Shaded regions = σs , 2σs envelopes (cf col 2). Solid red dots = odor response FRs (cf col 3). Hollow red dots
= odor response FRs during octopamine (cf col 4). Red lines show the change in odor response FRs due to
octopamine (cf broadened response). Black stars (*) = spontaneous FRs during octopamine (cf col 5). In panel
A cols 3, 4, 5, the x-axes are expressed in units of σs , while in panel B the y-axis measures raw spikes/sec FR.
2.2. Learning Experiments: PN and KC Behaviors.
PN activity in the AL, and KC activity in the MB, from typical Network Model simulations
are shown in Fig 1.1 as heatmaps, evolved over the time course of a simulation in which the
system was exposed to two different odors and trained on one of them. The AL is stimulated
with octopamine during training. Each row of the heatmap represents a distinct PN or KC
as it evolves in time (left to right columns of heat map). All the timescales are aligned.
5
PNs. In the AL heatmap, the light blue region corresponds to PN FRs within 2.5σs of
their respective mean spontaneous FRs µs , warm colors correspond to very high FRs, and
dark blues correspond to strongly inhibited FRs. The simulations demonstrate a number
of key PN behaviors, including (i) absent odor, PN FRs stay within their noise envelopes
(by definition), (ii) the two odors have distinct excitation/inhibition signatures on PNs, (iii)
octopamine alone (without odor) results in more PNs being excited beyond their usual noise
envelopes, and also results in some PNs being inhibited beyond their usual envelopes, (iv)
octopamine and odor, applied together, result in an overall excitation of PNs, and (v) the AL
behavior returns to baseline after octopamine is withdrawn, since AL connection weights are
not plastic. This last observation is in contrast to what occurs in the MB.
KCs. In the MB, the KCs fire sparsely due to global inhibition from the Lateral Horn.
The only plastic connections in the AL-MB system involve the KCs: Between PNs (and QNs)
and KCs (M P K , M QK in Methods Section); and between KCs and extrinsic readout neurons
(ENs) (M KE in Methods Section). Thus the KC odor signatures are modulated with training.
Black regions indicate FRs < 1 spike/sec, white regions indicate FRs > 1 spike/sec. The white
regions have been dilated to make the sparsely-firing KCs easier to see.
The simulations demonstrate a number of key KC behaviors, including (i) the baseline KC
FR response absent any odor is essentially zero, (ii) the two odors excite distinct sets of KCs
with varying consistency from noise trial to noise trial, (iii) for a given odor, some KCs fire
reliably in response to an odor stimulation and some fire only occasionally, (iv) when subject
to octopamine but no odor, KCs are unresponsive, a benefit during learning since it prevents
environmental noise from being encoded as meaningful, (v) when subject to octopamine and
odor, KCs respond strongly to the odor with high trial-to-trial consistency, and (vi) the global
inhibition from the LH controls the level of sparseness in the KCs, both their silence absent
any odor (with or without octopamine), and their sparse firing in response to odors.
Statistics of KC responses to odors pre-, during, and post-training are shown in Fig 2.2.
Naive moths have low KC response to odors, in both percentage of KCs activated and their
consistency of response to odor stimulations (Fig 2.2, blue dots and curves). During training
octopamine induces high KC response, in both percentage and consistency (Fig 2.2, red dots
and curves). After octopamine is withdrawn, KC response is lower than during training, but
remains higher than naive levels in both percentage and consistency (Fig 2.2, green dots and
curves) for the trained odor only. Thus the newly-learned importance of the trained odor is
encoded as broader and stronger KC responses by means of strengthened synaptic connections.
EN (readout neuron) activity is also shown over time at the bottom of Fig 1.1. Learning
is evidenced by the increased EN response to the trained odor even after octopamine has been
withdrawn, due to Hebbian growth of synaptic connections into and out of the MB.
The FR activity of the PNs in the AL, the KCs in the MB, and the ENs, as illustrated in
Figs. 1.1 and 2.2, demonstrate the entire learning process that occurs under the influence of
octopamine stimulation. Without octopamine learning does not occur. Interestingly, although
the AL does not itself experience plasticity changes, it is the AL’s increased FR activity
(induced by octopamine) which enables permanent rewiring of the MB via Hebbian plastic
updates.
6
Figure 2.2. KC responses to odor during training: KCs respond sparsely to odor pre- and posttraining, ie absent octopamine (blue and green dots and curves). Octopamine induces transient increased
responsivity (red dots and curves). Training results in permanent increases in response to the trained odor, but
no increase in response to control odor (green dots and curves).
A: KC response to an odor before, during, and after training. x-axis: indexed KCs (500 shown). y-axis:
consistency of response (in %). The plots are for odor 1 as the trained odor (ie same data as panel B). Blue =
pre-training (no octopamine). Red = during training (with octopamine); note the heightened transient response.
Green = post-training (no octopamine). There is a permanent increase in the number of KCs that respond to
the trained odor.
B: Response rate vs. percentage of active KCs for trained and control odors before, during, and after training.
x-axis: percentage of KCs responding at the given rate. y-axis: consistency of response (in %). Blue = pretraining. Red = during octopamine (transient). Green = post-training. The LH plot shows odor 1 as the
reinforced odor. The scatterplots in (A) correspond to the three curves in this plot. Note that the permanent
KC response curve shifts up and to the right (blue→green) in the trained odor, ie more KCs respond to the odor
(right shift) and they respond more consistently (upward shift). The RH plot shows odor 2 as a control. The
control’s permanent KC response curve does not shift.
C: As (B) above, but in this experiment odor 1 is now the control (LH plot), and odor 2 is reinforced (RH plot).
In this case, the response curve of odor 2 (reinforced) shifts to the right (blue→green), while the response curve
of odor 1 (control) is unchanged.
2.3. Learning Experiments: EN Behavior.
A key finding of this paper is that the AL-MB model demonstrates robust learning behavior.
Here “learning” is defined as rewiring the system so that the reinforced odor yields a significantly stronger response in the readout neuron (EN) post-training, relative to naive (i.e.
pre-training) response to that odor, and also relative to the post-training responses to control
odors.
Learning Experiment Structure. Moths were randomly generated from a fixed parameter
template, which included randomly-assigned input maps (odor→AL) for four odors. The
odors projected broadly onto the AL, each odor hitting ∼20 out of 60 glomeruli. As a result,
their projections onto the AL overlapped substantially. On average, a given odor uniquely
7
projected onto about 6 glomeruli, and shared the other 14 glomeruli with other odors. Each
generated moth was put through a series of training experiments, with each run in the series
structured as follows:
1. The moth first received a series of stimulations from each odor, to establish a baseline
(naive) EN response. The stimulations were 0.2 seconds long and separated by gaps
of several seconds.
2. The moth was trained on one of the odors for 1 to 4 sessions (one session = 5 odor
stimulations), by applying odor and octopamine concurrently. The MB plastic weights
were updated according to a Hebbian rule.
3. Upon completion of training, the four odors were each again applied as a series of odor
stimulations, to establish post-training EN response.
For each {odor, #sessions} pair, this experiment was conducted 11 times (i.e. 11 noise
realizations), for a total of 176 experiments on each moth. These results were aggregated to
assess the particular moth’s learning response.
Learning Experiment Results. As a general rule, Network Model moths consistently demonstrated strong learning behavior: Training increased EN response to the trained odor well
beyond naive levels, and also much more than it affected EN response to control odors. Fig
2.3 summarizes the changes in EN responses in a typical experiment on a moth with four
odors. Panel A shows a typical noise realization timecourse, where one odor was reinforced
with octopamine and the other three odors were controls. Panel B shows the statistics of EN
response modulation, according to {odor, #sessions} pairs.
For ease of interpretation, the moth shown in Fig 2.3 had naive EN responses of roughly
equal magnitude for all four odors. When naive EN response magnitudes were highly uneven
(> 3x), robust learning still occurred, but the interpretation of the results is more complex due
to scaling issues. A typical experiment using a moth with odor responses of highly unequal
magnitude is shown in Fig 5.1.
Points of interest (EN responses to learning).
1. Because EN response is driven solely by feed-forward signals from KCs, ENs had
response ≈ 0 in the absence of odor, with or without octopamine, as expected (since
KCs are silent absent any odor). Thus Hebbian growth during training did not increase
EN baseline (no-odor) response.
2. The EN response to odor + octopamine was always very strong, as seen in Fig 2.3
(A), where EN responses to odor + octopamine extend above the top of the figure.
Note that this effect follows automatically from the calibration of the Network Model
to in vivo data. Its functional value to the moth is addressed in the Discussion.
3. Training consistently increased the EN response to the reinforced odor much more
than EN response to control odors, measured as percentage of naive odor response.
Since the Network Model did not include a Hebbian decay dynamic (for simplicity,
absent clear evidence), this was the key indicator of robust learning. That is, focused
learning was expressed by substantially higher increase in EN response to reinforced
vs control odors. We assume that an added Hebbian decay term would have knocked
smaller increases back, thus returning control odor responses to baseline.
8
Figure 2.3. Effect of training on EN FRs: A: Typical timecourse of EN responses from an experiment
with a single moth. First, 16 stimulations of each odor were delivered, to establish naive odor responses. Note
EN response variability due to noise in the system, especially in the AL. Next, the moth was trained on the
first (blue) odor trained over 2 sessions (10 stimulations), by delivering odor and octopamine concurrently This
timecourse corresponds to the {odor, #sessions} pair in the first column in panel B, at index 2 on the x-axis.
Octopamine was then withdrawn, and the four odors were again delivered in series of stimulations, to establish
post-training changes in EN response. The long green line represents a hypothetical trigger threshold, such that
EN response > threshold would induce a distinct behavior.
B: EN response changes due to training, aggregated results with 11 noise realizations for each {odor, #sessions}
pair. Each column shows results of training a given odor, color coded: blue, purple, red, green. x-axis = number
of training sessions.
First row: The y-axis measures percent change in EN FR. The line shows mean percent change. The error bars
show ±1, 2 stds.
Second row: The y-axis measures percent changes in EN response, relative to the trained odor (ie subtracting
the trained odor’s change from all odors). This shows how far each control odor lags behind the trained odor.
The line shows mean percent lag. The error bars show ±1, 2 stds.
Results of ANOVA analysis for differential effects of training on reinforced vs unreinforced
odors shows that when naive odor EN response magnitudes were similar (within 3x of each
other) p-values were consistently < 0.01. ANOVA analysis results are given in Section 5.3.
2.4. MB Sparsity Experiments.
Projection into a high-dimensional, sparse layer is a common motif in biological neural
systems [31, 32]. To explore the role of MB sparsity during learning, we ran Network Model
experiments that varied the level of generalized inhibition imposed on the MB (the lateral
horn, LH, controls MB sparsity level). Each experiment set a certain level of LH inhibition,
then ran simulations (see Methods) that trained moths on one odor with 15 odor stimulations
9
and left one control odor untrained. EN responses to both trained and control odors were
recorded, as well as the percentage of KCs active in response to odor.
Too little damping from the LH resulted in a high percentage of KCs being active (low
sparsity). This regime gave consistent EN responses to odor. But it also caused EN responses
to both control odor and noise to increase significantly during training, reducing the contrast
between EN responses to trained and control odors and also increasing spontaneous EN noise.
Too much damping resulted in a very low percentage of KCs being active (high sparsity).
This ensured that training gains were focused on the trained odor while EN response to control
odors and noise were not boosted. However, in this regime EN responses to all odors, both
pre- and post-training, were generally unreliable because too few KCs were activated.
Thus sparseness in the high-dimensional MB fulfilled a vital role in the Network Model’s
learning system. LH inhibition of the MB had an optimal regime that balanced opposing
demands, viz for reliable odor response and for well-targeted Hebbian growth, such that on
one hand EN odor responses were reliable, and on the other hand training gains were focused
on the trained odor only. Timecourses illustrating this effect are seen in Fig 2.4 (A). Fig 2.4
(B) shows how this trade-off varied with MB sparsity, by plotting two figures-of-merit:
(2.1)
Signal-to-Noise Ratio (SNR) =
µ(f )
where f = EN odor response;
σ(f )
and
(2.2)
Learning Focus =
µ(fT )
, where µ(fT ) = mean post-training EN response to
µ(fC )
trained odor, µ(fC ) = mean post-training EN response to control odor.
3. Discussion.
Our discussion centers around observations and implications of our Network Model experiments, which offer insights into the moth olfactory network and how it learns, including: (i)
predictions about aspects of the AL-MB still unclear in the literature, (ii) the role of sparse
layers, (iii) the role of octopamine, (iv) the value of randomness, and (v) the value of noise.
In addition, we consider these insights in the context of machine learning algorithms.
3.1. Predictions re details of AL-MB structure.
Model simulations enable us to make predictions about some unresolved aspects of the moth’s
AL-MB system. Some examples:
Do LNs inhibit PNs and LNs as well as RNs. In the AL, LNs have a net inhibitory effect on
PNs [33, 34], but the exact means to this end are not clear. In particular, while LNs are known
to inhibit RNs [33], it is less clear whether or to what degree LNs also directly inhibit PNs
and LNs. Efforts to calibrate our Network Model to in vivo data indicate that LNs need to
inhibit not just RNs, but also (to a lesser degree) LNs and PNs. The model weight strengths
for LN→RN, →LN, and →PN are in the ratio of 6:2:1. That LNs would inhibit LNs makes
sense when the goal is maximizing PN output of the active glomerulus: By inhibiting the LNs
10
Figure 2.4. Effects of sparsity on learning and EN reliability Results for a typical experiment
on a moth with two odors. A: EN responses timecourses for two odors, at varying levels of KC activation
(a, b: <1%. c, d: 5 to 15%. e, f: 20 to 45%. Order of events: 3 stimulations of each odor as baseline,
train on first odor (only one session shown), then 3 stimulations each post-training. At very sparse levels
(a, b) training is focused but odor response is not reliable. At low sparsity levels (e, f ) training is unfocused,
boosting EN response to control odor and to background noise. B: Two Figures of Merit (FoMs) plotted against
MB sparsity. Low KC activation (high sparsity) correlates with well-focused learning, but low odor response
SNR. High KC activation (low sparsity) correlates with poorly-focused learning, but high odor response SNR.
µ(f )
The FoMs are each normalized for easier plotting. y-axis: Blue data: σ(f
, a measure of odor EN response
)
µ(fT )
SNR, where f = EN odor response. Red data: µ(f
, a measure of learning focus, where µ(fT ) = mean EN
C)
post-training response to reinforced odor; µ(fC ) = mean EN post-training response to control odor (values are
thresholded at 1 for plotting). A high value indicates that increases in EN response due to training were focused
on the trained odor; low values indicate that irrelevant signal (FC ) was also boosted by training. The points are
experimental data, the curves are cubic fits. Vertical green lines indicate the 5 - 15% sparsity region, typical in
biological neural systems.
of rival glomeruli, the active glomerulus reduces the amount of inhibition directed at itself.
Similarly, that LNs would inhibit PNs makes sense if the goal is to reduce the PN output of
rival glomeruli.
11
Octopamine’s effects on different neuron types. Octopamine increases the responsivity of a
neuron to incoming signals. It is unclear how or whether octopamine affects various neuron
types (ie RNs, PNs, LNs, KCs). Calibration of the Network Model’s AL behavior, and tuning
of KC behavior to enable learning, indicate that octopamine needs to target RNs and LNs,
but not PNs, KCs, or ENs. Logical arguments support these findings:
RNs: Because RNs initially receive the odor signal, these are logical neurons to stimulate
with octopamine, because it sharpens their response to the exact signature being trained,
which in turn sharpen the AL’s output code for that odor.
LNs: LNs have the dual roles of inhibiting rival glomeruli and limiting overall PN output
in the AL. For the first role, increased LN response to RNs will tend to sharpen AL response
to the trained odor, by accentuating inhibition of rival glomeruli PNs. For the second role,
increased LN activity mitigates the risk that increased RN activity (due to octopamine) might
blow up the overall PN output of the AL.
PNs: Network Model simulations suggest that PNs should receive little or no octopamine
stimulation. While increasing PN responsivity would benefit RN-induced sharpening of the
trained odor’s signature, there are three downsides. First, RN input to PNs is intrinsically
noisy, so higher PN responsivity amplifies noise as well as signal. Second, since PNs respond
to LNs, higher PN activity tends to reduce the impact of LN inhibition, and thus reduces
the inhibition-induced sharpening of the AL odor response caused by octopamine. Third,
increasing PN responsivity can have an outsize effect on overall PN firing rates, ie it is a
“high-gain” knob and therefore risky.
KCs: Network Model simulations indicate that direct octopamine stimulation of KCs
greatly reduces sparseness in the MB, which can be disastrous to learning. Thus we expect
that octopamine has no, or only slight, direct stimulative effect on KCs.
3.2. Noise filtering role of the sparse, high-dimensional stage.
Projection from a dense, low-dimensional coding space (eg the AL) to a sparse, high-dimensional
coding space (eg KCs in the MB) is a widespread motif of biological neural systems, with size
shifts are routinely on the order of 20x to 100x [31, 35, 32]. The reasons for this pattern are not
fully understood. Some proposed reasons include information capacity, long-range brain communication, and reduced training data needs [31], as well as improved discrimination ability
[11, 32].
Network Model experiments bring to light another, central, role of sparseness: It acts as
a robust noise filter, to protect the Hebbian growth process from amplifying upstream noise
to out-of-control levels. Though noise may be useful (or unavoidable) in upstream networks
such as the AL, noise that reaches the neurons on both sides of a synaptic connection will
be amplified by Hebbian growth during learning, swamping the system’s downstream neurons
(eg ENs) with noise. However, the “fire together, wire together” principle of Hebbian learning
is an AND gate. Thus it suffices to remove noise from just one of the two connected neurons
to prevent synaptic growth. Sparsity does precisely this.
Network Model experiments show that sparseness in the MB ensures that noise does not
get amplified during training, so that post-training EN spontaneous firing rates and EN control
odor responses remain unchanged. Conversely, when KC response is not sufficiently sparse,
any training leads rapidly to noisy EN spontaneous response levels and amplified EN responses
12
to control odor. This implies that the noise filtering induced by MB sparseness is necessary
for a workable Hebbian learning mechanism. This finding agrees with [11], where experiments
with a computational MB model indicated that sparse MB response gave stronger built-in
discriminating ability.
Setting aside the particular demands of Hebbian plasticity, robust noise filtering may be
a core function of sparse, high-dimensional stages within any network cascade where noise
accumulates due to (beneficial) use in upstream stages.
3.3. Roles of octopamine.
The levels of octopamine stimulation in our Network Model were calibrated to in vivo data
on PN responses to octopamine. Thus, the simulations give insights into downstream effects
of octopamine on plasticity, KC responses, EN responses, and Hebbian learning.
Accelerant. Moths can learn to respond to new odors remarkably quickly, in just a few
exposures. Simulations indicates that while Hebbian growth can occur without octopamine,
it is so slow that actionable learning, ie in terms of amplified EN responses, does not occur.
This implies that octopamine, through its stimulative effect, acts as a powerful accelerant to learning. Perhaps it is a mechanism that allows the moth to work around intrinsic
organic constraints on Hebbian growth of new synapses, constraints which would otherwise restrict the moth to an unacceptably slow learning rate. To the degree that octopamine enabled
a moth to learn more quickly, with fewer training samples, it would clearly be highly adaptive.
Active learning. Simulations indicate that octopamine strongly stimulates the EN response
to even an unfamiliar odor. Since octopamine is delivered as a reward, this has a beneficial
effect in the context of reinforcement learning [36], with the moth as the learning agent. An
agent (the moth) can in some cases learn more quickly when it has choice as to the sequence
of training samples (Active Learning [37]).
In particular, when a certain class of training sample is relatively rare, it benefits the agent
to actively seek out more samples of that class [38]. Octopamine enforces high EN response
to a reinforced odor, ensuring that ENs will consistently exceed their “take action” threshold
during training. If the action is to “approach”, the moth is more likely to again encounter the
odor, thus reaping the benefits predicted by Active Learning theory. This advantage applies
in the context of positively-reinforced odors.
In the case of aversive learning, the high EN responses to unfamiliar but objectionable
odors, due to dopamine, would cause the moth to preferentially avoid further examples of the
odor. This would slow learning of aversive responses (a drawback), but would also minimize
the moth’s exposure to bad odors (danger avoidance, a benefit).
Exploration of optimization space. A limitation of Hebbian growth is that it can only reinforce what already exists. That is, it only strengthens channels that are transmitting signals
deemed (by association) relevant to the stimulus being reinforced. Absent a mechanism like
octopamine, this constrains growth to channels that are already active. Simulations indicate
that octopamine induces much broader activity, both upstream from and within the plastic
layer, thus activating new transmitting channels. This allows the system to strengthen, and
13
bring permanently online, synaptic connections that were formerly silent. This expands the
solution space the system can explore during learning. This function may be particularly
important given the constraint of sparsity placed on odor codes in the MB.
Injury compensation. There is evidence that many forms of injury to neurons result in
dropped spikes and thus lower firing rates in response to odors [39]. This injury-induced
drop in the signals reaching the ENs could induce behavioral consequences, by lowering EN
responses to below key behavioral action thresholds. In this case, the generalized stimulation
induced by octopamine might compensate for the reduced signal strength sufficiently to lift
the EN responses back above threshold during training. This in turn would allow Hebbian
growth to strengthen the synaptic connections to those ENs, such that even the reduced signals
from the still-injured upstream network would be sufficient to push EN response above the
action threshold. This mechanism (octopamine stimulation plus Hebbian synaptic growth)
would allow the moth to regain behavioral function lost due to damage to the upstream neural
system.
3.4. The value of randomness.
The principle of randomness permeates the moth olfactory network, for example in random
neural connection maps [9] and in highly variable performance characteristics of chemical
receptors in the antennae [33]. A key result from Network Model experiments is that a
biologically-based neural net permeated with the random principle can robustly learn. This is
in marked contrast to engineered computers, where poorly-spec’ed components are a liability.
Particular benefits of randomness include:
Random KC→EN connections. These guarantee that projections onto ENs are incoherent
relative to whatever low-dimensional manifold in KC-space, which ensures (by the JohnsonLindenstrauss lemma) that the EN readouts will preserve distinctions between elements in
KC-space [31].
Variable gaba-sensitivity in glomeruli. This increases the range of sensitivity to odor concentration, because some glomeruli simply don’t turn off when a different, stronger odor tries
to mask them through lateral inhibition (LNs) [40].
Variable sensitivity of antennae receptors. This gives a natural sensitivity to concentration,
as progressively stronger odor will progressively activate the less-sensitive receptors, increasing the total RN input to glomeruli.
Resistance to injury. A randomly-connected network is more robust to damage. When
exact connection maps and strengths are not required in the first place, damage has less
impact on the fundamental nature of the system.
Most importantly, randomness (of connections and component properties) is evolutionarily
cheap, easy, and available. So perhaps the core benefit of randomness to the moth olfactory
network is that it works at all.
14
3.5. The value of noise.
Noise in biological neural systems is believed to have value, for example by encoding probability distributions and enabling neural Baysian estimations of posteriors [41]. In addition,
injection of noise while training engineered NNs can improve trained accuracy [42]. Simulations indicate that in the AL-MB, noise has potential benefits, coupled with caveats. Figure
2.2 shows how noise in the AL adds an extra dimension to MB odor encoding, increasing the
granularity of its odor responses.
The MB responds to odors in two ways: (i) by the number of KCs that are responsive,
and (ii) by the reliability (eg from 10% to 100%) of their responses. This can be seen in the
effect of octopamine on KC odor response, Fig 2.2 (B). Octopamine boosts MB odor response
by increasing the number of active KCs (horizontal shift in response curves) and also by
increasing the reliability of responsive KCs (vertical shift in responsivity curves). Both these
shifts represent a stronger MB response and translate into stronger EN response.
Taken together, they provide a finer granularity of the response range, versus the binary
response of a noise-free system. Looked at another way, the MB response to noisy inputs
from the AL is a concrete example of a mechanism used by a neural system to translate the
probability distributions encoded by noisy neurons into actionable signals with high dynamic
range and granularity.
A caveat is that noise in the AL-MB must be confined to the AL, i.e. upstream from
the encoding layer, in order to protect the readout neurons and Hebbian learning mechanism
from noise. The system’s success depends on robust noise filtering at the MB layer, via global
inhibition from the LH. So the three-stage architecture consisting of:
Noisy pre-amplifier layer → Sparse noise-reduction layer → Decision layer
is an interdependent system well-suited to nuanced decision-making.
Given a layer (the MB) that effectively protects the decision neurons from upstream noise,
the system is also potentially robust to noisy stimuli. In the neural net context, input samples
(i.e. inputs to the feature-reading layer) can be thought of as a de facto “first layer” of the
neural net. A system that is robust to upstream noise may also be naturally robust to noisy
inputs, a further potential advantage of judicially-placed sparse layers.
3.6. Applications to Machine Learning.
The model and simulations in this paper characterize key features of the AL-MB system which
might usefully be ported to machine learning algorithms. These features include: Generalized
stimulation during training; Hebbian growth; sparse layers to control plastic connections and
filter noise; and noisy initial layers. Advantages of this biological toolkit include:
Fast learning. Moths can reliably learn a new odor in less than 10 exposures. Biological
brains in general can learn given few training samples. This contrasts by orders of magnitude
with the voracious demands of DNNs, where assembling sufficient training data can be a serious chokepoint in deployment.
Robustness to noise. The sparse layer in the AL-MB acts as an effective noise filter, protecting the readout neurons from a noisy upstream layer (the AL). Since the system is designed to
accommodate upstream noise, it is possible that it can also readily accommodate noisy input
15
samples. NNs have a troublesome property, that input-output score functions are not locally
continuous [43]. Biological neural nets seem to avoid this particular fault. The noisy layer →
sparse layer motif may be one reason for this. It may thus be a useful motif to apply in ML
architectures.
Novel training mechanism. Hebbian growth, combined with generalized stimulation via octopamine, is (in the context of ML) a novel mechanism to explore a solution space and train a
classifier. In particular, it works on a fundamentally different principle than backpropagation
algorithms: It does not minimize a loss function via gradient descent; rather, it selectively
strengthens only those connections that are transmitting meaningful signals. The addition of
a sparse layer acts to control and focus learning, by leveraging the AND-function nature of
Hebbian growth.
Biological plausibility. One characteristic (not criticism) of backprop optimization is that it
is biologically implausible, since it requires a neuron to have more than local knowledge of the
system. The search for neural network architectures (for example with recurrent connections
to transport non-local information) and variants of backprop which are biologically plausible,
and which thus might narrow the gap between biological and engineered NNs, is currently an
area of interest, especially in the context of DNNs [44]. This paper demonstrates that the
triad of octopamine stimulation + Hebbian growth + sparse layers can efficiently train a NN,
and is thus a possible candidate to address the biological plausibility gap.
4. Materials and Methods.
In this section, we describe the biological moth olfactory network, as well as our Network
Model. We also provide a Glossary, and describe the in vivo data used for model calibration.
4.1. Moth olfactory system overview.
The parts of the AL-MB implicated in learning are organized as a feed-forward cascade of five
distinct networks, as well as a reward mechanism [21, 22]. Figure 1.1 gives a system schematic
along with typical firing rate (FR) timecourses (from simulation) for neurons in each network.
1. Antennae. Roughly 30,000 noisy chemical receptors detect odor and send signals to
the Antenna Lobe [6].
2. Antenna Lobe (AL). Contains roughly 60 units (glomeruli), each focused on a single
odor feature [21]. The AL essentially acts as a pre-amp, boosting faint signals and
denoising the antennae inputs [23]. AL neurons are noisy [10].
3. Lateral Horn (LH). Though not fully understood, one key function is global inhibition
of the Mushroom Body to enforce sparseness [25].
4. Mushroom Body (MB), here synonymous with the Kenyon Cells (KCs). About 4000
KCs are located in the calyx of the Mushroom Body (MB). These fire sparsely and
are believed to encode odor signatures [24, 4, 8].
5. Extrinsic Neurons (ENs), numbering ∼10’s, located downstream from the KCs. These
are believed to be “readout neurons” that interpret the KC codes and convey actionable
messages (such as “fly upwind”) [26, 27].
6. Reward Mechanism. A large neuron sprays octopamine globally over the AL and MB,
16
in response to reward, such as sugar at the proboscis. Learning does not occur without
this octopamine input [28, 1].
7. Inter-network connections: In the AL-MB these are strictly feed-forward, either excitatory or inhibitory. In particular, Antennae→AL, AL→LH, KCs→ENs are all excitatory. LH→KCs is inhibitory. AL→KCs have both excitatory and inhibitory channels.
8. Plasticity: The connections into the KCs (AL→KCs) and out of the KCs (KCs→ENs)
are known to be plastic during learning [29, 6]. The AL is not plastic.
4.2. Glossary.
Antenna lobe (AL). A collection of neurons innervated by odor receptors in the antennae.
It sends signals to the mushroom body via projection neurons. Connections in the AL are not
plastic.
Mushroom body (MB). A collection of neurons (Kenyon cells - KCs) downstream from the
antenna lobe. The MB is believed to store odor codes that serve as a memory, allowing the
moth to recognize odors. Connections in the MB are plastic.
Lateral horn (LH). A collection of neurons which receives input from the AL and sends
inhibitory output to the MB. One of its roles is to enforce sparse firing in MB neurons.
Receptor neuron (RN). These neurons respond to odors (volatiles) at the antennae and
stimulate the antenna lobe. RNs respond to different, distinct odors.
Glomerulus. The antenna lobe is divided into about 60 glomeruli, each of which is a selfcontained collection of neurons (projection and lateral), innervated by RNs that respond to
particular odors.
Projection neuron (PN). Each glomerulus contains projection neurons, whose output innervates the KCs and also the lateral horn, but not other glomeruli in the AL, ie they are
feed-forward only. Most PNs start in one glomerulus and are excitatory. A few PNs arborize
in several glomeruli and are inhibitory (we refer to inhibitory PNs as “QNs”). Each glomerulus
initiates about five PNs.
Lateral neuron (LN). Each glomerulus contains lateral neurons, which innervate other
glomeruli in the AL. LNs are inhibitory. One function is competitive inhibition among
glomeruli. Another function is gain control, ie boosting low signals and damping high signals.
Kenyon cell (KC). Neurons in the calyx of the MB. These have very low FRs, and tend to
respond to particular combinations of PNs. KCs respond sparsely to a given odor. There are
about 4000 KCs, ie a two-orders-of-magnitude increase over the number of glomeruli. Each
KC synapses with about ten PNs. Connections into and out of KCs are plastic.
Extrinsic neuron (EN). A small number of neurons downstream from the KCs. ENs are
thought to be “readout” neurons. They interpret the odor codes of the KCs, deciding to eg
“ignore”, “approach”, or “avoid”.
Firing rate (FR). The number of spikes/second at which a neuron fires. Typically FRs are
counted using a window (eg 500 ms). The moth’s response to odor stimulations is episodic,
with FR spikes in FR and rapid return to spontaneous FRs. Neurons respond to relative
changes in FR, rather than to raw magnitude changes. A neuron’s relative change in FR is
scaled by its spontaneous FR (see section 4.5 below).
Octopamine. A neuromodulator which stimulates neural firing. The moth spritzes octopamine on both the AL and MB in response to sugar, as a feedback reward mechanism.
17
Dopamine has a similar stimulating effect on both AL and MB, but it reinforces adverse rather
than positive events.
4.3. Component networks and their Network Model representations.
This subsection offers a more detailed discussion of the constituent networks in the biological
AL-MB, and details about how they are modeled in our Network Model.
Antennae and receptor neurons. The Antennae receptors, activated by chemical molecules
in the air, send excitatory signals to Receptor Neurons (RNs) in the AL. Several thousand
antennae converge onto 60 units (glomeruli) in the AL [45]. All the receptors for a given
atomic volatile converge onto the same glomerulus in the AL, so the glomeruli each have
distinct odor response profiles [46]. Since natural odors are a blend of atomic volatiles, a
natural odor stimulates several units within the AL [47].
Our model does not explicitly include antennae. Rather, the first layer of the model
consists of the RNs entering the glomeruli. Though ∼500 RNs feed a given glomerulus, the
model assumes one RN. The benefit of many RNs converging appears to be noise reduction
through averaging [33]. This can be simulated by one RN with a smaller noise envelope. Each
glomerulus’ RN has a spontaneous FR and is excited, according to random weights, by odor
stimuli.
Antenna lobe and projection neurons. The AL is fairly well characterized in both structure
and dynamics, with a few important gaps. Moths and flies are similar enough that findings in
flies (Drosophila) can generally be transferred to the moth (in contrast, locusts and honeybees
are more complex and findings in these insects do not safely transfer) [48].
The AL contains about 60 glomeruli, each a distinct unit which receives RN input and
projects to the KCs via excitatory PNs. The same PN signal also projects to the LH [25]. The
AL, unique among the networks, has inhibitory lateral neurons (LNs) [49], the only neurons
that are not strictly feed-forward. (There is some evidence of excitatory LNs, eg [50]; the
Network Model excludes this possibility.) The LNs act as a gain control on the AL, and also
allow odors to mask each other by inhibiting other glomeruli’s RNs [51, 40]. It is not known
whether LNs also inhibit PNs and LNs. Based on calibrations to in vivo data, in Network
Model LNs inhibit all neuron types (cf section 3.1). Thus each glomerulus contains dendrites
(ie outputs) for PNs and LNs, and axons (ie inputs) for RNs and LNs, as shown in Figure 4.1.
Each glomerulus does the following: Receives RN input from the antennae receptors upstream; inhibits other glomeruli within the AL via LNs; and sends excitatory signals downstream via Projection Neurons (PNs).
In general, each PN is innervated in a single glomerulus. In moths, there are ∼5 PNs
rooted in each glomerulus (60 glomeruli, ∼300 PNs). The Network Model assumes all PNs
from a given glomerulus carry the same signal (because they share the same glomerulus and
therefore inputs, and perhaps also because of ephaptic binding) [52].
Glomeruli also initiate pooled Inhibitory Projection Neurons (QNs) that send inhibitory
signals downstream to the KCs.
The AL contains a powerful macro-glomerulal complex (MGC), which processes pheromone.
Because pheromone response has fundamentally different dynamics than food odor response
[53], the model ignores it. Only the glomeruli associated with non-pheromone (food) odors
18
Figure 4.1. Schematic of AL glomeruli. Detail of neural connections within a glomerulus. Red =
inhibitory, green = excitatory, blue = increases responsiveness. RNs enter from the antennae. LNs enter from
other glomeruli; one full LN is shown. It is not known if octopamine modulates LNs and PNs (see section 3.1).
are modeled.
Connections in the AL are not plastic with long-term persistence [5]. While some evidence
of short-term plasticity exists, the Network Model ignores this option.
Lateral horn. The LH receives input from the PNs. It then sends an inhibitory signal to
the KCs. This inhibition from the LH appears to ensure that the KCs fire very sparsely and
thus act as coincidence detectors for signals from the AL [52, 54, 55].
The LH is also suspected of containing a parallel system for processing certain intrinsicallyknown odors in short-cut fashion (labeled lines) [56]. Since this parallel system is (by definition) not involved with learning, the Network Model ignores it. The LH is modeled solely as
a simple sparsifying inhibition on the KCs.
(Note: The locust and honeybee, which have more complex olfactory systems and different
use-cases in terms of odor processing, have a time-oscillating depolarization mechanism (local
potential fields, LPF) [57] which serves a similar purpose to LH inhibition in the moth. LPF
oscillations are absent in the moth [21].)
Mushroom body and Kenyon cells. The KCs (∼4000) in the MB are believed to encode
odor memories in a high-dimensional, sparse space [30]. Odors with no meaning to the moth
still have non-zero codes in the KCs.
KCs receive excitatory input from the PNs and inhibitory input from QNs, both of which
vary greatly between KCs, since each KC is innervated by only ∼10 PNs [21]. The connection
map appears to be random [58]. The KCs also receive generalized damping inhibition from
the LH. (There is some evidence in drosophila of an MB→MB global inhibitory neuron [54],
with the same essential effect as LH inhibition; the Network Model excludes this possibility.)
KCs fire very sparsely, generally respond to only a single odor, and are silent absent that odor
[8]. KCs are treated as noise-free. Their output is an excitatory signal sent to the extrinsic
neurons (ENs) [26].
In addition to olfactory input, the KCs receive input signals from other parts of the moth
(eg hearing) [52]. Because the Network Model targets olfactory learning, it ignores these other
inputs and uses a reduced number of KCs (∼2000 instead of ∼4000).
19
The synaptic connections in the MB (PNs→KCs, QNs→KCs, and KCs→ENs) are plastic,
ie they can be modified during training [59]. The generalized inhibition from LH→KCs is
modeled as non-plastic (actual physiology is not known).
Extrinsic neurons. Though located in the lobes of the MB, here ENs are not considered part
of the MB, which is taken to be synonymous with the KCs. ENs are few in number compared
to the KCs (∼10s ) [26, 27]. They are believed to be “readout” neurons, that interpret the KC
codes as actionable signals (eg “approach”, “avoid”) [6]. We assume that ENs trigger actions
when their output FRs exceed some threshold.
We define Learning as: Permanently boosting EN responses beyond their naive (untrained)
level, so that EN responses to reinforced stimuli can consistently exceed an action-triggering
threshold. This is tantamount to modifying the moth’s behavior.
Octopamine (reward circuit). A large neuron delivers octopamine to the entire AL and MB,
in response to positive stimuli, eg sugar at the proboscis. It acts as a reward feedback to the
system. A similar neuron delivers dopamine to the AL and MB in response to negative stimuli,
and acts as an aversive feedback signal [5]. Learning does not occur without octopamine (or
dopamine) [1].
Despite their opposite reward values, both octopamine and dopamine act in the same way
when sprayed on a neuron: They increase the neuron’s general tendency to fire [60]. In the
Network Model this effect is modeled as making a neuron more responsive to excitatory inputs
(eg from odors and RNs) and less responsive to inhibitory inputs (eg from LNs). Details of
octopamine’s effects if any on various neural types are not well-characterized. In the Network
Model octopamine directly affects RNs and LNs but not PNs in the AL (cf section 3.1); has
no direct effect on KCs or ENs (though there are strong indirect effects); and has no effect on
the LH inhibitory signal.
It is unclear whether octopamine delivery to both the MB and AL is necessary and sufficient for learning [1, 5]. The Network Model assumes that octopamine controls an “on/off”
switch for Hebbian growth, ie there is no plasticity in the MB (and therefore no learning)
without octopamine.
4.4. Network Model description.
This section describes our Network Model in detail. It covers the firing rate measure used to
compare model output to in vivo data; model dynamics; plasticity and other details; model
parameters; and moth generation. All coding was done in Matlab. Computer code for this
paper will be found at:
https://github.com/charlesDelahunt/SmartAsABug
4.5. Firing rate measure.
To compare PN firing rate statistics from in vivo experiments and from Network Model
simulations (ie model calibration), we use a measure of firing rate (FR) based on Mahalanobis
distance, similar to the measure DF
F common in the literature [26, 40, 30, 61]. The premise
is that neurons downstream respond to a ±1 std change in FRs equally (modulo different
connection weights), independent of the sometimes large (up to 40x) magnitude differences in
the raw spontaneous FRs of different neurons. The FR measure is defined as follows:
1. Each PN has a spontaneous firing rate (FR) with a gaussian noise envelope.
20
2. PNs with FR < 1 spike/sec are ignored, on the assumption that such PNs represent
artifacts of experiment (also, the gaussian noise assumption fails). About 10% of PNs
in experimental data fall in this category.
3. Output FR activity of PNs is measured as M (t) = distance from mean spontaneous
FR, in units of time-varying std dev of spontaneous FR (ie Mahalanobis distance):
Let
F (t) = raw firing rate (spikes per second).
S(t) = spontaneous firing rate (no odor).
µS(t) = moving average of S (no odor).
µ̄S(t) = smoothed estimate of the moving average µS, eg a quadratic or spline fit.
σS (t) = standard deviation of S, calculated using S − µ̄S values within a moving window centered on t.
σS (t) and µS(t) are typically steady absent octopamine, but are often strongly modulated by octopamine.
Then the measure of FR activity M is:
(4.1)
M (t) =
F (t) − µ̄S(t)
σS (t)
4. M is related to the measure DF
F :
F (t)−µS
DF
∆F
DF
, ie F is change in FR, normalized by spontaneous FR. The
F = F =
µS
key difference between M and DF
F is whether or how σS is estimated, due to varying
exigencies of experiment. Our experimental data allow reasonable estimates of σS and
µS. Network Model simulations produce very good estimates, since computers are
more amenable to repeated trials than live moths.
4.6. Model dynamics.
Our Network Model uses standard integrate-and-fire dynamics [16], evolved as stochastic
differential equations [62].
Let x(t) = firing rate (FR) for a neuron. Then
(4.2)
τ
dx
= −x + s(Σwi ui ) = −x + s(w · u), where
dt
w = connection weights;
u = upstream neuron FRs;
s() is a sigmoid function or similar.
PN dynamics are given here as an example. Full model dynamics are given in Section 5.2.
PNs are excitatory, and project forward from AL→MB:
(4.3)
τ
dP
e + dWP where
= −P + s(P)
dt
W(t) = brownian motion process;
e = −(1 − γo(t)M OP )*M LP *uL + (1 + o(t)M OP )*M RP *uR ;
P
21
M OP = octopamine→PN weight matrix (diagonal nG × nG);
M LP = LN→PN weight matrix (nG × nG with trM LP = 0);
M RP = RN→PN weight matrix (diagonal nG × nG);
o(t) indicates if octopamine is active (o(t) = 1 during training, 0 otherwise).
uL = LN FRs, vector nG × 1;
uR = RN FRs (nG × 1);
γ = scaling factor for effects on inhibition.
Discretization. The discretization uses Euler-Maruyama, a standard step-forward method
for SDEs [62].
Euler (ie noise-free): xn+1 = xn + ∆tf (xn )
√
Euler-Maruyama: xn+1 = xn + ∆tf (xn ) + randn(0,1) ∆t, where controls the noise intensity.
Convergence. Timestep ∆t was chosen such that noise-free E-M evolution gives the same
timecourses as Runge-Kutta (4th order), via Matlab’s ode45 function. ∆t = 10 mSec suffices
to match E-M evolution to R-K in noise-free moths. Values of ∆t ≤ 20 mSec gives equivalent
simulations in moths with AL noise calibrated to match in vivo data. Values of ∆t ≥ 40 mSec
show differences in evolution outcomes given AL noise.
Plasticity. The model assumes a Hebbian mechanism for growth in synaptic connection
weights [63, 29]. That is, the synaptic weight wab between two neurons a and b increases
proportionally to the product of their firing rates (“fire together, wire together”): ∆wab (t) ∝
fa (t)fb (t).
Thus, synaptic plasticity is defined by:
(4.4)
∆wab (t) = γfa (t)fb (t), where γ is a growth parameter.
There are two layers of plastic synaptic weights, pre- and post-MB: AL→MB (M P,K , M Q,K ),
and MB→ENs (M K,E ) . Learning rate parameters of the Network Model were calibrated to
match experimental effects of octopamine on PN firing rates and known moth learning speed
(eg 4 - 8 trials to induce behavior modification) [60]. The Network Model does not decay unused synaptic weights. Training does not alter octopamine delivery strength matrices (M O,* ).
That is, the neuromodulator channels are not plastic (unlike, for example, the case in [64]).
Odor and octopamine injections. Odors and octopamine are modeled as hamming windows.
The smooth leading and trailing edges ensures low stiffness of the dynamic ODEs, and allows
a 10 mSec timestep to give accurate evolution of the SDEs in simulations.
Training. Training on an odor consists of simultaneously applying stimulations of the
odor, injecting octopamine, and “switching on” Hebbian growth. Training with 5 to 10 odor
stimulations typically produces behavior change in live moths.
4.7. Model parameters.
There is a risk, when modeling a system, of adding too many free parameters in an effort to fit
the system. Fewer free parameters are better, for the sake of generality and to avoid overfitting.
Conversely, we wish to reasonably match the physiological realities of the system. Because the
key goal of this paper was to demonstrate that a simple model, in terms of parameters and
structure, can reproduce the learning behavior of the AL-MB, we made efforts to minimize
22
the number of free parameters. For example, neuron-to-neuron connections in the model are
defined by their distributions, ie two parameters each. These are (usually) distinct for different
source-to-target pairs (eg LN→RN, LN→LN, etc).
Some mean and std dev parameters for distributions are shared among different neuron types
(eg LNs, PNs, and QNs all share the same variance scaling parameter).
Parameter list. There are in total 47 free parameters:
1. Structure: 5 (eg number of neurons in each network)
2. Dynamics: 12 (noise: 2. decay tau and sigmoid: 3. Hebbian growth: 6. misc: 1).
3. Spontaneous RN FRs: 3.
4. Connection matrices: 27 (to control non-zero connection ratios: 5; synaptic weights (eg
M P,K , M R,P ): means 12, std devs 4; octopamine weights (eg M O,R , M O,P ): means 6, std
devs 2).
Dynamics parameters. The differential equations of all neuron types share the same decay
rate, set to allow return to equilibrium in ∼1 second, consistent with in vivo data. Neurons
also share parameters of the sigmoid function within the differential equation. Noise added
via the SDE model is controlled by a single parameter , the same for all neuron types. It is
σS
determined by empirical constraint on µS
, as shown in column 2 of Figure 2.1.
Connection matrix generation. Connection weight matrices (eg M P,K etc) are generated in
a standard way, from Gaussian distributions with std dev σ defined proportional to the mean
µ, using a scaling factor v:
M ∗,∗ ∼ N (µc , σc2 ) where µc depends on the neuron types being connected, and σc = vµc .
Many connection types typically share the same v.
A special feature of the AL is that all the neurons in a given glomerulus share a common
environment. For example, all the neurons, of whatever type, in glomerulus A will share the
same strong (or weak) LN axon from glomerulus B. Thus, the RN, LN, and PNs in a given
glomerulus are all correlated. In addition, neuron types are correlated. To model this dual
set of correlations, connection matrices in the AL are generated as follows. As an example,
consider LN connection matrices in the AL:
1. A glomerulus-glomerulus connection matrix M L,G is created, which defines LN arborization at the glomerular level.
2. This connection matrix is multiplied by a neural type-specific value to give M L,P ,M L,L ,
and M L,R connection matrices. This is particularly important when tuning the various
inhibitory effects of LNs on RNs, PNs (QNs), and LNs.
3. Sensitivity to GABA: A separate variance factor determines glomerular sensitivity to
GABA (ie sensitivity to inhibition). This is tuned to match data in the literature [40],
and applies to LN-to-PN(QN) (ie M L,P ) connections only.
The goal of this two-stage approach is to enforce two types of similarity found in the AL: (i)
Connections to all neurons within a single glomerulus are correlated; and(ii) connections to
all neurons of a certain type (LN, PN, RN) are correlated.
Due to constraints of the biological architecture there are many zero connections. For
example, about 85% of entries in the AL→MB weight matrix are zero because MB neurons
connect to only ∼10 projection neurons [58]. All MB→EN weights are set equal at the start
of training. Training leads rapidly to non-uniform distributions as inactive connections decay
23
and active connections strengthen.
RN spontaneous firing rates. RNs in the glomeruli of the AL have noisy spontaneous firing
rates [23]. The Network Model simulates this by assigning spontaneous firing rates to RNs.
These spontaneous firing rates are drawn from a gamma distribution plus a bias:
β α α−1 −βx
γ(x|α, β, b) = b + Γ(α)
x
e
, where α, β are shape and rate parameters, and Γ(·) is the
Gamma function.
This can be thought of as a source of energy injected into the system, at the furthest upstream
point (absent odor). Other energy sources are odor signals and octopamine. The spontaneous
firing rates of all other neurons in the Network Model are the result of their integrate-and-fire
dynamics responding as RN spontaneous FRs propagate through the system.
4.8. Discrepancies between biology and model.
There are some known discrepancies between our Network Model and the moth AL-MB.
These are listed below.
Connection weight distributions. This model version uses gaussian distributions to generate
initial connection weights. However, moths used in live experiments are older and thus presumably have modified PN→KC and KC→EN connection weights. If this modification was
strong, we might expect the connection weight distributions to tend towards a scale-free rather
than gaussian distribution [65]. This represents an unknown discrepancy between structure
parameters of the live moths used in experiments vs the model.
Hebbian pruning. The Network Model contains no pruning mechanism to offset, via decay,
the Hebbian growth mechanism. Such pruning mechanisms are common in nature, so it
is reasonable to suppose that one might exist in the AL-MB. The moth has inhibitory as
well as excitatory feed-forward connections from AL to MB. In the Network Model, pruning
is functionally replaced by Hebbian growth of QN→KC inhibitory connections, which act
to inhibit KCs and thus offset the growth of excitatory PN→KC connections (this does not
directly offset KC→EN Hebbian growth). Thus omitting a separate Hebbian decay mechanism
is a matter of convenience rather than a match to known biology.
Non-olfactory input to KCs. In addition to olfactory input, the KCs receive signals from
other parts of the moth, eg hearing. Because this model targets only olfactory learning, it
ignores these other inputs to the KCs, and reduces the total number of KCs (from ∼4000 to
∼2000).
Number of QNs. There are believed to be about 3-6 QNs projecting from the AL to the
MB. This model sets their number at about 15. The reason is that, absent a Hebbian pruning
system in the model, the QNs function as the brake on runaway increases in KC responses
due to Hebbian growth. So the increased number of QNs is a compensation for the lack of a
weight-decay system.
Number of ENs. This model version has only one EN, since its goal is to demonstrate
simple learning. The moth itself possesses multiple ENs.
LH inhibition. The LH→KC inhibitory mechanism used in this chapter is modeled as a
time-invariant global signal, delivered equally to all KCs. This simplifies the model parameter
space while retaining the essential functionality of the LH. A more refined version of LH→KC
inhibition might vary in strength according to PN output, since the same PN signals that
excite the KCs also excite the LH. The actual dynamics of the AL→LH→KC linkage are not
24
known, beyond the principle that inhibition from the LH sparsifies the KC codes and makes
the individual KCs act as coincidence detectors.
4.9. in vivo neural recordings data.
Model parameters were calibrated by matching Network Model performance to in vivo
electrode readings from the ALs of live moths. The various performance metrics are described
in Results.
Electrode data was collected by the lab of Prof Jeff Riffell (Dept of Biology, UW). It
consists of timecourses of PN firing rates measured via electrode in the AL of live moths,
during a variety of regimes including:
1. Series of 0.2 sec odor stimulations delivered without octopamine. These experiments
gave data re PN response to odor relative to PN spontaneous (baseline) FRs, absent
octopamine.
2. Series of 0.2 sec odor stimulations delivered coincident with sugar reward (which delivers octopamine). This gave data re how PN odor response is modulated by octopamine,
relative to octopamine-free spontaneous FR. See Figure 4.2 panel A.
3. Series of 0.2 sec odor stimulations, delivered first without and then coincident with an
octopamine wash applied to the AL. This gave data re how PN spontaneous FR and
PN odor response are modulated by octopamine. See Figure 4.2 panel B.
In most cases the applied odor consisted of a collection of 5 volatiles, which taken together
stimulate many glomeruli in the AL. It was selected to ensure sufficient odor-responsive PNs,
such that inserted electrodes would detect interesting (ie responsive) PNs. Further details re
in vivo data collection can be found in [66] and in Sec. 5.1. Example timecourses are shown
in Figure 4.2.
4.10. Simulation setup.
For Network Model learning experiments, the time sequence of events for simulations, shown
in Fig 1.1, is as follows:
1. A period of no stimulus, to assess baseline spontaneous behavior.
2. Four odor stimuli are delivered, 16 stimulations each (two odors were used in MB
sparseness experiments).
3. A period of control octopamine, ie without odor or Hebbian training.
4. The system is trained (odor + octopamine + Hebbian mechanism) on one of the odors.
5. A period of no stimulus, to assess post-training spontaneous behavior.
6. The odors are re-applied (16 stimulations each), without octopamine, to assess effects
of training on odor response.
25
Figure 4.2. Time series of PN firing rates from in vivo experiments. x-axis = time, y-axis =
FR. Blue lines = mean spontaneous rate, shaded regions = ±1 and 2 std. Red dots are odor responses. Green
dots are response to control (mineral oil).
A: PN response, given odor plus coincident sugar reward, ie plus octopamine (time series for PNs with odor
only are similar, but with less strong odor responses). Top row: unresponsive to odor. Middle row: excited
response to odor. Bottom row: inhibited response to odor.
B: PNs with octopamine wash added in mid-experiment, then rinsed away (duration shown by black line).
Octopamine can alter (up, down, or not at all) the spontaneous FR and/or the odor response, so there are
9 possible modulation regimes. This grid of timecourses shows a typical PN from each regime. Top row:
spontaneous FR in unaffected. Middle row: spontaneous FR is boosted. Bottom row: spontaneous FR is
inhibited. First column: odor response is unaffected. Second column: odor response is boosted. Third column:
odor response is inhibited.
26
5. Supporting Information.
5.1. Moth AL neural recording datasets.
Multichannel Recording Methods. Adult male M. sexta were reared in the laboratory at
the University of Washington on an artificial diet under a long-day (17/7 hr light/dark cycle)
photoperiod and prepared for experiments 23 d after emergence. In preparation for recording,
the moth was secured in a plastic tube with dental wax, and the caudal end of the head
capsule was opened and the piece of cuticle that is the attachment site of the pharyngeal
dilator muscles is positioned forward and readhered to the headcapsule with myristic acid. The
proboscis was extended and adhered to a piece of teflon tubing, 7-cm in length, that allowed
movement and extension of the proboscis along its length and at the tip. Behavioral response
to odor stimuli was determined by the percentage of proboscis movement response between
treatments, allowing for a conservative estimate of a learned response. This preparation allows
full access to the exposed AL while having no effect on the moths ability to feed normally.
During recordings, the brain was superfused slowly with physiological saline solution (in mM:
150 NaCl, 3 CaCl2, 3 KCl, 10 N-tris[hydroxymethyl] methyl-2 aminoethanesulfonic acid buffer,
and 25 sucrose, pH 6.9) throughout the experiment.
Olfactory stimuli were delivered to the preparation by pulses of air from a constant air
stream were diverted through a glass syringe containing a piece of filter paper bearing odor
stimuli. The stimulus was pulsed by means of a solenoid-activated valve controlled by the
SciWorks acquisition software (Data Wave Technologies, Longmont, CO, USA). The outlet of
the stimulus syringe was positioned 2 cm from and orthogonal to the center of the antennal
flagellum ipsilateral to the ALs.
We used a classic conditioning paradigm to examine the effects on AL neurons while the
moth learned to associate an odor with a sugar reward. Odor stimuli were delivered in a
three second pulses, and one second after odor onset, the unconditioned stimulus (US; 1 L
of 20% sucrose solution) was applied to the proboscis for ca. 2 s. A ten minute inter-trial
interval separated each training trial, and moths were trained over eight trials. After the
conditioning trials were completed, a test trial was performed during which only the trained
odor was presented to assess the behavioral and odor-driven ensemble responses as a result of
the conditioning treatment.
AL Recordings were made with 16-channel silicon multielectrode recording arrays (a443mm50-177; NeuroNexus Technologies, Ann Arbor, MI, USA). The spatial distribution design
of the recording array suits the dimensions of the AL in M. sexta, with four shanks spaced
125 m apart, and each with four recording sites 50 m apart. The four shanks were oriented
in a line parallel to the antennal nerve. The probe was advanced slowly through the AL
using a micromanipulator (Leica Microsystems, Bannockburn, IL, USA) until the uppermost
recording sites were just below the surface of the AL. Extracellular activity was acquired
with a RZ2 base station (Tucker-Davis Technologies, Alachua, FL, USA) and a RP2.1 real
time processor (Tucker-Davis Technologies) and extracellular activity in the form of action
potentials, or spikes, were extracted from the recorded signals and digitized at 25 kHz using
the Tucker-Davis Technologies data-acquisition software.Threshold and gain settings were ad27
justed independently for each channel, and spikes were captured in the 4-channel, or ‘tetrode,
recording configuration: any spike that passed threshold on one channel triggered the capture
of spikes recorded on the other 3 channels on the same shank. Offline Sorter v.3 (Plexon Neurotechnology Research Systems, Dallas, TX, USA) was used to sort extracellular spikes based
on their waveform shape and spikes were assigned timestamps to create raster plots and calculate peri-stimulus time histograms (PSTH). Only those clusters that were separated in three
dimensional space (PC1PC3) after statistical verification (multivariate ANOVA; P < 0.05)
were used for further analysis (typically 618 units were isolated per ensemble). Each spike in
each cluster was time-stamped, and these data were used to create raster plots and to calculate peristimulus time histograms (PSTHs), interspike interval histograms, cross-correlograms,
and rate histograms.
List of in vivo data sets.
1. AL, odor only: PNs, one odor, no octopamine. 7 preps with 8 - 16 PNs each.
2. AL, odor + octopamine. PNs, one odor, sugar reward. 10 preps with 9 - 21 PNs each.
3. AL + MB, odor + octopamine. PNs and KCs, one odor, sugar reward. 1 prep, with
7 PNs and 12 KCs.
4. AL, odor + octo wash: PNs, one odor, octopamine directly applied to AL. 7 preps: 6
preps with 8 - 13 PNs each; 1 prep with one pheromone-responsive neuron
5. AL, odors only (BEA): PNs, several odors and concentrations. 12 preps with 14 - 17
PNs each.
6. AL, odors only (ESO): PNs, several odors and concentrations. 4 preps with 12 - 14
PNs each.
5.2. Full equations of model dynamics.
(5.1)
(5.2)
(5.3)
(5.4)
(5.5)
(5.6)
τR · duR = fR uR , uL , uS , M L,R , M S,R , M O,R , o(t) + dWR
τP · duP = fP uR , uP , uL , M L,P , M R,P , M O,P , o(t) + dWP
τQ · duQ = fQ uR , uQ , uL , M L,Q , M R,Q , M O,Q , o(t) + dWQ
τL · duL = fL uR , uL , M L,L , M R,L , M O,L , o(t) + dWL
τK · duK = fK uP , uQ , uD , M P,K , M Q,K + dWK
τE · duE = fE uK , uE , M K,E
where
fR = −uR + sigmoid −
fP = −uP + sigmoid −
fQ = −uQ + sigmoid −
fL = −uL + sigmoid −
fK = −uK + sigmoid −
fE = −uE + M K,E uK
I − γ · o(t) · M O,R M L,R uL + I + o(t) · M O,R M S,R uS
I − γ · o(t) · M O,P M L,P uL + I + o(t) · M O,P M R,P uR
I − γ · o(t) · M O,Q M L,Q uL + I + o(t) · M O,Q M R,Q uR
I − γ · o(t) · M O,L M L,L uL + I + o(t) · M O,L M R,L uR
uD + M Q,K uQ + M P,K uP
28
Table 5.1
Variables and parameters for neuronal network model
Symbol
Type
R
P
Q
L
K
E
O
superscript
superscript
superscript
superscript
superscript
superscript
superscript
nG
nS
nQ
nK
nE
scalar
scalar
scalar
scalar
scalar
2000
1
uR
uP
uQ
uL
uK
uE
vector
vector
vector
vector
vector
vector
nG × 1
nG × 1
nQ × 1
nG × 1
nK × 1
nE × 1
uS
uD
vector
vector
M S,R
matrix
nG × nS
Stimulus → uR connections.
M O,R
M O,L
matrix
matrix
nG × nG
nG × nG
Octopamine → uR connections. Diagonal matrix.
Octopamine → uL connections. Diagonal matrix.
M R,L
M R,P
M R,Q
M P,K
M Q,K
M L,R
M L,P
M L,Q
M L,L
M K,E
matrix
matrix
matrix
matrix
matrix
matrix
matrix
matrix
matrix
matrix
nG × nG
nG × nG
nQ × nG
nK × nG
nK × nQ
nG × nG
nG × nG
nQ × nG
nG × nG
nE × nK
Connection
Connection
Connection
Connection
Connection
Connection
Connection
Connection
Connection
Connection
o(t)
γ
function
scalar
0 or 1
0.5
τR
τP
τQ
τL
τK
τE
scalar
scalar
scalar
scalar
scalar
scalar
∗
∗
Size/Value
Description and Remarks
Refers
Refers
Refers
Refers
Refers
Refers
Refers
60
2-4
to
to
to
to
to
to
to
Number
Number
Number
Number
Number
FRs
FRs
FRs
FRs
FRs
FRs
of
of
of
of
of
of
the
the
the
the
the
the
the
of
of
of
of
of
the
the
the
the
the
the
receptor neurons subpopulation.
excitatory projection neurons subpopulation.
inhibitory projection neurons subpopulation.
lateral neurons subpopulation.
kenyon cells subpopulation.
readout extrinsic neurons subpopulation.
octopamine neurotransmitter.
glomeruli in the antenna lobe.
different stimuli (odors).
inhibitory projection neurons.
kenyon cells.
extrinsic neurons.
∗
receptor neurons subpopulation.
exc. projection neurons subpopulation.
inh. projection neurons subpopulation.
lateral neurons subpopulation.
kenyon cells subpopulation. Sparse.
extrinsic neurons subpopulation.
weights
weights
weights
weights
weights
weights
weights
weights
weights
weights
uR → uL .
uR → uP . Diagonal matrix.
uR → uQ .
uP → uK .
uQ → uK .
uL → uR .
uL → uP .
uL → uQ .
uL → uL .
uK → uE .
Flags when octopamine is active (typically during training).
Scaling factor for octopamine’s effects on inhibition. ∗
Each glomerulus receives one RN and one octopamine input, and initiates one PN and one LN.
Octopamine decreases the response to inhibition less than it increases the response to excitation
29
5.3. ANOVA analysis of Network Model learning.
The differential increase in EN response to trained vs control odors was almost always
significant to p < 0.01. When odors’ naive EN response magnitudes differed by > 3, sometimes
differences in either raw increases or percentage increases (not both) did not attain this level
of significance, while the other metric did. An example of learning results when odors’ naive
EN response magnitudes differ widely is shown in Fig 5.1.
The p-values of 336 trained odor/control odor pairs are plotted in Fig 5.2, against the ratio
of their mean naive odor responses µµCT , for 28 moths randomly generated from a template, with
three control odors and one trained odor. Each p-value is for the trained odor vs one control
odor (so there are 12 data points per moth). Column 1 shows p-values for the change in raw
EN response, as in Fig 5.1 (C), trained vs control. Trained odors with very low-magnitude
naive response often did not have raw increases larger than high-magnitude control odors.
Column 2 shows p-values for the percentage change in EN response, as in Fig 5.1 (B), trained
vs control.
Unless the naive EN responses for the two odors were highly disparate (eg by factor of
>3x), the differential increase in EN response of the trained vs control odors is almost always
significant, measured both as raw and as percentage. Fig 5.3 plots the percentage of 336
trained-control pairs that had p-values for both measures of EN response increase (ie as raw
and as percentage) below the listed threshold (eg p = 0.01), for 336 trained-control pairs
whose ratio ( µµCT or µµCT ) is within the bound given on the x-axis. Fig 5.3 shows how many
moths, generated from template with no constraint on unbalanced naive odor EN responses,
had differential post-training EN responses with significance p < 0.01, for both measures (as
raw and as a percentage).
30
Figure 5.1. Effect of training on EN FRs, given odors with unequal naive response magnitudes. When odors induced naive EN responses of very different magnitudes, then trained odor response
increased much more than control odor responses either in raw magnitude, or as a percentage, or both.
A: Typical timecourse showing magnitudes EN responses before and after training the third (red) odor, indicated
by red arrow, over 15 odor stimulations. This corresponds to the third column in panels B - D, at index 3 on
the x-axis. Note that only the third (red) odor’s EN response changes magnitude.
Panels B - D: Changes to ENs during training. x-axis = number of training sessions. Each column shows
results of training a given odor, color coded: blue, black, red. y-axis measures raw EN or percent change in EN.
21 trials per data point.
B: Percent change (from pre-training) in ENs, mean ±2 stds.
C: Raw EN FRs, mean ±2 stds.
D: Changes in raw EN FRs, normalized by trained odor (ie subtract the trained odor’s changes from all odors),
mean ±2 std devs. This shows how far each control odor lagged behind the trained odor.
Note that the trained odor dominates in either raw increase (panels C, D) if naive response to trained odor was
large, or in percent increase (panel B) if naive response to trained odor was small.
31
Figure 5.2. p-values for trained-control odor pairs: A: p-values for change in raw EN responses.
B: p-values for percentage change in EN responses. P-values are sometimes high (for one metric or the other)
when trained and control odors have highly disparately-scaled naive responses µT (= mean raw T ) µC (= mean
raw C). Plots show results given 20 training stimulations.
When µT is larger (right end of x-axis), the p-value for raw change (A) is consistently very low, but the p-value
for percentage change (rB) can be high, since even a small incidental change to a low-intensity odor can be a
large percentage change.
When µC is larger (left end of x-axis), the p-value for percentage change (B) is consistently very low, but the pvalue for raw change (A) can be high, since even a small percentage change to a high-response odor corresponds
to a large raw change.
When naive odor responses are roughly matched, eg within 3x (ie 0.33 to 3), p-values for both raw and percentage
change are very low.
Figure 5.3. Fractions of p-values below 0.01 for trained-control odor pairs In most cases, the
trained odor shows much larger increases in EN response magnitude. A: The percentage of trained-control odor
pairs with EN response magnitudes within the ratios given on the x-axis. B: The percentage of trained-control
pairs, with EN response magnitudes within the ratios given on the x-axis, whose training-induced changes in EN
responses were distinct with p-value < 0.01. Each curve is for a different number of training stimulations. More
training increases distinctions, up to 15 stimulations. But additional training actually hinders distinctions, as
control odor response reinforcement begins to overtake trained odor reinforcement.
32
REFERENCES
[1] Martin Hammer and Randolf Menzel. Multiple sites of associative odor learning as revealed by local brain
microinjections of octopamine in honeybees. Learn Mem, 5(1):146–156, May 1998. 10454379[pmid].
[2] Jeffrey A. Riffell, Leif Abrell, and John G. Hildebrand. Physical processes and real-time chemical measurement of the insect olfactory environment. Journal of Chemical Ecology, 34(7):837–853, Jul 2008.
[3] Rachel I Wilson. Neural and behavioral mechanisms of olfactory perception. Current Opinion in Neurobiology, 18(4):408 – 412, 2008. Sensory systems.
[4] Robert A.A. Campbell and Glenn C. Turner. The mushroom body. Current Biology, 20(1):R11 – R12,
2010.
[5] Andrew M. Dacks, Jeffrey A. Riffell, Joshua P. Martin, Stephanie L. Gage, and Alan J. Nighorn. Olfactory
modulation by dopamine in the context of aversive learning. Journal of Neurophysiology, 108(2):539–
550, 7 2012.
[6] Nicolas Y. Masse, Glenn C. Turner, and Gregory S.X.E. Jefferis. Olfactory information processing in
drosophila. Current Biology, 19(16):R700 – R713, 2009.
[7] Gilles Laurent. Olfactory network dynamics and the coding of multidimensional signals. Nature Reviews
Neuroscience, 3:884 EP –, Nov 2002. Review Article.
[8] Kyle S. Honegger, Robert A. A. Campbell, and Glenn C. Turner. Cellular-resolution population
imaging reveals robust sparse coding in the drosophila mushroom body. Journal of Neuroscience,
31(33):11772–11785, 2011.
[9] Sophie J. C. Caron. Brains don’t play dice—or do they? Science, 342(6158):574–574, 2013.
[10] C. Giovanni Galizia. Olfactory coding in the insect brain: data and conjectures. European Journal of
Neuroscience, 39(11):1784–1795, 2014.
[11] Fei Peng and Lars Chittka. A simple computational model of the bee mushroom body can explain
seemingly complex forms of olfactory learning and memory. Current Biology, 27(2):224 – 230, 2017.
[12] Mark Roper, Chrisantha Fernando, and Lars Chittka. Insect bio-inspired neural network provides new
evidence on how simple feature detectors can enable complex visual generalization and stimulus
location invariance in the miniature brain of honeybees. PLOS Computational Biology, 13(2):1–23,
02 2017.
[13] Thiago S Mosqueiro and Ramn Huerta. Computational models to understand decision making and
pattern recognition in the insect brain. Current Opinion in Insect Science, 6:80 – 85, 2014. Pests and
resistance/Parasites/Parasitoids/Biological control/Neurosciences.
[14] Paolo Arena, Luca Patan, Vincenzo Stornanti, Pietro Savio Termini, Bianca Zpf, and Roland Strauss.
Modeling the insect mushroom bodies: Application to a delayed match-to-sample task. Neural Networks, 41:202 – 211, 2013. Special Issue on Autonomous Learning.
[15] Faramarz Faghihi, Ahmed A. Moustafa, Ralf Heinrich, and Florentin Wrgtter. A computational model
of conditioning inspired by drosophila olfactory system. Neural Networks, 87:96 – 108, 2017.
[16] Peter Dayan and L. F. Abbott. Theoretical Neuroscience: Computational and Mathematical Modeling of
Neural Systems. The MIT Press, 2005.
[17] D. Hubel and T. Wiesel. Receptive fields, binocular interaction, and functional architecture in the cat’s
visual cortex. Journal of Physiology, 160:106–154, 1962.
[18] Kunihiko Fukushima. Neocognitron: A self-organizing neural network model for a mechanism of pattern
recognition unaffected by shift in position. Biological Cybernetics, 36:193–202, 1980.
[19] Jurgen Schmidhuber. Deep learning in neural networks: An overview. Neural Networks, 61(Supplement
C):85 – 117, 2015.
[20] Yann LeCun. Facebook ai director yann lecun on his quest to unleash deep learning and make machines
smarter. IEEE Spectrum, 2015.
[21] Joshua P. Martin, Aaron Beyerlein, Andrew M. Dacks, Carolina E. Reisenman, Jeffrey A. Riffell, Hong Lei,
and John G. Hildebrand. The neurobiology of insect olfaction: Sensory processing in a comparative
context. Progress in Neurobiology, 95(3):427 – 447, 2011.
[22] Pal Kvello, Bjarte Lofaldli, Jurgen Rybak, Randolf Menzel, and Hanna Mustaparta. Digital, threedimensional average shaped atlas of the heliothis virescens brain with integrated gustatory and olfactory neurons. Frontiers in Systems Neuroscience, 3:14, 2009.
[23] Vikas Bhandawat, Shawn R Olsen, Nathan W Gouwens, Michelle L Schlief, and Rachel I Wilson. Sensory
33
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
processing in the drosophila antennal lobe increases reliability and separability of ensemble odor
representations. Nature Neuroscience, 10:1474–1482, 2007.
Emmanuel Perisse, Christopher Burke, Wolf Huetteroth, and Scott Waddell. Shocking revelations and
saccharin sweetness in the study of drosophila olfactory memory. Curr Biol, 23(17):R752–R763, Sep
2013. S0960-9822(13)00921-4[PII], 24028959[pmid].
Maxim Bazhenov and Mark Stopfer. Forward and back: Motifs of inhibition in olfactory processing.
Neuron, 67(3):357 – 358, 2010.
RAA Campbell, KS Honegger, H Qin, W Li, E Demir, and GC Turner. Imaging a population code for
odor identity in the drosophila mushroom body. Journal of Neuroscience, 33(25):10568–81, 2013.
Toshihide Hige, Yoshinori Aso, Gerald M. Rubin, and Glenn C. Turner. Plasticity-driven individualization
of olfactory coding in mushroom body output neurons. Nature, 526:258 EP –, Sep 2015.
M Hammer and R Menzel. Learning and memory in the honeybee. Journal of Neuroscience, 15(3):1617–
1630, 1995.
Stijn Cassenaer and Gilles Laurent. Hebbian stdp in mushroom bodies facilitates the synchronous flow
of olfactory information in locusts. Nature, 448:709 EP –, Jun 2007.
Glenn C. Turner, Maxim Bazhenov, and Gilles Laurent. Olfactory representations by drosophila mushroom body neurons. Journal of Neurophysiology, 99(2):734–746, 2008.
Surya Ganguli and Haim Sompolinsky. Compressed sensing, sparsity, and dimensionality in neuronal
information processing and data analysis. Annual Review of Neuroscience, 35(1):485–508, 2012.
PMID: 22483042.
Ashok Litwin-Kumar, Kameron Decker Harris, Richard Axel, Haim Sompolinsky, and L.F. Abbott. Optimal degrees of synaptic connectivity. Neuron, 93(5):1153 – 1164.e7, 2017.
Shawn R. Olsen, Vikas Bhandawat, and Rachel Irene Wilson. Divisive normalization in olfactory population codes. Neuron, 66(2):287–299, Apr 2010. 20435004[pmid].
Hong Lei, Thomas A. Christensen, and John G. Hildebrand. Local inhibition modulates odor-evoked
synchronization of glomerulus-specific output neurons. Nature Neuroscience, 5:557 EP –, May 2002.
Article.
Baktash Babadi and Haim Sompolinsky. Sparseness and expansion in sensory representations. Neuron,
83(5):1213 – 1226, 2014.
Richard S Sutton and Andrew G Barto. Reinforcement Learning. MIT Press, 1998.
B. Settles. Active Learning. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan
& Claypool, 2012.
Josh Attenberg and Foster Provost. Why label when you can search?: Alternatives to active learning for
applying human resources to build classification models under extreme class imbalance. In Proceedings
of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD
’10, pages 423–432, New York, NY, USA, 2010. ACM.
Pedro D. Maia and J. Nathan Kutz. Reaction time impairments in decision-making networks as a diagnostic marker for traumatic brain injuries and neurological diseases. Journal of Computational
Neuroscience, 42:323–347, 2017.
Elizabeth J. Hong and Rachel I. Wilson. Simultaneous encoding of odors by channels with diverse
sensitivity to inhibition. Neuron, 85(3):573 – 589, 2015.
Wei Ji Ma, Jeffrey M. Beck, Peter E. Latham, and Alexandre Pouget. Bayesian inference with probabilistic
population codes. Nature Neuroscience, 9:1432 EP –, Oct 2006. Article.
Guozhong An. The effects of adding noise during backpropagation training on a generalization performance. Neural Comput., 8(3):643–674, April 1996.
Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow,
and Rob Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2013.
Yoshua Bengio and Asja Fischer. Early inference in energy-based models approximates back-propagation.
arXiv e-prints, abs/1510.02777, October 2015.
Katherine I. Nagel and Rachel I. Wilson. Biophysical mechanisms underlying olfactory receptor neuron
dynamics. Nature Neuroscience, 14:208 EP –, Jan 2011. Article.
Nina Deisig, Martin Giurfa, Harald Lachnit, and Jean-Christophe Sandoz. Neural representation of
olfactory mixtures in the honeybee antennal lobe. European Journal of Neuroscience, 24(4):1161–
1174, 2006.
34
[47] Jeffrey A. Riffell, H. Lei, and John G. Hildebrand. Neural correlates of behavior in the moth manduca
sexta in response to complex odors. Proceedings of the National Academy of Sciences, 106(46):19219–
19226, 2009.
[48] Jeffrey A. Riffell, Hong Lei, Thomas A. Christensen, and John G. Hildebrand. Characterization and
coding of behaviorally significant odor mixtures. Current Biology, 19(4):335 – 340, 2009.
[49] Rachel I. Wilson and Gilles Laurent. Role of gabaergic inhibition in shaping odor-evoked spatiotemporal
patterns in the drosophila antennal lobe. Journal of Neuroscience, 25(40):9069–9079, 2005.
[50] Shawn R. Olsen, Vikas Bhandawat, and Rachel I. Wilson. Excitatory interactions between olfactory
processing channels in the ¡em¿drosophila¡/em¿ antennal lobe. Neuron, 54(4):667, 2008.
[51] Shawn R. Olsen and Rachel I. Wilson. Lateral presynaptic inhibition mediates gain control in an olfactory
circuit. Nature, 452:956 EP –, Mar 2008. Article.
[52] Marcus Sjoholm. Structure and function of the moth mushroom body. Swedish Univ of Agricultural
Sciences, Alrarp, 2006. PhD thesis.
[53] Gregory S.X.E. Jefferis, Christopher J. Potter, Alexander M. Chan, Elizabeth C. Marin, Torsten Rohlfing,
Calvin R. Maurer Jr., and Liqun Luo. Comprehensive maps of drosophila higher olfactory centers:
Spatially segregated fruit and pheromone representation. Cell, 128(6):1187 – 1203, 2007.
[54] Andrew C. Lin, Alexei M. Bygrave, Alix de Calignon, Tzumin Lee, and Gero Miesenböck. Sparse, decorrelated odor coding in the mushroom body enhances learned odor discrimination. Nature Neuroscience,
17:559 EP –, Feb 2014. Article.
[55] Eyal Gruntman and Glenn C. Turner. Integration of the olfactory code across dendritic claws of single
mushroom body neurons. Nature Neuroscience, 16:1821 EP –, Oct 2013. Article.
[56] Sean X. Luo, Richard Axel, and L. F. Abbott. Generating sparse and selective third-order responses in the
olfactory system of the fly. Proceedings of the National Academy of Sciences, 107(23):10713–10718,
2010.
[57] Javier Perez-Orive, Ofer Mazor, Glenn C. Turner, Stijn Cassenaer, Rachel I. Wilson, and Gilles Laurent.
Oscillations and sparsening of odor representations in the mushroom body. Science, 297(5580):359–
365, 2002.
[58] SJ Caron, V Ruta, LF Abbott, and R Axel. Random convergence of olfactory inputs in the drosophila
mushroom body. Nature, 497(5):113–7, 2013.
[59] Randolf Menzel and Gisela Manz. Neural plasticity of mushroom body-extrinsic neurons in the honeybee
brain. Journal of Experimental Biology, 208(22):4317–4332, 2005.
[60] Jeffrey A. Riffell, Hong Lei, Leif Abrell, and John G. Hildebrand. Neural basis of a pollinator’s buffet:
Olfactory specialization and learning in manduca sexta. Science, 2012.
[61] Ana F. Silbering and C. Giovanni Galizia. Processing of odor mixtures in the drosophila antennal
lobe reveals both global inhibition and glomerulus-specific interactions. Journal of Neuroscience,
27(44):11966–11977, 2007.
[62] Desmond J. Higham. An algorithmic introduction to numerical simulation of stochastic differential equations. SIAM Rev., 43(3):525–546, March 2001.
[63] D. O. Hebb. The organization of behavior : a neuropsychological theory. Wiley New York, 1949.
[64] W. Shane Grant, James Tanner, and Laurent Itti. Biologically plausible learning in neural networks with
modulatory feedback. Neural Networks, 88(Supplement C):32 – 48, 2017.
[65] Albert-Laszlo Barabasi and Reka Albert. Emergence of scaling in random networks. Science,
286(5439):509–512, 1999.
[66] Eli Shlizerman, Jeffrey A. Riffell, and J. Nathan Kutz. Data-driven inference of network connectivity
for modeling the dynamics of neural codes in the insect antennal lobe. Frontiers in Computational
Neuroscience, 8:70, 2014.
35
| 9cs.NE
|
The Causal Role of Astrocytes in Slow-Wave Rhythmogenesis:
A Computational Modelling Study
Leo Kozachkov, Konstantinos P. Michmizos
Laboratory for Computational Brain
Department of Computer Science, Rutgers University, Piscataway, NJ, USA
Corresponding author: Konstantinos P. Michmizos
Laboratory for Computational Brain
Department of Computer Science
Rutgers University
address: 110 Frelinghuysen Road, Piscataway, NJ, 08854
email: [email protected]
tel: 848-445-8841
fax: 732-445-0537
1. Abstract
Finding the origin of slow and infra-slow oscillations could reveal or explain brain mechanisms in
health and disease. Here, we present a biophysically constrained computational model of a neural
network where the inclusion of astrocytes introduced slow and infra-slow-oscillations. We show how
astrocytes can modulate the fast network activity through their slow inter-cellular Ca2+ wave speed and
amplitude and possibly cause the oscillatory imbalances observed in diseases commonly known for such
abnormalities, namely Alzheimer’s disease, Parkinson’s disease, epilepsy, depression and ischemic stroke.
This work aims to increase our knowledge on how astrocytes and neurons synergize to affect brain
function and dysfunction.
2. Introduction
Brain oscillations are periodic fluctuations in neuronal excitability that convey information in
space and time [1, 2]. Slow oscillations (SOs; 0.1–1 Hz) and infra-slow oscillations (ISOs; 0.01–0.1 Hz) [3]
characterize the idle “default” state of the brain [4] and orchestrate other default-state patterns such as
delta-waves, sleep-spindles, and K-complexes [5, 6]. While the origin of these oscillations remains elusive
[7], they have been associated with primitive brain functions, such as temperature regulation [8],
hormone secretion [9] and memory consolidation [10], whereas their disturbance is often related to
neurological disorders [11, 12]. Interestingly, SOs and ISOs share several distinctive patterns with the
oscillatory activity exercised by the networks of the most abundant type of glial cells, the astrocytes [13,
14]: The neural activity recurs in a wave-like pattern with the same frequencies that astrocytes use to
communicate with each other [15-17]; it can also be present even in the absence of stimulation, similarly
to spontaneous inter-astrocyte communication [18-21].
Astrocytes, long believed to only give structural and nutritional support to neurons, are now
regarded as key cells in neural information processing [22-26]. Astrocytes use processes, the long, thin
tendrils protruding from their main cell bodies, to make contact with local neural synapses [26-29]. This
three-body arrangement among pre- and post-synaptic neurons and the astrocyte process is named
tripartite synapse [23, 26]. An individual astrocyte can form up to 105 tripartite synapses [23, 26] and
envelop up to 8 neuronal somata [30], with a single neuron belonging exclusively to a single astrocyte
[31].
Contrary to neurons, which transmit information quickly by spikes, astrocytes are believed to use
slowly propagating intercellular Ca2+ waves [16, 18, 32-36]. Ca2+ waves appear to recur at discrete
origination sites with an infra-slow frequency [20, 33]. Although the mechanism by which intercellular
Ca2+ waves propagate is still unclear, both experimental and theoretical studies suggest that it involves a
combination of direct, gap-junctional communication [32, 34] as well as extracellular, diffusion-driven
communication [16, 37]. The average speed of astrocyte Ca2+ wave varies considerably between studies
[36], with values ranging from 8 μm/s [38] to 60 μm/s [37, 39]. Reported propagation distances for Ca2+
waves vary between 4 astrocytes [16] and 100 astrocytes [36]. While the exact role of Ca2+ waves remains
unclear, even early studies showed that they can have both excitatory and inhibitory neuromodulatory
effects [34, 35].
Astrocytes form dense syncytia, functional networks spanning across brain regions. Such regions
include the retina [18, 34, 35], hippocampus [24, 27, 30, 40, 41], cortex [7, 42, 43], cerebellum [33] and
thalamus [15]. Astrocyte networks tile the brain in a circumferential grid pattern [30] forming synaptic
1
islands, which are non-overlapping domains of influence over nearby neurons [31]. In the synaptic island
view of neural organization, “clusters of synapses are coordinated at nodes that correspond to the domain
of individual astrocytes” [31]. However, to date, despite the multitude of studies that experimentally
relate the astrocytic function with the slow oscillatory activity, there is still scarcity of studies, either
experimental or theoretical, that have proposed a generation mechanism for these oscillations and
therefore aim to understand how astrocytes modulate local neural activity in biological networks.
Being regarded as “a primary source for generating neural activity” [19], astrocytes have long
been linked to SOs and ISOs through interacting with two synaptic sites [7, 41-43]. Specifically, astrocytes
are found to modulate oscillations through co-activation of neuronal Adenosine A1 and N-methyl-Daspartate (NMDA) receptors [24, 44]. However, up to now there is no suggested mechanism that connects
the neural modulation to the long been observed slow oscillatory patterns of activity.
Toward this goal, we built a neurophysiologically plausible model of an astrocyte-neuron network
that proposes a possible mechanistic explanation for the origination of SOs and ISOs as a result of
astrocyte-neuron interactions during Ca2+ wave events. By combining astrocytic pre-synaptic inhibition
[22] with post-synaptic excitation [45], and replicating the connectome of astrocyte networks [30, 31],
our proposed model faithfully reproduced the slow oscillations of neural activity, namely the ISO, derived
from the Ca2+ wave origination rate, and the SO, coming from the wave propagation speed. By varying the
parameters of our model within their reported range, we also aimed to infer the possible key mechanisms
for the oscillatory imbalances observed in case of diseases commonly known for exhibiting slow oscillatory
abnormalities, namely Alzheimer’s disease (AD), Parkinson’s disease (PD), epilepsy, ischemic stroke and
depression.
3. Methods
a. Network Architecture
The astrocyte-neural network model consists of generic excitatory neurons organized in a
feedforward fashion. The first layer of neurons served as the input layer to the network. The neurons at
that layer were driven by a current drawn from a uniform distribution between 0 and 360 pA and exhibited
an average firing rate of ~10 spikes per second throughout the duration of the simulation. In the
subsequent layers, synaptic islands were formed between a single astrocyte and 4 neurons, to replicate
the average number of neurons that create tripartite synapses with a single astrocyte [31]. We fixed the
number of layers at 5, to replicate the approximate number of radially positioned astrocytes on which a
single Ca2+ wave has been reported to spread in in-vitro connectivity studies [30, 36]. The amplitude of
the Ca2+ wave decreased as it propagated through the 5 astrocytes by 7.9% per astrocyte. Each neuron
was fully connected to the neurons of the next layer. Neurons in layers other than the input layer were
additionally driven by a small noisy basal current, drawn from a uniform distribution between 0 and 180
pA to represent all the other inputs that a neuron receives from sources outside the network [15, 46];
this, in conjunction with astrocytic stimulation, resulted in them having average firing rates of firing rates
of roughly [14, 20, 26, 32, 38] spikes/sec throughout the course of the simulation, for each layer,
respectively [47].
The astrocytes were modelled as point cells positioned 50μm away from one another, in
accordance with experimental data [30]. The Ca2+ wave spontaneously originated from a discrete
origination site, as reported in in-vitro studies, e.g., [19, 33], and propagated away with a constant velocity
2
of 25 μm/s, a value at the center of the physiological in-vitro range (8μm/s to 60μm/s) [37-39] which
coincides with the reported average wave speed [36].
b. Spiking Neuron Model
The spiking neurons were simulated as Leaky-Integrate and Fire (LIF) models [48]. The parameters
for membrane resistance, R, and capacitance, C, were adopted from [49, 50] and are given in Table 1.
c. Astrocyte Model
Modeling the Intercellular Ca2+ Wave
The intracellular Ca2+ level in astrocyte 𝑖𝑖 at time t, 𝑐𝑐𝑖𝑖 (𝑡𝑡) was modeled by the separable equation:
(1)
𝑐𝑐𝑖𝑖 (𝑟𝑟𝑖𝑖 , 𝑡𝑡) = 𝑎𝑎𝑖𝑖 ∗ 𝑔𝑔𝑖𝑖 (𝑟𝑟𝑖𝑖 , 𝑡𝑡)
where 𝑟𝑟𝑖𝑖 is the distance of astrocyte 𝑖𝑖 from the origination site (𝑟𝑟0 ≡ 0), 𝑎𝑎𝑖𝑖 (𝑡𝑡) and 𝑔𝑔𝑖𝑖 (𝑡𝑡) are, respectively,
the amplitude and the shape of the intercellular Ca2+ wave at astrocyte 𝑖𝑖 and time 𝑡𝑡. Following
experimental evidence, we simulated the form of the Ca2+ wave, 𝑔𝑔𝑖𝑖 (𝑟𝑟𝑖𝑖 , 𝑡𝑡), as the bi-exponential function
[51]:
𝑎𝑎𝑖𝑖 as a decaying exponential:
𝑔𝑔𝑖𝑖 (𝑟𝑟𝑖𝑖 , 𝑡𝑡) = 𝑒𝑒𝑒𝑒𝑒𝑒 �−
𝑟𝑟𝑖𝑖 −𝜇𝜇∙𝑣𝑣∙𝑡𝑡
�−
𝜏𝜏𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝑒𝑒𝑒𝑒𝑒𝑒 �−
𝑎𝑎𝑖𝑖 = 𝑎𝑎0 ∗ 𝑛𝑛 ∗ exp �−
𝑟𝑟𝑖𝑖 −𝜇𝜇∙𝑣𝑣∙𝑡𝑡
�
𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
𝑖𝑖
�,
𝜏𝜏𝑑𝑑
(2)
(3)
where 𝜏𝜏𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 , 𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 were adopted from [29, 51] (see Table 1) and control the rise and fall time of the Ca2+
wave; 𝜏𝜏𝑑𝑑 controls the magnitude of amplitude fall-off between astrocytes, and µ (µ=1 unless otherwise
stated) is a parameter aimed to capture the permeability of the cells through which the wave propagates
with speed 𝑣𝑣 --lumping together all the sub-mechanisms of gap-junctional and extracellular
communication. The constant n is the normalization constant for biexponential distribution, which is a
function of the rise and decay parameters (supplementary material). Following modelling studies that
showed a relatively small, linear fall-off in amplitude for intercellular Ca2+ waves, we made 𝜏𝜏𝑑𝑑
comparatively large (Table 1) [38]. We calculated the time it takes for the inter-cellular Ca2+ wave peak to
propagate between astrocytes 𝑖𝑖 and 𝑖𝑖 + 1, 𝛥𝛥𝑡𝑡𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 , as follows: We first solved for the maxima of 𝑔𝑔:
and then calculated 𝛥𝛥𝑡𝑡𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 as:
𝜕𝜕𝑔𝑔𝑖𝑖 (𝑟𝑟𝑖𝑖 ,𝑡𝑡)
� 𝑟𝑟𝑗𝑗
𝜕𝜕𝜕𝜕
= 0, 𝑗𝑗 = {𝑖𝑖, 𝑖𝑖 + 1},
𝑖𝑖+1
𝑖𝑖
− 𝑡𝑡𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
=
𝛥𝛥𝑡𝑡𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 = 𝑡𝑡𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
Tripartite Synapse Model
𝑟𝑟𝑖𝑖+1 − 𝑟𝑟𝑖𝑖
𝜇𝜇𝜇𝜇
(4)
(5)
To model the effects of the astrocytes on the synaptic level, we used a modification of the TUM
dynamic synapse model [50] known as the Gatekeeper model of the tripartite synapse [22]. Specifically,
the dynamics of synaptic vesicle release and recovery in this model are described by:
3
Table 1. Parameter values for the neural-astrocytic network.
Neuron Parameters
[49, 50]
𝑅𝑅 = .6 𝐺𝐺𝐺𝐺
𝐶𝐶 = 100 𝑝𝑝𝑝𝑝
𝑣𝑣𝑡𝑡ℎ𝑟𝑟𝑟𝑟𝑟𝑟ℎ = 25 𝑚𝑚𝑚𝑚
𝑣𝑣𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 = −65 𝑚𝑚𝑚𝑚
Astrocyte Parameters
[29, 30, 36, 52, 53]
𝜏𝜏𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 = 1.1 𝑠𝑠
𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 = 0.5 𝑠𝑠
𝑟𝑟𝑖𝑖+1 − 𝑟𝑟𝑖𝑖 = 50 𝜇𝜇𝜇𝜇
𝜏𝜏𝑑𝑑 = 12 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐
𝑆𝑆𝑆𝑆𝑆𝑆
𝜏𝜏𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 = 300𝑚𝑚𝑚𝑚
𝑆𝑆𝑆𝑆𝑆𝑆
𝜏𝜏𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
= 50𝑚𝑚𝑚𝑚
𝑎𝑎0 = 500𝑛𝑛𝑛𝑛
Tripartite Synapse Parameters
[22, 50]
𝜏𝜏𝐶𝐶𝐶𝐶 = 4 𝑠𝑠
𝜏𝜏𝑖𝑖𝑖𝑖 = 50 𝑚𝑚𝑚𝑚
𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟 = 100 𝑚𝑚𝑚𝑚
𝑢𝑢 = 0.1 (𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢)
𝜅𝜅 = 0.5 𝑠𝑠
2+
�𝐶𝐶𝑎𝑎𝑡𝑡ℎ𝑟𝑟𝑟𝑟𝑟𝑟ℎ � = 196.69𝑛𝑛𝑛𝑛
𝐴𝐴 = 1000 𝑝𝑝A
𝑑𝑑𝑑𝑑
𝑓𝑓
2+
= �−
� + (1 − 𝑓𝑓) ∙ 𝜅𝜅𝜅𝜅�[𝐶𝐶𝑎𝑎2+ ] − �𝐶𝐶𝑎𝑎𝑡𝑡ℎ𝑟𝑟𝑟𝑟𝑟𝑟ℎ
��
𝑑𝑑𝑑𝑑
𝜏𝜏𝐶𝐶𝐶𝐶
(6)
𝑧𝑧
𝑑𝑑𝑑𝑑
=
− (1 − 𝑓𝑓) ∙ 𝑢𝑢 ∙ 𝑥𝑥 ∙ 𝛿𝛿�𝑡𝑡 − 𝑡𝑡𝑠𝑠𝑠𝑠 � − 𝑥𝑥 ∗ 𝜂𝜂(𝑓𝑓)
𝑑𝑑𝑑𝑑 𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟
(7)
𝑦𝑦
𝑧𝑧
𝑑𝑑𝑑𝑑
=
−
𝑑𝑑𝑑𝑑 𝜏𝜏𝑖𝑖𝑖𝑖 𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟
(9)
𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠 = 𝐴𝐴 ∗ 𝑦𝑦(𝑡𝑡),
(10)
𝑦𝑦
𝑑𝑑𝑑𝑑
=−
+ (1 − 𝑓𝑓) ∙ 𝑢𝑢 ∙ 𝑥𝑥 ∙ 𝛿𝛿�𝑡𝑡 − 𝑡𝑡𝑠𝑠𝑠𝑠 � + 𝑥𝑥 ∗ 𝜂𝜂(𝑓𝑓)
𝜏𝜏𝑖𝑖𝑖𝑖
𝑑𝑑𝑑𝑑
(8)
where 𝑓𝑓 is a gating variable that models Ca2+-dependent presynaptic inhibition at the tripartite
synapse, 𝜏𝜏𝐶𝐶𝐶𝐶 is the decay constant of 𝑓𝑓, 𝜅𝜅 controls the rise time of 𝑓𝑓, 𝛩𝛩(𝑥𝑥) is the Heaviside function, and
2+
] is the concentration of intracellular Ca2+, [𝐶𝐶𝑎𝑎2+ ], in a given astrocyte needed to activate 𝑓𝑓. The
[𝐶𝐶𝑎𝑎𝑡𝑡ℎ𝑟𝑟𝑟𝑟𝑟𝑟ℎ
variables 𝑥𝑥, 𝑦𝑦, and 𝑧𝑧 represent the fractions of synaptic resources in a recovered, active, and inactive state,
respectively; 𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟 and 𝜏𝜏𝑖𝑖𝑖𝑖 are the characteristic times of postsynaptic currents (PSCs) decay, and recovery
time from synaptic depression, respectively, and 𝑢𝑢 is the fraction of 𝑥𝑥 released when a spike arrives at the
synapse at time 𝑡𝑡𝑠𝑠𝑠𝑠 . The noise term, 𝜂𝜂(𝑓𝑓), is a stochastic term introduced into the model to fit the
experimentally observed increase of PSCs during a Ca2+ wave event and to compensate for the lack of
direct NMDAR-mediated effects on the postsynaptic neuron from astrocyte-derived gliotransmission in
the Gatekeeper model [22]. The total synaptic current in the gatekeeper model is given by:
where A is a constant multiplicative factor (Table 1), which controls for the network connectivity: When
𝐴𝐴 = 0, then 𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠 = 0 and the neurons do not communicate at all; when 𝐴𝐴 > 0, pre-synaptic spikes
trigger the associated post-synaptic neuron with strength increasing with 𝐴𝐴.
Modifications to the Gatekeeper Model
Since both excitatory and inhibitory tripartite modulations have been implicated in
rhythmogenesis [42] ,we directly incorporated excitatory NMDAR-mediated effects into the synapse. To
do so, we modified the Gatekeeper model by removing the noise term, 𝜂𝜂(𝑓𝑓), in an approach similar to
[52], and directly modeled the dynamics of astrocyte-derived NMDAR slow-mediated currents (SICs) into
the post-synaptic neuron, as explained below.
4
The shape, amplitude and time correlations of SICs with respect to astrocyte Ca2+ levels were
modeled based on experimental data. Specifically, in-vitro and in-vivo studies reveal that SICs are well fit
to bi-exponential distributions, have a rapid rise time (on the order of tens of milliseconds), and a
comparatively larger decay time (on the order of hundreds of milliseconds) [27, 53-55]. Furthermore, SICs
are correlated with Ca2+ wave peaks in both time [56] and amplitude [45, 53]. The amplitude dependence
of SICs on astrocyte Ca2+ followed the experimentally fit function [45, 53]:
𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 2.11 ∙
where
𝑢𝑢𝑢𝑢
𝑐𝑐𝑚𝑚2
(11)
∙ 𝑙𝑙𝑙𝑙(𝑤𝑤) ∙ 𝛩𝛩�𝑙𝑙𝑙𝑙(𝑤𝑤)�,
𝑤𝑤 = [𝐶𝐶𝑎𝑎2+ ]/𝑛𝑛𝑛𝑛 − 196.11
(12)
and 𝛩𝛩 (x) was the Heaviside function. The total equation for SIC dynamics, activated at every Ca2+ peak,
was:
2+
𝐼𝐼𝑆𝑆𝑆𝑆𝑆𝑆 (𝑡𝑡) = 𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 �[𝐶𝐶𝑎𝑎2+ ] = �𝐶𝐶𝑎𝑎𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
� � ∗ 𝑛𝑛 ∗ �𝑒𝑒𝑒𝑒𝑒𝑒 �−
𝑖𝑖
𝑡𝑡
𝑆𝑆𝑆𝑆𝑆𝑆 � −
𝜏𝜏𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝑒𝑒𝑒𝑒𝑒𝑒 �−
𝑡𝑡
𝑆𝑆𝑆𝑆𝑆𝑆 �� ,
𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
(13)
where n is a normalization constant for exponential distributions, which assures that the peak of the biexponential distribution before being multiplied by 𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 is equal to one—a derivation is provided in the
supplementary material. In other words, upon reaching a Ca2+ peak, an astrocyte releases gliotransmitters
that lead to NMDAR mediated SICs, with the SIC amplitude being logarithmically proportional to the Ca2+
wave amplitude. Overall, the total current injected into a post-synaptic neuron of a tripartite
synapse, 𝐼𝐼𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 , was the sum of the synaptic current, 𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠 , the astrocyte-derived SIC, 𝐼𝐼𝑆𝑆𝑆𝑆𝑆𝑆 , and the basal
current 𝐼𝐼𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 :
𝐼𝐼𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠 + 𝐼𝐼𝑆𝑆𝑆𝑆𝑆𝑆 + 𝐼𝐼𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 .
Modeling Astrocytic Dysfunction
(14)
Astrocytes have long been linked to the pathophysiology of brain diseases [57, 58]. Despite the
highly variant experimental results, the intrinsic variability of any brain and the multifaceted complexity
of its diseases, a striking observation is that the amplitude as well as the speed of the Ca+2 wave are
consistently involved in these diseases. While there is no consensus of what constitutes a “physiological”
Ca2+ wave amplitude in vivo, numerous studies support a range for its value to be [250-1000] nM [23, 53,
59]. Therefore, we modeled the effects of astrocyte Ca2+ wave amplitude on brain pathophysiology, at
least in terms of power differences observed in ISO and SO, by varying the amplitude of the initial Ca2+
wave, 𝑎𝑎0 , in the reported range. The Python model code, allowing the full replication of the simulations
for Figures 3-6, is available at http://combra.cs.rutgers.edu/astrocyence.
4. Results
a. Network Connectivity and Dynamics
Following the above biological constrains, our neuron-astrocyte network had astrocytes that were
radially distributed around a central astrocyte, which initiated a Ca2+ wave (Fig. 1a). Ca2+ waves can be
5
initiated by an astrocyte either spontaneously, as in our current implementation, or upon sensing the
nearby neural activity. Each astrocyte defined a synaptic island, as has been found [31]. The neurons in a
particular synaptic island connected in a feedforward manner to the next island (Fig. 1b); Astrocytes were
connected via gap junctions (Fig. 1c), forming a fully connected unidirectional network. Neuron-to-neuron
connections occurred in the presence of an astrocyte process forming a tripartite synapse (Fig. 1d) [23].
The Ca2+ wave decreased in amplitude as it propagated across the 5 astrocytes (Fig. 2 and eqs. 1-3).
b. Slow Oscillations Emerging from Astrocyte-Neuron Interactions as a Network
Characteristic
As the Ca2+ wave propagated away from the central astrocyte (Fig. 3a), it activated pre-synaptic
Adenosine A1 receptors (eqs. 6-9) and post-synaptic NMDAR-mediated SICs (eqs. 12, 13). As the wave
propagated along the layers of the network via the synaptic islands, it caused a transient increase in the
post-synaptic activity of neurons on each layer (Fig. 3b). An elevation in the neural spiking activity of an
island propagated between subsequent islands; this caused a similar transient increase of the spiking
activity on all subsequent islands (Fig. 3c), which depended on the strength of neural connectivity between
layers. Critically, there were two frequency components, one inside the ISO and the other inside the SO
range, that were observed in the firing rate of the network, ranging from the single-neuron to the singlelayer to, ultimately, the aggregate spike activity (Fig. 4). The two frequencies emerged from two distinct
astrocyte-derived processes, as explained below. In our network, the ISO increased with the frequency of
the Ca2+ wave origination rate while the SO increased with the speed of Ca2+ wave propagation. In this
study, we kept the origination rate constant at 0.1 Hz, a value commonly reported in in-vivo studies [20].
We also chose a wave speed of 25μm/s, which is in the middle of the physiological range (Table 1). This
resulted to an SO of 0.5Hz; Fig. 5a further examines the relationship between the SO frequency and Ca2+
wave speed velocity, within the range commonly reported for normal brain tissues, showing the slope of
the linear relationship to be equal to the inverse of the inter-astrocyte distance.
c. Frequency analysis of aggregate firing activity motivated by pathophysiology
Pathophysiological brain states exhibit disease-specific changes in oscillations [60] as well as
astrocytic dysfunctions. We used our model to explore the potential role of the astrocyte Ca2+ wave
dynamics in pathophysiological oscillations. In this study, we focused on the effect of relevant model
parameters to the oscillations’ power, which are commonly imbalanced in various brain diseases.
We first examined how the ISO and SO power depended on two variables that control the synaptic
dynamics: f, the gating variable, which is related to the activation of the A1 adenosine receptors,
and 𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 , which is the amplitude of the SICs (Fig. 5b-c). When we increased 𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 , the transient increase
in the firing rate became more prominent and this resulted to an SO power increase. The same effect was
achieved by increasing the gating strength, f, which inhibited transmission of the pre-synaptic spiking
activity and, therefore, made the astrocyte-derived SIC signal more prominent (Fig. 5c). The ISO power
was mainly related to the gating strength and to a lesser extent to the 𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 (Fig. 5b).
We also examined how the ISO and SO power was modulated by network connectivity and the
amplitude of the astrocyte Ca2+ waves. While the ISO power linearly increased with the network
connectivity (𝑅𝑅 2 = 0.99), the SO power stayed relatively constant (Fig. 6a). The SO and ISO power
depended linearly (𝑅𝑅 2 = 0.977, 𝑅𝑅 2 = 0.95 ) on the Ca2+ wave amplitude (Fig. 6b). These results are
interesting in light of the fact that a multitude of experimental studies show a positive correlation between
the amplitude of astrocyte Ca2+ waves and slow-/delta-wave power in pathophysiological brain states.
6
Figure 1. The architecture of the proposed astrocyte-neuron model a. The characteristic tiling pattern of
astrocytes in the brain. The Ca2+ wave progenitor astrocyte in denoted in black. 1b. The synaptic island
organization of astrocyte-neuron networks. An astrocyte for a given layer entirely envelops up to eight
neurons and is associated with up to 105 tripartite synapses c. An astrocyte-astrocyte gap junction; d. A
tripartite synapse.
Figure 2. Exponential decay of the Ca2+ wave amplitude per astrocyte. The peak amplitude, 𝑎𝑎𝑖𝑖 , for
𝑖𝑖
𝜏𝜏𝑑𝑑
astrocyte i was given by 𝑎𝑎𝑖𝑖 = 𝑎𝑎0 ∗ exp �− �, where 𝑎𝑎0 was the initial Ca2+ wave amplitude and 𝜏𝜏𝑑𝑑 was
the fall-off constant. 𝑎𝑎𝑖𝑖 is shown as a dotted line.
7
Figure 3. Astrocyte Ca2+ waves trigger excitatory post-synaptic NMDAR-mediated SICs along the wave
path, while simultaneously inhibiting pre-synaptic transmission via Adenosine A1 receptor activation. This
results in a rhythmic modulation of neuronal firing patterns, primarily in the slow and infra-slow frequency
bands (smoothing method: Hanning Window Length of 1s).
8
Figure 4. Time-series and time-frequency analysis of the aggregate neural network activity indicating
two distinct frequencies: the ISO (0.1 Hz) and the SO (0.5 Hz). The ISO corresponds to the frequency of
the Ca2+ wave initiation for the first astrocyte, while the SO is dependent on the Ca2+ wave propagation
speed. Bottom subplot: The thick black line is a smooth version of the network firing rate (grey thin line,
bin size = 20ms) (Hanning Window Length = 1s); Right subplot: The periodogram of the smoothed firing
rate (Length of FFT= 106 ).
9
Figure 5. a) Linear dependence of peak SO position on velocity; the independence of ISO position on
velocity. The slope of the SO line is proportional to the inverse of the inter-astrocyte distance, as expected
from theoretical considerations (see eq. 5); the standard error for both regressions was < 10-6 Hz. b) The
ISO power as a function of presynaptic gating strength and SIC amplitude, showing the dominance of the
SO over frequency space at high SIC amplitudes. c) The same plot for SO. As the presynaptic signal is gated,
the astrocyte-derived SIC is expressed more clearly.
Figure 6. The dependence of the ISO and the SO on a) network connectivity and b) Ca2+ wave amplitude,
motivated by brain pathophysiology.
10
5. Discussion
Our study proposes a mechanistic explanation for the long-suspected causal role of astrocytes on
oscillations in the infra-slow and slow frequency bands. The main contribution of this work is that ISO and
SO, two prominent and wide-spread slow oscillations modulating faster oscillations, arise as a property of
neural-astrocytic network interactions. Our neurophysiologically plausible model suggests that these
interactions are influenced by factors such as astrocyte connectomics, Ca2+ wave dynamics, and synaptic
efficiency. As a validation method for our model, we show that it can consistently support a linkage of the
widely reported oscillatory imbalances in brain disorders to known astrocytic dysfunctions in these
diseases.
Questions concerning the origin and the mechanisms of slow oscillations arose even from the very
first observation of their presence in the cortex [61]. In cortical areas, slow oscillations appear to be
generated locally, i.e. “in cortico-cortical networks, since they survive thalamectomy, but not the
disruption of cortico-cortical connections” and are further found to modulate “delta, theta, spindle, alpha,
beta, gamma and ripple oscillations” [62]. In healthy humans, each individual slow wave cycle has a
distinct origin and propagates uniquely across brain areas [63]; Similar patterns were also found in animal
models [64, 65]. These findings corroborate that slow oscillations emerge as a network mechanism
radiating across large brain areas and starting from distinct sources, a structural and functional pattern
that is consistent with that found in astrocytes.
Indirect evidence on how astrocytes participate in the brain’s computations suggests that they
organize neurons into functional, synchronous clusters, and integrate information across large brain
regions [43]. Alongside the structural similarity, the similarity in frequency profiles of astrocytic Ca2+
activity and neural spiking activity during ISOs is striking. Lőrincz et al. have reported thalamic astrocytes
exhibiting in situ spontaneous slowly propagating intracellular Ca2+ oscillations in an almost identical range
of frequencies (0.003–0.1 Hz) to the neuronal ISO, given rise by a slow ‘wave-like’ co-activation of different
neurons and a sometimes biphasic nature of the long-lasting hyperpolarizing potentials [15]. The extent
of the above similarities spreading across the structural and functional domains has rarely been
overlooked as a coincidence.
In fact, the role of astrocytes on brain function has long been speculated [66]. More than half a
century ago, Robert Galambos recognized the potential of “an alternate theory in which a glia-neural
complex of cells forming the functional unit” of brain computation [67]. In recent years, both astrocytes
and slow oscillations have, in separate studies, been implicated in cognitive performance: Stimulating
slow wave oscillations via transcranial stimulation is known to improve performance in memory tasks [68];
it is also known to increase astrocyte Ca2+ activity [69]. Interestingly, when human glial progenitor cells
were engrafted into the forebrains of neonatal rodents, the maturated recipient rodent brain “exhibited
large numbers and high proportions of both human glial progenitors and astrocytes” and, consequently,
outperformed their normal peers significantly on memory and learning tasks [70]. A potential explanation
for the observed enhancement of cognitive abilities is that astrocyte might evoke a phase-of-firing
modulation of neural activity by astrocytes [71]. Our model aligns with this hypothesis as astrocytes can
increase the firing of neurons at a certain phase within a period (Fig. 3) and also increase the SO and ISO
power by increasing their Ca2+ wave amplitude (Fig. 6b): This transient increase of firing rate could
facilitate the firing of the postsynaptic neurons for a given input by decreasing the entropy of the signal
passing through the synapses at these moments.
11
Several experimental studies have targeted astrocytes as a possible source of slow-wave neural
oscillations [15, 42, 43]. In agreement with our modeling approach (eq. 11-14; Fig. 5b-c), these studies
have emphasized the contribution of astrocytes on simultaneous presynaptic inhibition and postsynaptic
excitation and further suggested that their combination might generate the slow wave oscillations.
Astrocytes have also been associated with blood-oxygen-level-dependent (BOLD) signals, which govern
the ISO and SO frequency range, both in resting-state [3, 72, 73], and stimulus-evoked functional magnetic
resonance imaging (fMRI) studies [74]. Interestingly, a change in gliotransmission is found to significantly
alter the power of slow-wave oscillations [42] and, relatedly, modulate cortical UP and DOWN states [7].
Interestingly enough, our model suggests that the astrocytic network can modulate the firing rate at the
SO range by simply varying the inter-cellular Ca2+ wave propagation speed (Fig. 5a). This proposes a
potential mechanism where different astrocytic networks can change their own Ca2+ wave propagation
speed to allow neural encoding of information through amplitude, frequency or amplitude-frequency
modulation, a hypothesis that has been proposed by previous computational studies [16].
The type of generation mechanism that we propose here seems to be also present in glia-like
support cells, which behave similarly to astrocytes. During early development, glia-like inner support cells
within Kolliker’s organ spontaneously generate Ca2+ waves which propagate radially outward at a speed
comparable to that of astrocytes [75]. They are also organized into clusters roughly 60μm apart [76] , close
to the distance between synaptic islands (50μm). Notably, neurons within the developing auditory system
were found to receive slow inward currents with a frequency of roughly 0.2Hz; these SIC’s were coincident
with inner support cell Ca2+ wave events 85% of the time [75]. The reported Ca2+ waves speeds were
between 5 and 15 μm/s. Based on our model (eq. 5; Fig. 5a), for an inter-astrocyte distance of 60μm, a
frequency of 0.2 Hz would be generated by a Ca2+ wave propagation speed of 12 μm/s, which is within the
experimental reported range of values.
Implications to brain pathophysiology and its treatment
Alongside the mounting evidence that neuron-astrocyte interaction plays an important role in
brain function, many studies are also suggesting that a dysfunction in their interactions may accompany
neurological disorders. On one hand, the most commonly reported astrocytic dysfunction is Ca2+ wave
amplitude changes; on the other hand, the most commonly examined slow-oscillation imbalances are
power changes. Our model suggests that an increase of the Ca2+ wave amplitude results to an SO and ISO
power increase (Fig. 6b); it further suggests that increasing the connectivity among neurons would cause
the ISO, but not the SO, power to increase (Fig. 6a). ISO and SO power changes can also be achieved
through intrinsic astrocytic or synaptic mechanisms which would affect the SIC amplitude or the
presynaptic gating strength (Fig. 5b-c). It is interesting to see how well our modeling results align with
experimental studies that have examined, separately, how brain disorders affect the astrocytic Ca2+ wave
changes and the oscillatory imbalances.
Evidence for astrocyte involvement in major depression disorder (MDD) suggests that the
amplitude of astrocyte Ca2+ waves is altered as a result of the illness. Specifically, depressive rats exhibit
decreased levels of ATP, which is known to trigger and enhance astrocyte Ca2+ waves [77]; in line with our
modeling results, depressive patients exhibit lower power in delta/slow wave activity than the normal
population [78, 79]. A striking observation is that trans-cranial stimulation, when used to alleviate the
symptoms of MDD, is reported to both increase delta/slow oscillation activity and increase astrocyte Ca2+
activity in separate studies [69, 80].
12
The causal role of astrocytes in the onset of AD and epilepsy has long been speculated and
confirmed [81-85]. In AD, the cognitive loss is linked to the dysfunction of synaptic plasticity which is
known to be maintained, at least partially, by the astrocytes. A pathological hallmark of AD is the increased
levels of β-Amyloid (Αβ) [86-88] which is associated with a decrease of the astrocytic capacity to
internalize Αβ [89, 90]. The effect of blocking astrocytic purinergic receptors has long been known to cause
a reduction in Ca2+ wave velocity and amplitude, resulting to a significant decrease in the wave spread
[91]. Studies have shown that cultured astrocytes respond to elevation in environmental Aβ by increasing
the speed and amplitude of their calcium waves: the former by up to 280% and the latter by up to 207%
[92]. In alignment with our modeling results, a multitude of studies report an increase of slow oscillations
in AD; for a review see [93]. Interestingly, some of these studies include results for the ISO and the SO
(low delta) band, e.g., [94-96]. Similar to AD, astrocytes in epileptic brain tissue also exhibit velocities and
amplitudes above the normal range. For instance, studies in which seizures were induced by the
application of 4-Aminopyridine (4-AP), a known convulsant, found a significant increase in astrocyte
activity [84]. This brain disease, too, is typically accompanied by an increase of the SO power, at least for
temporal lobe epilepsy [97] and the same side of the seizure source [98].
Two other astrocyte-implicated brain diseases are ischemic stroke and PD. In the case of the acute
ischemic stroke, a significant [100-500]% increase in the astrocyte Ca2+ activity has been reported in the
nearby brain tissue with the peak value taken 80 minutes after photothrombosis [99]. The same disease
has been related to a [63-120]% increase of the delta/SO power, at least for the most severe cases [100].
Involving deep brain structures, PD cannot faithfully be linked to cortical oscillations. Nevertheless, initial
epidemiological studies on the relationship of the disease and caffeine consumption, further supported
by experimental studies on adenosine neurotransmission, have revealed astrocytes as a key mechanism
that can affect the onset or progression of PD [101]. In an animal model of PD, astrocytes located in the
Globus Pallidus externus have been found to exhibit increased Ca2+ waves in both frequency and
amplitude [102]. Based on our model, an increase of the Ca2+ wave results to an SO and ISO power increase
(Fig. 6b). Awake EEG studies in patients with PD have shown an increase in slow delta activity across the
cortex, with the oscillation increasing with cognitive impairment [103-105]. Conspicuously, adenosine
receptors on astrocytes have recently been proposed as potential targets for the treatment of PD [106].
To speculate further, since PD is a network disease, as it affects a multitude of closely connected nuclei,
it is not hard to imagine a neuroprotective approach that takes into account astrocytes as a functional
unit integrated into the neural network, and not just as a cell that mediates neuronal survival and
therefore has a limited number of possible interventions [107].
Overall, our modeling results are in agreement with the experimental studies, where slow-wave
oscillations are altered in astrocyte-implicated brain disorders. Whether our model can be used to shed
light on the astrocytes’ role on brain diseases remains to be seen.
6. Conclusion
Slow-wave oscillations, whose precise origin and function are not known, are considered a
fundamental mode of brain activity which is generalizable across cortical areas and species and typically
appears when the brain enters a resting-state, such as sleep or coma. Astrocytes have long been known
to be actively involved in neural circuits necessary for key brain functions, including sleep, memory, and
sensory processing. In this paper, we presented a biologically plausible computational model of a simple
astrocytic-neural network to propose a possible generation mechanism for ISO and SO. The model
13
qualitatively predicts the slow oscillation imbalances seen in astrocyte-implicated brain disorders. This
and other efforts aiming to reveal the computational role of the astrocytes in brain function and
dysfunction have the potential to shape a new research field, which one could call “computational astrocyence.”
7. Appendix
Derivation of Biexponential Normalization Constant
Let 𝑏𝑏(𝑥𝑥) be a biexponential function with time constants a and b and normalization constant 𝑛𝑛:
𝑥𝑥
𝑥𝑥
𝑏𝑏(𝑥𝑥) = 𝑛𝑛 ∗ (𝑒𝑒𝑒𝑒𝑒𝑒 �− � − 𝑒𝑒𝑒𝑒𝑒𝑒 �− �)
𝑎𝑎
𝑏𝑏
We wish to solve for 𝑛𝑛 such that 𝑏𝑏(𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 ) = 1. To determine 𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 , we take the partial derivative of
𝑏𝑏(𝑥𝑥) with respect to x and set this equal to zero:
𝜕𝜕𝜕𝜕
1
𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚
1
𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚
= 𝑛𝑛 ∗ �− � � ∗ exp �−
� + � � ∗ exp �−
�� = 0
𝜕𝜕𝜕𝜕
𝑏𝑏
𝑎𝑎
𝑏𝑏
𝑎𝑎
𝑏𝑏
𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 = (ln � � ∗ 𝑎𝑎 ∗ 𝑏𝑏) ∗ (𝑏𝑏 − 𝑎𝑎)−1
𝑎𝑎
Plugging 𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 back into 𝑏𝑏(𝑥𝑥) gives us the maximum amplitude of the function:
𝑏𝑏(𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 ) = 𝑛𝑛 ∗
𝑎𝑎
𝑏𝑏 𝑎𝑎−𝑏𝑏
(�𝑎𝑎�
𝑏𝑏
−
𝑏𝑏 𝑎𝑎−𝑏𝑏
� � )
𝑎𝑎
𝑏𝑏
−1
Therefore, the value of 𝑛𝑛 which assures that 𝑏𝑏(𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 ) = 1 is given by:
𝑎𝑎
𝑏𝑏 𝑎𝑎−𝑏𝑏
𝑏𝑏 𝑎𝑎−𝑏𝑏
−� � �
𝑛𝑛 = �� �
𝑎𝑎
𝑎𝑎
In the case of [𝑎𝑎 = 50, 𝑏𝑏 = 300] we get 𝑛𝑛 =
8. Acknowledgments
1
6∗ 65
.
5
We would like to thank Kevin T. Feigelis for his insightful suggestions and modeling work during
the early stages of this project.
9. References
1.
2.
3.
Sejnowski TJP, O. Network Oscillations: Emerging Computational Principles. Journal of
Neuroscience Journal of Neuroscience 2006;26:1673–6
Steriade M. Grouping of brain rhythms in corticothalamic systems. Neuroscience.
2006;137(4):1087-106.
Hiltunen T, Kantola J, Elseoud AA, Lepola P, Suominen K, Starck T, et al. Infra-slow EEG fluctuations
are correlated with resting-state network dynamics in fMRI. The Journal of Neuroscience.
2014;34(2):356-62.
14
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
Buckner RL, Andrews‐Hanna JR, Schacter DL. The brain's default network. Annals of the New York
Academy of Sciences. 2008;1124(1):1-38.
Buzsáki G, Chrobak JJ. Temporal structure in spatially organized neuronal ensembles: a role for
interneuronal networks. Current opinion in neurobiology. 1995;5(4):504-10.
Nir Y, Staba RJ, Andrillon T, Vyazovskiy VV, Cirelli C, Fried I, et al. Regional slow waves and spindles
in human sleep. Neuron. 2011;70(1):153-69.
Poskanzer KE, Yuste R. Astrocytic regulation of cortical UP states. Proceedings of the National
Academy of Sciences. 2011;108(45):18453-8.
McGinty D, Szymusiak R. Keeping cool: a hypothesis about the mechanisms and functions of slowwave sleep. Trends in neurosciences. 1990;13(12):480-7.
Jones S. Slow-wave sleep and the risk of type 2 diabetes in humans Tasali E, Leproult R, Ehrmann
DA, et al (Univ of Chicago, IL) Proc Natl Acad Sci USA 105: 1044-1049, 2008. Year Book of
Pulmonary Disease. 2009;2009:277-8.
Buzsaki G. Rhythms of the Brain: Oxford University Press; 2006.
Helkala E-L, Laulumaa V, Soikkeli R, Partanen J, Soininen H, Riekkinen PJ. Slow-wave activity in the
spectral analysis of the electroencephalogram is associated with cortical dysfunctions in patients
with Alzheimer's disease. Behavioral neuroscience. 1991;105(3):409.
Göder R, Boigs M, Braun S, Friege L, Fritzer G, Aldenhoff JB, et al. Impairment of visuospatial
memory is associated with decreased slow wave sleep in schizophrenia. Journal of psychiatric
research. 2004;38(6):591-9.
Herculano-Houzel S. The human brain in numbers: a linearly scaled-up primate brain. Frontiers in
human neuroscience. 2009;3:31.
Verkhratsky A, Butt A. Neuroglia: definition, classification, evolution, numbers, development. glial
physiology and pathophysiology. 2013:73-104.
Lőrincz ML, Geall F, Bao Y, Crunelli V, Hughes SW. ATP-dependent infra-slow (< 0.1 Hz) oscillations
in thalamic networks. PLoS One. 2009;4(2):e4447.
Goldberg M, De Pittà M, Volman V, Berry H, Ben-Jacob E. Nonlinear gap junctions enable longdistance propagation of pulsating calcium waves in astrocyte networks. PLoS Comput Biol.
2010;6(8):e1000909.
Khakh BS, McCarthy KD. Astrocyte calcium signaling: from observations to functions and the
challenges therein. Cold Spring Harb Perspect Biol. 2015;7:a020404.
Kurth-Nelson ZL, Mishra A, Newman EA. Spontaneous glial calcium waves in the retina develop
over early adulthood. The Journal of Neuroscience. 2009;29(36):11339-46.
Parri HR, Gould TM, Crunelli V. Spontaneous astrocytic Ca2+ oscillations in situ drive NMDARmediated neuronal excitation. Nature neuroscience. 2001;4(8):803-12.
Hirase H, Qian L, Barthó P, Buzsáki G. Calcium dynamics of cortical astrocytic networks in vivo. PLoS
Biol. 2004;2(4):e96.
Bazargani N, Attwell D. Astrocyte calcium signaling: the third wave. Nature neuroscience.
2016;19(2):182-9.
Volman V, Ben-Jacob E, Levine H. The astrocyte as a gatekeeper of synaptic information transfer.
Neural computation. 2007;19(2):303-26.
Araque A, Parpura V, Sanzgiri RP, Haydon PG. Tripartite synapses: glia, the unacknowledged
partner. Trends in neurosciences. 1999;22(5):208-15.
Florian C, Vecsey CG, Halassa MM, Haydon PG, Abel T. Astrocyte-derived adenosine and A1
receptor activity contribute to sleep loss-induced deficits in hippocampal synaptic plasticity and
memory in mice. The Journal of neuroscience. 2011;31(19):6956-62.
15
25. Lee HS, Ghetti A, Pinto-Duarte A, Wang X, Dziewczapolski G, Galimi F, et al. Astrocytes contribute to
gamma oscillations and recognition memory. Proceedings of the National Academy of Sciences.
2014;111(32):E3343-E52.
26. Perea G, Navarrete M, Araque A. Tripartite synapses: astrocytes process and control synaptic
information. Trends in neurosciences. 2009;32(8):421-31.
27. Araque A, Parpura V, Sanzgiri RP, Haydon PG. Glutamate‐dependent astrocyte modulation of
synaptic transmission between cultured hippocampal neurons. European Journal of Neuroscience.
1998;10(6):2129-42.
28. Clarke LE, Barres BA. Emerging roles of astrocytes in neural circuit development. Nature Reviews
Neuroscience. 2013;14(5):311-21.
29. De Pitta M, Volman V, Berry H, Parpura V, Volterra A, Ben-Jacob E. Computational quest for
understanding the role of astrocyte signaling in synaptic transmission and plasticity. Frontiers in
computational neuroscience. 2012;6.
30. Sul J-Y, Orosz G, Givens RS, Haydon PG. Astrocytic connectivity in the hippocampus. Neuron glia
biology. 2004;1(01):3-11.
31. Halassa MM, Fellin T, Takano H, Dong J-H, Haydon PG. Synaptic islands defined by the territory of a
single astrocyte. The Journal of neuroscience. 2007;27(24):6473-7.
32. Cornell-Bell A, Finkbeiner S. Ca2+ waves in astrocytes. Cell calcium. 1991;12(2-3):185-204.
33. Hoogland TM, Kuhn B, Göbel W, Huang W, Nakai J, Helmchen F, et al. Radially expanding transglial
calcium waves in the intact cerebellum. Proceedings of the National Academy of Sciences.
2009;106(9):3496-501.
34. Newman EA. Propagation of intercellular calcium waves in retinal astrocytes and Müller cells. The
Journal of Neuroscience. 2001;21(7):2215-23.
35. Newman EA, Zahs KR. Calcium waves in retinal glial cells. Science. 1997;275(5301):844-7.
36. Scemes E, Giaume C. Astrocyte calcium waves: what they are and what they do. Glia.
2006;54(7):716-25.
37. Nimmerjahn A, Bergles DE. Large-scale recording of astrocyte activity. Current opinion in
neurobiology. 2015;32:95-106.
38. Höfer T, Venance L, Giaume C. Control and plasticity of intercellular calcium waves in astrocytes: a
modeling approach. The Journal of neuroscience. 2002;22(12):4850-9.
39. Kuga N, Sasaki T, Takahara Y, Matsuki N, Ikegaya Y. Large-scale calcium waves traveling through
astrocytic networks in vivo. The Journal of Neuroscience. 2011;31(7):2607-14.
40. Nett WJ, Oloff SH, McCarthy KD. Hippocampal astrocytes in situ exhibit calcium oscillations that
occur independent of neuronal activity. Journal of neurophysiology. 2002;87(1):528-37.
41. Verderio C, Bacci A, Coco S, Pravettoni E, Fumagalli G, Matteoli M. Astrocytes are required for the
oscillatory activity in cultured hippocampal neurons. European Journal of Neuroscience.
1999;11(8):2793-800.
42. Fellin T, Halassa MM, Terunuma M, Succol F, Takano H, Frank M, et al. Endogenous nonneuronal
modulators of synaptic transmission control cortical slow oscillations in vivo. Proceedings of the
National Academy of Sciences. 2009;106(35):15037-42.
43. Poskanzer KE, Yuste R. Astrocytes regulate cortical state switching in vivo. Proceedings of the
National Academy of Sciences. 2016;113(19):E2675-E84.
44. Nadjar A, Blutstein T, Aubert A, Laye S, Haydon PG. Astrocyte‐derived adenosine modulates
increased sleep pressure during inflammatory response. Glia. 2013;61(5):724-31.
45. Nadkarni S, Jung P. Dressed neurons: modeling neural–glial interactions. Physical biology.
2004;1(1):35.
46. Fourcaud N, Brunel N. Dynamics of the firing probability of noisy integrate-and-fire neurons. Neural
computation. 2002;14(9):2057-110.
16
47. Bennett MR, Farnell L, Gibson WG, Lagopoulos J. Cortical Network Models of Firing Rates in the
Resting and Active States Predict BOLD Responses. PloS one. 2015;10(12):e0144796.
48. Lapicque L. Recherches quantitatives sur l’excitation électrique des nerfs traitée comme une
polarisation. J Physiol Pathol Gen. 1907;9(1):620-35.
49. Tal D, Schwartz EL. Computing with the leaky integrate-and-fire neuron: logarithmic computation
and multiplication. Neural Computation. 1997;9(2):305-18.
50. Tsodyks M, Pawelzik K, Markram H. Neural networks with dynamic synapses. Neural computation.
1998;10(4):821-35.
51. Winship IR, Plaa N, Murphy TH. Rapid astrocyte calcium signals correlate with neuronal activity and
onset of the hemodynamic response in vivo. The Journal of neuroscience. 2007;27(23):6268-72.
52. Wade JJ, McDaid LJ, Harkin J, Crunelli V, Kelso JS. Bidirectional coupling between astrocytes and
neurons mediates learning and dynamic coordination in the brain: a multiple modeling approach.
PloS one. 2011;6(12):e29445.
53. Parpura V, Haydon PG. Physiological astrocytic calcium levels stimulate glutamate release to
modulate adjacent neurons. Proceedings of the National Academy of Sciences. 2000;97(15):862934.
54. Fellin T, Pascual O, Gobbo S, Pozzan T, Haydon PG, Carmignoto G. Neuronal synchrony mediated by
astrocytic glutamate through activation of extrasynaptic NMDA receptors. Neuron. 2004;43(5):72943.
55. Haydon PG, Carmignoto G. Astrocyte control of synaptic transmission and neurovascular coupling.
Physiological reviews. 2006;86(3):1009-31.
56. Pasti L, Zonta M, Pozzan T, Vicini S, Carmignoto G. Cytosolic calcium oscillations in astrocytes may
regulate exocytotic release of glutamate. The Journal of Neuroscience. 2001;21(2):477-84.
57. Eddleston M, Mucke L. Molecular profile of reactive astrocytes—implications for their role in
neurologic disease. Neuroscience. 1993;54(1):15-36.
58. Barres BA. The mystery and magic of glia: a perspective on their roles in health and disease.
Neuron. 2008;60(3):430-40.
59. Fiacco TA, McCarthy KD. Astrocyte calcium elevations: properties, propagation, and effects on brain
signaling. Glia. 2006;54(7):676-90.
60. Schnitzler A, Gross J. Normal and pathological oscillatory communication in the brain. Nature
reviews neuroscience. 2005;6(4):285-96.
61. Steriade M, Nunez A, Amzica F. A novel slow (< 1 Hz) oscillation of neocortical neurons in vivo:
depolarizing and hyperpolarizing components. Journal of neuroscience. 1993;13(8):3252-65.
62. Csercsa R, Dombovári B, Fabó D, Wittner L, Erőss L, Entz L, et al. Laminar analysis of slow wave
activity in humans. Brain. 2010:awq169.
63. Murphy K, Birn RM, Handwerker DA, Jones TB, Bandettini PA. The impact of global signal regression
on resting state correlations: are anti-correlated networks introduced? Neuroimage.
2009;44(3):893-905.
64. Mohajerani MH, McVea DA, Fingas M, Murphy TH. Mirrored bilateral slow-wave cortical activity
within local circuits revealed by fast bihemispheric voltage-sensitive dye imaging in anesthetized
and awake mice. Journal of Neuroscience. 2010;30(10):3745-51.
65. Ferezou I, Haiss F, Gentet LJ, Aronoff R, Weber B, Petersen CC. Spatiotemporal dynamics of cortical
sensorimotor integration in behaving mice. Neuron. 2007;56(5):907-23.
66. Desmedt JE, La Grutta G. The effect of selective inhibition of pseudocholinesterase on the
spontaneous and evoked activity of the cat's cerebral cortex. The Journal of physiology.
1957;136(1):20.
67. Galambos R. A glia-neural theory of brain function. Proceedings of the National Academy of
Sciences. 1961;47(1):129-36.
17
68. Marshall L, Helgadóttir H, Mölle M, Born J. Boosting slow oscillations during sleep potentiates
memory. Nature. 2006;444(7119):610-3.
69. Monai H, Ohkura M, Tanaka M, Oe Y, Konno A, Hirai H, et al. Calcium imaging reveals glial
involvement in transcranial direct current stimulation-induced plasticity in mouse brain. Nature
communications. 2016;7.
70. Han X, Chen M, Wang F, Windrem M, Wang S, Shanz S, et al. Forebrain engraftment by human glial
progenitor cells enhances synaptic plasticity and learning in adult mice. Cell stem cell.
2013;12(3):342-53.
71. Zucker SW. Local field potentials and border ownership: a conjecture about computation in visual
cortex. Journal of Physiology-Paris. 2012;106(5):297-315.
72. Figley CR, Stroman PW. The role (s) of astrocytes and astrocyte activity in neurometabolism,
neurovascular coupling, and the production of functional neuroimaging signals. European Journal
of Neuroscience. 2011;33(4):577-88.
73. Palva JM, Palva S. Infra-slow fluctuations in electrophysiological recordings, blood-oxygenationlevel-dependent signals, and psychophysical time series. Neuroimage. 2012;62(4):2201-11.
74. Schulz K, Sydekum E, Krueppel R, Engelbrecht CJ, Schlegel F, Schröter A, et al. Simultaneous BOLD
fMRI and fiber-optic calcium recording in rat neocortex. Nature methods. 2012;9(6):597-602.
75. Tritsch NX, Yi E, Gale JE, Glowatzki E, Bergles DE. The origin of spontaneous activity in the
developing auditory system. Nature. 2007;450(7166):50-5.
76. Dayaratne M, Vlajkovic SM, Lipski J, Thorne PR. Kölliker’s organ and the development of
spontaneous activity in the auditory system: Implications for hearing dysfunction. BioMed research
international. 2014;2014.
77. Cao X, Li L-P, Wang Q, Wu Q, Hu H-H, Zhang M, et al. Astrocyte-derived ATP modulates depressivelike behaviors. Nature medicine. 2013;19(6):773-7.
78. Meerwijk EL, Ford JM, Weiss SJ. Resting-state EEG delta power is associated with psychological pain
in adults with a history of depression. Biological psychology. 2015;105:106-14.
79. Brenner RP, Ulrich RF, Spiker DG, Sclabassi RJ, Reynolds CF, Marin RS, et al. Computerized EEG
spectral analysis in elderly normal, demented and depressed subjects. Electroencephalography and
clinical neurophysiology. 1986;64(6):483-92.
80. Massimini M, Ferrarelli F, Esser SK, Riedner BA, Huber R, Murphy M, et al. Triggering sleep slow
waves by transcranial magnetic stimulation. Proceedings of the National Academy of Sciences.
2007;104(20):8496-501.
81. Seifert G, Carmignoto G, Steinhäuser C. Astrocyte dysfunction in epilepsy. Brain research reviews.
2010;63(1):212-21.
82. de Lanerolle NC, Lee T-S, Spencer DD. Astrocytes and epilepsy. Neurotherapeutics. 2010;7(4):42438.
83. Hinterkeuser S, SchroÈder W, Hager G, Seifert G, BluÈmcke I, Elger CE, et al. Astrocytes in the
hippocampus of patients with temporal lobe epilepsy display changes in potassium conductances.
European Journal of Neuroscience. 2000;12(6):2087-96.
84. Tian G-F, Azmi H, Takano T, Xu Q, Peng W, Lin J, et al. An astrocytic basis of epilepsy. Nature
medicine. 2005;11(9):973-81.
85. Verkhratsky A, Olabarria M, Noristani HN, Yeh C-Y, Rodriguez JJ. Astrocytes in Alzheimer's disease.
Neurotherapeutics. 2010;7(4):399-412.
86. Scheuner D, Eckman C, Jensen M, Song X, Citron M, Suzuki N, et al. Secreted amyloid beta-protein
similar to that in the senile plaques of Alzheimer's disease is increased in vivo by the presenilin 1
and 2 and APP mutations linked to familial Alzheimer's disease. Nature medicine. 1996;2(8):864-70.
87. Duff K, Eckman C, Zehr C, Yu X, Prada C-M, Perez-Tur J, et al. Increased amyloid-β42 (43) in brains of
mice expressing mutant presenilin 1. Nature. 1996;383(6602):710-3.
18
88. Schmechel D, Saunders A, Strittmatter W, Crain BJ, Hulette C, Joo S, et al. Increased amyloid betapeptide deposition in cerebral cortex as a consequence of apolipoprotein E genotype in late-onset
Alzheimer disease. Proceedings of the National Academy of Sciences. 1993;90(20):9649-53.
89. Zhao J, O'Connor T, Vassar R. The contribution of activated astrocytes to Aβ production:
implications for Alzheimer's disease pathogenesis. Journal of neuroinflammation. 2011;8(1):150.
90. Pihlaja R, Koistinaho J, Kauppinen R, Sandholm J, Tanila H, Koistinaho M. Multiple cellular and
molecular mechanisms are involved in human Aβ clearance by transplanted adult astrocytes. Glia.
2011;59(11):1643-57.
91. Scemes E, Dermietzel R, Spray DC. Calcium waves between astrocytes from Cx43 knockout mice.
Glia. 1998;24(1):65.
92. Haughey NJ, Mattson MP. Alzheimer’s amyloid β-peptide enhances ATP/gap junction-mediated
calcium-wave propagation in astrocytes. Neuromolecular medicine. 2003;3(3):173-80.
93. Jeong J. EEG dynamics in patients with Alzheimer's disease. Clinical neurophysiology.
2004;115(7):1490-505.
94. Hier DB, Mangone CA, Ganellen R, Warach JD, Van Egeren R, Perlik SJ, et al. Quantitative
measurement of delta activity in Alzheimer's disease. Clinical EEG and Neuroscience.
1991;22(3):178-82.
95. Fonseca LC, Tedrus GMAS, Prandi LR, Andrade A. Quantitative electroencephalography power and
coherence measurements in the diagnosis of mild and moderate Alzheimer's disease. Arquivos de
neuro-psiquiatria. 2011;69(2B):297-303.
96. Neto E, Allen EA, Aurlien H, Nordby H, Eichele T. EEG spectral features discriminate between
Alzheimer’s and vascular dementia. Frontiers in neurology. 2015;6.
97. Bernasconi A, Cendes F, Lee J, Reutens DC, Gotman J. EEG background delta activity in temporal
lobe epilepsy: correlation with volumetric and spectroscopic imaging. Epilepsia. 1999;40(11):15806.
98. Englot DJ, Yang L, Hamid H, Danielson N, Bai X, Marfeo A, et al. Impaired consciousness in temporal
lobe seizures: role of cortical slow activity. Brain. 2010;133(12):3764-77.
99. Ding S. Ca2+ signaling in astrocytes and its role in ischemic stroke. Glutamate and ATP at the
Interface of Metabolism and Signaling in the Brain: Springer; 2014. p. 189-211.
100. Finnigan SP, Rose SE, Walsh M, Griffin M, Janke AL, McMahon KL, et al. Correlation of quantitative
EEG in acute ischemic stroke with 30-day NIHSS score comparison with diffusion and perfusion
MRI. Stroke. 2004;35(4):899-903.
101. Morelli M, Carta AR, Kachroo A, Schwarzschild MA. Pathophysiological roles for purines: adenosine,
caffeine and urate. Progress in brain research. 2010;183:183-208.
102. Cui Q, Pitt JE, Pamukcu A, Poulin J-F, Mabrouk OS, Fiske MP, et al. Blunted mGluR Activation
Disinhibits Striatopallidal Transmission in Parkinsonian Mice. Cell Reports. 2016;17(9):2431-44.
103. Soikkeli R, Partanen J, Soininen H, Pääkkönen A, Riekkinen P. Slowing of EEG in Parkinson's disease.
Electroencephalography and clinical neurophysiology. 1991;79(3):159-65.
104. Neufeld M, Blumen S, Aitkin I, Parmet Y, Korczyn A. EEG frequency analysis in demented and
nondemented parkinsonian patients. Dementia and Geriatric Cognitive Disorders. 1994;5(1):23-8.
105. Cozac VV, Gschwandtner U, Hatz F, Hardmeier M, Rüegg S, Fuhr P. Quantitative EEG and Cognitive
Decline in Parkinson’s Disease. Parkinson’s Disease. 2016;2016.
106. Miyazaki I, Asanuma M. Serotonin 1A Receptors on Astrocytes as a Potential Target for the
Treatment of Parkinson’s Disease. Current medicinal chemistry. 2016;23(7):686-700.
107. Rappold PM, Tieu K. Astrocytes and therapeutics for Parkinson's disease. Neurotherapeutics.
2010;7(4):413-23.
19
| 9cs.NE
|
arXiv:1705.06693v1 [cs.NE] 18 May 2017
Limited-Memory Matrix Adaptation
for Large Scale Black-box Optimization
Ilya Loshchilov
Research Group on Machine Learning
for Automated Algorithm Design
University of Freiburg, Germany
[email protected]
Tobias Glasmachers
Institut für Neuroinformatik
Ruhr-Universität Bochum, Germany
[email protected]
Hans-Georg Beyer
Research Center Process and Product Engineering
Vorarlberg University of Applied Sciences, Dornbirn, Austria
[email protected]
Abstract
The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is a popular
method to deal with nonconvex and/or stochastic optimization problems when the
gradient information is not available. Being based on the CMA-ES, the recently
proposed Matrix Adaptation Evolution Strategy (MA-ES) provides a rather surprising result that the covariance matrix and all associated operations (e.g., potentially
unstable eigendecomposition) can be replaced in the CMA-ES by a updated transformation matrix without any loss of performance. In order to further simplify
MA-ES and reduce its O n2 time and storage complexity to O n log(n) , we
present the Limited-Memory Matrix Adaptation Evolution Strategy (LM-MA-ES)
for efficient zeroth order large-scale optimization. The algorithm demonstrates
state-of-the-art performance on a set of established large-scale benchmarks. We
explore the algorithm on the problem of generating adversarial inputs for a (nonsmooth) random forest classifier, demonstrating a surprising vulnerability of the
classifier.
1
Introduction
Evolution Strategies (ESs) are optimization methods originally inspired by mutation of organic beings
and designed to establish “a reward-based system, to increase the probability of those changes, which
lead to improvements of quality of the system” [31]. Going far beyond their biologically inspired
roots, they have been developed into state-of-the-art zeroth order search methods [12]. Evolution
Strategies [32] consider an objective function f : IRn 7→ IR, x 7→ f (x) to be minimized by sampling
i ∈ {1, . . . , λ} candidate solutions at iteration t as
(t)
xi ← y(t) + σ (t) · N 0, C(t) ,
(1)
where y(t) is the current estimate of the optimum, C(t) ∈ IRn×n is a covariance matrix initialized
to the identity matrix I, and σ (t) is a scaling factor for the mutation step, often referred to as the
global step size. Both y(t) and σ (t) are to be adapted or learned over time in Evolution Strategies.
Modern ESs such as the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) also include
the adaptation of C(t) [15, 14] to the shape of the local landscape, resembling second order methods.
Recent theoretical studies of ES and CMA-ES from the prospective of information geometry [39,
4, 29, 5], connecting the method to natural gradient learning, have made significant progress in
understanding the principles underpinning the state-of-the-art performance of the algorithm [13].
The variety of algorithms [12] derived from and inspired by the theoretical studies helped to notice
that the core component of CMA-ES, the covariance matrix itself (and covariance matrix square
root operations) can be removed from the algorithm without any loss of performance [8].1 The final
algorithm called Matrix Adaptation Evolution Strategy (MA-ES [8]) is conceptually simpler and
involves only matrix-matrix and matrix-vector operations,
which, however, lead to O n3 / log(n)
2
time complexity per sample (here, we show its O n implementation), and O(n2 ) space complexity.
In machine learning, ESs are mainly used for direct policy search in reinforcement learning [11, 18,
36, 34], hyperparameter tuning in supervised learning, e.g., for Support Vector Machines [10, 19] and
Deep Neural Networks [27]. With the steadily increasing dimensionality of real-world optimization
problems, the new challenges of large-scale black-box optimization become more pronounced for
CMA-ES and MA-ES due to their O(n2 ) complexity. To address them, a number of large-scale
CMA-ES variants has been proposed [23, 33, 37, 1, 26, 3] including the Limited-Memory CMA-ES
(LM-CMA-ES [26]) that matches the performance of quasi-Newton methods such as L-BFGS [35]
when dealing with large-scale black-box problems at a moderate cost of O(n log(n)) time and space
complexity.
In this work, we combine the best of two worlds: inspired by LM-CMA-ES we present the LimitedMemory Matrix Adaptation Evolution Strategy (LM-MA-ES), which matches state-of-the-art results
while reducing the time and space complexity of MA-ES to O(n log(n)) per sample.
2
Variable Metric Evolution Strategies: CMA-ES and MA-ES
Our discussion on variable metric ESs is based on Algorithm 1, which highlights the similarities and
differences between CMA-ES, MA-ES, and the proposed LM-MA-ES.
The sampling of the λ candidate solutions
in CMA-ES is described by eq. (1) and involves a matrix√
vector product between a matrix C and a vector zi sampled from the
√ n-dimensional standard
normal distribution. This operation (see line 7 in Algorithm 1) requires C to be stored (hence, the
quadratic in n space cost) and the matrix-vector multiplication to be performed (hence, the quadratic
in n time cost). The resulting vector di represents a direction of the so-called mutation operation.
The i-th candidate solution is obtained by changing (mutating) the current estimate of the optimum y
by di multiplied by the global mutationstep-size σ (line 11). The rationale behind parameterizing
the sampling distribution by N y, σ 2 C and not just N (y, C), i.e., decoupling C and σ, lies in the
observation that σ can be learned more quickly and more robustly than C and its adaptation alone
enables linear convergence on scale-invariant problems [20].
ESs are invariant to rank-preserving / strictly monotonic transformations of f -values because all
operations are based on ranks of evaluated solutions. The estimate of the optimum y is updated by
a weighted sum of mutation steps taken by the top ranked µ out λ solutions (line 12). The recent
analysis [2] demonstrated that the optimal recombination weights w for Sphere function (see Table 1)
are independent of the Hessian matrix, and hence optimal for all convex quadratic functions.
The currently most commonly applied adaptation rule for the step size is the cumulative step-size
adaptation (CSA) mechanism [16]. It is based on the length of an evolution path pσ , an exponentially
fading record of recent most successful steps zi:λ (see line 13). If the path becomes
too long (the
√
expected path length of a Gaussian random walk can be approximated by n when n is large),
indicating that recent steps tend to move into the same direction, then the step size is increased. On
the contrary, a too short path indicating oscillations due to overjumping the optimum results in a
reduction of the step size. Rigorous analysis of CSA with and without cumulation is given in [7].
The seminal CMA-ES algorithm [15, 14] introduced adaptation of the covariance matrix, which
renders the algorithm invariant to linear transformations of the search space (achieved in practice after
an initial adaptation phase) and hence enables a fast convergence rate independent of the conditioning
of the problem, resembling second order methods. The covariance matrix is adapted towards a
weighted maximum likelihood estimate of the µ most successful samples (rank-µ update) with
learning rate cµ and a second evolution path (rank-1 update) with learning rate c1 (see line 15); this
1
Of course, the transformation matrix needed for sampling the multivariate Gaussian is kept, enabling
variable metric optimization.
2
update has an alternative interpretation as a stochastic gradient step on the information geometric
manifold forming the algorithm’s state space [4, 29]. While the default hyperparameter values of
CMA-ES given in Algorithm 1 are known to be robust, their optimal values can be adapted online
during the optimization process attempting to provide an additional level of invariance [28].
Most implementations√of CMA-ES consider eigendecomposition procedures of O(n3 ) time complexity per call to obtain C from C only every n/λ iterations (see line 7) to achieve amortized O(n2 )
time complexity per sampled solution. Numerical stability of the O(n2 ) update can be ensured by
maintaining a triangular Cholesky factor [24].
The recently proposed Matrix Adaptation Evolution Strategy (MA-ES) greatly simplifies CMA-ES
by avoiding the construction
√ of the covariance matrix C. Instead it maintains only a transformation
matrix M representing C. After removing the approximate redundancy of pσ and pc , M can be
updated multiplicatively (line 16). Matrix multiplication is an O(n3 ) operation, therefore we propose
to replace the multiplicative update at iteration t by the equivalent additive update
µ
cµ (t) c1 (t) (t) T cµ X
c1
(t) (t)
M + dσ (pσ ) +
−
wi di:λ (zi:λ )T ,
M(t+1) ← 1 −
2
2
2
2 i=1
(t)
(2)
(t)
(t)
which achieves O(n2 ) time cost thanks to precomputing d(t)
σ = M pσ and reusing di:λ vectors.
The resulting algorithm is referred to as fast MA-ES.
3
The Limited-Memory Matrix Adaptation Evolution Strategy
A number of methods was proposed to
reduce the space and time complexity per sample from
O(n2 ) to O(n) or at least O n log(n) while still modeling the most relevant aspects of the full
covariance matrix. Simple approaches like [33] restrict the covariance matrix to its diagonal, while
more elaborate methods use a low-rank approach [37, 26]. Both approaches can be combined [1].
Inspired by the Limited-Memory CMA-ES [26] which in turn is inspired by the L-BFGS method
[25], we show how to scale up MA-ES to high-dimensional problems. The derivation given below is
based on the multiplicative update, the final result for the additive update (2) is equivalent when d(t)
σ
(t)
is not stored but reconstructed as M(t) pσ . At iteration t, the main update equation of MA-ES reads
!#
"
µ
c
X
c1 (t+1) (t+1) T
µ
(t) (t) T
(t+1)
(t)
pσ
(pσ
) −I +
wi zi:λ (zi:λ ) − I
,
(3)
M
←M
I+
2
2 i=1
where M(t) is adapted multiplicatively based on the rank-one update weighted by c21 and the rank-µ
c
update weighted by 2µ , starting from M(t=0) = I. By omitting the rank-µ update for the sake of
simplicity (i.e., by setting cµ = 0), we obtain
c1
c1
c1 (1) (1) T
pσ (pσ ) − I = 1 −
I + p(1)
(p(1) )T
(4)
M(1) ← I +
2
2
2 σ σ
(1)
The sampling procedure of the i-th solution xi
(1)
xi
(1)
follows
(1)
← y(1) + σ (1) di
(1)
= y(1) + σ (1) M(1) zi ,
(1)
(5)
(1)
∼ N (0, I ). One can rewrite di = M(1) zi based on equation (4) as
c1
c1
c1 c1 (1) (1) T (1)
(1)
(1)
(1)
(1)
(1) T
di = M(1) zi = 1 −
I + p(1)
(p
)
z
=
z
1
−
+ pσ (pσ ) zi
i
i
2
2 σ σ
2
2
(6)
(1) T (1)
Importantly, (pσ ) z
is a scalar (see line 10 in Algorithm 1) and thus equation (6) does not
where zi
require M(1) to be stored in memory. One generally obtains
c1
c1
(t)
(t)
(t)
(t)
(t) T
di = M(t) zi = M(t−1) P(t) zi = M(t−1) 1 −
I + p(t)
(p
)
zi
σ
σ
2
2
|
{z
}
:=P(t)
3
(7)
Algorithm 1 CMA-ES , MA-ES and LM-MA-ES
1: given n ∈ N+ , λ = 4 + b3 ln nc, µ = bλ/2c, wi =
Pµ 1
i=1
for i = 1, . . . , µ, µw =
2(µw −2+1/µw )
µw +2
4
2
1
−
c
,
,
, cσ = n+µ
,
c
=
,
c
=
2 +µ , cµ = min
2 +µ
1
c
1
+5
n+4
(n+1.3)
(n+2)
w2
w
w
w
i
2λ
n , cd,i
=
1
1.5i−1 n , cc,i
(t=0)
2: initialize t ← 0, y(t=0) ∈ IRn , σ (t=0) > 0, pσ
(t=0)
6:
7:
(t)
for j ← 1, . . . , min(t, m) do
15:
16:
17:
(t=0)
= 0, pc
= 0, C(t=0) = I , M(t=0) = I ,
(t)
(t)
. CMA-ES
. CMA-ES and MA-ES
(t)
. LM-MA-ES
(t)
di ← (1 − cd,j )di + cd,j mj
10:
14:
for i = 1, . . . , m
(t)
9:
13:
λ
4i−1 n
(t=0)
di ← M(t) di
12:
=
mi
∈ IRn , mi
= 0 for i = 1, . . . , m
repeat
for i ← 1, . . . , λ do
(t)
zi ← N (0, I )
(t)
(t)
di ← zi
√
if t mod nλ = 0 then M (t) ← C(t) else M (t) ← M (t−1)
8:
11:
j)
m = 4 + b3 ln nc, cσ =
3:
4:
5:
ln(µ+ 21 )−ln i
1
j=1 (ln(µ+ 2 )−ln
Pµ
(t)
(t)
(mj )T di
. LM-MA-ES
(t)
fi ← f (y(t) + σ (t) di )
Pµ
(t)
y(t+1) ← y(t) + σ (t) i=1 wi di:λ
. the symbol i : λ denotes i-th best sample on f
p
Pµ
(t)
(t+1)
(t)
← (1 − cσ )pσ + µw cσ (2 − cσ ) i=1 wi zi:λ
pσ
p
Pµ
(t)
(t+1)
(t)
← (1 − cc )pc + µw cc (2 − cc ) i=1 wi di:λ
pc
. CMA-ES
P
(t)
(t)
(t)
µ
C(t+1) ← (1 − c1 − cµ )C(t) + c1 pc (pc )T + cµ i=1 wi di:λ (di:λ )T
. CMA-ES
h
P
i
(t) (t)
(t) (t) T
c
µ
M(t+1) ← M(t) I + c21 pσ (pσ )T − I + 2µ
. MA-ES
i=1 wi zi:λ (zi:λ ) − I
for i ← 1, . . . , m do
(t+1)
18:
mi
19:
σ (t+1)
. LM-MA-ES
(t)
p
← (1 − cc,i )mi + µw cc,i (2 − cc,i )
(t+1) 2
kpσ k
cσ
(t)
−1
← σ · exp 2
n
Pµ
j=1
(t)
wj zj:λ
. LM-MA-ES
20:
t←t+1
21: until stopping criterion is met
leading to a sequence of products
c1
c1
(t)
(1) T
di =
1−
I + p(1)
· ...
σ (pσ )
2 c 2 c
c1
c1
1
1
(t)
(t) T
... · 1 −
I + pσ(t−1) (pσ(t−1) )T · 1 −
I + p(t)
zi
σ (pσ )
2
2
2
2
(t)
(8)
(t)
which is to be treated from right to left. Thus, the sampling procedure for di = M(t) zi does
neither require matrix-matrix-product operations nor does it require the storage of M(t) ∈ IRn×n ,
(t)
but can be performed based on t vectors pσ used to construct M(t) . However, this is efficient only
for t n. Therefore, in order to reduce the cost of the sampling procedure, equation (8) must be
(t)
approximated in one way or another by artificially limiting the number m of supporting pσ vectors
such that m n. In this work we pick m ∈ O(log(n)).
LM-CMA-ES[26]
addresses a similar problem of compactly representing the covariance matrix with
m ∈ O log(n) direction vectors: instead of considering the last m vectors, it samples them in a
certain temporal distance in terms of iterations t. The same approach works for LM-MA-ES (see
4
section 1 and Algorithm 1 in the supplementary material). However, the rather complicated procedure
of ensuring a temporal distance between pσ vectors can be simplified by considering different time
horizons of their
Pµupdate. This procedure is a viable alternative since pσ itself is anyway incrementally
updated with i=1 wi zi:λ. Thus, instead of a full transformation matrix M ∈ IRn×n , LM-MA-ES
maintains m ∈ O log(n) vectors mi (see lines 17-18 in Algorithm 1), modeling the deviation of
the transformation matrix from the identity as a rank-m matrix. The learning rates cc,i and cd,i for
applying and updating the vectors mi are chosen to be exponentially decaying, hence the mi are
fading records of mean update steps on exponentially differing time scales. This is in contrast to
CMA-ES and MA-ES, which update their matrices C and M only with two different learning rates
for the rank-1 and rank-µ updates, and hence operate on a single time scale. LM-MA-ES learns some
directions very quickly, while others are kept more stable. This can be advantageous in particular in
high dimensions where learning rates are generally small due to the sub-linear sample size.
The proposed LM-MA-ES method features all invariance properties of modern ESs, namely invariance to translation and rotation, as well as invariance to strictly monotonic (rank-preserving)
transformations of objective values.
4
Experimental validation
With our experimental evaluation we aim to answer the following questions:
• How does LM-MA-ES compare to MA-ES, i.e., what is the effect of modeling only a
O(log(n)) dimensional subspace?
• How does LM-MA-ES compare to other algorithms designed for high-dimensional blackbox optimization?
• Is LM-MA-ES suitable for solving problems in machine learning?
To answer these questions, we investigate the performance on standard benchmark problems of
varying dimension, and we generate adversarial inputs for a random forest classifier.
Ellipsoid
Rosenbrock
Discus
Cigar
Different Powers
Function f (x)
Pn
2
i=1 xi
Pn
6 i−1 2
n−1 x
i
i=1 10
Pn−1
100
· (x2i − xi+1 )2 + (xi − 1)2
i=1
Pn
6 2
2
10 x1 + i=2 xi
P
2
x21 + 106 n
i=2 xi
Pn
2+4(i−1)/(n−1)
i=1 |xi |
Table 1: Test functions used in this study.
4.1
10−2
seconds per sample
Name
Sphere
10−3
10−4
10−5
LM−MA−ES
fast MA−ES
−6
10
102
103
104
number of variables
Figure 1: Internal algorithm cost.
Performance on Benchmark Problems
We experimentally validate the proposed LM-MA-ES algorithm on large-scale variants (n ∈
{128, 256, . . . , 8192}) of well established benchmark problems [9] (see Table 1). Starting from
the initial region [−5, 5]n containing the optimum with initial step size σ = 3 we optimize until
reaching the (rather exact) target precision of ftar = 10−10 .
All hyperparameters of LM-MA-ES and MA-ES are given in Algorithm 1. We use LM-CMA-ES
[26], VD-CMA-ES [1] and the active (µ/µw , λ)-CMA-ES [17, 21] (aCMA-ES, known to be up to
2 times more efficient than the default CMA-ES) as baselines. The source code of LM-MA-ES is
available in the supplementary material.
LM-MA-ES does not show performance degradation when applied to translated and rotated test
functions (see
also Figure 1 in the supplementary material). Figure 1 shows the effect of the
O n log(n) scaling of the runtime per sample as compared to O(n2 ) of fast-MA-ES (in the following
denoted as MA-ES), measured for implementations of the algorithms in plain C.
The much better internal scaling is of value only if the algorithm does not pay a too high price in terms
of an increased number of function evaluations required to reach ftar . Therefore we test LM-MA-ES
against MA-ES with full rank transformation matrix. Figure 2 shows that LM-MA-ES performs
5
fast MA−ES
N = 256
100
10−5
−10
10
N = 512
100
101
102
103
104
105
100
10−5
10
106
103
104
105
Discus
107
108
10−10 3
10
109
103
104
105
106
107
objective
10−5
10−10 −3
10
10−2
10−1
100
objective
10−5
103
104
101
102
105
106
107
10−10 2
10
108
10
−5
10
101
102
103
104
105
101
102
103
time in seconds
104
105
108
101
102
103
104
105
106
104
105
time in seconds
Different Powers
105
objective
105
100
10−5
10−10 −1
10
107
10−5
Cigar
objective
100
106
100
10−10 0
10
106
1010
10−5
105
10
time in seconds
100
104
Rosenbrock
0
Discus
10−10 −1
10
103
5
time in seconds
105
109
number of function evaluations
105
10−10 0
10
103
108
10−5
Ellipsoid
1010
100
107
100
number of function evaluations
Sphere
106
Different Powers
100
10−10 2
10
108
105
105
105
number of function evaluations
105
104
number of function evaluations
objective
10−10 2
10
10−5
Cigar
objective
objective
106
1010
10−5
N = 8192
100
number of function evaluations
100
N = 4096
105
−10
105
objective
N = 2048
Rosenbrock
105
number of function evaluations
objective
N = 1024
Ellipsoid
1010
objective
objective
N = 128
Sphere
objective
LM−MA−ES
105
100
101
102
103
time in seconds
104
105
100
10−5
10−10 −1
10
100
101
102
103
time in seconds
Figure 2: Runtime in number of function evaluations (rows 1–2) and seconds (rows 3–4) of LM-MA-ES in 128
to 8192 dimensions and fast MA-ES in 128 to 1024 dimensions on six standard benchmark problems.
surprisingly well: in many cases it is actually faster and this tends to happen more often for larger n.
LM-MA-ES is always faster in terms of wall clock time, in some cases by a factor of 100.
Figure 3 shows that LM-MA-ES scales favorably compared to LM-CMA-ES achieving better scaling
on the Rosenbrock and Discus functions, but a worse scaling on Cigar. The latter result might be due
to an improper setting of the hyperparameters, a problem that can most probably be fixed with the
technique proposed in [28]. VD-CMA is not able to solve some rotated functions efficiently due to
the restrictions on the covariance matrix that the algorithm assumes [1, 26].
4.2
Adversarial Inputs
Several standard classifiers like the k-nearest-neighbor predictor (without distance-based weights),
decision trees, and random forests are not differentiable with respect to their inputs. The predictions
are piecewise constant and hence trivially piecewise differentiable, however, the gradient is zero and
hence uninformative. This is a significant complication when generating adversarial inputs [38] for
such classifiers, a problem for which only rather weak attacks exist [22, 40]. We compare MA-ES
and LM-MA-ES on this task. To this end we train a random forest consisting of 1000 trees on the
MNIST data set (784 dimensional input, 10 classes, 60,000 training points, 10,000 test points). It
achieves a test error of 97.21%, which is clearly worse than the results obtained with convolutional
neural networks, however, the predictor is highly non-trivial in the sense of being far from guessing
performance. Let h denote the random forest predictor, let x0 denote a correctly classified test point
(of which we have 9721), and let y0 denote the corresponding label. For an input x, the random forest
predictor outputs a probability vector h(x) ∈ R10 . We create an adversarial version of (x0 , y0 ) by
6
4
10
10 2
10
4
10
4
10
aCMA−ES
MA−ES
LM−CMA−ES
LM−MA−ES
3
10
2
10 2
10
4
10
3
3
10
Dimension
Rotated Discus
Rosenbrock
5
10
aCMA−ES
MA−ES
VD−CMA−ES
LM−CMA−ES
LM−MA−ES
4
10
3
3
10 2
10
4
10
Dimension
Rotated Cigar
10
4
10
aCMA−ES
MA−ES
LM−CMA−ES
VD−CMA−ES
LM−MA−ES
Function Evaluations / Dimension
1
aCMA−ES
MA−ES
VD−CMA−ES
LM−CMA−ES
LM−MA−ES
Rotated Ellipsoid
aCMA−ES
MA−ES
LM−CMA−ES
LM−MA−ES
Function Evaluations / Dimension
Function Evaluations / Dimension
2
10
10 2
10
Function Evaluations / Dimension
5
10
Function Evaluations / Dimension
Function Evaluations / Dimension
Sphere
3
10
3
10
3
10
10
Dimension
Rotated Different Powers
4
aCMA−ES
MA−ES
LM−CMA−ES
LM−MA−ES
3
10
2
3
10
Dimension
4
10
10 2
10
3
10
Dimension
4
10
2
10
10
Dimension
4
Figure 3: Median number of function evaluations (out of 5 runs) required to achieve a target objective function
value of ftar = 10−10 . The results of VD-CMA-ES are shown only on functions where the algorithm succeeded
achieving ftar . Some results for CMA-ES and MA-ES are missing due to their extremely long runtimes (note
that the vertical axis shows function evaluations, not runtime).
minimizing the objective function
(
h(x)y0 − max{(h(x)i | i 6= y0 }
f (x) =
−1
kx−x0 k
if x is classified as y0
otherwise
(9)
starting from x = x0 with initial step size σ = 1, where an MNIST digit is encoded as a vector
x ∈ R784 , with gray values in the interval [0, 255]. The case distinction encodes a preference for
wrongly classified points: correctly classified points have a positive value, wrongly classified points
have a (better) negative value. For positive values the incentive is to reduce the difference between the
fraction of trees voting for the correct label y0 and the strongest alternative. For a wrongly classified
point the continuous trend moves the point back as close as possible to the original image, in terms of
Euclidean distance. Note that due to invariance under rank-preserving transformations, the second
term is exactly equivalent to minimization of the distance between x and x0 .
Each of the 9721 optimization runs was restricted to a very low budget of 1000 objective
function evaluations. This seems reasonable
from a security perspective, since querying the
classifier too frequently renders a remote attack
inefficient. We used the sklearn implementation
of the random forest [30] and implementations
of MA-ES and LM-MA-ES based on numpy.
The overall experiment takes about two hours Figure 4: Original image (left) and corresponding adversarial image (right), created through a run of LMfor LM-MA-ES and about 7 hours for MA-ES MA-ES. The images are visually indistinguishable. The
on a laptop. The optimizers managed to turn (boosted) difference image is shown in the center; mid
6152 (MA-ES) and 6321 (LM-MA-ES) images gray corresponds to a difference of zero.
into wrongly classified inputs. In all cases, they
were visually indistinguishable from the original test inputs (see Figure 4). The surprisingly high rate
of about 65% of correctly classified test inputs that were successfully turned into adversarial versions
demonstrates that random forests can be highly unstable and rather easy to fool by an adversarial.
From an optimization point of view, LM-MA-ES performed clearly better than plain MA-ES. It ran
about 3.5 times faster and even produced better results: in 7171 cases LM-MA-ES reached a lower
objective value, while MA-ES was better in only 2550 out of 9721 cases. This is also reflected by the
larger number of cases in which an adversarial input was found (6321 vs. 6152, see above).
7
Optimization progress over time is plotted in Figure 5 for four subsets of runs of varying difficulty.
In all cases, LM-MA-ES is slightly faster. Significant differences can be observed only for the
easiest problems, where MA-ES fails to improve the solution, presumably due to the small budget of
objective function evaluations. In a control experiment we verified that the performance of MA-ES is
only insignificantly improved compares to a simple ES without matrix adaptation, which is restricted
to the adaptation of the global step size.
The better quality of the results given the same number of objective function evaluations can be
explained by the low-rank structure of the LM-MA-ES steps. Adapting a full 784×784 transformation
matrix with only 1000 function evaluations is a hopeless undertaking, while the top 23 = b4 +
3 log(784)c directions are much easier to adapt. This difference allows LM-MA-ES to use learning
rates that are up to several orders of magnitude larger than for MA-ES, resulting in faster adaptation
to the problem at hand. From the difference image in Figure 4 we clearly see a bright blob, indicating
that LM-MA-ES has identified a subspace with a meaningful interpretation for the problem at hand.
In contrast, the plain ES without matrix adaptation applies only white noise distortions, and this is
also what’s encoded by the initial search distribution of MA-ES and LM-MA-ES.
Returning to our initial questions we conclude
that LM-MA-ES successfully marries the simplicity of the update mechanisms of MA-ES
0.6
with the graceful scaling of large-scale optimizers like LM-CMA-ES to large n. LM-MAES is competitive with MA-ES in terms of the
0.4
required number of objective function evaluations, while reducing the algorithm internal
cost
0.2
per sample considerably from O n2 down to
O n log(n) . Taken together this yields in sig0
nificant speed-ups for high-dimensional problems. By applying our algorithm to the problem −0.01
of generating adversarial inputs for a random
forest classifier, we demonstrate the value of
LM-MA-ES for this domain, and uncover a sur- −0.02
prising vulnerability of random forests with re−0.03
spect to the existence of adversarial inputs.
LM−MA−ES
MA−ES
0− 99
1000−1099
2000−2099
5000−5099
0
10
20
30
40
Figure 5: Progress over 1000 function evaluations (44
iterations) of LM-MA-ES and MA-ES on the task of
generating adversarial inputs. The figure shows four
5 Conclusion
objective function curves corresponding to equation (9)
for each algorithm, referring to medians over 100 runs
each. The 9721 problems were sorted by their initial
objective value. We consider problems 0 to 99 (the 100
The recently proposed Matrix Adaptation Evo- simplest problems), problems 1000 to 1099, 2000 to
lution Strategy is a simpler variant of the Co- 2099, and 5000 to 5099.
variance Matrix Adaptation Evolution Strategy.
We showed that the O(n2 ) time and space complexity of MA-ES, which is prohibiting
for large
n, can be reduced to O n log(n) adopting the
approach used in [26]. The proposed LimitedMemory Matrix Adaptation Evolution Strategy
matches the state-of-the-art results on largescale optimization problems while being algorithmically simpler than LM-CMA-ES.
Future work should investigate to which extent the inclusion of the rank-µ update can improve the
performance. The learning rates of LM-MA-ES can be optimized online as it is commonly done in
self-adaptive evolutionary algorithms [6] or based on the maximum-likelihood principle [28].
A promising venue for LM-MA-ES would be to accelerate Stochastic Gradient Descent (SGD) for
training deep neural networks by replacing the evolution path vectors by momentum vectors based
on noisy batch gradients. The method could potentially represent an alternative to L-BFGS and
numerous SGD variants with adaptive learning rates.
8
References
[1] Youhei Akimoto, Anne Auger, and Nikolaus Hansen. Comparison-based natural gradient
optimization in high dimension. In Genetic and Evolutionary Computation Conference, pages
373–380. ACM, 2014.
[2] Youhei Akimoto, Anne Auger, and Nikolaus Hansen. Quality gain analysis of the weighted
recombination evolution strategy on general convex quadratic functions. In Proceedings of the
14th ACM/SIGEVO Conference on Foundations of Genetic Algorithms, pages 111–126. ACM,
2017.
[3] Youhei Akimoto and Nikolaus Hansen. Online model selection for restricted covariance matrix
adaptation. In International Conference on Parallel Problem Solving from Nature, pages 3–13.
Springer, 2016.
[4] Youhei Akimoto, Yuichi Nagata, Isao Ono, and Shigenobu Kobayashi. Bidirectional relation
between cma evolution strategies and natural evolution strategies. In International Conference
on Parallel Problem Solving from Nature, pages 154–163. Springer, 2010.
[5] Hans-Georg Beyer. Convergence analysis of evolutionary algorithms that are based on the
paradigm of information geometry. Evolutionary Computation, 22(4):679–709, 2014.
[6] Hans-Georg Beyer and Kalyanmoy Deb. On self-adaptive features in real-parameter evolutionary
algorithms. IEEE Transactions on evolutionary computation, 5(3):250–270, 2001.
[7] Hans-Georg Beyer and Michael Hellwig. The Dynamics of Cumulative Step-Size Adaptation
on the Ellipsoid Model. Evolutionary Computation, 24(1):25–57, 2016.
[8] Hans-Georg Beyer and Bernhard Sendhoff. Simplify your covariance matrix adaptation evolution strategy. IEEE Transactions on Evolutionary Computation, 2017.
[9] Finck Finck, Nikolaus Hansen, Raymond Ros, and Anne Auger. Real-parameter black-box
optimization benchmarking 2010: Experimental setup. Technical Report 2009/21, Research
Center PPE, 2010.
[10] Tobias Glasmachers and Christian Igel. Uncertainty handling in model selection for support
vector machines. In G. Rudolph, T. Jansen, S. Lucas, C. Poloni, and N. Beume, editors, Parallel
Problem Solving from Nature (PPSN), pages 185–194. Springer, 2008.
[11] Faustino Gomez, Jürgen Schmidhuber, and Risto Miikkulainen. Accelerated neural evolution
through cooperatively coevolved synapses. Journal of Machine Learning Research, 9(May):937–
965, 2008.
[12] Nikolaus Hansen, Dirk V Arnold, and Anne Auger. Evolution strategies. In Springer Handbook
of Computational Intelligence, pages 871–898. Springer, 2015.
[13] Nikolaus Hansen, Anne Auger, Raymond Ros, Steffen Finck, and Petr Pošík. Comparing results
of 31 algorithms from the black-box optimization benchmarking BBOB-2009. In Proceedings
of the 12th annual conference companion on Genetic and evolutionary computation, GECCO
2010, pages 1689–1696, 2010.
[14] Nikolaus Hansen, Sibylle D. Müller, and Petros Koumoutsakos. Reducing the time complexity of
the derandomized evolution strategy with covariance matrix adaptation (CMA-ES). Evolutionary
Computation, 11(1):1–18, 2003.
[15] Nikolaus Hansen and Andreas Ostermeier. Adapting arbitrary normal mutation distributions in
evolution strategies: The covariance matrix adaptation. In Evolutionary Computation, 1996.,
Proceedings of IEEE International Conference on, pages 312–317. IEEE, 1996.
[16] Nikolaus Hansen and Andreas Ostermeier. Completely Derandomized Self-Adaptation in
Evolution Strategies. Evolutionary Computation, 9(2):159–195, June 2001.
[17] Nikolaus Hansen and Raymond Ros. Benchmarking a weighted negative covariance matrix
update on the BBOB-2010 noiseless testbed. In Genetic and Evolutionary Computation
Conference, pages 1673–1680. ACM, 2010.
9
[18] Verena Heidrich-Meisner and Christian Igel. Hoeffding and bernstein races for selecting policies
in evolutionary direct policy search. In Proceedings of the 26th Annual International Conference
on Machine Learning, pages 401–408. ACM, 2009.
[19] Christian Igel. Evolutionary kernel learning. In Encyclopedia of Machine Learning, pages
369–373. Springer, 2011.
[20] Jens Jägersküpper. Rigorous runtime analysis of the (1+ 1) ES: 1/5-rule and ellipsoidal fitness
landscapes. In International Workshop on Foundations of Genetic Algorithms, pages 260–281.
Springer, 2005.
[21] Graheme A. Jastrebski and Dirk V. Arnold. Improving Evolution Strategies through Active
Covariance Matrix Adaptation. In IEEE Congress on Evolutionary Computation, pages 2814–
2821, 2006.
[22] Alex Kantchelian, Doug Tygar, and Anthony Joseph. Evasion and hardening of tree ensemble
classifiers. In International Conference on Machine Learning, pages 2387–2396, 2016.
[23] James N Knight and Monte Lunacek. Reducing the space-time complexity of the CMA-ES. In
Genetic and Evolutionary Computation Conference, pages 658–665. ACM, 2007.
[24] Oswin Krause, Dídac Rodríguez Arbonès, and Christian Igel. CMA-ES with optimal covariance
update and storage complexity. In Advances in Neural Information Processing Systems, pages
370–378, 2016.
[25] Dong C Liu and Jorge Nocedal. On the limited memory BFGS method for large scale optimization. Mathematical programming, 45(1-3):503–528, 1989.
[26] Ilya Loshchilov. LM-CMA: An alternative to L-BFGS for large-scale black box optimization.
Evolutionary Computation, 25(1):143–171, 2017.
[27] Ilya Loshchilov and Frank Hutter. CMA-ES for Hyperparameter Optimization of Deep Neural
Networks. arXiv preprint arXiv:1604.07269, 2016.
[28] Ilya Loshchilov, Marc Schoenauer, Michele Sebag, and Nikolaus Hansen. Maximum Likelihoodbased Online Adaptation of Hyper-parameters in CMA-ES. In Parallel Problem Solving from
Nature–PPSN, pages 70–79. Springer, 2014.
[29] Yann Ollivier, Ludovic Arnold, Anne Auger, and Nikolaus Hansen. Information-geometric
optimization algorithms: A unifying picture via invariance principles. Journal of Machine
Learning Research, 18(18):1–65, 2017.
[30] Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion,
Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard
Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research,
12:2825–2830, 2011.
[31] Leonard Rastrigin. In the world of random events. Latvian Academy of Science, USSR, Riga,
1963.
[32] Ingo Rechenberg. Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der
biologischen Evolution. Frommann-Holzboog, 1973.
[33] Raymond Ros and Nikolaus Hansen. A simple modification in CMA-ES achieving and space
complexity. In Parallel Problem Solving from Nature–PPSN, pages 296–305. Springer, 2008.
[34] Tim Salimans, Jonathan Ho, Xi. Chen, and Ilya Sutskever. Evolution strategies as a scalable
alternative to reinforcement learning. Technical Report arXiv:1703.03864, arxiv.org, 2017.
[35] David Shanno. Conditioning of Quasi-Newton Methods for Function Minimization. Math. of
Computation, 24(111):647–656, 1970.
[36] Freek Stulp and Olivier Sigaud. Robot skill learning: From reinforcement learning to evolution
strategies. Paladyn, Journal of Behavioral Robotics, 4(1):49–61, 2013.
10
[37] Yi Sun, Faustino Gomez, Tom Schaul, and Juergen Schmidhuber. A linear time natural evolution
strategy for non-separable functions. arXiv preprint arXiv:1106.1998, 2011.
[38] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian
Goodfellow, and Rob Fergus. Intriguing properties of neural networks. Technical Report
arXiv:1312.6199, arxiv.org, 2013.
[39] Daan Wierstra, Tom Schaul, Tobias Glasmachers, Yi Sun, Jan Peters, and Jürgen Schmidhuber.
Natural evolution strategies. The Journal of Machine Learning Research, 15(1):949–980, 2014.
[40] Weilin Xu, Yanjun Qi, and David Evans. Automatically evading classifiers. In Proceedings of
the 2016 Network and Distributed Systems Symposium, 2016.
11
Supplementary Material for the paper Limited-Memory Matrix Adaptation
for Large Scale Black-box Optimization
I
LM-MA-ES with the same procedure to store direction vectors as in
LM-CMA-ES
As it is mentioned in the main text of the paper, our original approach to design LM-MA-ES was
to employ the same procedure to storage direction vectors as used in LM-CMA-ES [26]. The
corresponding variant of LM-MA-ES with O(mmax n) time and space complexity is given below in
Algorithm 1.
Instead of M(t) ∈ IRn×n used in the original MA-ES, we employ a storage matrix M(t) ∈ IRmmax ×n
whose lines store the evolution path vectors computed at different iterations of the algorithm. While
(t)
the original MA-ES employs pσ to adapt both σ (t) and M(t) , we note that the optimal time horizon
(t)
of the two processes can be different. Therefore, we separately adapt pσ with the learning rate
(t)
1
cσ = mmax
and pc with the learning rate cc = n1 (see lines 14 and 15, respectively). At the first
mmax iterations of the algorithm, LM-MA works as MA-ES without the rank-µ update, i.e., with
cµ = 0, and a different setting for hyperparameters, e.g., cc 6= cσ . The equivalence is due to the
(t)
use of the first mmax evolution path vectors pc to reproduce M(t=mmax ) based on the the rank-one
update exactly. When mmax < n, one can greatly reduce the time and space complexity of the
final algorithm. However, the use of the most recent mmax evolution path vectors would lead to a
degenerative sampling. Therefore, following [26], we adopt a reference array ref (thought of as a
(t)
matrix, storing one vector per row) to access pc vectors in M(t) such that the i-th line of M(t) , i.e.,
(t)
Mrefi , belongs to the i-th oldest vector (in terms of its iteration index t) among the stored ones (see
line 25). In order to well approximate the effect that would be obtained with the full matrix based on
the rank-one update, we force the storage to support a temporal distance between the evolution path
vectors such that the temporal distance timerefi+1 − timerefi between i + 1-th and i-th vectors is not
greater than Ni = n2 /mmax (see lines 20-22). This procedure replaces the vector with the smallest
(t+1)
temporal distance to its older neighbor by the most recent vector pc
(see line 25). Periodically,
the oldest vector is also removed to constrain the distance according to N (see line 22). The step-size
adaptation of LM-MA-ES is the same as in MA-ES and CMA-ES, i.e., based on the cumulative
step-size adaptation (CSA) rule of [15] (see line 26).
As it is mentioned in section 3 of the main paper, we found that instead of a rather complicated
procedure described in this section, one can continuously update a set of m vectors that is distantly
similar to computing momentum vectors in Stochastic Gradient Descent. The idea was born while
attempting to implement Algorithm 1 for training deep neural networks. Since the performance of
both algorithms is comparable, we decided to present the simpler one (given in the main paper) as
our main method.
II
Invariance
LM-MA-ES, just like MA-ES, is invariant to translation, rotation and scaling of the objective function
in search space, provided that the initial search distribution is transformed accordingly. However,
in order to justify the use of separable test functions in the experimental evaluation, we validate
this property empirically, by showing median runs (out of 5 runs) of the algorithm on separable
and rotated problems. Figure 1 shows convergence curves. The only systematic effect is due to
initialization in the hypercube [−5, 5]n , which is not rotated. The deviations are minimal, and no
larger than the usual deviations due to randomized initialization and operation of the algorithm. This
is in contrast to VD-CMA-ES [1], which is unable to solve rotated versions of some of the test
problems.
1
Algorithm 1 LM-MA-ES with the same procedure to store direction vectors as in LM-CMA-ES
1: given n ∈ N+ , mmax
ln(µ+ 12 )−ln i
Pµ
1
j=1 (ln(µ+ 2 )−ln j)
2: initialize t ← 0, y
mmax ×n
3:
4:
5:
6:
7:
8:
9:
10:
11:
= 4 + b3 ln nc, λ = 4 + b3 ln nc, µ = bλ/2c, wi =
1
, cc = n1 , dσ = 0.5,
for i = 1, . . . , µ, µw = Pµ 1 w2 , c1 = n1 , cσ = mmax
i=1
(t=0)
n
(t=0)
i
(t=0)
(t=0)
(t=0)
∈ IR , m
= 0, σ
> 0, pσ
= 0, pc
max −1
max
, N ∈ Nm
= n2 /mmax
∈ Nm
+
+
= 0, M ∈
max
, time
Nm
+
IR
, ref ∈
repeat
for i ← 1, . . . , λ do
(t)
zi ← N (0, I )
(t)
(t)
di ← zi
for j ← m(t) , . . . , 1 do
0
c ← c1
if j < m(t) then
0
c1 ← c1 (timerefj+1 − timerefj )/Nj
P
0
0
(t)
(t)
(t)
n
di ← (1 − c1 )di + c1 Mrefj
M
d
refj ,k i,k
k=1
(t)
(t)
22:
23:
24:
fi ← f (y(t) + σ (t) di )
Pµ
(t)
(t+1)
y
← y(t) + σ (t) i=1 wi di:λ // the symbol i : λ denotes i-th best sample on f
p
Pµ
(t)
(t+1)
(t)
← (1 − cσ )pσ + µw cσ (2 − cσ ) i=1 wi zi:λ
pσ
p
Pµ
(t)
(t+1)
(t)
← (1 − cc )pc + µw cc (2 − cc ) i=1 wi zi:λ
pc
m(t+1) ← min(t + 1, mmax )
if (t < mmax ) then
refm(t+1) ← t + 1
else
imin ← argmini timerefi+1 − timerefi − Ni , |1 ≤ i ≤ (m(t+1) − 1)
if timerefi +1 − timerefi
− Nimin ≥ 0 then
min
min
imin ← 0
remove (imin + 1)-th element of ref and insert it at now empty mt+1 -th position
timerefm(t+1) ← t + 1
25:
Mrefm(t+1) ← pc
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
26:
(t+1)
σ
(t+1)
←σ
(t)
exp
cσ
dσ
2
k
kp(t+1)
σ
n
−1
27:
t←t+1
28: until stopping criterion is met
2
Separable Ellipsoid (bold) vs Rotated Ellipsoid (dotted)
5
10
Separable Discus (bold) vs Rotated Discus (dotted)
5
10
dim 256
dim 512
0
dim 256
dim 512
0
Objective
10
Objective
10
−5
−5
10
10
−10
10
5
−10
10
2
4
6
8
Number of function evaluations
10
6
x 10
Separable Cigar (bold) vs Rotated Cigar (dotted)
10
0
0.5
1
1.5
2
2.5
3
3.5
4
6
Number of function evaluations
x 10
Separable DiffPow (bold) vs Rotated DiffPow (dotted)
10
dim 256
dim 512
dim 256
dim 512
−2
10
0
Objective
Objective
10
−5
−4
10
−6
10
10
−8
10
−10
10
−10
2
10
4
6
8
10
12
5
Number of function evaluations
x 10
2
4
6
8
Number of function evaluations
10
5
x 10
Figure 1: The trajectories show the median of 5 runs of LM-MA-ES on separable (bold lines) and rotated
(dotted lines) versions of Ellipsoid, Discus, Cigar and Different Powers functions (see Table 1 of the main paper)
for problem dimensions 256 and 512. The rotated functions f (Rx) are obtained from the original separable
functions f (x) by applying R, an orthogonal n × n matrix with each column vector qi being a uniformly
distributed unit vector implementing an angle-preserving transformation. LM-MA-ES numerically demonstrates
invariance to rotations. The observed difference is due to the fact that the rotations effectively change the initial
range of the search domain.
3
| 9cs.NE
|
Robust Multivariate Nonparametric Tests via
Projection-Pursuit
Ilmun Kim†
Sivaraman Balakrishnan†
Larry Wasserman†
arXiv:1803.00715v1 [math.ST] 2 Mar 2018
Department of Statistics and Data Science†
Carnegie Mellon University
Pittsburgh, PA 15213
Abstract
In this work, we generalize the Cramér-von Mises statistic via projection pursuit to
obtain robust tests for the multivariate two-sample problem. The proposed tests are
consistent against all fixed alternatives, robust to heavy-tailed data and minimax rate
optimal. Our test statistics are completely free of tuning parameters and are computationally efficient even in high dimensions. When the dimension tends to infinity, the
proposed test is shown to have identical power to that of the existing high-dimensional
mean tests under certain location models. As a by-product of our approach, we introduce
a new metric called the angular distance which can be thought of as a robust alternative
to the Euclidean distance. Using the angular distance, we connect the proposed to the
reproducing kernel Hilbert space approach. In addition to the Cramér-von Mises statistic,
we show that the projection pursuit technique can be used to define robust, multivariate
tests in many other problems.
1
Introduction
Let X and Y be random vectors in Rd with distribution functions FX and GY , respectively.
Given two independent samples Xm = {X1 , . . . , Xm } and Yn = {Y1 , . . . , Yn } from FX and
GY , we want to test
H0 : FX = GY
versus
H1 : FX 6= GY .
(1)
This fundamental testing problem has received considerable attention in statistics with a
wide range of applications (see, e.g., Thas, 2010, for a review). A common statistic for the
univariate two-sample testing is the Cramér-von Mises (CvM) statistic (Anderson, 1962):
Z ∞
2
mn
b
b Y (t) dH(t),
FbX (t) − G
m+n ∞
b Y (t) are the empirical distribution functions of Xm and Yn , respectively,
where FbX (t) and G
b
b Y (t). At the population level, the CvM statistic can be
and (m + n)H(t)
= mFbX (t) + nG
generalized by
Z ∞
2
W (α) =
(FX (t) − GY (t))2 d (αFX (t) + (1 − α)GY (t))
(2)
−∞
for some α ∈ (0, 1). In fact, as pointed out by Bergsma and Dassios (2014), W 2 (α) is invariant
to the choice of α for continuous random variables; hence, we simply focus on α = 1/2 and
1
let W 2 (1/2) = W 2 for ease of notation. Lehmann (1951) provides an interesting expression
for W 2 given by
P max{X, X 0 } < min{Y, Y 0 } + P max{Y, Y 0 } < min{X, X 0 }
(3)
1
+ 2W 2 ,
=
3
i.i.d.
i.i.d.
where X, X 0 ∼ FX and Y, Y 0 ∼ GY . Using this expression, a probabilistic interpretation
of W 2 is available in terms of concordance and discordance probabilities of four points; see
Baringhaus and Henze (2017).
Another popular two-sample test statistic, which is often called the energy statistic (Székely
and Rizzo, 2013), is an estimate of the squared energy distance:
E 2 = 2EX,Y [|X − Y |] − EX,X 0 [|X − X 0 |] − EY,Y 0 [|Y − Y 0 |].
The energy distance is well-defined assuming a finite first moment and it can be written in a
form that is similar to the Cramér’s distance, namely,
Z ∞
2
(FX (t) − GY (t))2 dt.
E =2
−∞
For univariate two-sample testing, the CvM statistic has several advantages over the energy
statistic. For instance, when FX and GY are continuous, the CvM statistic is distribution-free
under H0 (Anderson, 1962) and its population quantity is well-defined without any moment
assumptions. Unlike the energy statistic, it also has an intuitive probabilistic interpretation
according to (3).
Despite its benefits in the univariate problem, the CvM statistic has rarely been studied
for multivariate testing. A primary reason is that the CvM statistic is rank-based, which leads
to a challenge to generalize it in a multivariate space. In contrast, the energy statistic can be
easily applied in arbitrary dimensions as in Baringhaus and Franz (2004); Székely and Rizzo
(2004). Specifically, they defined the squared multivariate energy distance by
Ed2 = 2EX,Y [||X − Y ||] − EX,X 0 [||X − X 0 ||] − EY,Y 0 [||Y − Y 0 ||],
(4)
where || · || is the Euclidean norm in Rd . As in the univariate case, the multivariate energy
distance maintains the characteristic property that it is always non-negative and equal to zero
if and only if FX = GY . It is worth mentioning that Ed2 can also be viewed as the average of
univariate Cramér’s distances of projected random variables (Baringhaus and Franz, 2004):
√
Z
Z
2
π(d − 1)Γ( d−1
2
2 )
Ed =
Fβ > X (t) − Gβ > Y (t) dtdλ(β),
(5)
d
Γ( 2 )
Sd−1 R
where λ represents the uniform probability measure on the d-dimensional unit sphere Sd−1 =
{x ∈ Rd : ||x|| = 1}. For each direction β ∈ Sd−1 , Fβ > X and Gβ > Y denote the distribution
functions of β > X and β > Y , respectively.
Although the multivariate energy distance can be simply estimated in any dimension, it
still requires the finite moment assumption as in the univariate case; thereby, the resulting
test might suffer from low power when the underlying distributions are heavy-tailed or contain
outliers. Given that outlying observations arise frequently in practice with high-dimensional
data, there is a need to develop a robust counterpart of the energy distance. The primary goal
2
of this work is to introduce a robust, tuning parameter free, two-sample testing procedure that
is easily applicable in arbitrary dimensions and consistent against all alternatives. Specifically,
we extend the univariate CvM distance W to an arbitrary dimension based on averaging over
all one-dimensional projections. Specifically, we define the squared multivariate CvM distance
as follows:
Z
Z
2
Wd2 =
Fβ > X (t) − Gβ > Y (t) dωβ (t)dλ(β),
(6)
Sd−1
R
where
dωβ (t) = d(Fβ > X (t)/2 + Gβ > Y (t)/2).
Throughout this paper, we refer to the process of averaging over all projections as projection
pursuit.
The proposed multivariate CvM distance shares some attractive properties of the energy
distance while being robust to heavy-tailed distributions or outliers. For example, Wd satisfies
the characteristic property (Lemma 2.1) and is invariant to orthogonal transformations. More
importantly, from Theorem 2.1, it is straightforward to estimate Wd without using any tuning
parameters.
For the first part of this paper, we investigate properties of the multivariate CvM distance
as well as the resulting testing procedures. Our main results are summarized as follows:
• Closed form expression (Section 2) We show that the test statistic has a simple
closed-form expression.
• Robustness (Section 3) We show that the test based on the CvM distance maintains
good power in the Huber contamination model, while the energy test becomes completely
powerless in this setting.
• Minimax optimality (Section 4) We analyze the statistical power of the proposed
tests over a wide class of distributions and prove their minimax rate optimality.
• Adaptivity in high-dimensions (Section 5) We consider a high-dimensional regime
where the dimension tends to infinity. Under this regime, we show that the limiting
power of the proposed test, the energy test and the high-dimensional mean tests introduced by Chen and Qin (2010) and Chakraborty and Chaudhuri (2017) are the same
against certain location models.
• Angular distance (Section 6) We introduce the angular distance between two vectors
and use this to show that the multivariate CvM distance is a member of the generalized
energy distance (Sejdinovic et al., 2013) as well as the maximum mean discrepancy
(MMD) (Gretton et al., 2012) associated with the angular distance.
Beyond the CvM statistic, the projection pursuit technique can be widely applicable to
other nonparametric statistics. For the second part of this study, we revisit some famous
univariate sign- or rank-based statistics and propose their multivariate counterparts via projection pursuit. Although there has been much effort to extend univariate sign- or rank-based
statistics in a multivariate space (see, e.g., Liu, 2006; Oja, 2010, for a survey), they are either
computationally expensive to implement or less intuitive to understand. Our projection pursuit approach addresses these issues by providing a tractable calculation form of statistics and
by having a direct interpretation in terms of projections. In Section 7, we demonstrate the
generality of the projection pursuit approach by presenting multivariate extensions of several
existing univariate statistics.
3
Table 1: The comparisons between different two-sample testing methods: 1) Cramér-von Mises
distance (CvM), 2) Energy distance (Energy), 3) Gaussian MMD (MMD) and 4) k-nearest
neighbor (NN).
Characteristic
1. Robustness
2. No tuning parameters
3. Interpretability
4. Computational cost
5. Minimax optimality
1.1
CvM
X
X
X
X
X
Energy
×
X
×
X
?
MMD
×
×
×
X
?
NN
×
×
X
X
?
Literature review
There are a number of multivariate two-sample testing procedures available in the literature.
To list a few: Anderson et al. (1994) propose the two-sample statistic based on the integrated
square distance between two kernel density estimates. The energy statistic is introduced by
Baringhaus and Franz (2004) and Székely and Rizzo (2004) independently. Gretton et al.
(2012) introduce a class of distances between two probability distributions, called the maximum mean discrepancy (MMD). They apply the MMD to multivariate two-sample testing
with the Gaussian kernel. Sejdinovic et al. (2013) show that the energy distance is a special
case of the MMD associated with the kernel induced by the Euclidean distance. Another
line of work is based on graph constructions including the k-nearest neighbor (NN) (Schilling,
1986; Henze, 1988; Mondal et al., 2015) and the minimum spanning tree (Friedman and Rafsky, 1979; Chen and Friedman, 2017). However, there are certain drawbacks associated with
the use of these methods. For example, one needs to choose tuning parameters such as the
bandwidths for the Gaussian MMD and the number of the nearest neighbors in the k-NN.
In addition, certain types of graph-based tests require an algorithm with high computational
complexity. On the other hand, the proposed approach is completely free of tuning parameters and can be computed in O(m + n) if a linear-type statistic is used (see Section 4) and in
O((m + n)3 ) if a U - or V -statistic is used.
The projection pursuit approach to CvM-type statistics can be found in other statistical
problems. For example, Zhu et al. (1997) and Cui (2002) consider the CvM statistic using projection pursuit to investigate one-sample goodness-of-fit tests for multivariate distributions.
Escanciano (2006) proposes the CvM-based goodness-of-fit test for parametric regression models. To the best of our knowledge, however, this is the first study that investigates the CvM
statistic for the multivariate two-sample problem using projection pursuit.
The advantages of our approach are summarized in Table 1 where we compare the proposed
test to some popular existing tests.
Outline. The rest of this paper is organized as follows. In Section 2, we introduce
test statistics and study their limiting distributions. In Section 3, we compare the power
of the CvM test with that of the energy test and we highlight the robustness of the CvM
test. Section 4 establishes the minimax rate optimality of the proposed tests. In Section 5,
we study the limiting power of the CvM test in high-dimensions and prove that it becomes
equivalent to that of the existing high-dimensional mean tests against certain location models.
We introduce the angular distance between two vectors in Section 6 to show that the CvM
4
distance is a special case of the generalized energy distance associated with the introduced
distance. In Section 7, the projection pursuit technique is applied to other sign- or rank-based
statistics and and this allows us to provide new multivariate extensions. Simulation results
are presented in Section 8 to demonstrate the competitive power performance of the proposed
approach with finite sample size. Throughout this paper, we assume the dimension d is an
arbitrary but fixed constant except in Section 5 where we allow the dimension to increase with
sample size. All proofs not contained in the main text are in the supplementary material.
Notation. The following notation is used throughout this paper. For U1 , U2 ∈ Rd , we
denote the angle between U1 and U2 by Ψ(U1 , U2 ) = arccos(U1 /||U1 ||, U2 /||U2 ||) where || · || is
the Euclidean norm in Rd . Let Sd−1 = {x ∈ Rd : ||x|| = 1} and let λ(·) be the uniform uniform
probability measure on Sd−1 . For given β ∈ Sd−1 , Fβ > X (t), Gβ > (t) denote the distribution
functions of β > X and β > Y , respectively. For 1 ≤ q < p, we let (p)q = p(p − 1) · · · (p − q + 1).
Let P0 and P1 be the probability measures under H0 and H1 , respectively. Similarly, E0
and E1 stand for the expectations under H0 and H1 . We denote the distribution function of
N (0, 1) by Φ and let zα be the upper α quantile of N (0, 1). We use an bn if there exist
constants C, C 0 > 0 such that C < |an /bn | < C 0 . Similarly, an & bn if there exists C > 0 such
that Cbn ≥ an .
2
Projection Pursuit-Type Cramér-von Mises Statistics
In this section, we present the basic properties of the CvM distance and introduce our test
statistics. To begin, we establish the characteristic property of the CvM distance, meaning
that Wd is nonnegative and equal to zero if and only if FX = GY . As a consequence, the
resulting tests are consistent against all fixed alternatives.
Lemma 2.1. Wd is nonnegative and has the characteristic property:
Wd (FX , GY ) = 0
if and only if
FX = GY .
Note that Wd involves integration over the unit sphere. Zhu et al. (1997) suggest an
approximation of this integral by considering a subset of Sd−1 , namely {β1 , . . . , βk }, and then
taking the sample mean over k different univariate CvM statistics. However, this approach
has a clear trade-off between accuracy and computational time depending on the choice of
k. Our approach does not suffer from this issue by explicitly calculating the integral over
Sd−1 . The explicit form of the integration is mainly due to Escanciano (2006) who provides
the following lemma:
Lemma 2.2. (Escanciano, 2006) For any two vectors U1 , U2 ∈ Rd ,
Z
1
1
U1> U2
>
>
I(β U1 ≤ 0)I(β U2 ≤ 0)dλ(β) = −
arccos
.
2 2π
||U1 ||||U2 ||
Sd−1
The angle between U1 and U2 is denoted by
Ψ (U1 , U2 ) := arccos
U1> U2
||U1 ||||U2 ||
.
Based on Lemma 2.2, we present another representation for Wd2 in terms of the expected
angle involving three random vectors.
5
Theorem 2.1. Let X, X 0 , X 00
distance can be written as
1
Wd2 =
E Ψ
2π
1
− E Ψ
4π
1
− E Ψ
4π
i.i.d.
i.i.d.
∼ FX and Y, Y 0 , Y 00 ∼ GY . The squared multivariate CvM
1
E Ψ X − Y 0, Y − Y 0
2π
1
X − X 00 , X 0 − X 00 −
E Ψ X − Y, X 0 − Y
4π
1
Y − Y 00 , Y 0 − Y 00 −
E Ψ Y − X, Y 0 − X .
4π
X − X 0, Y − X 0
+
When FX and GY are continuous, we have that
Wd2 =
2.1
1
1
1
−
E Ψ X − Y, X 0 − Y −
E Ψ Y − X, Y 0 − X .
3 2π
2π
Test Statistics
Theorem 2.1 leads to a straightforward empirical estimate of Wd2 based on either a U -statistic
or a V -statistic. Consider the kernel of order three:
1
1
h(x1 , x2 , x3 ; y1 , y2 , y3 ) =
Ψ (x1 − x2 , y1 − x2 ) +
Ψ (x1 − y2 , y1 − y2 )
2π
2π
−
1
1
Ψ (x1 − x3 , x2 − x3 ) −
Ψ (x1 − y1 , x2 − y1 )
4π
4π
−
1
1
Ψ (y1 − y3 , y2 − y3 ) −
Ψ (y1 − x1 , y2 − x1 ) .
4π
4π
Using this kernel, the U - and V -statistics (the latter is equivalent to a plug-in estimator;
see Lemma B.1) are defined as
X
X
1
Um,n =
h(Xi1 , Xi2 , Xi3 ; Yj1 , Yj2 , Yj3 ),
(m)3 (n)3
i1 ,i2 ,i3
j1 ,j2 ,j3
all distinct all distinct
Vm,n =
X
1
m 3 n3
X
h(Xi1 , Xi2 , Xi3 ; Yj1 , Yj2 , Yj3 )
(7)
i1 ,i2 ,i3 j1 ,j2 ,j3
Z
Z
=
Sd−1
b β > Y (t)
Fbβ > X (t) − G
2
b β > Y (t)/2 dλ(β).
d Fbβ > X (t)/2 + G
R
When the random vectors are continuous, the statistic can be simplified to a kernel of
order two:
1
1
1
Ψ (x1 − y1 , x2 − y1 ) −
Ψ (y1 − x1 , y2 − x1 ) ,
h∗ (x1 , x2 ; y1 , y2 ) = −
3 2π
2π
and the corresponding U - and V -statistics are defined as
∗
Um,n
m
X
1
=
(m)2 (n)2
n
X
h∗ (Xi1 , Xi2 ; Yj1 , Yj2 ),
i1 ,i2 =1 j1 ,j2 =1
i1 6=i2 j1 6=j2
∗
Vm,n
m
X
1
= 2 2
m n
n
X
h∗ (Xi1 , Xi2 ; Yj1 , Yj2 ).
i1 ,i2 =1 j1 ,j2 =1
6
Focusing on the case where FX and GY are continuous, we explore the asymptotic null
∗
∗ .
distribution of Um,n
and Vm,n
Theorem 2.2. Let m/N → π1 ∈ (0, 1) and n/N → π2 ∈ (0, 1) with N = m + n as N →
∞. Let λk be the eigenvalues with the corresponding eigenfunction φk satisfying the integral
equation
(
)
h
i
h∗ (x1 , X2 ; Y1 , Y2 ) φk (X2 ) = λk φk (x1 ) for k = 1, 2, . . . ,
EX EY ,Y e
2
1
2
∗
∗
where e
h∗ (x1 , x2 ; y1 , y2 ) = 1/2h∗ (x1 , x2 ; y1 , y2 ) + 1/2h∗ (x1 , x2 ; y2 , y1 ). Then Um,n
and Vm,n
have the same limiting null distribution as
∞
∗ d X
mn
∗
Um,n
− E0 Um,n
−→
λk (Zk2 − 1),
m+n
k=1
mn
∗
V ∗ − E0 Vm,n
m + n m,n
d
−→
∞
X
λk (Zk2 − 1),
k=1
∗
∗
1
1
i.i.d.
+
.
where Zk ∼ N (0, 1), E0 Um,n
= 0 and E0 Vm,n
=
6m 6n
∗
∗ are non-degenerate U - and V -statistics. Using
Remark 2.1. When FX 6= GY , Um,n
and Vm,n
∗
the asymptotic theory of non-degenerate U - and V -statistics (see e.g., Lee, 1990), Um,n
and
∗
Vm,n converge to a normal distribution under the alternative hypothesis.
2.2
Critical Values
∗
∗
Unfortunately, when d ≥ 2, neither Um,n
nor Vm,n
is distribution-free. The distributions of
the test statistics rely on the dependence structure in a multivariate distribution. For this
reason, we adaptively choose a critical value using the permutation approach as follows:
1. We regard the data as i.i.d. samples {Z1 , . . . , Zm+n } from the mixture π1 FX + π2 GY
where m/N → π1 and n/N → π2 .
2. Let Tm,n be the test statistic of interest calculated based on Xm = {Z1 , . . . , Zm } and
Yn = {Zm+1 , . . . , Zm+n }.
3. Define cα,m,n by
cα,m,n := inf{t ∈ R : 1 − α ≤ P(Tm,n ≤ t)}.
(8)
Now, π1 FX + π2 GY is unknown and thus cα,m,n needs to be estimated. In Proposition 2.1,
we show that cα,m,n of the proposed statistics can be consistently estimated via the permutation procedure. First, consider the centered and scaled test statistics
∗
mn
∗
em,n := mn U ∗ − E U ∗
U
and Vem,n :=
Vm,n
− E Vm,n
.
m,n
m,n
m+n
m+n
Then we have the following result when m/N → π1 ∈ (0, 1) and n/N → π2 ∈ (0, 1) as
N → ∞:
7
em,n , Vem,n , respectively,
Proposition 2.1. Let cuα,m,n , cvα,m,n be the critical values based on U
u
v
decided by (8). Also, let ĉα,m,n , ĉα,m,n be the corresponding critical values from the permutation
em,n , Vem,n so that P0 (U
em,n ≥ ĉuα,m,n ) ≤ α and P0 (Vem,n ≥ ĉvα,m,n ) ≤ α for any
distributions of U
finite sample size. Then
p
cuα,m,n − ĉuα,m,n −→ 0
p
cvα,m,n − ĉvα,m,n −→ 0.
and
Proof. This result is a simple application of Theorem A.1 in Appendix A. A detailed proof is
provided in the supplementary material.
Proposition 2.1 also implies that the power of the permutation tests converge to that of
the unconditional tests as the sample size tends to infinity. Therefore, the permutation tests
maintain asymptotic power of the corresponding unconditional tests while being exact under
finite sample size. We formally state this in the following:
Corollary 2.1. Consider the permutation test functions
φ∗Ue
m,n
em,n ≥ ĉuα,m,n )
:= I(U
φ∗Ve
and
m,n
:= I(Vem,n ≥ ĉvα,m,n ),
em,n and
where ĉuα,m,n and ĉvα,m,n are critical values based on permutation distributions of U
Vem,n , respectively. Similarly, define the unconditional tests φUem,n and φVem,n by replacing
ĉuα,m,n , ĉvα,m,n with cuα,m,n , cvα,m,n . Then
E1 [φ∗Ue
m,n
] − E1 [φUem,n ] → 0 and E1 [φ∗Ve
m,n
] − E1 [φVem,n ] → 0 as m, n → ∞.
Proof. The result follows by Slutsky’s theorem.
Remark 2.2. In Appendix A, we establish general conditions under which the permutation
distribution based on a degenerate U -statistic is asymptotically equivalent to the corresponding
unconditional distribution. This general result includes the energy statistic, Gaussian MMD
and Proposition 2.1 as a corollary.
3
Robustness
Recall that the energy distance and the CvM distance can be represented by integrals of the
L2 -type difference between two distribution functions. In view of this, the main difference
between the energy distance and the CvM distance is in their weight function. More precisely,
the energy distance is defined with dt, which gives the uniform weight on the whole real line.
On the other hand, the CvM distance is defined with d(Fβ > X (t)/2 + Gβ > Y (t)/2), which gives
the most weight on high-density regions. As a result, the test based on the CvM distance is
more robust to extreme observations than the one based on the energy distance. It is also
important to note that the CvM distance does not require any moment conditions. On the
other hand, the energy distance is only well-defined assuming a finite first moment. When
the moment condition is violated or there exist extreme observations, the test based on the
energy distance can have low power. In this section, we demonstrate this point by using
contaminated distribution models.
Suppose we observe samples from an -contamination model:
X ∼ Pm := (1 − )FX + Hm
and Y ∼ Qn := (1 − )GY + Hn0 ,
8
(9)
where FX and GY are fixed but Hm and Hn0 can change arbitrarily with m, n. With this model
assumption, we would like to assess whether a test can maintain adequate power even when
Hm , Hn0 have an adverse impact on power performance. We mainly focus on statistical power
to study robustness because one can always employ the permutation procedure to control the
type-I error under H0 : Pm = Qn .
To state the main result, consider the energy statistic based on a U -statistic:
Em,n
m n
2 XX
1 X
=
||Xi1 − Xi2 ||
||Xi − Yj || −
mn
(m)2
−
(10)
i1 6=i2
i=1 j=1
1 X
||Yj1 − Yj2 ||.
(n)2
j1 6=j2
Under the given setting, the following result shows that the CvM test is uniformly powerful
over different types of contaminations as the sample size tends to infinity. However, the energy
test becomes asymptotically powerless against certain contaminations.
Theorem 3.1. Suppose we observe samples Xm and Yn from the contaminated model in (9)
with a fixed contamination ratio 0 < < 1. Assume that there exist δ1 , δ2 > 0 such that
Ed (Pm , Qn ) ≥ δ1
and
Wd (Pm , Qn ) ≥ δ2 ,
for every m, n,
(11)
and `, u ∈ (0, 1) such that ` ≤ m/N, n/N ≤ u for every m, n. Consider the tests based on
Um,n and Em,n , defined by
φCvM := I(Um,n ≥ cα,m,n )
and
φEnergy := I(Em,n ≥ c0α,m,n ),
where cα,m,n and c0α,m,n are α level critical values. Then the power of the CvM test is asymptotically powerful uniformly over any choice of (Hm , Hn0 ), whereas the power of the energy test
becomes asymptotically powerless against certain choices of (Hm , Hn0 ), that is
lim
inf E1 [φCvM ] = 1 and
0
m,n→∞ Hm ,Hn
lim
inf E1 [φEnergy ] ≤ α.
0
m,n→∞ Hm ,Hn
(12)
To further illustrate Theorem 3.1 with finite sample size, we carried out simulation studies
based on the normal contamination models. We consider two alternatives with 1) location
differences and 2) scale differences. For both examples, we let Hm = Hn0 have multivariate
normal distributions given by
Hm = Hn0 := N ((0, . . . , 0)> , σ 2 Id ),
where σ controls the degree of heavy-tailedness. Note that the condition in (11) is easily met
whenever Hm = Hn0 , but FX 6= GY (with the extra first moment condition for Ed ), which can
be seen from the representations for Ed2 and Wd2 in (5) and (6).
Example 3.1 (Location difference). For the location alternative, we compare two multivariate
normal distributions, where the means are different but the covariance matrices are identical.
Specifically, we set
FX = N ((−0.5, . . . , −0.5)> , Id ),
and
GY = N ((0.5, . . . , 0.5)> , Id ),
with = 0.05. We then change σ = 1, 40, 80, 120, 160, 200 and 240 to investigate the robustness
of the tests against heavy-tail contaminations.
9
Contamination (Location)
1.0
●
●
●
●
●
0.8
Contamination (Scale)
1.0
●
CvM
Energy
●
●
●
●
0.6
0.4
0.2
●
●
●
0.8
Power
Power
●
●
●
CvM
Energy
0.6
0.4
0.2
0.0
0.0
0
50
100
150
200
0
50
100
150
200
Figure 1: Empirical power of φCvM (solid blue) and φEnergy (dashed red) under the contamination models with = 0.05. See Example 3.1 and 3.2 for details.
Example 3.2 (Scale difference). Similar to the location alternative, we again choose multivariate normal distributions which differ in their scale but not in their location parameters.
In detail, we have
FX = N ((0, . . . , 0)> , 0.12 × Id ),
and
GY = N ((0, . . . , 0)> , Id ),
with = 0.05. Again, we change σ = 1, 40, 80, 120, 160, 200 and 240 to assess the effect of
heavy-tail contaminations.
Experiments were run with m = n = 40 and d = 10, and the critical values were estimated
via Monte Carlo simulations at significance level α = 0.05. As can be seen from the result in
Figure 1, the power of the CvM test is consistently robust to the change of σ, while the power
of the energy test drops down significantly as σ increases. Roughly speaking, when σ is large,
it is likely that the energy statistic is dominated by extreme observations from Hm and Hn0 .
Furthermore, when Hm and Hn0 are identical (i.e. they are in favor of the null hypothesis), the
null and alternative distributions of the energy statistic tend to be close as σ increases. As a
consequence, the resulting test based on the energy statistic performs poorly. In contrast, the
CvM statistic puts less weight on extreme observations, which leads to robust performance of
the CvM test under contaminations. A mathematical justification can be found in the proof
of Theorem 3.1.
4
Minimax Optimality
In this section, we investigate the minimax optimality of the proposed tests based on the U and V -statistics. To formulate the minimax problem, we define the set of two multivariate
distributions which are at least far apart in terms of the Cramér-von Mises distance, i.e.
n
o
F() := (FX , GY ) : Wd (FX , GY ) ≥ .
10
For a given significance level α ∈ (0, 1), let Φm,n (α) be the set of measurable functions
φ : {Xm , Yn } 7→ {0, 1} such that
Φm,n (α) = {φ : P0 (φ = 0) ≤ α}.
We then define the minimax type II error as follows:
1 − βm,n () =
inf
sup
φ∈Φm,n (α) FX ,GY ∈F ()
Pm,n
1 (φ = 0).
(13)
Our primary interest is in finding the minimum separation rate m,n satisfying
n
o
m,n = inf : 1 − βm,n () ≤ ξ ,
for some 0 < ξ < 1 − α. Throughout this section, we assume that there exists `, u ∈ (0, 1)
such that ` ≤ m/N, n/N ≤ u for every m, n with N = m + n.
We start with presenting a lower bound of the multivariate CvM distance.
Lemma 4.1. When X, Y have continuous distributions, the multivariate CvM distance is
lower bounded by
Z
1
(14)
− PX,Y β > X ≤ β > Y dλ(β).
Wd (FX , GY ) ≥
Sd−1 2
Consider two multivariate normal distributions FX = N (−δN , Id ) and GY = N (δN , Id )
where δN,i = N −1/2 for i = 1, . . . , d. Based on Lemma 4.1, we show that the CvM distance
between FX and GY are at least N −1/2 far apart up to some scaling factor. In the end,
this result together with Neyman-Pearson lemma establishes a lower bound for the minimum
separation rate in Theorem 4.1.
Lemma 4.2. Let X ∼ N (−δN , Id ) and Y ∼ N (δN , Id ) where δN,i = N −1/2 for i = 1, . . . , d.
Then
Wd (FX , GY ) & N −1/2 .
Proof. Note that for any given direction β ∈ Sd−1 , β > (X − Y ) has an normal distribution:
β > (X − Y ) ∼ N (2β > δN , 2),
and consequently,
1
− PX,Y (β > X < β > Y ) =
2
√
1
− Φ − 2β > δN
2
1
>
2
≥ |β > δN | √ e−(β δN )
π
1
> 2
≥ N −1/2 √ |β > 1|e−(β 1) .
π
By combining (14) with (15), we conclude that
Z
1
Wd (FX , GY ) ≥
− PX,Y β > X ≤ β > Y dλ(β)
Sd−1 2
Z
> 2
−1/2 1
√
≥ N
|β > 1|e−(β 1) dλ(β) & N −1/2 ,
π Sd−1
which completes the proof.
11
(15)
Theorem 4.1 (Lower Bound). For 0 < ξ < 1 − α, there exists some constant b = b(α, ξ)
such that m,n = bN −1/2 and the minimax type II error is lower bounded by ξ, i.e.
1 − βm,n (m,n ) ≥ ξ.
Theorem 4.1 shows that no test can have considerable power against all alternatives when
m,n is of order N −1/2 ; thereby, it provides a lower bound of the minimum separation rate.
Now, we show that the lower bound is asymptotically tight by establishing matching upper
bounds. In particular, the upper bounds are obtained based on the U - and V -statistics.
Theorem 4.2 (Upper Bound). Consider the U- and V- statistics, given in (7). Define the
tests
φU := I(Um,n ≥ cuα,m,n )
and
φV := I(Vm,n ≥ cvα,m,n )
where cuα,m,n and cvα,m,n are the corresponding α level critical values. Let ?m,n be the radius of
interest defined by
bN
?m,n := √ .
N
Then the type II error of φU and φV tend to zero uniformly over F(?m,n ) when bN → ∞ as
m, n → ∞. In other words,
lim
sup
bN →∞ FX ,GY ∈F (?
P1 (φU = 0) = lim
sup
bN →∞ FX ,GY ∈F (?
m,n )
P1 (φV = 0) = 0
m,n )
as m, n → ∞.
Remark 4.1. The minimax upper bound can also be obtained via the projection pursuit-type
Kolmogorov-Smirnov test whose statistic is defined as
dK,m,n :=
sup
(β,t)∈Sd−1 ×R
b β > (t) .
Fbβ > X (t) − G
Details can be found in in the supplementary material. A major problem with dK,m,n , however,
is that the exact calculation of the test statistic is nontrivial in practice, especially when d is
large. On the other hand, the multivariate Cramér-von Mises statistics are computationally
straightforward in arbitrary dimensions due to Theorem 2.1.
It is worth noting that there are computationally more efficient ways of estimating Wd2 .
For example, when X and Y are continuous, one can use the linear-type statistic defined as
Lm,n =
M
1 X1
[h∗ (X2i−1 , X2i ; Y2i−1 , Y2i ) + h∗ (X2i , X2i−1 ; Y2i−1 , Y2i )] ,
M
2
(16)
i=1
where M = bn/2c and m = n for simplicity. While Lm,n is also an unbiased estimator of Wd2
and can be computed in linear time, it is notably sub-optimal in terms of minimax power.
In particular, the following lemma shows that Lm,n can only be powerful against alternatives
slower than N −1/4 , whereas the minimax optimal rate is N −1/2 . This illustrates a certain
trade-off between computational cost and statistical power.
12
m = n = 20, d = 2
1.0
●
●
●
●
●
m = n = 20, d = 30
1.0
●
●
●
●
●
●
●
●
●
●
0.8
0.6
●
Power
Power
0.8
0.4
0.6
0.4
●
●
0.2
0.2
●
●
0.0
0.2
0.4
0.6
0.8
U−Stat
Linear
Hotelling
●
●
0.0
1.0
0.1
0.2
µ
0.3
0.4
U−Stat
Linear
Hotelling
0.5
µ
Figure 2: Empirical power of the multivariate CvM tests based on the U -statistic (solid
blue), the linear-type statistic (dashed red) and Hotelling’s test (dotted gray) where X ∼
N ((−µ, . . . , −µ), Id ) and Y ∼ N ((µ, . . . , µ), Id ).
Lemma 4.3. Consider a sequence of alternatives such that
Wd (FX , GY ) N −ε .
Then no test based on Lm,n has power converging to one as m, n → ∞ when ε > 1/4.
In addition to the rate optimality, the proposed tests have comparable power to existing
optimal tests against certain directions. To illustrate this point, we consider an example of
multivariate normal distributions where X ∼ N (µX , Id ) and Y ∼ N (µY , Id ). Under this
particular setting, it is well-known that Hotelling’s test is asymptotically the most powerful
test when the dimension is fixed. On the other hand, when the dimension is close to the
sample size, Hotelling’s test performs poorly due to a high variance from the sample precision
matrix (Bai and Saranadasa, 1996). We carried out simulation studies with m = n = 20 to
compare the power of the multivariate CvM test with Hotelling’s T 2 . For our simulation, we
set µX = (−µ, . . . , −µ)> and µY = (µ, . . . , µ)> at significance level α = 0.05. According to
the results presented in Figure 2, when the dimension is small (d = 2), the CvM test based
on Um,n (denoted by φU ) shows comparable (in fact, almost identical) power to Hotelling’s
test. However, it is apparent to see that φU outperforms Hotelling’s test when the dimension
(d = 30) is close to the sample size. Indeed, under the given setting, φU has the same
limiting power as the high-dimensional mean tests introduced by Chen and Qin (2010) and
Chakraborty and Chaudhuri (2017). This is our next topic in Section 5. In general, the test
based on the linear statistic shows substantially lower power than φU as we expected.
5
Adaptivity of Power in High-Dimensions
Up to now, we have assumed that the dimension d is an arbitrary but fixed constant. In
this section, we study the power of the multivariate CvM test when the dimension tends to
infinity. In particular, we will focus on a simple location shift model:
X = µX + W
and Y = µY + W 0 ,
13
(17)
>
d
where
8 W = (W10, . . . , Wd ) ∈ R and Wi ’s are i.i.d. having a continuous distribution with
E W1 < ∞. W is independent and identically distributed with W. In addition, we consider
µX = (µX , . . . , µX )> ∈ Rd and µY = (µY , . . . , µY )> ∈ Rd such that
d1/4− (µX − µY ) → 0
for any > 0 as d → ∞.
(18)
There are many types of high-dimensional inference procedures for testing
H0 : µX = µY
versus
H1 : µX 6= µY .
(19)
For example, Chen and Qin (2010) suggest the test statistic based on the sample means.
Specifically, their statistic is based on the U -statistic defined by
TCQ
m
X
1
=
(m)2 (n)2
n
X
(Xi1 − Yj1 )> (Xi2 − Yj2 ).
i1 ,i2 =1 j1 ,j2 =1
i1 6=i2 j1 6=j2
Recently, Chakraborty and Chaudhuri (2017) define their test statistic based on spatial ranks
by
TWMW =
m
X
1
(m)2 (n)2
n
X
i1 ,i2 =1 j1 ,j2 =1
i1 6=i2 j1 6=j2
(Xi1 − Yj1 ) > (Xi2 − Yj2 )
.
||Xi1 − Yj1 || ||Xi2 − Yj2 ||
Somewhat surprisingly, under this given scenario, the limiting power of the tests based on
Um,n , Em,n , TCQ and TWMW tend to be the same when d → ∞ and then m, n → ∞. (In fact,
the latter condition is unnecessary from our proof). This finding is striking because the tests
based on TCQ and TWMW (CQ test and WMW test for short, respectively) are specifically
designed for high-dimensional mean inference, while the proposed test as well as energy test
can be sensitive to much broader alternatives. In other words, the CvM test and the energy
test are adaptive to high-dimensional location alternatives. In particular, we present the
following result.
Theorem 5.1. Suppose that X and Y satisfy the model assumption in (17) with
µ X − µY =
d1/4
ψ
√
m+n
and
V [X] = V [Y ] = σ 2 I.
In addition, we assume that m/N → π1 ∈ (0, 1) and n/N → π2 ∈ (0, 1) as m, n → ∞. Then
the power of the tests based on Um,n , Em,n , TCQ and TWMW are asymptotically the same if
we let d → ∞ and then m, n → ∞. In other words, let βCvM , βEnergy , βCQ and βWMW be the
power functions of these tests at α level. Then
lim lim βCvM = lim lim βEnergy = lim lim βCQ = lim lim βWMW ,
N →∞ d→∞
N →∞ d→∞
N →∞ d→∞
and they converge to
π1 π2 2
Φ −zα + √
ψ .
2σ 2
14
N →∞ d→∞
We would like to point out that the restricted model assumption in (17) is mainly due
to technical convenience. We believe that the main result in Theorem 5.1 is still valid under
much weaker assumptions supported by our simulations in Section 8 and the supplementary
material. Nevertheless, when the moment assumption is violated, the power of these tests can
be entirely different. For instance, our simulations in Section 8 further demonstrate that the
CQ and (especially) energy tests perform poorly when X and Y have Cauchy distributions
with different location parameters. In contrast, the CvM and WMW tests maintain robust
power against the same Cauchy alternative.
6
Connection to the Generalized Energy Distance and MMD
Recall that, under the finite first moment condition, the energy distance is defined with the
Euclidean distance. By considering a semimetric space (Z, ρ) of negative type, Sejdinovic
et al. (2013) generalize the energy distance by
Eρ2 = 2EX,Y [ρ(X, Y )] − EX,X 0 ρ(X, X 0 ) − EY,Y 0 ρ(Y, Y 0 ) .
They further establish the equivalence between the generalized energy distance and the MMD
with a kernel induced by ρ(·, ·). Given a distance-induced kernel k(·, ·), the squared MMD is
given by
MMD2k = EX,X 0 k(X, X 0 ) + EY,Y 0 k(Y, Y 0 ) − 2EX,Y [k(X, Y )] .
In this section, we will show that the multivariate CvM distance is a member of the
generalized energy distance by the use of the angular distance and thus also a member of
the MMD. To begin, let SX and SY be the support of X and Y , respectively and let S =
SX ∪ SY ⊆ Rd . Then we define the angular distance as follows:
Definition 6.1. Let W∗ be a random vector having mixture distribution (1/2)FX + (1/2)GY .
For z, z 0 ∈ S, denote the scaled angle between z − W∗ and z 0 − W∗ by
ρAngle (z, z 0 ; W∗ ) =
1
Ψ z − W∗ , z 0 − W∗ .
π
The angular distance is defined as the expected value of the scaled angle by
ρAngle (z, z 0 ) = EW∗ ρAngle (z, z 0 ; W∗ ) .
(20)
The next lemma shows that ρAngle is a metric of negative type defined on S.
Lemma 6.1. For ∀z, z 0 , z 00 ∈ S and ρAngle : S × S 7→ [0, ∞), the following conditions are
satisfied:
1. ρAngle (z, z 0 ) ≥ 0 and ρAngle (z, z 0 ) = 0 if and only if z = z 0 .
2. ρAngle (z, z 0 ) = ρAngle (z 0 , z).
3. ρAngle (z, z 0 ) ≤ ρAngle (z, z 00 ) + ρAngle (z 0 , z 00 ).
In addition, for ∀n ≥ 2, z1 , . . . , zn ∈ S, and α1 , . . . , αn ∈ R, with
n X
n
X
αi αj ρAngle (zi , zj ) ≤ 0.
i=1 j=1
15
Pn
i=1 αi
= 0,
By the use of the angular distance, we establish the identity between the generalized
energy distance and the CvM distance in the next lemma. As a result, we conclude that the
multivariate CvM distance is a special case of the generalized energy distance based on the
angular distance.
Lemma 6.2. Let us consider the angular distance defined in (20). Then the following identity
holds:
Wd2 = 2EX,Y [ρAngle (X, Y )] − EX,X 0 ρAngle (X, X 0 ) − EY,Y 0 ρAngle (Y, Y 0 ) .
Proof. The proof is based on the expression of Wd2 in Theorem 2.1 and Fubini’s theorem;
details are omitted.
Remark 6.1. The angular distance can be generalized by taking the expectation with respect
to a different measure. For instance, when the expectation is taken with respect to the Lebesgue
measure ν, the generalized angular distance is proportional to the Euclidean distance, i.e.
Eν ρAngle (z, z 0 ; ν) = γd ||z − z 0 ||,
where γd solely depends on the dimension (see the proof of Lemma 6.1 for more details).
The main difference between the Euclidean distance and the proposed distance is that the
latter takes into account information from the underlying distribution and is less sensitive to
outliers. In this aspect, the introduced angular distance can be viewed as a robust alternative
for the Euclidean distance.
7
Other Multivariate Extensions via Projection Pursuit
The projection pursuit approach used for the multivariate CvM statistic can be applied to
other univariate robust statistics. In this section, we illustrate the utility of the projection
pursuit approach by considering several examples including Kendall’s tau, the coefficient
by Blum et al. (1961) and the sign covariance (Bergsma and Dassios, 2014). We begin by
considering one-sample and two-sample robust statistics. Given a pair of random variables
(X, Y ), define Z = X − Y . The univariate sign test statistic is an estimate of θsign := P(Z >
0) − 1/2, and it is used to test whether
H0 : P(Z > 0) = 1/2
versus
H1 : P(Z > 0) 6= 1/2.
The projection pursuit technique extends θsign to a multivariate case as follows:
Proposition 7.1 (One-sample sign test statistic). For i.i.d. random vectors Z1 , Z2 ∈ Rd , the
projection pursuit approach generalizes θsign as
Z
Sd−1
1
P(β Z1 > 0) −
2
>
2
dλ(β) =
1
1
−
E [Ψ (Z1 , Z2 )] .
4 2π
Proof. Given β ∈ Rd−1 , note that
1
P(β Z1 > 0) −
2
>
2
=
h
i
h
i
1
− E I(β > Z1 > 0) + E I(β > Z1 > 0, β > Z2 > 0) .
4
16
(21)
Apply Lemma 2.2 with Fubini’s theorem to have
Z
1
>
E
I(β Z1 > 0)dλ(β) =
2
Sd−1
Z
1
1
>
>
E [Ψ (Z1 , Z2 )] .
E
I(β Z1 > 0, β Z2 > 0)dλ(β) = −
2
2π
d−1
S
This completes the proof.
Given independent two samples Xm = {X1 , . . . , Xm } and Yn = {Y1 , . . . , Yn }, WilcoxonMann-Whitney test is aim to distinguish
H0 : P(X > Y ) =
1
2
versus
1
H1 : P(X > Y ) 6= .
2
Its test statistic is based on an estimate of θWMW := P(X > Y ) − 1/2. The next proposition
extends θWMW to a multivariate case via projection pursuit.
i.i.d.
Proposition 7.2 (Two-sample Wilcoxon-Mann-Whitney test statistic). Let X1 , X2 ∼ FX
i.i.d.
and Y1 , Y2 ∼ GY . The projection pursuit approach generalizes θWMW as
Z
1 2
1
1
>
>
P(β X > β Y ) −
dλ(β) = −
E [Ψ (X1 − Y1 , X2 − Y2 )] .
2
4 2π
Sd−1
(22)
Proof. The result follows by replacing Z1 , Z2 with X1 − Y1 , X2 − Y2 in Proposition 7.1.
Remark 7.1. The first order Taylor approximation of the arccosine function shows that the
representations given in the right-side of (21) and (22) are related to the spatial sign-statistics
introduced by Wang et al. (2015) and Chakraborty and Chaudhuri (2017), respectively. We
believe, however, that our projection pursuit-type statistics — which can be viewed as the
average of univariate statistics based on projected random variables — is more intuitive to
understand.
The same technique can be further applied to some robust statistics for independence
testing. To test for independence between two random variables, Kendall’s tau statistic is
defined as an estimate of τ = 4P (X1 < X2 , Y1 < Y2 ) − 1. We present a multivariate extension
of τ as follows:
Theorem 7.1 (Kendall’s tau). For given i.i.d. pairs of random vectors (X1 , Y1 ), . . . , (X4 , Y4 ) ∈
Rp×q , the multivariate extension of τ using projection pursuit is given by
Z
Z
h
i2
4P α> (X1 − X2 ) < 0, β > (Y1 − Y2 ) < 0 − 1 dλ(α)dλ(β)
Sp−1
Sq−1
2
2
= E 2 − Ψ (X1 − X2 , X3 − X4 )
2 − Ψ (Y1 − Y2 , Y3 − Y4 ) − 1.
π
π
Kendall’s tau has been frequently used in practice due to its robustness, simplicity and
interpretability. Nonetheless, the main limitation of Kendall’s tau is that it can be zero even
when there exists a certain association between random variables. There have been alternative
approaches to resolve this issue in the literature. For a multivariate case, Zhu et al. (2017)
17
extend Hoeffding’s coefficient (Hoeffding, 1948) via projection pursuit. Specifically, they define
the projection correlation between X ∈ Rp and Y ∈ Rq as
Z
Z
Z
2
Fα> X,β > Y (u, v) − Fα> X (u)Fβ > Y (v) dω1 (u, v, α, β),
(23)
Sp−1
Sq−1
R2
where dω1 (u, v, α, β) = dFα> X,β > Y (u, v)dλ(α)dλ(β). Although the projection correlation is
more broadly sensitive than Kendall’s tau is in detecting dependence among random variables,
it can still be zero even when X and Y are dependent. A counterexample for the univariate
case can be found in Hoeffding (1948).
On the other hand, the coefficient introduced by Blum et al. (1961) overcomes this issue
by replacing dFX,Y with dFX dFY . The univariate Blum-Kiefer-Rosenblatt (BKR) coefficient
(Blum et al., 1961) is defined by
Z
[FXY (u, v) − FX (u)FY (v)]2 dFX (u)dFY (v).
R2
The BKR coefficient can also be generalized in a multivariate space via projection pursuit.
Theorem 7.2 (Blum–Kiefer–Rosenblatt (BKR) coefficient). Let us consider weight function
dω2 (u, v, α, β) = dFα> X (u)dFβ > Y (v)dλ(α)dλ(β). For i.i.d. (X1 , Y1 ), . . . , (X6 , Y6 ), the univariate BKR coefficient can be extended to a multivariate case by
Z
Z
Z
2
Fα> X,β > Y (u, v) − Fα> X (u)Fβ > Y (v) dω2 (u, v, α, β)
Sp−1
Sq−1
R2
1
1
= E
−
Ψ (Y1 − Y4 , Y2 − Y4 )
2 2π
1
1
1
1
−
Ψ (X1 − X5 , X2 − X5 )
−
Ψ (Y3 − Y6 , Y4 − Y6 )
+E
2 2π
2 2π
1
1
1
1
−
Ψ (X1 − X4 , X2 − X4 )
−
Ψ (Y1 − Y5 , Y3 − Y5 ) .
−2E
2 2π
2 2π
1
1
−
Ψ (X1 − X3 , X2 − X3 )
2 2π
Recently, Bergsma and Dassios (2014) introduce a modification of Kendall’s tau, which
is zero if and only if random variables are independent. Let τ ∗ denote Bergsma-Dassios sign
covariance. The univariate τ ∗ is given by
τ ∗ = E [asign (X1 , X2 , X3 , X4 )asign (Y1 , Y2 , Y3 , Y4 )] ,
(24)
with asign (z1 , z2 , z3 , z4 ) = sign (|z1 − z2 | + |z3 − z4 | − |z1 − z3 | − |z2 − z4 |). Motivated by the
projection pursuit approach, we propose the multivariate τ ∗ as follows:
Definition 7.1 (Multivariate τ ∗ ). Suppose (X1 , Y1 ), . . . , (X4 , Y4 ) are i.i.d. random vectors
from a joint distribution FXY where X ∈ Rp and Y ∈ Rq . We define the multivariate τ ∗ by
Z
Z
∗
τp,q
=
E asign (α> X1 , α> X2 , α> X3 , α> X4 )
Sp−1
Sq−1
×asign (β > Y1 , β > Y2 , β > Y3 , β > Y4 ) dλ(α)dλ(β).
18
Since the kernel of τ ∗ is sign-invariant, i.e.
asign (z1 , z2 , z3 , z4 ) = asign (−z1 , −z2 , −z3 , −z4 ),
∗ becomes the univariate τ ∗ when p = q = 1. Also, note that since
it is easy to see that τp,q
X and Y are independent if and only if α> X and β > Y are independent for all α ∈ Sp−1 and
∗ follows by that of the univariate τ ∗ .
β > ∈ Sq−1 , the characteristic property of τp,q
∗
To have an expression for τp,q without requiring integration over the unit sphere, we first
generalize Lemma 2.2 with three indicator functions presented in Lemma 7.1. Then based on
∗ in Theorem 7.3.
this result, we provide an alternative expression for τp,q
Lemma 7.1. For arbitrary vectors U1 , U2 , U3 ∈ Rd , the following identity holds
Z
I(β > U1 ≤ 0)I(β > U2 ≤ 0)I(β > U3 ≤ 0)dλ(β)
Sd−1
=
1
1
−
[Ψ (U1 , U2 ) + Ψ (U1 , U3 ) + Ψ (U2 , U3 )] .
2 4π
For U1 , U2 , U3 ∈ Rd , define gd (U1 , U2 , U3 ) and hd (Z1 , Z2 , Z3 , Z4 ) by
1
1
[Ψ (U1 , U2 ) + Ψ (U1 , U3 ) + Ψ (U2 , U3 )]
gd (U1 , U2 , U3 ) = −
2 4π
and
hd (Z1 , Z2 , Z3 , Z4 )
= gd (Z1 − Z2 , Z2 − Z3 , Z3 − Z4 ) + gd (Z2 − Z1 , Z1 − Z3 , Z3 − Z4 )
+ gd (Z1 − Z2 , Z2 − Z4 , Z4 − Z3 ) + gd (Z2 − Z1 , Z1 − Z4 , Z4 − Z3 ).
∗ as follows:
Based on the kernel hd , we present an alternative expression for τp,q
Theorem 7.3. For i.i.d. random vectors (X1 , Y1 ), . . . , (X4 , Y4 ) from a joint distribution FXY
∗ can be written as
where X ∈ Rp and Y ∈ Rq , τp,q
∗
τp,q
= E [hp (X1 , X2 , X3 , X4 )hq (Y1 , Y2 , Y3 , Y4 )]
+E [hp (X1 , X2 , X3 , X4 )hq (Y3 , Y4 , Y1 , Y2 )]
−2E [hp (X1 , X2 , X3 , X4 )hq (Y1 , Y3 , Y2 , Y4 )] .
∗ based on a U - or V Theorem 7.3 leads to a straightforward empirical estimate of τp,q
statistic. This is also true for the other multivariate generalizations introduced in this section.
Using these estimates, some theoretical and empirical properties of the proposed measures can
be further investigated. These topics are reserved for future work.
Remark 7.2. In order to estimate the projection correlation and the multivariate BKR coefficient, one needs at least five and six independent pairs of samples, respectively. On the
∗ requires four independent pairs of samples; thereby τ ∗ has a
other hand, an estimate of τp,q
p,q
computational advantage over the other robust measure of independence.
Remark 7.3. Dhar et al. (2016) highlight the robustness of the univariate τ ∗ and contrasted it
to the non-robustness of the distance covariance (Székely et al., 2007). One can similarly show
∗ also retains the robustness in terms of the maximum bias functional
that the multivariate τp,q
considered in Dhar et al. (2016).
19
8
Simulations
In this section, we present some simulation examples to support the argument in Section 5
as well as to illustrate the finite sample performance of the proposed tests. Throughout
our experiments, the significance level was set at 0.05 and the permutation procedure was
used to set up the critical value. Since the computational costs of U -statistics together with
permutation tests are expensive for large n and d, we set the sample size and the dimension
relatively small by m = n = 20 and d = 200, but which are sufficient to emphasize our claim.
First, we consider several examples where the power of the considered tests in Section 5 are
approximately equivalent to each other. Here we use multivariate normal distributions with
different means
µ(0) = (0, . . . , 0)> , µ(1) = (0.1, . . . , 0.1)>
√
µ(2) = 2(0.1, . . . , 0.1, 0, . . . , 0 )>
| {z } | {z }
and
d/2 elements d/2 elements
and covariance matrices:
1. Identity matrix (denoted by I) where σi,i = 1 and σi,j = 0 for i 6= j.
2. Banded matrix (denoted by ΣBand ) where σi,i = 1, σi,j = 0.6 for |i − j| = 1, σi,j = 0.3
for |i − j| = 2 and σi,j = 0 otherwise.
3. Autocorrelation matrix (denoted by ΣAuto ) where σi,i = 1 and σi,j = 0.2|i−j| when i 6= j.
4. Block diagonal matrix (denoted by ΣBlock ) where the 5 × 5 main diagonal blocks A are
defined by ai,i = 1 and ai,j = 0.2 when i 6= j, and the off-diagonal blocks are zeros.
Then we generate random samples from X ∼ N (µ(0) , Σ) and either Y ∼ N (µ(1) , Σ) or Y ∼
N (µ(2) , Σ). The results are summarized in Table 2. As we expected, the empirical power of
the considered tests are close when the moment assumptions are met and the covariates are
independent or weakly dependent. More empirical evidence can be found in the supplementary
material.
Table 2: Empirical power of the considered tests
α = 0.05.
I
ΣBand
methods µ(1)
µ(2)
µ(1)
µ(2)
CvM
0.251 0.271 0.266 0.191
Energy 0.246 0.261 0.261 0.191
CQ
0.256 0.251 0.251 0.196
WMW 0.241 0.266 0.261 0.196
against the normal location models at
ΣAuto
µ(1)
µ(2)
0.246 0.241
0.246 0.241
0.256 0.241
0.246 0.236
ΣBlock
µ(1)
µ(2)
0.257 0.251
0.257 0.251
0.276 0.256
0.267 0.261
In our second experiment, we consider several examples where the moment conditions
are not satisfied. Simply, we focus on random samples generated from multivariate Cauchy
distributions. Let Cauchy(γ, s) denote the univariate Cauchy distribution where γ, s are the
location parameter and the scale parameter, respectively. Let X = (X (1) , . . . , X (d) ) and Y =
i.i.d.
i.i.d.
(Y (1) , . . . , Y (d) ) be random vectors where X (i) ∼ Cauchy(0, 1) and Y (i) ∼ Cauchy(γ, s)
for i = 1, . . . , d. We first consider location differences where γ is not zero but the scale
parameters are identical, i.e. s = 1. Similarly, we consider scale differences where the scale
parameter s changes, but the location parameters are identical, i.e. γ = 0. We also report
20
Table 3: Empirical power of the considered tests against multivariate Cauchy distributions at
α = 0.05 where γ, s represent the location and scale parameter, respectively.
Location
Scale
γ
2
3
4
5
0
0
0
0
s
1
1
1
1
2
3
4
5
CvM
0.146
0.231
0.593
0.874
0.628
0.940
0.990
1.000
Energy
0.075
0.065
0.136
0.131
0.342
0.678
0.844
0.859
CQ
0.136
0.246
0.322
0.442
0.070
0.060
0.055
0.090
WMW
0.296
0.643
0.915
0.985
0.090
0.060
0.055
0.106
MMD
0.101
0.101
0.161
0.171
0.492
0.794
0.945
0.970
NN
0.281
0.683
0.869
0.970
0.221
0.211
0.196
0.241
the power results based on the Gaussian MMD with the median heuristic (Gretton et al.,
2012) and k-NN with k = 3 (Schilling, 1986) as reference points. These tests are denoted by
MMD and NN, respectively. From the results presented in Table 3, we can see that, unlike
the multivariate normal cases, there are significant differences between power performance
among CvM, Energy, CQ and WMW tests. In particular, the tests based on the Energy and
CQ statistics have relatively low power against the heavy-tail location alternatives, whereas
the tests based on the CvM and WMW statistics show better performance than the others.
Turning to the scale problems, it can be seen that the CQ and WMW tests have no power to
detect scale differences, which makes sense because they are specifically designed for location
problems. On the other hand, the CvM and Energy tests are reasonably sensitive to these
alternatives. Among the omnibus tests, the MMD test shows a good performance against the
scale differences, but not against the location differences. The NN test outperforms the other
omnibus tests against the location alternatives; however, it has low power against the scale
alternatives. In summary, the overall ranks among the omnibus tests against the Cauchy
alternatives are
1. Cauchy location: NN CvM MMD Energy.
2. Cauchy scale: CvM MMD Energy NN.
9
Concluding Remarks
In this work, we extended the univariate Cramér-von Mises statistic for two-sample testing to
the multivariate case using projection pursuit. The proposed statistics have straightforward
calculation formulas in arbitrary dimensions and the resulting tests have good properties.
Throughout this paper, we demonstrated their robustness, minimax optimality and highdimensional power properties. In addition, we applied the same projection technique to other
robust statistics and presented their multivariate extensions.
Beyond nonparametric testing problems, we believe that our approach can be used for
other problems. For example, our work can be viewed as an application of the angular distance to the two-sample problem. The angular distance is closely connected to the Euclidean
distance (Remark 6.1) but is more robust to outliers by incorporating information from the
underlying distribution. Given that the use of distances is of fundamental importance in many
statistical applications (including clustering, classification and regression), we expect that the
21
angular distance can be applied to other statistical problems as a robust alternative for the
Euclidean distance.
References
Anderson, N. H., Hall, P., and Titterington, D. M. (1994). Two-sample test statistics for measuring discrepancies between two multivariate probability density functions using kernelbased density estimates. Journal of Multivariate Analysis, 50(1):41–54.
Anderson, T. W. (1962). On the distribution of the two-sample cramer-von mises criterion.
The Annals of Mathematical Statistics, pages 1148–1159.
Bai, Z. and Saranadasa, H. (1996). Effect of high dimension: by an example of a two sample
problem. Statistica Sinica, pages 311–329.
Baraud, Y. (2002). Non-asymptotic minimax rates of testing in signal detection. Bernoulli,
8(5):577–606.
Baringhaus, L. and Franz, C. (2004). On a new multivariate two-sample test. Journal of
Multivariate Analysis, 88(1):190–206.
Baringhaus, L. and Henze, N. (2017). Cramér–von mises distance: probabilistic interpretation,
confidence intervals, and neighbourhood-of-model validation. Journal of Nonparametric
Statistics, 29(2):167–188.
Bergsma, W. and Dassios, A. (2014). A consistent test of independence based on a sign
covariance related to kendall’s tau. Bernoulli, 20(2):1006–1028.
Bhat, B. V. (1995). Theory of U-statistics and its applications. PhD thesis, Karnatak University.
Blum, J. R., Kiefer, J., and Rosenblatt, M. (1961). Distribution free tests of independence
based on the sample distribution function. The Annals of Mathematical Statistics, pages
485–498.
Bogomolny, E., Bohigas, O., and Schmit, C. (2007). Distance matrices and isometric embeddings. arXiv preprint arXiv:0710.2063.
Chakraborty, A. and Chaudhuri, P. (2017). Tests for high-dimensional data based on means,
spatial signs and spatial ranks. The Annals of Statistics, 45(2):771–799.
Chen, H. and Friedman, J. H. (2017). A new graph-based two-sample test for multivariate
and object data. Journal of the American Statistical Association, 112(517):397–409.
Chen, S. X. and Qin, Y.-L. (2010). A two-sample test for high-dimensional data with applications to gene-set testing. The Annals of Statistics, pages 808–835.
Chung, E. and Romano, J. P. (2016). Asymptotically valid and exact permutation tests based
on two-sample u-statistics. Journal of Statistical Planning and Inference, 168:97–105.
Cui, H. (2002). Average projection type weighted cramér-von mises statistics for testing some
distributions. Science in China Series A: Mathematics, 45(5):562–577.
22
Dhar, S. S., Dassios, A., and Bergsma, W. (2016). A study of the power and robustness of a
new test for independence against contiguous alternatives. Electronic Journal of Statistics,
10(1):330–351.
Escanciano, J. C. (2006). A consistent diagnostic test for regression models using projections.
Econometric Theory, 22(6):1030–1051.
Friedman, J. H. and Rafsky, L. C. (1979). Multivariate generalizations of the wald-wolfowitz
and smirnov two-sample tests. The Annals of Statistics, pages 697–717.
Giné, E. and Nickl, R. (2009). An exponential inequality for the distribution function of the
kernel density estimator, with applications to adaptive estimation. Probability Theory and
Related Fields, 143(3-4):569–596.
Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., and Smola, A. (2012). A kernel
two-sample test. Journal of Machine Learning Research, 13(Mar):723–773.
Henze, N. (1988). A multivariate two-sample test based on the number of nearest neighbor
type coincidences. The Annals of Statistics, pages 772–783.
Hoeffding, W. (1948). A non-parametric test of independence. The Annals of Mathematical
Statistics, pages 546–557.
Hoeffding, W. (1952). The large-sample power of tests based on permutations of observations.
The Annals of Mathematical Statistics, pages 169–192.
Lee, J. (1990). U-statistics: Theory and practice.
Lehmann, E. L. (1951). Consistency and unbiasedness of certain nonparametric tests. The
Annals of Mathematical Statistics, pages 165–179.
Liu, R. Y. (2006). Data depth: robust multivariate analysis, computational geometry, and
applications, volume 72. American Mathematical Soc.
McDiarmid, C. (1989). On the method of bounded differences. Surveys in combinatorics,
141(1):148–188.
Mondal, P. K., Biswas, M., and Ghosh, A. K. (2015). On high dimensional two-sample tests
based on nearest neighbors. Journal of Multivariate Analysis, 141:168–178.
Oja, H. (2010). Multivariate nonparametric methods with R: an approach based on spatial
signs and ranks. Springer Science & Business Media.
Ramdas, A., Reddi, S. J., Poczos, B., Singh, A., and Wasserman, L. (2015). Adaptivity and
computation-statistics tradeoffs for kernel and distance based high dimensional two sample
testing. arXiv preprint arXiv:1508.00655.
Schilling, M. F. (1986). Multivariate two-sample tests based on nearest neighbors. Journal of
the American Statistical Association, 81(395):799–806.
Sejdinovic, D., Sriperumbudur, B., Gretton, A., and Fukumizu, K. (2013). Equivalence of
distance-based and RKHS-based statistics in hypothesis testing. The Annals of Statistics,
pages 2263–2291.
23
Serfling, R. J. (2009). Approximation theorems of mathematical statistics, volume 162. John
Wiley & Sons.
Székely, G. J. and Rizzo, M. L. (2004). Testing for equal distributions in high dimension.
InterStat, 5(16.10).
Székely, G. J. and Rizzo, M. L. (2013). Energy statistics: A class of statistics based on
distances. Journal of Statistical Planning and Inference, 143(8):1249–1272.
Székely, G. J., Rizzo, M. L., and Bakirov, N. K. (2007). Measuring and testing dependence
by correlation of distances. The Annals of Statistics, 35(6):2769–2794.
Thas, O. (2010). Comparing distributions. Springer.
Wang, L., Peng, B., and Li, R. (2015). A high-dimensional nonparametric multivariate test
for mean vector. Journal of the American Statistical Association, 110(512):1658–1669.
Zhou, W.-X., Zheng, C., and Zhang, Z. (2017). Two-sample smooth tests for the equality of
distributions. Bernoulli, 23(2):951–989.
Zhu, L., Xu, K., Li, R., and Zhong, W. (2017). Projection correlation between two random
vectors. Biometrika.
Zhu, L.-X., Fang, K.-T., and Bhatti, M. I. (1997). On estimated projection pursuit-type
cramer–von mises statistics. Journal of Multivariate Analysis, 63(1):1–14.
24
A
Permutation Tests
Permutation tests are among the most widely used conditional procedures of nonparametric
inference. An essential feature of the permutation tests is that they are guaranteed to obtain
an exact rejection probability whenever the exchangeability condition is satisfied under H0 .
However, except for the finite sample exactness, there has been little theory for permutation
tests. The goal here is to enhance our understanding of permutation tests for the two-sample
problem. Specifically, we establish fairly general conditions under which the permutation
distribution is asymptotically equivalent to the corresponding unconditional null distribution
based on a U -statistic. Thereby, under the given conditions, both the permutation and
unconditional tests have the same asymptotic power (see Corollary 2.1).
To distinguish between H0 and H1 in (1), consider a symmetric kernel h(X1 , . . . , Xr ; Y1 , . . . , Yr )
of degree (r, r) such that
E [h(X1 , . . . , Xr ; Y1 , . . . , Yr )] = θ,
E {h(X1 , . . . , Xr ; Y1 , . . . , Yr )}2 < ∞.
(25)
Let Um,n be the corresponding U -statistic:
Um,n =
X X
1
h(Xα1 , . . . , Xαr ; Yβ1 , . . . , Yβr ),
(m)r (n)r
α∈Am β∈Bn
where Am and Bn denote the set of all subsets that consist of r unique elements from
{1, . . . , m} and {1, . . . , n}, respectively.
Let hc,d (x1 , . . . , xc ; y1 , . . . , yd ) be the conditional expectation given by
E h(x1 , . . . , xc , Xc+1 , . . . , Xr ; y1 , . . . , yd , Yd+1 , . . . , Yr ) .
Further define
e
hc,d (x1 , . . . , xc ; y1 , . . . , yd ) := hc,d (x1 , . . . , xc ; y1 , . . . , yd ) − θ and
i
h
ξc,d = V [hc,d (X1 , . . . , Xc ; Y1 , . . . , Yd )] := E {e
hc,d (X1 , . . . , Xc ; Y1 , . . . , Yd )}2 ,
where ξ0,0 = 0. We say the kernel h is non-degenerate if both ξ0,1 and ξ1,0 are positive, and
degenerate if ξ0,1 = ξ1,0 = 0.
It is well-known that a U -statistic has different limiting behaviors depending on the existence of degeneracy (see e.g. Lee, 1990). When the kernel is non-degenerate, Chung and
Romano (2016) provide a sufficient condition under which the permutation distribution approximates the unconditional distribution of Um,n . Their condition is basically equivalent
to
h1,0 (z) = −h0,1 (z),
(26)
so that the U -statistic is approximated by a linear-statistic:
m
n
X
X
√
√
r
r
N (Um,n − E [Um,n ]) = N
h1,0 (Xi ) −
h1,0 (Yj ) + oP (1).
m
n
i=1
25
j=1
Then the desired result follows by Theorem 3.2 of Hoeffding (1952) and the central limit
theorem. The goal of this section is to develop a similar result for a degenerate U -statistic
including the proposed CvM statistic as an example.
To begin, consider the centered U -statistic multiplied by N = m + n:
em,n (X1 , . . . , Xm ; Y1 , . . . , Yn ) = N (Um,n − E0 [Um,n ]).
U
Let {Z1 , . . . , Zm+n } = {X1 , . . . , Xm ; Y1 , . . . , Yn }. Then denote the permutation distribution
em,n given by
of U
R̂m,n (t) =
1 X e
I Um,n (Zπ1 , . . . , ZπN ) ≤ t ,
N!
π∈GN
where GN is the set of all permutations among {1, . . . , N }. Also, let R(t) be the unconditional
em,n . Then we present the following Theorem.
limiting distribution of U
Theorem A.1. Suppose the kernel h is degenerate under H0 and satisfies
Condition 1. h0,2 (z1 , z2 ) = h2,0 (z1 , z2 ) and h1,1 (z1 , z2 ) =
1−r
r h0,2 (z1 , z2 )
Condition 2. ξ0,1 = ξ1,0 = 0 and ξ0,2 , ξ2,0 , ξ1,1 > 0.
Condition 3. m/N → π1 and n/N → π2 as N → ∞.
Then under the moment condition in (25),
p
sup R̂m,n (t) − R(t) −→ 0.
t∈R
Proof. The proof can be found in the supplementary material.
Example A.1 (The kernel MMD in Gretton et al. (2012)). Given a symmetric kernel k(·, ·),
consider h(x1 , x2 ; y1 , y2 ) defined by
h(x1 , x2 ; y1 , y2 ) := k(x1 , x2 ) + k(y1 , y2 )
o
1n
−
k(x1 , y1 ) + k(x1 , y2 ) + k(x2 , y1 ) + k(x2 , y2 ) .
2
Then the square of the MMD is
MMD2k (FX , GY ) = EX,X 0 ,Y,Y 0 h(X, X 0 ; Y, Y 0 ) ,
i.i.d.
i.i.d.
where X, X 0 ∼ FX and Y, Y 0 ∼ GY . Now, h(x1 , x2 ; y1 , y2 ) satisfies the conditions in
Theorem A.1 with r = 2;
• h0,1 (x) = h1,0 (y) = 0
for all x and y almost surely
• h0,2 (z1 , z2 ) = h2,0 (z1 , z2 ) = −2h1,1 (z1 , z2 ).
As a result, the permutation distribution of the MMD based on the U -statistic is asymptotically equivalent to the corresponding unconditional
distribution with additional conditions that
m/N → π1 and n/N → π2 as N → ∞, and E h2 < ∞.
26
B
B.1
Proofs
Proof of Lemma 2.1
From the definition of Wd2 , it is clear to see that Wd2 ≥ 0 and it becomes zero if FX = GY . For
the other direction, we will show that if Wd2 = 0, then X and Y have the same characteristic
function:
h > i
h > i
EX eitβ X = EY eitβ Y
for all (β, t) ∈ Sd−1 × R,
which implies FX = GY .
1. Univariate case
We begin with the univariate case. Suppose that X and Y are univariate random variables.
We claim that the univariate CvM distance has the characteristic property:
Z
2
W =
(FX (t) − GY (t))2 d (FX (t)/2 + GY (t)/2) = 0
R
if and only if FX (t) = GY (t) for all t ∈ R. Define E to be the smallest set E ⊆ R such that
Z
Z
dFX =
dGY = 1.
E
E
Since P(X ∈ E c ) = P(Y ∈ E c ) = 0, we only need to show that W 2 = 0 implies FX (t) = GY (t)
for all t ∈ E. Suppose there exists t0 ∈ E such that FX (t0 ) 6= GY (t0 ). Suppose further that
either FX or GY is discontinuous at t0 . Then, there is a nonzero probability mass at t0
and thus W 2 > 0. This contradicts the assumption. Next, suppose that both FX and GY
are continuous at t0 . Then W 2 > 0 follows similarly from Lemma 4.1 of Lehmann (1951).
Therefore, we conclude that W 2 = 0 implies that X and Y have the same distribution.
2. Multivariate case
Recall that λ(·) is the uniform measure on Sd−1 . From the characteristic property of the
univariate CvM distance, Wd2 = 0 implies that β > X and β > Y are identically distributed for
λ-almost all β ∈ Sd−1 . Now, by the continuity of the characteristic function, we conclude that
h > i
h > i
EX eitβ X = EY eitβ Y
for all (β, t) ∈ Sd−1 × R.
B.2
Proof of Theorem 2.1
The key ingredient of the proof is Lemma 2.2 originally provided in Escanciano (2006). In
the original paper, the integration was with respect to dβ rather than dλ(β):
(
)
Z
>U
U
2
1
I(β > U1 ≤ 0)I(β > U2 ≤ 0)dβ = c(d) π − arccos
,
||U1 ||||U2 ||
Sd−1
where c(d) = π d/2−1 /Γ(d/2) as corrected in Zhu et al. (2017). Our expression in Lemma 2.2
follows by the fact that the area of the surface of the d-dimensional unit ball is 2π d/2 /Γ(d/2).
We start with a general case and then simplify the expression when FX and GY are
continuous.
27
1. General case
Recall that
Wd2 =
Z
Z
Sd−1
= Eβ,W∗
|
Fβ > X (t) − Gβ > Y (t)
2
dwβ (t)dλ(β)
R
>
Fβ > X (β W∗ )
{z
(I)
2
+ Eβ,W∗
} |
>
Gβ > Y (β W∗ )
{z
(II)
2
}
h
i
− 2 Eβ,W∗ Fβ > X (β > W∗ )Gβ > Y (β > W∗ ) ,
|
{z
}
(III)
where W∗ ∼ 1/2FX + 1/2GY . By the Fubini’s theorem and the definition of W∗ , the first
term (I) has the identity
h
i
(I) = Eβ,W∗ ,X,X 0 I(β > X ≤ β > W∗ , β > X 0 ≤ β > W∗ )
h
i
1
>
> 00
> 0
> 00
0
00
= Eβ,X,X ,X I(β X ≤ β X , β X ≤ β X )
2
h
i
1
+ Eβ,X,X 0 ,Y I(β > X ≤ β > Y, β > X 0 ≤ β > Y ) .
2
Similarly,
h
i
(II) = Eβ,W∗ ,Y,Y 0 I(β > Y ≤ β > W∗ , β > Y 0 ≤ β > W∗ )
h
i
1
= Eβ,Y,Y 0 ,Y 00 I(β > Y ≤ β > Y 00 , β > Y 0 ≤ β > Y 00 )
2
h
i
1
+ Eβ,Y,Y 0 ,X I(β > Y ≤ β > X, β > Y 0 ≤ β > X)
2
and
h
i
(III) = Eβ,X,Y,W∗ I(β > X ≤ β > W∗ , β > Y ≤ β > W∗ )
h
i
1
= Eβ,X,X 0 ,Y I(β > X ≤ β > X 0 , β > Y ≤ β > X 0 )
2
h
i
1
+ Eβ,X,Y,Y 0 I(β > X ≤ β > Y 0 , β > Y ≤ β > Y 0 ) .
2
We then apply Lemma 2.2 to obtain the desired result.
2. Continuous case
28
When FX , GY are continuous, then for each fixed β ∈ Sd−1 , the following identities hold:
Z
1
2
Fβ > X (t) dFβ > X (t) = P max{β > X, β > X 0 } ≤ β > X 00 = ,
3
Z
2
1
Gβ > Y (t) dGβ > Y (t) = P max{β > Y, β > Y 0 } ≤ β > Y 00 = ,
3
(27)
Z
2
>
> 0
>
Fβ > X (t) dGβ > Y (t) = P max{β X, β X } ≤ β Y ,
Z
Gβ > Y (t)
2
dFβ > X (t) = P max{β > Y, β > Y 0 } ≤ β > X .
In addition, from the following two identities:
• P max{β > X, β > X 0 } ≤ β > Y + P max{β > X, β > Y } ≤ β > X 0
+ P max{β > X 0 , β > Y } ≤ β > X = 1,
• P max{β > X, β > Y } ≤ β > X 0 = P max{β > X 0 , β > Y } ≤ β > X ,
we have
Z
Fβ > X (t)Gβ > Y (t)dFβ > X (t) = P max{β > X, β > Y } ≤ β > X 0
1 1
= − P max{β > X, β > X 0 } ≤ β > Y .
2 2
(28)
Similarly,
Z
Fβ > X (t)Gβ > Y (t)dGβ > Y (t) = P max{β > Y, β > X} ≤ β > Y 0
1 1
= − P max{β > Y, β > Y 0 } ≤ β > X .
2 2
Now, combine (27), (28) and (29) together to have
Z
Z
2
Fβ > X (t) − Gβ > Y (t) d(Fβ > X (t)/2 + Gβ > Y (t)/2)dλ(β)
Sd−1
R
Z
=
Sd−1
P max{β > X, β > X 0 } ≤ β > Y dλ(β)
2
P max{β > Y, β > Y 0 } ≤ β > X dλ(β) − .
3
d−1
S
Z
+
Hence,
h
i
Wd2 = Eβ,X,X 0 ,Y I(β > X ≤ β > Y, β > X 0 ≤ β > Y )
h
i 2
+ Eβ,Y,Y 0 ,X I(β > Y ≤ β > X, β > Y 0 ≤ β > X) − .
3
Then apply Lemma 2.2 to obtain the result.
29
(29)
Lemma B.1 (V -statistic as a plug-in estimator). The V -statistic is equivalent to the plug-in
estimator:
X
1
m3 n3
Vm,n =
X
h(Xi1 , Xi2 , Xi3 ; Yj1 , Yj2 , Yj3 )
i1 ,i2 ,i3 j1 ,j2 ,j3
Z
Z
=
Sd−1
2
b β > Y (t)/2 dλ(β).
b β > Y (t) d Fbβ > X (t)/2 + G
Fbβ > X (t) − G
R
b β > X as
Proof. Consider the decomposition of the weighted L2 difference between Fbβ > X and G
Z
Z
Sd−1
2
b β > Y (t)/2 dλ(β)
b β > Y (t) d Fbβ > X (t)/2 + G
Fbβ > X (t) − G
R
= Iˇ(1) + Iˇ(2) + Iˇ(3) + Iˇ(4) + Iˇ(5) + Iˇ(6) ,
where
1
Iˇ(1) =
2
Z
1
Iˇ(2) =
2
Z
1
Iˇ(3) =
2
Z
1
Iˇ(4) =
2
Z
Z
Sd−1
R
Z
Sd−1
2
b β > X (t) dFbβ > X (t)dλ(β),
G
R
Z
Sd−1
R
Z
Z
Sd−1
Z
Iˇ(6) = −
2
b β > X (t)dλ(β),
Fbβ > X (t) dG
R
Z
Sd−1
Iˇ(5) = −
2
Fbβ > X (t) dFbβ > X (t)dλ(β),
2
b β > X (t) dG
b β > X (t)dλ(β),
G
b β > X (t)dFbβ > X (t)dλ(β),
Fbβ > X (t)G
R
Z
Sd−1
b β > X (t)dG
b β > X (t)dλ(β).
Fbβ > X (t)G
R
Note that
Z
Z
Sd−1
Z
=
Sd−1
=
=
1
m3
2
Fbβ > X (t) dFbβ > X (t)dλ(β)
R
m
1 X
m
i2 =1
X Z
i1 ,i2 ,i3
Sd−1
!2
m
1 X
I(β > Xi1 ≤ β > Xi2 ) dλ(β)
m
i1 =1
I(β > Xi1 ≤ β > Xi3 , β > Xi2 ≤ β > Xi3 )dλ(β)
1
1 X 1
− 3
Ψ (Xi1 − Xi3 , Xi2 − Xi3 ) .
2 m
2π
i1 ,i2 ,i3
The last identity is due to Lemma 7.1. The rest terms can be calculated similarly; hence, the
result follows.
30
B.3
Proof of Theorem 2.2
We first show that h∗ is degenerate under H0 . Then apply the limit theorem of U -statistics.
1. Degeneracy
(1)
To begin, let us focus on univariate cases where x1 , x2 , y1 , y2 ∈ R. Let h∗ denote the
symmetrized h∗ for the univariate case, that is
(1)
h∗ (x1 , x2 ; y1 , y2 ) :=
1n
I(max{x1 , x2 } ≤ y1 ) + I(max{x1 , x2 } ≤ y2 )
2
o 2
+ I(max{y1 , y2 } ≤ x1 ) + I(max{y1 , y2 } ≤ x2 ) − .
3
Due to the identity
I(max{x1 , x2 } ≤ y1 ) + I(max{x1 , x2 } ≤ y2 )
+ I(max{y1 , y2 } ≤ x1 ) + I(max{y1 , y2 } ≤ x2 ) − 1
= I(max{x1 , x2 } ≤ min{y1 , y2 }) + I(max{y1 , y2 } ≤ min{x1 , x2 }),
the symmetrized kernel becomes
(1)
2h∗ (x1 , x2 ; y1 , y2 ) = I(max{x1 , x2 } ≤ min{y1 , y2 })
1
+ I(max{y1 , y2 } ≤ min{x1 , x2 }) − ,
3
which is equivalent to the kernel for Lehmann’s two-sample statistic (Lehmann, 1951). In
addition, when FX = GY , we have
h
i
(1)
(1)
h∗1,0 (x1 ) := E h∗ (x1 , X2 ; Y1 , Y2 ) = 0,
(30)
h
i
(1)
(1)
h∗0,1 (y1 ) := E h∗ (X1 , X2 ; y1 , Y2 ) = 0.
See Chapter 4 of Bhat (1995) for details.
Now, come back to multivariate cases where x1 , x2 , y1 , y2 ∈ Rd . Let h∗ denote the inte(1)
gration of the univariate kernel h∗ over the unit sphere, i.e.
Z
(1)
h∗ (x1 , x2 , y1 , y2 ) =
h∗ (β > x1 , β > x2 ; β > y1 , β > x2 )dλ(β).
Sd−1
Due to (30), for any β ∈ Sd−1 , we have
h
i
h
i
(1)
(1)
E h∗ (β > x1 , β > X2 ; β > Y1 , β > Y2 ) = E h∗ (β > X1 , β > X2 ; β > y1 , β > Y2 ) = 0,
which results in
h∗1,0 (x1 ) := E [h∗ (x1 , X2 ; Y1 , Y2 )]
Z
h
i
(1)
=
E h∗ (β > x1 , β > X2 ; β > Y1 , β > Y2 ) dλ(β) = 0
Sd−1
31
h∗0,1 (y1 ) := E [h∗ (X1 , X2 ; y1 , Y2 )]
Z
i
h
(1)
E h∗ (β > X1 , β > X2 ; β > y1 , β > Y2 ) dλ(β) = 0.
=
Sd−1
On the other hand,
h∗2,0 (x1 , x2 ) := E [h∗ (x1 , x2 ; Y1 , Y2 )]
Z
2
1
=
1 − Fβ > X (max{β > x1 , β > x2 }) dλ(β)
2 Sd−1
Z
1
1
F 2> (min{β > x1 , β > x2 })dλ(β) − ,
+
2 Sd−1 β X
6
h∗0,2 (y1 , y2 ) := E [h∗ (X1 , X2 ; y1 , y2 )]
Z
2
1
1 − Gβ > Y (max{β > y1 , β > y2 }) dλ(β)
=
2 Sd−1
Z
1
1
+
G2β > Y (min{β > y1 , β > y2 })dλ(β) − ,
2 Sd−1
6
h∗1,1 (x1 , y1 ) := E [h∗ (x1 , X2 ; y1 , Y2 )]
1
= − h∗2,0 (x1 , y1 ).
2
Note that h∗2,0 (x1 , x2 ) 6= 0 for some (x1 , x2 ). For example, when x1 = x2 , it is seen that
h∗2,0 (x1 , x1 ) ≥ 1/12 and by the continuity of h∗2,0 at (x1 , x1 ), there exist a set E with nonzero
measure such that h∗2,0 (z1 , z2 ) > 0 for (z1 , z2 ) ∈ E. Therefore, we conclude that h∗ has a
degeneracy of order one under H0 .
2. Limiting distribution of the U -statistic
For the U -statistic, we apply the result given in Chapter 3 of Bhat (1995) to have
d
∗
N Um,n
−→
∞
∞
∞
1 X
2 X
1 X
λk (Zk2 − 1) +
λk (Zk02 − 1) − √
λk Zi Zi0 ,
π1
π2
π1 π2
k=1
k=1
k=1
i.i.d.
where Zk , Zk0 ∼ N (0, 1). Next, by Slutsky’s theorem,
∞
i
X h √
2
√
mn
d
∗
π2 Zk − π1 Zk0 − 1
Um,n
−→
λk
m+n
k=1
and
√
π 2 Zk −
√
π1 Zk0 ∼ N (0, 1),
the result follows.
3. Limiting distribution of the V -statistic
∗
∗ converges to a constant
We will show that the residual of the difference between Um,n
and Vm,n
in probability; thereby the desired result follows by Slutsky’s theorem.
32
From the relationship between U - and V -statistics, we can decompose the V -statistic as
∗
Vm,n
=
(m)2 (n)2 ∗
1
Um,n + 2 2 Rm,n ,
2
2
m n
m n
(31)
where
Rm,n =
m
X
n
X
h∗ (Xi1 , Xi2 ; Yj1 , Yj2 ) +
+
n
X
n
X
h∗ (Xi1 , Xi2 ; Yj1 , Yj2 )
i1 ,i2 =1 j1 =j2 =1
i1 6=i2
i1 =i2 =1 j1 ,j2 =1
j1 6=j2
m
X
m
X
h∗ (Xi1 , Xi2 ; Yj1 , Yj2 ).
i1 =i2 =1 j1 =j2 =1
From the boundness of h∗ , we apply McDiarmid’s inequality (McDiarmid, 1989) to have
!
1
C2
P
(Rm,n − E [Rm,n ]) > < 2 exp − PN
,
2
m 2 n2
c
i=1 i
where C > 0 is an universal constant and
2
2
N
X
1
2
2
1
2
2
2
ci = n
+
−
+m
+
−
.
n2 mn mn2
m2 mn m2 n
i=1
Based on this result, we obtain
1
(Rm,n − E [Rm,n ]) = oP
2
m n2
m+n
mn
.
∗
Under H0 , the expected value of Um,n
becomes zero and thereby
∗
E0 Vm,n
=
Note that by the Fubini’s theorem,
∗
E0 Vm,n
"Z
m
X
1
= E0
2 2
Sd−1 m n
n
X
1
m 2 n2
E0 [Rm,n ] .
(
I(β > Xi1 , β > Xi2 ≤ β > Yj1 )
i1 ,i2 =1 j1 ,j2 =1
+ I(β > Yj1 , β > Yj2
"
Z
=
Sd−1
Z
=
Sd−1
E0
m
X
1
m2 n2
n
X
)
#
2
dλ(β)
≤ β > Xi1 ) −
3
(
I(β > Xi1 , β > Xi2 ≤ β > Yj1 )
i1 ,i2 =1 j1 ,j2 =1
2
+ I(β > Yj1 , β > Yj2 ≤ β > Xi1 ) −
3
1
mn(n − 1) nm(m − 1) mn
+
+
dλ(β)
m2 n2
6
6
3
33
)#
dλ(β)
=
1m+n
.
6 mn
Therefore, we conclude that
∗
mn
mn
∗
Vm,n
− E0 Vm,n
=
U ∗ + oP (1),
m+n
m + n m,n
(32)
and the proof is complete.
B.4
Proof of Proposition 2.1
Consider the symmetrized kernel of h∗ denoted by
h∗ (x1 , x2 ; y1 , y2 ) :=
1
1
h∗ (x1 , x2 ; y1 , y2 ) + h∗ (x2 , x1 ; y2 , y1 ).
2
2
Followed by the proof of Theorem 2.2, the symmetrized kernel h∗ satisfies the first and second
conditions in Theorem A.1. Therefore, the result of the U -statistic follows directly based
on Theorem A.1 by replacing N with mn/(m + n). On the other hand, according to (32),
em,n and Vem,n are asymptotically equivalent when X and Y have the same distribution.
U
Applying Slutsky’s theorem, we can similarly show that the permutation distribution of Vem,n
approximates the corresponding unconditional distribution for a large sample size. Hence, the
result follows.
B.5
Proof of Theorem 3.1
The boundedness property of the multivariate CvM statistic guarantees that the variance of
the U -statistic is bounded regardless of the presence of contamination. Consequently, when
there exists a fixed signal, we have
E1 [Um,n ] − E0 [Um,n ]
p
→ ∞,
max{V0 [Um,n ] , V1 [Um,n ]}
as the sample size increases. On the other hand, the energy statistic can be adversely affected
by outliers, which lead to an extremely high variance of the test statistic. In fact, as the same
size tends to infinity, the energy statistic can be completely dominated by these outliers and
the resulting test becomes powerless in the end. We will make this statement more rigorous
below.
1. Multivariate CvM statistic
From Theorem 2 of Lee (1990), the explicit formula of the variance of Um,n is given by
V [Um,n ] =
k1 X
k2
X
k1
c
c=0 d=0
k2
d
m−k1 n−k2
k1 −c
n k2 −d
m
k1 k2
2
σc,d
,
(33)
where
2
σc,d
= V [hc,d (X1 , . . . , Xc ; Y1 , . . . , Yd )] ,
and hc,d is defined in Appendix A. In our case, we have k1 = k2 = 3 or k1 = k2 = 2 when
X, Y have continuous distributions.
34
2 is also bounded for any random
Note that the boundedness of a kernel implies that σc,d
vectors. Therefore, based on the variance formula in (33), one can find an universal constant
C such that
C
V [Um,n ] ≤ ,
(34)
N
when 0 < ` ≤ m/N, n/N ≤ u < 1.
Suppose that we reject the null hypothesis when
Um,n ≥ E0 [Um,n ] + cα,m,n = cα,m,n ,
where cα,m,n is a critical value that controls the type I error and thus the test function is
defined as
I(Um,n ≥ cα,m,n ).
One naive choice (which is larger than the one based on (8)) can be chosen as
r
C
cnaive
α,m,n =
Nα
by Chebyshev’s inequality.
Next, we bound the power function. Note that since cnaive
α,m,n → 0 as N → ∞ and E1 [Um,n ] ≥
2
δ2 for every m, n, there exists N0 such that for N ≥ N0 ,
1
E1 [Um,n ] ≥ cnaive
α,m,n .
2
Therefore, for N ≥ N0 , the power function is lower-bounded by
(i)
P1 (Um,n ≥ cα,m,n ) ≥ P1 Um,n ≥ cnaive
α,m,n
(35)
cnaive
Um,n − E1 [Um,n ]
E1 [Um,n ]
α,m,n
p
= P1
≥p
−p
V1 [Um,n ]
V1 [Um,n ]
V1 [Um,n ]
!
(ii)
Um,n − E1 [Um,n ]
E1 [Um,n ]
p
≥ P1
≥− p
V1 [Um,n ]
2 V1 [Um,n ]
!
E1 [Um,n ]
−Um,n + E1 [Um,n ]
p
= 1 − P1
≥ p
V1 [Um,n ]
2 V1 [Um,n ]
(iii)
≥ 1−
(iv)
≥ 1−
!
4V1 [Um,n ]
(E1 [Um,n ])2
4C
,
N δ24
where we used cnaive
α,m,n ≥ cα,m,n for (i), (ii) is due to (35), (iii) is based on the Chebyshev’s
inequality, and (iv) follows by (11) and (34). Note that the lower bound is uniform; therefore,
for N ≥ N0 , we have
inf β(Um,n ) = inf 0 P1 (Um,n ≥ cα,m,n ) ≥ 1 −
0
Hm ,Hn
Hm ,Hn
35
4C
N δ24
and conclude
lim
inf β(Um,n ) = 1.
0
m,n→∞ Hm ,Hn
2. Energy statistic
For the energy statistic, we will pick FX , GY , Hm , Hn0 properly to show that the energy statistic
is dominated by samples from the contaminations Hm and Hn0 . When Hm and Hn0 behave
in favor of H0 (i.e. Hm = Hn0 ), the resulting test based on the energy statistic has no power
even if FX 6= GY for all m, n.
To begin, we choose
FX = N ((µ, . . . , µ)> , Id ), GY = N ((−µ, . . . , −µ)> , Id ) and
2
Hm = Hn0 = N ((0, . . . , 0)> , σN
Id ),
2 = (m + n)p with p > 2. It is worth noting that there is nothing special about
where σN
the choice of normal distributions. The rest of the proof can be similarly applied to other
distributions with some moment conditions. Under this setting, it is clear to see that
Z
Z
2
2
2
Fβ > X (t) − Gβ > Y (t) dtdλ(β) = δ12 > 0.
Ed = (1 − )
Sd−1
R
e Ye as follows:
Next, for given X and Y , we define the truncated random vectors X,
(
>
e = (0, . . . , 0) ,
X
X/σN ,
if X ∼ FX ,
if X ∼ Hm ,
(
(0, . . . , 0)> ,
Ye =
Y /σN ,
if Y ∼ GY ,
if Y ∼ Hn0 .
and
e and Ye have the same mixture distribution as:
By the construction, X
e Ye ∼ (1 − )θ0 + N (0, Id ),
X,
where θ0 has the degenerating distribution at (0, . . . , 0)> . At a high-level, when the energy
statistic of the original samples is asymptotically equivalent to that of the truncated samples
from the identical distributions, then it implies that the resulting test based on the energy
statistic is asymptotically powerless.
Define the difference between the Euclidean distances based on the original samples
e1 , X
e2 ) by
(X1 , X2 ) and the truncated samples (X
∆m,n (X1 , X2 ) =
1
e1 − X
e2 ||
||X1 − X2 || − ||X
σN
e1 − X
e2 ||.
= ||X1 /σN − X2 /σN || − ||X
36
Note that there are four possible cases of the difference ∆m,n (X1 , X2 ):
if X1 , X2 ∼ FX ,
(a) σ1N ||X1 − X2 ||,
(b) 1 ||X − X || − 1 ||X ||,
if X1 ∼ FX , X2 ∼ Hm ,
1
2
2
σN
σN
∆m,n (X1 , X2 ) =
1
1
(c) σN ||X1 − X2 || − σN ||X1 ||,
if X1 ∼ Hm , X2 ∼ FX ,
(d) 0,
if X1 , X2 ∼ Hm ,
and we will show that each case of ∆m,n (X1 , X2 ) is asymptotically negligible. First, note that
−1
if X ∼ FX , then X/σN = OP (σN
) and whereas if X ∼ Hm , then X/σN = OP (1). Based on
−1
this observation, we see that the remainder term in (a) becomes
OP (σN
). For the remainders
√
in (b) and (c), we apply the Taylor expansion of f (x) = x + a around x = 0 to show that
−1
both remainders in (b) and (c) are OP (σN
). Therefore, we conclude that
−1
∆m,n (X1 , X2 ) = OP (σN
).
In a similar fashion, one can obtain
−1
),
∆m,n (Y1 , Y2 ) = OP (σN
−1
)
∆m,n (X1 , Y1 ) = OP (σN
and
−1
),
∆m,n (X2 , Y2 ) = OP (σN
and thus
1
1
1
1
||X1 − Y1 || +
||X2 − Y2 || −
||X1 − X2 || −
||Y1 − Y2 ||
σN
σN
σN
σN
e1 − Ye1 || + ||X
e2 − Ye2 || − ||X
e1 − X
e2 || − ||Ye1 − Ye2 || + OP (σ −1 ).
= ||X
N
In the end, the following approximation holds
1
em,n + OP (σ −1 ),
Em,n = E
N
σN
where
m
n
n
X
X
1
e
ei − Yej || + ||X
ei − Yej ||
Em,n =
||X
1
1
2
2
(m)2 (n)2
i1 ,i2 =1 j1 ,j2 =1
i1 6=i2 j1 6=j2
o
ei − X
ei || − ||Yej − Yej || .
−||X
1
2
1
2
em,n has degeneracy of order one with the finite variance of the kernel, N E
em,n
Since E
converges to an infinite weighted sum of chi-square random variables:
d
em,n −→
NE
∞
X
λk (Zk2 − 1),
k=1
2
p
for some {λk }∞
k=1 . Under the assumption that σN = (m + n) with p > 2, we now conclude
that
N
em,n + OP (N σ −1 ) = N E
em,n + oP (1),
Em,n = N E
N
σN
and N Em,n /σN converges to the same null distribution:
∞
X
N
d
Em,n −→
λk (Zk2 − 1).
σN
k=1
Hence, the resulting test becomes asymptotically powerless.
37
B.6
Proof of Lemma 4.1
Let β > Z ∼ 12 Fβ > X + 12 Gβ > Y . From the definition of the multivariate CvM distance,
Wd2 = Eβ,Z
(i)
Fβ > X (β > Z) − Gβ > Y (β > Z)
2
)2
(
Eβ,Z Fβ > X (β > Z) − Gβ > Y (β > Z)
≥
(
=
1
Eβ,X 0 Fβ > X (β > X 0 ) − Gβ > Y (β > X 0 )
2
1
+ Eβ,Y 0 Fβ > X (β > Y 0 ) − Gβ > Y (β > Y 0 )
2
(ii)
(
≥
1
Eβ EX 0 Fβ > X (β > X 0 ) − Gβ > Y (β > X 0 )
2
1
+ Eβ EY 0 Fβ > X (β > Y 0 ) − Gβ > Y (β > Y 0 )
2
(
=
)2
)2
1
Eβ EX,X 0 ,Y I(β > X ≤ β > X 0 ) − I(β > Y ≤ β > X 0 )
2
1
+ Eβ EX,Y,Y 0 I(β > X ≤ β > Y 0 ) − I(β > Y ≤ β > Y 0 )
2
(
)
2
1
(iii)
Eβ − PX,Y β > X ≤ β > Y
=
2
(Z
)2
1
>
>
− PX,Y β X ≤ β Y dλ(β) ,
=
Sd−1 2
)2
where (i), (ii) are due to the Jensen’s inequality and (iii) is based on the fact that P(β > X <
β > X 0 ) = P(β > Y < β > Y 0 ) = 1/2. Therefore, the lower bound in (14) holds.
B.7
Proof of Theorem 4.1
The minimax lower bound is based on an application of Neyman-Pearson lemma also known
as the Le Cam’s lemma (see, e.g. Baraud, 2002):
inf
sup
φ∈ΦN,α FX ,GY ∈F (?
m,n )
Pm,n
(φ = 0) ≥ 1 − α − sup P0m,n (A) − P1m,n (A)
1
A
1
≥ 1−α−
2
= 1−α−
38
1
2
s
Z
Rd
dP1m,n
dP0m,n
2
q
E0 L2m,n − 1.
dP0m,n − 1
For our calculation, it is more convenient to use the following bound (by switching P0m,n with
P1m,n ):
q
1
Pm,n
(φ
=
0)
≥
1
−
α
−
inf
sup
E1 L−2
m,n − 1.
1
φ∈ΦN,α FX ,GY ∈F (? )
2
m,n
Now, if we can show that
2
E1 L−2
m,n ≤ 1 + 4(1 − α − ξ) ,
(36)
then it follows
inf
sup
φ∈ΦN,α FX ,GY ∈F (?
m,n )
Pm,n
(φ = 0) ≥ ξ.
1
Under the alternative, we choose
X ∼ FX = N ((−δN , . . . , −δN ), Id )
and Y ∼ GY = N ((δN , . . . , δN ), Id ),
where
s
δN =
1
log
4d(m + n)
4
2
(1 − α − ξ) + 1 ,
ω1
(37)
and ω1 = π14 + 2π13 π2 + π24 + 2π1 π23 + π12 π22 . From Lemma 4.2, we know that our choice of two
distributions satisfies (FX , GY ) ∈ F(?m,n ). Under the null, we choose the mixture of FX and
GY :
X, Y ∼ π1 FX + π2 GY
with π1 + π2 = 1.
Then after straightforward calculations, one can have
2
2
2
4
3
4ndδN
E1 L−2
+ (π24 + 2π1 π23 )e4mdδN + π12 π22 e4(m+n)dδN
m,n = (π1 + 2π1 π2 )e
+ π12 π22 + 2π13 π2 + 2π1 π23 + 4π12 π22
2
≤ (π14 + 2π13 π2 + π24 + 2π1 π23 + π12 π22 ) e4(m+n)dδN
|
{z
}
ω1
+ π12 π22 + 2π13 π2 + 2π1 π23 + 4π12 π22 .
|
{z
}
ω2
Note that ω1 + ω2 = 1 and so (36) holds with our choice of δN in (37), which yields the claim.
B.8
Proof of Theorem 4.2
We first present the result based on the V -statistic and then come back to the U -statistic.
1. V -statistic
To begin, define ρ(FX∗ , G∗Y ) as
ρ(FX∗ , G∗Y )
39
Z
:=
Sd−1
Z
2
Fβ∗> X (t) − G∗β > Y (t) d(Fβ > X (t)/2 + Gβ > Y (t)/2)dλ(β).
R
ρm,n (FX∗ , G∗Y )
Further define
correspondence, that is,
by changing the weight function of ρ(FX , GY ) with the empirical
ρm,n (FX∗ , G∗Y )
Z
Z
2
b β > Y (t)/2)dλ(β),
=
Fβ∗> X (t) − G∗β > Y (t) d(Fbβ > X (t)/2 + G
Sd−1
R
b β > Y (t) are the empirical distribution functions, corresponding to Fβ > X (t),
where Fbβ > X (t), G
Gβ > Y (t). From the definition, we have
b Y ).
Vm,n = ρm,n (FbX , G
By Minkowski’s inequality,
q
q
q
b
b Y , GY )
ρm,n (FX , GY ) ≤
ρm,n (FX , GY ) + ρm,n (G
≤
q
q
q
b Y , GY ) + ρm,n (FbX , G
b Y ).
ρm,n (FX , FbX ) + ρm,n (G
Notice that
)
(
q
2 1/2
>
>
ρm,n (FX , FbX ) = Eβ,Zm,n Fβ > X (β Zm,n ) − Fbβ > X (β Zm,n )
(
≤
sup
(t,β)∈R×Sd−1
=
Fβ > X (t) − Fbβ > X (t)
2
)1/2
Fβ > X (t) − Fbβ > X (t)
sup
(t,β)∈R×Sd−1
:= dK (FX , FbX ),
b Y . Similarly,
where Zm,n are random vectors generated from 1/2FbX + 1/2G
b Y ) :=
dK (GY , G
sup
(t,β)∈R×Sd−1
≥
b β > Y (t)
Gβ > Y (t) − G
q
b Y ).
ρm,n (GY , G
Hence we have
q
q
b Y ) ≥ ρm,n (FX , GY ) − dK (FX , FbX ) − dK (GY , G
b Y ).
ρm,n (FbX , G
This in turn implies that
P1 Vm,n ≥ cvα,m,n
q
p
b Y ) ≥ cv
≥ P1
ρm,n (FX , GY ) − dK (FX , FbX ) − dK (GY , G
α,m,n
40
q
p
b Y ) ≥ ρm,n (FX , GY ) − cvα,m,n
= 1 − P1 dK (FX , FbX ) + dK (GY , G
q
1
1p v
b
≥ 1 − P1 dK (FX , FX ) ≥
ρm,n (FX , GY ) −
cα,m,n
2
2
q
1
1p v
b
ρm,n (FX , GY ) −
cα,m,n ,
− P1 dK (GY , GY ) ≥
2
2
(i)
where (i) follows from the fact that {A + B ≥ } implies {A ≥ /2} ∪ {B ≥ /2}.
Next, consider the event A defined by
(
)
1
A = ρm,n (FX , GY ) − ρ(FX , GY ) ≤ ρ(FX , GY ) .
2
Under this event, it is clear to see that ρm,n (FX , GY ) is lower bounded by
1
ρm,n (FX , GY ) ≥ ρ(FX , GY ).
2
p
√
Then, combining with the condition that ρ(FX , GY ) ≥ bN / N , we have
q
1
1p v
b
P1 dK (FX , FX ) ≥
ρm,n (FX , GY ) −
cα,m,n
2
2
q
1p v
1
b
= P1 dK (FX , FX ) ≥
ρm,n (FX , GY ) −
cα,m,n , A
2
2
q
1p v
1
ρm,n (FX , GY ) −
cα,m,n , Ac
+ P1 dK (FX , FbX ) ≥
2
2
bN
1p v
≤ P dK (FX , FbX ) ≥ √
−
cα,m,n + P1 (Ac ) .
2
2 2N
Similarly, one can obtain
q
p
b Y ) ≥ 1 ρm,n (FX , GY ) − 1 cv
P1 dK (GY , G
α,m,n
2
2
bN
1p v
bY ) ≥ √
≤ P dK (GY , G
−
cα,m,n + P1 (Ac ) .
2
2 2N
Note that ρm,n (FX , GY ) is the sample mean of i.i.d. random variables f (Z1 ), . . . , f (ZN )
where
ρm,n (FX , GY )
Z Z
2
b Y /2)
=
Fβ > X (β > Zi ) − Gβ > Y (β > Zi ) dλ(β)d(FbX /2 + G
Rd
:=
β∈Sd−1
N
1 X
f (Zi ),
N
i=1
41
and its expected value is ρ(FX , GY ). With this in mind, we apply Chebyshev’s inequality to
bound
1
c
P1 (A ) = P1 ρm,n (FX , GY ) − ρ(FX , GY ) > ρ(FX , GY )
2
≤
4
1
4
V1 [ρm,n (FX , GY )] =
V1 [f (Z)]
2
2
(ρ(FX , GY ))
(ρ(FX , GY )) N
1
4
E1 f 2 (Z)
2
(ρ(FX , GY )) N
4 E1 f 2 (Z)
=
from ρ(FX , GY ) = E1 [f (Z)]
N (E1 [f (Z)])2
≤
≤
4 E1 [f (Z)]
N (E1 [f (Z)])2
=
4
1
4 N
≤
N E1 [f (Z)]
N b2N
=
4
.
b2N
because f (Z) ∈ [0, 1],
by using the condition ρ(FX , GY ) ≥ b2N /N
As remarked in the main text, the critical value cα,m,n is chosen based on the mixture of
FX and GY and it changes with the distributions; thereby the resulting test is not distributionfree. However, since Vm,n is a degenerate V -statistic under H0 with the bounded kernel, one
can find an universal constant Bα such that
Bα
≥ cvα,m,n for any (FX , GY ),
N
Bα
(ii) P0 Vm,n ≥
≤ α, whenever FX = GY .
N
(i)
Given Bα , we take bN such that b2N ≥ 4Bα . Then
√
bN
1p v
b
P dK (FX , FX ) ≥ √
−
cα,m,n ≤ P
mdK (FX , FbX ) ≥
2
2 2N
√
bN
1p v
b
bY ) ≥
P dK (GY , GY ) ≥ √
−
cα,m,n ≤ P
ndK (GY , G
2
2 2N
bN
10
r
bN
10
r
m
N
n
N
,
Note that dK (FX , FbX ) can be considered as the supremum of an empirical process
dK (FX , FbX ) =
sup
(t,β)∈R×Sd−1
Fβ > X (t) − Fbβ > X (t)
m
1 X
= sup
g(Xi ) − E [g(X)] ,
g∈G m
i=1
indexed by
G = {x 7→ gt,β = I(β > x ≤ t) : (t, β) ∈ R × Sd−1 }.
42
.
Then, followed by Theorem 7.3 of Giné and Nickl (2009) and Lemma 7.2 of Zhou et al. (2017),
h
i
P dK (FX , FbX ) ≥ E dK (FX , FbX ) +
r
×
t
t
+
m 3m
(
h
i
1
+ 4E dK (FX , FbX )
2
)1/2
(38)
!
≤ 2e−t
with
i rd
d
b
E dK (FX , FX ) .
+ .
m m
h
b Y ) can be obtained. As a result, there exist universal
Similarly, the tail bound of dK (GY , G
0
0
constants C1 , C2 , C1 , C2 such that
r
r
√
bN m
m
b
P
≤ C1 exp −C2
bN
mdK (FX , FX ) ≥
10 N
N
r
r
√
bN n
n
0
0
b
P
mdK (GY , GY ) ≥
≤ C1 exp −C2
bN .
10 N
N
Combining the results, we have obtained the uniform lower bound of the power function:
r
m
P1 (Vm,n ≥ cα,m,n ) ≥ 1 − C1 exp −C2
bN
N
r
n
8
0
0
− C1 exp −C2
bN − 2 .
N
bN
Therefore,
inf
(FX ,GY )∈F (?m,n )
P1 Vm,n ≥ cvα,m,n → 1
as
bN → ∞.
In other words,
lim
P1 φVm,n = 0 = 0,
sup
bN →∞ (FX ,GY )∈F (?
m,n )
as m, n → ∞, which completes the proof.
2. U -statistic
Let cuα,m,n be the α level critical value of the test based on Um,n . Now, consider the residual
term Rm,n such that
m3 n3
1
Vm,n − 3 3 Rm,n .
Um,n =
(m)3 (n)3
m n
Since the kernel h is bounded, it is be seen that
1
m 3 n3
|Rm,n | .
43
1
.
N
Based on this result, we can find an universal constant Cα > 0 such that
m 3 n3
1
u
u
P1 Um,n ≥ cα,m,n = P1
Vm,n − 3 3 Rm,n ≥ cα,m,n
(m)3 (n)3
m n
Cα
.
≥ P1 Vm,n ≥
N
We then apply the previous result on the V -statistic to show that
P1 φUm,n = 0 = 0,
lim
sup
bN →∞ (FX ,GY )∈F (?
m,n )
as m, n → ∞.
3. Projection Pursuit-type Kolmogorov–Smirnov statistic
We also consider the projection pursuit-type Kolmogorov–Smirnov (KS) test statistic defined
by
bY ) =
dK,m,n := dK (FbX , G
sup
(t,β)∈R×Sd−1
b β > Y (t) .
Fbβ > X (t) − G
The test based on the KS statistic also achieves the optimal minimax rate.
Proposition B.1. Let us define the test φKS = I dK,m,n ≥ cα,m,n where cα,m,n is a critical
value of the null distribution. Under the same assumptions used in Theorem 4.2, we have
lim
sup
P1 (φKS = 0) = 0.
bN →∞ (FX ,GY )∈F (?
m,n )
Proof. The proof is similar to that of Theorem 4.2. Using the triangle inequality,
b Y ) + dK (FbX , G
b Y ).
dK (FX , GY ) ≤ dK (FX , FbX ) + dK (GY , G
Since the KS distance is lower-bounded by
p
b
√N ≤ ρ(FX , GY ) ≤ dK (FX , GY ),
N
it follows that
P1 (dK,m,n
bN
b
b
≥ cα,m,n ) ≥ P1 √ − dK (FX , FX ) − dK (GY , GY ) ≥ cα,m,n
N
bN
b
b
= 1 − P1 dK (FX , FX ) + dK (GY , GY ) > √ − cα,m,n
N
r
√
bN m 1 √
b
≥ 1 − P1
mdK (FX , FX ) >
−
mcα,m,n
2
N
2
r
√
bN n
1√
b
− P1
ndK (GY , GY ) >
−
ncα,m,n .
2
N
2
Now, we apply the multivariate DKW-inequality in (38) to finish the proof.
44
B.9
Proof of Lemma 4.3
Let σ02 and σ12 be the variance of
1
[h∗ (X2i−1 , X2i ; Y2i−1 , Y2i ) + h∗ (X2i , X2i−1 ; Y2i−1 , Y2i )] ,
2
respectively. From the boundedness of h∗ , we have 0 < σ02 , σ12 < ∞ for all m, n. By the
central limit theorem,
√
√
M Lm,n d
−→ N (0, 1)
σ0
under H0
√
M Lm,n − Wd2 (FX,N , GY,N )
M Lm,n
d
=
+ o(1) −→ N (0, 1)
σ1
σ1
under H1 ,
because Wd2 (FX,N , GY,N ) N −2ε where ε > 1/4. Hence, the asymptotic null and alternative
distributions are both normally distributed with the same mean zero but with (possibly)
different positive variances. This means that the two asymptotic distributions cannot be
perfectly separated even when m, n → ∞. This leads to inconsistency of the resulting test.
Therefore, the result follows.
B.10
Proof of Theorem 5.1
Chakraborty and Chaudhuri (2017) show that the power based on the CQ and WMW tests
are asymptotically identical under some location models, which include ours (see Theorem
2.2 of Chakraborty and Chaudhuri, 2017). Here, we will establish the rest of identities. Since
TCQ is easier to deal with than TWMW , we will consider 1) the relationship between the CvM
statistic and the CQ statistic and 2) the relationship between the energy statistic and the CQ
statistic in order and will show their asymptotic identities.
Remark B.1. Ramdas et al. (2015) observe that the tests based on TCQ , the Gaussian MMD
and a variant of the energy statistic involving a tuning parameter are the same under highdimensional location models (Theorem 3 of Ramdas et al., 2015). Our approach is different
to theirs in that we do not consider the variant of the energy statistic but the original energy
statistic without tuning parameters.
The proof is lengthy and involves Taylor approximations repeatedly. We believe that the
same result can be obtained under much weaker assumptions (as evidenced by our simulations). For example, one can relax the independent assumption among covariates to some
mixing conditions as in Chakraborty and Chaudhuri (2017). However, it will make the proof
much longer and less straightforward. Hence, we prove the result admittedly under restricted
assumptions and provide simulation results to support our claim.
Overview of the proof.
1. In Lemma B.2, we establish the asymptotic identity between the CvM statistic and the
CQ statistic (after properly centering and scaling). We prove the result by applying
Taylor approximations repeatedly and in Lemma B.3, we show that the remainder term
of the approximation is negligible.
45
2. In Lemma B.4, we establish the asymptotic identity between the energy statistic and
the CQ statistic (after properly centering and scaling) by following similar procedures
applied for Lemma B.2.
3. At the end of this section, we collect the results and establish the desired identities.
We begin with establishing the relationship between the CvM statistic and the CQ statistic.
Lemma B.2. Consider the same location-shifted model assumption used in Theorem 5.1 with
µX − µY = µ. Then for every fixed m, n and as d → ∞,
√
2π 3d ∗
µ2
1
d
Um,n − √
=√
TCQ − dµ2 + oP (1) −→ N (0, 1),
γm,n
2π 3σ 2
dσ 2 γm,n
where
s
γm,n =
2
4
2
+
+
.
m(m − 1) n(n − 1) mn
(39)
Proof. To simplify the notations, let us denote the inner products by
K1 =
(X1 − Y1 )> (X2 − Y1 )
,
||X1 − Y1 ||||X2 − Y1 ||
K2 =
(X1 − Y2 )> (X2 − Y2 )
||X1 − Y2 ||||X2 − Y2 ||
K3 =
(Y1 − X1 )> (Y2 − X1 )
,
||Y1 − X1 ||||Y2 − X1 ||
K4 =
(Y1 − X2 )> (Y2 − X2 )
.
||Y1 − X2 ||||Y2 − X2 ||
∗
Then the symmetrized kernel of Um,n
becomes
2πh∗ (X1 , X2 ; Y1 , Y2 ) =
2π 1
− (arccos(K1 ) + arccos(K2 ) + arccos(K3 ) + arccos(K4 )) .
3
2
Since h∗ is location invariant, without loss of generality, we assume E [X] = 0 and E [Y ] =
µ> 1. Using the Taylor expansion of arccos(x) around x = 1/2 and the central limit theorem
(C.L.T.), one can show that
Ψ (X1 − Y1 , X2 − Y1 ) =
π
1
2 (X1 − Y1 )> (X2 − Y1 )
+√ −√
+ oP (d−1/2 )
3
||X
−
Y
||||X
−
Y
||
3
3
1
1
2
1
(40)
under the assumption of µX − µY = µ given in (18). Hence, the kernel h∗ is approximated by
1
2
2πh∗ (X1 , X2 ; Y1 , Y2 ) = √ (K1 + K2 + K3 + K4 ) − √ + oP (d−1/2 ).
3
3
For the rest of the proof, we will show that K1 + K2 + K3 + K4 is dominated by
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 )
= (X1 − Y1 )> (X2 − Y2 ) + (X1 − Y2 )> (X2 − Y1 )
46
and therefore Um,n is essentially the same as TCQ as d → ∞. In order to show this, further
denote
%x1 y1 = ||X1 − Y1 ||,
%x1 y2 = ||X1 − Y2 ||,
%x2 y1 = ||X2 − Y1 ||,
%x2 y2 = ||X2 − Y2 ||.
Note then by the law of large number and the continuous mapping theorem, we have
1
p
%x1 y1 %x2 y1 −→ 2σ 2 + µ2 .
d
(41)
In addition, the central limit theorem provides
1
√ X1> X2 = OP (1),
d
1
√ X1> Y1 = OP (1),
d
1
1
√ Y1> Y2 − dµ2 = √ Y1> Y2 + o(1)
d
d
= OP (1)
(42)
from the condition in (18).
By (41) and (42), we can approximate
K1 + K2 + K3 + K4
n
1
=
X1> X2 − X1> Y1 − X2> Y1 + X1> X2 − X1> Y2
2dσ 2 + dµ2
− X2> Y2 + Y1> Y2 − Y1> X1 − Y2> X1 + Y1> Y2 − Y1> X2 − Y2> X2
+
=
o
Y1> Y1
Y2> Y2
X1> X1
X2> X2
+
+
+
+ oP (d−1/2 )
%x1 y1 %x2 y1
%x1 y2 %x2 y2
%x1 y1 %x1 y2
%x2 y1 %x2 y2
n
o
1
>
>
(X
−
Y
)
(X
−
Y
)
+
(X
−
Y
)
(X
−
Y
)
1
1
2
2
1
2
2
1
2dσ 2 + dµ2
n
o
1
>
>
>
>
Y
X
+
Y
X
+
Y
X
+
Y
X
−
1
1
2
2
2
1
2
2dσ 2 + dµ2 1
+
Y1> Y1
Y2> Y2
X1> X1
X2> X2
+
+
+
+ oP (d−1/2 ).
%x1 y1 %x2 y1
%x1 y2 %x2 y2
%x1 y1 %x1 y2
%x2 y1 %x2 y2
Hence,
K1 + K2 + K3 + K4
n
o
1
=
h
(X
,
X
;
Y
,
Y
)
+
h
(X
,
X
;
Y
,
Y
)
1
2 1 2
1
2 2 1
CQ
CQ
2dσ 2 + dµ2
+ Rd + oP (d−1/2 ),
47
where
Rd =
Y2> Y2
X1> X1
X2> X2
Y1> Y1
+
+
+
%x1 y1 %x2 y1
%x1 y2 %x2 y2
%x1 y1 %x1 y2
%x2 y1 %x2 y2
n
o
1
>
>
>
>
−
Y
X
+
Y
X
+
Y
X
+
Y
X
.
1
1
2
2
1
2
1
2
2dσ 2 + dµ2
Now if we can show
Rd = oP (d−1/2 ),
then K1 +K2 +K3 +K4 is asymptotically identical with hCQ (X1 , X2 ; Y1 , Y2 )+hCQ (X1 , X2 ; Y2 , Y1 )
up to some scaling part as d → ∞. However, it is hard to handle (Xi> Xi )/(%xi y1 %xi y2 ) and
(Yi> Yi )/(%x1 yi %x2 yi ) directly. So we need to further approximate Rd .
Again, by the C.L.T. and the weak law of large number, we have
Y1> Y1
σ 2 + µ2
d(σ 2 + µ2 )
Y1> Y1
−
+
=
+ oP (d−1/2 ).
%x1 y1 %x2 y1
d(2σ 2 + µ2 ) 2σ 2 + µ2
%x1 y1 %x2 y1
In order to deal with the inverse of %x1 y1 %x2 y1 , once again use the Taylor expansion of f (x) =
√
1/ x around at x = 4 to show
%2x1 y1 %2x2 y1
d(σ 2 + µ2 )
1 1
+ oP (d−1/2 ).
= + −
%x1 y1 %x2 y1
2 4 16d2 (σ 2 + µ2 )2
Hence, we have
%2x1 y1 %2x2 y1
Y1> Y1
σ 2 + µ2
Y1> Y1
3
−
+
+ oP (d−1/2 ).
= − 2
%x1 y1 %x2 y1
4 2σ + µ2 16d2 (σ 2 + µ2 )2 d(2σ 2 + µ2 )
Similarly,
%2x1 y2 %2x2 y2
Y2> Y2
σ 2 + µ2
Y2> Y2
3
−
+
+ oP (d−1/2 )
= − 2
%x1 y2 %x2 y2
4 2σ + µ2 16d2 (σ 2 + µ2 )2 d(2σ 2 + µ2 )
%2x1 y1 %2x1 y2
X1> X1
σ 2 + µ2
X1> X1
3
= − 2
−
+
+ oP (d−1/2 )
%x1 y1 %x1 y2
4 2σ + µ2 16d2 (σ 2 + µ2 )2 d(2σ 2 + µ2 )
%2x2 y1 %2x2 y2
X2> X2
3
σ 2 + µ2
X2> X2
= − 2
−
+
+ oP (d−1/2 ).
%x2 y1 %x2 y2
4 2σ + µ2 16d2 (σ 2 + µ2 )2 d(2σ 2 + µ2 )
Let us denote
$(X1 , X2 , Y1 , Y2 ) := Y1> Y1 + Y2> Y2 + X1> X1 + X2> X2
−Y1> X1 − Y2> X1 − Y1> X2 − Y2> X2 .
By combining the above results and letting Rd be
1
2σ 2 − µ2
$(X
,
X
,
Y
,
Y
)
+
1
2
1
2
d(2σ 2 + µ2 )
2σ 2 + µ2
n
o
1
2
2
2
2
2
2
2
2
−
%
%
+
%
%
+
%
%
+
%
%
x1 y2 x2 y2
x1 y1 x1 y2
x2 y1 x2 y2 ,
16d2 (σ 2 + µ2 )2 x1 y1 x2 y1
Rd =
48
(43)
we have
Rd = Rd + oP (d−1/2 ).
Let ηk be the kth moment of each coordinate of W, that is
h
i
ηk = E W1k .
After straightforward calculation, one can obtain the following moments:
E %2x1 y1 %2x2 y1 = E %2x1 y2 %2x2 y2
= d2 µ4 + 4d2 µ2 σ 2 + 4d2 σ 4 + 4dµ2 σ 2 − dσ 4 + 4dη3 µ + dη4 ,
and similarly
E %2x1 y1 %2x1 y2 = E %2x2 y1 %2x2 y2
= d2 µ4 + 4d2 µ2 σ 2 + 4d2 σ 4 + 4dµ2 σ 2 − dσ 4 − 4dη3 µ + dη4 .
Based on these moments, the expected value of Rd is calculated by
2σ 2 − µ2
1
E Rd =2 + 2
−
2
2
2
2σ + µ
16d (σ + µ2 )2
n
o
× 4d2 µ4 + 16d2 µ2 σ 2 + 16d2 σ 4 + 16dµ2 σ 2 − 4dσ 4 + 4dη4 .
Then, under the condition in (18), one can show that
√
d E Rd − 2
√
√
d(2σ 2 − µ2 )
d
=
−
2
2
2
2
2σ + µ
16d (σ + µ2 )2
o
n
× 4d2 µ4 + 16d2 µ2 σ 2 + 16d2 σ 4 + 16dµ2 σ 2 + o(1)
√
√
n
o
d(2σ 2 − µ2 )
d
2 2 2
2 4
=
−
16d
µ
σ
+
16d
σ
+ o(1)
2σ 2 + µ2
16d2 (σ 2 + µ2 )2
√
√
− dµ6 − dµ4 σ 2
=
+ o(1)
(2σ 2 + µ2 )(σ 2 + µ2 )2
= o(1).
Now, from Lemma B.3, the remainder term becomes negligible:
√
d Rd − 2 = oP (1).
In the end, we have established the following approximation:
√
2π dh∗ (X1 , X2 ; Y1 , Y2 )
n
o
1
= √
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 ) + oP (1)
3d(2σ 2 + µ2 )
49
(44)
=
n
o
1
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 ) + oP (1).
2 3dσ 2
√
Combining this result with Theorem 2.1 in Chakraborty and Chaudhuri (2017), we finish the
proof by obtaining
√
2π 3d ∗
µ2
1
d
Um,n − √
TCQ − dµ2 + oP (1) −→ N (0, 1).
=√
2
2
γm,n
2π 3σ
dσ γm,n
The next lemma shows that the variance of Rd is smaller order of d−1 ; thereby, the
residual term is asymptotically negligible. We first upper-bound the variance of Rd by using
the Efron-Stein inequality and then establish the result.
Lemma B.3. Consider the same model assumption in Theorem 5.1 with µ = o(1). Then
V Rd = o(d−1 ),
where Rd is given in (43).
Proof. First, define
Z = Rd −
2σ 2 − µ2
2σ 2 + µ2
= f (X11 , . . . , X1d , X21 , . . . , X2d , Y11 , . . . , Y1d , Y21 , . . . , Y2d ).
|
{z
} |
{z
} |
{z
} |
{z
}
X1
X2
Y1
Y2
We will show the variance of Z is of o(d−1 ).
Based on the Efron-Stein inequality, the variance of Rd is upper bounded by
V(Z) ≤
4d
X
E
(i)
Z −E Z
2
i=1
= 2dE
(1)
Z −E
Z
2
+ 2dE
(2d+1)
Z −E
Z
2
,
where E(i) [·] is the conditional expectation conditioned on everything but ith element of
(X11 , . . . , X1d ,
X21 , . . . , X2d , Y11 , . . . , Y1d , Y21 , . . . , Y2d ). Without loss of generality, assume E [X1i ] = E [X2i ] =
0 and E [Y1i ] = E [Y2i ] = µ.
h
h
2 i
2 i
For the rest of the proof, we will upper bound (i) 2dE Z − E(1) Z
and (ii) 2dE Z − E(2d+1) Z
in sequence.
Bounding the first term.
Let us begin with dealing with Z − E(1) Z. We define
ϑd = 16d3 (σ 2 + µ2 )2 (2σ 2 + µ2 )
50
so that
ϑ2d
E
(1)
Z −E
Z
2
i
h
2
2
− 2I(1) I(2)
+ I(2)
= E (I(1) − I(2) )2 = E I(1)
where
n
o
2
I(1) = 16d2 (σ 2 + µ2 )2 X11
− σ 2 − X11 Y11 − X11 Y21 ,
(
I(2) = d(2σ 2 + µ2 ) ∆1 + ∆2
+ ∆3
d
X
d
X
(Y2i − X1i )2 +
(Y1i − X2i )2
i=2
i=1
d
X
d
X
(Y1i − X1i )2 +
i=2
!
!)
(Y2i − X2i )2
,
i=1
and
2
2
2
2
2
3
∆1 = Y11
X11
− 2Y11
Y21 X11 − 2Y11 Y21
X11 + 4Y11 Y21 X11
− 2Y11 X11
2
2
3
4
2
2
+ Y21
X11
− 2Y21 X11
+ X11
− σ 2 Y11
− σ 2 Y21
− η4 ,
2
∆2 = X11
− 2Y11 X11 − σ 2 ,
2
∆3 = X11
− 2Y21 X11 − σ 2 .
Since the goal is to show
2dE
Z − E(1) Z
2
= o(d−1 ),
and ϑ2d is order of d6 , it is enough to show that
E (I(1) − I(2) )2 = o(d4 ).
For this reason, we ignore the terms that are of smaller order than d4 . Based on the fact that
2 is approximated by
µ = o(1), the expected value of I(1)
h
i
2
E I(1)
= (256σ 12 + 256η4 σ 8 )d4 + o(d−4 ).
Similarly, one can have
h
i
2
E I(2)
= (256σ 12 + 256η4 σ 8 )d4 + o(d−4 ).
On the other hand,
−2E I(1) I(2) = −(512σ 12 + 512η4 σ 8 )d4 + o(d−4 ),
therefore, we conclude
2dE
(1)
Z −E
Z
51
2
= o(d−1 ).
(45)
Bounding the second term.
Next, we will prove
2dE
Z − E(2d+1) Z
2
= o(d−1 ).
Similar to the previous step, define
ϑd = 16d3 (σ 2 + µ2 )2 (2σ 2 + µ2 )
so that
ϑ2d
E
(2d+1)
Z −E
Z
2
i
h
2
2
− 2J(1) J(2) ,
+ J(2)
= E (J(1) − J(2) )2 = E J(1)
where
n
o
2
J(1) = 16d2 (σ 2 + µ2 )2 Y11
− (σ 2 + µ2 ) − X11 Y11 + X11 µ − X21 Y11 + X21 µ
d
X
(
2
2
J(2) = d(2σ + µ )
∆01
+
∆02
2
(X2i − Y1i ) +
i=2
+
d
X
∆03
d
X
!
2
(X1i − Y2i )
i=1
2
(X1i − Y1i ) +
i=2
d
X
!)
2
,
(X2i − Y2i )
i=1
and
2
2
2
2
2
3
∆01 = X11
Y11
− 2X11
X21 Y11 − 2X11 X21
Y11 + 4X11 X21 Y11
− 2X11 Y11
2
2
3
4
2
2
+ X21
Y11
− 2X21 Y11
+ Y11
− µ4 − µ2 X11
− µ2 X21
− 6µ2 σ 2
2
2
− 4µη4 − σ 2 X11
− σ 2 X21
− η4 ,
2
∆02 = Y11
− 2X11 Y11 + 2X11 µ − µ2 − σ 2 ,
2
∆03 = Y11
− 2X21 Y11 + 2X21 µ − µ2 − σ 2 .
Once again, after ignoring small order terms, we have
h
i
2
E J(1)
= (256σ 12 + 256η4 σ 8 )d4 + o(d−4 )
h
i
2
E J(2)
= (256σ 12 + 256η4 σ 8 )d4 + o(d−4 )
−2E J(1) J(2) = −(512σ 12 + 512η4 σ 8 )d4 + o(d−4 ),
and therefore we conclude
2dE
(2d+1)
Z −E
Z
2
= o(d−1 ).
(46)
Combining (45) with (46),
V Rd = V [Z] ≤ 2dE
(1)
Z −E
Z
2
which completes the proof.
52
+ 2dE
(2d+1)
Z −E
Z
2
= o(d−1 ),
The following lemma is about the high-dimensional approximation of the energy statistic
to the CQ statistic:
Lemma B.4. Suppose X and Y satisfy the assumptions in Theorem 5.1 with µX − µY = µ.
Then for every fixed m, n and d → ∞,
r
√
2
d µ2
1
d
Em,n −
=√
TCQ − dµ2 + oP (1) −→ N (0, 1),
2
γm,n σ
2 σ
dσ γm,n
where Em,n is the energy statistic given in (10) and γm,n is given in (39).
Proof. Consider the symmetrized kernel of the energy statistic Em,n :
χ∗Energy (X1 , X2 ; Y1 , Y2 )
=
1
||X1 − Y1 || + ||X2 − Y1 || + ||X1 − Y2 || + ||X2 − Y2 ||
2
− ||X1 − X2 || − ||Y1 − Y2 ||.
Based on the C.L.T. and Taylor expansion of f (x) =
||X1 − Y1 || =
√
x at x = d(2σ 2 + µ2 ),
1p
1
d(2σ 2 + µ2 ) + p
||X1 − Y1 ||2 + oP (1).
2
2 d(2σ 2 + µ2 )
Similarly,
1√
2dσ 2 +
2
1√
||Y1 − Y2 || =
2dσ 2 +
2
||X1 − X2 || =
1
√
||X1 − X2 ||2 + oP (1)
2
2 2dσ
1
√
||Y1 − Y2 ||2 + oP (1).
2 2dσ 2
Therefore,
χ∗Energy (X1 , X2 ; Y1 , Y2 )
o
n
1
= p
||X1 − Y1 ||2 + ||X2 − Y1 ||2 + ||X1 − Y2 ||2 + ||X2 − Y2 ||2
4 d(2σ 2 + µ2 )
n
o
√
p
1
+ d(2σ 2 + µ2 ) − 2dσ 2 − √
||X1 − X2 ||2 + ||Y1 − Y2 ||2 + oP (1)
2 2dσ 2
n
o
1
= p
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 )
2 d(2σ 2 + µ2 )
b d + oP (1),
+R
where
√
p
b d = d(2σ 2 + µ2 ) − 2dσ 2 +
R
1
1
p
− √
d(2σ 2 + µ2 ) 2 2dσ 2
n
o
× ||X1 − X2 ||2 + ||Y1 − Y2 ||2 .
53
2
!
Note that
o
1 n
√ ||X1 − X2 ||2 + ||Y1 − Y2 ||2 − 4dσ 2 = OP (1),
d
and thus the remainder term becomes negligible under the assumption in (18), i.e.
!
√
p
√ 2
1
1
b d = d(2σ 2 + µ2 ) − 2dσ 2 + 4 dσ
p
R
− √
+ oP (1)
2 2σ 2 + µ2 2 2σ 2
p
√
√ (4σ 2 + µ2 ) 2σ 2 − 4σ 2 2σ 2 + µ2
p
= d
+ oP (1)
2 σ 4 + µ2 σ 2
√
√
√ 2
√
2σ 2 2
4)
d (4σ + µ2 ) 2σ 2 − 4σ 2 2σ 2 + 4σ
µ
+
O(µ
2
p
+ oP (1)
=
2 σ 4 + µ2 σ 2
√
= O( dµ4 ) + oP (1) = oP (1).
Furthermore, it is clear to see that
n
o
1
p
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 )
2 d(2σ 2 + µ2 )
n
o
1
= √
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 ) + oP (1).
2 2dσ 2
Finally, we have established
χ∗Energy (X1 , X2 ; Y1 , Y2 )
n
o
1
hCQ (X1 , X2 ; Y1 , Y2 ) + hCQ (X1 , X2 ; Y2 , Y1 ) + oP (1)
= √
2 2dσ 2
and after centering and scaling properly, we can obtain the result.
Finally, we collect the previous results and prove the main result in Theorem 5.1.
Main Proof of Theorem 5.1
Suppose X and Y satisfy the model assumptions in Theorem 5.1 with
µ=
d1/4
ψ
√
.
m+n
Then for every fixed m, n as d → ∞, followed by Lemma B.2, Lemma B.4 and Theorem 2.1
of Chakraborty and Chaudhuri (2017), we have
lim βCvM (δ) = lim βEnergy (δ) = lim βCQ (δ) = lim βW M W (δ),
d→∞
d→∞
d→∞
d→∞
and they converge to
Φ −zα +
ψ2
(m + n)γm,n σ 2
.
Lastly, we finish the proof by noting that
√
(m + n)γm,n →
2
π1 π2
54
as m, n → ∞.
B.11
Proof of Lemma 6.1
For given w ∈ Rd , it is seen that
Z
I(β > z ≤ β > w) − I(β > z 0 ≤ β > w) dλ(β)
(47)
Sd−1
Z
=
I(β > z ≤ β > w < β > z 0 ) + I(β > z 0 ≤ β > w < β > z)dλ(β)
Sd−1
)
)
(
(
1
1
1
(z − w)> (w − z 0 )
1
(z 0 − w)> (w − z)
=
+ −
−
arccos
arccos
2 2π
||z − w||||w − z 0 ||
2 2π
||z 0 − w||||w − z||
)
(
1
(z − w)> (w − z 0 )
= 1 − arccos
π
||z − w||||w − z 0 ||
(
)!
1
(z − w)> (w − z 0 )
=
π − arccos
π
||z − w||||w − z 0 ||
(
)
(z − w)> (z 0 − w)
(i) 1
:= ρAngle (z, z 0 ; w),
= arccos
π
||z − w||||z 0 − w||
where (i) is due to arccos(x) + arccos(−x) = π. Then ρAngle (z, z 0 ) is the expected value of
ρAngle (z, z 0 ; W∗ ) over W∗ ∼ 1/2FX + 1/2GY , i.e.
ρAngle (z, z 0 ) = EW∗ ρAngle (z, z 0 ; W∗ )
"
(
)#
(z − W∗ )> (z 0 − W∗ )
1
= EW∗ arccos
.
π
||z − W∗ ||||z 0 − W∗ ||
Now, if z = z 0 , it is trivial to see ρAngle (z, z 0 ) = 0. In addition, if ρAngle (z, z 0 ) = 0, then
we have z = z 0 . In order to show the second direction, note that arccos is a positive and
monotone decreasing function and so ρAngle (z, z 0 ) = 0 implies that
(z − W∗ )> (z 0 − W∗ )
= 1,
||z − W∗ ||||z 0 − W∗ ||
almost surely with respect to 1/2FX + 1/2GY . By the Cauchy-Schwarz, the inner product
becomes one if and only if (z − W∗ ) or (z 0 − W∗ ) is a multiple of the other. This is only
possible when z − W∗ = z 0 − W∗ almost surely, which implies z = z 0 . The symmetry property
follows easily by the definition of ρAngle . In addition, from triangle inequality, we have
Z
I(β > z ≤ β > w) − I(β > z 0 ≤ β > w) dλ(β)
Sd−1
Z
I(β > z ≤ β > w) − I(β > z 00 ≤ β > w) dλ(β)
≤
Sd−1
Z
+
I(β > z 00 ≤ β > w) − I(β > z 0 ≤ β > w) dλ(β),
Sd−1
55
and therefore by the equality in (47), we can establish
ρAngle (z, z 0 ; w) ≤ ρAngle (z, z 00 ; w) + ρAngle (z 0 , z 00 ; w).
Now, by taking the expectation over W∗ , we conclude that
ρAngle (z, z 0 ) ≤ ρAngle (z, z 00 ) + ρAngle (z 0 , z 00 ).
Next, we will show that for ∀n ≥ 2, z1 , . . . , zn ∈ S, and α1 , . . . , αn ∈ R, with
n X
n
X
Pn
i=1 αi
= 0,
αi αj ρAngle (zi , zj ) ≤ 0.
i=1 j=1
The result follows from Section 6 of Bogomolny et al. (2007) who showed that for each fixed
w,
n X
n
X
αi αj ρAngle (zi , zj ; w) ≤ 0,
(48)
i=1 j=1
P
for any α1 , . . . , αn ∈ R, with ni=1 αi = 0. Therefore, by taking the expected value over W∗
in (48), we conclude that ρAngle is of negative-type.
Regarding Remark 6.1, note that
Eν ρAngle (z, z 0 ; ν)
Z
Z
=
I(β > z ≤ ν < β > z 0 ) + I(β > z 0 ≤ ν < β > z)dνdλ(β)
Sd−1
(i)
Z
=
R
|β > z − z 0 |dλ(β)
Sd−1
(ii)
= γd ||z − z 0 ||,
where (i) and (ii) are due to Lemma 2.1 and Lemma 2.3 of Baringhaus and Franz (2004) and
√
π(d − 1)Γ ((d − 2)/2)
.
γd =
2Γ(d/2)
Therefore, the generalized angular distance with the Lebesgue measure corresponds to the
Euclidean distance.
B.12
Proof of Theorem 7.1
Given α ∈ Sp−1 , β ∈ Sq−1 , expand the square term to have
n
o2
4P α> (X1 − X2 ) < 0, β > (Y1 − Y2 ) < 0 − 1
h
= 16E I(α> (X1 − X2 ) < 0, α> (X3 − X4 ) < 0)
i
× I(β > (Y1 − Y2 ) < 0, β > (Y3 − Y4 ) < 0)
56
h
i
−8E I(α> (X1 − X2 ) < 0) × I(β > (Y1 − Y2 ) < 0) + 1.
By applying Lemma 2.2, the first term becomes
2
2
E 2 − Ψ (X1 − X2 , X3 − X4 )
2 − Ψ (Y1 − Y2 , Y3 − Y4 )
π
π
and the remainder terms become −1, which yields the expression.
B.13
Proof of Theorem 7.2
Given α ∈ Sp−1 and β ∈ Sq−1 ,
Z h
i2
Fα> X,β > Y (u, v) − Fα> X (u)Fβ > Y (v) dFα> X (u)dFβ > Y (v)
R2
h
= E I(α> (X1 − X3 ) ≤ 0, α> (X2 − X3 ) ≤ 0)
i
× I(β > (Y1 − Y4 ) ≤ 0, β > (Y2 − Y4 ) ≤ 0)
h
+ E I(α> (X1 − X5 ) ≤ 0, α> (X2 − X5 ) ≤ 0)
i
× I(β > (Y3 − Y6 ) ≤ 0, β > (Y4 − Y6 ) ≤ 0)
h
−2E I(α> (X1 − X4 ) ≤ 0, α> (X2 − X4 ) ≤ 0)
i
× I(β > (Y1 − Y5 ) ≤ 0, β > (Y3 − Y5 ) ≤ 0) .
Then apply Lemma 2.2 to obtain the expression.
B.14
Proof of Lemma 7.1
The proof is based on the inclusion–exclusion principle. For the sets A, B, C, the principle
presents
|A ∩ B ∩ C|
= |A ∪ B ∪ C| − |A| − |B| − |C| + |A ∩ B| + |A ∩ C| + |B ∩ C|
(49)
= |S| − |Ac ∩ B c ∩ C c | − |A| − |B| − |C| + |A ∩ B| + |A ∩ C| + |B ∩ C|,
where the second equality is due to De Morgan’s laws with the entire set S. For our purpose,
define the sets Aβ , Bβ , Cβ by
n
o
n
o
n
o
Aβ = β > U1 ≤ 0 , Bβ = β > U2 ≤ 0 and Cβ = β > U3 ≤ 0 .
Let β have the uniform distribution over Sd−1 . Since β and −β have the same distribution,
we have
Z
Z
dλ(β).
dλ(β) =
Acβ ∩Bβc ∩Cβc
Aβ ∩Bβ ∩Cβ
57
Furthermore, since
Z
dλ(β) =
dλ(β) =
dλ(β) =
Bβ
Aβ
Sβ
Z
Z
Z
dλ(β) = 1,
Bβ
1
2
and
dλ(β) =
1
1
−
Ψ (U1 , U2 ) ,
2 2π
dλ(β) =
1
1
−
Ψ (U1 , U3 ) ,
2 2π
dλ(β) =
1
1
−
Ψ (U2 , U3 ) ,
2 2π
Z
Aβ ∩Bβ
Z
Aβ ∩Cβ
Z
Bβ ∩Cβ
the identity of (49) presents
Z
I(β > U1 ≤ 0)I(β > U2 ≤ 0)I(β > U3 ≤ 0)dµ(β)
Sd−1
=
1
1
−
[Ψ (U1 , U2 ) + Ψ (U1 , U3 ) + Ψ (U2 , U3 )] ,
2 4π
which completes the proof.
B.15
Proof of Theorem 7.3
From Bergsma and Dassios (2014), the univariate τ ∗ can be written as
τ ∗ = 4P (X1 , X2 < X3 , X4 & Y1 , Y2 < Y3 , Y4 )
+ 4P (X1 , X2 < X3 , X4 & Y1 , Y2 > Y3 , Y4 )
− 8P (X1 , X2 < X3 , X4 & Y1 , Y3 < Y2 , Y4 ) ,
where Z1 , Z2 < Z3 , Z4 denotes max{Z1 , Z2 } < min{Z3 , Z4 }.
Note that
I(X1 , X2 < X3 , X4 )
= I(X1 < X2 < X3 < X4 ) + I(X2 < X1 < X3 < X4 )
+ I(X1 < X2 < X4 < X3 ) + I(X2 < X1 < X4 < X3 )
= I(X1 < X2 )I(X2 < X3 )I(X3 < X4 ) + I(X2 < X1 )I(X1 < X3 )I(X3 < X4 )
+ I(X1 < X2 )I(X2 < X4 )I(X4 < X3 ) + I(X2 < X1 )I(X1 < X4 )I(X4 < X3 ).
Similarly, we have
I(Y1 , Y2 < Y3 , Y4 )
= I(Y1 < Y2 )I(Y2 < Y3 )I(Y3 < Y4 ) + I(Y2 < Y1 )I(Y1 < Y3 )I(Y3 < Y4 )
58
+ I(Y1 < Y2 )I(Y2 < Y4 )I(Y4 < Y3 ) + I(Y2 < Y1 )I(Y1 < Y4 )I(Y4 < Y3 ).
Therefore, the product I(X1 , X2 < X3 , X4 )I(Y1 , Y2 < Y3 , Y4 ) can be expressed as the linear
combination of
I(Xi1 < Xi2 )I(Xi2 < Xi3 )I(Xi3 < Xi4 )I(Yj1 < Yj2 )I(Yj2 < Yj3 )I(Yj3 < Yj4 ).
Using Lemma 7.1,
Z
I(α> Xi1 < α> Xi2 )I(α> Xi2 < α> Xi3 )I(α> Xi3 < α> Xi4 )dλ(α)
Sp−1
1
1
−
[Ψ (U1 , U2 ) + Ψ (U1 , U3 ) + Ψ (U2 , U3 )] ,
2 4π
=
where U1 = Xi1 − Xi2 , U2 = Xi2 − Xi3 and U3 = Xi3 − Xi4 .
Similarly,
Z
Sq−1
=
I(β > Yj1 < β > Yj2 )I(β > Yj2 < β > Yj3 )I(β > Yj3 < β > Yj4 )dλ(β)
1
1
−
[Ψ (V1 , V2 ) + Ψ (V1 , V3 ) + Ψ (V2 , V3 )] ,
2 4π
where V1 = Yj1 − Yj2 , V2 = Yj2 − Yj3 and V3 = Yj3 − Yj4 .
As a result, we have
Z
Z
Sp−1
Sq−1
P(α> X1 , α> X2 < α> X3 , α> X4
& β > Y1 , β > Y2 < β > Y3 , β > Y4 )dλ(α)dλ(β)
= E [hp (X1 , X2 , X3 , X4 )hq (Y1 , Y2 , Y3 , Y4 )] .
Applying the same argument to the rests, we can obtain
∗
τp,q
=E [hp (X1 , X2 , X3 , X4 )hq (Y1 , Y2 , Y3 , Y4 )]
+E [hp (X1 , X2 , X3 , X4 )hq (Y3 , Y4 , Y1 , Y2 )]
−2E [hp (X1 , X2 , X3 , X4 )hq (Y1 , Y3 , Y2 , Y4 )] ,
which finishes the proof.
B.16
Proof of Theorem A.1
Let Zm,n = (Z1 , . . . , ZN ) = (X1 , . . . , Xm , Y1 , . . . , Yn ) and
TN (Zm,n ) = N (Um,n − θ),
where θ = E [Um,n ]. Also, let GN be the collection of all permutations of (1, . . . , N ) and let
GN = (π(1), . . . , π(N )) and G0N = (π 0 (1), . . . , π 0 (N )) be two independent random samples
uniformly chosen from GN . Denote TN (Zπ(1) , . . . , Zπ(N ) ) by TN (GN Zm,n ).
59
Our goal is to show
d
TN (GN Zm,n ), TN (G0N Zm,n ) −→ (T, T 0 )
where T, T 0 are independent and identically distributed with common distribution R(t). Then
the result follows by Theorem 3.2 of Hoeffding (1952). The proof consists of several pieces
and closely follows the proof of the limiting distribution of a degenerating U -statistic in Bhat
(1995); Serfling (2009). We start with the projection of the U -statistic.
Projection of the U -statistics.
Consider the projection of the degenerate U -statistic:
Ûm,n − θ
=
r(r − 1)
m(m − 1)
+
r2
mn
X
1≤i1 <i2 ≤m
m X
n
X
r(r − 1)
e
h2,0 (Xi1 , Xi2 ) +
n(n − 1)
X
e
h0,2 (Yj1 , Yj2 )
1≤j1 <j2 ≤n
e
h1,1 (Xi , Yj ),
i=1 j=1
where
e
h2,0 (Xi1 , Xi2 ) = h2,0 (Xi1 , Xi2 ) − θ,
e
h0,2 (Yj1 , Yj2 ) = h0,2 (Yj1 , Yj2 ) − θ
and
e
h1,1 (Xi , Yj ) = h1,1 (Xi , Yj ) − θ.
From Lemma 3.4.2 of Bhat (1995), the projection of the two-sample U -statistic is asymptotically the same as the original U -statistic:
N (Um,n − θ) = N (Ûm,n − θ) + oP (1).
Furthermore, under the assumption E h2 < ∞, we can write
e
h2,0 (X1 , X2 ) =
∞
X
(50)
λi φi (X1 )φi (X2 ),
i=1
e
h0,2 (Y1 , Y2 ) =
∞
X
γi ψi (Y1 )ψi (Y2 ),
i=1
e
h1,1 (X1 , X2 ) =
∞
X
αi φ∗i (X1 )ψi∗ (X2 ),
i=1
where {φi (·)}, {ψi (·)}, {φ∗ (·), ψ ∗ (·)} are orthonormal eigenfunctions and the corresponding
eigenvalues {λi }, {γi }, {αi }, associated with e
h2,0 , e
h0,2 and e
h1,1 , respectively (see e.g., Bhat,
1995, for details). From the given assumptions, the following is true:
φi (z) = ψi (z) = φ∗i (z) = ψi∗ (z)
γi = λi
and αi =
60
1−r
λi .
r
Therefore,
∞
X
X
1
N (Ûm,n − θ) = â1
λi φi (Xi1 )φi (Xi2 )
m
1≤i1 6=i2 ≤m i=1
∞
X X
1
+ â2
λj φj (Yj1 )φj (Yj2 )
n
1≤j1 6=j2 ≤n j=1
m X
n X
∞
X
1
+ â3 √
λk φk (Xi1 )φk (Yj1 )
mn
i1 =1 j1 =1 k=1
00
= â1 Tm + â2 Tn0 + â3 Tmn
,
where
â1 =
r(r − 1) N
,
2
m
â2 =
r(r − 1) N
2
n
N
.
and â3 = −r(r − 1) √
mn
Let
eN = N (Û GN − θ)
U
m,n
0
GN
e 0 = N (Ûm,n
and U
− θ),
N
which are the projections of two permutation statistics TN (GN Zm,n ) and TN (G0N Zm,n ), respectively. Then, based on (50), it is enough to show that
d
eN , U
e 0 −→
U
(T, T 0 ).
N
Truncation of the U -statistics.
Now, define a truncated version of N (Ûm,n − θ) by
K
X
X
1
λi φi (Xi1 )φi (Xi2 )
N (Ûm,n,K − θ) = â1
m
i=1
1≤i1 6=i2 ≤m
K
X X
1
λj φj (Yj1 )φj (Yj2 )
+ â2
n
1≤j1 6=j2 ≤n j=1
m X
n X
K
X
1
+ â3 √
λk φk (Xi1 )φk (Yj1 )
mn
i1 =1 j1 =1 k=1
0
00
= â1 TmK + â2 TnK
+ â3 TmnK
.
We may write
0
00
â1 TmK + â2 TnK
+ â3 TmnK
"K
#
"K
#
"K
#
X
X
X
2
02
0
0
=â1
λk Wkm − Vkm + â2
λk Wkn − Vkn + â3
λk Wkm Wkn
k=1
k=1
k=1
61
r(r − 1)
=
2
(
K
X
r
λk
k=1
N
Wkm −
m
r
N 0
W
n kn
!2
−
K
X
λk
k=1
N 0
N
Vkm + Vkn
m
n
)
,
where
m
1 X
Wkm = √
φk (Xi1 ),
m
i1 =1
Vkm =
m
1 X 2
φk (Xi1 ),
m
n
1 X
0
Wkn
=√
φk (Yj1 )
n
j1 =1
0
Vkn
=
i1 =1
n
1 X 2
φk (Yj1 ),
n
j1 =1
for k = 1, . . . , K.
By strong law of large numbers,
a.s.
∗>
0
0
Vmn
= (V1m , . . . , VKm , V1n
, . . . , VKn
)> −→ V ∗> = (V1 , . . . , VK , V10 , . . . , VK0 )>
and by the assumption that m/N → π1 , n/N → π2 ,
N (Ûm,n,K − θ)
( K
r(r − 1) X
=
λk
2
r
k=1
N
r(r − 1)
Wkm −
m
2
r
N 0
W
n kn
!2
K
1 X
−
λk
π1 π2
)
+ oP (1)
k=1
2
( K
)
m
n
K
X
X
X
X
1
1
1
r(r − 1)
N
λk
φk (Xi ) −
φk (Yj ) −
λk + oP (1)
=
2
m
n
π1 π2
i=1
k=1
( K
X
r(r − 1)
=
λk
N
2
k=1
N
X
j=1
!2
i φk (Zi )
i=1
k=1
K
1 X
λk
−
π1 π2
)
+ oP (1)
k=1
where
(1 , . . . , m , m+1 , . . . , m+n ) = (m−1 , . . . , m−1 , −n−1 , . . . , −n−1 ).
|
{z
} |
{z
}
m terms
n terms
Proving independence of the truncated U -statistics.
Consider the truncated permutation statistics
eN K = N (Û GN − θ)
U
m,n,K
( K
X
r(r − 1)
=
N
λk
2
k=1
N
X
!2
π(i) φk (Zi )
i=1
K
1 X
−
λk
π1 π2
)
+ oP (1)
k=1
G0N
0
eN
U
K = N (Ûm,n,K − θ)
( K
X
r(r − 1)
=
N
λk
2
k=1
N
X
!2
π0 (i) φk (Zi )
i=1
K
1 X
−
λk
π1 π2
k=1
62
)
+ oP (1).
Note that
Cov π(i) φk (Zi ), π0 (i) φk (Zi ) = E π(i) E π0 (i) E φ2k (Zi ) = 0,
because π(i) and π0 (i) are independent random variables having either 1/m or −1/n with
m/N and n/N probabilities. By the multivariate central limit theorem,
√
N
N
X
π(i) φ1 (Zi ), . . . ,
i=1
N
X
π(i) φK (Zi ),
i=1
N
X
π0 (i) φ1 (Zi ), . . . ,
i=1
N
X
!>
π0 (i) φK (Zi )
i=1
d
−→ N (0, π1 −1 π2 −1 I2K )
and thus the components are asymptotically independent to each other. Then by continuous
mapping theorem, we have
d
0
eN K , U
e 0 ) −→
(U
(TK , TK
)
NK
0 are independent and have the same distribution with
where TK and TK
K
r(r − 1) X
λk (Wk2 − 1).
2π1 π2
k=1
Bounding the difference between characteristic functions.
We will use the characteristic functions to show
d
0
eN , U
eN
U
−→ (T, T 0 ).
More specifically, we will show that for any x, y ∈ R and any > 0 and sufficiently large n, m,
i
i
h
h
0
e
e0
E ei(xUN +yUN ) − E ei(xT +yT ) ≤ A + B + C <
where
h
i
h
i
e
e0
e
e0
A = E ei(xUN +yUN ) − E ei(xUN K +yUN K ) ,
h
i
h
i
0
e
e0
B = E ei(xUN K +yUN K ) − E ei(xTK +yTK ) ,
h
i
h
i
0
0
C = E ei(xTK +yTK ) − E ei(xT +yT ) .
1. Bounding A.
Using the inequality |eiz | ≤ 1 and |eiz − 1| ≤ |z|, we bound A as
i
h
i
h
e
e0
e
e0
E ei(xUN +yUN ) − E ei(xUN K +yUN K ))
2 1/2
2 1/2
eN K − U
eN
e0 − U
e0
≤ |x| E U
+ |y| E U
NK
N
63
(
!1/2
∞
X
r(r − 1)
2π̂1
r(r − 1)
2
≤ (|x| + |y|)
λ2k
+
2π̂2
k=K+1
!1/2 )
∞
X
r(r − 1)
2
− √
λ
π̂1 π̂2 k=K+1 k
r(r − 1)
= (|x| + |y|) √
2
r(r − 1)
≤ (|x| + |y|) √
2π̂1 π̂2
∞
X
2
1
1
√ −√
π̂1
π̂2
∞
X
2
!1/2
λ2k
k=K+1
!1/2
λ2k
k=K+1
!1/2
∞
X
λ2k
k=K+1
where π̂1 = m/N and π̂2 = n/N .
Now, for fixed x and y and any given > 0, we choose K large enough to bound
∞
X
r(r − 1)
(|x| + |y|) √
2π1 π2
!1/2
λ2k
k=K+1
< .
3
(51)
Since π̂1 → π1 and π̂2 → π2 , for sufficiently large m, n, we have
r(r − 1)
A ≤ (|x| + |y|) √
2π̂1 π̂2
∞
X
k=K+1
!1/2
λ2k
< .
3
2. Bounding B.
By central limit theorem and continuous mapping theorem,
i
i
h
h
0
e
e0
E ei(xUN K +yUN K )) − E ei(xTK +yTK ) <
for all sufficiently large m, n.
3
3. Bounding C.
From Chapter 3 of Bhat (1995) with the given conditions on the kernel, the asymptotic
distribution of a degenerate U -statistic converges to
∞
∞
k=1
k=1
r(r − 1) X
r(r − 1) X
N (Um,n − θ) −→
λk (Wk2 − 1) +
λk (Wk02 − 1)
2π1
2π2
d
∞
(52)
r(r − 1) X
− √
λk Wk Wk0
π1 π2
k=1
where {Wk } and {Wk0 } are independent standard normal random variables and {λk } are
eigenvalues associated with the kernel. Note that the right-side of (52) can be re-written as
∞
√
r(r − 1) X √
λk ( π2 Wk − π1 Wk0 )2 − 1 ,
2π1 π2
k=1
64
where
√
π2 Wk −
√
π1 Wk0 ∼ N (0, 1). Therefore, T, T 0 are identically distributed as
∞
r(r − 1) X
λk (Wk2 − 1).
2π1 π2
k=1
0 have the same distribution with
Recall that TK , TK
K
r(r − 1) X
λk (Wk2 − 1).
2π1 π2
k=1
Consequently,
h
h
i1/2
h
i
h
i
2 i1/2
0
0
0
+ |y| E TK
− T0
E ei(xTK +yTK ) − E ei(xT +yT ) ≤ |x| E (TK − T )2
r(r − 1)
≤ (|x| + |y|) √
2π1 π2
∞
X
k=K+1
!1/2
λ2k
< ,
3
with the same choice of x, y, , K in (51).
Combining the bounds.
From the previous results, we conclude that for any x, y ∈ R and any > 0 with sufficiently
large m, n,
i
i
h
h
0
e
e0
E ei(xUN +yUN ) − E ei(xT +yT ) < ,
and therefore
d
0
eN , U
eN
U
−→ (T, T 0 ).
The last step is a simple application of Polya-Cantelli lemma.
C
Additional Simulations
In this section, we provide additional simulation results to support our claim in Section 5. The
simulation setting is the same as that in the first part of Section 8, but now the dimension is
set at d = 1, 000. We generate samples from X, Y ∼ N (µ(0) , Σ) under H0 , and X ∼ N (µ(0) , Σ)
and Y ∼ N (µ(1) , Σ) under H1 . Here, we provide scatter plots between the considered statistics
including the CvM, Energy and CQ statistics to demonstrate their asymptotic identities. The
results are presented in Figure 3 and 4. From the results, it is seen that pairwise observations
of (i) the CvM statistic vs. the CQ statistic or (ii) the Energy statistic vs. the CQ statistic
approximately lie on a linear line, meaning that after centering and scaling, the two statistics
are identical with small stochastic errors. These results confirm Lemma B.2 and Lemma B.4,
and further illustrate that the resulting tests have approximately the same power.
65
H0 , ⌃Band
H1 , ⌃Band
●
●
●
●
●
●●
●
●●
0.000
0.001
10 15 20 25
TCQ
5
●
●
●
●
●●
●
●
●
●
●
●●
0
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
−5
10
20
●
0
TCQ
30
●
0.002
●
●●
●
●
●
●
●
●●
●●
●
●
●●
●●
●●
●
●
●
●●
●●
●●
●
●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
−0.0005
0.003
0.0005
CvM
30
H1 , ⌃Auto
30
●
●
●
●●
●
10
0
TCQ
20
10
●
20
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●●
●●
●
●
●
●
●●
●
●
●●
0
TCQ
0.0025
●
●
●
−10
0.0015
CvM
H0 , ⌃Auto
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
−0.0005
0.0010
0.0025
0.000
CvM
0.002
0.003
H1 , ⌃Block
25
10 15 20 25
0.001
CvM
H0 , ⌃Block
●
●●
●
0
−5
5
10
15
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
0
●
●
●
●
●●
●●
20
●●
●●
TCQ
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
5
●
●
●
●
−5
TCQ
●
●
−0.0005
0.0005
0.0015
●
−0.0005
CvM
0.0005
0.0015
CvM
Figure 3: The scatter plots between the CvM statistic Um,n and the CQ statistic TCQ when
m = n = 20 and d = 1, 000. The red lines are based on the least square estimates.
66
H0 , ⌃Band
H1 , ⌃Band
●
●
●
●
●
0.0
0.2
0.4
10 15 20 25
0
5
●
●
●
●
●●
−5
●●
●
●●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
TCQ
10
20
●
0
TCQ
30
●
0.6
●
●●
●●
●
0.0
0.8
0.4
0.6
30
●
30
●
●
●
●●
10
0
TCQ
20
10
●
20
●
●●
●●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
0
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
−0.2
0.0
0.2
0.4
0.6
−0.2
●
0.2
0.4
0.6
Energy
H1 , ⌃Block
25
10 15 20 25
Energy
H0 , ⌃Block
0.0
●●
●
0
−5
20
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
15
TCQ
●
●
●
●
0
●
●
●
●●
●
●
●●
5
5
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
10
TCQ
0.2
Energy
H1 , ⌃Auto
●
●
●
−10
●
●●
●
●●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●●
●
●●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
Energy
H0 , ⌃Auto
TCQ
●
●
●
●
−5
●
●
−0.1
0.1
0.3
0.5
●
−0.1
Energy
0.1
0.3
0.5
Energy
Figure 4: The scatter plots between the Energy statistic Em,n and the CQ statistic TCQ when
m = n = 20 and d = 1, 000. The red lines are based on the least square estimates.
67
| 10math.ST
|
Enhanced Max-Min SINR for Uplink Cell-Free
Massive MIMO Systems
Manijeh Bashar∗, Kanapathippillai Cumanan∗, Alister G. Burr∗ , Mérouane Debbah † , and Hien Quoc Ngo‡
∗ Department
of Electronic Engineering, ,University of York, Heslington, York, UK
Networks and Systems Group, CentraleSupélec,
Université Paris-Saclay, France
,
‡ School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, Belfast, UK
Email: {mb1465, kanapathippillai.cumanan, alister.burr}@york.ac.uk, [email protected], [email protected]
arXiv:1801.10188v1 [cs.IT] 30 Jan 2018
† Large
Abstract—In this paper, we consider the max-min signal-tointerference plus noise ratio (SINR) problem for the uplink
transmission of a cell-free Massive multiple-input multipleoutput (MIMO) system. Assuming that the central processing
unit (CPU) and the users exploit only the knowledge of the
channel statistics, we first derive a closed-form expression for
uplink rate. In particular, we enhance (or maximize) user
fairness by solving the max-min optimization problem for user
rate, by power allocation and choice of receiver coefficients,
where the minimum uplink rate of the users is maximized with
available transmit power at the particular user. Based on the
derived closed-form expression for the uplink rate, we formulate
the original user max-min problem to design the optimal receiver
coefficients and user power allocations. However, this maxmin SINR problem is not jointly convex in terms of design
variables and therefore we decompose this original problem
into two sub-problems, namely, receiver coefficient design and
user power allocation. By iteratively solving these sub-problems,
we develop an iterative algorithm to obtain the optimal receiver coefficient and user power allocations. In particular, the
receiver coefficients design for a fixed user power allocation
is formulated as generalized eigenvalue problem whereas a
geometric programming (GP) approach is utilized to solve the
power allocation problem for a given set of receiver coefficients.
Numerical results confirm a three-fold increase in system rate
over existing schemes in the literature.
Keywords: Cell-free Massive MIMO, convex optimization,
max-min SINR problem, geometric programming, generalized
eigenvalue problem.
I. I NTRODUCTION
Massive multiple-input multiple-output (MIMO) is one
of the most promising techniques for 5th Generation (5G)
networks due to its potential for significant rate enhancement
and spectral as well as energy efficiency [1]–[6]. In cell-free
Massive MIMO, randomly distributed access points (APs)
jointly serve distributed users. In this paper, we propose
a max-min signal-to-interference plus noise ratio (SINR)
approach for an uplink cell-free Massive MIMO system. In
[7], the authors investigate the problem of max–min SINR in a
single-cell Massive MIMO system. A similar max-min SINR
problem refereed to SINR balancing has been considered for
cognitive radio network in [8]–[12]. In [13], the same maxmin SINR problem is considered through appropriate user
The work of K. Cumanan and A. G. Burr was supported by H2020MSCA-RISE-2015 under grant number 690750. The work on which this
paper is based was carried out in collaboration with COST Action CA15104
(IRACON).
power allocation where a bisection search method is utilized
to determine the optimal solution. However, a novel approach
to significantly improve all users’ performance is proposed
in this paper by designing optimal receiver coefficients and
user power allocation. By employing maximal ratio combining (MRC) at the receiver, we first derive a closed-form
expression for the average uplink rate of the users. Based
on these user power allocations and receiver coefficients, we
formulate the corresponding max-min SINR problem, which
is not jointly convex in terms of the design parameters.
In order to realize a solution for this non-convex problem,
we decompose the original problem into two sub-problems:
receiver coefficient design and user power allocation. An
iterative algorithm is proposed that successively solves these
two sub-problems while one of the design variables (i.e.,
user power allocation or receiver coefficients) is fixed. The
receiver coefficient design is formulated into a generalized
eigenvalue problem [14] whereas a geometric programming
(GP) approach [15] is exploited to solve the user power
allocation problem. The performance of the proposed scheme
in terms of the user rate is significantly higher than that of the
scheme proposed in [13]. The contributions and the results of
our work are summarized as follows: 1) For the considered
cell-free Massive MIMO system, we derive the average user
rate in the uplink. 2) Based on the derived user rate, we
propose a novel max-min SINR approach to significantly
improve the SINR performance in terms of the achieved
user rate. The original max-min problem formulation is not
convex and therefore we decompose the original problem into
two sub-problems and propose an iterative algorithm to yield
the optimal solution. 3) The user power allocation and the
receiver coefficient design sub-problems are solved through
the GP approach and the generalized eigenvalue problem,
respectively. 4) Numerical results are provided to validate the
superiority of the proposed algorithm in comparison with the
scheme proposed in [13].
II. SYSTEM MODEL
We consider uplink transmission in a cell-free Massive
MIMO system with M randomly distributed single-antenna
APs and K randomly distributed single-antenna users in the
area, as shown in Fig. 1. The channel coefficient between
at the kth user. The receivedP
signal at the mth AP from all
√
√
K
users is given by ym = ρ k=1 gmk qk sk + nm , where
nm ∼ CN (0, 1) is the noise at the mth AP. In addition, MRC
is employed at the APs. More precisely, the received signal
∗
at the mth AP, ym , is first multiplied with ĝmk
. The resulting
∗
ĝmk ym is then forwarded to the CPU for signal detection.
In order to improve achievable rate, the forwarded signal is
further multiplied by a receiver filter coefficient at the CPU.
The aggregated received signal at the CPU can be written as
Figure 1. The uplink of a cell-free Massive MIMO system with K users
and M APs. The dashed lines denote the uplink channels and the solid lines
present the backhaul links from the APs to the central processing unit (CPU).
M
X
rk =
∗
umk ĝmk
ym
(2)
m=1
M
K
M
X
√ X X
√
∗
∗
′
′
′
= ρ
umk ĝmk
nm .
umk ĝmk gmk qk sk +
m=1
k′ =1 m=1
the kth√user and the mth AP, gmk , is modeled as [13]
gmk = βmk hmk , where βmk denotes the large-scale fading
and hmk ∼ CN (0, 1) represents small-scale fading between
the kth user and the mth AP.
A. Uplink Channel Estimation
In order to estimate the channel coefficients in the uplink,
the APs employ an minimum mean square error (MMSE)
estimator. All pilot sequences used in the channel estimation
phase are collected in a matrix Φ ∈ Cτ ×K , where τ is the
length of the pilot sequence for each user and the kth column,
φk , represents the pilot sequence used for the kth user. After
performing a de-spreading operation, the MMSE estimate of
the channel coefficient between the kth user and the mth AP
is given by [13]
K
X
√
√
H
ĝmk = cmk τ pp gmk + τ pp
gmk′ φ H
k φ k′ +φ k np,m , (1)
k′ 6=k
where each element of np,m , np,m ∼ CN (0, 1), denotes
the noise at the mth antenna, pp represents the normalized
signal-to-noise ratio (SNR) of each pilot sequence (which
we define √in Section V), and cmk is given by cmk =
τ pp βmk
. The estimated channels in
PK
2
φ
+
1
τ pp k′ =1 βmk′ φ H
′
k
k
(1) are used by the APs to design the receiver coefficients and
determine power allocations at users to maintain user fairness.
In this paper, we investigate the cases of both random pilot
assignment and orthogonal pilots in cell-free Massive MIMO.
Here the term ”orthogonal pilots” refers to the case where
unique orthogonal pilots are assigned to all users, while in
”random pilot assignment” each user is randomly assigned a
pilot sequence from a set of orthogonal sequences of length
τ (< K), following the approach of [13], [16].
B. Uplink Data Transmission
In this subsection, we consider the uplink data transmission, where all users send their signals to the APs.
The transmitted signal from the kth user is represented by
√
xk = qk sk , where sk (E{|sk |2 } = 1) and qk denote
respectively the transmitted symbol and the transmit power
By collecting all the coefficients umk , ∀ m corresponding
to the kth user, we define uk = [u1k , u2k , · · · , uMk ]T and
without loss of generality, it is assumed that k uk k= 1. The
optimal solution of uk , qk , ∀ k for the considered max-min
SINR approach is investigated in Section IV.
III. P ERFORMANCE A NALYSIS
In this section, we derive the average user rate for the considered system model in the previous section by following a
similar approach to that in [13]. Note that the main difference
between the proposed approach and the scheme in [13] is
the new set of receiver coefficients which are introduced at
the CPU to improve the achievable user rates. The benefits
of the proposed approach in terms of achieved user uplink
rate is demonstrated through numerical simulation results in
Section V. In deriving the achievable rates of each user, it
is assumed that the CPU exploits only the knowledge of
channel statistics between the users and APs in detecting data
from the received signal in (3). Without loss of generality, the
aggregated received signal in (3) can be written as
√
= ρE
rk
(
|
M
X
√
∗
umk ĝmk
gmk qk
m=1
{z
)
sk
(3)
}
DSk
)!
(M
M
X
√ X
√
√
∗
∗
+ ρ
umk ĝmk gmk qk sk
umk ĝmkgmk qk −E
m=1
m=1
+
|
{z
}
BUk
M
K
M
X
X
√ X
√
∗
∗
umk ĝmk
nm ,
ρ
umk ĝmk
gmk′ qk′ sk′ +
k′ 6=k
m=1
m=1
|
{z
IUIkk′
}
|
{z
TNk
}
where DSk and BUk denote the desired signal (DS) and
beamforming uncertainty (BU) for the kth user, respectively,
and IUIk represents the inter-user-interference (IUI) caused
by the k ′ th user. In addition, TNk accounts for the total noise
(TN) following the MRC detection. The average SINR of
the received signal in (3) can be defined by considering the
worst-case of the uncorrelated Gaussian noise as follows [13]:
SINRUP
k =
E{|BUk |2}+
PK
|DSk |2
k′ 6=k
. (4)
E{|IUIkk′ |2}+E{|TNk |2}
Based on the SINR definition in (4), the achievable uplink
rate of the kth user is defined in the following theorem.
Theorem 1. By employing MRC detection at APs, the achievable uplink rate of the kth user in the Cell-free Massive
MIMO system with K randomly distributed single-antenna
users and M single-antenna APs is given by (5) (defined at
the beginning of the next page).
Algorithm 1 Proposed Algorithm to Solve P1
(0)
(0)
(0)
1. Initialize q(0) = [q1 , q2 , · · · , qK ], i = 1
2. Repeat
3. i = i + 1
4. Set q(i) = q(i−1) and find the optimal receiver coefficients
(i) (i)
(i)
U(i) = [u1 , u2 , · · · , uK ] through solving the generalized
eigenvalue Problem P2 in (7)
5. Compute q(i+1) through solving Problem P4 in (9).
6. Go back to Step 3 and repeat until required accuracy.
s.t. k uk k= 1, ∀ k.
(7)
Note
that
in
(5),
we
have
Γk
=
[γ1k , γ2k , · · · , γMk ]T , uk
=
[u1k , u2k , · · · , uMk ]T ,
γ1k β1k′ γ2k β2k′
γMk βMk′ T
=
[
∆kk′
,
,··· ,
] ,Rk
=
β1k
β2k
βMk
=
diag [γ1k , γ2k , · · · , γMk ],
and
Dkk′
diag [β1k′ γ1k , β2k′ γ2k , · · · , βMk′ γMk ].
Problem P2 is a generalized eigenvalue problem [14],
where the optimal solutions can be obtained by determining
the generalized
of the matrix P
pair Ak = qk Γk ΓH
k
PK eigenvalue
H
1
2
H
φk φ k′ | ∆kk′∆kk′+ K
and Bk = k′ 6=k qk′ |φ
k′ =1 qk′ Dkk′+ρRk
corresponding to the maximum generalized eigenvalue.
Proof: Please refer to Appendix A.
B. Power Allocation
IV. P ROPOSED M AX -M IN SINR S CHEME
In this section, we formulate the umax-min SINR problem
in cell-free massive MIMO, where the minimum uplink user
rate between users is maximized while satisfying the transmit
power constraint at each user. This max-min rate problem can
be formulated as follows:
P1 : max
qk ,uk
s.t.
min
k=1,··· ,K
Rk
(6)
In this subsection, we solve the power allocation problem
for a set of fixed receiver coefficients. The power allocation
problem can be formulated into the following max-min problem:
min
SINRk
(8)
P3 : max
qk
s.t.
P4 :
(k)
pmax
A. Receiver Coefficients Design
In this subsection, we solve the receiver coefficient design
problem to maximize the uplink rate of each user for a
given set of transmit power allocation at all users. These
coefficients (i.e., uk , ∀ k) can be obtained by interdependently maximizing the uplink SINR of each user. Hence, the
optimal coefficients for all users for a given set of transmit
power allocation can be determined by solving the following
optimization problem:
0 ≤ qk ≤ p(k)
max , ∀ k,
Without loss of generality, Problem P3 can be rewritten by
introducing a new slack variable as
||uk || = 1, ∀ k, 0 ≤ qk ≤ p(k)
max , ∀ k,
where
is the maximum transmit power available at
user k. Problem P1 is not jointly convex in terms of uk and
power allocation qk , ∀ k. Therefore, this problem cannot be
directly solved through existing convex optimization software.
To tackle this non-convexity issue, we divide the original
Problem P1 into two sub-problems: receiver coefficient design
(i.e. uk ) and the power allocation problem. To obtain a
solution for Problem P1 , these sub-problems are alternately
solved as explained in the following subsections.
k=1,··· ,K
s.t.
max
t,qk
t
(9)
0≤ qk ≤ p(k)
max , ∀ k, SINRk ≥ t, ∀ k.
Proposition 1: Problem P4 can be formulated into a GP.
Proof: Please refer to Appendix B.
Therefore, this problem can be efficiently solved through existing convex optimization software. Based on these two subproblems, an iterative algorithm is developed by alternately
solving each sub-problem in each iteration. The proposed
algorithm is summarized in Algorithm 1.
V. C ONVERGENCE
ANALYSIS
In this section, the convergence analysis of the proposed
Algorithm 1 is provided. Two sub-problems are alternately
solved to determine the solution to Problem P1 . At each
iteration, one of the design parameters is determined by
solving the corresponding sub-problem while other design
variable is fixed. Note that each sub-problem provides an
optimal solution for the other given design variable. At
(i)
the ith iteration, the receiver filter coefficients uk , ∀k are
(i)
determined for a given power allocation q and similarly, the
P2 :
power
allocation q(i+1) is updated for a given set of receiver
H
H
(i)
uk qk Γk Γk uk
P
, filter coefficients uk , ∀k. The optimal power allocation
max
P
(i)
2
K
K
uk
1
H
φH
q(i+1) obtained for a given uk achieves an uplink rate greater
uH
k φ k′ | ∆kk′∆kk′ +
k
k′ =1 qk′ Dkk′ + ρRk uk
k′ 6=k qk′ |φ
Rk = log2
1 +
2
∆kk′ ∆H
kk′ +
PK
k′ =1 qk′ Dkk′
1
The scheme
in [4]
0.9
Cumulative distribution
PK
H
H
uk
k′ 6=k qk′ φ k φ k′
qk Γk ΓH
uk
k
0.8
0.7
Proposed
scheme
0.6
0.5
0.4
Orthogonal pilots
Random pilot
assignment, =10
Random pilot
assignment, =5
0.3
0.2
0.1
1
+ Rk uk
ρ
.
(5)
0.8
0.7
Proposed
scheme
0.6
0.5
0.4
0.3
Orthogonal pilots
Random pilot,
assignment, =20
0.2
0.1
0
The scheme
in [4]
0.9
Cumulative distribution
1
uH
k
0
0
0.5
1
1.5
2
2.5
3
0.5
Min-user uplink rate (bits/s/Hz)
than or equal to that of the previous iteration. In addition, the
power allocation q(i) is also a feasible solution in determin(i+1)
ing q(i+1) as the receiver filter coefficients uk
, ∀k are
(i)
determined for a given q . This reveals that the achieved
uplink rate monotonically increases with each iteration, which
can be also observed from the simulation results presented
in Fig. 4. As the achievable uplink max-min rate is upper
bounded by a certain value for a given set of per-user power
constraints, the proposed algorithm converges to a particular
solution. Fortunately, the proposed Algorithm 1 converges to
the optimal solution, as we will prove by establishing the
uplink-downlink duality in the following section.
1.5
2
2.5
3
Figure 3. The cumulative distribution of the per-user uplink rate, with
random pilots for M = 100, K = 40, τ = 20, and D = 1 km.
1.5
Min-user uplink rate (bits/s/Hz)
Figure 2. The cumulative distribution of the per-user uplink rate, with
orthogonal and random pilots for M = 60, K = 20 and D = 1 km.
1
Min-user uplink rate (bits/s/Hz)
1.4
1.3
1.2
1.1
Channel 1
Channel 2
Channel 3
Channel 4
Channel 5
1
0.9
0.8
1
2
3
4
5
Number of iterations
VI. N UMERICAL R ESULTS AND D ISCUSSION
Figure 4. The convergence of the proposed max-min SINR approach for
M = 60, K = 20, τ = 10, and D = 1 km.
In this section, we provide numerical simulation results to
validate the performance of the proposed max-min SINR approach with different parameters. A cell-free Massive MIMO
system with M APs and K single antenna users is considered
in a D × D simulation area, where both APs and users are
randomly distributed. In the following subsections, we define
the simulation parameters and then present the corresponding
simulation results. To model the channel coefficients between
users andσ APs,
the coefficient βmk is given by βmk =
sh zmk
where PLσmk zis the path loss from the kth
PLmk .10 10
sh mk
user to the mth AP, and 10 10
denotes the shadow fading
with standard deviation σsh , and zmk ∼ N (0, 1) [13]. The
noise power is given by Pn = BWkB T0 W, where BW = 20
MHz denotes the bandwidth, kB = 1.381 × 10−23 represents
the Boltzmann constant, and T0 = 290 (Kelvin) denotes the
noise temperature. Moreover, W = 9dB, and denotes the
noise figure [13]. It is assumed that that P̄p and ρ̄ denote
the pilot sequence and the uplink data, respectively, where
P̄
Pp = Pnp and ρ = Pρ̄n . In simulations, we set P̄p = 100mW
and ρ̄ = 100mW. Similar to [13], we suppose the simulation
area is wrapped around at the edges which can simulate an
area without boundaries. Hence, the square simulation area
has eight neighbors. We evaluate the rate of the system over
300 random realizations of the locations of APs, users and
shadowing. In this subsection, we investigate the effect of the
max-min SINR problem on the system performance. Fig. 2
compares the cumulative distribution of the achievable uplink
rates for our proposed algorithm with the power allocation
scheme in [13], for three cases of orthogonal pilots, random
pilots with τ = 10 and τ = 5 for the length of pilot sequence.
In Fig. 2, M = 60 APs and K = 20 users are randomly
distributed through the simulation area of size 1 × 1 km2 . As
the figure shows, the performance of the proposed scheme is
almost three times than that of the scheme in [13].
In Fig. 3, we compare the performance of the proposed
max-min SINR approach with the scheme in [13] for the
case of M = 100 APs, K = 40 users and τ = 20 as the
length of the pilot sequence. Fig. 3 shows the superiority
of the proposed iterative algorithm over the power allocation
scheme in [13]. Moreover, Fig. 3 demonstrates that the rate of
the proposed max-min SINR approach is more concentrated
around the median. Fig. 4 investigates the convergence of
the proposed max-min SINR algorithm for a set of different
channel realizations. The figure shows that the proposed
algorithm converges after a few iterations, while the minimum
rate of the users increases with the iteration number.
VII. C ONCLUSIONS
We have considered the max-min optimization problem in
cell-free Massive MIMO systems, and propose an efficient
solution that maximizes the smallest of the uplink rate of the
users. We propose to divide the original max-min problem
into two sub-problems which can be iteratively solved by
exploiting generalized eigenvalue problem and GP. The simulation results showcased the effectiveness of the proposed
scheme in terms of maximising the smallest of the uplink
rate of the users compared with existing schemes.
A PPENDIX A: P ROOF OF T HEOREM 1
The desired signal for the user k is given by
(M
)
M
X
√
√ X
∗
DSk = E
umk ĝmk gmk qk = qk
umkγmk . (10)
m=1
2
m=1
Hence, |DSk | = qk
2
P
M
m=1
umk γmk
2
− ρE
m=1
m=1
where the last equality comes from the analysis in [13, Ap2
pendix A]] and using the following fact; γmk = E{|ĝmk | } =
√
2
τ pp βmk cmk . The term E{|IUIkk′ | } is obtained as
2
M
X
√
∗
umk ĝmk
gmk′ qk′
E {| IUIkk′ |2 } = ρE
m=1
( M
X
√
= ρE
cmk umk gmk′ qk′
√
τ pp
K
X
φH
gmiφ H
k φ i +φ
k np,m
i=1
= ρ qk′ E
|
M
X
cmk umk gmk′ ñ∗mk
m=1
{z
A
M
X
+ ρ τ pp E qk′
cmk umk gmk′
m=1
{z
|
B
|
m=1
2
cmk umk |gmk′ |2 φ H
k φ k′
{z
!∗ 2
2
}
K
X
i=1
(13)
}
C
M
X
∗ 2
.
cmk umk gmk′
gmiφ H
φ
+ τ pp qk′ E
k i
m=1
i6=k′
{z
}
|
K
X
D
The first term in (13) is given by
C = τ pp qk′ E
M
X
m=1
= τ pp qk′ φ H
k φ k′
2
2
cmk umk |gmk′ | φ H
k φ k′
2
M
X
2
c2mk u2mk βmk
′
m=1
M
X
βmk′
umk γmk
βmk
m=1
2
(11)
) 2
M
X
√
∗
umk ĝmk
gmk qk
= ρqk
u2mk γmk βmk ,
m=1
+ qk′ φ H
k φ k′
m=1
M
X
B = τ pp qk′ E
M
X
. Moreover, the term
E{|BUk | } can be obtained as
( M
o
n
X
√
2
∗
E |BUk | = ρE
umk ĝmk
gmk qk
(
where the third equality in (12) is due to the fact that for two
independent random variables X and Y and E{X} = 0, we
have E{|X + Y |2 } = E{|X|2 } + E{|Y |2 } [13, Appendix A].
Since ñmk = φ H
k np,m ∼ CN (0, 1) is independent from the
term gmk′ similar to [13, AppendixPA], the term A in (12)
2
2
immediately is given by A = qk′ M
m=1 cmk umk βmk′ .The
term B in (12) can be obtained as
!2
,
(14)
where the last equality is derived based on the fact γmk =
√
τ pp βmk cmk . The second term in (13) can be obtained as
M
X
∗ 2
H
cmk umk gmk′
gmiφ k φ i
D = τ pp qk′ E
m=1
i6=k′
= τ pp
M X
K
X
K
X
2
qk′ c2mk u2mk βmk′ βmi φ H
k φi .
(15)
m=1 i6=k′
Hence, (12) can be written as
M
o
n
X
2
c2mk u2mk βmk′
E |IUIkk′ | = qk′
(16)
m=1
(12)
+ τ pp qk′ φ H
k φ k′
|
2
{z
C1
M
X
2
c2mk u2mk βmk
′
}
m=1
!∗ 2
gmiφ H
k φi
}
+ τ pp qk′
M X
K
X
c2mk u2mk βmk′ βmi φ H
k φi
2
m=1 i6=k′
,
+ qk′
2
φH
k φ k′
M
X
βmk′
umk γmk
βmk
m=1
!2
,
(17)
and
C2 = τ pp qk′ φ H
k φ k′
2
M
X
2
c2mk u2mk βmk
′
m=1
+ τ pp qk′
M X
M
X
c2mk u2mk βmk′ βmi φ H
k φi
2
. (18)
m=1 i6=k′
|
{z
For the last term of (18), we have
C3 = τ pp qk′
M X
K
X
c2mk u2mk βmk′ βmi φ H
k φi
2
m=1 i6=k′
=
−
M
M
X
X
√
τ pp qk′
u2mk cmk βmk′ βmk − qk′
u2mk c2mk βmk′
m=1
M
X
τ pp qk′
u2mk c2mk βmk′
m=1
m=1
2
φH
k φ k′
,
(19)
where in the last step,Pwe used equation (II-A). As a result,
√
2
C1 + C2 = τ pp qk′ M
m=1 umk cmk βmk′ βmk . Then finally
we have
!
M
o
n
X
2
2
(20)
umk βmk′ γmk
= ρqk′
E |IUIkk′ |
m=1
+ ρqk′
M
X
βmk′
umk γmk
βmk
m=1
2
φH
k φ k′
!2
.
The total noise for the user k is given by
2
M
M
X
X
o
n
2
∗
umk ĝmk nm
E |TNk | = E
=
u2mk γmk ,(21)
m=1
m=1
where the last equality is due to the fact that the terms ĝmk
and nm are uncorrelated. Finally, by substituting (VII), (12),
(21) and (21) into (4), SINR of kth user is obtained by (5),
which completes the proof of Theorem 1.
A PPENDIX B: P ROOF
OF
P ROPOSITION 1
The standard form of GP is defined as follows [15]:
P5 : min
s.t.
f0 (x)
(22)
fi (x) ≤ 1, i = 1, · · · , m,
gi (x) = 1, i = 1, · · · , p,
where f0 and fi are posynomial and gi are monomial functions. Moreover, x = {x1 , · · · , xn } represent the optimization
variables. The SINR constraint in (9) is not a posynomial
function in its form, however it can be rewritten into the
following posynomial function:
uH
k
P
K
k′ 6=k qk
k 6=k
where a
}
C3
By applying a simple transformation, (23) is equivalent to the
following inequality:
K
K
X
X
1
−1
qk
(24)
bkk′ qk′ + ck < ,
akk′ qk′ +
t
′
′
PK
H
2
1
′
′
|φ H
k φ k′ | ∆kk′∆kk′ +
1
k′ =1 qk Dkk + pRk uk
< ,
H
t
uH
q
Γ
Γ
u
k k k
k
k
′
∀k. (23)
k =1
H
2
H
uH
k (|φ k φ k′ | ∆kk′∆kk′ )uk
kk′
H )u
uH
(Γ
Γ
k
k
k
k
uH
k Rk u k
H
H
puk (Γk Γk )uk
=
, bkk′ =
uH
k Dkk′ uk
,
H
uH
k (Γk Γk )uk
and ck =
. The transformation in (24) shows that
the left-hand side of (23) is a polynomial function. Therefore,
the power allocation Problem P4 is a standard GP (convex
problem), where the objective function and constraints are
monomial and polynomials, which completes the proof of
Proposition 1.
R EFERENCES
[1] A. Zappone, L. Sanguinetti, G. Bacci, E. A. Jorswieck, and M. Debbah,
“Energy-efficient power control: a look at 5G wireless technologies,”
IEEE Trans. Signal Process., vol. 64, no. 7, pp. 1668–1683, Apr. 2016.
[2] E. Björnson, E. A. Jorswieck, M. Debbah, and B. Ottersten, “Multiobjective signal processing optimization: the way to balance conflicting
metrics in 5G systems,” IEEE Signal Process. Mag., vol. 31, no. 6, pp.
14–23, Oct. 2014.
[3] A. G. Burr, M. Bashar, and D. Maryopi, “Cooperative access networks:
Optimum fronthaul quantization in distributed Massive MIMO and
cloud RAN,” in Proc. IEEE VTC, Jun. 2018, pp. 1–5.
[4] M. Bashar, K. Cumanan, A. G. Burr, H. Q. Ngo, and M. Debbah,
“Cell-free Massive MIMO with limited backhaul,” in Proc. IEEE ICC,
May 2018, pp. 1–7.
[5] M. Bashar, A. G. Burr, K. Haneda, and K. Cumanan, “Robust user
scheduling with COST 2100 channel model for Massive MIMO networks,” Accepted to publish in IET microwave antenna and propagation, Feb. 2018.
[6] Y. Lejosne, M. Bashar, D. Slock, and Y. Yuan-Wu, “From MU Massive
MISO to pathwise MU Massive MIMO,” in Proc. IEEE SPAWC, Jun.
2014, pp. 16–20.
[7] H. Sifaou, A. Kammoun, L. Sanguinetti, M. Debbah, and M. Alouini,
“Max–min SINR in large-scale single-cell MU-MIMO: asymptotic
analysis and low-complexity transceivers,” IEEE Trans. Signal Process., vol. 65, no. 7, pp. 1841–1854, Apr. 2017.
[8] K. Cumanan, R. Krishna, L. Musavian, and S. Lambotharan, “Joint
beamforming and user maximization techniques for techniques for
cognitive radio networks based on branch and bound method,” IEEE
Trans. Wireless Commun., vol. 9, no. 10, pp. 3082–3092, Oct. 2010.
[9] K. Cumanan, L. Musavian, S. Lambotharan, and A. B. Gershman,
“SINR balancing technique for downlink beamforming in cognitive
radio networks,” IEEE Signal Process. Lett., vol. 17, no. 2, pp. 133–
136, Feb. 2010.
[10] K. Cumanan, R. Krishna, Z. Xiong, and S. Lambotharan, “SINR
balancing technique and its comparison to semidefinite programming
based QoS provision for cognitive radios,” in Proc. IEEE VTC, Apr.
2009, pp. 1–5.
[11] K. Cumanan, Y. Rahulamathavan, S. Lambotharan, and Z. Ding, “A
mixed quality of service based linear transceiver design for a multiuser
MIMO network with linear transmit covariance constraints,” in Proc.
IEEE WCNC, Apr. 2013, pp. 3895–3899.
[12] K. Cumanan, J. Tang, and S. Lambotharan, “Rate balancing based linear
transceiver design for multiuser MIMO system with multiple linear
transmit covariance constraints,” in Proc. IEEE ICC, Jun. 2011, pp.
1–5.
[13] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
“Cell-free Massive MIMO versus small cells,” IEEE Trans. Wireless
Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017.
[14] G. Golub and C. V. Loan, Matrix Computations, 2nd ed. Baltimore,
MD: The Johns Hopkins Univ. Press, 1996.
[15] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, UK:
Cambridge University Press, 2004.
[16] H. Ahmadi, A. Farhang, N. Marchetti, and A. MacKenzie, “A game theoretic approach for pilot contamination avoidance in Massive MIMO,”
IEEE Wireless Commun. Lett., vol. 5, no. 1, pp. 12–15, Feb. 2016.
| 7cs.IT
|
Complexity and algorithms for finding
a perfect phylogeny from mixed tumor samples∗
Ademir Hujdurovića,b†
Urša Kačarc‡
Martin Milaniča,b§
Bernard Riesd¶
arXiv:1506.07675v4 [q-bio.PE] 7 Jul 2016
Alexandru I. Tomescuek
February 26, 2018
a
University of Primorska, UP IAM, Muzejski trg 2, SI 6000 Koper, Slovenia
University of Primorska, UP FAMNIT, Glagoljaška 8, SI 6000 Koper, Slovenia
c Celtra Inc., Ljubljana, Slovenia
University of Fribourg, Department of Informatics Decision Support and Operations Research
Group, Fribourg, Switzerland
e Helsinki Institute for Information Technology HIIT, Department of Computer Science,
University of Helsinki, Finland
b
d
Abstract
Recently, Hajirasouliha and Raphael (WABI 2014) proposed a model for deconvoluting mixed
tumor samples measured from a collection of high-throughput sequencing reads. This is related to understanding tumor evolution and critical cancer mutations. In short, their formulation asks to split each row of a binary matrix so that the resulting matrix corresponds to a
perfect phylogeny and has the minimum number of rows among all matrices with this property. In this paper we disprove several claims about this problem, including an NP-hardness
proof of it. However, we show that the problem is indeed NP-hard, by providing a different
proof. We also prove NP-completeness of a variant of this problem proposed in the same paper.
On the positive side, we propose an efficient (though not necessarily optimal) heuristic algorithm based on coloring co-comparability graphs, and a polynomial time algorithm for solving
the problem optimally on matrix instances in which no column is contained in both columns
of a pair of conflicting columns. Implementations of these algorithms are freely available at
https://github.com/alexandrutomescu/MixedPerfectPhylogeny.
Keywords: perfect phylogeny, graph coloring, NP-hard problem, heuristic algorithm
∗
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice,
after which this version may no longer be accessible.
†
E-mail: [email protected]
‡
E-mail: [email protected]
§
E-mail: [email protected]
¶
E-mail: [email protected]
k
E-mail: [email protected]
1
1
Introduction
Tumor progression is assumed to follow a phylogenetic evolution in which each tumor cell passes its
somatic mutations to its daughter cells as it divides, with new mutations being accumulated over
time. It is important to discover what tumor types are present in the sample, at what evolutionary
stage the tumor is in, or what are the “founder” mutations of the tumor, mutations that trigger
an uncontrollable growth of the tumor. These can lead to better understanding of cancer [2, 23],
better diagnosis, and more targeted therapies [22].
DNA sequencing is one method for discovering the somatic mutations present in each tumor
sample. The most accurate possible observation would come from sampling and sequencing every
single cell. However, because of single-cell sequencing limitations, and the sheer number of tumor
cells, one usually samples populations of cells. Even though the samples are taken spatially and
morphologically apart, they can still contain millions of different cancer cells. Moreover, this
mixing is not consistent across different collections of samples. Therefore, studying only these
mixed samples poses a serious challenge to understanding tumors, their evolution, or their founding
mutations.
Solutions for overcoming this limitation can come from a computational approach, as one could
deconvolute each sample by exploiting some properties of the tumor progression. One common
assumption is that all mutations in the parent cells are passed to the descendants. Another one,
called the “infinite sites assumption”, postulates that once a mutation occurs at a particular site, it
does not occur again at that site. These two assumptions give rise to the so-called perfect phylogeny
evolutionary model. Hajirasouliha and Raphael proposed in [10] a model for deconvoluting each
sample into a set of tumor types so that the multi-set of all resulting tumor types forms a perfect
phylogeny, and is minimum with this property. Even though this model has some limitations, for
example it assumes no errors, and only single nucleotide variant mutations, it is a fundamental
problem whose understanding can lead to more practical extensions.
Other major approaches for deconvoluting tumor heterogeneity include methods based on somatic point mutations, such as PyClone [27], SciClone [21], PhyloSub [15], CITUP [20], LICHeE [26],
and methods based on somatic copy number alterations, such as THetA [25], TITAN [9] and MixClone [18].
Let us review two methods from the first category mentioned above. CITUP [20] exhaustively
enumerates through all possible phylogenetic trees (up to maximum number of vertices) and tries
to decompose each sample into several nodes of the phylogeny. The fit between each sample and
the phylogenetic tree is one minimizing a Bayesian information criterion on the frequencies of each
mutation. This is computed either exactly, with quadratic integer programming, or with a heuristic
iterative method. The tree achieving an optimal fit is output, together with the decompositions of
each sample as nodes (i.e., sets of mutations) of this tree.
Method LICHeE [26] also tries to fit the observed mutation frequencies to an optimal phylogenetic tree, but with an optimized search for such a tree. Mutations are first assigned to clusters
based on their frequencies (a mutation can belong to more clusters). These clusters form the
nodes of a directed acyclic graph (DAG). Directed edges are added to these graphs from a node to
all its possible descendants, based on inclusions among their mutation sets and on compatibility
among their observed frequencies. Spanning trees of this DAG are enumerated, and the ones best
compatible with the mutation frequencies are output.
As opposed to the problem proposed in [10] and considered in this paper, these two methods heavily rely on the mutation frequencies. Frequencies appear at the core of their problem
2
formulations, and without them they would probably output an arbitrary phylogenetic tree and
a decomposition of samples into nodes of this tree compatible with the observed data. The two
problems in this paper have the same output, but only assume data on absence or presence of mutations in each sample. The objective function of our first problem requires that the sum, over each
sample, of the number of leaves of the phylogeny that the sample is decomposed to, is minimum.
The second problem formulation only requires that output phylogeny has the minimum number of
leaves.
In this paper we show that several proofs from [10] related to the optimization problem proposed
therein are incorrect, including an NP-hardness proof of it.1 However, the NP-hardness claim turns
out to be correct: in this paper we provide a different NP-hardness proof. We also adapt this proof
to a variant of the problem also proposed in [10] but whose complexity was left open. This problem
asks to minimize the set (instead of multi-set) of all tumor types of the perfect phylogeny. The two
problems, formally defined in Section 2, are referred to as the Minimum Conflict-Free Row
Split and the Minimum Distinct Conflict-Free Row Split problem, respectively.
Moreover, we obtain a polynomial time algorithm for a collection of instances of the Minimum
Conflict-Free Row Split problem, which can be biologically characterized as follows. Say that
two mutations i and j are exclusive if i is present in a sample in which j is absent, and j is present
in a sample in which i is absent. Observe that exclusive mutations cannot both be present in the
same vertex of a perfect phylogeny. Thus, we say that a sample is a mixture at exclusive mutations
i and j if both i and j are present in that sample. The instances for which we can solve the problem
in polynomial time are such that for any two exclusive mutations i and j, no mutation is present
only in the samples mixed at i and j.
We also propose an efficient (though not necessarily optimal) heuristic algorithm for the
Minimum Conflict-Free Row Split problem, based on coloring co-comparability graphs,
and provide implementations of both algorithms, freely available at https://github.com/
alexandrutomescu/MixedPerfectPhylogeny.
Paper outline. In Section 2 we give all formal definitions and review the approach of [10]. In
Section 3 we give a complete characterization of the so-called row-conflict graphs, the class of graphs
considered in [10]. The complexity results are presented in Section 4, and the above-mentioned
polynomial time algorithm is given in Section 5. In Section 6 we discuss the heuristic algorithm for
general instances, and in Sections 7 and 8 we present experimental results on the binary matrices
from clear cell renal cell carcinomas (ccRCC) from [5]. We conclude the paper with a discussion in
Section 9.
Some of the results in this paper appeared in the proceedings of WABI 2015 [13]. In addition
to the material presented in [13], this paper contains all the missing proofs (complete proofs of
Theorems 2, 3, and 4, a more detailed proof of Lemma 2), a time complexity analysis of the
algorithm presented in Section 5, a discussion following the proof of Theorem 5 on the necessity of
the assumptions for the algorithm given in Section 5, and three additional sections (Sections 6, 7
and 8) describing a polynomial time heuristic for general instances and experimental results.
1
Hajirasouliha and Raphael mentioned during their WABI 2014 talk that their claim about every graph being
a row-conflict graph (Theorem 4 in [10]) contained a flaw and proposed a correction stating that for every binary
matrix M with an all-zeros row and an all-ones row, the complement of GM,r (for any row r of M ) is transitively
orientable (cf. Section 2 for the definition of GM,r and Theorem 2 below). In particular, the fact that Theorem 4
in [10] is incorrect implies that the NP-hardness proof from [10] is incorrect.
3
2
Problem formulation
As mentioned in the introduction, we assume that we have a set of sequencing reads from each
tumor sample, and that based on these reads we have discovered the sample variants with respect
to a reference (e.g., by using a somatic mutation caller such as VarScan 2 [17]). This gives rise to
an m × n matrix M whose m rows are the different samples, and whose n columns are the genome
loci where a mutation was observed with respect to the reference. The entries of M are either 0 or
1, with 0 indicating the absence of a mutation, and 1 indicating the presence of the mutation. We
assume that the matrix has no row whose all entries are 0.
Under ideal conditions, e.g., each mutation was called without errors, and the samples do not
contain reads from several leaves of the perfect phylogeny, M corresponds to a perfect phylogeny
matrix. Such matrices are characterizable by a simple property, called conflict-freeness.
Definition 1. Two columns i and j of a binary matrix M are said to be in conflict if there exist
three rows r, r0 , r00 of M such that Mr,i = Mr,j = 1, Mr0 ,i = Mr00 ,j = 0, and Mr0 ,j = Mr00 ,i = 1. A
binary matrix M is said to be conflict-free if no two columns of M are in conflict.
It is well known that the rows of M are leaves of a perfect phylogenetic tree if and only if M is
conflict-free (see [3, 8]). Moreover, if this is the case, then the corresponding phylogenetic tree can
be retrieved from M in time linear in the size of M [7].
However, in practice, each tumor sample is a mixture of reads from several tumor types, and
thus possibly M is not conflict-free. If we are not allowed to edit the entries of M as done e.g. by
methods such as [29], [28], Hajirasouliha and Raphael proposed in [10] to turn M into a conflict-free
matrix M 0 by splitting each row r of M into some rows r1 , . . . , rk such that r is the bitwise OR of
r1 , . . . , rk ; that is, for every column c, Mr,c = 1 if and only if Mri ,c = 1 for at least one ri . The rows
r1 , . . . , rk can be seen as the deconvolution of the mixed sample r into samples from single vertices
of a perfect phylogeny. One can then build the perfect phylogeny corresponding to M 0 and carry
further downstream analysis. Let us make this row split operation precise.
Definition 2. Given a binary matrix M ∈ {0, 1}m×n with rows labeled r1 , r2 , . . . , rm , we say that
0
a binary matrix M 0 ∈ {0, 1}m ×n is a row split of M if there exists a partition of the set of rows of
0 such that for all i ∈ {1, 2, . . . , m}, r is the bitwise OR of the binary
M 0 into m sets R10 , R20 , . . . , Rm
i
vectors given by the rows of Ri0 . The set Ri0 of rows of M 0 is said to be a set of split rows of row ri .
Observe that a simple strategy for obtaining a conflict-free row split of M is to split every row r
into as many rows as there are 1s in r, with a single 1 per row. While this might be an informative
solution for some instances (cf. also Corollary 2 on p. 14), Hajirasouliha and Raphael proposed
in [10] as criterion for obtaining a meaningful conflict-free row split M 0 the requirement that the
number of rows of M 0 is minimum among all conflict-free row splits of M .
In this paper we consider the following problem, which we call Minimum Conflict-Free
Row Split problem. For a binary matrix M , we denote by γ(M ) the minimum number of rows
in a conflict-free row split M 0 of M . This notation is in line with notation γ(M ) used in [10]
to denote the minimum number of additional rows in a conflict-free row split M 0 of M , that is,
γ(M ) = γ(M ) − m, where m is the number of rows of M .
4
Minimum Conflict-Free Row Split:
Input: A binary matrix M , an integer k.
Question: Is it true that γ(M ) ≤ k?
The optimization version of the above problem (in which only a given subset of rows needs to
be split) was called the Minimum-Split-Row problem in [10], however, all results from [10] deal
with the variant of the problem in which all rows need to be split (some perhaps trivially by setting
Ri0 = {ri }), which is equivalent to the Minimum Conflict-Free Row Split problem.
Given a binary matrix M and a row r of M , the conflict graph of (M, r) is the graph GM,r
defined as follows: with each entry 1 in r, we associate a vertex in GM,r , and two vertices in GM,r
are connected by an edge if and only if the corresponding columns in M are in conflict. Denoting by
χ(G) the chromatic number of a graph G, Hajirasouliha and Raphael proved in [10] the following
lower bound on the value of γ(M ):
Lemma 1. [10] Let M be a binary matrix M with a conflict-free row split M 0 . Then, for every
row ri of M with a set Ri0 of split rows of M 0 , we have |Ri0 | ≥ χ(GM,ri ).
P
Corollary 1. For every binary matrix M , we have γ(M ) ≥ r χ(GM,r ).
Hajirasouliha and Raphael also claimed in [10] the following hardness result.
Theorem 1. [10] The Minimum Conflict-Free Row Split problem is NP-hard.
To recall their approach for proving Theorem 1, we need one more definition. We denote the
fact that two graphs G and H are isomorphic by G ∼
= H.
Definition 3. A graph G is a row-conflict graph if there exists a binary matrix M and a row r of
M such that G ∼
= GM,r .
The proof of Theorem 1 was based on a reduction from the chromatic number problem in graphs
and relied on three ingredients: the lower bound given by Corollary 1, Theorem 4 from [10] stating
that every graph is a row-conflict graph, and an algorithm based on graph coloring, also proposed
in [10], for optimally solving the Minimum Conflict-Free
Row Split problem by constructing
P
a conflict-free row split of M with exactly r χ(GM,r ) rows. In particular, their results would
imply that the lower bound on γ(M ) given by Corollary 1 is always attained with equality.
Contrary to what was claimed in [10], we show that there exist graphs that are not row-conflict
graphs. In fact, we give a complete characterization of row-conflict graphs, showing that a graph
is a row-conflict graph if and only if its complement is transitively orientable (see ??). Using a
reduction from 3-edge-colorability of cubic graphs, we show that it is NP-complete to test whether
a given binary matrix M has a conflict-free row split M 0 with number of rows achieving the lower
bound given by Corollary 1 (see ??). This implies that there exist infinitely many matrices for
which this bound is not achieved.
A corollary of our characterization of row-conflict graphs is that the chromatic number is polynomially computable for this class of graphs.PThis fact with the assumption that P 6= NP, as well
as the existence of matrices M with γ(M ) > r χ(GM,r ), each individually imply that the claimed
NP-hardness proof of the Minimum Conflict-Free Row Split problem given in [10] is flawed.
Nevertheless, our NP-completeness proof (see Theorem 3) implies that Theorem 1 is correct.
5
On the positive side, we give a polynomial time algorithm for the Minimum Conflict-Free
Row Split problem on input matrices M in which no column is contained in both columns of a
pair of conflicting columns (see Theorem 5).
We also consider a variant of the problem, also proposed in [10], in which we are only interested
in minimizing the number of distinct rows in a conflict-free row split of M . This problem is similar
to the Minimum Perfect Phylogeny Haplotyping problem [1], in which we need to explain a set
of genotypes with a minimum number of haplotypes admitting a perfect phylogeny. For a binary
matrix M , we denote by η(M ) the minimum number of distinct rows in a conflict-free row split
M 0 of M . We establish NP-completeness of the following problem (see Theorem 4), which was left
open in [10].
Minimum Distinct Conflict-Free Row Split:
Input: A binary matrix M , an integer k.
Question: Is it true that η(M ) ≤ k?
3
A characterization of row-conflict graphs
Definition 4. Given a binary matrix M and two columns i and j of M , column i is said to be
contained in column j if Mk,i ≤ Mk,j holds for every k. The undirected containment graph HM
is the undirected graph whose vertices correspond to the columns of M and in which two vertices i
and j, i 6= j, are adjacent if and only if the column corresponding to vertex i is contained in the
column corresponding to vertex j or vice versa.
Recall that an orientation of an undirected graph G = (V, E) is a directed graph D = (V, A)
such that for every edge uv ∈ E, either (u, v) ∈ A or (v, u) ∈ A, but not both. An orientation is
said to be transitive if the presence of the directed edges (u, v) and (v, w) implies the presence of the
directed edge (u, w). A graph is said to be transitively orientable if it has a transitive orientation.
The complement of a graph G is a graph G with the same vertex set as G in which two distinct
vertices are adjacent if and only if they are non-adjacent in G. Transitively orientable graphs
appeared in the literature under the name of comparability graphs (and their complements under
the name of co-comparability graphs). Transitively orientable graphs and their complements form
a subclass of the well known class of perfect graphs [6]. Therefore, odd cycles of length at least 5
and their complements are examples of graphs that are not transitively orientable.
Observation 1. For every binary matrix M , the graph HM is transitively orientable.
Proof. We say that column i is properly contained in column j if i is contained in j and Mk,i < Mk,j
for some k. Fix an ordering {c1 , . . . , cn } of the columns of M . Let us define a binary relation @
on the set of columns on M by setting, for every two columns ci and cj of M , ci @ cj if and only
if either ci is properly contained in cj , or i < j and each of ci and cj is contained in the other one
(that is, as binary vectors they are the same). Observe that for a pair of columns ci and cj with
ci cj ∈ E(HM ) we have either ci @ cj or cj @ ci but not both. The binary relation @ defines an
orientation of HM , by orienting each edge ci cj as going from ci to cj if and only if ci @ cj . This
orientation can be easily verified to be transitive.
6
In the next theorem, we characterize row-conflict graphs (cf. Definition 3).
Theorem 2. A graph G is a row-conflict graph if and only if G is transitively orientable.
Proof. (⇒) Let M be an arbitrary binary matrix, r an arbitrary row of M , and let G = GM,r . Let
N be the submatrix of M consiting of the columns of M that have 1 in row r. It is now easy to see
that GM,r ∼
= GN,r . Moreover, any two columns of N are either in conflict or their corresponding
vertices are adjacent in HN . Therefore, HN ∼
= GN,r . Since HN is transitively orientable (by
Observation 1), it follows that G is transitively orientable as well.
(⇐) We follow the strategy of the proof of Theorem 4 in [10] (which works for complements
of transitively orientable graphs). For the sake of completeness, we include here a short proof of
this implication. Let G be a graph such that H = G is transitively orientable, with a transitive
→
−
→
−
orientation H . It can be easily seen that H is acyclic, thus we may assume that vertices of G are
→
−
topologically ordered as V (G) = {v1 , . . . , vn }, that is, for every directed edge (vi , vj ) in H , we have
i < j. Let E(G) = {e1 , e2 , . . . , em }. We construct a matrix M with n columns and 2m + 1 rows,
such that GM,1 ∼
= G. The first row of M is defined to have all entries equal to 1. For every edge
ek = vi vj , i < j, of G, the 2k-th row of M has entry 0 in the column corresponding to vertex vi ,
and entry 1 in the column corresponding to vj . Additionally, the (2k + 1)-st row of M has entry 1
in the column corresponding to vertex vi , and entry 0 in the column corresponding to vj . Since the
first row has all entries equal to 1, after filling in these entries of M , the two columns corresponding
to vi and vj , respectively, are in conflict.
We need to fill in the remaining entries of M so that we do not introduce any new conflicts. For
→
−
every i, we fill in the remaining entries so that whenever (vi , vj ) is a directed edge in H , the column
corresponding to the vertex vi is contained in the column corresponding to the vertex vj . This can
→
−
be achieved by examining the columns one by one, following the topological order (v1 , . . . , vn ) of H ,
and filling each unfilled entry with a 0, unless this would violate the above containment principle.
At the end of this procedure, there are no conflicts between columns corresponding to vertices
→
−
vi and vj , whenever (vi , vj ) is a directed edge in H . Therefore, GM,1 ∼
= G.
Theorem 2 implies that odd cycles of length at least 5 and their complements are not rowconflict graphs. The reader not familiar with transitively orientable graphs might find it useful to
verify that the cycle of length 5 cannot be transitively oriented.
4
Complexity results
Theorem 3. The following two problems are NP-complete:
• The Minimum Conflict-Free Row Split problem.
P
• Given a binary matrix M , is it true that γ(M ) = r χ(GM,r )?
Proof. The Minimum Conflict-Free Row Split problem is in NP, since testing if a given binary matrix M 0 with at most k rows, equipped with a partition of its rows into m sets, satisfies the
condition in the definition of a row split, as well as the conflict-freeness, can be done in polynomial
time. To argue that the second problem is in NP, we proceed
Psimilarly as above, performing an ad0 equals
ditional test
checking
that
the
number
of
rows
of
M
r χ(GM,r ). (In this
P
Pcase, we will have
γ(M ) ≤ r χ(GM,r ) and equality will follow from Corollary 1.) The value of r χ(GM,r ) can be
7
computed in polynomial time, since each graph GM,r is the complement of a transitively orientable
graph (by Theorem 2), and the chromatic number of complements of transitively orientable graphs
can be computed in polynomial time (see, e.g., [6]).
We prove hardness of both problems at once, making a reduction from the following NP-complete
problem [11]: Given a simple cubic graph G = (V, E), is G 3-edge-colorable? (A graph is cubic,
or 3-regular, if every vertex is incident with precisely three edges. A matching in a graph is a
set of pairwise disjoint edges. A graph is 3-edge-colorable if its edge set can be partitioned into 3
matchings.)
Given a simple cubic graph G = (V, E), we construct an instance (M, k) of the Minimum
Conflict-Free Row Split problem as follows:
• M is a (|V | + 3) × (|E| + 3) binary matrix, with rows indexed by V ∪ {r1 , r2 , r3 }, columns
indexed by E ∪ {c1 , c2 , c3 }, and entries defined as follows (see Fig. 1 for an example):
– For every row indexed by a vertex v ∈ V and every column indexed by an edge e, we
have
1, if v is an endpoint of e;
Mv,e =
0, otherwise.
– For every row indexed by a vertex v ∈ V and every column indexed by some c ∈
{c1 , c2 , c3 }, we have Mv,c = 1.
– For every row indexed by some r ∈ {r1 , r2 , r3 } and every column indexed by an edge
e ∈ E, we have Mr,e = 0.
– For every row indexed by some ri ∈ {r1 , r2 , r3 } and every column indexed by some
cj ∈ {c1 , c2 , c3 }, we have
1, if i = j
Mri ,cj =
0, otherwise.
• k = 3|V | + 3.
G = (V, E)
v1
e4
e3
v4
v3
v1
v2
v3
v4
r1
r2
r3
e1
e5
e6
e2
k = 15
M
v2
e1 e2 e3 e4 e5 e6 c1 c2 c3
1
1
0
0
0
0
0
0
1
1
0
0
0
0
1
0
1
0
0
0
0
1
0
0
1
0
0
0
0
1
0
1
0
0
0
0
0
1
1
0
0
0
1
1
1
1
1
0
0
1
1
1
1
0
1
0
1
1
1
1
0
0
1
Figure 1: An example construction of (M, k) from G.
Note that for each row indexed by a vertex v ∈ V , the graph GM,v is isomorphic to the disjoint
union of two complete graphs with three vertices each, hence χ(GM,v ) = 3. For each row indexed
by some rP
∈ {r1 , r2 , r3 }, the graph GM,r consists in a single vertex, thus χ(GM,r ) = 1. It follows
that k = r χ(GM,r ) and
P therefore M is a yes instance to the second problem (“Given a binary
matrix M , is γ(M ) =
r χ(GM,r )?”) if and only if (M, k) is a yes instance for the Minimum
8
Conflict-Free Row Split problem. Hardness of both problems will therefore follow from the
following claim, which we prove next: G is 3-edge-colorable if and only if γ(M ) ≤ k.
Suppose first that G is 3-edge-colorable, and let E = E1 ∪ E2 ∪ E3 be a partition of E into
3 matchings. We obtain a row split M 0 of M by replacing each row of M indexed by a vertex
v ∈ V with three rows and keeping each row of M indexed by some r ∈ {r1 , r2 , r3 } unchanged.
Clearly, this will result in a matrix with k rows. For every v ∈ V , we replace the row of M
indexed by v as follows. Vertex v is incident with precisely three edges in G, say e1 , e2 , e3 . Since
E1 , E2 , E3 are matchings partitioning E, we may assume, without loss of generality, that ei ∈ Ei
for all i ∈ {1, 2, 3}. The three rows replacing in M 0 the row of M indexed by v are indexed by v 1 ,
v 2 , v 3 and defined as follows: for every i ∈ {1, 2, 3} and every column c ∈ E ∪ {c1 , c2 , c3 }, we have
1, if c = ei or c = ci ;
Mv0 i ,c =
0, otherwise.
By construction, M 0 is a row split of M with k rows. We claim that M 0 is conflict-free. No pair of
columns indexed by two edges in E agree on value 1 in any row, hence they cannot be in conflict.
The same holds for any pair of columns indexed by two elements of {c1 , c2 , c3 }. Consider now
two columns, one indexed by an edge e ∈ E and one indexed by ci ∈ {c1 , c2 , c3 }. Without loss of
generality, we may assume that e ∈ E1 . There are only two rows in which the column indexed by e
has value 1, namely the rows indexed by copies of the endpoints of e, say u1 and v 1 (with u, v ∈ V ).
The values of M 0 in column ci at rows u1 and v 1 are both 1 (if i = 1), otherwise they are both
0. Consequently, the two columns cannot be in conflict. Since M 0 is a conflict-free row split of M
with k rows, this establishes γ(M ) ≤ k.
For the converse direction, let M 0 be a conflict-free row split of M with at most k rows. Let
0
V = V ∪ {r1 , r2 , r3 } and consider a partition {Ri0 | i ∈ V 0 } of the set of rows of M 0 into |V | + 3
sets indexed by elements of V 0 such that for all i ∈ V 0 , the row of M indexed by i is the bitwise
OR of the rows of Ri0 . Since k is a lower bound on γ(M ), matrix M 0 has exactly k rows. This fact
and Corollary 1 imply that each row in M indexed by a vertex v ∈ V has |Rv0 | = 3 and each row
indexed by some r ∈ {r1 , r2 , r3 } has |Rr0 | = 1.
We must have that for all i ∈ V 0 , the row of M indexed by i isPthe bitwise sum of the rows of
0
0 . Indeed, if for some
Ri , that is, for every column c ∈ E ∪ {c1 , c2 , c3 }, we have Mi,c = r∈R0 Mr,c
i
P
0 > 1, then i is a vertex of G.
i ∈ V 0 and some column c ∈ E ∪ {c1 , c2 , c3 }, we have that r∈R0 Mr,c
i
Furthermore, since |Ri0 | = 3, there are either two edges of G, say e and f , incident with i such that
0 = M 0 = 1, or there are two distinct elements e, f ∈ {c , c , c } with
for some r ∈ Ri0 , we have Mr,e
1 2 3
r,f
the same property. In the former case, considering the rows replacing the rows of M indexed by
the endpoints of e and f other than i, respectively, we find two distinct rows r0 and r00 of M such
that Mr0 0 ,e = Mr0 00 ,f = 1 and Mr0 00 ,e = Mr0 0 ,f = 0, which contradicts the fact that M 0 is conflict-free.
In the latter case, the argument is similar.
By permuting the rows of M 0 if necessary, we may assume that each set of the form Rv0 is
ordered as Rv0 = {v 1 , v 2 , v 3 } so that
1, if i = j;
0
Mvi ,cj =
0, otherwise.
We claim that for every edge e = uv ∈ E and every i ∈ {1, 2, 3}, we have that Mu0 i ,e = Mv0 i ,e . If
this was not the case, then we would have Mu0 i ,e = Mv0 j ,e = 1 for a distinct pair i, j ∈ {1, 2, 3}. But
9
then the columns of M 0 indexed by e and ci would both agree in value 1 in row indexed by ui and
disagree (in opposite directions) in rows indexed by v j and v i . Thus, they would be in conflict,
contrary to the fact that M 0 is conflict-free.
Since for every edge e = uv ∈ E and every i ∈ {1, 2, 3}, we have that Mu0 i ,e = Mv0 i ,e , we can
partition the edges of E into three pairwise disjoint sets E1 , E2 , E3 by placing every edge e = uv ∈ E
into Ei if and only if i ∈ {1, 2, 3} is the unique index such that Mu0 i ,e = Mv0 i ,e = 1. We claim that
each Ei is a matching in G. This will imply that G is 3-edge-colorable and complete the proof.
If some Ei is not a matching, then there exist two distinct edges, say e, f ∈ Ei with a common
endpoint. Let e = xy and f = xz. The columns of M 0 indexed by e and f agree in value 1 at
row indexed by xi , while they disagree (in opposite directions) in rows indexed by y i and z i . Thus,
they are in conflict, contrary to the conflict-freeness of M 0 .
Hajirasouliha and Raphael proposed in [10] an algorithm based on graph coloring for optimally
solving the Minimum Conflict-Free
Row Split problem by constructing a conflict-free row
P
split of M with exactly r χ(GM,r ) rows. Since there are infinitely many cubic graphs that are
not 3-edge-colorable (see, e.g., [14]),
P the proof of Theorem 3 implies that there exist infinitely many
matrices M such that γ(M ) > r χ(GM,r ). On such instances, the algorithm from [10] will not
produce a valid (that is, conflict-free) solution.
Since the smallest
P cubic 4-edge-chromatic graph is the Petersen graph, the smallest matrix
M with γ(M ) > r χ(GM,r ) that can be obtained using the construction given in the proof of
Theorem 3 is of order 13 × 18. A smaller matrix M for which the bound from Corollary 1 is not
tight can be obtained by applying a similar construction starting from the complete graph of order
3 (which is a 2-regular 3-edge-chromatic graph):
1
1
0
1
1
1 0 1 1 1
M = 0 1 1 1 1 .
0 0 0 1 0
0
0
0
0
1
P
We leave it as an exercise for the reader to verify that r χ(GM,r ) = 8 and γ(M ) ≥ 9 (in fact,
γ(M ) =P
9). Let us also remark that in [16, Section 4.2.1] a binary matrix M is given with
γ(M ) = r χ(GM,r ), on which the algorithm from [10] fails to produce a conflict-free solution.
We conclude this section with another hardness result.
Theorem 4. The Minimum Distinct Conflict-Free Row Split problem is NP-complete.
Proof. Membership in NP of the Minimum Distinct Conflict-Free Row Split problem can
be argued similarly as for the Minimum Conflict-Free Row Split problem. It suffices to argue
that there is a polynomially-sized conflict-free matrix M 0 such that M 0 is a row split of M with at
0 of rows of M 0 into m sets
most k distinct rows. We may assume that for a partition R10 , . . . , Rm
satisfying the condition in the definition of a row split, the rows within each Ri0 are pairwise distinct.
Recall (e.g. from [8]) that a conflict-free matrix with d distinct rows and n columns corresponds to
a perfect phylogenetic rooted tree T such that: T has d leaves (the rows of the matrix), all internal
vertices of T are branching, and all edges from a vertex to its children are injectively labeled with
a column of M , with the exception of at most one edge which is unlabeled. Thus T has at most 2n
edges, and we infer that d ≤ 2n. Therefore, the total number of rows of M 0 does not exceed 2nm,
where m and n are the numbers of rows and columns of M , respectively.
10
The hardness proof is based on a slight modification of the reduction used in the proof of
Theorem 3. (See Fig. 2 for an example.) Given a cubic graph G = (V, E), we map it to (M , k)
where
• M is the binary matrix obtained from the binary matrix M described in the proof of Theorem 3 by adding to it three columns d1 , d2 , d3 , which on the rows indexed by V equal 0, and
on the rows indexed by r1 , r2 , r3 , each di equals ci , i ∈ {1, 2, 3}.
• k = |E| + 3.
G = (V, E)
e4
e3
v4
v3
e1
e5
e6
e2
k=9
M
v1
v2
v1
v2
v3
v4
r1
r2
r3
e1 e2 e3 e4 e5 e6 c1 c2 c3 d1 d2 d3
1
1
0
0
0
0
0
0
1
1
0
0
0
0
1
0
1
0
0
0
0
1
0
0
1
0
0
0
0
1
0
1
0
0
0
0
0
1
1
0
0
0
1
1
1
1
1
0
0
1
1
1
1
0
1
0
1
1
1
1
0
0
1
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
Figure 2: An example construction of (M , k) from G.
We claim that (M , k) is an instance of the Minimum Distinct Conflict-Free Row Split
problem such that G is 3-edge-colorable if and only if η(M ) ≤ k.
Suppose that G is 3-edge-colorable. Given a partition of E into three matchings E = E1 ∪E2 ∪E3 ,
we construct the same matrix M 0 as described in the proof of Theorem 3, to which we append the
three columns indexed by d1 , d2 , d3 which are all 0s on the rows indexed by vertices, and which are
the same as in M on the rows r1 , r2 , r3 . By the same argument given in the proof of Theorem 3, M 0
is conflict-free. Each row ri , i ∈ {1, 2, 3}, is distinct from all other rows of M 0 . Let v i , i ∈ {1, 2, 3},
be a row corresponding to a vertex v and suppose Mv0 i ,e = 1, where e = uv is one of the three edges
incident to v, and e ∈ Ei . By construction, the only other row having a 1 in column e is ui . Thus,
row v i is different from all other rows, except ui . In fact, we can see that row v i is identical to row
ui , since they have no other entry 1 on the columns indexed by edges. Additionally, they both have
1 in column ci , since e ∈ Ei , and 0 in the other five columns in {c1 , c2 , c3 , d1 , d2 , d3 } \ {ci }. Hence,
the number of distinct rows of M 0 is at most 3|V |/2 + 3 = |E| + 3 = k, since G is cubic, and thus
η(M ) ≤ k.
For the converse direction, suppose that M 0 is a conflict-free row split of M with at most
k = |E| + 3 distinct rows. Let V 0 = V ∪ {r1 , r2 , r3 } and consider a partition {Ri0 | i ∈ V 0 } of the
set of rows of M 0 into |V | + 3 sets indexed by elements of V 0 such that for all i ∈ V 0 , the row of
M indexed by i is the bitwise OR of the rows of Ri0 . We will prove that (1) the number of pairwise
distinct rows in Rv0 is 3 for all v ∈ V , and that (2) the number of pairwise distinct rows in Rr is 1
for all r ∈ {r1 , r2 , r3 }. Applying the same approach as in the proof of Theorem 3 will then imply
that G is 3-edge-colorable.
As argued in the proof of Theorem 3, no row in Rv0 has two 1s on two columns indexed by two
edges, say e and f , because each of e and f has an endpoint which is not an endpoint of the other
edge (and thus a row with two 1s on two columns indexed by two edges would imply a conflict in
M 0 ). Moreover, no row in Rv0 has two 1s on two columns indexed by c1 , c2 , c3 .
11
Let us associate with each row of M 0 belonging to some Ri0 with i ∈ V the edge column where
it has a 1 (if there is any). Since each edge column contains a 1 and no row has two 1s on the
columns indexed by edges, the number of pairwise distinct rows of M 0 indexed by a vertex is at
least |E|. Since in each Rr0 i , i ∈ {1, 2, 3}, we must have at least one row distinct from all other rows
of M 0 (because of the 1s in columns d1 , d2 , d3 ), and M 0 has at most |E| + 3 pairwise distinct rows,
the number of distinct rows of M 0 is exactly |E| + 3. This directly implies (2), more precisely, that
each Rr consists only of a row identical to the corresponding row of M .
In order to prove property (1), suppose now that there is a row of M indexed by a vertex v such
that Rv0 contains at least 4 pairwise distinct rows. Observe first that there is no row in Rv0 having
a 1 only in one column among {c1 , c2 , c3 } (and only 0s in the columns indexed by edges). Indeed,
besides being distinct from the row in each Rr0 , r ∈ {r1 , r2 , r3 }, it would also be distinct from each
of the set of at least |E| rows of M 0 having a 1 on a column indexed by an edge. Thus this would
contradict the fact that M 0 has at most |E| + 3 pairwise distinct rows. This implies that there are
two distinct rows v 0 and v 00 in Rv0 such that v 0 and v 00 both contain a 1 on the same column indexed
by an edge, say e, but on a column among {c1 , c2 , c3 }, say ci , v 0 contains 1 and v 00 contains 0. This
shows that there is a conflict in M 0 , since Mr0 i ,e = 0 and Mr0 i ,ci = 1, a contradiction.
5
A polynomially solvable case
In this section we consider the binary matrices in which no column is contained in both columns of
a pair of conflicting columns, and derive a polynomial time algorithm for the Minimum ConflictFree Row Split problem on such matrices. The main idea behind the algorithm is the fact that
on such matrices the lower bound from Corollary 1 is achieved, and the bound can be expressed
in terms of parameters of a set of derived digraphs, the so-called directed containment graph (see
Definition 5 below).
Let M be a binary matrix such that no column of M is contained in two or more conflicting
columns. If there are duplicated columns in M , then we form a new matrix where we take just one
copy of the columns that are duplicated. Since an optimal solution of the reduced instance can be
mapped to an optimal solution of the original instance (by duplicating the columns corresponding
to the copies of the duplicated columns in M kept by the reduction), we may assume that there
are no duplicated columns in M .
Definition 5. Given a binary matrix M with distinct columns c1 , . . . , cn and a row r of M , the
→
−
directed containment graph of (M, r) is the graph H M,r whose vertex set is the set of columns of
M having a 1 in row r, in which there is a directed edge from ci to cj if and only if i 6= j and ci is
contained in cj .
→
−
We will use the notation ci @r cj as a shorthand for the fact that (ci , cj ) is a directed edge of H M,r .
→
−
→
−
We say that ci is a source of H M,r if ci ∈ V ( H M,r ) and there is no cj with cj @r ci . Let σ(M, r)
→
−
denote the number of sources in H M,r .
Lemma 2. If there are no duplicated columns in M , then σ(M, r) ≤ χ(GM,r ) holds for any row r
of M .
Proof. Two vertices in the complement of GM,r are adjacent if and only if the corresponding columns
→
−
of M are either disjoint or one contains the other one. However, since the vertices of both H M,r
12
and GM,r correspond to columns in which M has value 1 in row r, no two such columns can be
→
−
disjoint. Consequently, the underlying undirected graph of H M,r is equal to the complement of
→
−
GM,r . The set of all sources of H M,r forms an independent set in its underlying undirected graph.
This set corresponds to a clique in GM,r . Therefore σ(M, r) ≤ ω(GM,r ) ≤ χ(GM,r ) (where ω(GM,r )
denotes the maximum size of a clique in GM,r ).
Our algorithm is the following one (see also Fig. 3 for an example).
Input: An m × n binary matrix M with columns c1 , c2 , . . . , cn , without duplicated columns,
and such that no column of M is contained in both columns of a pair of conflicting columns.
Output: A conflict-free row split M 0 of M with γ(M ) rows.
Algorithm:
1. Define a new matrix M 0 with columns c01 , c02 , . . . , c0n .
0
2. For each row r of M , add the rows r10 , . . . , rσ(M,r)
to M 0 , defined as:
→
−
Let cr,1 , . . . , cr,σ(M,r) be the sources of H M,r .
1, if cr,i = cj or cr,i @r cj ;
Mr0 0 ,c0 =
i j
0, otherwise.
3. Return M 0 .
M c1
1
r
1
0
0
0
c2
0
1
1
1
1
c3
0
1
0
1
0
c4
0
1
1
0
0
c5
0
0
0
1
0
M0
0
r1
r20
0
r3
c04 c05
c01
c02
c03
1
0
0
0
1
1
0 0
1 0
0 1
0
0
0
→
−
H M,r
c1
cr,1
c2
c3
cr,2
c4
cr,3
Figure 3: An example of a matrix M in which no column is contained in both columns of a
pair of conflicting columns (c1 , c2 and c3 , c4 are conflicting). The rows r10 , r20 , r30 constructed by
the algorithm corresponding to row r of M are shown in the center. On the right, the directed
containment graph of (M, r).
Theorem 5. For any m×n binary matrix M without duplicated columns such that noPcolumn of M
is
Pcontained in both columns of a pair of conflicting0 columns, it holds that γ(M ) = r χ(GM,r ) =
r σ(M, r). Moreover, a conflict-free row split M of M with γ(M ) rows can be constructed in
time O(mn2 ).
Proof. We claim that the matrix M 0 produced by the above algorithm is a conflict-free row split of
M with number of rows equal to γ(M ).
It is clear that M 0 is a row split of M . Let us prove that M 0 is conflict-free. Suppose the
contrary, that is, let c0i and c0j be two columns of M 0 which are in conflict. Then, there exists a row
rk0 of M 0 (obtained by splitting a row r of M ) which has 1 in columns c0i and c0j .
13
We will first show that ci is contained in cj or vice versa. If cr,k = ci (resp. cr,k = cj ) then
ci @r cj (resp. cj @r ci ) and therefore column ci is contained in column cj (resp. cj is contained in
ci ). Suppose now that cr,k 6∈ {ci , cj }. Since rk0 has 1 in columns c0i and c0j it follows that cr,k @r ci
and cr,k @r cj . This implies that column cr,k is contained in both column ci and column cj . By the
assumption on M , ci and cj cannot be in conflict, hence, one of them is contained in the other one.
Thus, we may assume without loss of generality that ci is contained in cj . Since c0i and c0j are
in conflict it follows that there exists a row w`0 of M 0 which has 1 in column c0i and 0 in column
c0j . This implies that the corresponding row w of M has 1 in column ci , and consequently also in
→
−
cj , since ci is contained in cj . Therefore, both ci and cj are vertices of H M,w . If ci = cw,` , then
w`0 has value 1 in column c0j (since ci is contained in cj ), which contradicts the choice of w`0 . Thus,
→
−
ci 6= cw,` and cw,` @w ci . However, since ci is contained in cj and H M,w is transitive, it follows that
cw,` @w cj . This implies that row w`0 has value 1 in column c0j , which again contradicts the choice
of w`0 . This finally shows that M 0 is conflict-free.
P
0
0
r σ(M, r) and
P Since the number of rows in M is
P M is conflict
P free, we have γ(M ) ≤
σ(M,
r).
By
Corollary
1
and
Lemma
2
we
have
σ(M,
r)
≤
r
r
r χ(GM,r ) ≤ γ(M ). This
implies equality.
It remains to justify the time complexity. First, we compute, in time O(mn2 ), the transitive
→
−
orientation H M of the undirected containment graph HM as specified by Observation 1 (that
→
−
is, (ci , cj ) is an arc of H M if and only if ci is properly contained in cj ). Since for each row
→
−
→
−
→
−
r of M , the graph H M,r is an induced subdigraph of H M , the σ(M, r) sources of H M,r can
→
−
be computed from H M in the straightforward way in time O(n2 ). The corresponding σ(M, r)
0
new rows of M
P can be computed in2 time O(σ(M, r)n), which results in total time complexity of
2
O(mn ) + O( r σ(M, r)n) = O(mn ), as claimed.
Note that the correctness of the algorithm crucially relies on the assumption that no column of
the input matrix is contained in both columns of a pair of conflicting
For example, the
columns.
1 1 1
algorithm fails to resolve the conflict in the 3 × 3 input matrix M = 0 1 0 , which violates
0 0 1
the assumption (column c1 is contained in both columns c2 and c3 , which are in conflict). Given
the above matrix M , the output matrix M 0 computed by the algorithm is in fact equal to M .
It is also worth mentioning that if the input matrix satisfies the stronger property that no column
is contained in another one, Theorem 5 implies that the naı̈ve solution obtained by splitting each
row r into as many 1s as it contains always produces an optimal solution. This is true since all
→
−
vertices of H M,r are sources. We thus obtain:
Corollary 2. For any binary m × n matrix M such that no column of M is contained in another
one, it holds that γ(M ) = m0 , where m0 equals the number of 1s in M . Moreover, a conflict-free
row split M 0 of M of size m0 × n can be constructed in time O(m0 n).
6
A heuristic algorithm based on coloring co-comparability graphs
As pointed out in Section 4, the graph theoretic algorithm from [10, Section 4] fails to always
produce a conflict-free row split of the input matrix. In this section, we propose a polynomial time
heuristic algorithm for the Minimum Conflict-Free Row Split problem, that is, an algorithm
14
that always produces a conflict-free row split of the input matrix. This algorithm is also based on
graph colorings.
Before presenting the algorithm, we describe the intuition behind it. The lower bound on γ(M )
given by Corollary 1 follows from the fact that in every conflict-free row split M 0 of the input
matrix M , the rows replacing row r in the split can be used to produce a valid vertex coloring of
GM,r , the conflict graph of (M, r). The difficulty in reversing this P
argument in order to obtain a
row split of M having a number of rows close to the lower bound r χ(GM,r ) is due to the fact
that we cannot independently combine the splits of rows r of M according to optimal colorings of
their conflict graphs, as new conflicts may arise.
We can guarantee that the corresponding row splits will be pairwise compatible (in the sense
that no new conflicts will be generated) as follows. We color GM , the conflict graph of the whole
input matrix (which we will define in a moment), and split each row r according to the coloring
of its conflict graph GM,r given by the restriction of the coloring of GM to the vertex set of GM,r .
The graph GM , the conflict graph of M , is defined as follows: with each column of M , we associate
a vertex in GM . Two vertices in GM are connected by an edge if and only if the corresponding
columns in M are in conflict. Note that each conflict graph GM,r of an individual row is an induced
subgraph of GM , hence the restriction of any proper coloring c of GM to V (GM,r ) is a proper
coloring of GM,r .
P The above approach will result in a row split having number of rows given by the value of
r |c(V (GM,r ))|, where |c(V (GM,r ))| denotes the number of colors used by c on V (GM,r ). As a
first heuristic attempt to minimize this quantity, Kačar proposed in [16, Section 4.2.2] to choose
a coloring c of GM with χ(GM ) colors. However, this is computationally intractable. While rowconflict graphs are characterized (in Theorem 2) as exactly the co-comparability graphs (that is,
as complements of transitively orientable graphs)—for which the coloring problem is polynomially
solvable [6]—, conflict graphs of binary matrices do not enjoy such nice features. Indeed, if G is any
graph of minimum degree at least 2, then G ∼
= GM , where M ∈ {0, 1}E(G)×V (G) is the edge-vertex
incidence matrix of G (defined by Me,v = 1 if and only if vertex v is an endpoint of edge e).
This can be amended as follows. We can “restore” the structure of co-comparability graphs by
observing that GM is a spanning subgraph of HM , the complement of the undirected containment
graph HM (cf. Definition 4), and working with HM instead. Recall that HM is the undirected
graph whose vertices correspond to the columns of M and in which two vertices i and j, i 6= j,
are adjacent if and only if one the corresponding columns is contained in the other one. To show
that GM is a spanning subgraph of HM , note first that we may assume that V (GM ) = V (HM ) (as
both vertex sets are in bijective correspondence with the set of columns of M ). Moreover, if two
vertices i and j of GM are adjacent, then the corresponding columns are in conflict, which implies
that neither of them is contained in the other one; consequently, they are adjacent in HM .
Since GM is a spanning subgraph of HM , any proper coloring of HM is also a proper coloring
of GM . Moreover, even though the graph HM might have more edges than GM , these additional
edges (if any) will not be contained in any of the graphs GM,r . Indeed, for every row r, its conflict
graph GM,r coincides both with the subgraph of GM induced by U := V (GM,r ) as well as with the
subgraph of HM induced by U . This is because for any two vertices i and j in U that are adjacent
in HM , the corresponding columns cannot be disjoint, therefore, since i and j are not adjacent in
HM , the corresponding columns must be in conflict.
In view of the above observations, we propose choosing an optimal P
coloring c of the cocomparability graph HM as a heuristic approach to minimizing the value of r |c(V (GM,r ))| for a
15
coloring c of GM . A row split of M is then defined according to the coloring c.
This leads to the following algorithm (see Fig. 4 for an example):
Input: An m × n binary matrix M with columns labeled with 1, . . . , n.
Output: A conflict-free row split M 0 of M .
Algorithm:
1. Define a new matrix M 0 with n columns labeled with 1, . . . , n.
2. Compute HM , the complement of the undirected containment graph HM .
3. Compute an optimal coloring c of HM .
4. For each row r of M :
Let c(V (GM,r )) = {sr1 , . . . , srt }.
Add the rows r10 , . . . , rt0 to M 0 , defined as:
1, if Mr,j = 1 and c(j) = sri ;
0
Mr0 ,j =
i
0, otherwise.
5. Return M 0 .
M1 2
1 0
r
1 1
0 1
0 1
0 1
3
0
1
0
1
0
4 5
0 0
1 0
1 0
0 1
0 0
M0
0
r1
r20
0
r3
1
1
0
0
2
0
0
1
3
0
0
1
4
0
1
0
5
0
0
0
1 s1
HM
s3
2
5 s3
4
s2
s3
3
s1 , s2 , s3 – colors
as used by c
Figure 4: An example of a binary matrix M , the complement of its undirected containment graph
together with an optimal coloring c, and a split of a row of M according to the above algorithm.
Theorem 6. For any m × n binary matrix M , the above algorithm can be implemented to run in
time O(n2 (n1/2 + m)). The matrix M 0 output by the algorithm is a conflict-free row split of M .
Proof. We first show that the matrix M 0 produced by the above algorithm is a conflict-free row
split of M . Clearly, M 0 is a row split of M . We say that a row r0 of M 0 is an r-row if r is
the row of M such that r0 was added to M 0 in step 4) of the algorithm when considering row
r. Suppose for a contradiction that M 0 is not conflict-free, and let {j, j 0 } be a pair of conflicting
columns of M 0 . Then, there exist rows p, q, and r of M and rows p0i , qk0 , and r`0 of M 0 such that
p0i is a p-row, qk0 is a q-row, and r`0 is an r-row, Mp0 0 ,j = Mp0 0 ,j 0 = 1, Mq0 0 ,j = 1, Mq0 0 ,j 0 = 0, and
i
i
k
k
Mr0 0 ,j = 0, Mr0 0 ,j 0 = 1. Since Mp0 0 ,j = Mp0 0 ,j 0 = Mq0 0 ,j = Mr0 0 ,j 0 = 1, the definition of M 0 implies that
`
`
i
i
k
16
`
Mp,j = Mp,j 0 = Mq,j = Mr,j 0 = 1 and c(j) = c(j 0 ) = spi , c(j) = sqk , and c(j 0 ) = sr` . Consequently,
spi = sqk = sr` . Moreover, since Mq0 0 ,j 0 = 0 and c(j 0 ) = sqk , we infer that Mq,j 0 = 0 and, similarly,
k
since Mr0 0 ,j = 0 and c(j) = sr` , we infer that Mr,j = 0. It follows that columns j and j 0 of M are
`
in conflict. On the other hand, since c(j) = c(j 0 ) and c is a proper vertex coloring of HM , vertices
corresponding to j and j 0 are non-adjacent in HM . Hence, they are adjacent in HM , which means
that one of the columns j and j 0 is contained in the other one, contradicting the fact that they are
in conflict. This completes the proof that M 0 is conflict-free.
It remains to justify the time complexity. The graph HM can be computed in time O(mn2 ) by
comparing every ordered pair of columns for containment. In the same time, we can also compute
a transitive orientation @ of HM (as done for example in Observation 1). An optimal coloring
of HM corresponds to a minimum chain partition of the partially ordered set P = (V (HM ), @
), and can be computed in time O(n5/2 ) as follows (cf. [24] and [19, p. 73-74]). Applying the
approach of Fulkerson [4], a minimum chain partition of P can be computed by solving a maximum
matching problem in a derived bipartite graph having 2n vertices. This can be done in time O(n5/2 )
using
and Karp [12]. Step 4) of the algorithm can be executed in time
P the algorithm of Hopcroft
2
O( r |c(V (GM,r )|n) = O(mn ). The claimed time complexity of O(n2 (n1/2 + m)) follows.
7
Implementation and experimental results
C++ implementations of both algorithms are available at https://github.com/alexandrutomescu/
MixedPerfectPhylogeny. The input matrices must be in .csv format, and are allowed to have duplicate columns. In addition to a conflict-free row split matrix, we also output its perfect phylogeny
tree. We tested our implementation on the ten binary matrices constructed from clear cell renal
cell carcinomas (ccRCC) from [5]: EV001-EV003, EV005-EV007, RMH002, RMH004, RMH008,
RK26. The phylogenetic trees constructed by [5] from these matrices appear in [5, Fig.3].
Table 1: The numbers of rows, columns, and pairwise distinct columns in the input matrices, the
lower bounds on the minimum number of rows in conflict-free row splits of the input matrices given
by Corollary 1, and the numbers of rows and pairwise distinct rows in the matrices output by the
heuristic algorithm. The distinct rows form the leaves of the output perfect phylogeny.
Name
EV001
EV002
EV003
EV005∗
EV006
EV007
RHM002
RHM004
RHM008
RK26
#rows
10
7
8
7
9
8
5
6
8
11
Input
#cols #distinct cols
122
22
103
17
56
12
83
10
76
11
66
15
54
11
140
17
81
10
75
17
∗ Input
lower bound
on #rows in output
37
18
13
7
13
19
9
16
10
18
is conflict-free
17
#rows
51
29
19
7
25
25
13
21
20
26
Output
#distinct rows
22
17
12
7
11
15
11
17
10
17
These ten matrices have between 5 and 11 rows, and between 55 and 140 columns. One of
them is already conflict-free, while the other nine do not belong to the polynomially-solvable case
discussed in Section 5. On each of these nine matrices, our heuristic algorithm from Section 6
runs in less than one second. On ten random binary matrices with 50 rows and 1000 columns,
the heuristic algorithm runs on average in 97 seconds. Due to the restricted structure of the input
matrices on which the polynomial time algorithm from Section 5 works correctly, we did not test
the implementation of this algorithm on random inputs. However, since this algorithm is simpler
than the heuristic one, it is plausible to expect that it will run at least as fast. Of course one might
want to first check whether the input m × n binary matrix satisfies the assumption that no column
is contained in both columns of a pair of conflicting columns. This can be tested straightforwardly
in time O(n2 (m + n)) by first classifying each ordered pair of columns as conflicting, disjoint, or in
containment, and then testing each triple of columns for the condition that the first one is contained
in each of the other two, which are in conflict. The running time of this check is asymptotically
worse than that of the heuristic algorithm. However, we expect the running times to differ little on
practical instances of moderate sizes, since the constant hidden in the O() notation for the above
check is small.
In Table 1 we list the numbers of rows (i.e., samples) in the ten original matrices, together with
numbers of columns and pairwise distinct columns, the lower bounds on the minimum number of
rows in a conflict-free row split of each of the matrices given by Corollary 1, and the numbers of
rows and pairwise distinct rows in the matrices output by the heuristic algorithm. The similarity
between the numbers of pairwise distinct columns in the input and of pairwise distinct rows in the
output can be explained by the observation that if the input matrix consists of n pairwise distinct
columns, then there will be at most n distinct rows in the naive solution (split each row into as
many rows of the identity matrix as the number of 1s it contains). Thus n is an upper bound
for an optimal solution to the Minimum Distinct Conflict-Free Row Split problem, which
explains why our heuristic algorithm applied to the Minimum Distinct Conflict-Free Row
Split problem performs similarly as the naive one. In Fig. 5 we show four perfect phylogeny trees
corresponding to the matrices output by the heuristic algorithm. The results on all matrices are
available online, linked from https://github.com/alexandrutomescu/MixedPerfectPhylogeny.
We also ran our heuristic algorithm on the same datasets, with the difference that we removed
from the input matrices those columns that appear (as binary vectors) strictly less than 2 times
(this value is a parameter to our implementation). This is the same idea and default threshold used
by Popic et al. [26] in their tool LICHeE. Their motivation is that several mutations accumulate
before a new tumor branch separates, and thus such rare mutation patterns may be due to errors
in the data. We show four output trees in Fig. 6.
Finally, we also ran LICHeE on the same ten patients from [5]. Since LICHeE uses the variant
allele frequency (VAF) of every mutation, we used the matrices containing VAF values linked from
https://github.com/viq854/lichee/tree/master/LICHeE/data/ccRCC and ran LICHeE with
the parameters indicated therein. As referred to in the above, LICHeE starts by grouping the
robust mutations into clusters of size at least a given number, by default 2. (This is the same
experiment as the one done in the paper [26] introducing LICHeE; see [26] for further details.) We
show four trees produced by LICHeE in Fig. 7. Note LICHeE does not necessarily output binary
trees.
18
Equalities among split rows:
R1_1 = R3_1 = R4_1
R3_2 = R4_2
R11_2 = R10_2 = R5_4
R9_3 = R6_2 = R7_2 = R8_3
R7_1 = R8_1
R1_1
R11_1
Equalities among split rows:
R9_2 = R6_3 = R3_2 = R1_3 = R7_2
R5_1 = R2_1
R5_3 = R2_2
R4_1 = R3_1
Equalities among split rows:
R1_2 = R2_1 = R4_1
R1_3 = R3_2 = R6_3 = R4_3 = R5_2 = R7_2 = R8_3
R6_2 = R5_1 = R7_1
Equalities among split rows:
R1_2 = R2_2
R2_1 = R6_2
R4_1
R5_2
R6_1
R8_2
R5_1
R8_1
R1_2
R5_3
R1_1
R1_1
R1_1
R7_1
R10_1
R9_2
R9_3
R2_1
R9_1
R7_1
R7_1
R11_2
R6_1
R5_2
R3_1
R3_3
R8_2
R2_1
R3_2
R9_1
R6_2
R6_1
R5_1
R1_2
R9_2
R3_2
R6_2
R7_2
R4_2
R1_3
R2_2
R1_2
R2_2
R5_3
R6_3
R4_2
(a) EV003
R3_1
R7_3
(b) RMH002
R1_2
R6_1
(c) RMH008
(d) RK26
Figure 5: Four perfect phylogeny trees corresponding to the conflict-free row split matrices output
by our heuristic algorithm from Section 6. The naming convention is: R1 is an original row (i.e.,
sample) name, and R1 1, R1 2, R1 3 are the row (i.e., samples) names corresponding to R1 in
the output matrix. Equal split rows form a single node of the perfect phylogeny (equalities are
indicated in boxes).
Equalities among split rows:
R1 = R3 = R4
R11_2 = R10_2 = R9_2 = R5_3 = R6_2 = R7_2 = R8_3
R6_1 = R7_1 = R8_1
Equalities among split rows:
R1_2 = R2_1 = R4_1
R1_3 = R3_2 = R6_3 = R4_3 = R5_2 = R7_2 = R8_3
R6_2 = R5_1 = R7_1
R2_1
R11_1
R1_1
R8_2
Equalities among split rows:
R5 = R2
R3 = R1
R8_1
R1
R3
R7
R6_2
R6
R5
R4
R9
R6
(a) EV003
R5_2
R8_2
R2
R3
R4_2
R1_3
R2_2
(b) RMH002
R5_1
R1_2
R3_1
R7
(c) RMH008
R6_1
R6_1
R10_1
R11_2
R9_1
R2_2
R1
(d) RK26
Figure 6: The perfect phylogeny trees output by our heuristic algorithm, after removing from
the input matrices those columns that appear only once. The naming convention is as in Fig. 5.
Compare these trees to the trees from [5, Fig. 3] and to the ones from Fig. 7.
19
GL
12
GL
GL
GL
2
22
27
9
6
5
10
2
8
4
1
8
2
R5
R2
R4
1
1
2
R6
R9
R7
4
R1
R3
(a) EV003
R1
R6
4
10
4
R3
8
R7
6
R2
R8
(b) RMH002
R5
3
5
7
4
R7
R4
R6
9
2
21
4
R3
(c) RMH008
R2
R1
R2
3
2
R4
R1
R3
1
1
R9
R10
2
3
6
R5
R6
R7
R8
(d) RK26
Figure 7: The tumor evolutions predicted by LICHeE [26]. Numbers in the circles represent the
number of mutations at that node. See [26] for further details.
8
Discussion
After filtering the unique columns of the matrices EV003 and RMH002, the resulting matrices
are conflict-free. The perfect phylogeny for RMH002 (Fig. 6) is the same as the one produced by
LICHeE (Fig. 7) and the one from [5, Fig. 3]. The perfect phylogeny for EV003 (Fig. 6) is almost
identical to the one produced by LICHeE (Fig. 7) and the one from [5, Fig. 3], the only difference
being that each of the pairs of samples {R2, R5} and {R1, R3} is collapsed into a single leaf of
the phylogeny. This may be due to the fact that LICHeE and the method used by [5] exploit VAF
values, not only binary values. Overall, these two matrices suggest that filtering out rare columns
may be a relevant strategy.
All the columns of the matrix RMH008 appear at least two times, thus there are no columns
to be filtered out. While LICHeE finds that only samples R4 and R6 are a combination of more
leaves of the tumor phylogeny, our algorithm finds that all eight samples are combinations of two
or more leaves. However, there are similarities to the prediction of LICHeE. For example, both
samples R4 and R6 have mutations in common with R5 and R7 (R6 2 = R5 1 = R7 1, (R1 3 = )
R4 3 = R5 2 = R7 2). Our subclones R6 2 and R4 3 also correspond to the subclones R6 dom and
R4 min, respectively, found in [5]. Samples R4 and R6 also have mutations in common with R1
(R1 3 = R6 3 = R4 3), and with R2 (nodes R2 2 and R1 3 = R6 3 = R4 3 are both siblings of a
node at distance 4 from the root). The subclones R6 3 and R4 3 also correspond to the subclones
R6 min and R4 dom, respectively, found in [5].
In matrix RK26, 6 columns are unique, and they have been filtered out in Fig. 6. LICHeE
finds that only sample R5 is a combination of more leaves, while our algorithm again finds that all
samples are combinations of two or more leaves. However, there are again similarities in the results.
Sample R5 has mutations shared with R6, R7, R8 (nodes R5 2 and R6 1 = R7 1 = R8 1 are siblings
of a node at depth 4). Subclone R5 2 also corresponds to subclone R5 dom found in [5]. Sample R5
also has mutations shared with R9 and R10 (nodes R5 1, R10 1 and R9 1 have the lowest common
ancestor at depth 5). Subclone R5 1 also corresponds to subclone R5 min found in [5].
20
R11
9
Conclusion
In this paper we showed hardness of the Minimum Conflict-Free Row Split and the Minimum
Distinct Conflict-Free Row Split problems, and gave a polynomial time algorithm for the
Minimum Conflict-Free Row Split problem on instances such that no column is contained
in both columns of a pair of conflicting columns. More general tractable instances could be found
by inspecting further dependencies between column containment and conflictness. For example, it
remains open whether the Minimum Conflict-Free Row Split problem is tractable on matrices
in which no pair of conflicting columns is contained in both columns of a pair of conflicting columns.
It would also be interesting to identify polynomially solvable cases of the Minimum Distinct
Conflict-Free Row Split problem and to explore variations of the problems in which we are
also allowed to edit the entries of the input matrix.
In the paper we also gave a polynomial time heuristic algorithm for the Minimum ConflictFree Row Split problem based on graph coloring. We leave as a question for future research
to determine the (in-)approximability status of the optimization variants of the two problems.
Finally, we remark that in [10] it was assumed that the matrices have no duplicated columns,
which was not necessary in this paper.
References
[1] V. Bafna et al. A note on efficient computation of haplotypes via perfect phylogeny. Journal
of Computational Biology, 11(5):858–866, 2004.
[2] P. J. Campbell et al. Subclonal phylogenetic structures in cancer revealed by ultra-deep
sequencing. Proceedings of the National Academy of Sciences, 105(35):13081–13086, 2008.
[3] G. F. Estabrook et al. An idealized concept of the true cladistic character. Mathematical
Biosciences, 23(3-4):263–272, 1975.
[4] D. R. Fulkerson. Note on Dilworth’s decomposition theorem for partially ordered sets. Proc.
Amer. Math. Soc., 7:701–702, 1956.
[5] M. Gerlinger, S. Horswell, J. Larkin, A. J. Rowan, M. P. Salm, I. Varela, R. Fisher, N. McGranahan, N. Matthews, C. R. Santos, P. Martinez, B. Phillimore, S. Begum, A. Rabinowitz,
B. Spencer-Dene, S. Gulati, P. A. Bates, G. Stamp, L. Pickering, M. Gore, D. L. Nicol,
S. Hazell, P. A. Futreal, A. Stewart, and C. Swanton. Genomic architecture and evolution of
clear cell renal cell carcinomas defined by multiregion sequencing. Nat Genet, 46(3):225–233,
03 2014.
[6] M. C. Golumbic. Algorithmic Graph Theory and Perfect Graphs, volume 57 of Annals of
Discrete Mathematics. Elsevier Science B.V., Amsterdam, second edition, 2004.
[7] D. Gusfield. Efficient algorithms for inferring evolutionary trees. Networks, 21(1):19–28, 1991.
[8] D. Gusfield. Algorithms on Strings, Trees and Sequences: Computer Science and Computational Biology. Cambridge University Press, 1997.
[9] G. Ha et al. Titan: inference of copy number architectures in clonal cell populations from
tumor whole-genome sequence data. Genome Research, 24(11):1881–1893, 2014.
21
[10] I. Hajirasouliha and B. J. Raphael. Reconstructing mutational history in multiply sampled
tumors using perfect phylogeny mixtures. In Algorithms in Bioinformatics - 14th International
Workshop, WABI 2014, volume 8701 of LNCS, pages 354–367. Springer, 2014.
[11] I. Holyer. The NP-completeness of edge-coloring. SIAM J. Comput., 10(4):718–720, 1981.
[12] J. E. Hopcroft and R. M. Karp. An n5/2 algorithm for maximum matchings in bipartite graphs.
SIAM J. Comput., 2:225–231, 1973.
[13] A. Hujdurović, U. Kačar, M. Milanič, B. Ries, and A. I. Tomescu. Finding a perfect phylogeny
from mixed tumor samples. In Algorithms in Bioinformatics - 15th International Workshop,
WABI 2015, volume 9289 of LNCS, pages 80–92. Springer, 2015.
[14] R. Isaacs. Infinite families of nontrivial trivalent graphs which are not Tait colorable. Amer.
Math. Monthly, 82:221–239, 1975.
[15] W. Jiao et al. Inferring clonal evolution of tumors from single nucleotide somatic mutations.
BMC Bioinformatics, 15:35, 2014.
[16] U. Kačar. Problemi popolne filogenije (Perfect Phylogeny Problems). 2015. Final project
paper. University of Primorska, Faculty of Mathematics, Natural Sciences and Information
Technologies, Koper, Slovenia, 2015. Available online at http://www.famnit.upr.si/sl/
izobrazevanje/zakljucna_dela/view/276.
[17] D. C. Koboldt et al. VarScan 2: Somatic mutation and copy number alteration discovery in
cancer by exome sequencing. Genome Research, 22:568–576, 2012.
[18] Y. Li and X. Xie. Mixclone: a mixture model for inferring tumor subclonal populations. BMC
Genomics, 16(S-2):S1, 2015.
[19] V. Mäkinen, D. Belazzougui, F. Cunial, and A. I. Tomescu. Genome-Scale Algorithm Design.
Cambridge University Press, 2015.
[20] S. Malikic, A. W. McPherson, N. Donmez, and C. S. Sahinalp. Clonality inference in multiple
tumor samples using phylogeny. Bioinformatics, 31(9):1349–1356, 2015.
[21] C. A. Miller et al. SciClone: Inferring Clonal Architecture and Tracking the Spatial and
Temporal Patterns of Tumor Evolution. PLoS Comput Biol, 10(8):e1003665+, Aug. 2014.
[22] D. E. Newburger et al. Genome evolution during progression to breast cancer. Genome
Research, 23(7):1097–1108, 2013.
[23] S. Nik-Zainal et al. The life history of 21 breast cancers. Cell, 149(5):994–1007, 2012.
[24] S. C. Ntafos and S. L. Hakimi. On path cover problems in digraphs and applications to program
testing. IEEE Trans. Software Engrg., 5(5):520–529, 1979.
[25] L. Oesper et al. THetA: inferring intra-tumor heterogeneity from high-throughput DNA sequencing data. Genome Biology, 14(7):R80+, 2013.
[26] V. Popic, R. Salari, I. Hajirasouliha, D. Kashef-Haghighi, R. B. West, and S. Batzoglou. Fast
and scalable inference of multi-sample cancer lineages. Genome Biology, 16(1):1–17, 2015.
22
[27] A. Roth et al. PyClone: statistical inference of clonal population structure in cancer. Nature
methods, 11(4):396–398, Apr. 2014.
[28] R. Salari et al. Inference of tumor phylogenies with improved somatic mutation discovery.
Journal of Computational Biology, 20(11):933–944, 2013.
[29] K. E. van Rens et al. SNV-PPILP: Refined SNV calling for tumor data using perfect phylogenies and ILP. Bioinformatics, 31(7):1133–1135, 2015.
23
| 8cs.DS
|
Preventing Atomicity Violations with Contracts
Diogo G. Sousa
Ricardo J. Dias
Carla Ferreira
João M. Lourenço
arXiv:1505.02951v1 [cs.DC] 12 May 2015
NOVA LINCS — NOVA Laboratory for Computer Science and Informatics
Departamento de Informática, Faculdade de Ciências e Tecnologia
Universidade NOVA de Lisboa, Portugal
May 13, 2015
Abstract
Software developers are expected to protect concurrent accesses to shared regions of
memory with some mutual exclusion primitive that ensures atomicity properties to a
sequence of program statements. This approach prevents data races but may fail to provide all necessary correctness properties.The composition of correlated atomic operations
without further synchronization may cause atomicity violations. Atomic violations may be
avoided by grouping the correlated atomic regions in a single larger atomic scope. Concurrent programs are particularly prone to atomicity violations when they use services
provided by third party packages or modules, since the programmer may fail to identify
which services are correlated. In this paper we propose to use contracts for concurrency,
where the developer of a module writes a set of contract terms that specify which methods are correlated and must be executed in the same atomic scope. These contracts are
then used to verify the correctness of the main program with respect to the usage of the
module(s). If a contract is well defined and complete, and the main program respects
it, then the program is safe from atomicity violations with respect to that module. We
also propose a static analysis based methodology to verify contracts for concurrency that
we applied to some real-world software packages. The bug we found in Tomcat 6.0 was
immediately acknowledged and corrected by its development team.
1
Introduction
The encapsulation of a set of functionalities as services of a software module offers strong
advantages in software development, since it promotes the reuse of code and ease maintenance
efforts. If a programmer is unacquainted with the implementation details of a particular set
of services, she may fail to identify correlations that exist across those services, such as data
and code dependencies, leading to an inappropriate usage. This is particularly relevant in a
concurrent setting, where it is hard to account for all the possible interleavings between threads
and the effects of these interleaved calls to the module’s internal state.
One of the requirements for the correct behavior of a module is to respect its protocol,
which defines the legal sequences of invocations to its methods. For instance, a module that
offers an abstraction to deal with files typically will demand that the programmer start by
calling the method open(), followed by an arbitrary number of read() or write() operations,
and concluding with a call to close(). A program that does not follow this protocol is incorrect
and should be fixed. A way to enforce a program to conform to such well defined behaviors is to
use the design by contract methodology [21], and specifying contracts that regulate the module
usage protocol. In this setting, the contract not only serves as useful documentation, but may
also be automatically verified, ensuring the client’s program obeys the module’s protocol [8,15].
The development of concurrent programs brings new challenges on how to define the protocol of a module. Not only it is important to respect the module’s protocol, but is also necessary
1
void schedule () {
Task t = taskQueue . next ();
if ( t . isReady ())
t . run ();
}
Figure 1: Program with an atomicity violation.
to guarantee the atomic execution of sequences of calls that are susceptible of causing atomicity violations. These atomicity violations are possible, even when the individual methods
in the module are protected by some concurrency control mechanism. Figure 1 shows part
of a program that schedules tasks. The schedule() method gets a task, verifies if it is ready
to run, and execute it if so. This program contains a potential atomicity violation since the
method may execute a task that is not marked as ready. This may happen when another
thread concurrently schedules the same task, despite the fact the methods of Task are atomic.
In this case the isReady() and run() methods should be executed in the same atomic context
to avoid atomicity violations. Atomicity violations are one of the most common source of bugs
in concurrent programming [20] and are particularly susceptible to occur when composing calls
to a module, as the developer may not be aware of the implementation and internal state of
the module.
In this paper we propose to extend module usage protocols with a specification of the
sequences of calls that should be executed atomically. We will also present an efficient static
analysis to verify these protocols.
The contributions of this paper can be summarized as:
1. A proposal of contracts for concurrency addressing the issue of atomicity violations;
2. A static analysis methodology to extract the behavior of a program with respect to the
sequence of calls it may execute;
3. A static analysis to verify if a program conforms to a module’s contract, hence that the
module’s correlated services are correctly invoked in the scope of an atomic region.
The remaining of this paper is organized as follows. In Section 2 we provide a specification
and the semantics for the contract. Section 3 contains the general methodology of the analysis.
Section 4 presents the phase of the analysis that extracts the behavior of the client program
while Section 5 shows how to verify a contract based on the extracted information. Section 9
follows with the presentation and discussion of the results of our experimental validation. The
related work is presented in Section 10, and we conclude with the final remarks in Section 11.
2
Contract Specification
The contract of a module must specify which sequences of calls of its non-private methods must
be executed atomically, as to avoid atomicity violations in the module’s client program. In the
spirit of the programming by contract methodology, we assume the definition of the contract,
including the identification of the sequences of methods that should be executed atomically is
a responsibility of the module’s developer.
Definition 1 (Contract). The contract of a module with public methods m1 , · · ·, mn is of the
form,
2
1. e1
2. e2
..
.
k. ek .
where each clause i is described by ei , a star-free regular expression over the alphabet {m1 , · · ·, mn }.
Star-free regular expressions are regular expressions without the Kleene star, using only the
alternative (|) and the (implicit) concatenation operators.
Each sequence defined in ei must be executed atomically by the program using the module,
otherwise there is a violation of the contract. The contract specifies a finite number of sequences
of calls, since it is the union of star-free languages. Therefore, it is possible to have the same
expressivity by explicitly enumerating all sequences of calls, i.e., without using the alternative
operator. We chose to offer the alternative operator so the programmer can group similar
scenarios under the same clause. Our verification analysis assumes the contract defines a finite
number of call sequences.
Example Consider the array implementation offered by Java standard library, java.util.ArrayList.
For simplicity we will only consider the methods add(obj), contains(obj), indexOf(obj), get(idx),
set(idx, obj), remove(idx), and size().
The following contract defines some of the clauses for this class.
1. contains indexOf
2. indexOf (remove | set | get)
3. size (remove | set | get)
4. add indexOf.
Clause 1 of ArrayList’s contract denotes the execution of contains() followed by indexOf()
should be atomic, otherwise the client program may confirm the existence of an object in the
array, but fail to obtain its index due to a concurrent modification. Clause 2 represents a
similar scenario where, in addition, the position of the object is modified. In clause 3 we deal
with the common situation where the program verifies if a given index is valid before accessing
the array. To make sure the size obtained by size() is valid when accessing the array we
should execute these calls atomically. Clause 4 represents scenarios where an object is added
to the array and then the program tries to obtain information about that object by querying
the array.
Another relevant clause is contains indexOf (set | remove), but the contract’s semantic already enforces the atomicity of this clause as a consequence of the composition of clauses 1 and 2,
as they overlap in the indexOf() method.
3
Methodology
The proposed analysis verifies statically if a client program complies with the contract of a
given module, as defined in Section 2. This is achieved by verifying that the threads launched
by the program always execute atomically the sequence of calls defined by the contract.
This analysis has the following phases:
1. Determine the entry methods of each thread launched by the program.
2. Determine which of the program’s methods are atomically executed. We say that a
method is atomically executed if it is atomic1 or if the method is always called by atomically executed methods.
1 An
atomic method is a method that explicitly applies a concurrency control mechanism to enforce atomicity.
3
3. Extract the behavior of each of the program’s threads with respect to the usage of the
module under analysis.
4. For each thread, verify that its usage of the module respects the contract as defined in
Section 2.
In Section 4 we introduce the algorithm that extracts the program’s behavior with respect
to the module’s usage. Section 5 defines the methodology that verifies whether the extracted
behavior complies to the contract.
4
Extracting the Behavior of a Program
The behavior of the program with respect to the module usage can be seen as the individual
behavior of any thread the program may launch. The usage of a module by a thread t of a
program can be described by a language L over the alphabet m1 , · · ·, mn , the public methods
of the module. A word m1 · · · mn ∈ L if some execution of t may run the sequence of calls
m1 , · · ·, mn to the module.
To extract the usage of a module by a program, our analysis generates a context-free
grammar that represents the language L of a thread t of the client program, which is represented
by its control flow graph (CFG) [1]. The CFG of the thread t represents every possible path the
control flow may take during its execution. In other words, the analysis generates a grammar
Gt such that, if there is an execution path of t that runs the sequence of calls m1 , · · ·, mn , then
m1 · · · mn ∈ L(Gt ). (The language represented by a grammar G is denoted by L(G).)
A context-free grammar is especially suitable to capture the structure of the CFG since
it easily captures the call relations between methods that cannot be captured by a weaker
class of languages such as regular languages. The first example bellow will show how this
is done. Another advantage of using context-free grammars (as opposed to another static
analysis technique) is that we can use efficient algorithms for parsing to explore the language
it represents.
Definition 2 (Program’s Thread Behavior Grammar). The grammar Gt = (N, Σ, P, S) is
build from the CFG of the client’s program thread t.
We define,
• N , the set of non-terminals, as the set of nodes of the CFG. Additionally we add nonterminals that represent each method of the client’s program (represented in calligraphic
font);
• Σ, the set of terminals, as the set of identifiers of the public methods of the module under
analysis (represented in bold);
• P , the set of productions, as described bellow, by rules 1–5;
• S, the grammar initial symbol, as the non-terminal that represents the entry method of
the thread t.
For each method f() that thread t may run we add to P the productions respecting the
rules 1–5. Method f() is represented by F . A CFG node is denoted by α : JvK, where α is the
non-terminal that represents the node and v its type. We distinguish the following types of
nodes: entry, the entry node of method F ; mod.h(), a call to method h() of the module mod
under analysis; g(), a call to another method g() of the client program; and return, the return
point of method F . The succ : N → P(N ) function is used to obtain the successors of a given
CFG node.
4
if α : JentryK,
if α : Jmod.h()K,
if α : Jg()K,
if α : JreturnK,
if α : JotherwiseK,
{F → α} ∪ {α → β | β ∈ succ(α)} ⊂ P
{α → h β | β ∈ succ(α)} ⊂ P
(1)
(2)
{α → G β | β ∈ succ(α)} ⊂ P
where G represents g()
(3)
{α → ǫ} ⊂ P
{α → β | β ∈ succ(α)} ⊂ P
(4)
(5)
No more productions belong to P .
Rules 1–5 capture the structure of the CFG in the form of a context-free grammar. Intuitively this grammar represents the flow control of the thread t of the program, ignoring
everything not related with the module’s usage. For example, if f g ∈ L(Gt ) then the thread t
may invoke, method f(), followed by g().
Rule 1 adds a production that relates the non-terminal F , representing method f(), to the
entry node of the CFG of f(). Calls to the module under analysis are recorded in the grammar
by the Rule 2. Rule 3 handles calls to another method g() of the client program (method g() will
have its non-terminal G added by Rule 1). The return point of a method adds an ǫ production
to the grammar (Rule 4). All others types of CFG nodes are handled uniformly, preserving
the CFG structure by making them reducible to the successor non-terminals (Rule 5). Notice
that only the client program code is analyzed.
The Gt grammar may be ambiguous, i.e., offer several different derivations to the same
word. Each ambiguity in the parsing of a sequence of calls m1 · · · mn ∈ L(Gt ) represents
different contexts where these calls can be executed by thread t. Therefore we need to allow
such ambiguities so that the verification of the contract can cover all the occurrences of the
sequences of calls in the client program.
The language L(Gt ) contains every sequence of calls the program may execute, i.e., it
produces no false negatives. However L(Gt ) may contain sequences of calls the program does
not execute (for instance calls performed inside a block of code that is never executed), which
may lead to false positives.
Examples Figure 2 (left) shows a program that consists of two methods that call each other
mutually. Method f() is the entry point of the thread and the module under analysis is
represented by object m. The control flow graphs of these methods are shown in Figure 2 (right).
According to Definition 2, we construct the grammar G1 = (N1 , Σ1 , P1 , S1 ), where
N1 = {F , G, A, B, C, D, E, F, G, H, I, J, K, L, M },
Σ1 = {a, b, c, d},
S1 = F ,
and P1 has the following productions:
F →A
G→G
A→B
B → aC
H → cI
I→J |M
C→D|E
D→GE
J →GK
K → dL
E → bF
F →ǫ
L→FM
M → ǫ.
5
f()
g()
G
entry
A
entry
void f () {
m . a ();
if ( cond )
g ();
m . b ();
}
H
m.c()
B
m.a()
I
cond
C
void g () {
m . c ();
if ( cond ) {
g ();
m . d ();
f ();
}
}
cond
J
g()
D
g()
K
m.d()
E
m.b()
L
f()
F
return
M
return
Figure 2: Program with recursive calls using the module m (left) and respective CFG (right).
A second example, shown in Figure 3, exemplifies how the Definition 2 handles a flow
control with loops. In this example we have a single function f(), which is assumed to be the
entry point of the thread. We have G2 = (N2 , Σ2 , P2 , S2 ), with
N2 = {F , A, B, C, D, E, F, G, H},
Σ2 = {a, b, c, d},
S2 = F .
The set of productions P2 is,
F →A
A→B
E → cF
F →B
B → aC | aG
C→D|E
G → dH
H →ǫ
D → b F.
5
Contract Verification
The verification of a contract must ensure all sequences of calls specified by the contract are
executed atomically by all threads the client program may launch. Since there is a finite
number of call sequences defined by the contract we can verify each of these sequences to check
if the contract is respected.
The idea of the algorithm is to generate a grammar the captures the behavior of each thread
with respect to the module usage. Any sequence of the calls contained in the contract can then
be found by parsing the word (i.e. the sequence of calls) in that grammar. This will create a
parsing tree for each place where the thread can execute that sequence of calls. The parsing
tree can then be inspected to determine the atomicity of the sequence of calls discovered.
Algorithm 1 presents the pseudo-code of the algorithm that verifies a contract against a
client’s program. For each thread t of a program P , it is necessary to determine if (and where)
any of the sequences of calls defined by the contract w = m1 , · · ·, mn occur in P (line 4). To do
so, each of the these sequences are parsed in the grammar G′t (line 5) that includes all words
and sub-words of Gt . Sub-words must be included since we want to take into account partial
6
A
entry
void f () {
while ( m . a ()) {
if ( cond )
m . b ();
else
m . c ();
count ++;
B
m.a()
C
cond
D
m.b()
E
m.c()
}
count++
m . d ();
F
G
m.d()
}
H
return
Figure 3: Program using the module m (left) and respective CFG (right).
Algorithm 1 Contract verification algorithm.
Require: P , client’s program;
C, module contract (set of allowed sequences).
1: for t ∈ threads(P ) do
2:
Gt ← make_grammar(t)
3:
G′t ← subword_grammar(Gt )
4:
for w ∈ C do
5:
T ← parse(G′t , w)
6:
for τ ∈ T do
7:
N ← lowest_common_ancestor(τ, w)
8:
if ¬run_atomically(N ) then
9:
return ERROR
10: return OK
traces of the execution of thread t, i.e., if we have a program m.a(); m.b(); m.c(); m.d(); we
are able to verify the word b c by parsing it in G′t . Notice that G′t may be ambiguous. Each
different parsing tree represents different locations where the sequence of calls m1 , · · ·, mn may
occur in thread t. Function parse() returns the set of these parsing trees. Each parsing tree
contains information about the location of each methods call of m1 , · · ·, mn in program P (since
non-terminals represent CFG nodes). Additionally, by going upwards in the parsing tree, we
can find the node that represents the method under which all calls to m1 , · · ·, mn are performed.
This node is the lowest common ancestor of terminals m1 , · · ·, mn in the parsing tree (line 7).
Therefore we have to check the lowest common ancestor is always executed atomically (line
8) to make sure the whole sequence of calls is executed under the same atomic context. Since
it is the lowest common ancestor we are sure to require the minimal synchronization from
the program. A parsing tree contains information about the location in the program where a
contract violation may occur, therefore we can offer detailed instructions to the programmer
on where this violation occurs and how to fix it.
Grammar Gt can use all the expressivity offered by context-free languages. For this reason
it is not sufficient to use the LR(·) parsing algorithm [17], since it does not handle ambiguous
grammars. To deal with the full class of context-free languages a GLR parser (Generalized
LR parser) must be used. GLR parsers explore all the ambiguities that can generate different
derivation trees for a word. A GLR parser was introduced by Tomita in [24]. Tomita presents
7
R
void atomic run () {
f ();
m . c ();
}
void f () {
m . a ();
g ();
}
void g () {
while ( cond )
m . b ();
}
F
G
R→F c
F →aG
A
G→A
B
A→B|ǫ
B→bA
A
B
a
b
A
b ǫ c
Figure 4: Program (left), simplified grammar (center) and parsing tree of a b b c (right).
a non-deterministic versions of the LR(0) parsing algorithm with some optimizations in the
representation of the parsing stack that improve the temporal and spacial complexity of the
parsing phase.
Another important point is that the number of parsing trees may be infinite. This is due
to loops in the grammar, i.e., derivations from a non-terminal to itself (A ⇒ · · · ⇒ A), which
often occur in Gt (every loop in the control flow graph will yield a corresponding loop in the
grammar). For this reason the parsing algorithm must detect and prune parsing branches that
will lead to redundant loops, ensuring a finite number of parsing trees is returned. To achieve
this the parsing algorithm must detect a loop in the list of reduction it has applied in the
current parsing branch, and abort it if that loop did not contribute to parse a new terminal.
Examples Figure 4 shows a program (left), that uses the module m. The method run() is the
entry point of the thread t and is atomic. In the center of the figure we shown a simplified
version of the Gt grammar. (The G′t grammar is not shown for the sake of brevity.) The
run(), f(), and g() methods are represented in the grammar by the non-terminals R, F , and
G respectively. If we apply Algorithm 1 to this program with the contract C = {a b b c} the
resulting parsing tree, denoted by τ (line 6 of Algorithm 1), is represented in Figure 4 (right).
To verify all calls represented in this tree are executed atomically, the algorithm determines
the lowest common ancestor of a b b c in the parsing tree (line 7), in this example R. Since
R is always executed atomically (atomic keyword), it complies to the contract of the module.
Figure 5 exemplifies a situation where the generated grammar is ambiguous. In this case
the contract is C = {a b}. The figure shows the two distinct ways to parse the word a b
(right). Both these trees will be obtained by our verification algorithm (line 5 of Algorithm 1).
The first tree (top) has F as the lowest common ancestor of a b. Since F corresponds to the
method f(), which is executed atomically, so this tree respects the contract. The second tree
(bottom) has R as the lowest common ancestor of a b, corresponding to the execution of the
else branch of method run(). This non-terminal (R) does not correspond to an atomically
executed method, therefore the contract is not met and a contract violation is detected.
6
Analysis with Points-to
In an object-oriented programming language the module is often represented as an object,
in which case we should differentiate the instances of the class of the module. This section
explains how the analysis is extended to handle multiple instances of the module by using
points-to information.
To extend the analysis to points-to a different grammar is generated for each allocation site
of the module. Each allocation site represents an instance of the module, and the verification
8
void run () {
if (...)
f ();
else {
m . a ();
g ();
}
}
R
F
G
R→aG|F
F →aG
void atomic f () {
m . a ();
g ();
}
a
b
G→b
R
G
void atomic g () {
m . b ();
}
a
b
Figure 5: Program (left), simplified grammar (center) and parsing tree of a b (right).
Algorithm 2 Contract verification algorithm with points-to information.
Require: P , client’s program;
C, module contract (set of allowed sequences).
1: for t ∈ threads(P ) do
2:
for a ∈ mod_alloc_sites(t) do
3:
Gta ← make_grammar(t, a)
4:
G′ta ← subword_grammar(Gta )
5:
for w ∈ C do
6:
T ← parse(G′ta , w)
7:
for τ ∈ T do
8:
N ← lowest_common_ancestor(τ, w)
9:
if ¬run_atomically(N ) then
10:
return ERROR
11: return OK
algorithm verifies the contract words for each allocation site and thread (whereas the previous
algorithm verified the contract words for each thread). The new algorithm is shown in Algorithm 2. It generated the grammar Gta for a thread t and module instance a. This grammar
can be seen as the behavior of thread t with respect to the module instance a, ignoring every
other instance of the module.
To generate the grammar Gta we adapt the Definition 2 to only take into account the
instance a. The grammar generation is extended in the following way:
Definition 3 (Program’s Thread Behavior Grammar with points-to). The grammar Gt =
(N, Σ, P, S) is build from the CFG of the client’s program thread t and an object allocation
site a, which represents an instance of the module.
We define N , Σ, P and S in the same way as Definition 2.
The rules remain the same, except for rule 2, which becomes:
9
void replace ( int o , int n )
{
if ( array . contains ( o ))
{
int idx = array . indexOf ( o );
array . set ( idx , n );
}
}
Figure 6: Examples of atomic violation with data dependencies.
if α : Jmod.h()K and mod can only point to a
(6)
{α → h β | β ∈ succ(α)} ⊂ P
if α : Jmod.h()K and mod can point to a
(7)
{α → h β | β ∈ succ(α)} ⊂ P
{α → β | β ∈ succ(α)} ⊂ P
if α : Jmod.h()K and mod cannot point to a
{α → β | β ∈ succ(α)} ⊂ P
(8)
Here we use the points-to information to generate the grammar, and we should consider
the places where a variable can point-to. If it may point-to our instance a or another instance
we consider both possibilities in the Rule 7 of Definition 3.
7
Contracts with Parameters
Frequently contract clauses can be refined by considering the flow of data across calls to the
module. For instance Listing 6 shows a procedure that replaces an item in an array by another.
This listing contains two atomicity violations: the element might not exist when indexOf() is
called; and the index obtained might be outdated when set() is executed. Naturally, we can
define a clause that forces the atomicity of this sequence of calls as contains indexOf set,
but this can be substantially refined by explicitly require that a correlation exists between
the indexOf() and set() calls. To do so we extend the contract specification to capture the
arguments and return values of the calls, which allows the user to establish the relation of
values across calls.
The contract can therefore be extended to accommodate this relations, in this case the
clause might be
contains(X) Y=indexOf(X) set(Y,_).
This clause contains variables (X, Y) that must satisfy unification for the clause to be applicable. The underscore symbol (_) represents a variable that will not be used (and therefore
requires no binding). Algorithm 1 can easily be modified to filter out the parsing trees that
correspond to calls that do not satisfy the unification required by the clause in question.
In our implementation we require a exact match between the terms of the program to satisfy
the unification, since it was sufficient for most scenarios. It can however be advantageous to
generalize the unification relation. For example, the calls
array . contains ( o );
idx = array . indexOf ( o +1);
array . set ( idx , n );
also imply a data dependency between the first two calls. We should say that A unifies
with B if, and only if, the value of A depends on the value of B, which can occur due to value
10
manipulation (data dependency) or control-flow dependency (control dependency). This can
be obtained by an information flow analysis, such as presented in [5], which can statically infer
the variables that influenced the value that a variable hold on a specific part of the program.
This extension of the analysis can be a great advantage for some types of modules. As
an example we rewrite the contract for the Java standard library class, java.util.ArrayList,
presented in Section 2:
1. contains(X) indexOf(X)
2. X=indexOf(_) (remove(X) | set(X,_) | get(X))
3. X=size() (remove(X) | set(X,_) | get(X))
4. add(X) indexOf(X).
This contract captures in detail the relations between calls that may be problematic, and
excludes from the contract sequences of calls that does not constitute atomicity violations.
8
Prototype
A prototype was implemented to evaluate our methodology. This tool analyses Java programs
using Soot [25], a Java static analysis framework. This framework directly analyses Java
bytecode, allowing us to analyse a compiled program, without requiring access to its source
code. In our implementation a method can be marked atomic with a Java annotation. The
contract is also defined as an annotation of the class representing the module under analysis.
The prototype is available in https://github.com/trxsys/gluon.
8.1
Optimizations
To achieve a reasonable time performance we implemented a few optimizations. Some of
these optimizations reduced the analysis run time by a few orders of magnitude in some cases,
without sacrificing precision.
A simple optimization was applied to the grammar to reduce its size. When constructing the
grammar, most control flow graph nodes will have a single successor. Rule 5 (Definition 2) will
always be applied to these kind of nodes, since they represent an instruction that does not call
any function. This creates redundant ambiguities in the grammar due to the multiple control
flow paths that never use the module under analysis. To avoid exploration of redundant parsing
branches we rewrite the grammar to transform productions of the form A → βBδ, B → α to
A → βαδ, if no other rule with head B exists. For example, an if else that do not use the
module will create the productions A → B, A → C, B → D, C → D. This transformation will
reduce it to A → D, leaving no ambiguity for the parser to explore here. This optimization
reduced the analysis time by at least one order of magnitude considering the majority of the
tests we performed. For instance, the Elevator test could not be analyzed in a reasonable time
prior to this optimization.
Another optimization was applied during the parsing phase. Since the GLR parser builds
the derivation tree bottom-up we can be sure to find the lowest common ancestor of the terminals as early as possible. The lowest common ancestors will be the first non-terminal in the
tree covering all the terminals of the parse tree. This is easily determined if we propagate,
bottom-up, the number of terminals each node of the tree covers. Whenever a lowest common
ancestor is determined we do not need further parsing and can immediately verify if the corresponding calls are in the same atomic context. This avoids completing the rest of the tree
which can contain ambiguities, therefore a possibly large number of new branches is avoided.
Another key aspect of the parsing algorithm implementation is the loop detection. To
achieve a good performance we should prune parsing branches that generated unproductive
loop as soon as possible. Our implementation guarantees the same non-terminal never appears
twice in a parsing tree without contributing to the recognition of a new terminal.
11
Table 1: Optimization Improvements.
Optimization
Improvement
Grammar Simplification
Stop Parsing at LCA
Subword Parser
428%
?
3%
To achieve a better performance we also do not explicitly compute the subword grammar (G′t ). We have modified our GLR parser to parse subwords as described in [23]. This
greatly improves the parser performance because constructing G′t introduces many irrelevant
ambiguities the parser had to explore.
The Table 1 show how much the of the optimizations improve the analysis performance.
These results are build from an test made to stress the performance of gluon but is consistent
with real applications. The Improvement column show how much of an improvement that
particular optimization contributes to the analysis. The Stop Parsing at LCA cause an improvement that we were not able to measure since the test was unable to complete in reasonable
time.
8.2
Class Scope Mode
Gluon normally analyzes the entire program, taking into account any sequence of calls that can
spread across the whole program (as long as they are consecutive calls to a module). However
this is infeasible for very large programs so, for these programs, we ran the analysis with for
each class, ignoring calls to other classes. This will detect contract violations where the control
flow does not escape the class, which is reasonable since code locality indicates a stronger
correlations between calls.
This mode of operation can be useful to analyze large programs as they might have very
complex control flow graphs and thus are infeasible to analyze with the scope of the whole
program.
In this mode the grammar is build for each class instead of each thread. The methods of
the class will create non-terminals F1 , · · ·, Fn , just as before. The only change in creating this
grammar is that we create the productions S → F1 | · · · | Fn as the starting production of
the grammar (S being the initial symbol). This means that we consider the execution of all
methods of the class being analyzed.
9
Validation and Evaluation
To validate the proposed analysis we analyzed a few real-world programs (Tomcat, Lucene,
Derby, OpenJMS and Cassandra) as well as small programs known to contain atomicity violations. These small programs were adapted from the literature [2–4, 16, 19, 22, 27] and are
typically used to evaluate atomicity violation detection techniques. We modified these small
programs to employ a modular design and we wrote contracts to enforce the correct atomic
scope of calls to that module. Some additional clauses were added that may represent atomicity violations in the context of the module usage, even if the program do not violate those
clauses.
For the large benchmarks analyzed we aimed to discover new, unknown, atomicity violations. To do so we had to create contracts in an automated manner, since the code base was too
large. To automate the generation of contracts we employ a very simplistic approach that tries
to infer the contract’s clauses based on what is already synchronized in the code. This idea is
that most sequences of calls that should be atomic was correctly used somewhere. Having this
in mind we look for sequences of calls done to a module that are used atomically at least two
points of the program. If a sequence of calls is done atomically in two places of the code that
might indicate that these calls are correlated and should be atomic. We used these sequences
12
Table 2: Validation results.
Clauses
Contract
Violations
False
Positives
Potential
AV
Real
AV
SLOC
Time (s)
ICFinder
Static
ICFinder
Final
Allocate Vector [16]
Coord03 [2]
Coord04 [3]
Jigsaw [27]
Local [2]
Knight [19]
NASA [2]
Store [22]
StringBuffer [3]
UnderReporting [27]
VectorFail [22]
Account [27]
Arithmetic DB [19]
Connection [4]
Elevator [27]
1
4
2
1
2
1
1
1
1
1
2
4
2
2
2
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
183
151
35
100
24
135
89
621
27
20
70
42
243
74
268
0.120
0.093
0.039
0.044
0.033
0.219
0.035
0.090
0.032
0.029
0.048
0.041
0.272
0.058
0.333
121
-
2
-
OpenJMS 0.7
Tomcat 6.0
Cassandra 2.0
Derby 10.10
Lucene 4.6
6
9
1
1
3
54
157
60
19
136
10
16
24
5
21
28
47
15
7
76
4
3
2
1
0
163K
239K
192K
793K
478K
148
3070
246
522
151
126
365
122
391
15
12
16
2
Benchmark
as our contracts, after manually filtering a few irrelevant contracts. This is a very simple way
to generate contracts, which should ideally be written by the module’s developer to capture
common cases of atomicity violations, so we can expect the contracts to be more fine-tuned to
better target atomicity violations if the contracts are part of the regular project development.
Since these programs load classes dynamically it is impossible to obtain complete points-to
information, so we are pessimistic and assumed every module instance could be referenced
by any variable that are type-compatible. We also used the class scope mode described in
Section 8.2 because it would be impractical to analyze such large programs with the scope of
the whole program. This restrictions did not apply to the small programs analyzed.
Table 2 summarizes the results that validate the correctness of our approach. The table
contains both the macro benchmarks (above) and the micro benchmarks (bellow). The columns
represent the number of clauses of the contract (Clauses); the number of violations of those
clauses (Contract Violations); the number of false positives, i.e. sequences of calls that in fact
the program will never execute (False Positives); the number of potential atomicity violations,
i.e. atomicity violations that could happen if the object was concurrently accessed by multiple
threads (Potential AV); the number of real atomicity violations that can in fact occur and
compromise the correct execution of the program (Real AV); the number lines of code of the
benchmark (SLOC); and the time it took for the analysis to complete (the analysis run time
excludes the Soot initialization time, which were always less than 179s per run).
Our tool was able to detect all violation of the contract by the client program in the
microbenchmarks, so no false negatives occurred, which supports the soundness of the analysis.
Since some tests include additional contract clauses with call sequences not present in the test
programs we also show that, in general, the analysis does not detect spurious violations, i.e.,
false positives.2 A corrected version of each test was also verified and the prototype correctly
detected that all contract’s call sequences in the client program were now atomically executed.
Correcting a program is trivial since the prototype pinpoints the methods that must be made
atomic, and ensures the synchronization required has the finest possible scope, since it is the
method that corresponds to the lowest common ancestor of the terminals in the parse tree.
The large benchmarks show that gluon can be applied to large scale programs with good
results. Even with a simple automated contract generation we were able to detect 10 atomicity
violations in real-world programs. Six of these bugs where reported (Tomcat 3 , Derby 4 ,
2 In these tests no false positives were detected. However it is possible to create situations where false
positives occur. For instance, the analysis assumes a loop may iterate an arbitrary number of times, which
makes it consider execution traces that may not be possible.
3 https://issues.apache.org/bugzilla/show_bug.cgi?id=56784
4 https://issues.apache.org/jira/browse/DERBY-6679
13
Cassandra 5 ), with two bugs already fixed in Tomcat 8.0.11. The false positives incorrectly
reported by gluon were all due to conservative points-to information, since the program loads
and calls classes and methods dynamically (leading to an incomplete points-to graph).
ICFinder [18] uses a static analysis to detect two types of common incorrect composition
patterns. This is then filtered with a dynamic analysis. Of the atomicity violations detected
by gluon none of them was captured by ICFinder, since they failed to match the definition of
the patterns.
It is hard to directly compare both approaches since they use very different approaches.
Loosely speaking, in Table 2 the ICFinder Static column corresponds to the Contract Violations, since they both represent the static methodology of the approaches. The ICFinder Final
column cannot be directly compared with the Real AV column because “ICFinder Final” may
contain scenarios that not represent atomicity violations (in particular if ICFinder does not correctly identify the atomic sets). ICFinder Final cannot also be compared with “Potencial AV”
since “Potencial AV” is manually obtained from “Contract Violations”, and “ICFinder Final” is
the dynamic filtration of “ICFinder Static”. In the end the number of bugs reported by gluon
was 6, with 2 bugs confirmed and with fixes already applied, 1 bug considered highly unlikely,
and 3 bugs pending confirmation; ICFinder has 3 confirmed and fixed bugs on Tomcat. 6
The performance results show our tool can run efficiently. For larger programs we have to
use class scope mode, sacrificing precision for performance, but we still can capture interesting
contract violations. The performance of the analysis depends greatly on the number of branches
the parser explores. This high number of parsing branches is due to the complexity of the
control flow of the program, offering a huge amount of distinct control flow paths. In general
the parsing phase will dominate the time complexity of the analysis, so the analysis run time
will be proportional to the number of explored parsing branches. Memory usage is not a
problem for the analysis, since the asymptotic space complexity is determined by the size of
the parsing table and the largest parsing tree. Memory usage is not affected by the number
of parsing trees because our GLR parser explores the parsing branches in-depth instead of
in-breadth. In-depth exploration is possible because we never have infinite height parsing trees
due to our detection of unproductive loops.
9.1
ICFinder
ICFinder tries to infer automatically what a module is, and incorrect compositions of pairs of
calls to modules.
Two patterns are used to detect potencial atomicity violations in method calls compositions:
• USE: Detects stale value errors. This pattern detects data or control flow dependencies
between two calls to the module.
• COMP: If a call to method a() dominates b() and b() post-dominates a() in some place,
that is captured by this pattern. This means that, for each piece of code involving two
calls to the module (a() and b()), if a() is always executed before b() and b() is always
executed after a(), it is a COMP violation.
Both this patterns are extremely broad and contain many false positives. To deal with this
the authors filter this results with a dynamic analysis that only consider violations as defined
in [26]. This analysis assumes that the notion of atomic set was correctly inferred by ICFinder.
10
Related Work
The methodology of design by contract was introduced by Meyer [21] as a technique to write
robust code, based on contracts between programs and objects. In this context, a contract
5 https://issues.apache.org/jira/browse/CASSANDRA-7757
6 https://issues.apache.org/bugzilla/show_bug.cgi?id=53498
14
specifies the necessary conditions the program must met in order to call the object’s methods,
whose semantics is ensured if those pre-conditions are met.
Cheon et al. proposes the use of contracts to specify protocols for accessing objects [8].
These contracts use regular expressions to describe the sequences of calls that can be executed
for a given Java object. The authors present a dynamic analysis for the verification of the
contracts. This contrasts to our analysis which statically validates the contracts. Beckman et
al. introduce a methodology based on typestate that statically verifies if a protocol of an object
is respected [4]. This approach requires the programmer to explicitly unpack objects before
it can be used. Hurlin [15] extends the work of Cheon to support protocols in concurrent
scenarios. The protocol specification is extended with operators that allow methods to be
executed concurrently, and pre-conditions that have to be satisfied before the execution of a
method. This analysis is statically verified by a theorem prover. Theorem proving, in general,
is very limited since automated theorem proving tend to be inefficient.
Peng Liu et al. developed a way to detect atomicity violations caused by method composition [18], much like the ones we describe in this paper. They define two patterns that
are likely to cause atomicity violations, one capturing stale value errors and the other one by
trying to infer a correlation between method calls by analyzing the control flow graph (if a() is
executed before b() and b() is executed after a()). This patterns are captured statically and
then filtered with a dynamic analysis.
Many works can be found about atomicity violations. Artho et al. in [2] define the notion of
high-level data races, that characterize sequences of atomic operations that should be executed
atomically to avoid atomicity violations. The definition of high-level data races do not totally
capture the violations that may occur in a program. Praun and Gross [27] extend Artho’s
approach to detect potential anomalies in the execution of methods of an object and increase
the precision of the analysis by distinguish between read and write accesses to variables shared
between multiple threads. An additional refinement to the notion of high-level data races was
introduced by Pessanha in [10], relaxing the properties defined by Artho, which results in a
higher precision of the analysis. Farchi et al. [11] propose a methodology to detect atomicity
violations in the usage of modules based on the definition of high-level data races. Another
common type of atomicity violations that arise when sequencing several atomic operations
are stale value errors. This type of anomaly is characterized by the usage of values obtained
atomically across several atomic operations. These values can be outdated and compromise
the correct execution of the program. Various analysis were developed to detect these types
of anomalies [3, 6, 10]. Several other analysis to verify atomicity violations can be found in
the literature, based on access patterns to shared variables [19, 26], type systems [7], semantic
invariants [9], and other specific methodologies [12–14].
11
Concluding Remarks
In this paper we present the problem of atomicity violations when using a module, even when
their methods are individually synchronized by some concurrency control mechanism. We
propose a solution based on the design by contract methodology. Our contracts define which
call sequences to a module should be executed in an atomic manner.
We introduce a static analysis to verify these contracts. The proposed analysis extracts the
behavior of the client’s program with respect to the module usage, and verifies whether the
contract is respected.
A prototype was implemented and the experimental results shows the analysis is highly
precise and can run efficiently on real-world programs.
References
[1] Frances E. Allen. Control flow analysis. SIGPLAN Not., 5(7):1–19, July 1970.
15
[2] Cyrille Artho, Klaus Havelund, and Armin Biere. High-level data races. Software Testing,
Verification and Reliability, 13(4):207–227, December 2003.
[3] Cyrille Artho, Klaus Havelund, and Armin Biere. Using block-local atomicity to detect
stale-value concurrency errors. Automated Technology for Verification and Analysis, pages
150–164, 2004.
[4] Nels E. Beckman, Kevin Bierhoff, and Jonathan Aldrich. Verifying correct usage of atomic
blocks and typestate. SIGPLAN Not., 43(10):227–244, October 2008.
[5] Jean-Francois Bergeretti and Bernard A Carré. Information-flow and data-flow analysis of
while-programs. ACM Transactions on Programming Languages and Systems (TOPLAS),
7(1):37–61, 1985.
[6] M. Burrows and K.R.M. Leino. Finding stale-value errors in concurrent programs. Concurrency and Computation: Practice and Experience, 16(12):1161–1172, 2004.
[7] Luís Caires and João C. Seco. The type discipline of behavioral separation. SIGPLAN
Not., 48(1):275–286, January 2013.
[8] Yoonsik Cheon and Ashaveena Perumandla. Specifying and checking method call sequences of java programs. Software Quality Control, 15(1):7–25, March 2007.
[9] R. Demeyer and W. Vanhoof. A framework for verifying the application-level race-freeness
of concurrent programs. In 22nd Workshop on Logic-based Programming Environments
(WLPE 2012), page 10, 2012.
[10] Ricardo J. Dias, Vasco Pessanha, and João M. Lourenço. Precise detection of atomicity violations. In Hardware and Software: Verification and Testing, Lecture Notes in Computer
Science. Springer Berlin / Heidelberg, November 2012. HVC 2012 Best Paper Award.
[11] Eitan Farchi, Itai Segall, João M. Lourenço, and Diogo Sousa. Using program closures to
make an application programming interface (api) implementation thread safe. In Proceedings of the 2012 Workshop on Parallel and Distributed Systems: Testing, Analysis, and
Debugging, PADTAD 2012, pages 18–24, New York, NY, USA, 2012. ACM.
[12] Cormac Flanagan and Stephen N Freund. Atomizer: a dynamic atomicity checker for
multithreaded programs. SIGPLAN Not., 39(1):256–267, January 2004.
[13] Cormac Flanagan and Stephen N. Freund. FastTrack: efficient and precise dynamic race
detection. Commun. ACM, 53(11):93–101, November 2010.
[14] Cormac Flanagan, Stephen N. Freund, and Jaeheon Yi. Velodrome: a sound and complete
dynamic atomicity checker for multithreaded programs. SIGPLAN Not., 43(6):293–303,
June 2008.
[15] Clément Hurlin. Specifying and checking protocols of multithreaded classes. In Proceedings
of the 2009 ACM symposium on Applied Computing, SAC ’09, pages 587–592, New York,
NY, USA, 2009. ACM.
[16] IBM’s Concurrency Testing Repository.
[17] Donald E Knuth. On the translation of languages from left to right. Information and
control, 8(6):607–639, 1965.
[18] Peng Liu, Julian Dolby, and Charles Zhang. Finding incorrect compositions of atomicity.
In Proceedings of the 2013 9th Joint Meeting on Foundations of Software Engineering,
pages 158–168. ACM, 2013.
16
[19] J. Lourenço, D. Sousa, B. Teixeira, and R. Dias. Detecting concurrency anomalies in
transactional memory programs. Computer Science and Information Systems/ComSIS,
8(2):533–548, 2011.
[20] Shan Lu, Soyeon Park, Eunsoo Seo, and Yuanyuan Zhou. Learning from mistakes: a
comprehensive study on real world concurrency bug characteristics. SIGPLAN Not.,
43(3):329–339, March 2008.
[21] Bertrand Meyer. Applying "design by contract". Computer, 25(10):40–51, October 1992.
[22] Vasco Pessanha. Verificação prática de anomalias em programas de memória transaccional
(Practical verification of anomalies in transactional memory programs). Master’s thesis,
Universidade Nova de Lisboa, 2011.
[23] Jan Rekers and Wilco Koorn. Substring parsing for arbitrary context-free grammars.
ACM Sigplan Notices, 26(5):59–66, 1991.
[24] Masaru Tomita. An efficient augmented-context-free parsing algorithm. Comput. Linguist., 13(1-2):31–46, January 1987.
[25] Raja Vallée-Rai, Phong Co, Etienne Gagnon, Laurie Hendren, Patrick Lam, and Vijay
Sundaresan. Soot - a java bytecode optimization framework. In Proceedings of the 1999
conference of the Centre for Advanced Studies on Collaborative research, CASCON ’99,
pages 13–. IBM Press, 1999.
[26] M. Vaziri, F. Tip, and J. Dolby. Associating synchronization constraints with data in an
object-oriented language. In ACM SIGPLAN Notices, volume 41, pages 334–345. ACM,
2006.
[27] C. Von Praun and T.R. Gross. Static detection of atomicity violations in object-oriented
programs. Journal of Object Technology, 3(6):103–122, 2004.
17
12
12.1
Appendix
Grammar Example
@Contract ( clauses =...)
class Module
{
public Module () { }
public void a () { }
public void b () { }
public void c () { }
}
public class Main
{
private static Module m ;
private static void f ()
{
m . c ();
}
@Atomic
private static void g ()
{
m . a ();
m . b ();
f ();
}
public static void main ( String [] args )
{
m = new Module ();
for ( int i =0; i < 10; i ++)
if ( i %2 == 0)
m . a ();
else
m . b ();
f ();
g ();
}
}
Figure 7: Program.
18
Start: A
A -> B
D -> E F
E -> X
F -> G
G -> H
B -> C
C -> D
L -> M
L -> N
M -> O
N -> R
O -> a P
H -> I
I -> K
I -> J
J -> L
K -> T U
BF -> BG
U -> V W
BG -> b BH
T -> BA
BH -> T BI
W -> epsilon
BI -> epsilon
V -> BD
BB -> c BC
Q -> S
BC -> epsilon
P -> Q
BD -> BE
S -> H
BE -> a BF
R -> b Q
Y -> Z
X -> Y
Z -> epsilon
BA -> BB
Figure 8: Non-optimized grammar.
19
Start: A’
A’ -> A
A -> I
L -> b I
L -> a I
I -> L
I -> T V
T -> c
V -> a b T
Figure 9: Optimized grammar.
20
| 6cs.PL
|
arXiv:1708.06012v2 [cs.IT] 28 Dec 2017
Product Matrix Minimum Storage Regenerating
Codes with Flexible Number of Helpers
Kaveh Mahdaviani
Soheil Mohajer
Ashish Khisti
ECE Dept., University of Toronto
Toronto, ON M5S3G4, Canada
Email: [email protected]
ECE Dept., University of Minnesota
Minneapolis, MN 55404, USA
Email: [email protected]
ECE Dept., University of Toronto
Toronto, ON M5S3G4, Canada
Email: [email protected]
Abstract—In coding for distributed storage systems, efficient
data reconstruction and repair through accessing a predefined
number of arbitrarily chosen storage nodes is guaranteed by
regenerating codes. Traditionally, code parameters, specially the
number of helper nodes participating in a repair process, are
predetermined. However, depending on the state of the system
and network traffic, it is desirable to adapt such parameters
accordingly in order to minimize the cost of repair. In this work
a class of regenerating codes with minimum storage is introduced
that can simultaneously operate at the optimal repair bandwidth,
for a wide range of exact repair mechanisms, based on different
number of helper nodes.
I. I NTRODUCTION
For a distributed storage system (DSS) reliability and accessibility are the most important features. With large scale
DSSs nowadays it is common to lose access to a storage node
or part of its content. Hence, both reliability and accessibility
depend on system’s capability to replace a failed node by a
new one, and recover its content. This procedure is referred to
as repair. For the DSS to be capable of repair, it is necessary
to store redundancy. It is shown that there exits an information
theoretic tradeoff between the amount of redundancy (i.e. storage overhead), and the amount of data transmission required
for a repair, referred to as repair bandwidth [1].
Among various methods of storing redundancy and performing repair, a specific class of erasure codes, named
regenerating codes, offers the efficient performance [1]. More
precisely, a regenerating code on a Galois field Fq for a DSS
with n storage nodes, maps the source data of size F symbols
into n pieces of size α symbols each, and stores them in n
separate nodes, such that any k out of n nodes suffice to
recover the data. Such system is capable to tolerate up to
(n − k) node failures. Moreover, upon failure of one node, it
can be replaced by a new node whose content is determined by
connecting to an arbitrary set of d (where d ≥ k) helper nodes,
and downloading β symbols form each (where β ≤ α). Ideally,
one would like to minimize the storage overhead, and repair
bandwidth simultaneously. It turns out that for a given file size
F , there is a tradeoff between the per-node storage capacity α
and the repair bandwidth γ = dβ, and one can be minimized
only at the cost of a penalty for the other [1]. In particular,
The work of S. Mohajer is supported by the National Science Foundation
under Grant CCF-1617884.
at one extreme point of this tradeoff, one could first minimize
the per-node storage, α, and then minimize the per-node
repair bandwidth, β, to obtain a minimum storage regenerating
(MSR) code. As a result, MSR codes have the maximum
distance separable (MDS) property, and also minimize the
repair bandwidth for the given α [1], which means for an
MSR code we have F = kα, and
F
.
(1)
β=
k(d − k + 1)
In other words, in MSR codes the repair bandwidth is decreasing super-linearly as d grows.
Reversing the order of minimization between α, and β
results in another extreme point of the tradeoff, which provides
the minimum repair bandwidth (MBR) regenerating codes.
Our focus in this work is on MSR codes as they minimize
the storage cost. Moreover, we only consider the repair mechanisms in which the replacement node contains exactly the same
content as stored in the failed node. Such repair mechanisms
are referred to as exact repair, and enable the code to be
systematic, which is a significant advantage in practice.
The common adopted model in regenerating codes considers
a predetermined number d (where k ≤ d ≤ n − 1) of helpers
required for any repair procedure. Each of these helpers is also
assumed to provide β = γ/d repair bandwidth. This sets a
threshold for the system’s capability to perform repair. On the
other hand, in practice the state of system dynamically changes
as a function of various factors including traffic load, available
bandwidth, etc. Therefore, runtime adaptation would be of
great value towards optimizing the performance. For instance,
when the system is heavily loaded by many read requests, there
might be only very few nodes available to serve as helpers.
In this situation we are interested in optimal repair based on
the available helpers. Likewise, when there are many helpers
available it is beneficial to use a large number of helpers as
increasing d reduces both γ and β in optimal repair mechanism
characterized by equation (1). This could then reduce both the
total network traffic as well as the transmission delay. We refer
to such property as bandwidth adaptive.
The design of such codes has been of interest and the
significance of bandwidth adaptivity in the performance of
the system has been emphasised in [2]–[5]. However, it is
a challenging problem to design such coding scheme with a
large flexibility degree since it needs to satisfy many optimality
conditions simultaneously. As a result, this problem has only
been considered for the MSR [3], [6], and MBR [7], [8]
extreme points of the tradeoff. For the MBR case, [7], [8]
provided a solution for a wide range of practical parameters
based on the Product Matrix framework introduced in [9]. In
[3] a solution is provided based on interference alignment,
which only achieves the MSR characteristics when both α
and β tend to infinity. The first explicit exact repair MSR
code constructions which satisfy the bandwidth adaptivity are
introduced in [6]. These constructions work for any parameters
k, n, and all values of d such that k < d < n. Although
these constructions can achieve optimality for finite values
of α and β, but the required value for these parameters are
still very huge (i.e. exponentially large in n), and hence they
only achieve optimality for extremely large contents. Recently,
[10] introduced a modified version of the codes in [6] which
achieves MSR optimality for much lower values of α, at the
cost of loosing bandwidth adaptivity. Indeed the MSR code in
[10] works only for d = n−1. In [11], d < n−1, and practical
α is achieved for MSR codes without bandwidth adaptivity.
In this work we address the design of MSR codes with
bandwidth adaptive exact repair for small α, and β, following
the Product Matrix framework [9]. The code allows us to
choose the number of helper nodes for each repair scenario
independently, and it is capable to adjust the per-node repair
bandwidth to its optimum value based on the number of
selected helpers as in (1). Compared to the constructions
proposed in [6] for a DSS with n storage nodes the required
values for α and β in the presented code is reduced to the
nth root for the same set of other parameters. The main
contributions of this work are explained in the next section,
after formally defining the problem setup.
II. M ODEL AND M AIN R ESULTS
A. Model
The first element we consider for the model of our bandwidth adaptive distributed storage system is a predefined
Galois field alphabet, Fq of size q. Hereafter we assume all
the symbols stored or transmitted through the network are
elements of Fq . Besides, we will consider a homogeneous
group of n storage nodes, each capable of storing α symbols.
Definition 1 (Bandwidth Adaptive Regenerating Code). Consider the set of parameters α, n, k, δ, a set D = {d1 , · · · , dδ },
with d1 < · · · < dδ , and a total repair bandwidth function
γ : D → [α, ∞). A bandwidth adaptive regenerating code
C(n, k, D, α, γ) is a regenerating code with per-node storage
capacity α, such that in each repair process the number of
helpers, d, can be chosen arbitrarily from the set D. The
choice of helper nodes is also arbitrary, and each of the
chosen helpers then provides β(d) = γ(d)/d repair symbols.
Moreover, the data collector recovers the whole source data
by accessing any arbitrary set of k nodes.
Note that the flexibility of the repair procedure depends on
the parameter δ, such that for a larger δ, there are more options
to select the number of helpers. In general, it is appealing to
have small choices such as d1 , to guarantee the capability
of code to perform repair when the number of available
helpers is small, and also large choices such as dδ , to provide
the capability of reducing the per-node repair bandwidth and
hence the transmission delay whenever a larger number of
helpers are available. The coding scheme we present in this
work allows to design such a range for the elements in D.
Definition 2 (Total Storage Capacity). For the set of parameters α, n, k, δ, a set D = {d1 , · · · , dδ }, and a given function
γ : D → [α, ∞), the total storage capacity of a bandwidth
adaptive distributed storage system is the maximum size of
a file that could be stored in a network of n storage nodes
with per-node storage capacity α, using a bandwidth adaptive
regenerating code C(n, k, D, α, γ). We will denote the storage
capacity of such a system by F (n, k, D, α, γ), or simply F
when the parameters could be inferred from the context.
Definition 3 (Bandwidth Adaptive MSR Codes, and the Flexibility Degree). For any choice of parameters α, n, k, δ, and
set D = {d1 , · · · , dδ }, the bandwidth adaptive regenerating
codes that realize both the MDS property defined by F (n, k,
D, α, γ) = kα, as well as the the MSR characteristic equation
simultaneously for all d ∈ D, given as,
α = (d − k + 1)β(d), ∀d ∈ D,
(2)
is referred to as bandwidth adaptive MSR codes. Moreover, the
number of elements in the set D is referred to as flexibility
degree of the code, and is denoted by δ.
B. Main Results
The main contribution of this work is to provide a bandwidth
adaptive MSR coding scheme with small per-node storage
requirement. This coding scheme also guarantees exact repair
for different choices of the number of helpers. This result is
formally stated in the form of the following theorem. In this
paper lcm() denotes the least common multiple.
Theorem 1. For arbitrary positive integers n, k, and δ, there
exists an adaptive bandwidth MSR code, with a finite per-node
storage capacity α and total storage capacity F , satisfying
α = (k − 1)lcm (1, 2, · · · , δ) ,
F = kα,
which is capable of performing exact repair using any arbitrary di helpers, for
di = (i + 1)(k − 1), i ∈ {1, · · · , δ},
and simultaneously satisfies the MSR characteristic equation
(2) for any di . i.e.,
α
, i ∈ {1, · · · , δ}.
β(di ) =
(di − k + 1)
Section III provides a constructive proof for this theorem.
III. C ODING S CHEME
The coding scheme presented in this work is closely related
to the Product Matrix MSR code introduced in [9], and could
be considered as an extension of the Product Matrix code, that
achieves bandwidth adaptivity. To demonstrate this connection
we will try to follow the notation used in [9].
In the design of the proposed coding scheme, we chose a
design parameter µ, and the required flexibility degree δ. All
the other parameters of the code including α, F , k, D =
{d1 , · · · , dδ }, and β(di ) will be then determined based on µ,
and δ as follows. The per-node storage capacity is
α = µ · lcm (1, · · · , δ) .
(3)
Moreover, we have k = µ + 1, and F = (µ + 1)α, which
satisfies the MDS property. Finally, for D we have
D = {d1 , · · · , dδ }, di = (i + 1)µ, i ∈ {1, · · · , δ}. (4)
and for any di ∈ D, the associated per-node and total repair
bandwidths denoted by β(di ), and γ(di ) respectively are
(i + 1)α
α
γ(di ) = di β(di ) =
.
(5)
β(di ) = ,
iµ
i
A. Coding for Storage
We begin the introduction of the coding scheme by describing the process of encoding the source symbols and deriving
the encoded symbols to be stored in the storage nodes. Similar
to the product matrix codes, the first step in encoding for
storage in this scheme is to arrange the information symbols
in a matrix, denoted by M , which we refer to hereafter as the
data matrix. Let
zδ = lcm (1, · · · , δ) .
The data matrix in our coding scheme is structured as follows,
S1 S2 O O O
O
···
O
S2 S3 S4 O O
O
···
O
O S4 S5 S6 O
O
···
O
O O S6 S7 S8
O
···
O
M = .
.. , (6)
..
..
.
.
O ···
O
S
S
S
2zδ −4 2zδ −3 2zδ −2
O ···
O
O S2zδ −2 S2zδ −1
O ···
O
O
O
S2zδ
where, each Si , i ∈ {1, · · · , 2zδ } is a symmetric µ × µ matrix
filled with µ(µ + 1)/2 source symbols, and O is a µ × µ zero
matrix. Therefore, M ’s dimensions are (zδ + 1)µ × zδ µ. Note
that the total number of distinct source symbols is
µ(µ + 1)
F = 2zδ
= kα.
2
The source encoder then creates the vector of coded symbols
for each of the n storage nodes, by calculating the product
of a node-specific coefficient vector and the data matrix. To
describe this process, we first need the following definition.
Definition 4 (Generalized Vandermonde Matrix). For distinct,
non-zero elements e1 , · · · , em in Fq , and some integer c ≥ 0,
a matrix Am×ℓ with entries Ai,j = eic+j−1 is referred to as a
generalized Vandermonde matrix.
In particular, for distinct, non-zero elements ei ’s in Fq , with
i ∈ {1, · · · , n} we define a generalized Vandermonde matrix
of size n × (zδ + 1)µ
as
(z +1)µ
e1 e21 · · · e1 δ
(z +1)µ
e2 e22 · · · e2 δ
.
Ψ=
..
.
(zδ +1)µ
2
en en · · · en
Note that submatrices of Ψ are also generalized Vandermonde
matrices. Moreover, one can show that any square generalized
Vandermonde matrix is invertible [12].
We denote the j th row of Ψ by ψ j . Then the vector of
encoded symbols to be stored on node j, j ∈ {1, · · · , n},
denoted by xj , is calculated as
xj = ψ j M.
Note that the per-node storage capacity requirement for this
coding scheme is then zδ µ as given by (3).
B. Data Reconstruction
In order to reconstruct all the information stored in the
system, the data collector accesses k arbitrary nodes in the
network and downloads all their contents. To describe the
details of the decoding we use the following lemma.
Lemma 1. Let X and Ψ be two known generalized Vandermonde matrices of size (µ+1)×µ, and ∆ be a known diagonal
matrices of size (µ+1)×(µ+1) with non-zero distinct diagonal
elements. Then one can uniquely solve the equation
X = ΨA + ∆ΨB,
for unknown µ × µ symmetric matrices A, and B.
The proof of this lemma is presented in [12]. The following
theorem explains the data reconstruction procedure.
Theorem 2. For the coding scheme presented in subsection
III-A, there exists a decoding scheme to reconstruct all source
symbols arranged in the data matrix M from the encoded
content of any arbitrary group of k = µ + 1 storage nodes.
Proof. Let’s assume the set of accessed nodes is {ℓ1 , · · · , ℓk }.
Moreover, let’s denote the k × (zδ + 1)µ submatrix of Ψ
associated with the nodes ℓ1 , · · · , ℓk , by ΨDC . We will further
denote the submatrix of ΨDC consisting of columns (i−1)µ+1
through iµ, by ΨDC (i). In other words, we have a partitioning
of ΨDC ’s columns as
ΨDC = [ΨDC (1), · · · , ΨDC (zδ + 1)] .
As a result, defining the
µdiagonal matrix
e ℓ1 0 0 · · · 0
0 eµℓ 0 · · · 0
2
ΛDC = .
.. ,
..
..
. .
0
0 0 · · · eµℓk
we have
ΨDC (i + 1) = ΛDC ΨDC (i).
(7)
Similarly, let’s denote the matrix consisting of the collected
encoded vectors by XDC , and its partitioning to k × µ submatrices XDC (i), i ∈{1, ·· · , zδ } as follows
xℓ1
..
XDC = . = [XDC (1), · · · , XDC (zδ )] .
xℓk
The decoding procedure for data reconstruction consists of
zδ consecutive steps. The first step uses only the submatrix
XDC (1). Similar to the data reconstruction for product matrix
MSR codes, using (7) we have,
S
XDC (1) = [ΨDC (1), ΨDC (2)] 1
S2
= ΨDC (1)S1 + ΛDC ΨDC (1)S2 .
Using Lemma 1, the decoder recovers both S1 , and S2 ,
using XDC (1), in step one. Then, for i ∈ {2, · · · , zδ }, the
decoder performs step i by using submatrix XDC (i), and
decodes submatrices S2i−1 , and S2i , as follows.
In step i of the data reconstruction decoding, having the
submatrix S2(i−1) already recovered from step i − 1, the
decoder first calculates
X̂DC (i) = XDC (i) − ΨDC (i − 1)S2(i−1) .
S2i−1
.
(8)
= [ΨDC (i), ΨDC (i + 1)]
S2i
Then from (7), and (8), we have
X̂DC (i) = ΨDC (i)S2i−1 + ΛDC ΨDC (i)S2i .
Again using Lemma 1, decoder recovers S2i−1 , and S2i at the
end of the step i of the decoding. Hence, by finishing step zδ ,
decoder reconstructs all the submatrices in M .
C. Bandwidth Adaptive Exact Repair
We now describe the bandwidth adaptive repair procedure,
by assuming that node f is failed and the set of helpers
selected for the repair are H = {ℓ1 , · · · , ℓd }, for some
arbitrary d ∈ D. The following theorem describes the repair
procedure in this bandwidth adaptive MSR code.
Theorem 3. Consider the coding scheme presented in subsection III-A, with design parameters µ, and δ, and D as defined
in (4). For any arbitrary failed node f , and any arbitrary set
of helpers H = {ℓ1 , · · · , ℓd }, for some d ∈ D, there exists
a repair scheme for recovering the content of node f with
per-node repair bandwidth,
α
.
(9)
β(d) =
d−µ
Proof. Without loss of generality let d = (m + 1)µ, for some
m ∈ {1, · · · , δ}. Each helper node h ∈ H, creates β(d) =
α/(d − µ) repair symbols to repair node f as follows. First h
partitions its encoded content into β(d) equal segments as
(10)
xh = [xh (1), · · · , xh (β(d))] .
Note that (3), and (4) guarantee that for any d ∈ D, α is
an integer multiple of d − µ, hence β(d) is an integer. Each
segment xh (i) is then of size d − µ = mµ. Similarly, we split
the first α entries of a coefficient vector assigned to node ℓ,
namely ψ ℓ , into β(d) equal segments as
h
i
ψ ℓ (1 : α) = ψ ℓ (1), · · · , ψ ℓ (β(d)) ,
(11)
where each segment ψ ℓ (i) is of size d − µ = mµ.
Now each helper node h ∈ H, creates
its repair symbols as
r(h, f ) = r1 (h, f ), · · · , rβ(d) (h, f )
h
⊺
⊺ i
= xh (1) ψ f (1) , · · · , xh (β(d)) ψ f (β(d))
.
(12)
The repair decoder then receives
a
d
×
β(d)
matrix
r(ℓ1 , f )
..
ΥH =
.
.
r(ℓd , f )
We then introduce the following partitioning of the matrix ΥH ,
into β(d) submatrices, as follows
ΥH = [ΥH (1), · · · , ΥH (β(d))] ,
(13)
where ΥH (i), i ∈ {1, · · · , β} is the ith column of ΥH .
Before starting to describe the repair decoding procedure,
we need to introduce some notations associated to a given repair scenario. Consider a repair procedure with d = (m + 1)µ,
d ∈ D. For the corresponding β(d) = α/(d − µ) we
will partition matrix M as depicted in Fig. 1. Note that
this results in β(d) non-overlapping diagonal submatices Mi ,
i ∈ {1, · · · , β(d)}, each of size mµ × mµ, along with µ × µ
submatrices S2m , S4m , · · · , S2β(d)m = S2zδ as shown in the
figure. From the construction of the data matrix, introduced in
(6), each Mi submatrix will be symmetric. As a result, the data
matrix M could be interpreted in terms of the submatrices Mi ,
and S2i for i ∈ {1, · · · , β(d)}, associated to a repair procedure
with d = (m + 1)µ, d ∈ D.
In addition, for any node ℓ, we introduce the vector φℓ as,
φℓ = [eℓ , · · · , eµℓ ] .
(14)
Finally the last notations we use to describe the adaptive
repair decoding scheme, using a given set of helpers H =
{ℓ1 , · · · , ℓd },is,
i
h
φ
ψ ℓ (i), eimµ
ℓ1
ℓ1
1
.
.
ΩH (i) =
, i ∈ {1, · · · , β}. (15)
.
h
i
imµ
ψ ℓ (i), eℓd φℓ
d
d
Note that, ΩH (i), i ∈ {1, · · · , β(d)}, is a d × d generalized
Vandermonde matrix and hence is invertible. We also use the
following notations for submatrices
of the
inverse of ΩH (i),
ΘH (i)
−1
(ΩH (i)) =
,
(16)
ΞH (i)
where ΘH (i) represents the top (d − µ) × d submatrix, and
ΞH (i), the bottom µ × d submatrix.
The decoding procedure for the repair of node f is performed in β(d) sequential steps. In the first step, the decoder
only uses the first repair symbol received from each of the
helpers, namely r1 (ℓi , f ), for i ∈ {1, · · · , d}.
Using (10), (11), (13), and (15), and the partitioning denoted
in Fig. 1, the submatrix ΥH (1), introduced in (13) can be
written as
⊺
M1
ΥH (1) = ΩH (1)
ψ f (1) . (17)
O · · · O S2m
Multiplying the inverse of ΩH (1) from right to the both sides
of (17), and using
(16) thedecoder derives
⊺
(18)
M1 ψ f (1) = ΘH (1)ΥH (1),
and similarly, using
(14),
⊺
(m−1)µ
= ΞH (1)ΥH (1).
(19)
S2m ef
φf
Since both M1 , and S2m are symmetric, from (18) we have,
⊺
(20)
ψ f (1)M1 = (ΘH (1)ΥH (1)) ,
µ
and from (19), by multiplying the scalar ef , we get
⊺
µ
emµ
(21)
f φf S2m = ef (ΞH (1)ΥH (1)) .
Using a partitioning similar to (10) for xf , from (20), and (21)
the decoder then recovers hxf (1) as,
i
xf (1) = ψ f (1)M1 + O, · · · , O, emµ
f φf S2m
µ×mµ
,
0
M2
M=
S2(i−1)m+1
S2(i−1)m+2
O
..
, Mi =
.
O
M1
Mi
0
Mβ
S2(i−1)m+2
S2(i−1)m+3
S2(i−1)m+4
S2(i−1)m+4
S2(i−1)m+5
O
O
O
O
O
O
···
···
O
O
S2(i−1)m+6
..
.
O
···
O
..
.
···
O
S2im−4
S2im−3
S2im−2
···
O
O
S2im−2
S2im−1
.
Fig. 1: In the above figure β represents β(d). Moreover, each of the small coloured squares represent a non-zero submatrix of M ; blue:
S2m , red: S4m , light green: S2(i−1)m , yellow: S2im , dark green: S2(β(d)−1)m , and black: S2zδ .
where, the rightmost term in the above expression is derived
by padding m − 1, µ × µ zero matrices, O, to the left of the
matrix calculated in (21).
In step i for i = 2 through β(d) of the repair decoding,
the decoder then recovers xf (i), using ΥH (i) received from
(i−1)mµ
the helpers, along with ef
φf S2(i−1)m , recovered from
the step i − 1 of decoding. To this end, the decoder first
removes the contribution of the S2(i−1)m submatrix in the
repair symbols inΥH (i) by calculating
(i−1)mµ−µ
φℓ
e ℓ1
1
⊺
..
S2(i−1)m e(i−1)mµ φ
.
Υ̂H (i)=ΥH (i)−
f
.
f
(i−1)mµ−µ
e ℓd
φℓ
d
⊺
(i−1)mµ
is itself
In the above expression, S2(i−1)m ef
φf
(i−1)mµ
derived by transposing ef
φf S2(i−1)m . Hence we have,
⊺
Mi
Υ̂H (i) = ΩH (i)
ψ f (i) .
O · · · O S2im
Therefore, similar to (18) through (21) the decoder derives,
⊺
(22)
ψ f (i)Mi = (ΘH (i)ΥH (i)) ,
and
eimµ
φf S2im = eµf (ΞH (i)ΥH (i))⊺ .
(23)
f
Finally, using (22) and (23),
we
have
h
i
φf S2im
xf (i) = ψ f (i)Mi + O, · · · , O, eima
.
f
µ×mµ
Remark 1. In a DSS with n nodes, for D = {d1 , · · · , dδ },
the bandwidth adaptive MSR codes presented in [6], although
support any rate, require
α = (lcm (d1 − k + 1, · · · , dδ − k + 1))n .
(24)
Comparing (24) with (3), one could see that the presented
scheme reduces the required α (and β) values to the nth root.
However, this scheme works only for 2k − 1 < di , ∀di ∈ D.
Hence, the design of high-rate bandwidth adaptive MSR codes
with small α and β still remains an open problem.
IV. C ONCLUSION
We presented an alternative solution for exact-repair MSR
codes in which optimal exact repair is guaranteed simultaneously with a range of choices, D = {d1 , · · · , dδ }, for
the number of helpers. Comparing to the only other explicit
constructions, presented in [6], we showed that when di ≥
2k − 1, ∀di ∈ D, the required values for α, and β are reduced
to the nth root for a DSS with n nodes.
R EFERENCES
[1] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. J. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems,” IEEE
Transactions on Information Theory, vol. 56, no. 9, pp. 4539–4551,
Sep. 2010.
[2] A.-M. Kermarrec, N. L. Scouarnec, and G. Straub, “Repairing multiple
failures with coordinated and adaptive regenerating codes,” in Proc.
IEEE International Symposium on Network Coding (NetCod), Beijing,
China, July 2011, pp. 1–6.
[3] V. Cadambe, S. Jafar, H. Maleki, K. Ramchandran, and C. Suh,
“Asymptotic interference alignment for optimal repair of MDS codes in
distributed storage,” IEEE Transactions on Information Theory, vol. 59,
no. 5, pp. 2974–2987, May 2013.
[4] V. Aggarwal, C. Tian, V. A. Vaishampayan, and Y.-F. R. Chen, “Distributed data storage systems with opportunistic repair,” in Proc. IEEE
International Conference on Computer Communications (INFOCOM),
Toronto, Canada, April–May 2014, pp. 1833–1841.
[5] M. Hajiaghayi and H. Jafarkhani, “MDS codes with progressive
engagement property for cloud storage systems,” Available online:
https://arxiv.org/abs/1605.06927, 2016.
[6] M. Ye and A. Barg, “Explicit constructions of high-rate MDS array
codes with optimal repair bandwidth,” IEEE Transactions on Information Theory, vol. 63, no. 4, pp. 2001–2014, April 2017.
[7] K. Mahdaviani, A. Khisti, and S. Mohajer, “Bandwidth adaptive & error
resilient regenerating codes with minimum repair bandwidth,” in Proc.
IEEE International Symposium on Information Theory (ISIT), Barcelona,
Spain, July 2016, pp. 235–239.
[8] ——, “Bandwidth adaptive & error resilient MBR exact repair regenerating codes,” Available online: https://arxiv.org/pdf/1711.02770.pdf, 2017.
[9] K. V. Rashmi, N. B. Shah, and P. V. Kumar, “Optimal exact regenerating
codes for distributed storage at the MSR and MBR points via a productmatrix construction,” IEEE Transactions on Information Theory, vol. 57,
no. 8, pp. 5227–5239, August 2011.
[10] M. Ye and A. Barg, “Explicit constructions of optimal-access MDS
codes with nearly optimal sub-packetization,” IEEE Transactions on
Information Theory, vol. 63, no. 10, pp. 6307–6317, Oct. 2017.
[11] B. Sasidharan, M. Vajha, and P. V. Kumar, “An explicit, coupledlayer construction of a high-rate MSR code with low sub-packetization
level, small field size and d < (n − 1),” in Proc. IEEE International
Symposium on Information Theory (ISIT), Aachen, Germany, June 2017,
pp. 2048–2052.
[12] K. Mahdaviani, S. Mohajer, and A. Khisti, “Product matrix MSR
codes with bandwidth adaptive exact repair,” Available online:
https://arxiv.org/abs/1708.03402, 2017.
| 7cs.IT
|
arXiv:1508.04865v6 [math.GR] 30 May 2017
SUBGROUPS OF RELATIVELY HYPERBOLIC GROUPS OF BREDON
COHOMOLOGICAL DIMENSION 2
EDUARDO MARTÍNEZ-PEDROZA
Abstract. A remarkable result of Gersten states that the class of hyperbolic groups of cohomological dimension 2 is closed under taking finitely presented (or more generally FP2 )
subgroups. We prove the analogous result for relatively hyperbolic groups of Bredon cohomological dimension 2 with respect to the family of parabolic subgroups. A class of groups
where our result applies consists of C ′ (1/6) small cancellation products. The proof relies
on an algebraic approach to relative homological Dehn functions, and a characterization of
relative hyperbolicity in the framework of finiteness properties over Bredon modules and
homological Isoperimetric inequalities.
1. Introduction
It is a remarkable result of Gersten that the class of hyperbolic groups of cohomological
dimension at most 2 is closed under taking finitely presented subgroups, see Theorem 1.1.
The dimension assumption is sharp since Brady exhibited a hyperbolic group of cohomological dimension 3 containing a finitely presented subgroup which is not hyperbolic [5].
Theorem 1.1 (Gersten). [15, Theorem 5.4] Let G be a hyperbolic group and suppose
cd(G) ≤ 2. If H is a subgroup of type FP2 , then H is hyperbolic.
In this article we show that such subgroup phenomena hold in the class of relatively
hyperbolic groups of Bredon cohomological dimension 2 with respect to parabolic subgroups. We use the approach to relative hyperbolicity by Bowditch [4] which is equivalent
to the approach by Osin [27] for finitely generated groups [27, Theorem 6.10].
Definition 1.2. (Bowditch relatively hyperbolic groups)[4] A graph Γ is fine if every edge is
contained in only finitely many circuits (embedded closed paths) of length n for any integer
n. A finitely generated group G is hyperbolic relative to a finite collection of subgroups
P if G acts on a connected, fine, hyperbolic graph Γ with finite edge stabilizers, finitely
many orbits of edges, and P is a set of representatives of distinct conjugacy classes of
vertex stabilizers (such that each infinite stabilizer is represented). Subgroups of G which
are conjugate to a subgroup of some P ∈ P are called parabolic subgroups.
A family of subgroups F of a group G is a non-empty collection of subgroups closed
under conjugation and subgroups. The theory of (right) modules over the orbit category
OF (G) was established by Bredon [8], tom Dieck [29] and Lück [20]. In the case that F
is the trivial family, the OF (G)-modules are ZG-modules. In this context, the notions of
cohomological dimension cdF (G) and finiteness properties FPn,F for the pair (G, F ) are
defined analogously as their counterparts cd(G) and FPn , see Section 3.
2000 Mathematics Subject Classification. 20F67, 20F65, 20J05, 57S30, 57M60, 55N25.
Key words and phrases. Isoperimetric Functions, Dehn functions, Hyperbolic Groups, Relatively Hyperbolic
groups, Finiteness properties, Cohomological dimension, Homological isoperimetric inequalities, Orbit category,
Bredon modules.
1
2
E. MARTÍNEZ-PEDROZA
To simplify the notation in our statements we introduce the following non-standard
terminology. A group G is hyperbolic relative to a family of subgroups F if there is a finite
collection of subgroups P generating the family F , such that G is hyperbolic relative to P.
If H ≤ G and F is a family of subgroups, we denote by F ∩ H the family of subgroups
{K ∩ H : K ∈ F }. Our main result is the following.
Theorem 1.3. Let G be hyperbolic relative to a family F containing all finite subgroups.
If cdF (G) ≤ 2 and H ≤ G is of type FP2,F ∩H , then H is hyperbolic relative to F ∩ H.
Given a group G and a family of subgroups F , a model for EF G is a G-complex X
such that stabilizers of all the isotropy groups of X belong to F , and for every Q ∈ F the
fixed point set X Q is a non-empty contractible subcomplex of X. The geometric dimension
gdF (G) is defined as the smallest dimension of a model for EF G. It is known that
cdF (G) ≤ gdF (G) ≤ max{3, cdF (G)},
see the work of Lück and Meintrup [21, Theorem 0.1].
Examples of Brady, Leary and Nucinkis [6], and Leary and Fluch [12] show that the
equality of cdF (G) and gdF (G) fails for some non-trivial families. For the case that F is the
trivial family, whether this equality holds is the outstanding Eilenberg-Ganea conjecture.
The use of cohomological dimension instead of geometric dimension in Theorem 1.3 is
justified by the unknown Eilenberg-Ganea phenomena for general families.
In practice, we expect most applications of Theorem 1.3 to be in the geometric case
gdF (G) ≤ 2 which lies outside the abstract framework of Bredon modules.
1.1. A version of Theorem 1.3 outside the framework of Bredon modules. By a toral
relatively hyperbolic group we mean a finitely generated torsion-free group G which is
hyperbolic relative to a finite collection P of abelian subgroups. By [27, Theorem 1.1], all
subgroups in P are finitely generated.
Theorem 1.4. Let G be toral relatively hyperbolic group, let F be the family of parabolic
subgroups, and suppose gdF (G) ≤ 2. If H ≤ G is finitely presented with finitely many
conjugacy classes of maximal parabolic subgroups, then H is toral relatively hyperbolic.
Theorem 1.4 follows from cdF (G) ≤ gdF (G) and Proposition 1.5 below. Recall that
a collection of subgroups P of a group G is malnormal (almost malnormal) if for any
P, P′ ∈ P and g ∈ G, either g−1 Pg ∩ P′ is trivial (respectively, finite), or P = P′ and g ∈ P.
Proposition 1.5 (Proposition 3.10). If G is finitely presented and F is a family generated
by a malnormal finite collection of finitely generated subgroups, then G is FP2,F .
Proof of Theorem 1.4. Since G is torsion-free, any two maximal parabolic subgroups of
G have either trivial intersection or they are equal [25, Lemma 2.2]. Hence the same
statement holds for the maximal parabolic subgroups of H. Since all parabolic subgroups
of H are finitely generated, Proposition 1.5 implies that H is FP2,F ∩H . Then Theorem 1.3
implies that H is toral relatively hyperbolic.
We raise the following question on the hypothesis of Theorem 1.4, remark on an simple
example that might shed some light on the situation, and quote a related result.
Question 1.6. Let G be a toral relatively hyperbolic group, let F be the family of parabolic
subgroups, and suppose cdF (G) ≤ 2. Suppose that H is finitely presented. Is H toral
relatively hyperbolic?
3
Remark 1.7. There is a toral relatively hyperbolic G with cdF (G) ≤ 2, where F is the
family of parabolic subgroups, and such that G contains a free group of finite rank with
infinitely many conjugacy classes of maximal parabolic subgroups. In particular, H is
not hyperbolic relative to F ∩ H, the group H is not FP2,F ∩H as a consequence of [28,
Proposition 3.6.1], but H is hyperbolic and hence toral relatively hyperbolic.
Indeed, let H be a free group of rank 2, and let G be hyperbolic free by cyclic group
H ⋊ϕ Z; for the existence of an automorphism ϕ ∈ Aut(H) such that the resulting semidirect
product is hyperbolic we refer the reader to [2]. Let g be a non-trivial element of H that is
not a proper power, and let F be the family generated by P = hgi. Then G is hyperbolic
relative to F by [4, Theorem 7.11]. Since H is normal in G, the subgroup H has non-trivial
intersection with any conjugate of P. Hence H has infinitely many conjugacy classes of
maximal parabolic subgroups, and in particular is not hyperbolic relative to F ∩ H. It
is left to remark that gdF (G) ≤ 2. Since H has rank two, G is the fundamental group
of a finite 2-dimensional locally CAT(0) complex X, see [7]. Let γ be a closed geodesic
in X representing the conjugacy class of g, and let Y be the space obtained by taking the
e of X and coning-off each geodesic of X
e covering γ. Observe that G acts
universal cover X
on Y and that Y is a 2-dimensional model for EF G.
A CAT(-1) version of Theorem 1.4 was known by Hanlon and the author [18, Theorem
5.15]. This result addresses Question 1.6 in a particular geometric setting. We quote this
result in the framework of toral relatively hyperbolic groups, and remark that the statement
in the cited reference is slightly more general.
Theorem 1.8. [18, Theorem 5.15] Let G be a toral relatively hyperbolic group, let F
be the family of parabolic subgroups, and suppose there is a cocompact 2-dimensional
CAT(-1) model of EF G with trivial edge stabilizers. If H ≤ G is finitely presented, then H
is toral relatively hyperbolic.
1.2. An application to small cancellation products. The studies on small cancellation
groups by Greendlinger imply that a torsion-free C ′ (1/6) group has cohomological dimension 2, see [11]. An analogous statement holds for all small cancellation products as
defined in [22, Chapter V].
Theorem 1.9 (Theorem 6.3). Let G be the quotient of a free product ∗m
i=1 Pi /hhRii where R
is a finite symmetrized set. If R is satisfies the C ′ (1/6)-small cancellation hypothesis and
G is torsion-free, then gdF (G) ≤ 2 where F is the family generated by {Pi : 1 ≤ i ≤ m}.
A group G as in the proposition is hyperbolic relative to {Pi }m
i=1 , see [27, Page 4, (II)].
Theorems 1.3 and 1.9 yield Corollary 1.10 below. The statement of the corollary for finitely
presented C ′ (1/6)-small cancellation groups was known [15, Theorem 7.6].
m
Corollary 1.10. Let G be a torsion-free quotient ∗m
i=1 Pi /hhRii of a free product ∗i=1 Pi where
R is a finite set satisfying the C ′ (1/6)-small cancellation condition. Let F be the family
generated by {Pi }m
i=1 . If H ≤ G is FP2,F ∩H , then H is hyperbolic relative to F ∩ H.
1.3. A characterization of relative hyperbolicity. The proof of Theorem 1.3, relies on a
characterization of relatively hyperbolic groups stated below.
The homological Dehn function of a space is a generalized isoperimetric function describing the minimal volume required to fill cellular 1–cycles with cellular 2–chains, see
Section 4 for definitions. The homological Dehn function FVG : N → N of a finitely presented group G is the homological Dehn function of the universal cover of a K(G, 1) with
finite 2-skeleton; the growth rate of this function is an invariant of the group [1, 13, 15].
4
E. MARTÍNEZ-PEDROZA
Recall that a pair of functions f and g from N to N ∪ {∞} have equivalent growth rate
if f g and g f , where f g means that there exists 0 < C < ∞ such that
f (n) ≤ Cg(Cn + C) + Cn + C for all n ∈ N.
Gersten observed that one can define FVG only assuming that G is an FP2 group, and
discovered the following characterization of word-hyperbolic groups.
Theorem 1.11 (Gersten). [15, Theorem 5.2] A group G is hyperbolic if and only if G is
FP2 and FVG (n) has linear growth.
Given a group G and a family of subgroups F , under the assumption that G is FP2,F ,
we define the homological Dehn function of G relative to F which we denote by FVG,F .
The growth rate of FVG,F is an invariant of the pair (G, F ), see Theorem 4.4. In the case
that F consists of only the trivial subgroup and G is finitely presented (or more generally
FP2 ), the definition of FVG,F coincides with the homological Dehn function FVG .
Theorem 1.12. Let G be a group and let P be a finite collection of subgroups generating
a family F .
(1) Suppose that P is an almost malnormal collection. If G is FP2,F and FVG,F (n)
has linear growth, then G is hyperbolic relative to P.
(2) Suppose that F contains all finite subgroups of G. If G is hyperbolic relative to P
then G is F2,F and FVG,F (n) has linear growth.
This result could be interpreted as an algebraic version of [17, Prop. 2.50] or [23,
Theorem 1.8]. The proof of Theorem 1.12(1) is given in Section 6. Theorem 1.12(2) is a
result of Przytycki and the author [24, Corollary 1.5].
Remark 1.13. The assumption that P is an almost malnormal collection in Theorem 1.12(1)
can not be removed from the statement. Let G = Z ⊕ Z, let P = Z ⊕ 0, and let F be the
family of subgroups of P. The real line is a cocompact model for EF G when considering
the action on which Z ⊕ 0 acts trivially and 0 ⊕ Z acts by translations. In this case, G is
FP2,F , the function FVG,F is trivial, and G is not hyperbolic relative to P.
1.4. The proof of Theorem 1.3. The main step in the proof of Theorem 1.3 is the following subgroup theorem for homological Dehn functions. The proof of Theorem 1.14 will be
given in Section 5.2.
Theorem 1.14. Let G be a group and F a family of subgroups. Suppose that G admits a
cocompact EF G and cdF (G) ≤ 2. If H ≤ G is of type FP2,F ∩H , then FVH,F ∩H FVG,F .
Remark 1.15. In the case that F is the trivial family, a version of this result is implicit in the
work of Gersten [15]. Another version in the case that F is the trivial family is [19, Theorem 1.1] which assumes that there is a 2-dimensional compact K(G, 1) and H is finitely
presented. Recall that finite presentability is not equivalent to FP2 , see [3, Example 6(3)].
Theorem 1.16. [24, Theorem 1.1] Let G be a hyperbolic group relative to a family F
containing all finite subgroups. Then there is a cocompact model for EF G.
Proof of Theorem 1.3. By Theorems 1.12(2) and 1.16, the function FVG,F (n) has linear
growth, and that G admits a cocompact EF G, respectively. Theorem 1.14 implies that
FVH,F ∩H FVG,F , and hence FVH,F ∩H has linear growth. Since H is FP0,F ∩H , the family
F ∩ H is generated by a finite collection Q of subgroups [28, Prop. 3.6.1]. Assume that the
collection Q has minimal size; observe that no pair of subgroups of Q are conjugate in H.
Since the collection of maximal parabolic subgroups of G is almost malnormal [25, Lemma
2.2], then Q is an almost malnormal collection of subgroups in H. By Theorem 1.12(1), H
is hyperbolic relative to Q.
5
1.5. Outline. The rest of the article is organized as follows. Section 2 contains preliminary material including a brief review of Bredon modules. Section 3 contains a combinatorial proof that FP1,F is equivalent to F1,F , an expected result for which we could not
find a proof in the literature. Section 4 introduces the notion of homological Dehn function
FVG,F for a pair (G, F ) and discuss its algebraic and topological approaches. Section 5
contains the proof of Theorem 1.14. Section 6 discuss relative hyperbolicity and contains
the proofs of Theorem 1.9 and Theorem 1.12.
Acknowledgments. The author thanks Gaelan Hanlon, Tomaz Prytuła and Mario Velasquez for comments. We would also like to thank Dieter Degrijse for pointing out an
error in an earlier version of the article, Ian Leary for consultation at the Conference on
Finiteness Conditions in Topology and Algebra 2015 and Denis Osin for a conversation
suggesting the example in Remark 1.7 at the Geometric and Asymptotic Group Theory
with Applications Conference 2016. We specially thank the referees for suggestions, feedback, and pointing out necessary corrections. The author acknowledges funding by the
Natural Sciences and Engineering Research Council of Canada, NSERC.
2. Preliminaries
2.1. G-spaces and G-maps. In this article, all group actions are from the left, all spaces
are combinatorial complexes and all maps between complexes are combinatorial. All group
actions on complexes are by combinatorial maps. Moreover, all group actions are assumed
to have no inversions; this means that if a cell is fixed setwise, then it is fixed pointwise. A
G-complex X is cocompact if there are finitely many orbits of cells. For a G-complex X,
the G-stabilizer of a cell σ is denoted by Gσ , and the fixed point set of H ≤ G is denoted
by X H . Since there are no inversions, the fixed point set X H is a subcomplex of X. A path
I → X in a complex X is a combinatorial map from an oriented subdivided interval; in
particular a path has an initial vertex and a terminal vertex.
Definition 2.1 (Combinatorial Complexes and Maps). [9, Chapter I Appendix] A map
X → Y between CW-complexes is combinatorial if its restriction to each open cell of X
is a homeomorphism onto an open cell of Y. A CW-complex X is combinatorial provided
that the attaching map of each open cell is combinatorial for a suitable subdivision.
2.2. Modules over the Orbit Category. Let F be a family of subgroups of G. The orbit
category OF (G) is the small category whose objects are the G-spaces of left cosets G/H
with H ∈ F and morphisms are the G-maps among them. The set of morphisms from G/K
to G/L is denoted by [G/K, G/L].
The category of right OF (G)-modules is the functor category of contravariant functors
from OF (G) into the category of abelian groups AB. More concretely, a right module over
OF (G) is a contravariant functor OF (G) → AB, and a morphism M → N between right
modules over OF (G) is a natural transformation from M to N. For a OF (G)-module M
and H ∈ F , we denote by M[G/H] the image of the object G/H by the functor M. The
abelian group M[G/H] is called the component of M at G/H. For the rest of this article by
an OF (G)-module we shall mean a right OF (G)-module.
Since the category of OF (G)-modules is a functor category, several constructions of the
category of abelian groups AB can be carried out in OF (G)–modules by performing them
componentwise. As a consequence it is a complete, cocomplete, abelian category with
enough projectives. In particular, computing kernels, images, and intersections reduces to
6
E. MARTÍNEZ-PEDROZA
componentwise computations in AB. For instance, a sequence K → L → M of OF (G)–
modules is exact if and only if for every H ∈ F the sequence K[G/H] → L[G/H] →
M[G/H] of abelian groups is exact.
2.3. Free OF (G)-modules. Let S be a (left) G-set. Consider the contravariant functor
S
OF (G) → SET that maps the object G/H to the fixedpoint set S H and maps a morϕ
phism G/H → G/K given by ϕ(H) = gK to the function X K → X H given by x 7→ g.x.
s
The OF (G)-module induced by S , denoted by Z[·, S ], is the composed functor OF (G) →
SET → AB where the second functor is the free abelian group construction X 7→ Z[X].
For a subgroup K ≤ G, the module induced by the G-set G/K is denoted by Z[·, G/K].
The module Z[·, G/G] is denoted by ZG,F . Note that each component Z[G/K, G/G] is
an infinite cyclic group with canonical generator the only element of G/G.
An OF (G)-module F is free if F is isomorphic to Z[·, S ] for S a G-set with isotropy
groups in F . If, in addition, the G-set S has finitely many orbits then Z[·, S ] is a finitely
generated free OF (G)-module.
2.4. Augmentation maps. Let S be a G-set and consider the OF (G)-module Z[·, S ]. For
K ∈ F , the component Z[G/K, S ] is by definition the free abelian group Z[S K ] with S K as
a free basis. The augmentation map of Z[·, S ] is the morphism ǫ : Z[·, S ] → ZG,F whose
components ǫG/K : Z[S K ] → Z are defined by mapping each element of S K to the canonical
generator of Z[G/K, G/G]. The main observation is that ǫ is surjective if and only if the
fixed point set S K is not empty for every K ∈ F .
2.5. Schanuel’s Lemma. The following proposition is a well known result of homological algebra that is used several times through the article.
Proposition 2.2 (Schanuel’s Lemma). [10, Ch.VIII, 4.2] Let
0 → Pn → Pn−1 → · · · → P0 → M → 0
and
0 → P′n → P′n−1 → · · · → P′0 → M → 0
be exact sequences of OF (G)–modules such that Pi and P′i are projective for i ≤ n − 1.
• If n is even then
P0 ⊕ P′1 ⊕ P2 ⊕ P′3 ⊕ · · · ⊕ Pn ≃ P′0 ⊕ P1 ⊕ P′2 ⊕ P3 ⊕ · · · ⊕ P′n
• If n is odd then
P0 ⊕ P′1 ⊕ P2 ⊕ P′3 ⊕ · · · ⊕ P′n ≃ P′0 ⊕ P1 ⊕ P′2 ⊕ P3 ⊕ · · · ⊕ Pn .
2.6. The cellular OF (G)-chain complex of a G–CW-complex. Let G be a group, let
F be a family of subgroups, and let X be a G-complex. These data induce a canonical
contravariant functor
OF (G) −→ TOP
that maps the object G/H to the fixed point set X H = {x ∈ X : ∀h ∈ H h.x = x}, and maps a
ϕ
G-map G/H → G/K given by ϕ(H) = gK to the cellular map X K → X H given by x 7→ g.x.
Observe that X H is a subcomplex of X since we are assuming that our actions are cellular
and without inversions. The composition of the contravariant functor OF (G) −→ TOP with
any covariant functor TOP → AB (as a homological functor) is a right OF (G)-module.
7
The augmented cellular OF (G)–chain complex of X, denoted by C∗ (X, OF (G)), is the
contravariant functor
OF (G) −→ TOP −→ Chain–Complexes
G/H
7→
XH
C∗ (X H , Z),
7→
where the second functor is the standard cellular augmented Z–chain complex functor.
Remark 2.3. If X is a cocompact G–cell-complex such that X K is n-acyclic for every K ∈ F
and all the isotropy groups of X belong to F , then the induced chain complex of OF (G)modules
(2.1)
∂n
∂2
∂1
ǫ
Cn (X, OF (G)) −→ . . . −→ C1 (X, OF (G)) −→ C0 (X, OF (G)) −→ ZG,F → 0,
is an exact sequence of finitely generated free OF (G)-modules.
3. Finiteness properties
Let G be a group and let F be a family of subgroups.
Definition 3.1. The group G is of type Fn,F if there is a cocompact G–complex X such
that all the isotropy groups of X belong to F , and for every K ∈ F the fixed point set
X K is a non-empty and (n − 1)-connected subcomplex of X. A complex X satisfying these
conditions is called an an Fn,F -complex for G. The group G is of type FF if there is a
cocompact model for EF (G).
Definition 3.2. The group G is of type FPn,F if there is a resolution of OF (G)–modules
∂n
∂2
∂1
· · · → Pn −→ . . . −→ P1 −→ P0 −→ Z → 0,
such that for 0 ≤ i ≤ n the module Pi is finitely generated and projective. Such a resolution
is called an FPn,F –resolution. It is equivalent to require that for 0 ≤ i ≤ n, the module Pi
is finitely generated and free, see for example [10, Ch.VIII 4.3].
The main result of the section is the following proposition whose proof, as well as the
proof of its corollary, are postponed to Subsection 3.1.
Proposition 3.3. If G is FP1,F then G is F1,F .
Definition 3.4. The family generated by a collection of subgroups P of G consists of all
subgroups K ≤ G such that that g−1 Kg ≤ P for some g ∈ G and P ∈ P.
Corollary 3.5. If G is FP1,F and F is a family generated by a finite and almost malnormal
collection P, then there is an F1,F -graph with finite edge G-stabilizers.
The second part of the section will recall the definition of the coned-off Cayley complex
of a relative presentation, and observe that these complexes provide F2,F -complexes in
some cases.
3.1. Proof of Proposition 3.3. Take an FP1,F –resolution consisting of only free OF (G)modules, see for example [10, Ch.VIII 4.3],
p
ǫ
Z[·, T ] −→ Z[·, V] −→ ZG,F → 0
ǫ
where Z[·, V] → ZG,F is the augmentation map. Then T and V are G-sets with isotropy
groups in F and finitely many G-orbits. For each K ∈ F the sequence of abelian groups
arising by taking the G/K-component
p
ǫ
Z[T K ] −→ Z[V K ] −→ Z → 0
8
E. MARTÍNEZ-PEDROZA
is exact.
Step 1. For each t ∈ T there is a finite directed Gt -graph Γt = (Vt , Et ) such that Vt is a
subset of V and
(1) Γt is connected when considered as an undirected graph,
(2) the Gt -action on Vt is the restriction of the G-action on V,
(3) the Gt -stabilizer of an edge e ∈ Et is the intersection of the Gt -stabilizers of its
endpoints, and
P
(4) e∈Et (e+ − e− ) = p(t).
Proof. Since T has isotropy groups in F , we have that Gt ∈ F and hence the sequence
p
ǫ
Z[T Gt ] −→ Z[V Gt ] −→ Z → 0
is exact. Therefore the fixed point set V Gt is non-empty. Choose ⋆t ∈ V Gt . Consider the
expression
X
p(t) =
nt,v v,
v∈V
where each nt,v is an integer. Define Wt = {v ∈ V : nt,v , 0}. Define
Vt = Wt ∪ {⋆t },
and observe that Vt is a finite Gt -invariant subset of V. The edge set Et is defined as the
following subset of Vt × N,
Et = v × j : v ∈ Vt and 1 ≤ j ≤ |nt,v| .
For e = v × j define e+ = v and e− = ⋆t if nt,v > 0, and e+ = ⋆t and e− = v if nt,v < 0. Since
each edge of Et has ⋆t as one of its endpoints, Γt = (Vt , Et ) is connected (when considered
as a 1-complex). The Gt -action on Vt induces an action on Et by
g.(v × j) = g.v × j
respecting the adjacency relation. Since ǫ ◦ p(t) = 0 and ǫ is the augmentation map, we
P
have that v∈Vt nt,v = 0. Therefore,
X
X
nt,v (v − ⋆t ) = p(t).
(e+ − e− ) =
e∈Et
v∈Vt
Step 2. Let K be a subgroup of G, and let E be a K-set. Then there is G-set Ê and an
injective and K-equivariant map ı : E ֒→ Ê with the following properties:
(1) The map ı induces a bijection of orbit spaces E/K → Ê/G.
(2) For each e ∈ E, the K-stabilizer Ke equals the G-stabilizer Gı(e) .
(3) If V is a G-set and f : E → V is K-equivariant, then there is a unique G-map
fˆ : Ê → V such that fˆ ◦ ı = f .
`
Proof. Observe that E is isomorphic to the K-space i∈I K/Ki where I is a complete set
of representaives of K-orbits of E and K/Ki denotes the K-space of left cosets of Ki in
`
K. Define Ê as i∈I G/Ki and observe that the assignment Ki 7→ Ki induces an injective
K-map E ֒→ Ê. The three statements of the lemma are observations.
Step 3. There is a directed G-graph Γ with vertex set V with the following properties.
(1) For each K ∈ F , the fixed point set ΓK is a non-empty connected subgraph of Γ.
(2) Γ has finitely many orbits of vertices and edges.
(3) All isotropy groups of Γ belong to F .
In particular, as a complex, Γ is an F1,F -space for G.
9
Proof. Let Γ be the directed G-graph whose vertex set is the G-set V, and edge G-set E is
defined as follows.
Let RT be a complete set of representatives of G-orbits of T . For t ∈ RT , let Γt = (Vt , Et )
be the directed Gt -graph of Step 1. Let Êt be the G-set provided by Step 2 for the Gt -set
Et . The Gt -maps (−)t : Et → V and (+)t : Et → V defining initial and terminal vertices
extend to G-maps (−)t : Êt → V and (+)t : Êt → V. Let E be the disjoint union of G-sets
`
E = t∈RT Êt and let (+) : E → V and (−) : E → V be the induced maps. Note that E has
finitely many G-orbits.
By construction, the group G acts on the directed graph Γ = (V, E) with finitely many
orbits of vertices and edges. Moreover, the G-action on Γ as a complex has no inversions
(if an edge is fixed setwise, the it is fixed pointwise). The stabilizer Ge of an edge e ∈ E
is the intersection of two subgroups in F , hence Ge ∈ F . It remains to show that the fixed
point sets ΓK are non-empty and connected for K ∈ F .
P
Let h : Z[T ] → Z[E] be the morphism of ZG-modules given by t 7→ e∈Et e for t ∈ RT .
Note that h is well defined since the subgroup Gt preserves Et setwise. Since
X
(e+ − e− ) = p(t)
e∈Et
by definition of Et , it follows that for each K ∈ F the following diagram commutes
(3.1)
Z[T✤ K ]
✤
✤h
Z[E K ]
p
/ Z[V K ]
ǫ
/Z
/0
Id
∂
/ Z[V K ]
where ∂ is the standard boundary map from 1-chains to 0-chains of Γ. Since the top row is
an exact sequence, it follows that H0 (ΓK , Z) Z and hence ΓK is non-empty and connected.
`
Proof of Corollary 3.5. Consider the G-set V = P∈P G/P. Observe that V/G is finite,
every isotropy group of V is in F , and for every K ∈ F the fixed point set V K is non-empty.
The first two properties imply that Z[·, V] is a finitely generated free OF (G)-module, and
ǫ
the third property implies that the augmentation map Z[·, V] → ZG,F is surjective. Since
p
ǫ
G is FP1,F , there is a partial FP1,F -resolution Z[·, T ] → Z[·, V] → ZG,F → 0. Then the
proof of Proposition 3.3 provides an F1,F -graph Γ with vertex set V. Since P is an almost
malnormal collection, the intersection of the G-stabilizers of two distinct vertices is finite,
and hence stabilizers of edges of Γ are finite.
3.2. Coned-off Cayley complexes and F2,F . For this subsection, let G be a group, let P
be a collection of subgroups of G, and let F be the family generated by P.
Definition 3.6 (Finite relative presentations). [27] The group G is finitely presented relative
to P if there is finite subset S of G, and a finite subset R of words over the alphabet S ⊔
F
P∈P P such that the natural homomorphism from the the free product F = (∗P∈P P) ∗ F(S )
into G is surjective and its kernel is normally generated by the elements of R (considered
as elements of F); here F(S ) denotes the free group on the set of letters S . In this case, the
data hS , P|Ri is called a finite relative presentation of G.
Definition 3.7 (Cayley graph). Let S be a subset of G closed under inverses. The directed
Cayley graph Γ = Γ(G, S ) is the directed graph with vertex set G and edge set define
as follows: for each g ∈ G and s ∈ S , there is an edge from g to gs labelled s. It is an
10
E. MARTÍNEZ-PEDROZA
observation that if S generates G then Γ is connected, and that if S is finite then the quotient
Γ/G is a finite graph.
The undirected Cayley graph is obtained from Γ(G, S ) by first ignoring the orientation
of the edges, and then identifying edges between the same pair vertices. Observe that any
combinatorial path in the undirected Cayley graph is completely determined by its starting
vertex and a word in the alphabet S . If the path is locally embedded, the corresponding
word is reduced. From here on, we will only consider the undirected Cayley graph which
we will denote by Γ(G, S ), and we will call it the Cayley graph of G with respect to S .
The definition of coned-off Cayley complex as well as Lemma 3.9 below are based
on [17, Definition 2.47, Lemma 2.48] respectively. In the cited article, the definition of
coned-off Cayley complex has some minor differences, and the constructions assume that
each P ∈ P is finitely generated. We do not require this last assumption for our arguments
Definition 3.8 (Coned-off Cayley Graph and Coned-off Cayley Complex). Let hS , P|Ri be
a finite relative presentation of G.
The coned-off Cayley graph Γ̂ of G relative to P and S is the G-graph obtained from the
Cayley graph of G with respect to S ∪ S −1 , by adding a new vertex v(gP) for each left coset
gP with g ∈ G and P ∈ P, and edges from v(gP) to each element of gP. The vertices of the
form v(gP) are called cone-vertices, and any other vertex a non-cone vertex. The G-action
on the cone-vertices is defined using the G-action on the corresponding left cosets, and this
action extends naturally to the edge set.
Observe that each locally embedded path in Γ̂ between elements of G has a unique label
F
corresponding to a word in the alphabet S ⊔ P P, since edges between vertices in G have a
label in S ∪S −1 , and for each oriented and embedded path of length two with middle vertex
a cone point v(gP) corresponds a unique element of P. The label of a path that consists of
a single non-cone vertex is defined as the empty word.
The coned-off Cayley complex X induced by hS , P|Ri is the 2-complex obtained by
equivariantly attaching 2-cells to the coned-off Cayley graph Γ̂ as follows. Attach a 2cell with trivial stabilizer to each loop of Γ̂ corresponding to a relator r ∈ R in a manner
equivariant under the G-action on Γ̂.
Lemma 3.9. Let hS , P|Ri be a finite relative presentation of G. The induced coned-off
Cayley complex X is simply-connected.
Proof. Let T be the coned-off Cayley graph of F = (∗P∈P P) ∗ F(S ) with respect to P and
S . Observe that T is simply-connected since it is a tree. Let Γ̂ be the coned-off Cayley
graph of G with respect to P and S . Observe that Γ̂ is recovered as the quotient T/N.
The main observation is that T → Γ̂ is a normal covering space with N as a group of deck
transformations. Indeed, the edge stabilizers of T with respect to N are trivial by definition;
the vertex stabilizers of T with respect to N are trivial since each factor P of F embeds in
G an therefore N ∩ P is trivial for every P ∈ P. It follows that the 1-skeleton of X has
fundamental group isomorphic to N, and since N is the normal closure of {r1 , . . . , rm } in F
the definition of X implies that π1 (X) is trivial.
Proposition 3.10. Suppose P is a finite and malnormal collection of subgroups of G, and
hS , P|Ri is a finite relative presentation of G. Then the induced coned-off Cayley complex
X is an F2,F -space with trivial 1-cell G-stabilizers. In particular, for each K ∈ F the fixed
point set X K is either a point or X.
11
Proof. The space X is simply-connected by 3.9, and since the relative presentation is finite,
the G-action on X is cocompact. Let K ∈ F be a non-trivial subgroup. Since the Gstabilizers of 1-cells and 2-cells of X are trivial, the fixed point set X K consists only of
0-cells. Since P is a malnormal collection and K fixes at least one cone-vertex, it follows
that X K is a single cone-vertex of X. Therefore X is an F2,F -space with trivial 1-cell
stabilizers.
4. The relative homological Dehn function
Thoughtout this section, G is a group, F is a family of subgroups, and G is FP2,F .
Definition 4.1. If F is a free ZG–module with ZG−basis B, then {gb : g ∈ G, b ∈ B} is a
basis for F as a free Z-module inducing a G-equivariant ℓ1 -norm k · k1 . We call a free
ZG−module based if it is understood to have a fixed ZG-basis and we use this basis for the
induced ℓ1 –norm k · k1 .
Definition 4.2. Let η : F → M be a surjective homomorphism of ZG-modules and suppose
that F is free, finitely generated, and based. The filling norm on M induced by η and the
based free module F is the G-equivariant norm defined by
kmkη = min { kxk1 : x ∈ F, η(x) = m} .
Definition 4.3 (Algebraic Definition of Relative Homological Filling Function). The homological filling function of G relative to F , is the function
FVG,F : N → N
defined as follows. Let P∗ be a resolution of OF (G)–modules for ZG,F of type FP2,F and
let
∂2
∂1
P2 −→ P1 −→ P0 −→ Z → 0
be the resolution of finitely generated ZG–modules obtained by evaluating P∗ at G/1.
Choose filling norms for the ZG–modules P1 and P2 , denoted by k · kP1 and k · kP2 respectively. Then
FVG,F (k) = max kγk∂2 : γ ∈ ker(∂1 ), kγkP1 ≤ k
where k · k∂2 is the filling norm on ker(∂1 ) given by
kγk∂2 = min kµkP2 : µ ∈ P2 , ∂2 (µ) = γ .
Theorem 4.4. The growth rate of FVG,F is an invariant of the pair (G, F ).
The rest of the section is divided into two parts, first the proof of Theorem 4.4, and then
a discussion on a topological approach to FVG,F .
4.1. Proof of Theorem 4.4. The proof in the case that F consists only of the trivial subgroup was proved in [19, Theorem 3.5]. The argument below is essentially a translation
of the argument in [19, Proof of Theorem 3.5] from the category of ZG-modules to the
category of OF (G)-modules up to using Lemma 4.6 instead of [19, Lemma 2.8].
Lemma 4.5. [15, Lemma 4.1][19, Lemma 2.9] Any two filling norms k · kη and k · kθ on
a finitely generated ZG–module M are equivalent in the sense that there exists a constant
C ≥ 0 such that C −1 kmkη ≤ kmkθ ≤ Ckmkη for all m ∈ M.
Lemma 4.6. Let ϕ : P → Q be a homomorphism between finitely generated ZG−modules.
Let k · kP and k · kQ denote filling norms on P and Q respectively. There exists a constant
C > 0 such that kϕ(p)kQ ≤ CkpkP for all p ∈ P.
12
E. MARTÍNEZ-PEDROZA
Proof. Consider the commutative diagram
ϕ̃
/B
A❄
❄❄
ψ
❄
❄❄
ρ
❄❄
P ϕ /Q
constructed as follows. Let A and B be finitely generated and based free ZG-modules, and
let A → P and B → Q be surjective morphisms inducing the filling norms k · kP and k · kQ .
Since A is free and B → Q is surjective, there is a lifting ϕ̃ : A → B of ϕ. Since ϕ̃ is
morphism between finitely generated, free, based ZG−modules, by [19, Lemma 2.7], there
is a constant C such that kϕ̃(a)k1 ≤ Ckak1 for every a ∈ A. Let p ∈ P and let a ∈ A such
that ρ(a) = p. It follows that
kϕ(p)kQ ≤ kϕ̃(a)k1 ≤ Ckak1 .
Since the above inequality holds for any a ∈ A with ρ(a) = p, it follows that
kϕ(p)kQ ≤ C · min {kak1 } = CkpkP .
ρ(a)=p
Proof of Theorem 4.4. Let P∗ and Q∗ be two resolutions of OF (G)–modules for ZG,F of
type FP2,F . For i = 1, 2, let k · kPi and k · kQi be filling norms of the components Pi [G/1]
Q
P
and Qi [G/1] respectively. Denote by FVG,F
and FVG,F
the homological filling functions
induced by P∗ and Q∗ respectively. By symmetry, it is enough to prove that there is C ≥ 0
such that for every k ∈ N,
Q
P
FVG,F
(k) ≤ C · FVG,F
(Ck + C) + Ck + C.
(4.0)
Define C as follows. Since the category of OF (G)-modules is an abelian category with
enough projectives, any two projective resolutions of an OF (G)–module are chain homotopy equivalent. Therefore, there are chain maps fi : Qi → Pi , gi : Pi → Qi , and a map
hi : Qi → Qi+1 such that
∂i+1 ◦ hi + hi−1 ◦ ∂i = gi ◦ fi − id.
From here on, we only consider the components at G/1 of these resolutions and morphisms, so we are in the category of ZG-modules. We have two exact sequences of finitely
generated ZG-modules
δ2
δ1
P2 → P1 → P0 → Z → 0,
∂2
∂1
Q2 → Q1 → Q0 → Z → 0
and the morphisms fi : Qi → Pi , gi : Pi → Qi , and hi : Qi → Qi+1 . Let C be the maximum
of the constants given by Lemma 4.6 for g2 , h1 , and f1 and the filling-norms provided .
Inequality (4.0) is proved exactly as in [19, Proof of Theorem 3.5] by using Lemma 4.6
instead of [19, Lemma 2.8]. We include the short argument for completeness.
Fix k. Let α ∈ ker(∂1 ) be such that kαkQ1 ≤ k. Choose β ∈ P2 such that δ2 (β) = f1 (α)
and k f1 (α)kδ2 = kβkP2 . Then
α = g1 ◦ f1 (α) − ∂2 ◦ h1 (α) + h0 ◦ ∂1 (α) = g1 ◦ δ2 (β) − ∂2 ◦ h1 (α)
= ∂2 ◦ g2 (β) − ∂2 ◦ h1 (α)
= ∂2 (g2 (β) − h1 (α))
13
It follows that
kαk∂2 ≤ kg2 (β) − h1 (α)kQ2 ≤ kg2 (β)kQ2 + kh1 (α)kQ2
≤ C · kβkP2 + C · kαkQ1
= C · k f1 (α)kδ2 + C · kαkQ1
P
≤ C · FVG,F
(k f1 (α)kP1 ) + CkαkQ1
P
≤ C · FVG,F CkαkQ1 + CkαkQ1
P
(Ck) + Ck
≤ C · FVG,F
Since α was arbitrary, inequality (4.0) holds for every k ∈ N completing the proof.
4.2. Topological approach to FVG,F .
Definition 4.7. Let X be a cell complex. The homological Dehn function of X is the function FVX : N → N ∪ {∞} defined as
FVX (k) = max { kγk∂ : γ ∈ Z1 (X, Z), kγk1 ≤ k } ,
where
kγk∂ = min { kµk1 : µ ∈ C2 (X, Z), ∂(µ) = γ } ,
and the maximum and minimum of the empty set are defined as zero and ∞ respectively.
Recall that a cell complex X is n-acyclic if its reduced homology groups with integer
coefficients H̃k (X, Z) are trivial for 0 ≤ k ≤ n.
Definition 4.8. The group G is FHn,F if there is a cocompact G–complex X such that all
isotropy groups of X belong to F and for every K ∈ F the fixed point set X K is a non-empty
(n − 1)-acyclic subcomplex of X. Such a complex is called an FHn,F -complex.
Definition 4.9 (Topological Definition of Relative Homological Dehn Function). Let X be
an FH2,F -complex for G. The homological filling function FVG,F of G relative to F , is the
(equivalence class of the) homological filling function FVX : N → N ∪ {∞}.
Proposition 4.10. Suppose that G is of type FH2,F . Then the algebraic and topological
definitions of relative homological Dehn function are equivalent.
Proof. Let X be an FHn,F -complex for G, see Definition 4.8. It follows that induced
augmented cellular chain complex of OF (G)-modules C∗ (X, OF (G)) is an FP2,F resolution
of ZG,F . By Theorem 4.4, FVX and FVG,F are equivalent.
Proposition 4.11. Suppose that G is of type FP2,F
(1) If X is an F1,F -graph then there is a 1-acyclic F1,F -space Y with 1-skeleton X and
finitely many G-orbits of 2-dimensional cells.
(2) If Y is a 1-acyclic F1,F -space then FVG,F and FVY are equivalent.
The proof of Proposition 4.11 requires the following lemma.
Lemma 4.12. If G is of type FP2,F and X is an F1,F -space then the OF (G)-module of
cellular 1-cycles Z1 (X, OF (G)) is finitely generated.
Proof. Consider the exact sequence of OF (G)-modules
0 → Z1 (X, OF (G)) → C1 (X, OF (G)) → C0 (X, OF (G)) → ZG,F → 0.
14
E. MARTÍNEZ-PEDROZA
The assumptions on X imply that each Ci (X, OF (G)) is finitely generated and free. Since
G is FP2,F , there is an exact sequence of OF (G)-modules
0 → K → P1 → P0 → ZG,F → 0
where each Pi is finitely generated and projective, and K is finitely generated. Applying
Schanuel lemma 2.2 to these sequences shows that Z1 (X, OF (G)) is finitely generated.
Proof of Proposition 4.11. We claim that there are finitely many circuits γ1 , . . . , γn of X
such that the corresponding 1-cycles form a generating set of the ZG-module Z1 (X, Z). By
Lemma 4.12 the OF (G)-module of cellular 1-cycles Z1 (X, OF (G)) is finitely generated,
hence Z1 (X, Z) is finitely generated. The claim follows by observing that each cellular
P
1-cycle of X can be expressed as a finite sum of 1-cycles i αi where each αi is induced
by a circuit [16, Lemma A2]. Let Y by the 2-dimensional G-complex with 1-skeleton X
obtained as follows. For each circuit γi , attach to X a 2-dimensional cell with boundary γi
and trivial G-stabilizer, and extend equivariantly. By construction, Y is an F1,F -space with
trivial H1 (Y, Z), and finitely many G-orbits of 2-cells.
To prove the second statement, Lemma 4.12 implies that Z1 (Y, OF (G)) is finitely generated. Hence there is an FP2,F resolution of the form
· · · → Q2 → C1 (Y, OF (G)) → C0 (Y, OF (G)) → ZG,F → 0
where Q2 is finitely generated and free. The expression defining FVG,F using this resolution and the definition of FVY differ only on the chosen filling norm on Z1 (Y, Z). Since
any two filling norms are equivalent, Lemma 4.5, we have that FVG,F and FVY are equivalent.
5. Proof of Theorem 1.14.
The section is divided into two parts. In the first part, we introduce the notion of
submodule distortion and prove a technical result relating retractions and module distortion, Proposition 5.6. In the second part, this proposition is used to prove Theorem 1.14.
Through the rest of the section G is a group and F is a family of subgroups.
5.1. Submodule distortion and retractions.
Definition 5.1. The distortion function of a finitely generated ZG-submodule P in a finitely
generated ZG-module L is defined by
DistLP (k) = max { kγkP : γ ∈ P and kγkL ≤ k } ,
where k · kP and k · kL are filling norms for P and L, and DistLP (k) is allowed to be ∞.
Remark 5.2. The equivalence of filling norms on a finitely generated ZG-module, Lemma 4.5,
implies that the growth rate of DistLP (k) is independent of the choice of filling norms.
Remark 5.3. In the framework of Definition 4.9, the homological filling function of G
P1
.
relative to F is the distortion function of Ker(∂1 ) in P1 , specifically, FVG,F = DistKer
(∂1 )
Definition 5.4. A ZG-module M is F -free if there is a G-set S with all isotropy groups in
F such that the induced ZG-module Z[S ] is isomorphic to M.
Remark 5.5. Let M be a free OF (G)-module. Then the G/1-component M[G/1] is an F free ZG-module. If M is finitely generated as an OF (G)-module, then M[G/1] is finitely
generated as a ZG-module.
15
Proposition 5.6. Let H be a subgroup of G. Consider a commutative diagram
/L
O
MO
Q
ı
/P
where
•
•
•
•
L and P are finitely generated ZG-modules and P is a submodule of L,
M and Q are finitely generated ZH-modules and Q is a submodule of M,
as ZH-modules, Q is a submodule of P, and M is a submodule of L,
all the arrows in the diagram are inclusions.
Let k · kL and k · k M be filling norms on L and M respectively. Suppose there is an integer
C0 ≥ 1 such that for every x ∈ M
kxkL ≤ C0 · kxk M .
Suppose that P is F -free and there is a commutative diagram of ZH-modules
P
⑧? ❅❅❅
⑧
❅❅ρ
ı ⑧⑧
❅❅
⑧
⑧
❅
⑧⑧
/ Q.
Q
Id
Then
M
DistLP .
DistQ
Proof of Proposition 5.6. Since P is a finitely generated F -free ZG-module, there is a Gset S with isotropy groups in F and finitely many G-orbits such that P Z[S ]. Denote by
k · kP the ℓ1 -norm on P induced by the free Z-basis S ; observe that this norm is a filling
norm on P as a ZG-module.
Since Q is a finitely generated ZH-module, there is a partition S = S 1 ∪ S 2 where each
S i is an H-equivariant subset of S , the quotient S 1 /H is finite, and the inclusion ı : Q → P
factors through the inclusion P1 ֒→ P as in the commutative diagram below
P1
⑦?
⑦
⑦
⑦⑦
⑦⑦
Q ı /P
where P1 = Z[S 1 ], and P2 = Z[S 2 ]. Equip the ZH-module P1 with the ℓ1 -norm k · kP1
induced by the free Z-basis S 1 , and observe that this norm is a filling norm on P1 as a
finitely generated ZH-module. We have that
kxkP1 = kxkP
for every x ∈ P1 . Moreover, if p = x ⊕ y is an element of P P1 ⊕ P2 , we have that
kxkP ≤ kpkP .
Observe that we can assume that ρ : P → Q restricts to the trivial morphism when
restricted to P2 . Simply redefine ρ on P2 as the zero morphism and observe that the commutative diagram of the statement of the proposition still holds.
16
E. MARTÍNEZ-PEDROZA
Consider the morphism of finitely generated ZH-modules ρ : P1 → Q. By Lemma 4.6
there is an integer C1 such that for every x ∈ P1 ,
kρ(x)kQ ≤ C1 kxkP1 .
Let p ∈ P be arbitrary and suppose p = x ⊕ y where x ∈ P1 and y ∈ P2 . Then
kρ(p)kQ = kρ(x)kQ ≤ C1 kxkP1 = C1 kxkP ≤ C1 kpkP ,
where the first equality follows from ρ being trivial on P2 .
Let C = max{C0 , C1 }. Let x ∈ Q be an arbitrary element and observe that
kxkQ = kρ ◦ ı(x)kQ ≤ C · kı(x)kP ≤ C · DistLP (kı(x)kL ) ≤ C · DistLP (C · kı(x)k M ) ,
where the last inequality follows from the hypothesis that kxkL ≤ Ckxk M for every x ∈ M.
Therefore for every positive integer k,
M
(k) ≤ C · DistLP (Ck)
DistQ
5.2. Proof of Theorem 1.14.
Definition 5.7. The Bredon cohomological dimension of G relative to F is at most n,
denoted by cdF (G) ≤ n, if there is a projective resolution of OF (G)–modules
0 −→ Pn −→ · · · −→ P1 −→ P0 −→ ZG,F → 0.
Proposition 5.8. If G is of type FF and cdF (G) ≤ 2, then there is an F1,F -complex X such
that Z1 (X, OF (G)) is finitely generated and free.
The proof of Proposition 5.8 requires the following definition and lemma.
Definition 5.9. An OF (G)-module M is stably free if there is a finitely generated free
OF (G)-module F such that M ⊕ F is a free OF (G)-module.
Lemma 5.10. Let X be a G-graph and let v be a vertex of X with G-stabilizer K ∈ F . Let
X⋆ be the G-graph obtained from X by adding a new G-orbit of edges with representative
an edge e such that Ge = K and both endpoints are equal to v. Then the OF (G)-module of
cellular 1-cycles Z1 (X⋆ ) is isomorphic to Z1 (X) ⊕ Z[·, G/K].
Proof. For each J ∈ F , the component Z1 (X⋆ )[G/J] is identified with Z-module of cellular
1-cycles of the fixed point set X⋆J . Hence there are natural isomorphisms Z1 (X⋆ )[G/J] →
Z1 (X J ) if J is not a subgroup of a conjugate of K, and Z1 (X⋆ )[G/J] → Z1 (X J ) ⊕ Z if J is a
subgroup of a conjugate of K. One verifies that these isomorphisms of Z-modules are the
components of an isomorphism of OF (G)-modules Z1 (X⋆ ) → Z1 (X) ⊕ Z[·, G/K].
Proof of Proposition 5.8. Let X be a cocompact model for EF (G), and let n be its dimension. Consider the exact sequences of OF (G)-modules
(5.1)
0 → Z1 (X) → C1 (X) → C0 (X) → ZG,F → 0,
and
(5.2)
0 → Cn (X) → · · · → C1 (X) → C0 (X) → ZG,F → 0.
The assumptions on X imply that each Ci (X) is a finitely generated and free OF (G)-module.
Since cdF (G) ≤ 2, there is an exact sequence of projective OF (G)-modules
(5.3)
0 → P2 → P1 → P0 → ZG,F → 0.
Applying Schanuel’s lemma to the exact sequences (5.1) and (5.3) implies that Z1 (X) is
projective. Then Schanuel’s lemma can be applied to the exact sequences (5.1) and (5.2),
which yields that Z1 (X) is stably free and finitely generated. Let Y be the complex obtained
17
by adding finitely many new orbits of edges to X in such a way that Lemma 5.10 implies
that Z1 (Y) is a finitely generated free OF (G)-module. Then Y is an F1,F -space for G.
Proposition 5.11. Let Γ be an F1,F -complex for G. If H is a subgroup of G of type F1,F ∩H ,
then Γ contains an H-equivariant subgraph ∆ which is an F1,F ∩H -complex.
Proof. Since H is FP1,F ∩H , Proposition 3.3 implies that there is an F1,F ∩H -complex ∆.
Let {vi : i ∈ I} and {e j : i ∈ J} be complete sets of representatives of H-orbits of vertices
and edges of ∆ respectively. For each i ∈ I, choose a vertex wi in the fixed point set ΓHvi
which is non-empty since Hvi ∈ F . The set assignment vi 7→ wi induces an H-map φ from
the vertex set of ∆ to the one of Γ. For each j ∈ J, let H j be the (pointwise) H-stabilizer of
e j and let v− and v+ be the endpoints of e j . Let γ j be an edge path in the fixed point set ΓH j
from φ(v− ) to φ(v+ ); ΓH j is connected since H j ∈ F . After subdividing ∆, the assignment
e j 7→ γ j extends φ to an H-map of graphs φ : ∆ → Γ. Replacing ∆ with the image of φ
proves the claim.
Proof of Theorem 1.14. By Proposition 5.8, there is an F1,F -complex Γ for G such that
Z1 (Γ, OF (G)) is free and finitely generated. By Proposition 5.11, Γ contains an H-equivariant
subgraph ∆ which is an F1,F ∩H -complex for H. Since H is FP2,F ∩H , we have that Z1 (∆, OF ∩H (H))
is finitely generated.
From here on, all considered modules are OF ∩H (H)-modules. Consider Γ and ∆ as
H-graphs. Then there is a short exact sequence of cellular OF ∩H (H)–chain complexes
(5.4)
0 → C∗ (∆) → C∗ (Γ) → C∗ (Γ, ∆) → 0
where Ci (Γ, ∆) is the quotient Ci (Γ)/Ci (∆). Consider the induced long exact homology
sequence. Observing that H2 (Γ, ∆) is trivial, and that the reduced homology H̃0 (∆) is trivial
(since ∆K is connected for all K ∈ F ∩ H), we obtain that the sequence
(5.5)
0 → Z1 (∆) → Z1 (Γ) → Z1 (Γ, ∆) → 0
is exact. Considering once more the long exact sequence induced by (5.4), one verifies that
(5.6)
0 → Z1 (Γ, ∆) → C1 (Γ, ∆) → C0 (Γ, ∆) → 0
is an exact sequence of OF ∩H (H)-modules. Since ∆ is an H-equivariant subgraph of Γ, we
have that Ci (∆) is a free factor of Ci (Γ); hence Ci (Γ, ∆) is free. Therefore, the sequence (5.6)
implies that Z1 (Γ, ∆) is projective. The exact sequence (5.5) together with Z1 (Γ, ∆) being
projective implies that there is a commutative diagram of OF ∩H (H)-modules
Z1 (Γ)❍
❍❍ ρ
✇;
✇
❍❍
✇
ı ✇
❍❍
✇
✇
❍❍
✇
#
✇✇
/ Z1 (∆).
Z1 (∆)
Id
The proof concludes by invoking Proposition 5.6 as follows. Considering the H/1 component of the above commutative diagram, we obtain the analogous diagram for the ZHmodules of integral cellular 1-cycles of Γ and ∆. Since Z1 (Γ, OF (G)) is finitely generated and free, it follows that P := Z1 (Γ, Z) is a finitely generated and F -free ZG-module.
Analogously, since Z1 (∆, OF ∩H (H)) is finitely generated, Q := Z1 (∆, Z) is a finitely generated ZH-module. Moreover, Q is a submodule of the finitely generated ZH-module
M := C1 (∆, Z), and analogously P is a submodule of the finitely generated ZG-module
L := C1 (Γ, Z). Observe that we can assume that M ≤ L. Considering the natural ℓ1 -norms
k·kL and k·k M on L and M induced by the collection 1-cells (with a chosen orientation) of Γ
18
E. MARTÍNEZ-PEDROZA
and ∆ respectively, we have the equality kσkL = kσk M for every σ ∈ M. Then Remark 5.3
implies FVH,F ∩H = DistQM and FVG,F = DistLP . By Proposition 5.6, FVH,F ∩H FVG,F .
6. Relatively Hyperbolic Groups
6.1. Proof of Theorem 1.12(1). Let G be a group and let P be a collection of subgroups
generating a family F . Suppose that P is a finite and almost malnormal collection, that G
is FP2,F , and that FVG,F (n) has linear growth.
By Corollary 3.5, there is an F1,F -graph X with finite edge stabilizers. By Proposition 4.11, there is an F1,F -space Y with 1-skeleton X, with finitely many G-orbits of 2dimensional cells, and such that FVG,F and FVY are equivalent. To conclude that G is
hyperbolic relative to F , it is enough to prove that X is fine and hyperbolic. Since FVG,F
has linear growth, the proof concludes by invoking the following result.
Lemma 6.1. [23, Corollary 3.3] Let Y be a 1-acyclic 2-dimensional cell complex such that
there is a bound on the length of attaching maps of 2-cells. If FVY (n) has linear growth
then the 1-skeleton of Y is a fine hyperbolic graph.
6.2. A Class of Relatively Hyperbolic Groups of Bredon Cohomological Dimension 2.
This last subsection relies on small cancellation theory over free products; our main reference is Lyndon-Schupp texbook on Combinatorial group theory [22, Chapter V, Section
9]. The following remark recalls some well known facts on small cancellation products.
Remark 6.2. Suppose that P = {Pi } is a finite collection of groups, let F = ∗Pi be the
corresponding free product, and let R be a symmetrized subset of F satisfying the C ′ (1/6)small cancellation condition. Let N be the normal closure of R in F.
(1) The collection P can be regarded as a collection of subgroups of F/N. Indeed, Greendlinger’s
lemma [22, Ch. V. Theorem 9.3] implies that N does not contain elements of length
less than three. In particular, the natural map F → F/N embeds each factor Pi of F,
and no two factors are identified.
(2) The group F/N is hyperbolic relative to P; this is a direct consequence of [22, Ch. V.
Theorem 9.3] using Osin’s approach to relative hyperbolicity [27, Page 4, (II)].
(3) If each P ∈ P is torsion-free, then P is malnormal in F/N, [27, Proposition 2.36].
Theorem 6.3. Suppose that P = {Pi } is a finite family of torsion-free groups, that R
is a finite symmetrized subset of the free product F = ∗Pi satisfying the C ′ (1/6)-small
cancellation condition and with no proper powers, and that G = F/N where N is the
normal closure of R.
If X is the coned-off Cayley complex of a relative presentation hP|R′ i where R′ is a
minimal subset of R with normal closure N in F, then X is a cocompact EF G-complex
where F is the family generated by P. In particular gdF (G) ≤ 2.
Proof. In view of Proposition 3.10 and Remark 6.2, it is left to prove that X is contractible.
We argue that every spherical diagram S → X is reducible as explained below. It follows
that X has trivial second homotopy group by a remark of Gersten [14, Remark 3.2]. Since
X is 2-dimensional and simply-connected, an application of Hurewicz and Whitehead theorems from algebraic topology imply that X is contractible.
A spherical diagram is a combinatorial cellular map S → X where S is topologically
the 2-sphere. Roughly speaking, the diagram is reducible in the sense of Gersten if there
are two distinct faces (2-cells) of the 2-sphere S that have an edge in common and map
to the same 2-cell of X by a mirror image, see [14, Sec. 3.1] for a precise definition. The
19
small cancellation condition implies that every spherical diagram S → X is reducible; we
sketch an argument below following Ol’shanskiı̆ [26, Proof of Theorem 13.3].
Let γ be a path in X whose endpoints are non-cone vertices. We observed in DefiniF
tion 3.8 that γ has as a label a reduced word in the alphabet Pi . If α is a subpath of γ
then the label of α is defined as the label of the minimal length subpath of γ such that its
endpoints are non-cone vertices and it contains α as a subpath. In particular, if α is a trivial
path then either α is a non-cone vertex and its label is the empty word, or α is a cone-vertex
v(gP) and its label is a non-trivial element of the group P ∈ P.
Now we remark two direct consequences of Greendlinger’s lemma [22, Ch. V. Theorem
9.3] together with their interpretations/consequences on the cell structure of X. We leave
some of the details to the reader.
• If r ∈ R has semi-reduced form r = ab and a is not the empty word, then b < N.
As a consequence, any closed curve γ → X in the 1-skeleton of X with label r ∈ R is
an embedded circle.
• Let r, r′ ∈ R. If r and r′ have semi-reduced forms asbt and aubv respectively, and su−1
is a reduced form such that su−1 ∈ N, then r = r′ .
As a consequence, if D and D′ are a pair of 2-cells of X whose boundaries intersect,
then either D = D′ or their boundaries intersect in a single path (possibly a trivial path)
whose induced label (as a subpath of the boundary of D) is a piece.
This last statement relies on the minimality of R′ and the assumption that R has
no proper powers, and not only on the stated consequence of Greendlinger’s lemma.
Suppose that there is a pair of distinct 2-cells D and D′ of X such that their boundary
intersection contains a path whose label is not a piece. Then from the definition of piece,
it follows that the labels r and r′ of the boundaries of D and D′ respectively are the same
reduced word up to taking the inverse and a cyclic conjugation. By minimality of R′ , the
words r and r′ correspond to the same element of R′ , say r. By definition of X, since D
and D′ are distinct and have the same boundary path labelled by r, the word r is a proper
power. But this contradicts that R has no proper powers.
In particular, the consequence of Greendlinger’s lemma is only used to show that if
the boundary intersection of D and D′ is non-empty then it is connected. This is obtained
by observing that for a suitable orientations of the boundary paths of D and D′ , if their
intersection is not connected their labels have have semi-reduced forms asbt and aubv
respectively, and such that su−1 is the label of an embedded closed path.
Suppose that S → X is a spherical diagram that is not reducible. Let Φ be the graph
whose vertices are the 2-cells of S ; and given two vertices, there is an edge between them
for each connected component of the intersection of the boundary paths of the corresponding 2-cells. It is immediate that Φ is planar. Below we argue that every vertex of Φ has
degree at least six, and that Φ is simplicial, i.e., no double edges and no loops. Since every
finite planar simplicial graph has a vertex of degree at most five, the existence of Φ yields
a contradiction. Hence a non-reducible spherical diagram S → X can not exist.
The first bullet statement above implies that Φ has no edges that connect a vertex to
itself. Since S → X is not reducible, the second bullet statement above implies that if
D and D′ are two distinct 2-cells of S and their boundaries have non-empty intersection,
then they intersect in a single path whose induced label is a piece. This implies that Φ has
no two distinct edges connecting the same vertices. Hence Φ is simplicial. Finally, the
C ′ (1/6)-condition implies that every r ∈ R′ can not be the concatenation of less than six
pieces, and we observed that the intersection of any pair of 2-cells of S is along a path with
label a piece; it follows that every vertex of Φ has degree at least six.
20
E. MARTÍNEZ-PEDROZA
References
[1] Aaron Abrams, Noel Brady, Pallavi Dani, and Robert Young. Homological and homotopical Dehn functions
are different. Proc. Natl. Acad. Sci. USA, 110(48):19206–19212, 2013.
[2] M. Bestvina and M. Feighn. A combination theorem for negatively curved groups. J. Differential Geom.,
35(1):85–101, 1992.
[3] Mladen Bestvina and Noel Brady. Morse theory and finiteness properties of groups. Invent. Math.,
129(3):445–470, 1997.
[4] B. H. Bowditch. Relatively hyperbolic groups. Internat. J. Algebra Comput., 22(3):1250016, 66, 2012.
[5] Noel Brady. Branched coverings of cubical complexes and subgroups of hyperbolic groups. J. London Math.
Soc. (2), 60(2):461–480, 1999.
[6] Noel Brady, Ian J. Leary, and Brita E. A. Nucinkis. On algebraic and geometric dimensions for groups with
torsion. J. London Math. Soc. (2), 64(2):489–500, 2001.
[7] Thomas Brady. Complexes of nonpositive curvature for extensions of F2 by Z. Topology Appl., 63(3):267–
275, 1995.
[8] Glen E. Bredon. Equivariant cohomology theories. Lecture Notes in Mathematics, No. 34. Springer-Verlag,
Berlin-New York, 1967.
[9] Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature, volume 319 of
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences].
Springer-Verlag, Berlin, 1999.
[10] Kenneth S. Brown. Cohomology of groups, volume 87 of Graduate Texts in Mathematics. Springer-Verlag,
New York, 1994. Corrected reprint of the 1982 original.
[11] Ian M. Chiswell, Donald J. Collins, and Johannes Huebschmann. Aspherical group presentations. Math. Z.,
178(1):1–36, 1981.
[12] Martin G. Fluch and Ian J. Leary. An Eilenberg-Ganea phenomenon for actions with virtually cyclic stabilisers. Groups Geom. Dyn., 8(1):135–142, 2014.
[13] S. M. Gersten. Homological Dehn functions and the word problem. Available at
http://www.math.utah.edu/˜sg/Papers/df9.pdf.
[14] S. M. Gersten. Reducible diagrams and equations over groups. In Essays in group theory, volume 8 of Math.
Sci. Res. Inst. Publ., pages 15–73. Springer, New York, 1987.
[15] S. M. Gersten. Subgroups of word hyperbolic groups in dimension 2. J. London Math. Soc. (2), 54(2):261–
283, 1996.
[16] S. M. Gersten. Cohomological lower bounds for isoperimetric functions on groups. Topology, 37(5):1031–
1072, 1998.
[17] Daniel Groves and Jason Fox Manning. Dehn filling in relatively hyperbolic groups. Israel J. Math.,
168:317–429, 2008.
[18] Richard Gaelan Hanlon and Eduardo Martı́nez-Pedroza. Lifting group actions, equivariant towers and subgroups of non-positively curved groups. Algebr. Geom. Topol., 14(5):2783–2808, 2014.
[19] Richard Gaelan Hanlon and Eduardo Martı́nez Pedroza. A subgroup theorem for homological filling functions. Groups Geom. Dyn., 10(3):867–883, 2016.
[20] Wolfgang Lück. Transformation groups and algebraic K-theory, volume 1408 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1989. Mathematica Gottingensis.
[21] Wolfgang Lück and David Meintrup. On the universal space for group actions with compact isotropy. In
Geometry and topology: Aarhus (1998), volume 258 of Contemp. Math., pages 293–305. Amer. Math.
Soc., Providence, RI, 2000.
[22] Roger C. Lyndon and Paul E. Schupp. Combinatorial group theory. Classics in Mathematics. SpringerVerlag, Berlin, 2001. Reprint of the 1977 edition.
[23] Eduardo Martı́nez-Pedroza. A note on fine graphs and homological isoperimetric inequalities. Canad. Math.
Bull., 59(1):170–181, 2016.
[24] Eduardo Martı́nez-Pedroza and Piotr Przytycki. Dismantlable classifying space for the family of parabolic
subgroups of a relatively hyperbolic group. Journal of the Institute of Mathematics of Jussieu, 2017.
[25] Eduardo Martı́nez-Pedroza and Daniel T. Wise. Relative quasiconvexity using fine hyperbolic graphs. Algebr. Geom. Topol., 11(1):477–501, 2011.
[26] A. Yu. Ol’shanskiı̆. Geometry of defining relations in groups, volume 70 of Mathematics and its Applications
(Soviet Series). Kluwer Academic Publishers Group, Dordrecht, 1991. Translated from the 1989 Russian
original by Yu. A. Bakhturin.
[27] Denis V. Osin. Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems. Mem. Amer. Math. Soc., 179(843):vi+100, 2006.
21
[28] Simon St. John-Green. Cohomological Finiteness Properties of Groups. PhD thesis, University of
Southampton, UK, 2014.
[29] Tammo tom Dieck. Transformation groups, volume 8 of de Gruyter Studies in Mathematics. Walter de
Gruyter & Co., Berlin, 1987.
Memorial University, St. John’s, Newfoundland and Labrador, Canada
E-mail address: [email protected]
| 4math.GR
|
Efficient Minimization of Higher Order Submodular Functions
using Monotonic Boolean Functions
Srikumar Ramalingam1
Chris Russell2&3
L’ubor Ladický4
Philip H.S. Torr5
1
University of Utah, USA
Alan Turing Institute, UK
3
University of Edinburgh, UK
4
ETH Zurich, Switzerland
5
University of Oxford, Oxford, UK
arXiv:1109.2304v2 [cs.DS] 23 Jan 2017
2
Abstract. Submodular function minimization is a key problem in a wide variety of applications in
machine learning, economics, game theory, computer vision, and many others. The general solver
has a complexity of O(n3 log2 n.E + n4 logO(1) n) where E is the time required to evaluate the
function and n is the number of variables [32]. On the other hand, many computer vision and machine
learning problems are defined over special subclasses of submodular functions that can be written
as the sum of many submodular cost functions defined over cliques containing few variables. In
such functions, the pseudo-Boolean (or polynomial) representation [3] of these subclasses are of
degree (or order, or clique size) k where k n. In this work, we develop efficient algorithms
for the minimization of this useful subclass of submodular functions. To do this, we define novel
mapping that transform submodular functions of order k into quadratic ones. The underlying idea
is to use auxiliary variables to model the higher order terms and the transformation is found using
a carefully constructed linear program. In particular, we model the auxiliary variables as monotonic
Boolean functions, allowing us to obtain a compact transformation using as few auxiliary variables
as possible. The transformed quadratic function can be efficiently minimized using the standard maxflow algorithm with a time complexity of O((n + m)3 ) where m is the total number of auxiliary
variables involved in transforming all the higher order terms to quadratic ones. Specifically, we show
that our approach for fourth order function requires only 2 auxiliary variables in contrast to 30 or
more variables used in existing approaches. In the general case, we give an upper bound for the
number or auxiliary variables required to transform a function of order k using Dedekind number,
k
which is substantially lower than the existing bound of 22 .
Keywords: submodular functions, quadratic pseudo-Boolean functions, monotonic Boolean functions, Dedekind number, max-flow/mincut algorithm
1
Introduction
Many optimization problems in several domains such as operations research, computer vision, machine
learning, and computational biology involve submodular function minimization. Submodular functions
(See Definition 1) are discrete analogues of convex functions [33]. Examples of such functions include
cut capacity functions, matroid rank functions and entropy functions. Submodular function minimization
techniques may be broadly classified into two categories: algorithms for general submodular functions
and efficient and customized algorithms for subclasses of submodular functions. This paper falls under
the second category.
General solvers: The role of submodular functions in optimization was first discovered by Edmonds
when he gave several important results on the related poly-matroids [10]. Grötschel, Lovász, and Schrijver
first gave a polynomial-time algorithm for minimization of submodular function using ellipsoid method
2
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
[17]. Recently several combinatorial and strongly polynomial algorithms [13,22,24,45,36] have been developed based on the work of Cunningham [9]. The current best strongly polynomial algorithm for minimizing general submodular functions [32] has a run-time complexity of O(n3 log2 n.E + n4 logO(1) n),
where E is the time taken to evaluate the function, and n is the number of variables. Weakly polynomial
time algorithms with a smaller dependence on n also exist. For example, Lee et al. [32] shows a method
with a run-time complexity of O(n2 log nM.E + n3 logO(1) nM ), where M is the maximum absolute
value of the function values.
Specialized solvers: Higher order submodular functions are useful in modeling many computer vision
and machine learning problems [26,31,21]. Such problems typically involve millions of pixels making
the use of general solvers highly infeasible. Further, each pixel may take multiple discrete values and the
conversion of such a problem to a Boolean one introduces further variables. On the other hand, the cost
functions for many such optimization algorithms belong to a small subclass of submodular functions.
The goal of this paper is to provide an efficient approach for minimizing these subclasses of submodular
functions using a max-flow algorithm.
Notations: Let B denote the Boolean set {0, 1} and R the set of reals. Let the vector x = (x1 , ..., xn ) ∈
Bn , and V = {1, 2, ..., n} be the set of indices of x. We introduce a set representation to denote the
labelings of x. Let S4 = {1, 2, 3, 4} and let P be the power set of S4 . For example, a labeling {x1 =
1, x2 = 0, x3 = 1, x4 = 1} is denoted by the set {1, 3, 4}. For a subset A ⊆ V , let us denote by 1A ∈ Bn
its characteristic vector, i.e.
(
1 if j ∈ A,
1Sj =
(1)
0 otherwise.
Definition 1. Submodular functions map f : Bn → R and satisfy the following condition:
f (X) + f (Y ) ≥ f (X ∨ Y ) + f (X ∧ Y ),
(2)
where X and Y are elements of Bn and the symbols ∨ and ∧ denote union and intersection of sets
respectively.
In this paper, we use a pseudo-Boolean polynomial representation for denoting submodular functions.
Definition 2. Pseudo-Boolean functions (PBF) take a Boolean vector as argument and return a real number, i.e. f : Bn → R [3]. These can be uniquely expressed as multi-linear polynomials, i.e. for all f there
exists a unique set of real numbers {aS : S ∈ Bn } :
X
Y
f (x1 , ..., xn ) =
aS (
xj ),
aS ∈ R,
(3)
S⊆V
j∈S
where a∅ is said to be the constant term.
The term order refers to the maximum degree of the polynomial. A submodular function of second order
involving Boolean variables can be easily represented using a graph such that the minimum cut, computed
using a max-flow algorithm, also efficiently minimizes the function. However, max-flow algorithms can
not exactly minimize non-submodular functions or some submodular ones of an order greater than 3 [49].
There is a long history of research in solving subclasses of submodular functions both exactly and efficiently using max-flow algorithms [1,28,18,48,38]. In this paper, we propose a linear programming
formulation that is capable of answering this question: given any pseudo Boolean function, it can derive
a quadratic submodular formulation of the same cost or a closest quadratic submodular function (i.e., say
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
under L1 norm), if an exact derivation does not exist. The problem of using a linear program (LP) for
expressing a given function using other functions (with AVs) was already established in [6]. Compared
to the existing results, we also provide a smaller LP for submodular functions and also show that we need
only fewer AVs compared to existing methods.
Definition 3. F k denotes a class of pseudo-Boolean functions of order k such that every function f (x) ∈
F k satisfies the submodularity property given in Definition 2.
It was first shown in [18] that any function in F 2 can be minimized exactly using a max-flow algorithm. Billionnet and Minoux [1] showed that any function in F 3 can be transformed into a function
in F 2 using additional variables. While transforming a given higher order function to a function in F 2 ,
we use additional variables that we refer to as auxiliary variables (AV). In the course of this paper, you
will see that these AVs are often more difficult to handle than variables in the original function and our
algorithms are driven by the quest to understand the role of these auxiliary variables and to eliminate the
unnecessary ones.
Kolmogorov [27] improved the complexity of Iwata’s capacity scaling algorithm [23] for special functions which are represented as a sum of submodular terms. This is the first line of research that does not
use auxiliary variables to handle higher order terms. The formulation of Kolmogorov also closely resembles the approach of Cooper [7], who used a linear program with an exponential number of constraints
for solving the minimization of the submodular function. It was shown that we can have a algorithm that
can be parallelized for minimizing decomposable submodular functions, which can be decomposed into
sum of simple submodular functions. In [35], it was shown that the algorithm converges linearly, and
they also provide upper and lower bounds on the rate of convergence.
Recently, Zivny et al. [49] made substantial progress in characterizing the class of functions that can
be transformed to F 2 . Their most notable result is to show that not all functions in F 4 can be transformed
to a function in F 2 . This result stands in strong contrast to the third order case that was positively resolved more than two decades earlier [1]. Using Theorem 5.2 from [37] it is possible to decompose a
given submodular function in F 4 into 10 different groups Gi , i = {1..10}, where each Gi is shown in
Table 1. Zivny et al. showed that one of these groups (G10 ) can not be expressed using any function in F 2
employing any number of AVs. Most of these results were obtained by mapping the submodular function
minimization to a valued constraint satisfaction problem.
1.1
Problem Statement and main contributions
Largest subclass of submodular functions: We are interested in transforming a given function in F k into
a function in F 2 using AVs. As such a transformation is not possible for all submodular functions of order
four or more [49], our goal is to implicitly map the largest subclass F2k that can be transformed into F 2 .
This distinction between the two classes F2k and F k will be crucial in the remainder of the paper (see
Figure 1).
Definition 4. The class F2k is the largest subclass of F k such that every function f (x) ∈ F2k has an
equivalent quadratic function h(x, z) ∈ F 2 using AVs z = z1 , z2 , ..., zm ∈ Bm satisfying the following
condition:
f (x) = minm h(x, z), ∀x.
z∈B
(4)
In this paper, we are interested in developing an algorithm to transform every function in this class F2k to
a function in F 2 .
3
4
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
Fig. 1. All the function in the classes F 1 , F 2 , F 3 and F2k , k ≥ 2 can be transformed to functions in F 2 and minimized
using the maxflow/mincut algorithm.
Efficient transformation of higher order functions: We propose a linear programming algorithm to transform higher order submodular functions to quadratic ones using monotonic Boolean functions (MBF [8]).
This framework provides several advantages. First we show that the state of an AV in a minimum cost
labeling is equivalent to an MBF defined over the original variables. This provides an upper bound on the
number of AVs given by the Dedekind number [29], which is defined as the total number of MBFs over
a set of n binary variables. In the case of fourth order functions, there are 168 such functions. Using the
properties of MBFs and the nature of these AVs in our transformation, we prove that these 168 AVs can be
replaced by two AVs.
Minimal use of AVs: One of our goals is to use a minimum number (m) of AVs in performing the transformation of (4). Although, given a fixed choice of F2k , reducing the value of m does not change the
complexity of the resulting min/cut algorithm asymptotically, it is crucial in several machine learning and
computer vision problems. In general, most image based labeling problems involve millions of pixels and
in typical problems, the number of fourth order priors is linearly proportional to the number of pixels.
Such problems may be infeasible for large values of m. It was shown that the transformation of functions
in F24 can be achieved using about 30 auxiliary variables [50]. On the other hand, we show that we can
transform the same class of functions using only 2 additional nodes. Note that this reduction is applicable
to every fourth order term in the function. A typical vision problem may involve functions having 10000
F24 terms for an image of size 100 × 100. Under these parameters, our algorithm will use 20000 AVs,
whereas the existing approach [50] would use as many as 300000 AVs. In several practical problems, this
improvement will make a significant difference in the running time of the algorithm.
k
For a function in F2k , the maximum number of AVs required is given by 22 [6]. We show that one can
transform the function using substantially fewer number of AVs given by Dedekind number. In section 3.1,
k
we show that the Dedekind number is substantially lower than 22 . In [6], an LP based approach was used
k
to obtain the bound of 22 . We also use an LP-based approach, however the use of monotonic Boolean
functions enables us to improve this bound to Dedekind number. The idea of reducing the number of AVs
in an LP formulation has been done in other contexts [46]. In [46], a combinatorial structure commonly
referred to as gadgets were computed using linear programming. This enables the transformation of
constraints from one optimization problem to another. In this work, we show that we can transform a
function with several AVs to a function involving much fewer AVs using a linear programming approach.
1.2
Limitations of Current Approaches and Open Problems
Decomposition of submodular functions: Many existing algorithms for transforming higher order functions target the minimization of a single k-variable k th order function. However, the transformation framework is incomplete without showing that a given n-variable submodular function of k th order can be decomposed into several individual k-variable k th order sub-functions. Billionnet proved that it is possible
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
to decompose a function in F 3 involving several variables into 3-variable functions in F 3 [1]. To the
best of our knowledge, the decomposition of fourth or higher order functions is still an open problem
and it will remain a hard problem due to the following reasoning. In [16], it was proven that testing a
membership of a function f with n variables in F 4 is NP-complete. It is easy to test the submodularity of
a fourth order function with 4 variables. Thus if a function f with n variables is decomposed into several
4-variable fourth order functions and if each of these individual 4-variable functions are submodular, then
the function f is submodular. This seems to be most possible case when we know that a function is submodular. Thus it is very unlikely to know that a function is submodular and not know its decomposition.
Given this, it is likely that specialized solvers based on max-flow algorithms may never solve the general
class of submodular functions. However, this decomposition problem is not a critical issue in machine
learning and vision problems. This is because the higher order priors from natural statistics already occur
in different sub-functions of k nodes - in other words, the decomposition is known a priori. This paper
only focuses on the transformation of a single k-variable function in F k . As mentioned above, the solution to this problem is still sufficient to solve large functions with hundreds of nodes and higher order
priors in applications.
Non-Boolean problems: The results in this paper are applicable only to set or pseudo-Boolean functions.
Many real world problems involve variables that can take multiple discrete values. Ishikawa showed
that it is possible to transform a multi-label second order function to a Boolean second order function
using Boolean variables to encode multi-label variables [20]. To denote a single multi-label variable with
l labels, l Boolean variables were used. Ishikawa’s method considered functions with convex priors, a
class of functions that is slightly more restricted than general submodular functions. Schlesinger and
Flach later showed that it is possible to transform general submodular multi-label functions of second
order to Boolean second order functions [44]. This approach used l − 1 Boolean variables to encode an
l-label multi-label variable. Ramalingam et al. [39] generalized this work for transforming multi-label
higher order functions to Boolean second-order functions. In [39], the transformation does not preserve
submodularity for fourth or higher order functions [39]. Zivny et al. [49] proved that it is not possible to
have a submodularity preserving transformation for fourth or higher order functions.
Excess AVs: The complexity of an efficient max-flow algorithm is O((n + m)3 ) where n is the number of
variables in the original higher order function and m is the number of AVs. Typically in imaging problems,
the number of higher order terms is of O(n) and the order k is less than 10. Thus the minimization of the
function corresponding to an entire image with O(n) higher order terms will still have a complexity of
O((n + n)3 ). However when m becomes at least quadratic in n, for example, if a higher-order term is
defined over every triplet of variables in V , the complexity of the max-flow algorithm will exceed that of
a general solver being O((n + n3 )3 ). Thus in applications involving a very large number of higher order
terms, a general solver may be more appropriate.
2
Preliminaries
Definition 5. The (discrete) derivative of a function f (x1 , . . . , xn ) with respect to xi is given by:
δf
(x1 , . . . , xn ) = f (x1 , . . . , xi−1 , 1, xi+1 , . . . , xn ) − f (x1 , . . . , xi−1 , 0, xi+1 , . . . , xn ).
δxi
(5)
5
6
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
Definition 6. The second discrete derivative of a function ∆i,j (x) is given by
δ δf
∆i,j (x) =
(x1 , . . . , xn )
δxj δxi
(6)
= f (x1 ,...,xi−1 ,1,xi+1 ...,xj−1 ,1,xj+1 ...,xn )−f (x1 ,...,xi−1 ,0,xi+1 ...,xj−1 ,1,xj+1 ...,xn )
− f (x1 ,...,xi−1 ,1,xi+1 ...,xj−1 ,0,xj+1 ...,xn )−f (x1 ,...,xi−1 ,0,xi+1 ...,xj−1 ,0,xj+1 ...,xn ) .
Note that it follows from the definition of submodular functions (2), that their second derivative is always
non-positive for all x.
3
Transforming functions in F2n to F 2
Consider the following submodular function f (x) ∈ F2n represented as a multi-linear polynomial:
X
Y
f (x) =
aS (
xj ), aS ∈ R.
S∈Bn
(7)
j∈S
Let us consider a function h(x, z) ∈ F 2 where z is a set of AVs used to model functions in F2n . Any
general function in F 2 can be represented as a multi-linear polynomial (consisting of linear and bi-linear
terms involving all variables):
X
X
X
X
X
h(x, z) =
a i xi −
ai,j xi xj +
al zl −
al,m zl zm −
ai,l xi zl .
(8)
i
i,j:i>j
l
l,m:l>m
i,l
The negative signs in front of the bi-linear terms (xi xj , zl xi , zl zm ) emphasize that their coefficients
(−aij , −ail , −alm ) must be non-positive if the function is submodular. We are seeking a function h such
that:
f (x) = minn h(x, z), ∀x.
(9)
z∈B
Here the function f (x) is known. We are interested in computing the coefficients (a), and in determining
the number of auxiliary variables required to express a function as a pairwise submodular function. The
problem is challenging due to the inherent instability and dependencies within the problem – different
choices of parameters cause auxiliary variables to take different states. To explore the space of possible
solutions fully, we must characterize what states an AV takes.
3.1
Auxiliary Variables as Monotonic Boolean Functions
Definition 7. A monotonic (increasing) Boolean function (MBF) m : Bn → B takes a Boolean vector as
argument and returns a Boolean, s.t if yi ≤ xi , ∀i =⇒ m(y) ≤ m(x).
Lemma 1. Let zs (x) be a function that takes an argument x and returns a Boolean as shown below:
0
zs (x) = arg min min
h(x,
z
,
z
)
,
(10)
s
0
zs
z
where h(x, z0 , zs ) is a submodular function defined in Equation (8) and satisfying Equation (9). The
function zs (x) that maps a Boolean vector x to the Boolean state of zs is an MBF (See Definition 7),
where z0 is the set of all auxiliary variables except zs .
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
Proof. We consider a current labeling x with an induced labeling of zs = zs (x). We first note
h0 (x, zs ) = min
h(x, z0 , zs )
0
z
(11)
is a submodular function i.e. it satisfies (2). We now consider increasing the value of x, that is given a
current labeling x we consider a new labeling x(i) such that
(
1
if j = i
(i)
(12)
xj =
xj otherwise.
We wish to prove
zs (x(i) ) ≥ zs (x) ∀x, i.
(13)
Note that if zs (x) = 0 or xi = 1 this result is trivial. This leaves the case: zs (x) = 1 and xi = 0. It
follows from (6) that:
h0 (x1 , . . . , xi−1 , 1, xi+1 , . . . , 0) − h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 0) ≥
h0 (x1 , . . . , xi−1 , 1, xi+1 , . . . , 1) − h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 1).
(14)
Using Equation (10), we derive the following from our hypothesis zs (x) = 1 and xi = 0:
h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 0) ≥ h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 1).
(15)
Hence by replacing h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 0) with h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 1) in Equation (14),
we have
h0 (x1 , . . . , xi−1 , 1, xi+1 , . . . , 0) − h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 0) ≥
(16)
h0 (x1 , . . . , xi−1 , 1, xi+1 , . . . , 1) − h0 (x1 , . . . , xi−1 , 0, xi+1 , . . . , 0).
This implies the following:
h0 (x1 , . . . , xi−1 , 1, xi+1 , . . . , 0) ≥ h0 (x1 , . . . , xi−1 , 1, xi+1 , . . . , 1).
(17)
Therefore zs (x(i) ) = 1 as per the Equation (10). Repeated application of the statement gives yi ≤
xi , ∀i =⇒ zs (y) ≤ zs (x) as required
t
u
Definition 8. The Dedekind number M (n) is the number of MBFs of n variables. Finding a closed-form
expression for M (n) is known as the Dedekind problem [25,29].
The Dedekind number of known values are shown below: M (1) = 3, this corresponds to the set of
functions:
M1 (x1 ) ∈ {0, 1, x1 },
(18)
where 0 and 1 are the functions that take any input and return 0 or 1 respectively. M (2) = 6 corresponding
to the set of functions:
M2 (x1 , x2 ) = {0, 1, x1 , x2 , x1 ∨ x2 , x1 ∧ x2 }.
(19)
Similarly, M (3) = 20, M (4) = 168, M (5) = 7581, M (6) ≈ 7.8 × 106 , M (7) ≈ 2.4 × 1012 , and
M (8) ≈ 5.6 × 1023 .
Theorem 1. On transforming the largest graph-representable subclass of k th order function to pairwise
Boolean function, the upper bound on the maximal number of required AVs is given by the Dedekind
number D(k).
7
8
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
Proof. The proof is straightforward. Consider a general multinomial, of similar form to Equation (7)
with more than D(k) AVs. It follows from Lemma 1 that at least 2 of the AVs must correspond to the
same MBF, and always take the same values. Hence, all references to one of these AV in the pseudoBoolean representation can be replaced with references to the other, without changing the associated
costs. Repeated application of this process will leave us with a solution with at most D(k) AVs.
t
u
Although this upper bound is large for even small values of k, it is much tighter than the existing upper
k
bound of S(k) = 22 [6] (also see Proposition 24 in [51]).
k
Lemma 2. Let D(k) denote the Dedekind number for all positive values of k. Given S(k) = 22 and for
even values of k, we have:
P
k
k ( )
k
(20)
S(k) ≥ 2 i∈{0,1,...,k}\{ 2 −1, 2 } i D(k).
When k is odd, we have:
P
S(k) ≥ 2
i∈{0,1,...,k}\{ k−1 , k+1 }
2
2
(ki)
D(k).
(21)
Proof. For even small values of k = {3, ..., 8} the upper bound using Dedekind’s number is much tighter
compared to S(k):(M (3) = 20, S(3) = 256), (M (4) = 168, S(4) = 65536), (M (5) = 7581, S(5) ≈
4.29 × 109 ), (M (6) ≈ 7.8 × 106 , S(6) ≈ 1.85 × 1019 ), (M (7) ≈ 2.4 × 1012 , S(7) ≈ 3.4 × 1038 ), and
(M (8) ≈ 5.6 × 1023 , S(8) ≈ 1.156 × 1077 ). For k > 8, D(k) remains unknown, and the development
of a closed form solution remains an active area of research.
Several upper bounds have been derived for D(k) and we use the following bound by Hansel [19,25]
to prove our result.
( kk )
(22)
D(k) ≤ 3 b 2 c ,
k
D(k) ≤ 2
log2 (3)(b k c)
2
.
(23)
The proof is given for two different cases depending on whether k is even or odd. First let us consider the
case when k is even.
k
k
(24)
= k .
b k2 c
2
We can obtain the following:
k × (k − 1)...(k − k2 )
k × (k − 1)...(k − ( k2 − 1))
k
k
=
.
=
=
k
( k2 − 1)
1 × 2...( k2 )
1 × 2...( k2 − 1)
2
(25)
Using binomial theorem we know that
X
i∈{0,1,...k}
k
= 2k .
i
Using Equations( 25) and ( 26) we have the following Equation:
X
k
k
2k =
+2 k .
i
2
k
k
(26)
(27)
i∈{0,1,..., 2 −2, 2 +1,...,k}
Since log2 (3) < 2, it is easy to observe the following:
2k ≥
X
k
i∈{0,1,..., k
2 −2, 2 +1,...,k}
k
k
+ log2 (3) k .
i
2
(28)
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
Taking both sides to the exponent of 2, we have the following:
k
P
k
P
22 ≥ 2
22 ≥ 2
i∈{0,1,..., k −2, k +1,...,k}
2
2
i∈{0,1,..., k −2, k +1,...,k}
2
2
(ki)+log2 (3)( kk2 )
(ki)
P
i∈{0,1,...,k}\{ k −1, k }
2
2
S(k) ≥ 2
(29)
k
log2 (3)( k )
2
k
i
()
(30)
2
D(k).
(31)
This implies that S(k) is significantly larger than D(k). Let us consider the case when k is odd.
k
k
=
k−1 .
b k2 c
2
It is well known that:
k × (k − 1)...(k − k−1
k
2 )(k −
=
k+1
k−1 k+1
1
×
2...(
)(
2
2
2 )
k+1
2 )
=
k
k−1
=
k+1
k−1
2
k
k−1
2
(1 −
Using Equations( 33) and ( 26) we have the following Equation:
X
k
k
2
k
)
.
2 =
+ (1 + 1 −
k + 1 k−1
i
2
k−3 k+3
i∈{0,1,...,
Since log2 (3) < (1 + 1 −
2
k+1 )
,
2
2
).
k+1
(32)
(33)
(34)
,...,k}
2
for k > 8, it is easy to observe the following:
X
2k ≥
k+3
i∈{0,1,..., k−3
2 , 2 ,...,k}
k
k
+ log2 (3) k−1 .
i
2
(35)
By lifting both sides to the power of 2, we have the following relation:
k
P
k
P
22 ≥ 2
22 ≥ 2
i∈{0,1,..., k−3 , k+3 ,...,k}
2
2
i∈{0,1,..., k−3 , k+3 ,...,k}
2
2
P
S(k) ≥ 2
k
)
(ki)+log2 (3)( k−1
2
(ki)
(36)
k
log2 (3)( k−1 )
2
i∈{0,1,...,k}\{ k−1 , k+1 }
2
2
k
i
()
2
(37)
D(k).
(38)
t
u
We observe that S(k) is significantly larger than D(k) when k is odd.
In [52], the problem of improving this upper bound was mentioned as an open problem. In some
sense, both these upper bounds are not practically feasible for even small values of k. This number is
prohibitive because we are looking for an exact transformation that preserves submodularity. By using
auxiliary variables, we can also transform a given higher order function to a non-submodular one using
much fewer variables [21,12,15]. In section 5, we will further tighten the bound for fourth order functions.
Note that this representation of AVs as MBF is over-complete, for example if the MBF of a auxiliary
variable zi is the constant function zi (x) = 1 we can replace minz,zi h(x, z, zi ) with the simpler (i.e. one
containing less auxiliary variables) function minz h(x, z, 1).
Given any function f in F2k , the equivalent pairwise form f 0 ∈ F 2 can be found by solving a linear
program. The construction of the linear program is given in the following section.
9
10
4
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
The Linear Program
A sketch of the formulation can be given as follows: In general, the presence of AVs of indeterminate state,
given a labeling x makes the minimizing an LP non-convex and challenging to solve directly. Instead of
optimizing this problem containing AVs of unspecified state, we create an auxiliary variable associated
with every MBF. Hence given any labeling x the state of every auxiliary variable is fixed a priori, making
the problem convex. We show how the constraints that a particular AV must conform to a given MBF can
be formulated as linear constraints, and that consequently the problem of finding the closest member of
f 0 ∈ F 2 to any pseudo Boolean function is a linear program.
This program will make use of the max-flow linear program formulation to guarantee that the minimum cost labeling of the AVs corresponds to their MBFs. To do this we must first rewrite the cost of
Equation (8) in a slightly different form. We write:
X
X
X
f (x, z) = c∅ +
ci,s (1 − xi ) +
ct,i xi +
ci,j xi (1 − xj )
i
+
X
i
cl,s (1 − zl ) +
X
l
i,j:i>j
X
ct,l zl +
l
cl,m zl (1 − zm ) +
l,m:l>m
X
ci,l xi (1 − zl ).
(39)
i,l
where c∅ is a constant that may be either positive or negative and all other c are non-negative values
referred to as the capacity of an edge. By [11], this form is equivalent to that of (8), in that any function
that can be written in form (8), can also be written as (39) and visa versa.
4.1
The Max-flow Linear Program
Under the assumption that x is fixed, we are interested in finding a minimum of the Equation:
X
X
X
fx (z) = c∅ +
ci,s (1 − xi ) +
ct,i xi +
ci,j xi (1 − xj )
i
+
X
i
cl,s (1 − zl ) +
X
l
i,j:i>j
ct,l zl +
l
= dx,∅ +
X
X
l
dx,∅ = c∅ +
cl,m zl (1 − zm ) +
l,m:l>m
dx,l,s (1 − zl ) +
where
X
ci,s +
i:xi =0
dx,l,s = cl,s +
dx,t,l zl +
X
dx,l,m zl (1 − zm )
(40)
l,m:l>m
X
i:xi =1
X
ci,l xi (1 − zl )
i,l
l
X
X
X
ct,i +
ci,j
(41)
i,j:i>j∧xi =1∧xj =0
ci,l , dx,t,l = ct,l and dx,l,m = cl,m .
(42)
i:xi =1
Then the minimum cost of Equation (39) may be found by solving its dual max-flow program. Writing
∇x,s for flow from the sink, and ∇x,t for flow to the sink, we seek
max ∇x,s + dx,∅ ,
(43)
subject to the constraints that
P
P fx,ij − dx,ij
fx,ji − j:(i,j)∈E fx,ij
j:(j,i)∈E
P
P
∇x,s + j:(j,s)∈E fx,js − j:(s,j)∈E fx,sj
P
P
∇x,t + j:(j,t)∈E fx,jt − j:(t,j)∈E fx,tj
fx,ij
≤ 0,
≤ 0,
≤0
≤0
≥ 0,
∀(i, j) ∈ E
∀i 6= s, t
(44)
(i, j) ∈ E
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
where E is the set of all ordered pairs (l, m) : ∀l > m, (s, l) : ∀l and (l, t) : ∀t, and fx,i,j corresponds to
the flow through the edge (i, j).
We will not use this exact LP formulation, but instead rely on the fact that fx (z) is a minimal cost
labeling if and only if there exists a flow satisfying constraints (44) such that
fx (z) − ∇x,s − dx,∅ ≤ 0.
4.2
(45)
Choice of MBF as a set of linear constraints
We are seeking minima of a quadratic pseudo Boolean function of the form (39), where x is the variables
we are interested in minimizing and z the auxiliary variables. As previously mentioned, formulations that
allow the state of the auxiliary variable to vary tend to result in non-convex optimization problems. To
avoid such difficulties, we specify as the location of minima of z as a set of hard constraints. We want
that:
min fx (z) = fx ([m1 (x), m2 (x), . . . mD(k) (x)])
z
∀x.
(46)
where fx is defined as in (40), and m1 , . . . mD(k) are the set of all possible MBFs defined over x. By
setting all of the capacities di,j to 0, it can be seen that a solution satisfying (46) must exist. It follows
from the reduction described in Lemma 1, and that all functions that can be expressed in a pairwise form
can also be expressed in a form that satisfies these restrictions.
We enforce condition (46) by the set of linear constraints (44) and (45) for all possible choices of x.
Formally we enforce the condition
fx ([m1 (x), . . . , mD(k) (x)]) − ∇x,s − dx,∅ ≤ 0.
(47)
Substituting in (40) we have 2k sets of conditions, namely,
X
l
dx,l,s (1 − ml (x)) +
X
l
X
dx,t,l (1 − ml (x)) +
dx,l,m ml (x) (1 − mm (x)) − ∇x,s ≤ 0, (48)
l,m:l>m
subject to the set of constraints (44) for all x. Note that we make use of the max-flow formulation, and
not the more obvious min-cut formulation, as this remains a linear program even if we allow the capacity
of edges d1 to vary.
Submodularity Constraints We further require that the quadratic function is submodular or equivalently,
the capacity of all edges ci,j be non-negative. This can be enforced by the set of linear constraints that
ci,j ≥ 0,
4.3
∀i, j.
(49)
Finding the nearest submodular Quadratic Function
We now assume that we have been given an arbitrary function g(x) to minimize, that may or may not lie
in F k . We are interested in finding the closest possible function in F 2 to it. To find the closest function
to it (under the L1 norm), we minimize:
1
In itself d is just a notational convenience, being a sum of coefficients in c.
11
12
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
min
c
g(x)− min f (x, z) =
(50)
g(x)−f (x, m(x)) =
(51)
z
x∈Bk
min
c
X
x∈Bk
min
c
X
X
g(x)− c∅ +
X
ci,s (1 − xi ) +
X
i
x∈Bk
+
X
i
cl,s (1 − ml (x)) +
X
X
ci,j xi (1 − xj )
(52)
i,j:i>j
X
l
+
ct,i xi +
X
ct,l (1 − ml (x)) +
l
cl,m ml (x) (1 − mm (x))
l,m:l>m
ci,l xi (1 − ml (x))
i,l
where m(x) = [m1 (x), . . . , mD(k) (x)] is the vector of all MBFs over x, and P
subject to the family of
constraints
set
out
in
the
previous
subsection.
Note
that
expressions
of
the
form
i |gi | can be written as
P
t
u
i hi subject to the linear constraints hi > gi and hi > −gi and this is a linear program.
4.4
Discussion
Several results follow from the linear program described in the previous section. In particular, if we
consider a function g of the same form as Equation (3) such that
X
min
c
g(x) − min f (x, z) = 0.
z
x∈Bk
(53)
exactly defines a linear polytope for any choice of |x| = k, and this result holds for any choice of basis
functions.
Of equal note, the convex-concave procedure [47] is a generic move-making algorithm that finds
local optima by successively minimizing a sequence of convex (i.e. tractable) upper-bound functions
that are tight at the current location (x0 ). [34] showed how this could be similarly done for quadratic
Boolean functions, by decomposing them into submodular and supermodular components. The work [30]
showed that any function could be decomposed into a quadratic submodular function, and an additional
overestimated term. Nevertheless, this decomposition was not optimal, and they did not suggest how to
find a optimal overestimation. The optimal overestimation which lies in F 2 for a cost function defined
over a clique g may be found by solving the above LP subject to the additional requirements:
g(x) ≤ minz f (x, z),
0
0
g(x ) ≥ minz f (x , z).
∀x 6= x0
(54)
(55)
Efficiency concerns: As we consider larger cliques, it becomes less computationally feasible to use the
techniques discussed in this section, at least without pruning the number of auxiliary variables considered.
As previously mentioned, constant AVs and AVs that corresponds to that of a single variable in x i.e.
zl = xi can be safely discarded without loss of generality. In the following section, we show that a
function in F24 can be represented by only two AVs, rather than 168 as suggested by the Dedekind number.
However, in the general case a minimal form representation eludes us. As a matter of pragmatism, it
may be useful to attempt to solve the LP of the previous section without making use of any AV, and to
successively introduce new variables, until a minimum cost solution is found.
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
5
Tighter Bounds: Transforming functions in F24 to F 2
Consider the following submodular function f (x1 , x2 , x3 , x4 ) ∈ F 4 represented as a multi-linear polynomial:
X
X
X
f (x1 , x2 , x3 , x4 ) = a0 +
a i xi +
aij xi xj +
aijk xi xj xk + a1234 x1 x2 x3 x4 , ∆ij (x) ≤ 0
i
i>j
i>j>k
(56)
where i, j, k ∈ S4 and ∆ij (x) is the discrete second derivative of f (x) with respect to xi and xj .
Consider a function h(x1 , x2 , x3 , x4 , zs ) ∈ F 2 where zs is an AV used to model functions in F24 . In
general, we need several AVs to transform a function in F24 to a function in F2 . Any general function in
F 2 using one AV, can be represented as a multi-linear polynomial (consisting of linear and bilinear terms
involving all five variables):
h(x1 , x2 , x3 , x4 , zs ) = b0 +
X
i
bi x i −
X
bij xi xj + (gs −
i>j
4
X
gs,i xi )zs ,
bij ≥ 0, gs,i ≥ 0, i, j ∈ S4 .
i=1
(57)
The negative signs in front of the bilinear terms (xi xj , zs xi ) emphasize that their coefficients (−bij , −gs,i )
must be non-positive to ensure submodularity. We have the following condition from Equation (4), given
in page 3:
f (x1 , x2 , x3 , x4 ) = min h(x1 , x2 , x3 , x4 , zs ),
zs ∈B
∀x.
(58)
Here the coefficients (ai , aij , aijk , aijkl ) in the function f (x) are known. We wish to compute the coefficients (bi , bij , gs , gs,n ) where i, j ∈ V, i 6= j, n ∈ S4 . If we were given (gs , gs,i ) then from Equations (57) and (58) we would have:
(
P4
1 if gs − i=1 gs,i xi < 0,
zs =
(59)
0 otherwise.
Our main result is to prove that any function h ∈ F24 can be transformed to a function h0 (x1 , x2 , x3 , x4 , zt , zr ) ∈
F 2 involving only two auxiliary variables zt and zr as stated in Theorem 2.
Let A be the family of sets corresponding to labelings of x such that: zs = 0 = arg minzs h(x, zs ).
In the same way let B be the family of sets corresponding to labelings of x such that:zs = 1 =
arg minzs h(x, zs ). These sets A and B partition x, as defined below:
Definition 9. A partition divides P into sets A and B such that A = {S(x) : 0 = arg minz∈B h(x, z), x ∈
B4 } and B = P\A. Note that ∅ ∈ A. Here S(x) denotes the set corresponding to x.
In the rest of the paper, we say that the AV zs is associated with [A, B] or denote it by zs : [A, B]. We
illustrate the concept of a partition in Figure 2.
A few partitions that play a key role in our transformation are referred to as forward, backward, and
intermediate partitions.
Definition 10. The forward reference partition [Af , Bf ] takes the form:
B ∈ Bf ⇐⇒ |B| ≥ 3, Af = P\Bf
(60)
The backward reference partition [Ab , Bb ] is shown below:
B ∈ Bb ⇐⇒ |B| ≥ 2, Ab = P\Bb
(61)
13
14
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
(a)
(b)
Fig. 2. We show some examples of partitions using Hasse diagrams. Here, we use set representation for denoting the labelings of (x1 , x2 , x3 , x4 ). For example the set {1, 2, 4} is equivalent to the labeling {x1 =
1, x2 = 1, x3 = 0, x4 = 1}. In (a), A = {{}, {2}, {3}, {4}, {2, 3}, {2, 4}, {3, 4}, {2, 3, 4}} and B =
{{1}, {1, 2}, {1, 3}, {1, 4}, {1, 2, 3}, {1, 2, 4}, {1, 3, 4}, S4 }. (a) and (b) are examples of partitions. Any arbitrary
AV must be associated with one of these 168 partitions as given by the Dedekind number D(k).
Figure 3(a) and (b) show the forward and backward partitions respectively.
We consider a set of 18 partitions as intermediate partitions [Ai , Bi ] as shown in Figure 4. There
are 6 intermediate partitions where there are five sets in Bi that have cardinality 2 (one such partition
is shown in Figure 4(a)). There are 12 intermediate partitions where there are four sets in Bi that have
cardinality 2 (one such partition is shown in Figure 4(b)). One may expect more intermediate partitions
by considering all possible different sets in Bi having cardinality 2. However, we will see later that such
partitions are not necessary for transforming a function in F24 to a function in F 2 .
(a)
(b)
Fig. 3. The two reference partitions, referred to as forward and backward, are shown.
The basic idea in our work is to replace several AVs using the minimum number of AVs without changing
the values of the function at their respective minima.
Definition 11. We say that a function h(x, z) can be transformed to another function h0 (x, z0 ) where
z 6= z0 if the following condition is satisfied:
min h(x, z) = min
h0 (x, z0 ),
0
z
z
∀x
(62)
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
(a)
(b)
Fig. 4. We have a total of 18 intermediate partitions. In (a), we show one of the 6 intermediate partitions where five
sets in Bi have cardinality 2. We denote this as I(34), where the index refers to the only set that does not have
cardinality 2. In (b), we show one of the 12 intermediate partitions where four sets in Bi have cardinality 2. We
denote this as I(24, 34), where the indices refer to the sets that do not have cardinality 2.
where z and z0 are vectors of auxiliary variables with different partitions. The cardinality of z need not
be equal to the cardinality of z0 .
Through a sequence of transformations of the above form, we start with a general function h(x, z) and
finally compute a function h0 (x, zs , zt ) with only two AVs in reference partitions.
Lemma 3. Let za : [As , Bs ] and zb : [As , Bs ] be two AVs that have the same partition then h(x, za , zb ) ∈
F 2 can be transformed to some function h0 (x, zs ) ∈ F 2 involving only one AVzs .
Proof. According to the Equation 62, we can transform a function h(x, za , zb ) to h0 (x, zs ) if it satisfies
the following condition:
min h(x, za , zb ) = min h0 (x, zs ),
za ,zb
zs
∀x
(63)
Since the AVs za and zb take the same partition [As , Bs ] their Boolean values are equal for different
configurations of x. Thus we can replace all instances of za and zb with zs .
t
u
Theorem 2. Any function f (x) in F24 can be transformed to some function h(x, zf , zs ) in F 2 where
zf correspond to the forward partition and zs can either be the backward partition or one of the 18
intermediate partitions.
Proof. Using Theorem 5.2 from [37] we can decompose a given submodular function in F 4 into functions
in 10 different groups Gi , i = {1..10} where each Gi is shown in Table 1. As shown in [50] the functions
in G10 does not belong to F24 . It was also shown that any submodular function that has any functions from
group G10 does not belong to F24 according to Theorem 16(3) in [50]. Thus all the functions in F2k should
be composed of functions in the groups Gi , i ∈ {1, ..., 9}.
The number of distinct terms in each group Gi is given in Table 1. Overall, there are 31 distinct
functions in the groups Gi , i ∈ {1, ..., 9}. The terms in the first group G1 has only second degree terms.
Hence, the functions in this group does not require any AVs. The terms in the next 7 groups Gi , i ∈
{2, ..., 8} can each be represented by a single AV, which can be either zf or zb . Here zf and zb denote
AV s in the forward and backward partitions respectively. The 6 terms in G9 can be represented using two
AV s zf and zi , where zf and zi correspond to forward and intermediate reference partitions (denoted by
I(k, l) in Figure 4(a)) respectively. It is important to note that the functions in G9 involve interaction
between zf and zi , i.e., there exists a bilinear term zf zi in G9 .
We prove the result by considering two cases.
15
16
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
Group Gi
G1 (i, j)
G2 (i, j, k)
G3
|Gi |
6
4
1
G4
1
G5 (i, j, k) 4
G6 (i, j, k) 4
G7 (i)
4
G8
1
G9 (i, j)
6
f (x)
−xi xj
−xi xj xk
−xi xj xk xl
−xi xj xk xl + xi xj xk + xi xj xl + xi xk xl +
xj xk xl − xi xj − xi xk − xi xl −
xj xk − xj xl − xk xl
xi xj xk xl − xi xj xk − xi xl − xj xl −
xk xl
xi xj xk − xi xj − xi xk − xj xk
xi xj xk xl − xi xj xk − xi xj xl − xi xk xl
2xi xj xk xl − xi xj xk − xi xj xl − xi xk xl −
xj xk xl
xi xj xk xl − xi xj − xi xk xl − xj xk xl
minz1 ,z2 h(x, z1 , z2 ) where h ∈ F 2
−xi xj
minzf (2 − xi − xj − xk )zf
minzf (3 − xi − xj − xk − xl )zf
minzb (1 − xi − xj − xk − xl )zb
minzb (2 − xi − xj − xk − 2xl )zb
minzb (1 − xi − xj − xk )zb
minzf (3 − 2xi − xj − xk − xl )zf
minzf (2 − xi − xj − xk − xl )zf
minzf ,zi ((2 − xk − xl )zf +
(1 − xi − xj )zi
−zf zi )
−xi xj xk xl + xi xk xl + xj xk xl −
f (x) ∈
/ F24 as shown in [50]
xi xk − xi xl − xj xk − xj xl − xk xl
Table 1. The above table is adapted from Figure 2 of [51] where {i, j, k, l} = S4 . Each group Gi has several terms
depending on the values of {i, j, k, l}. The number of distinct terms in each group is given by |Gi |. Since the groups
G4 and G8 involve all four variables and are symmetric, they contain one function each. zf and zb correspond to
AV s for forward and backward partitions. zi corresponds to one of the intermediate partitions denoted by I(kl) in
Figure 4(a). For each group Gi , we also use an index (.) in the first column to identify a specific function from others
in its group.
G10
6
Absence of G9 functions: In the first case, we consider functions that can be expressed as a sum of
functions in the first 8 groups Gi , i = {1..8}. In other words, we study the scenario where we express
the function without using any function from G9 . Let us denote such a function as f 0(x) that can be
expressed as a sum of 25 functions from the 8 groups Gi , i = {1..8} as shown below:
f 0(x) = α1 G1 (i, j) + · · · α25 G8
(64)
The only AVs involved in all the functions are zf and zb . Using Lemma 3 we can obtain a function
that uses only two variables zf and zb as shown below:
f 0(x) = g(x) + min (gf (x)zf + gb (x)zb ),
zf ,zb
(65)
where g(x), gf (x) and gb (x) are functions involving x. This implies that any function that can be expressed without any function from G9 can be expressed using the only forward and backward partitions.
Presence of G9 functions: Let us consider an arbitrary function f (x) in F24 that is expressed as a sum of
functions from these 31 groups including functions from G9 :
f (x) = α1 G1 (i, j) + α2 G2 (i, j) + · · · + α31 G9 (k, l)
(66)
In Table 2, we show that the sum of two functions can always be represented using only two auxiliary
variables. In Tables 3, 4 and 5 we show the sum of functions with 3, 4 and 5 terms respectively. Different
combinations of functions lead to functions that can always be expressed with only 2 auxiliary variables.
Without loss of generality, we have avoided the repetition for all possible indices by treating them
using the set {i, j, k, l} = {1, 2, 3, 4}. We have already proved the case where where G9 is absent. Thus,
the tables only show summations that involve at least one function from G9 .
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
In some cases, we do not show the sums of functions with real coefficients (α, β, γ, δ, η) to demonstrate the special scenarios where the combination of two functions involving intermediate partition zi
can be transformed to a function that involves only zf and zb . In such cases, we can do the following
sequentially:
α1 f1 (x) + α2 f2 (x) = α1 (f1 (x) + f2 (x)) + (α2 − α1 )f2 (x),
α1 ≤ α2
(67)
Let β = α2 − α1 and let f3 (x) = f1 (x) + f2 (x) as per the Table 2. Now we can further use Table 2
on α1 f3 (x) + βf2 (x) to generate other functions.
As we see in Table 1, the function G9 (i, j) uses two auxiliary variables zf and zi ∈ I(k, l). As we
observe in Table 2, on adding the function G9 (i, j) with other functions we have the following scenarios:
1. The coefficient of zi is unaltered and we only change the coefficients of zf . This happens in 6 of
the additions in Table 2 as given by (G9 (i, j), G1 (i, j)), (G9 (i, j), G2 (j, k, l)), (G9 (i, j), G7 (k)) and
(G9 (i, j), G8 ).
2. We obtain a function that can be expressed with only one zf . This happens in 4 of the additions in
Table 2 as given by (G9 (i, j), G2 (i, j, k)), (G9 (i, j), G3 ),(G9 (i, j), G7 (i)), and (G9 (i, j), G9 (k, l)).
3. We obtain a function that can be expressed with only one zb . This happens in 2 of the additions shown
in Table 2 as given by (G9 (i, j), G4 ) and (G9 (i, j), G6 (j, k, l)).
4. We obtain a function that can be expressed with zf and zb . This happens in one of the additions
shown in Table 2 as given by (G9 (i, j), G5 (i, j, k)).
5. We obtain a function with zf and zi , whose coefficients are changed. This happens in 3 of the additions shown in Table 2 as given by (G9 (i, j), G5 (j, k, l)), (G9 (i, j), G6 (i, j, k)), and (G9 (i, j), G9 (i, k)).
There are 6 functions in group G9 . However, as shown in Table 2, additions involving G9 (i, j) and
G9 (k, l) produce functions involving only one auxiliary variable zf . In other words, sum of functions
involving G9 (i, k) can sometimes be represented using functions from the first 8 groups (G1 to G8 ). Out
of the 6 functions in G9 only two of them are necessary at a time. Without loss of generality, we rewrite
the f (x) using a maximum of 2 functions in group G9 as shown below:
f (x) = α1 G1 (i, j) + · · · + α26 G9 (i, j) + α27 G9 (i, k).
(68)
The remaining four terms in G9 are not necessary due to the following reasons:
– G9 (i, l) is not necessary because its addition to G9 (i, j) and G9 (i, k) will lead to a function involving
only zf and zb as per the second last row of Table 3.
– Any function G9 (j, k) is not necessary because its addition to G9 (i, j) and G9 (i, k) will lead to a
function involving only zf and zb as per the last row of Table 3.
– Any function G9 (j, l) is not necessary because its addition to G9 (i, k) can be represented using a
function that involves only zf as per the last row of Table 2.
– Any function G9 (k, l) is not necessary because its addition to G9 (i, j) can be represented using a
function that involves only zf as per the last row of Table 2.
We observed that we need a maximum of two functions from G9 to represent any function in F24 . So there
are two possibilities for f (x) and we denote them as f 1(x) and f 2(x) depending on whether we use one
or two of the functions from G9 as shown below:
f 1(x) =αG9 (i, j) + βG5 (i, k, l) + γG5 (j, k, l) + δG6 (i, j, k) + ηG6 (i, j, l) + σ min(gf (x)zf ) + g(x),
zf
(69)
f 2(x) =αG9 (i, j) + βG9 (i, k) + γG5 (j, k, l) + δG6 (i, j, k) + η min(gf (x)zf ) + g(x).
zf
(70)
17
18
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
We have represented f 1 and f 2 using 7 and 6 terms respectively. We show that we do not need any
other functions for representing f 1 and f 2:
– G1 : We can represent them using g(x).
– G2 : These functions involve zf and we can represent them using minzf gf (x, zf ) according to
Lemma 3.
– G3 : These functions involve zf and we can represent them using minzf gf (x, zf ) according to
Lemma 3.
– G4 : By adding this function to G9 (i, j) we obtain functions in G6 . The generated G6 functions can
be subsequently added to any functions in G9 (i, j) or G9 (i, k). If there is no G9 term, then we can
represent the function using zf and zb as explained earlier in the case of f 0.
– G5 : In the case of f 1, some functions in G5 can be added to G9 (i, j) to obtain G6 and G8 . The
generated functions can be subsequently added to any G9 (i, j) terms. If there is no G9 (i, j) term, then
we can represent the function using zf and zb as explained earlier in the case of f 0. The functions
G5 (i, k, l) and G5 (j, k, l) alter the coefficients of zi . For now, we keep these functions of G5 as separate
terms in function f 1(x).
In the case of f 2, some functions in G5 can be added to G9 (i, j) or G9 (i, k) to obtain G6 and G8
terms. The generated functions can be subsequently added to any G9 (i, j) or G9 (i, k) terms. If there
is no G9 (i, j) term, then we can represent the function using zf and zb as explained earlier in the case
of f 0. On adding the function G5 (j, k, l) to G9 (i, j) or G9 (i, k) we produce functions that alter the
coefficients of zi . For now, we keep this function in G5 to represent the original function f 2(x).
– G6 : In the case of f 1, some functions in G6 can be added to G9 (i, j) to obtain functions in G5 , which
can be subsequently added to any G9 (i, j) terms. If there is no G9 (i, j) term, then we can represent
the function using zf and zb as explained earlier in the case of f 0. The functions G6 (i, j, k) and
G6 (i, j, l) produce functions that alter the coefficients of zi . For now, we keep these two functions of
G6 as separate terms in f 1(x).
In the case of f 2, some functions in G6 can be added to G9 (i, j) or G9 (i, k to obtain G5 , which can
be subsequently added to any G9 (i, j) or G9 (i, k) terms. If there is no G9 (i, j) term, then we can
represent the function using zf and zb as explained earlier in the case of f 0. On adding the function
G6 (i, j, k) to G9 (i, j) or G9 (i, k) we produce functions that alter the coefficients of zi . For now, we
keep this function G6 (i, j, k) as separate term in f 2(x).
– G7 : Some functions in G7 can be added to G9 (i, j) or G9 (i, k) to generate functions in G8 , which
can be subsequently added to any G9 (i, j) or G9 (i, k) terms. In other cases, the functions in G7 only
modify the coefficients of zf terms that can be represented by the function minzf (gf (x)zf ).
– G8 : This function can be represented by minzf (gf (x)zf ) since it only has one AVzf .
Using Table 5 we rewrite f 1 as given below:
f 1(x) =g(x) + min ((α(2 − xk − xl ) + σg(x))zf +
zf ,zi
(71)
(α(1 − xi − xj )+
β(2 − xi − 2xj − xk − xl )+
γ(2 − 2xi − xj − xk − xl )+
δ(1 − xi − xj − xk )+
η(1 − xi − xj − xl ))zi −
αzf zi )
=g(x) + min (gf0 (x)zf + gi (x)zi − αzf zi )
zf ,zi
(72)
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
minz1 ,z2 h(x, z1 , z2 ),
AV s
where h(x, z1 , z2 ) = f1 (x) + f2 (x), ∀x
−βxi xj + α minzf ,zi ((2 − xk − xl )zf +
zf ,
(1 − xi − xj )zi −
αG9 (i, j)
βG1 (i, j)
zi ∈ I(kl)
zf zi )
−xi xj + G7 (k) =
G9 (i, j)
G2 (i, j, k)
zf
−xi xj + minzf (3 − xi − xj − 2xk − xl )zf
minzf ,zi ((α(2 − xk − xl ) + β(2 − xj − xk − xl ))zf +
zf ,
α(1 − xi − xj )zi −
αG9 (i, j)
βG2 (j, k, l)
zi ∈ I(kl)
αzf zi )
−xi xj + G2 (i, k, l) + G2 (j, k, l) =
G9 (i, j)
G3
zf
−xi xj + minzf ((2 − xi − xk − xl ) + (2 − xj − xk − xl ))zf
G6 (i, j, k) + G6 (i, j, l) - xk xl =
G9 (i, j)
G4
zb
−xk xl + minzb (2 − 2xi − 2xj − xk − xl )zb
G8 + G6 (i, j, l) - xk xl =
zf ,
−xk xl + minzf (2 − xi − xj − xk − xl )zf +
G9 (i, j)
G5 (i, j, k)
zb
minzb (1 − xi − xj − xl )zb
minzf ,zi ((α(2 − xk − xl )zf +
zf ,
(α(1 − xi − xj ) + β(2 − 2xi − xj − xk − xl ))zi −
αG9 (i, j)
βG5 (j, k, l)
zi ∈ I(kl)
αzf zi )
minzf ,zi ((α(2 − xk − xl )zf +
zf ,
(α(1 − xi − xj ) + β(1 − xi − xj − xk ))zi −
αG9 (i, j)
βG6 (i, j, k)
zi ∈ I(kl)
αzf zi )
G5 (i, k, l) − xk xl =
G9 (i, j)
G6 (j, k, l)
zb
−xk xl + minzb (2 − xi − 2xj − xk − xl )zb
G8 − xi xj + G2 (i, k, l) =
G9 (i, j)
G7 (i)
zf
−x1 x2 + minzf ((2 − xi − xj − xk − xl ) + (2 − xi − xk − xl ))zf
minzf ,zi ((α(2 − xk − xl ) + β(3 − xi − xj − 2xk − xl ))zf +
zf ,
α(1 − xi − xj )zi −
αG9 (i, j)
βG7 (k)
zi ∈ I(kl)
αzf zi )
minzf ,zi (α(2 − xk − xl ) + β(2 − xi − xj − xk − xl ))zf +
zf ,
α(1 − xi − xj )zi −
αG9 (i, j)
βG8
zi ∈ I(kl)
αzf zi )
minzf ,zi ((α(2 − xk − xl ) + β(2 − xj − xl ))zf +
zf ,
(α(1 − xi − xj ) + β(1 − xi − xk ))zi αG9 (i, j)
βG9 (i, k)
zi ∈ I(kl, jl)
αzf zi − βzf zi )
G8 − xi xj − xk xl =
G9 (i, j)
G9 (k, l)
zf
−xi xj − xk xl + minzf ((2 − xi − xj − xk − xl )zf
Table 2. We show the sum of a function G9 (i, j) with any other function in Table 1 can be expressed using two
auxiliary variables. Here the index set {i, j, k, l} = S4 denotes the four distinct integers S4 = {1, 2, 3, 4}.
f1 (x), α ∈ R+ f2 (x), β ∈ R+
19
20
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
minz1 ,z2 h(x, z1 , z2 ),
AV s
h(x, z1 , z2 ) = f1 (x) + f2 (x) + f3 (x), ∀x
minzf ,zi ((α(2 − xk − xl )zf +
(α(1 − xi − xj )+
zf ,
β(2 − xi − 2xj − xk − xl )+
αG9 (i, j)
βG5 (i, k, l)
γG5 (j, k, l)
zi ∈ I(kl)
γ(2 − 2xi − xj − xk − xl ))zi −
αzf zi )
minzf ,zi ((α(2 − xk − xl )zf
(α(1 − xi − xj )+
zf ,
β(1 − xi − xj − xk )+
αG9 (i, j)
βG6 (i, j, k)
γG6 (i, j, l)
zi ∈ I(kl)
γ(1 − xi − xj − xl ))zi −
αzf zi )
minzf ,zi ((α(2 − xk − xl )zf
(α(1 − xi − xj )+
zf ,
β(2 − 2xi − xj − xk − xl )+
αG9 (i, j)
βG5 (j, k, l)
γG6 (i, j, l)
zi ∈ I(kl)
γ(1 − xi − xj − xl ))zi −
αzf zi )
minzf ,zi ((α(2 − xk − xl ) + β(2 − xj − xl ))zf
(α(1 − xi − xj )+
zf ,
β(1 − xi − xk )+
αG9 (i, j)
βG9 (i, k)
γG5 (j, k, l)
zi ∈ I(kl, jl)
γ(2 − 2xi − xj − xk − xl ))zi −
αzf zi − βzf zi )
minzf ,zi ((α(2 − xk − xl ) + β(2 − xj − xl ))zf
(α(1 − xi − xj )+
zf ,
β(1 − xi − xk )+
αG9 (i, j)
βG9 (i, k)
γG6 (i, j, k)
zi ∈ I(kl, jl)
γ(1 − xi − xj − xk ))zi −
αzf zi − βzf zi )
−xj xk xl + G8 + G5 (j, k, l) =
zf ,
−xj xk xl + minzf (2 − xi − xj − xk − xl )zf +
G9 (i, j)
G9 (i, k)
G9 (i, l)
zb
minzb (2 − 2xi − xj − xk − xl )zb
G7 (l) + G8 + G6 (i, j, k) =
minzf ((3 − xi − xj − xk − 4xl )+
zf ,
G9 (i, j)
G9 (i, k)
G9 (j, k)
zb
(2 − xi − xj − xk − xl ))zf +
minzb (1 − xi − xj − xk )zb
Table 3. We show the sum of any three functions from Table 1 can be expressed using two auxiliary variables. Here
the index set {i, j, k, l} = S4 denotes the four distinct integers S4 = {1, 2, 3, 4}.
f1 (x), α ∈ R+ f2 (x), β ∈ R+ f3 (x), γ ∈ R+
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
minz1 ,z2 h(x, z1 , z2 ),
AV s
h(x, z1 , z2 ) = f1 (x) + f2 (x) + f3 (x) + f4 (x), ∀x
(G9 (i, j) + G5 (i, j, k))+
zf ,
(G5 (i, k, l) + G5 (j, k, l)) =
G9 (i, j) G5 (i, k, l) G5 (j, k, l) G5 (i, j, k)
zb
(G8 + G6 (i, j, l) + G1 (k, l))+
(G5 (i, k, l) + G5 (j, k, l))
(G9 (i, j) + G6 (j, k, l))+
(G6 (i, j, k) + G6 (i, j, l)) =
zf ,
G9 (i, j) G6 (i, j, k) G6 (i, j, l) G6 (j, k, l)
zb
(G5 (i, j, k) + G1 (k, l))+
(G6 (i, j, k) + G6 (i, j, l))
minzf ,zi (α(2 − xk − xl )zf +
(α(1 − xi − xj )+
zf ,
β(2 − xi − 2xj − xk − xl )+
αG9 (i, j) βG5 (i, k, l) γG5 (j, k, l) δG6 (i, j, k)
γ(2 − 2xi − xj − xk − xl )+ zi ∈ I(k, l)
δ(1 − xi − xj − xk ))zi −
αzf zi )
minzf ,zi (α(2 − xk − xl )zf +
(α(1 − xi − xj )+
zf ,
β(1 − xi − xj − xk )+
αG9 (i, j) βG6 (i, j, k) γG6 (i, j, l) δG5 (i, k, l)
γ(1 − xi − xj − xl )+ zi ∈ I(k, l)
δ(2 − xi − 2xj − xk − xl ))zi −
αzf zi )
minzf ,zi ((α(2 − xk − xl ) + β(2 − xj − xl ))zf +
(α(1 − xi − xj )+
zf ,
β(1 − xi − xk )+
αG9 (i, j) βG9 (i, k) γG5 (j, k, l) δG6 (i, j, k)
γ(2 − 2xi − xj − xk − xl )+ zi ∈ I(k, l)
δ(1 − xi − xj − xk ))zi −
αzf zi − βzf zi )
Table 4. We show the sum of any four functions from Table 1 can be expressed using two auxiliary variables. Here
the index set {i, j, k, l} = S4 denotes the four distinct integers S4 = {1, 2, 3, 4}.
f1 (x)
f2 (x)
f3 (x)
f4 (x)
minz1 ,z2 (f1 (x) + f2 (x) + f3 (x) + f4 (x) + f5 (x)),
∀x
minzf ,zi (α(2 − xk − xl )zf +
(α(1 − xi − xj )+
β(2 − xi − 2xj − xk − xl )+
γ(2 − 2xi − xj − xk − xl )+
αG9 (i, j) βG5 (i, k, l) γG5 (j, k, l) δG6 (i, j, k) ηG6 (i, j, l)
δ(1 − xi − xj − xk )+
η(1 − xi − xj − xl ))zi −
αzf zi )
Table 5. We show the sum of five functions from Table 1 can be expressed using two auxiliary variables zf and
zi ∈ I(k, l). Here the index set {i, j, k, l} = S4 denotes the four distinct integers S4 = {1, 2, 3, 4}.
f1 (x),
α ∈ R+
f2 (x),
α ∈ R+
f3 (x),
α ∈ R+
f4 (x),
α ∈ R+
f5 (x),
α ∈ R+
21
22
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
Using the last row of Table 4 we rewrite f 2 as given below:
f 2(x) =g(x) + min ((α(2 − xk − xl ) + β(2 − xj − xl ) + gf (x))zf +
zf ,zi
(73)
(α(1 − xi − xj )+
β(1 − xi − xk )+
γ(2 − 2xi − xj − xk − xl )+
δ(1 − xi − xj − xk ))zi −
αzf zi − βzf zi )
=g(x) + min (gf0 (x)zf + gi (x)zi − (α + β)zf zi )
zf ,zi
(74)
It is shown that f 1 and f 2 need only two AVs zf and zs . In the case of f 0, zs is a backward partition.
In the case of f 1 and f 2, zs belongs to one of the 18 intermediate partitions.
t
u
6
Linear Programming solution
For a given function f (x1 , x2 , x3 , x4 ) in F24 , our goal is to compute a function h(x, z) in F 2 . Theorem 2
shows that we need only two AVs (zf , zs ). Here zf corresponds to the forward reference partition. The
AV zs is either the backward partition or one of the 18 intermediate reference partitions. Unfortunately,
we do not know which one of these 19 partitions is required before we do the transformation. In what
follows, we will show the transformation assuming that we know the specific partition for zs . Note that
zb is a special case of zi and we do not use the bilinear term zf zs when zs = zb . In order to handle this
condition we use a Boolean variable that takes the value 0 when the intermediate partition is the backward
reference partition and 1 otherwise:
(
0 if zi ∈ [Ab , Bb ],
δ(zs ) =
(75)
1 otherwise.
The required function h(x, z) is the following:
h(x, zf , zs ) = b0 +
X
i
bi x i −
X
bij xi xj + (gf −
4
X
gf,i xi )zf + (gs −
i=1
i>j
4
X
gs,i xi )zs − δ(zs )jf s zf zs ,
i=1
(76)
such that bij , gf,i , gs,i , jf s ≥ 0 and i, j ∈ S4 . As we know the partitions of (zf , zs ), we know their
Boolean values for all labelings of x. We need the coefficients (bi , bij , jf s , gf , gs , gf,i , gs,i ), i ∈ S4 to
compute h(x1 , x2 , x3 , x4 , zf , zs ). These coefficients satisfy both submodularity constraints (that the coefficients of all bilinear terms (xi xj , xi zf , xj zs , zf zs ) are less than or equal to zero) and those imposed
by the reference partitions. First we list the submodularity conditions below:
T
bij
gf,i
≥ 0, i, j = S4 , i 6= j,
gs,i
jf s
| {z }
Sp
(77)
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
where 0 refers of a vector composed of 0’s of appropriate length. Next we list the conditions which
guarantee f (x) = minzf ,zs h(x, zf , zs ),
∀x. Let η(S) be the value of zf zs for S ∈ P. This can be
obtained using the partitions of zf and zs .
(
1
η(S) =
0
if S ∈ (Bf ∩ Bs )
otherwise.
(78)
Let us denote the value of AVzs for different subsets of S4 as given below:
zsS
(
1
=
0
if S ∈ Bs ,
if S ∈ As .
(79)
Let G and H denote values of functions f and h respectively:
G = f (1S1 , 1S2 , 1S3 , 1S4 ),
H = h(1S1 , 1S2 , 1S3 , 1S4 , 0, 0) + (gf −
4
X
∀S ∈ P
gf,i 1Si )zfS + (gs −
i=1
4
X
gs,i 1Si )zsS − δ(zs )η(S)jf s
(80)
(81)
i=1
As a result we have the following 16 linear Equations (N.B. there are 24 (16) different S):
G = H,
∀S ∈ P
(82)
We already know the partition of (zf , zs ) and their appropriate values a priori. The following constraints
ensure that zf and zs behave according to their associated partitions.
P4
gf − i=1 gf,i 1Si
P4
≥ 0, S ∈ Af , D ∈ As
gs − i=1 gs,i 1D
i
{z
}
|
(83)
P4
gf − i=1 gf,i 1Si − δ(zs )η(S)jf s
P4
≤ 0, S ∈ Bf , D ∈ Bs .
gs − i=1 gs,i 1D
i − δ(zs )η(D)jf s
{z
}
|
(84)
Gg
Gl
We need to compute the coefficients (bij , gf , gf,i , gs , gs,i , jf s ) that satisfy the Equations (77), (82), and
(84). This is equivalent to finding a feasible point in a linear programming problem:
min const
(85)
s.t Sp ≥ 0, G = H, Gg ≥ 0, Gl ≤ 0
(86)
In the above LP formulation we assumed that we know the partition of AVs zf and zs . However, zs can
be one of the 19 partitions. Before we do the transformation it is not easy to know which one of the 19
partitions is necessary. So we solve the LP 19 times and iterate over all the 19 partitions to identify the
necessary one. For the correct partition, will be able to find a solution that satisfies all the constraints.
23
24
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
Fig. 5. We generated submodular functions using non-weighted sum of functions from groups Gi , i = {2, ..., 9}. The
x-axis denotes the number of functions chosen from Gi , i = {2, ..., 8} in generating the submodular functions and
the y-axis gives the percentage of transformations requiring zi in intermediate partition.
7
Experiments
The functions in the class F24 can be transformed to functions in F 2 using 25 AVs according to existing
results [49]. We show that this transformation can be done using only two AVs using a linear program. In
Matlab, the transformation takes around 0.03 seconds and it can be further improved using efficient C++
implementation.
In our experiments as shown in Figure 5, we generated submodular functions using non-weighted
sum of functions from G9 and functions from groups Gi , i = {2, ..., 8}. We do not consider functions
from group G1 since they do not require any AVs. The number of functions nG used from groups Gi , i =
{2, ..., 8} is increased from 0 to 19. For each value of nG , we generated 1000 functions and the nonnegative weights are randomly generated in the interval [0, 1]. We observed that as we increase nG , the
generated submodular functions were less likely to use intermediate AVs. This also concurs with Table 2,
that many combinations of G9 with other functions can be represented using functions in the first 8 groups
(G1 to G8 ) that do not require any AVs in intermediate partition.
8
Discussion and open problems
The reduction of higher order functions to quadratic ones will be beneficial for developing efficient
minimization algorithms. These techniques can be broadly classified into two types: submodularitypreserving [1,28,18,48,38,14,40,50,42] and general techniques [41,12,15,2,21]. This paper belongs to
the submodular-preserving class of algorithms where higher order submodular functions are transformed
to quadratic submodular functions using AVs. The general techniques are usually employed in association with roof-duality approaches for minimizing non-submodular functions [4,3,5,43]. The general techniques also employ AVs and these AVs need not be MBFs. The existing upper bound for general reduction
techniques is given by G(k) = 2k−2 (k−3)+1 for a kth order function. We show the comparison between
the AVs used in general techniques and submodularity-preserving techniques in Table 6.
Note that the upper bound for the number of AVs required for submodularity-preserving transformation is much higher than for general reduction techniques. We have improved the upper bound for
k
submodular functions from 22 to Dedekind number D(k). In the case of fourth order functions we have
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
Type
Degree
34 5
6
7
8
General
Ishikawa [21] 1 5 17
49
129
321
Submodularity Preserving Dedekind [25] 1 2 7581 ≈ 7.8 × 106 ≈ 2.4 × 1012 ≈ 5.6 × 1023
Table 6. Comparison of the number of AVs used for general versus submodularity-preserving techiques. The
Dedekind number D(k) is unknown for k > 8.
further improved the upper bound from 168 (D(4)) to 2.
Acknowledgments: Srikumar Ramalingam would like to thank Mitsubishi Electric Research Laboratories (MERL) for the support. L’ubor Ladický is funded by Max Planck Center for Learning Systems
Fellowship. Philip H.S. Torr would like to acknowledge the financial support provided by ERC grant
ERC-2012-AdG 321162-HELIOS, EPSRC/MURI grant ref EP/N019474/1, EPSRC grant EP/M013774/,
and EPSRC Programme Grant Seebibyte EP/M013774/1.
References
1. A. Billionnet and M. Minoux. Maximizing a supermodular pseudo-boolean function: a polynomial algorithm
for supermodular cubic functions. Discrete Appl. Math., 12(1):1–11, 1985.
2. E. Boros and A. Gruber. On quadratization of pseudo-boolean functions. In ISAIM, 2012.
3. E. Boros and P. L. Hammer. Pseudo-boolean optimization. Discrete Appl. Math., 123(1-3):155–225, 2002.
4. E. Boros and P.L. Hammer. A max-flow approach to improved roof-duality in quadratic 0-1 minimization.
Technical Report RRR 15-1989, RUTCOR, Rutgers University, 1989.
5. E. Boros, P.L. Hammer, R. Sun, and G. Tavares. A max-flow approach to improved lower bounds for quadratic
unconstrained binary optimization (qubo). Discrete Optimization, 5(2):501âĂŞ529, 2008.
6. D.A. Cohen, M.C. Cooper, P. Creed, and S. Zivny P.G. Jeavons. An algebraic theory of complexity for discrete
optimization. In SIAM Journal of Computing, volume 42(5), pages 1915–193, 2013.
7. M.C. Cooper. Minimization of locally defined submodular functions by optimal soft arc consistency. Constraints,
13(4):437âĂŞ458, 2008.
8. Y. Crama and P.L. Hammer. Boolean Functions: Theory, Algorithms and Applications. Cambridge University
Press, 2011.
9. W.H. Cunningham. On submodular function minimization. Combinatorica, 5:185–192, 1985.
10. J. Edmonds. Submodular functions, matroids and certain polyhedra. Calgary International Conference on
Combinatorial Structures and their applications, page 69âĂŞ87, 1969.
11. M.L. Fisher, G.L. Nemhauser, and L.A. Wolsey. An analysis of approximation for maximizing submodular
setfunctions-i. Mathematical Programming Studies, 8:73–87, 1978.
12. A. Fix, A. Gruber, E. Boros, and R. Zabih. A graph cut algorithm for higher-order markov random fields. In
ICCV, pages 1020 – 1027, 2011.
13. L. Fleischer and S. Iwata. A push-relabel framework for submodular function minimization and applications to
parametric optimization. Discrete Applied Mathematics, 131(2):311âĂŞ322, 2001.
14. D. Freedman and P. Drineas. Energy minimization via graph cuts: Settling what is possible. In CVPR, volume 2,
pages 939–946, 2005.
15. A.C. Gallagher, D. Batra, and D. Parikh. Inference for order reduction in markov random fields. In CVPR, pages
1857 – 1864, 2011.
16. G. Gallo and B. Simeone. On the supermodular knapsack problem. Mathematical Programming: Series A and
B, 45(2):295–309, 1989.
17. M. Grötschel, L. Lovász, and A. Schrijver. The ellipsoid method and its consequences in combinatorial optimization. Combinatorica, 1(2):169–197, 1981.
18. P. L. Hammer. Some network flow problems solved with pseudo-boolean programming. Operations Research,
13:388–399, 1965.
19. G. Hansel. Sur le nombre des fonctions booleennes monotones de n variables. C.R. Acad. Sci. Paris, 1966.
25
26
Srikumar Ramalingam, Chris Russell, L’ubor Ladický, and Philip H.S. Torr
20. H. Ishikawa. Exact optimization for Markov random fields with convex priors. PAMI, 25:1333–1336, 2003.
21. H. Ishikawa. Transformation of general binary mrf minimization to the first-order case. PAMI, 13(6):1234 –
1249, 2011.
22. S. Iwata. A fully combinatorial algorithm for submodular function minimization. Journal of Combinatorial
Theory, Series B, 84(2):203–212, 2002.
23. S. Iwata. A faster scaling algorithm for minimizing submodular functions. SIAM J. Computing, 2337:1–8, 2003.
24. S. Iwata, L. Fleischer, and S. Fujishige. A combinatorial strongly polynomial algorithm for minimizing submodular functions. J. ACM, 48(4):761–777, 2001.
25. D. Kleitman. On dedekind’s problem: The number of monotone boolean functions. Proc. Amer. Math Soc.,
1969.
26. P. Kohli, M. P. Kumar, and P. H. S. Torr. P3 & beyond: Solving energies with higher order cliques. In CVPR,
pages 1 – 8, 2007.
27. V. Kolmogorov.
Minimizing a sum of submodular functions.
Discrete Applied Mathematics,
160(15):2246âĂŞ2258, 2012.
28. V. Kolmogorov and R. Zabih. What energy functions can be minimized via graph cuts? PAMI, 26(2), 2004.
29. A.D. Korshunov. The number of monotone boolean functions. Problemy Kibernet 38:5-108, 1981.
30. L. Ladicky, C. Russell, P. Kohli, and P. Torr. Graph cut based inference with co-occurrence statistics. In European
Conference on Computer Vision, pages 239–253. springer, 2010.
31. X. Lan, S. Roth, D. P. Huttenlocher, and M. J. Black. Efficient belief propagation with learned higher-order
Markov random fields. In ECCV, pages 269–282, 2006.
32. Y.T. Lee, A. Sidford, and S.C. Wong. A faster cutting plane method and its implications for combinatorial and
convex optimization. In In IEEE 56th Annual Symposium on Foundations of Computer Science, 2015.
33. L. Lovasz. Submodular functions and convexity. Mathematical Programming - The State of the Art, pages
235–257, 1983.
34. M. Narasimhan and J.A. Bilmes. A submodular-supermodular procedure with applications to discriminative
structure learning. In Uncertainty in Artificial Intelligence, pages 404–412, 2005.
35. R. Nishihara, S. Jegelka, and M.I. Jordan. On the convergence rate of decomposable submodular function
minimization. In NIPS, 2014.
36. J.B. Orlin. A faster strongly polynomial time algorithm for submodular function minimization. Mathematical
Programming, 118(2):237–251, 2009.
37. S. Promislow and V. Young. Supermodular functions on finite lattices. Order 22(4), 2005.
38. M. Queyranne. A combinatorial algorithm for minimizing symmetric submodular functions. SODA, pages
98–101, 1995.
39. S. Ramalingam, P. Kohli, K. Alahari, and P.H.S. Torr. Exact inference in multi-label crfs with higher order
cliques. In CVPR, pages 1–8, 2008.
40. J.M.W. Rhys. A selection problem of shared fixed costs and network flows. Management Science, pages 200 –
207, 1970.
41. I. G. Rosenberg. Reduction of bivalent maximization to the quadratic case. Cahiers du Centre d’Etudes de
Recherche Operationnelle, 17:71–74, 1975.
42. C. Rother, P. Kohli, W. Feng, and J. Jia. Minimizing sparse higher order energy functions of discrete variables.
In CVPR, 2009.
43. C. Rother, V. Kolmogorov, V. Lempitsky, and M. Szummer. Optimizing binary MRFs via extended roof duality.
In CVPR, 2007.
44. D. Schlesinger and B. Flach. Transforming an arbitrary minsum problem into a binary one. Technical Report
TUD-FI06-01, Dresden University of Technology, 2006.
45. A. Schrijver. A combinatorial algorithm minimizing submodular functions in strongly polynomial time. Journal
of Combinatorial Theory Series B, 80(2):346–355, 2000.
46. L. Trevisan G.B. Sorkin, M. Sudan, and D.P. Williamson. Gadgets, approximation, and linear programming.
SIAM Journal of Computing, 29(6):2074âĂŞ2097, 2000.
47. Alan Yuille, Anand Rangarajan, and A. L. Yuille. The concave-convex procedure (cccp). In Advances in Neural
Information Processing Systems 14. MIT Press, 2002.
48. B. Zalesky.
Efficient determination of gibbs estimators with submodular energy functions.
http://arxiv.org/abs/math/0304041v1, 2003.
Efficient Minimization of Higher Order Submodular Functions using Monotonic Boolean Functions
49. S. Zivny, D.A.Cohen, and P.G. Jeavons. The expressive power of binary submodular functions. Discrete Applied
Mathematics, 157(15):3347–3358, 2009.
50. S. Zivny and P.G. Jeavons. Classes of submodular constraints expressible by graph cuts. Proceedings of CP,
pages 112–127, 2008.
51. S. Zivny and P.G. Jeavons. Which submodular functions are expressible using binary submodular functions?
Oxford University Computing Laboratory Researc Report CS-RR-08-08, 2008.
52. S. Zivny and P.G. Jeavons. Classes of submodular constraints expressible by graph cuts. Constraints, 15(3):430–
452, 2010.
27
| 8cs.DS
|
Practical optimization of Steiner Trees via the cavity method
Alfredo Braunstein∗
DISAT, Politecnico di Torino, Corso Duca Degli Abruzzi 24, Torino, Italy
Human Genetics Foundation, Via Nizza 52, Torino, Italy and
Collegio Carlo Alberto, Via Real Collegio 30, Moncalieri, Italy
Anna Muntoni†
arXiv:1607.03866v1 [cs.DS] 13 Jul 2016
DISAT, Politecnico di Torino, Corso Duca Degli Abruzzi 24, Torino, Italy
The optimization version of the cavity method for single instances, called Max-Sum, has
been applied in the past to the Minimum Steiner Tree Problem on Graphs and variants.
Max-Sum has been shown experimentally to give asymptotically optimal results on certain
types of weighted random graphs, and to give good solutions in short computation times for
some types of real networks. However, the hypotheses behind the formulation and the cavity
method itself limit substantially the class of instances on which the approach gives good
results (or even converges). Moreover, in the standard model formulation, the diameter of
the tree solution is limited by a predefined bound, that affects both computation time and
convergence properties. In this work we describe two main enhancements to the Max-Sum
equations to be able to cope with optimization of real-world instances. First, we develop
an alternative “flat” model formulation, that allows to reduce substantially the relevant
configuration space, making the approach feasible on instances with large solution diameter,
in particular when the number of terminal nodes is small. Second, we propose an integration
between Max-Sum and three greedy heuristics. This integration allows to transform MaxSum into a highly competitive self-contained algorithm, in which a feasible solution is given
at each step of the iterative procedure. Part of this development participated on the 2014
DIMACS challenge on Steiner Problems, and we report the results here. The performance
on the challenge of the proposed approach was highly satisfactory: it maintained a small
gap to the best bound in most cases, and obtained best results on several instances on two
different categories. We also present several improvements with respect to the version of the
algorithm that participated in the competition, including new best solutions for some of the
instances of the challenge.
2
∗
†
[email protected]
[email protected]
3
CONTENTS
I. Introduction
A. Definition
II. The model
4
6
7
A. Variables
7
B. Constraints
7
III. Belief Propagation
8
A. Belief Propagation equations
8
B. The β → ∞ limit: Max-Sum equations
10
C. Reinforcement
11
D. Efficient computation of the equations
12
E. The Flat Model
13
IV. A Belief Propagation-Inspired Heuristics
A. Pruned Trees
17
17
1. Minimum Spanning Tree and Shortest Path Tree
18
2. Goemans-Williamson heuristics
19
B. Labeling
20
C. Rooting
21
1. SPG rooting
21
2. PCSPG rooting
21
D. Reinforcement
22
E. Depth
22
F. Running schemes
23
V. Results for Scale-free and Grid-like graphs
23
A. Performances of Max Sum against heuristics
24
B. Scale-free networks
26
4
1. SPG results
26
C. Grid graphs
28
1. 2d lattices
28
2. 3d grid-graphs
28
VI. Results of the DIMACS challenge and StenLib instances
A. SPG results
30
31
1. Preprocessing
31
2. Results for the D-increasing scheme
31
3. Results for the D-bounded scheme
32
B. PCSPG results
33
C. Non-reweighting scheme results
34
VII. Conclusion
34
Acknowledgments
35
A. Results for Scale free Graphs
36
B. Results for Grid Graphs
37
C. Dimacs and Stenlib results
40
References
48
I.
INTRODUCTION
The cavity method has been developed for the study of disordered systems in statistical physics and has been employed in recent years for the design of a family of algorithmic
techniques for combinatorial optimization. Among these, it has been shown that series
of network optimization problems including the Minimum Steiner Tree (MStT) problem
and variants, e.g. the Prize-Collecting Steiner Problem on Graphs (PCSPG), can successfully be described by a model with local constraints and solved (at least on certain
5
families of instances) with a variant of the cavity method for optimization, specifically the
reinforced Max-Sum algorithm. The MStT has applications in many areas ranging from
biology (e.g., finding protein associations in cell signaling [1–3]) to network technologies
(e.g., finding optimal ways to deploy fiber optic and heating networks for households and
industries [4]).
We will describe the main problem and variants in the rest of this section. On Section
II we will describe the model we will use to represent the optimization problems. The
formulation presented here is based on edge variables as opposed to vertex variables in
the “pointer-depth” formulation used in [2, 5–7]. On both representations, the diameter
of the solution tree is bounded by a parameter that both affects the computation time
of each iteration and is intimately related to convergence properties. However, specially
on 2-dimensional and 3-dimensional instances, the diameter of the solutions grows much
faster with the system size with respect to random graphs (as a power of the system size
instead of the logarithm of it). We propose an alternative formulation (the flat model)
in Subsection III E that overcomes the a priori bound on the diameter, and that is made
possible by the new edge-variables formulation.
In real-world instances, the de-correlation hypotheses behind the cavity equations
are seldom accurate. The reinforced Max-Sum equations are partially able to overcome
this inaccuracy by using intermediate results to slowly bootstrap the system into one
with large enough external fields on which the equations are more accurate. However,
in such a slow procedure no solution is found on intermediate steps before the final
convergence, if ever, arrives. This is a big drawback for real-world optimization, as
alternative algorithms, either based on local search heuristics or linear programming, can
on the contrary provide reasonably good solutions on intermediate steps. In this work
we study the practical problem of obtaining solutions within constricted time limits, by
implementing fast heuristic methods to construct good trees from partial results while
the main computation is in course. This will be described in Section IV.
We will present results on low-dimensional and scale-free graphs on Section V, and
finally describe the results of the 2014 DIMACS Challenge on Section VI, in which
6
the algorithm developed in a preliminary version of this work participated. In order
to participate on the challenge, a strategy to determine the various parameters of the
algorithm without human intervention was devised in order to make the optimization
algorithm fully automatic as per the challenge rules. We consider the results obtained
on the challenge to be very promising, specially considering the fact that competing
algorithms were based on industrial-grade commercial solver libraries while our approach
can be implemented in a few hundred lines of plain C code.
A.
Definition
In this section we will define the Rooted Steiner Tree Problem (RSTP) on graphs.
Given an undirected graph G =
The Rooted Prize-collecting Steiner tree problem.
(V, E) with positive real weights wij > 0 for (i, j) ∈ E associated with edges (wij could
be different from wji ), non-negative real prizes ci associated with i ∈ V and a single root
vertex r ∈ V , we consider the problem of finding a directed tree T = (VT ⊂ V, ET ⊂ E)
rooted at r that minimizes the cost or energy function:
H (T ) =
X
wij +
X
ci
(1)
i∈V \VT
(i,j)∈ET
That is, we seek
T ? = arg min H (T )
T
(2)
Without loss of generality, we can assume the minimum in (2) to be unique, by
eventually adding negligibly small random noise to weights wij . Nodes with ci > 0
are called profitable vertices, or sometimes generalized terminals (or even terminals) by
extension of the classical Steiner Problem on graphs where subset of terminal nodes must
be included in the tree.
The following problems can be trivially (polynomially) mapped into RSTP: Prizecollecting Steiner Tree Problem (PCSPG), Steiner Tree Problem on graphs (SPG), Group
Steiner Problem (GSPG), Minimum Weighted Steiner Trees (MWM). We will discuss
some specific mapping issues in Section IV C.
7
II.
THE MODEL
A.
Variables
Consider a rooted (directed) tree T , a feasible solution of the RSTP defined in Subsection I A, in which the maximum allowed distance between any leaf and the root is
parametrized by D (we say that the tree is D−bounded). From T , we will define for
each oriented (i, j) ∈ E an integer variable dij ∈ {−D, . . . , D} as follows. For each
directed edge (i, j) in ET (directed edges in ET will be assumed to point to r), dij will
be the distance (in hops), or depth, from i to r along the tree. That is, dij is the length
of the unique simple path v0 = i, v1 = j, . . . , vdij = r in T from i to r. For each edge
(i, j) such that (j, i) ∈ ET , define dij = −dji . The sign of dij will define unambiguously
the orientation of edge (i, j) in the tree with respect to the root r. For edges such that
both (i, j) , (j, i) ∈
/ ET , we define dij = 0. The vector d = {dij : (i, j) ∈ E} so defined
clearly satisfies an antisymmetric condition dij = −dji for each (i, j) ∈ E. Such d will
be henceforth called a representation of T . It is easy to verify that the mapping T 7→ d
is one-to-one: the inverse mapping is d 7→ (Vd , Ed ) with Vd = {i ∈ V : ∃k ∈ V : dki 6= 0}
and Ed = {(i, j) : dij 6= 0}.
B.
Constraints
We would like to switch to an optimization problem on the vector d that mirrors the
RSTP one. In order to ensure that the (Vd , Ed ) are trees, we will need to impose rigid
constraints on the vector d besides the antisymmetric condition. This will be implemented as a family of local constraints on sub-vectors of variables di = {dji : j ∈ V (i)}
where the symbol V (i) will denote the neighborhood of i, V (i) = {j ∈ V : (i, j) ∈ E}.
For each node i we define a proper compatibility function ψi (di ), that is equal to one
if the constraints are satisfied and zero otherwise. As the sign of dij represents the orientation of edges along the tree (with dij > 0 if j is closer to r than i), two mutually
excluding situations can occur in the neighbor of i. Either i does not belong to T , and so
8
dij = 0 for each j ∈ V (i), or else there exists exactly one neighbor k such that dik > 0,
and for the remaining neighbors l ∈ V (i) \ k, either dli = 0 or dli = dik + 1. The root
node r is special, as there is no neighbor closer to r than itself; i.e. for each k ∈ V (r),
dkr is either 0 or 1. Symbolically, admissible configurations of di can be encoded by the
nonzero arguments of the following compatibility function:
X X
Y
Y
δd ,−d
δdki ,0 +
(δdli ,d+1 + δdli ,0 ) for i 6= r
ψi (di ) =
ki
d>0 k∈V (i)
k∈V (i)
ψr (dr ) =
Y
(3)
l∈V (i)\k
(δdkr ,1 + δdkr ,0 )
(4)
k∈V (r)
where δ denotes the Kronecker delta. Once ψi are defined, a cost function H (we
will use the same symbols as the one for trees) can be defined for variables d such that
vectors d with finite H(d) represent some tree T , and in such case, H (T ) = H (d):
Q
∞
if i ψi (di ) = 0
H (d) =
o
n
Q
P c I [d ≡ 0] + P
w
I
[d
>
0]
if i ψi (di ) = 1
ij
ij
i
i
j∈V (i)
i
where I is the indicator function which takes value 1 if the argument is true or 0
otherwise. Then we can substitute (2) with
d? = arg min H (d)
d
III.
A.
BELIEF PROPAGATION
Belief Propagation equations
The Belief Propagation (BP) equations (or Replica-Symmetric Cavity Equations in
Statistical Physics), are a set of equations to describe approximately a Boltzmann-Gibbs
distribution in terms of some of its marginals. The Boltzmann distribution is in this case
a probability measure on the space of all Steiner Trees on G (represented by vectors d),
where lower cost trees have larger probability in a way that is dependent on a positive
value β called the inverse temperature. The corresponding Boltzmann distribution is in
9
this case:
1 −βH(d)
e
Z
P
P
1 Y
=
ψi (di ) e−β i (ci I[di ≡0]+ j∈V (i) wij I[dij >0])
Z
P (d) =
(5)
(6)
i
where the β-dependent constant Z (called the partition function) can be obtained by
the normalization condition of the probability measure P . When β → ∞, the distribution
concentrates on the optimal tree(s). The marginal function P (dij ) (defined with a slight
abuse of notation using the same symbol P ) consists in the quantity
P (dij ) =
X
P d0 δdij ,d0
(7)
ij
d0
Calculating marginals is generally hard in computational terms but extremely useful;
for example, in the β → ∞ limit, knowing the value of a marginal gives crucial information on the state of variable dij in the optimal configuration(s). It is easy to see that this
calculation can be used recursively to compute the optimum itself.
The BP equations are a set of fixed-point equations for modified or cavity marginal
functions. They can be shown to become asymptotically exact for some models on
certain types of random graphs in the limit |V | → ∞. On single instances, the Belief
Propagation equations are widely employed for inference in various contexts, including
telecommunication and visual stereo recognition. The equations are exact on any acyclic
graph and correspond to a dynamic programming solution for the corresponding inference
problem. We will not enter here into details on the nature of the approximation behind
the equations; we refer the interested reader to [8].
The cavity messages for our model are mij : {−D, . . . , D} 7→ [0, 1], also called messages, and its general expression [8] can be written for this particular model as follows:
mij (dij ) ∝
X
di\j
ψi (di ) e−β (ci I[di ≡0]+
P
k∈V (i)
wik I[dik >0])
Y
mki (dki )
(8)
k∈V (i)\j
where di\j = {dki : k ∈ V (i) \ j}. The proportional sign ∝ above hides a normalization constant that can be computed a posteriori after the computation of the rest of the
right-hand side of (8) for all values of dij . Notice that the computation of this constant
10
is fundamentally easier than the computation of the partition function Z in (5), as the
corresponding sum involves only 2D + 1 terms rather than an exponential number. On
a fixed point, an approximation for the marginals P (dij ) is given by
P (dij ) ∝ mij (dij ) mji (−dij )
(9)
The equations are normally employed as follows: a numerical solution for the fixedpoint equations (8) is sought by iteration from a random initial condition, and, on the
fixed point, (9) is applied to obtain (approximated) marginals.
B.
The β → ∞ limit: Max-Sum equations
As we are interested in the optimization problem, we will take the β → ∞ limit of
(5)-(6). As standard (see e.g. [8]), one performs a change of variables into cavity fields
hij (dij ) =
1
β
log mij (dij ) and local fields Hij (dij ) =
1
β
log P (dij ). For β → ∞, local
fields Hij give very valuable information about locally restricted optima:
Hij (dij ) =
min0
d0 :ψi (d )≡1
H d0 −
min0
d0 :ψi (d )≡1
d0ij =dij
H d0
from which a global optimum can be easily computed. We substitute the change of
variables into (8) − (9), to obtain in the β → ∞ limit the Max-Sum (MS) equations:
hij (dij ) = max
− ci I [di ≡ 0] +
di\j
ψi (di )=1
−
X
wik I [dik > 0] +
k∈V (i)
Hij (dij ) = hij (dij ) + hji (−dij ) + C
X
k∈V (i)\j
0
hki (dki )
+C
(10)
(11)
where C, C 0 are additive constants (i.e., that do not depend on dij ) that can be computed after the rest of the right-hand side, and ensure that maxd hij (d) = maxd Hij (d) =
0; such condition corresponds to the normalization constraint on messages mij and
marginals Pij for finite β. For shortness, we will drop from now on the additive constants
in the equations.
11
As with BP, MS equations are normally solved numerically by repeated iteration of
(10). On a fixed point, (11) is computed and then, for each Hij , we perform the maximum
over dij
d?ij = arg max Hij (dij ) .
(12)
If this maximum in (12) is unique, a tree can be reconstructed by using the inverse
mapping defined in Section II. Variables d?ij are called decisional variables.
Notably, it can be shown [5, 7] that MS equations are exact on arbitrary graphs for
the Spanning Tree problem, i.e. when a positive large enough constant c is associated
with each node i ∈ V . If a fixed point is found and the maximum d?ij are non degenerate
(i.e. the maximums are unique), then they form the representation of the (then unique)
Minimum Spanning Tree. The degeneracy requirement can be relaxed by adding to edge
weights random noise terms rij , negligibly small with respect to wij .
C.
Reinforcement
For the Minimum Steiner Tree or Prize-collecting Steiner Tree Problem, iteration of
Equations (10)-(11) very seldom converge. Nevertheless, there is still valuable information in local fields before convergence. The following strategy can be applied [6, 7, 9]:
add a reinforcement term to (10), that progressively bootstraps the model into an easy
one in the direction of a feasible configuration.
The dynamical equations are:
X
ht+1
(d
)
=
max
−c
I
[d
≡
0]
−
wik I [dik > 0] +
ij
i
i
ij
di\j
ψi (di )=1
k∈V (i)
t+1
t
Hijt+1 (dij ) = ht+1
ij (dij ) + hji (−dij ) + γt Hij (dij )
X
k∈V (i)\j
htki (dki ) (13)
(14)
where γt is called the reinforcement factor. The factor γt is increased in a linear regime
γt = γ1 t, where the parameter γ1 is typically small, for instance taken in the interval
[10−5 , 10−3 ]. Rather than waiting for convergence of the equations, decisional variables
d?ij are computed during iterations, and the process is stopped when decisional variables
12
are repeated a predefined number of times (e.g. 50–100). The number of iteration needed
is proportional to γ1−1 , as it is observed that the process generally stops when γt ' 1. For
best results, γ1 should be small (so the bootstrapping procedure is slow) and it is crucial
to have an extremely efficient computation of the 2D |E| values on the left of (13). As
we will see, these can be computed in time Θ (D |E|).
D.
Efficient computation of the equations
Equation (13) requires the computation of a maximum over a set of (2D + 1)|V (i)|
elements, that quickly becomes prohibitive even for modest values of D and |V (i)|.
Fortunately, the computation can be performed in amortized O (2D + 1) time as follows.
First, let us consider the root node r. In this case the compatibility function is particularly
simple: neighboring edges can be present with dkr = 1 or absent with dkr = 0 in the
tree. The equations consequently simplify enormously:
ht+1
rj (drj )
=
P
k∈V (r)\j
max htkr (1) , htkr (0)
−∞
for drj = −1, 0
(15)
for drj 6= −1, 0
Let us consider any i 6= r. Suppose first dij > 0. Then, in order for ψi to be 1, it
must be that dki = dij + 1 or dki = 0 for each k ∈ V (i) \ j:
ht+1
ij (dij ) = −wij +
X
max htki (dij + 1) , htki (0)
(16)
max htki (dij + 1) , htki (0) +
(17)
k∈V (i)\j
= −wij +
X
k∈V (i)
− max htji (dij + 1) , htji (0)
(18)
The above equation can be clearly computed for all j ∈ V (i) in Θ (D |V (i)|) operations (first computing the sum, then subtracting one term for each neighbor). For i 6= r,
dij = −1 is forbidden, so ht+1
ij (−1) = −∞. Suppose then dij < −1. In this case there
must exist k ∈ V (i) \ j such that dik + 1 = dji , and all other l ∈ V (i) \ {k, j}, dli = −dij
13
or dli = 0. In symbols,
ht+1
ij (dij ) = max
k∈V (i)\j
=
X
t
hki (dij + 1) − wik +
X
max htli (0) , htli (−dij )
(19)
l∈V (i)\k,j
max htli (0) , htli (−dij ) − max htji (0) , htji (−dij ) +
l∈V (i)
+ max
k∈V (i)\j
t
hki (dij + 1) − wik − max htki (0) , htki (−dij )
(20)
Note that A = maxk∈V (i)\j Ak , where Ak = htki (dij + 1)−wik −max htki (0) , htki (−dij )
in (20) can be computed in |V (i)| operations for all j ∈ V (i). First, in Θ (|V (i)|) operations, the first two maxima Ak1 , Ak2 can be computed along with k1 , the position of
the first maximum. Then, for j ∈ V (i), if j = k1 , then A = Ak2 and for j 6= k1 , A = Ak1 .
Finally, the case dij = 0 is similar to the one with dij < −1 and can be computed by
reusing the computation above:
ht+1
ij (0) = max
X
htki (0) − ci , max ht+1
ij (d)
k∈V (i)\j
d<−1
(21)
In summary, (13) can be computed for all neighbors of a vertex i in a time proportional
to D |V (i)|, which gives a total time per iteration proportional to D |E|. As this is
obviously the same time requirement of (14), each iteration of the MS equations requires
a number of elementary operations proportional to D |E|.
E.
The Flat Model
For variants of the problem in which the hop-length from the root is unlimited, variables dij should be unbounded. Fortunately, if dij ∈ {−D, . . . , D}, a value of D = |V |
would be clearly sufficient. However, the computation of the equations scales in time
as D |E|; so it is generally not desirable to allow too large values of D. As we will see,
a slightly modified model allows to significantly reduce the needed depth bound D to
D = |K| where K is the set of generalized terminals, i.e. nodes with ci > 0. The idea is
to allow in the representation d of the tree, to have chains of edges with identical depth
d, i.e. v0 , . . . , vk with dv0 v1 = dv1 v2 = · · · = dvk−1 vk . We will allow this situation (the
14
flat rule, to differentiate it from the normal rule (3)) for a node vs only if: (i) vs is not
a terminal and (ii) vs has degree exactly two in the tree (i.e. no other neighbor besides
vs−1 and vs+1 ). These two conditions ensure that (optimal) configurations satisfying this
relaxed set of constraints represent trees; extra cycles with identical depth, containing
no terminal, can of course be present, but are suboptimal in terms of cost. In symbols,
we would use instead of the compatibility function in (3),
ψi0 (di ) = ψi (di ) + ψif lat (di )
X X X
ψif lat (di ) = δci ,0
δdki ,−d δdli ,d
d>0 k∈V (i) l∈V (i)\k
(22)
Y
δdij ,0
(23)
j∈V (i)\k,l
Two remarks are in order: the correspondence between a tree T and a representation
d is no more one-to-one: different vectors d represent the same tree T (because in a nonbranching path inside T , depth can either increase or not increase). Moreover, as we have
seen above, some allowed configurations d now do not represent any tree (because they
may have extraneous disconnected cycles). Nevertheless, such apparent inconsistencies
are not problematic. To see this, consider the following two statements:
1. Given d such that ψi0 (di ) ≡ 1, consider S = (Vd , Ed ) with Ed = {(i, j) : dij 6= 0},
Vd = {i ∈ V : ∃k ∈ V : dki 6= 0}. Then the graph S consists of the disjoint union
of a tree T and zero or more disconnected components that are simple cycles that
do not own any terminal.
Along any directed cycle, depth cannot decrease (otherwise along the path we
would find a vertex j with dij < djk which is forbidden both by the normal and
the flat rule); therefore only the flat rule could have been employed. Thus, there
cannot exist any branching vertex (nor terminal, including the root node r) in
the cycle and the depth should be constant. This cycle with constant depth will
form a separate connected component with no profitable vertices. The connected
component of r, on the other hand, cannot have cycles, and being it acyclic and
connected, it is a tree.
2. Any optimal tree has a flat representation d with D = |K| where K = {v 6= r : cv > 0}.
15
For unbounded D, consider d the most compact representation for an optimal tree
T , i.e. no depth increases in any non-branching, non-terminal vertex on the tree.
Consider the unique simple path (r = v0 , . . . , vk = l) from the root to a leaf in T .
We will prove that dvi+1 ,vi ≤ 1+K−Ki for i = 0, . . . k−1, where Ki is the number of
terminals different from i in the subtree Ti rooted at i. As in an optimal tree, every
leaf should be a terminal, the proposition would be proved by considering i = k−1,
as dvk vk−1 ≤ 1 + K − 1 = K. Let us prove it by induction on i. The result is clearly
true for i = 0, as dv1 r = 1 ≤ 1 + K − Ki = 1. Suppose the result true for some
0 ≤ i < k − 1. There are two cases to consider: 1) dvi+2 vi+1 = dvi+1 vi . Then vi+1 is
non-branching and cvi+1 = 0 (otherwise the flat rule cannot be applied). But then
Ki+1 = Ki (because the tree Ti is just Ti+1 plus the edge (i, i + 1) and i + 1 is nonterminal). Clearly then dvi+2 vi+1 = dvi+1 vi ≤ 1 + K − Ki = 1 + K − Ki+1 . 2) Either
vi is terminal or vi is branching in T , i.e. vi has at least one children j different from
vi+1 in T . In both cases Ki+1 ≤ Ki −1: in first case, simply because vi is a terminal.
In the second one because the subtree Tj will necessarily have at least one leaf that
must be a terminal. Then dvi+2 vi+1 = 1 + dvi+1 vi ≤ 2 + K − Ki ≤ 1 + K − Ki+1 .
Thus by (1)-(2) we can deduce that the flat mode is not less convenient than the normal
model in terms of solutions, i.e.
min
d:ψi0 (d)≡1
H (d) ≤
min
H (d)
d:ψ i (d)≡1
and also that for D ≥ |K|, the flat model is complete, i.e.
min
d:ψi0 (d)≡1
H (d) = min H (T )
T tree
(24)
and d? = arg min H (d) is a representation of the optimal tree.
To give an example we plot in Fig. (1) a feasible solution of the RSTP for a small
graph in the branching and in the f lat representation. For sake of simplicity terminals
are displayed as squares, the root as a triangle, and numbers close to edges represent the
decisional variables different from zero. In the normal model, starting from leaves {0,1},
all hop-lengths decrease if we make a step in the direction of the root, in this case from
16
left to right. In the flat representation, proceeding from the root “2”, we first encounter
node “3” and, since it is a terminal, so the depth must increase. After that, being “4”
neither terminal nor a “branching node” and having degree exactly two within the tree,
the depth can remain equal to 2. Node “5” is not a terminal but has degree 3 and so we
must increase the depth on children branches. Nodes {6, 7, 8} are not terminals and in
the cycle of Fig. (1) have degree 2; thus this disconnected component can be present in
a feasible configuration for the flat model but is energetically inconvenient and it will be
discarded by minimization of the energy.
0
4
1
0
9
2
4
5
3
3
1
4
2
6
8
3
1
9
3
5
2
8
4
2
6
2
2
3
1
2
2
7
7
Figure 1. Left: A feasible solution of the RSTP for a small graph of ten nodes in the normal
representation. Right: A feasible configuration for the flat representation.
Finally, the computation for the corresponding (13) for the flat model can be also
carried out in overall time proportional to D |E| per iteration. The MS equation is
ht+1
ij (dij ) = max
di\j :
ψi0 (di )=1
− ci I [di ≡ 0] −
X
wik I [dik > 0] +
k∈V (i)
X
htki (dki )
(25)
k∈V (i)\j
n
o
= max Mij (dij ) , Mijf lat (dij )
(26)
where
X
Mij (dij ) = max
−ci I [di ≡ 0] −
wik I [dik > 0] +
di\j :
k∈V (i)
ψi (di )=1
Mijf lat (dij ) = −wij I [dij > 0] +
max
X
ψif lat (di )=1
k∈V (i)\j
di\j :
X
htki (dki )
k∈V (i)\j
t
hki (dki ) − wik I [dik > 0]
(27)
(28)
17
The term M corresponds to the MS equation for the normal model that can be
computed as described in Subsection III D. For i such that ci = 0, the term M f lat will
be computed as follows. For dij 6= 0
n
o
P
t
M f lat (dij ) = −wij + maxk∈V (i)\j ht (dij ) +
l∈V (i)\j,k hli (0)
ij
ki
n
o
P
M f lat (d ) = max
t (d ) − w +
t (0)
h
h
ij
ij
ik
k∈V
(i)\j
l∈V (i)\j,k li
ij
ki
for dij > 0
for dij < 0
(29)
The restricted maxima in (29) for all neighbors j ∈ V (i) can be computed in time
proportional to |V (i)| as in (20). For dij = 0 instead,
Mijf lat (0)
= max
max
d>0 k, l∈V (i)\j
k6=l
htki (d)
+
htli (−d)
− wil +
X
htmi (0)
(30)
m∈V (i)\j,k,l
For each d > 0, the internal max can be computed for all j ∈ V (i) again in time
proportional |V (i)|, in a way similar to the one described in (20) but this time recording
the first three maximums of the quantities in braces instead of first two. In summary, also
using the flat rule the number of needed elementary operations per iteration is Θ (D |E|).
IV.
A BELIEF PROPAGATION-INSPIRED HEURISTICS
A.
Pruned Trees
One drawback of the MS heuristics with respect to local search based ones consists
in the fact that until convergence of the algorithm, decisional variables are in a state
of inconsistency, incompatible with a single feasible solution of the problem. A way of
obtaining feasible trees after a few iterations, but long before convergence, is to use simple
heuristics for the SPG and for the PCSPG that use information carried by MS fields
instead of the original edge and node weights. This procedure will be carefully designed so
that in case of convergence of the MS equations the outcome of the heuristics is identical
to the solution given by the decisional variables of the MS solution. These strategies
become helpful and decisive when we deal with limited available time, especially for very
large instances, and of course in cases in which plain MS equations do not converge.
18
To give an example we plot in Fig. (2) the outcomes of the heuristics (labelled as “N”,
“J” and “W” which are described in Subsection IV B) and MS, before convergence of the
algorithm, for one of the instances of 11th DIMACS Implementation Challenge, the cc312nu instance. Points represent feasible solutions to the PCSPG at any time while we
trace the minimum of the energy provided by each variant using a dashed line. At time
zero we plot the energy of the trivial solution in which the tree only contains its root. We
see that after few hundreds of seconds the heuristics can give a more energetically favored
solution which improves in time until it coincides with the MS energy at convergence.
cc3−12nu
cc3−12nu
120
120
Max Sum
Heuristics W
115
115
110
110
Energy
Energy
Max Sum
Heuristics N
Heuristics J
105
100
105
100
95
95
0
20
40
60
80
100
Time [s]
120
140
160
180
0
100
200
300
400
500
600
Time [s]
Figure 2. Energy of the solution to the PCSPG as function of time on instance cc3-12nu
1.
Minimum Spanning Tree and Shortest Path Tree
t of the graph G = (V, E)
At each iteration t we compute a set of modified weights wij
which are functions of the local fields Hijt or cavity fields htij . We build a feasible Steiner
tree on this re-weighted graph as follows. First, we compute a spanning tree TH (VH , EH )
(using either (a) Minimum Spanning Tree (MST) or (b) the Shortest Path Tree (SPT) by
Prim or Dijkstra’s algorithm respectively). Afterwards, we apply the following pruning
procedure: starting from each leaf node i ∈ VH with V (i) = {j}, we check whether
wij > ci . In this case adding node i to the solution is energetically unfavorable and we
delete i and (i, j) from TH . We recursively repeat this procedure until no such leaf is
19
t will be computed in two alternative ways:
found. Weights wij
1. Reweighting edges. A first way uses only information contained in cavity fields
t = max
t
Hijt : we set wij
d6=0 Hij (d). This quantity will be strictly positive if the
decisional variable d?ij = 0 and will be zero if d?ij 6= 0.
2. Reweighting nodes. A second way of assigning auxiliary costs to edges takes into
account the prediction of MS regarding the presence of each vertex i in the solution.
From the equations, a decisional variable can be assigned to the presence of node
i at depth d ≥ 0 by setting
X
t
t
t
hi (d) = max hik (−d) +
max hli (d + 1) , hli (0)
k∈V (i)
l∈V (i)\k
X
hi (0) =
htki (0) − ci
for d > 0
(31)
k∈V (i)
We will thus force the presence of nodes i such that maxd>0 hi (d) > hi (0), by
adding a large prize C to edges connecting nodes not satisfying this property.
2.
Goemans-Williamson heuristics
For the PCSPG, in addition to the MST and the SPT, we implement the GoemansWilliamson (GW) algorithm. For the theoretical reasoning and a detailed description see
[10] and [11]. Here we briefly explain the main steps of the algorithm and how we modify
the weights and the prizes of the graph in order to include the (partial) MS result within
the heuristics.
The algorithm consists in two steps, the “growth” stage and the “pruning” stage. In
the first one we partition vertices in clusters that are merged and ignored during the
iterations until one significant cluster remains; the more a cluster contains profitable and
low cost connected nodes, the more the cluster will have the chance of being the final
one. The second stage finds a solution of minimum energy for the PCSPG within the
nodes of the final cluster.
20
Before applying the algorithm we modify prizes and weights in the following way. For
each node we compute hi = maxd hi (d) − hi (0) defined in (31). If hi is positive, node i is
present in the intermediate solution and so we increase the prize ci of a large constant C;
otherwise it keeps its original prize. In this way we favor those clusters containing nodes
with zero original prize but predicted by MS as Steiner nodes. Moreover, we modify
edges connecting nodes i : hi < 0 by adding C to the original weight so that clusters
whose members are not included in MS solution are penalized.
B.
Labeling
Several approaches and techniques have been proposed in Subsection IV A, most of
them were not present in the original algorithm that competed in DIMACS challenge.
Experiments are labelled depending on which model, heuristics and assignment of weights
and/or prizes have been used. All the features of the final algorithm correspond to the
following labels:
• “O”: this is the original version of the algorithm which competed in the challenge
and appear in the official results as “polito”. It consists in the Max-Sum algorithm
for the normal model joined to the MST; weights are modified as described in
Reweighting edges on the preceding page.
• “N”: we implement the MST heuristics in which weights are computed according
to Reweighting nodes on the previous page.
• “J”: here we use the SPT heuristics and weights are modified as in Reweighting
edges on the preceding page.
• “W”: the heuristics is the GW reported in Subsection IV A 2.
• “F”: we use the flat model described in Subsection III E. If no additional labels are
included, we refer to the MST heuristics with modified weights as in Reweighting
edges on the previous page.
21
Labels can be also combined, e.g. “F J” corresponds to the Flat model with Shortest
Path Tree heuristics.
C.
Rooting
The PCSPG and SPG variants of the problem are undirected and unrooted but the
formalism introduced in Section II needs to select a root node among the profitable or
terminal nodes, for the PCSPG and SPG variants respectively. It is always possible to
choose a random terminal for the SPG but there is no clear strategy for the PCSPG.
Moreover, the choice of a random terminal for the SPG using the normal model could lead
to a suboptimal solution for a limited depth D. Here we propose one rooting procedure
for each variant.
1.
SPG rooting
Since the running time per iteration is Θ (D|E|), it is convenient to select as root
the terminal that allows a representation with a small parameter D. A straightforward
heuristics consists then in finding the terminal for which the maximum hop distance
to other terminals is the smallest. This can be computed simply by an application of
Breadth-First Search for each root candidate, in time proportional to |K| |E| where |K|
is the number of terminals.
2.
PCSPG rooting
The procedure reported in this section is the same used in [7]. Add an extra root
node r and connect it to all profitable vertices with extra edges of identical very large
weight µ. The optimal PCSPG solution in this modified graph consists trivially in a
single node tree {r}, since the addition of anything else carries a cost µ that makes it
unprofitable. Nevertheless, the MS algorithm provides additional information besides the
non-informative optimal result. Consider the “second” optimum solution of the problem.
22
The root node r will be connected to one (and only one) vertex of the original graph,
since adding additional edge costs equal to µ will be cost-wise inconvenient. This vertex
can be identified by computing j ∗ ∈ arg maxj Hjr (1). Now clearly, nontrivial (unrooted)
solutions of the original problem are in one to one correspondence with r-rooted solutions
of the modified problem with only one neighbor of r (and the difference in cost is simply
µ). Unfortunately, the information contained in MS fields is insufficient to reconstruct
the full “second” optimal tree of the modified graph so a second run of the MS algorithm
in the original graph, using j ∗ as root is needed.
D.
Reinforcement
As explained in [2, 7, 9], the reinforcement procedure speeds up and aids convergence
of the MS algorithm but adds an additional parameter γ1 . Smaller values of γ1 lead
typically to better solutions at the cost of longer convergence times. We adopted the
following approach: we start the MS algorithm with a large value of γ1 ∼ 10−2 . Then we
iteratively halve γ1 and run again MS until γ1 ∼ 10−5 . We can stop the loop in γ1 if the
energy gap between the new solution and the old one is significantly small since reducing
again the parameter typically does generally not bring a significant improvement.
E.
Depth
The depth parameter D is fundamental in the algorithm described in this work. It
unequivocally delimits the space of the solutions since MS will provide D-bounded trees;
furthermore the computational time depends linearly on D. Choosing a small depth
reduces the running time but this can affect the quality of the solution: for the SPG if D
is not sufficiently large the connection of all terminals is impossible and in the PCSPG
some profitable nodes cannot be reached. Larger values of D implies a larger solution
space where the optimum may be better. Thus we need to guarantee that all profitable
vertices, or terminals, can be connected within the tree and then let the algorithm choose
the optimal subgraph.
23
The computation of the minimum value of the depth Dmin uses again a Breadth-First
Search procedure. We start searching from the predefined root and we save the distances,
in hops, of the shortest paths between the root and any profitable vertex. We then choose
as starting value of D, the maximum value among all these lengths. Notice that for the
flat representation Dmin can be taken equal to the number of profitable vertices.
F.
Running schemes
In the following we explain two operative procedures that will be used for different
experiments regarding the SPG and the PCSPG variants.
• D-increasing scheme. Once we have determined the proper value of Dmin as in
Subsection IV E we apply MS and the heuristics following the reinforcement scheme
described in Subsection IV D. The process is repeated for increasing values of D
until the predefined running time is over.
• D-bounded scheme. Here we run the algorithm using a unique value of D. Simulations stop either because the reinforcement scheme in Subsection IV D or the
available time is over.
V.
RESULTS FOR SCALE-FREE AND GRID-LIKE GRAPHS
In this Section we report the performances of our new developments for two classes
of graphs which model very well real-world networks, namely, the scale-free and the grid
graphs. We will show here that the introduction of heuristics in Section IV guarantees
feasible solutions after few iterations of MS even for loopy graphs and we will underline
the improvements carried by the flat model for particular instances. We show the results
of “N”, “J” and “W” (combined to the normal or to the f lat model) and we compare
them to the results obtained by the “old” heuristic, “O”; quantitatively we compute the
percentage gap between the energy given by algorithm x and algorithm y as:
Gap(x, y) =
x−y
· 100
y
(32)
24
where x represents one of our “new” enhancements and y the “O” algorithm; the more
the gap is negative the more x outperforms y.
Experiments were run on a Multi-Core AMD opteron 2600Mhz server, where most of
the cases we use the D-increasing scheme in Subsection IV F for a running time of 600
s; if different schemes are used they will be precised in each section.
Most relevant results for single-instance problems are reported in several tables in
Appendix A and B that have all the same structure. The first column contains the name
of the instances, the second one displays the best energy found by the best algorithm
that is reported in the third section. The remaining columns list the running time needed
by the best algorithm, the “O” result and the gap computed as in (32) between the best
energy and the “O” energy of the “old” algorithm. In some specific frameworks we also
propose averaged results over different realizations of the same graph to confirm evidences
suggested by single-instance energies.
Instances names give clear information about graph properties. The first letters identify the type of graph, i.e. can be either “SF” or “G” depending on we are dealing with
a scale-free network or a grid graph. They also contain the number of nodes and the
number of terminals that are followed by letters n and a respectively, while a suffix “-p”
is added for the PCSPG instances. Grid-graph names also reveal the nodes layout on the
graph. For instance, the graph G_100x100x2_a10 is a 3d grid graph of size 100x100x2
that contains 20000 nodes, 10 of which are terminals and we aim to solve a SPG.
A.
Performances of Max Sum against heuristics
To give an example of the efficiency of the MS-guided heuristics we create 5 grid graphs
of size 10x10x10 containing terminals in the range [10, 410] and 5 scale-free networks of
103 nodes and 100 terminals with edges in the interval [2991, 10879] on which we solve
the SPG. In Table I and II we report the energies achieved by “O” algorithm, MS and the
MST without the reweighting scheme introduced by “O”. Energies of the MST algorithm
are averaged over 10 realizations of each graph, instead, for both “O” and MS we run
25
the algorithms 10 times for each of the 10 instances with different initial conditions and
we collect the best energies among the 10 initialitations. Then these energies will be
averaged over the 10 instances. The fraction of successes over the 10×10 attempts is
reported in the left column of “MS conv”. In the right column we count how many times
MS converged at least one time over the 10 initializations, and we normalize the number
of successes with respect to the number of instances per graph.
Terminals “O”
MS
MS conv. MST
10
10.56 10.80 11/100 4/10 21.82
110
56.24 61.55
210
81.76 83.13 17/100 4/10 100.93
310
103.62 103.49 25/100 6/10 120.87
410
123.44 124.10 26/100 7/10 137.23
6/100 2/10 77.04
Table I. Average energies for “O”, MS, MST and MS convergences for grid graphs 10x10x10 as a
function of the number of terminals.
Edges “O”
MS
MS conv. MST
2991 34.09 33.89 30/100 4/10 40.72
4975 21.25 21.22 56/100 8/10 25.13
6951 16.53 16.92 59/100 7/10 18.52
8919 13.06 13.22 14/100 3/10 14.92
10879 10.85 10.76 30/100 3/10 12.49
Table II. Average energies for “O”, MS, MST and MS convergences for scale-free graphs of 1000
nodes and 100 terminals as a function of the number of edges.
Results suggest that MS very seldom converge on both families of networks with an
average fraction of success of 17/100 for grids and 38/100 for scale-free graphs. The
heuristic “O” always outputs a solution but sometimes is suboptimal in terms of energy
as we can notice from the comparison with the MS ones. Nevertheless these energies are
far below of the ones of the MST heuristic with original edge weights.
26
B.
Scale-free networks
The first class of networks on which we run our algorithms is the one of the scale-free
graphs. Instances are generated by the Barabási-Albert model also known as preferential
attachment scheme. Starting from a N0 nodes, new nodes are added, one at the time, to
the graph. Each new node is attached to m different vertices with a probability that is
proportional to the connectivity of the existing nodes.
In our experiments the number of nodes of the graphs is N = 104 for SPG instances,
while the initial set of nodes N0 has cardinality m; all graphs have weights distributed
according to the uniform distribution in the interval [0, 1].
1.
SPG results
To underline the efficiency of the flat model we perform experiments on sparse scalefree networks of 104 nodes and m = 3. We show in Fig. (3) the energies and the gaps
with respect to “O” of all the other algorithms as a function of the number of terminals
for three values of the depth D = {3, 5, 7} and unlimited running time. For very small
values of the depth (3 and 5) the “N” variant achieves gaps of the order of 10% which is
significantly larger than the ones obtained by all flat-model based heuristics (“F”, “NF”
and “JF”) that on average are equal to −15%, −10% and −6% for increasing values of
D. Only for D = 7 the “N” variant reaches the “O” performances while the “J” heuristic
gives negative gaps only for D = 3.
In addition we show the performances for very dense graphs where the parameter m
changes in the interval [10, 500] and we set a = 1000. As reported in Table III the “NF”
variant achieves, for single-instance runs, the best performance as long as the number
of edges is not very large. As m increases the “N” variant outperforms all the other
variants reaching gaps with respect to “O” that seem to increase, in absolute value, as
we increment the number of edges.
27
Scale Free n1e4 m3 D3 SPG
Scale Free n1e4 m3 D3 SPG
70
40
O
F
N
NF
J
JF
60
20
40
Gap %
Energy
50
Gap(F,O)
Gap (N,O)
Gap (NF,O)
Gap (J,O)
Gap(JF,O)
30
30
10
0
20
−10
10
−20
0
0
10
20
30
40
50
60
70
Number of terminals
80
90
100
110
0
10
20
Scale Free n1e4 m3 D5 SPG
30
40
50
60
70
Number of terminals
80
90
100
110
Scale Free n1e4 m3 D5 SPG
60
35
O
F
N
NF
J
JF
50
Gap(F,O)
Gap (N,O)
Gap (NF,O)
Gap (J,O)
Gap(JF,O)
30
25
20
40
Gap %
Energy
15
30
10
5
20
0
−5
10
−10
0
0
10
20
30
40
50
60
70
Number of terminals
80
90
100
110
−15
0
10
20
Scale Free n1e4 m3 D7 SPG
30
40
50
60
70
Number of terminals
80
90
100
110
Scale Free n1e4 m3 D7 SPG
60
12
O
F
N
NF
J
JF
50
Gap(F,O)
Gap (N,O)
Gap (NF,O)
Gap (J,O)
Gap(JF,O)
10
8
6
40
Gap %
Energy
4
30
20
2
0
−2
−4
10
−6
−8
0
0
10
20
30
40
50
60
70
Number of terminals
80
90
100
110
0
10
20
30
40
50
60
70
Number of terminals
80
90
100
110
Figure 3. Left: Energy of the solution for the SPG for scale-free graph as a function of the
number of terminals. Right: Energy gaps of the “F”, “N”, “NF”, “J” and “JF” heuristics with
respect to “O”. From up to the bottom the depth D increases.
28
C.
1.
Grid graphs
2d lattices
For these simulations we create 20 graph of 104 nodes lying in a 100x100 square lattice
and connected by edges whose weights distribution is uniform in the interval [0 1]. The
number of terminals a varies from 10 to 1000 and their prizes, for PCSPG variant only,
are picked uniformly from the interval [0 15] to ensure a non-trivial optimal solution
different from a single-node tree, the root, (in the case of very low prizes) and a spanning
tree (in the opposite case of very large prizes).
Regarding the SPG, best solutions for single-runs are obtained through the “F” variant
despite the gap with respect to the “O” algorithm is on average equals to -0.2 %. Besides,
for graph containing a large number of terminals, the most performing algorithm is the
“NF”. As shown in Table IV for the PCSPG the “F” and “O” variants achieve very close
results except for graph G_100x100_a10-p where the “W” heuristic reaches a gap of -5
%.
2.
3d grid-graphs
Instances are 100x100x2 grid-graphs whose links have weights distributed uniformly in
[0, 1] and whose terminals, for PCSPG only, have prizes in the range [0 3]. We investigate
different regimes depending on how many terminals are placed on the graph.
Regarding the SPG variant, as illustrated in Table V, the “F” variant more often
gives the best solution in the case of graphs with few terminals probably because here
the “depth” for the “F” algorithm is the number of terminals which is much smaller than
the parameter Dmin computed as in Section IV E.
This statement is confirmed by the statistical measure of the energies and the gaps
as a function of the number of terminals that is shown in Fig. (4). The “F” variant, a
part for a small positive gap in the range [20, 40], attains negative gaps that reach -10%
for instances with few terminals and a constant -1% for all the range a ∈ [80, 140].
29
Grid−graph 100x100x2 SPG
Grid−graph 100x100x2 SPG
400
40
O
F
N
NF
J
JF
350
300
30
20
Gap %
Energy
250
200
10
0
150
−10
100
0
Gap (F,O)
Gap (N,O)
Gap (NF,O)
Gap(J,O)
Gap(JF,O)
−20
50
0
20
40
60
80
Number of terminals
100
120
140
−30
0
20
40
60
80
Number of terminals
100
120
140
Figure 4. Left: Energy of the solution for the SPG for a grid-graph 100x100x2 as a function of
the number of terminals. Right: Energy gaps of the “F”, “N”, “NF”, “J” and “JF” heuristics with
respect to “O”.
For the PCSPG we investigate how energies changes when the number of terminals
is in the range [3000, 10000]. Single-instance results suggest that the “O” variant outperforms all the other up to a ∼ 4000 where we observe that “N” and “NF” algorithms
achieve best results with a gap that decreases for increasing a as reported in Table VI.
To fairly compare the results of the heuristics “O”, “N”, “J” and “W” in this regime, we
perform several simulations for each graph as a varies in [3000, 10000] and we compare
the averaged outcomes in Fig. (5). While the “J” variant attains energies as similar to
the “O” ones as a grows, the “N” and “W” heuristics increase their respective gaps as a
increases.
30
Grid−graph 100x100x2 PCSPG
Grid−graph 100x100x2 PCSPG
3200
0.5
O
N
J
W
3000
2800
0
−0.5
2400
Gap %
Energy
2600
2200
2000
−1
−1.5
1800
−2
Gap (N,O)
Gap(J,O)
Gap(W,O)
1600
1400
3000
4000
5000
6000
7000
Number of terminals
8000
9000
10000
−2.5
3000
4000
5000
6000
7000
Number of terminals
8000
9000
10000
Figure 5. Left: Energy of the solution for the PCSPG for a grid-graph 100x100x2 as a function
of the number of terminals. Right: Energy gaps of the “N”, “J” and “W” heuristics with respect
to “O”.
VI.
RESULTS OF THE DIMACS CHALLENGE AND STENLIB INSTANCES
Here we show the results of the MS algorithm combined to greedy heuristics for the
SPG and PCSPG benchmark chosen for DIMACS Challenge. These instances have been
selected to be particularly challenging (often the optimal solutions are not known) and
to be representative of the whole set available here: http://dimacs11.cs.princeton.
edu/competition.html. Additionally we report our results for a set of hard instances,
called PUC, where most of the optimal solutions are still not known.
For both SPG and PCSPG variants we will need to compare outcomes provided by
different algorithms or to measure the improvements of the new algorithm. According
to the rules of the official competition, we use as comparison metrics the Primal Bound
(PB) that is the best solution found in a fixed running time and the gap, computed as in
(32), between algorithms x and y that will be specified in the following sections. Absolute
values of the PB and the respectively gaps are collected in tables which are all displayed
in Appendix C. Notice that official results are often rounded to the first decimal place
and this may slightly affect the accuracy of the gaps.
As for the scale-free and grid graph, simulations were run on a Multi-Core AMD
31
opteron 2600Mhz server, which is slightly slower than the cluster on which the challenge
simulations were performed. Nevertheless, we compared the results obtained in the
challenge by our submitted code on our local server with the same time limit, and the gaps
of the primal bounds between “O” and “polito” are on average ±0.3% for all instances,
well inside the confidence interval for these simulations. Thus we are sufficiently confident
that new results can be also compared fairly with those of the challenge.
A.
SPG results
1.
Preprocessing
Regarding the SPG problem, a common approach is to apply some reduction or preprocessing techniques to instances. Preprocessing may significantly modify the original
graph but in a way that an optimal solution of the reduced graph can be easily mapped
into the minimum Steiner Tree of the original problem. The advantages of preprocessing consist in a speed up of the convergence and, often, a clear improvement of the
solutions. In this work we use the preprocessing feature of the nice package Bossa,
http://www.cs.princeton.edu/~rwerneck/bossa/.
We use the label “pre” for the results in which instances are preprocessed before the
application of the algorithm. In all cases in which it was measured, the running time
include the preprocessing time.
2.
Results for the D-increasing scheme
In the following we present the best Primal Bounds among all variants of the algorithm
labelled according to Section IV B. Experiments are performed as in the D-increasing
scheme in SectionIV F where the running time is set to be of 7200 s.
Results are displayed in Table VII. The second column reveals the best Primal Bound
found using our algorithms while the third one displays which model and/or heuristics
reach such results. In the third section we report the running time of our fastest performer
32
which is specified in brackets. The last two sections are devoted to the comparison
between the best new performance and the “polito” PB.
Generally we improved our old “polito” results with some significant gaps of -5.87 %
and -4.11 % for word666 and es10000fst01 instances respectively. Furthermore we notice
that, globally, all “N-like” options outperforms the other variants.
In Table VIII we compare our new results to the best results of the competition. For
each instance we show our PB, the best PB found in the challenge, where in brackets
we report the performer that attained such result, and the last column contains the gaps
between the two energies computed according to (32). Despite we are quite far from the
best known bound for some instances (for example alut2625, lin36 and lin37 whose gaps
are higher than 10 %), we generally approached the performance of the best challengers
of the competition. Moreover, we further improved the best known energies of some
solutions, like for the i640-341 and the “cc12 -like” instances; such primal bounds are
reported in bold letters.
Motivated by the performances of the “N-like” algorithm and by the improvements
obtained on the “cc12 -like” instances, we run “Npre” on the entire set of the PUC instances to which these graphs belong. Results shown if Table XIII and XIV reveals that
we achieve optimal performances since all gaps between Npre and the best known energies are smaller then 0.80 %. Moreover, we reach new best known bounds for cc10-2p,
cc11-2p and hc11p instances.
3.
Results for the D-bounded scheme
To underline the differences between the branching and the flat model, we run again
the algorithm using the D-bounded scheme in Section IV F on page 23 imposing the
depth equals to the number of terminals. The running time is set to 1200 s for these
experiments.
For each heuristic, “N”, “O” and “J”, we compute the time interval for which the same
heuristics combined with the flat model, “NF”, “F” and “JF”, provides a better solution
33
then the normal model; the same experiment is performed for the pre-processed instances.
In Fig. (6) we report such time thresholds.
1200
J
Jpre
N
Npre
O
Opre
1000
Time threshold [s]
800
600
400
200
I064ac
rt05
G106ac
rc09
w23c23
2r211c
wrp3-83
bipa2u
bipa2p
cc3-12u
cc3-12p
cc3-12n
cc12-2n
cc12-2u
hc12u
cc12-2p
lin37
hc12p
lin36
alut2625
es10000fst01
alue7080
world666
i640-341
fnl4461fst
i640-314
e18
i640-211
d18
0
Figure 6. Time threshold Flat vs Branching model
For several instances the time threshold is equal to zero, meaning that the branching
representation is always better than the flat one. A threshold of the order of hundreds
of seconds appears for several instances while for es1000fst101, a rectilinear graph, and
the G106ac, a “Vienna class” graph, the time interval in which the flat model (combined
to almost all the heuristics) provides the best solution is significantly large.
B.
PCSPG results
Here we show the performance of our new enhancements “F”, “J”, “N” and “W” for
the PCSPG variant. As for the SPG case we first collect our best primal bounds and we
compare them to the “polito” results of the challenge. Afterwards, we show the measure
of the gap with respect to the three best algorithms of the competition, “KTS”, “staynerd”
and “mozartballs”. We perform simulations using the operative scheme in Section IV F
for a running time of 7200 s.
As plotted in Table IX we find that the best performer is “F” that outputs the best
34
PB for 27 instances over 32 but the more significant improvements are given by “W”
for the K400-7, hands04 and handbd13 instances with a gap of -7 %, -10% and -27 %
respectively; the fastest performer instead is the “N” variant.
In Table X we show the results of our best performers compared to “KTS”, “staynerd”,
“mozartballs” and “polito” algorithms of the competition. Energies are now well comparable to the best known PB with an average gap of 0.1 % for 29 instances. Moreover,
we achieve the best known energy for hc10p, hc11u, hc12u and cc12-2nu by our new
algorithm; these results are reported in bold letters in Table X. Note that additionally,
for the instance hc12p, the best known energy is the one obtained by “polito”, i.e. the
“O” variant.
C.
Non-reweighting scheme results
In order to underline the improvements of the MS-based reweighting scheme for these
hard instances, we compare them to the PB provided by the heuristics (pruned Minimum
Spanning Tree, pruned Shortest Path Tree and Goemans-Williamson heuristics) for the
SPG and PCSPG instances with original weights on edges. As reported in Table XI
and XII energies in columns “MST”, “SPT” and “GW” (for Prize-collecting only) are far
from being comparable to the performances of the re-weighted scheme, labelled as “MS
reweighting”, and thus to the state-of-the-art algorithms of the challenge.
VII.
CONCLUSION
In this work, we presented several improvements for a message-passing approach to
several variants of the Steiner Tree Problem on graphs. A first improvement is the
incorporation of heuristics that are able to transform partial information coming from
an intermediate state of the messages before convergence into feasible solutions. This
may be useful when the computation time at disposal is short, but it is of particular
importance because it forces the algorithm to output solutions even in cases in which the
tree-like approximation is inaccurate and reinforced Max-Sum equations do not converge.
35
This resulted in the variant that we called “O” in this manuscript, that participated in
the 2014 DIMACS Challenge with encouraging results. We give a detailed report on the
outcome of the Challenge.
With respect to the results in the Challenge, we report here several further improvements, including two different heuristics, the “W”, “N” and “J” variants and their combinations. Many strategies have been explored and reported here, with the outcome that
the “N” variant is the best overall, but some of the other variants give advantages for
some instance types. During this development we derived an “edge variables” formulation, that allows to cope with a modified flat model that removes one impediment of
past approaches of this type; namely the need of a large maximum hop-distance D. This
results in the “F” variant of the algorithm, that give advantages in particular on structured networks including meshes and scale-free graphs. Moreover the “edge variables”
formulation presented here is also in principle able to accommodate other constraints,
such as degree ones, in a simple way.
ACKNOWLEDGMENTS
AB acknowledges support by Fondazione CRT under the initiative “La Ricerca dei
Talenti”. The authors acknowledge ERC grant No. 267915 for financial support in the
participation to the DIMACS challenge.
36
Appendix A: Results for Scale free Graphs
Instance
SF_n1e4_m10_a1e3
PB
Algo Time [s] “O” PB
115.92 N
Gap %
595.97
116.86
−0.80
SF_n1e4_m36_a1e3
33.54 NF
266.24
36.24
−7.45
SF_n1e4_m62_a1e3
19.28 NF
381.05
20.64
−6.59
SF_n1e4_m87_a1e3
14.21 NF
492.69
15.18
−6.36
SF_n1e4_m113_a1e3
11.37 NF
443.22
12.67
−10.26
SF_n1e4_m139_a1e3
9.14 NF
602.81
10.39
−12.00
SF_n1e4_m165_a1e3
8.89 N
487.00
8.92
−0.41
SF_n1e4_m191_a1e3
7.24 N
29.51
7.33
−1.24
SF_n1e4_m216_a1e3
6.80 N
325.07
6.90
−1.44
SF_n1e4_m242_a1e3
5.84 N
40.44
5.92
−1.48
SF_n1e4_m268_a1e3
5.67 N
454.76
5.75
−1.40
SF_n1e4_m294_a1e3
5.05 N
60.07
5.14
−1.79
SF_n1e4_m319_a1e3
4.58 N
55.55
4.76
−3.83
SF_n1e4_m345_a1e3
4.07 N
46.22
4.19
−3.05
SF_n1e4_m371_a1e3
4.05 N
49.51
4.29
−5.63
SF_n1e4_m397_a1e3
3.86 N
76.40
4.05
−4.69
SF_n1e4_m423_a1e3
3.44 N
106.89
3.74
−7.92
SF_n1e4_m448_a1e3
3.29 N
80.44
3.62
−9.01
SF_n1e4_m474_a1e3
3.14 N
96.80
3.42
−8.00
SF_n1e4_m500_a1e3
3.10 N
104.78
3.35
−7.37
Table III. Best Primal Bounds for SPG on Scale Free networks, fixed a
37
Appendix B: Results for Grid Graphs
Instance
PB
Algo Time [s] “O” PB
Gap %
−5.29
G_100x100_a10-p
50.36 W
13.977
53.18
G_100x100_a62-p
215.10 O
25.072
215.10
−
G_100x100_a114-p
305.92 O
8.24
305.92
−
G_100x100_a166-p
363.03 F
20.32
369.46
−1.74
G_100x100_a218-p
421.11 F
206.27
427.23
−1.43
G_100x100_a271-p
477.41 O
13.62
477.41
G_100x100_a323-p
532.13 F
34.91
538.51
G_100x100_a375-p
554.55 O
98.42
554.55
G_100x100_a427-p
581.97 F
168.98
586.18
G_100x100_a479-p
624.02 O
16.64
624.02
−
G_100x100_a531-p
654.29 O
16.22
654.29
−
G_100x100_a583-p
673.97 F
73.10
675.27
G_100x100_a635-p
725.31 O
12.44
725.31
−
G_100x100_a687-p
742.67 O
13.62
742.67
−
G_100x100_a739-p
765.91 F
247.15
770.46
−0.59
G_100x100_a792-p
788.36 F
261.86
795.58
−0.91
G_100x100_a844-p
817.39 O
14.45
817.39
−
G_100x100_a896-p
854.57 O
13.51
854.57
−
G_100x100_a948-p
872.88 O
11.95
872.88
−
G_100x100_a1000-p
890.94 O
10.96
890.94
−
−
−1.18
−
−0.72
−0.19
Table IV. Primal Bounds for the PCSPG on 2d grid graphs
38
Instance
PB
Algo Time [s] “O” PB
Gap %
G_100x100x2_a10
78.07 O
51.142
78.07
−
G_100x100x2_a62
203.13 O
14.114
203.13
−
G_100x100x2_a114
286.96 F
26.15
290.06
−1.07
G_100x100x2_a166
358.52 F
25.08
362.57
−1.12
G_100x100x2_a218
424.50 F
31.80
429.52
−1.17
G_100x100x2_a271
468.21 O
27.04
468.21
G_100x100x2_a323
495.39 F
57.39
503.96
−1.70
G_100x100x2_a375
552.07 F
32.98
556.24
−0.75
G_100x100x2_a427
578.32 F
39.57
586.57
−1.41
G_100x100x2_a479
608.99 F
55.25
609.44
−0.07
G_100x100x2_a531
635.29 F
43.12
644.63
−1.45
G_100x100x2_a583
662.13 F
44.29
665.76
−0.55
G_100x100x2_a635
692.11 F
47.99
697.51
−0.77
G_100x100x2_a687
716.63 F
50.44
719.79
−0.44
G_100x100x2_a739
755.78 O
30.24
755.78
G_100x100x2_a792
780.45 F
56.66
781.62
−0.15
G_100x100x2_a844
797.13 F
38.55
804.08
−0.86
G_100x100x2_a896
822.84 F
42.10
823.93
−0.13
G_100x100x2_a948
843.94 F
49.80
847.32
−0.40
−
−
Table V. Primal Bounds for the SPG on 3d grid graphs with few terminals
39
Instance
PB
Algo Time [s] “O” PB
Gap %
G_100x100x2_a1000-p
875.85 O
45.04
875.85
−
G_100x100x2_a1474-p
1057.83 O
46.75
1057.83
−
G_100x100x2_a1947-p
1234.77 O
43.86
1234.77
−
G_100x100x2_a2421-p
1395.98 O
46.19
1395.98
−
G_100x100x2_a2895-p
1554.88 O
47.62
1554.88
−
G_100x100x2_a3368-p
1697.28 O
46.85
1697.28
−
G_100x100x2_a3842-p
1824.59 N
34.68
1833.18
−0.47
G_100x100x2_a4316-p
1941.98 N
53.69
1955.73
−0.70
G_100x100x2_a4789-p
2036.87 NF
241.08
2056.08
−0.93
G_100x100x2_a5263-p
2122.72 NF
230.64
2152.41
−1.38
G_100x100x2_a5737-p
2229.83 NF
169.36
2264.99
−1.55
G_100x100x2_a6211-p
2339.31 N
40.47
2382.05
−1.79
G_100x100x2_a6684-p
2414.71 N
516.42
2457.86
−1.76
G_100x100x2_a7158-p
2524.14 N
52.56
2575.23
−1.98
G_100x100x2_a7632-p
2609.18 NF
243.35
2658.92
−1.87
G_100x100x2_a8105-p
2700.71 NF
456.12
2758.35
−2.09
G_100x100x2_a8579-p
2775.56 N
34.92
2834.97
−2.10
G_100x100x2_a9053-p
2863.76 NF
362.59
2919.05
−1.89
G_100x100x2_a9526-p
2953.47 N
211.22
3019.45
−2.19
G_100x100x2_a10000-p 3012.83 N
42.95
3074.09
−1.99
Table VI. Primal Bounds for the PCSPG on 3d grid graphs with many terminals
40
Appendix C: Dimacs and Stenlib results
Instance
New PB Algo
Time [s] “polito” PB Gap %
d18
223
All
2.97 (Fpre) 224
−0.45
e18
564
All
112.78 (NFpre) 565
−0.18
i640-211
11984
All
i640-314
35532
i640-341
851.79 (Opre) 11984
0.00
N NF
223.53 (N) 35532
0.00
32042
NFpre
2164.31 32047
−0.02
fnl4461fst
189367
Npre
3759.34 192847
−1.80
world666
122971
N
94.37 130516
−5.78
alue7080
66624
NF
alut2625
40183
Npre
es10000fst01 733237957 Npre
100.96 67847
−1.80
1042.46 41501
−3.18
1232.63 764631264
−4.11
0.68
lin36
64486
Npre
399.04 64052
lin37
112886
Npre
892.89 114001
−0.98
hc12p
236075
N Npre
3664.50 (N) 236042
0.01
hc12u
2269
Opre J Jpre N Npre
cc12-2p
121056
N
cc12-2u
1174
N Npre
cc12-2n
613
Npre
cc3-12n
111
All
cc3-12p
19003
N Npre
5750.94 (N) 18932
cc3-12u
185
All
2331.35 (N) 186
bipa2p
35285
Npre
bipa2u
338
All
2r211c
89000
Opre J Jpre N Npre
wrp3-83
8301057
N
w23c23
691
F Fpre JF JFpre NF NFpre
rc09
120450
Npre
rt05
58286
F
G106ac
40396063 N
7088.62 40346932
0.12
I064ac
188849475 J
4678.88 188479558
0.20
s5
25210
All
0.18
4883.88 (Npre) 2265
7077.52 121091
6330.21 (Npre) 1177
−0.03
−0.25
4425.05 615
−0.33
20.60 (Opre) 111
0.00
1927.36 35336
0.38
−0.54
−0.14
118.86 (NF) 339
−0.29
615.75 (Jpre) 90000
−1.11
0.00
1191.62 8301263
3541.60 (NF) 693
4176.39 122358
27.08 (F) 59147
−0.29
−1.56
−1.46
124.29 (NF) 25210
Table VII. Best PB for all the SPG instances and comparison with “polito” results
0.00
41
Instance
PB
Best PB of the Challenge Gap %
d18
223
223 (AB mozartballs scipjack staynerd)
0.00
e18
564
564 (mozartballs scipjack staynerd)
0.00
i640-211
11984
11984 (polito scipjack)
0.00
i640-314
35532
35532 (polito staynerd)
0.00
i640-341
32042
32047 (polito)
−0.02
fnl4461fst
189367
182527 (PUW)
3.75
world666
122971
122467 (mozartballs PUW scipjack staynerd)
0.41
alue7080
66624
62514 (PUW)
6.57
alut2625
40183
35471 (PUW)
13.28
716559567 (mozartballs)
2.33
es10000fst01 733237957
lin36
64486
55608 (PUW)
15.97
lin37
112886
99560 (PUW)
13.38
hc12p
236075
236042 (polito)
0.01
hc12u
2269
2262 (staynerd)
0.31
cc12-2p
121056
121091 (polito)
−0.03
cc12-2u
1174
1177 (polito)
−0.25
cc12-2n
613
615 (polito)
−0.33
cc3-12n
111
111 (mozartballs polito PUW scipjack staynerd)
0.00
cc3-12p
19003
18865 (mozartballs)
0.73
cc3-12u
185
185 (mozartballs PUW staynerd)
0.00
bipa2p
35285
35336 (polito)
−0.14
bipa2u
338
337 (mozartballs staynerd)
0.30
2r211c
89000
89000 (mozartballs PUW scipjack staynerd)
0.00
wrp3-83
8301057
8300906 (PUW)
0.00
w23c23
691
689 (mozartballs staynerd)
0.29
rc09
120450
111005 (mozartballs staynerd)
8.51
rt05
58286
51354 (PUW)
13.50
G106ac
40396063
36920936 (mozartballs staynerd)
9.41
I064ac
188849475
186852309 (PUW)
1.07
s5
25210
25210 (mozartballs polito PUW scipjack staynerd)
0.00
Table VIII. Best PB for all the SPG instances and comparison with the best results of the
Challenge
42
Instance
PB
Algo
Time [s] “polito” PB Gap %
C13-A
236
FJNW
1.23 (N) 237
−0.42
C19-B
146
FJNW
2.62 (N) 146
0.00
D03-B
1509
FJNW
5.79 (N) 1510
D20-A
536
FJNW
4.64 (N) 536
0.00
P400-3
2951725 F J N W
1.64 (N) 2951725
0.00
P400-4
2852956 F J N W
4.30 (N) 2852956
0.00
K400-7
485587
W
2289.60 523885
−7.31
K400-10
401032
W
1086.08 406365
−1.31
hc10p
59682
W
7140.50 59813
−0.22
hc11u
1115
FJN
1988.85 (N) 1120
hc12p
235132
FJN
5463.03 (N) 235043
hc12u
2216
FJN
6071.29 (N) 2227
bip52nu
222
F
3093.53 223
−0.45
bip62nu
214
FJNW
4.32 (N) 214
0.00
cc3-12nu
95
FJNW
26.43 (N) 95
cc12-2nu
565
FN
2789.17 (F) 567
i640-001
2932
FJNW
0.42 (N) 3053
−3.96
i640-221
8430
F
5466.72 8626
−2.27
i640-321
28790
F
6949.88 28821
−0.11
i640-341
29679
F
2037.28 29713
−0.11
a2000RandGraph-2
1483.84 F J N W
10.04 (N) 1484.2
−0.02
a4000RandGraph-3
3406.62 F J N W
6.74 (N) 3407.5
−0.03
−0.07
−0.45
0.04
−0.49
0.00
−0.35
a8000RandGraph-1-2 4719.97
FJN
5552.40 (N) 4722.8
−0.06
a14000RandGraph-1-5 9475.59
FJN
690.66 (N) 9475.6
0.00
handsd04
525.86
W
handbd13
13.23
FJNW
handsi03
56.23
handbi07
1470.15 584.1
−9.97
100.15 (N) 18.1
−26.91
FJN
31.95 (J) 56.3
−0.12
151.04
FJNW
88.46 (W) 151.1
−0.04
drosophila001
8273.98
W
HCMV
3491.73 8288.3
−0.17
7376.36 F J N W
4.62 (W) 7378.2
−0.02
lymphoma
3341.89 F J N W
344.67 (N) 3349.1
−0.22
metabol-expr-mice-1
11346.93 F J N W 2120.98 (W) 11901.9
−4.66
Table IX. Best PB for all the PCSPG instances and comparison with “polito” results
43
Instance
PB
Best PB of the Challenge Gap %
C13-A
236
236 (KTS mozartballs staynerd)
0.00
C19-B
146
146 (All)
0.00
D03-B
1509
1509 (All)
0.00
D20-A
536
536 (All)
0.00
P400-3
2951725
2951725 (All)
0.00
P400-4
2852956
2852956 (All)
0.00
K400-7
485587
474466 (KTS mozartballs staynerd)
2.34
K400-10
401032
394191 (KTS mozartballs staynerd)
1.74
hc10p
59682
59738 (polito)
−0.09
hc11u
1115
1116 (mozartballs staynerd)
−0.09
hc12p
235132
234977 (polito)
0.07
hc12u
2216
2221 (KTS)
−0.23
bip52nu
222
222 (mozartballs)
0.00
bip62nu
214
214 (All)
0.00
cc3-12nu
95
95 (KTS mozartballs staynerd)
0.00
cc12-2nu
565
567 (KTS polito)
−0.35
i640-001
2932
2932 (All)
0.00
i640-221
8430
8400 (KTS mozartballs)
0.36
i640-321
28790
28787 (KTS)
0.01
i640-341
29679
29666 (KTS)
0.04
a2000RandGraph-2
1483.84
1483.8 (mozartballs staynerd)
0.00
a4000RandGraph-3
3406.62
3406.6 (mozartballs staynerd)
0.00
a8000RandGraph-1-2 4719.97
4720.0 (mozartballs staynerd)
0.00
a14000RandGraph-1-5 9475.59
9475.6 (mozartballs staynerd)
0.00
493.80 (staynerd)
6.49
handsd04
525.86
handbd13
13.23
13.20 (All)
0.23
handsi03
56.23
56.10 (mozartballs staynerd)
0.23
handbi07
151.04
151 (KTS mozartballs staynerd)
0.03
drosophila001
8273.98
8274 (mozartballs staynerd)
0.00
HCMV
7376.36
7371.5 (KTS mozartballs staynerd)
0.07
lymphoma
3341.89
3341.9 (KTS mozartballs staynerd)
0.00
metabol-expr-mice-1
11346.93
11346.9 (mozartballs staynerd)
0.00
Table X. Best PB for all the PCSPG instances and comparison with the best results of the
Challenge
44
Instance
MS reweighting
MST
SPT
d18
223
335
358
e18
564
902
892
i640-211
11984
26928
15967
i640-314
35532
55479
41186
i640-341
32042
53826
43593
fnl4461fst
189367
203167
271681
world666
122971
172983
1356820
alue7080
66624
112478
188092
alut2625
40183
89813
157660
es10000fst01 733237957
760866530 990978452
lin36
64486
145363
130588
lin37
112886
223493
233031
hc12p
236075
332813
314607
hc12u
2269
2749
2878
cc12-2p
121056
237979
175225
cc12-2u
1174
1644
1624
cc12-2n
613
1130
1116
cc3-12n
111
165
155
cc3-12p
19003
44366
26078
cc3-12u
185
245
255
bipa2p
35285
56488
45663
bipa2u
338
421
420
2r211c
89000
155000
165000
wrp3-83
8301057
8301505
8402368
w23c23
691
828
918
rc09
120450
138800
226297
rt05
58286
66236
103866
G106ac
40396063
44792419 45158294
I064ac
188849475
190392002 192238839
s5
25210
25210
25210
Table XI. Heuristics vs MS guided heuristics. Comparison of the PB for the SPG instances
45
Instance
MS reweightng
MST
SPT
GW
C13-A
236
305
360
309
C19-B
146
215
237
204
D03-B
1509
1819
1799
1819
D20-A
536
661
638
806
P400-3
2951725
3633310 4108109 3475372
P400-4
2852956
3649056 3785708 3577568
K400-7
485587
528817
528817
528817
K400-10
401032
434488
434488
434488
hc10p
59682
77232
77232
75181
hc11u
1115
1318
1363
1264
hc12p
235132
308065
308065
308065
hc12u
2216
2581
2666
2543
bip52nu
222
284
281
278
bip62nu
214
239
242
244
cc3-12nu
95
112
112
112
cc12-2nu
565
697
697
697
i640-001
2932
3764
3764
3764
i640-221
8430
21117
12006
11630
i640-321
28790
49605
42401
41880
i640-341
29679
49707
38085
41826
a2000RandGraph-2
1483.84
1648.58 1822.47 1645.15
a4000RandGraph-3
3406.62
3616.03 4314.37 3615.58
a8000RandGraph-1-2 4719.97
4790.86 4790.86 4790.86
a14000RandGraph-1-5 9475.59
10514.12 10514.12 10514.12
handsd04
525.86
792.61
792.61
792.61
handbd13
13.23
13.24
13.24
13.24
handsi03
56.23
56.26
56.26
56.26
handbi07
151.04
151.06
151.06
151.06
drosophila001
8273.98
8296.30 8296.30 8296.30
HCMV
7376.36
7376.84 7376.84 7376.84
lymphoma
3341.89
3410.36 3410.36 3410.36
metabol-expr-mice-1
11346.93
11885.90 11885.90 11885.90
Table XII. Heuristics vs MS guided heuristics. Comparison of the PB for the PCSPG instances
46
Instance Best Known Npre Gap %
bip42p
24657 (opt)
bip42u
236 (opt)
bip52p
24526
bip52u
234
bip62p
22843
bip62u
219
bipe2p
5616 (opt)
bipe2u
54 (opt)
cc10-2p
35297
cc10-2u
342
cc11-2p
63491
cc11-2u
612
cc3-10p
12772
cc3-10u
125
cc3-11p
15582
cc3-11u
153
cc3-4p
2338 (opt)
cc3-4u
23 (opt)
cc3-5p
3661 (opt)
cc3-5u
36 (opt)
cc5-3p
7299 (opt)
cc5-3u
71 (opt)
cc6-2p
3271 (opt)
cc6-2u
32 (opt)
cc6-3p
20270 (opt)
cc6-3u
197 (opt)
cc7-3p
56799
cc7-3u
549
cc9-2p
17199
cc9-2u
167 (opt)
24657
0.00
236
0.00
24549
0.09
234
0.00
22843
0.00
219
0.00
5616
0.00
54
0.00
35269
−0.08
342
0.00
63405
−0.14
614
0.33
12870
0.77
125
0.00
15680
0.63
153
0.00
2338
0.00
23
0.00
3665
0.11
36
0.00
7302
0.04
71
0.00
3271
0.00
32
0.00
20298
0.14
198
0.51
56835
0.06
553
0.73
17225
0.15
167
0.00
Table XIII. Best known energies for PUC instances, “bip”-like and “cc”-like classes
47
Instance Best Known
Npre Gap %
hc10p
59797
59808
0.02
hc10u
575
575
0.00
hc11p
119492
119456
−0.03
hc11u
1145
1153
0.70
hc6p
4003 (opt)
4003
0.00
hc6u
39 (opt)
39
0.00
hc7p
7905 (opt)
7906
0.01
hc7u
77 (opt)
77
0.00
hc8p
15322 (opt)
15336
0.09
hc8u
148 (opt)
148
0.00
hc9p
30242
30313
0.23
hc9u
292
292
0.00
Table XIV. Best known energies for PUC instances, “hc”-like class
48
[1] Shao-shan Carol Huang and Ernest Fraenkel. Integrating Proteomic, Transcriptional, and
Interactome Data Reveals Hidden Components of Signaling and Regulatory Networks. Sci.
Signal., 2(81):ra40–ra40, July 2009.
[2] Marc Bailly-Bechet, Alfredo Braunstein, and Riccardo Zecchina. A prize-collecting steiner
tree approach for transduction network inference. In Pierpaolo Degano and Roberto Gorrieri, editors, Computational Methods in Systems Biology, number 5688 in Lecture Notes in
Computer Science, pages 83–95. Springer Berlin Heidelberg, January 2009.
[3] Nurcan Tuncbag, Alfredo Braunstein, Andrea Pagnani, Shao-ShanCarol Huang, Jennifer
Chayes, Christian Borgs, Riccardo Zecchina, and Ernest Fraenkel. Simultaneous reconstruction of multiple signaling pathways via the prize-collecting steiner forest problem. In
Benny Chor, editor, Research in Computational Molecular Biology, volume 7262 of Lecture
Notes in Computer Science, pages 287–301. Springer Berlin Heidelberg, 2012. Cited by
0002.
[4] Ivana Ljubić, René Weiskircher, Ulrich Pferschy, Gunnar W. Klau, Petra Mutzel, and Matteo Fischetti. An Algorithmic Framework for the Exact Solution of the Prize-Collecting
Steiner Tree Problem. Math. Program., 105(2-3):427–449, October 2005.
[5] Mohsen Bayati, A. Braunstein, and Riccardo Zecchina. A rigorous analysis of the cavity
equations for the minimum spanning tree. Journal of Mathematical Physics, 49(12):125206,
2008. Cited by 0012.
[6] M. Bayati, C. Borgs, A. Braunstein, J. Chayes, A. Ramezanpour, and R. Zecchina. Statistical mechanics of steiner trees. Physical Review Letters, 101(3):037208, July 2008.
[7] Indaco Biazzo, Alfredo Braunstein, and Riccardo Zecchina. Performance of a cavity-methodbased algorithm for the prize-collecting steiner tree problem on graphs. Phys. Rev. E,
86:026706, August 2012.
[8] Marc Mézard and Andrea Montanari. Information, Physics, and Computation. Oxford
University Press, January 2009.
[9] M. Bailly-Bechet, C. Borgs, A. Braunstein, J. Chayes, A. Dagkessamanskaia, J.-M. François,
and R. Zecchina. Finding undetected protein associations in cell signaling by belief propagation. Proceedings of the National Academy of Sciences, 108(2):882–887, January 2011.
[10] Michel X. Goemans and David P. Williamson. A general approximation technique for
constrained forest problems. In Proceedings of the Third Annual ACM-SIAM Symposium
49
on Discrete Algorithms, SODA ’92, pages 307–316, Philadelphia, PA, USA, 1992. Society
for Industrial and Applied Mathematics.
[11] Michel X. Goemans and David P. Williamson. Approximation algorithms for np-hard problems. chapter The Primal-dual Method for Approximation Algorithms and Its Application
to Network Design Problems, pages 144–191. PWS Publishing Co., Boston, MA, USA, 1997.
| 8cs.DS
|
1
An Empirical Study of the L2-Boost technique with
Echo State Networks
Sebastián Basterrech
arXiv:1501.00503v1 [cs.LG] 2 Jan 2015
IT4Innovations
National Supercomputing Center
VSB-Technical University of Ostrava
Ostrava, Czech Republic
[email protected]
Abstract—A particular case of Recurrent Neural Network
(RNN) was introduced at the beginning of the 2000s under the
name of Echo State Networks (ESNs). The ESN model overcomes
the limitations during the training of the RNNs while introducing
no significant disadvantages. Although the model presents some
well-identified drawbacks when the parameters are not well
initialized. The performance of an ESN is highly dependent on
its internal parameters and pattern of connectivity of the hiddenhidden weights Often, the tuning of the network parameters can
be hard and can impact in the accuracy of the models.
In this work, we investigate the performance of a specific
boosting technique (called L2 -Boost) with ESNs as single predictors. The L2 -Boost technique has been shown to be an effective
tool to combine “weak” predictors in regression problems. In
this study, we use an ensemble of random initialized ESNs
(without control their parameters) as “weak” predictors of the
boosting procedure. We evaluate our approach on five well-know
time-series benchmark problems. Additionally, we compare this
technique with a baseline approach that consists of averaging the
prediction of an ensemble of ESNs.
Keywords-L2 -boosting, Echo State Network, Time-series modeling, Reservoir Computing, Ensemble Methods
I. I NTRODUCTION
Boosting is a general procedure for improving the accuracy
of an ensemble of methods. It has been successful used
in supervised learning problems since its apparition in the
1990s [1]–[3]. Several variations of the original Boosting idea
have been introduced over the years [4], [5], one of the most
popular is called AdaBoost [3]. At the beginning, Boosting
was used in problems where the output features were label
or discrete responses (classification problems). An analogy
between AdaBoost and additive models was studied in [4].
This connection was essential for the extension of Boosting
for solving problems where the output features are continuous
variables (regression problems). Bühlmann et al. developed
a variation of the Boosting technique called L2 -Boost that is
constructed from an additive model and the functional gradient
descent method [5].
Since the early 2000s, a computational paradigm called
Reservoir Computing (RC) has gained prominence in the
Neural Computation community. In a RC model there are two
well-separated concepts: a dynamical system and a memoryless function. The purpose of the dynamical system is to
encode the spatio-temporal information of the input patterns
into a spatial representation. At each time this dynamical
system is characterized by its state that is called reservoir
in the RC literature. This non-linear transformation is most
often realized by a Recurrent Neural Network (RNN) with a
large pool of interconnected neurons. A distinctive principle
of a RC model is that the parameters of the dynamical system
(the RNN weights) do not participate in the training process.
That is, once the reservoir parameters are initialized, they
remain fixed during the training process. Another part of the
model is a memory-less supervised learning tool called readout
structure. This part is designed to be robust and fast in the
learning process.
The RC models has been applied in the neuroscience area
for processing cognitive information in the neural system. [6].
Furthermore, they have proven to be extremely effective tools
for time-series problems in the area of Machine Learning. For
instance, as far as we know one of the most popular RC models
the Echo State Networks (ESN) [7], has the best known learning performance on the Mackey-Glass time-series prediction
problem [8], [9]. In this article, we will concentrate in the ESN
model for solving time-series problems. The reservoir in the
ESN model is composed by a RNN with sigmoid neurons, and
the readout part of the model is a linear regression. The weight
connection between neurons in the reservoir are collected in
a matrix that we will call reservoir matrix. The main global
parameters of the ESN model are: the input scaling factor,
the spectral radius of the reservoir matrix and the pattern of
connectivity among the reservoir units. The setting of these
parameters often requires the human expertise and several
empirical trials [10]. As a consequence, the setting procedure
can be expensive in computational time. For instance, the time
complexity of an algorithm that computes the spectral radius
of a N × N matrix is equal to O(N 4 ) [11].
The goal of this article is to investigate the performance
of an automatic procedure to combine single weak ESNs and
the L2 -Boost technique. We develop an automatic technique
based on L2 -Boost, which combines the prediction of several
random initialized ESNs in order to produce a highly accurate
tool. We use the term weak ESN for an ESN without checking
and computing the spectral radius. We use the terminology
weak due to the fact that this particular ESN is not optimal.
The main advantages of the procedure presented in this article
are:
• Descend the computational effort. In order to gain in the
computational effort, the approach consists of combining
weak ESNs. The procedure avoids to tune the reservoir
parameters, which can often have a high computational
cost. It uses only a uniform random initialization of the
weights. Note that, there are not a control of the spectral
radius, then some single weak ESNs can have unstable
dynamics.
• The procedure is automatic. The procedure does not
require external human expertise for setting the model
parameters, and for evaluating the model performance.
• The technique has a new parameter used for overfitting
control. This parameter which we will that comes from
the L2 -Boost technique.
We present empirical results of the procedure introduced
in this paper on a wide range of benchmark problems. We
compare these performances with the accuracy obtained by a
single ESN. Furthermore, we realize a comparison with the
accuracy of a baseline approach that computes the average
among single ESN models. Each single ESN is independently
initialized and adjusted during the learning process. There are
empirical evidence in the Machine Learning literature that
show that this baseline approach sometimes performs better
than other ensemble methods [12].
This work is a revised and expanded version of the article [13].
The structure of this article is organized as follows. In
Section II, we start with a specification of supervised learning
problems with temporal data. Next, we present an overview
about the family of additive models. Subsection II-C introduces the L2 -Boost technique. In Section III, we present the
Reservoir Computing paradigm. Particularly, we focus on the
Echo State Network model in III-A. In Subsection III-B is
presented a formalization of the procedure introduced in this
article. Section IV describes the empirical results. This section
starts with a description of the benchmark problems. Next,
we present the reached results. Finally, last section provides
conclusions and future work.
II. BACKGROUND
In this Section, we start specifying the context where the
ESN model and the L2 -Boost technique are applied. Next, we
present the additive models and we introduce a description of
the L2 -Boost technique.
A. Problem Specification
We begin specifying a supervised learning problem. Given
a data set L = {(x(t) , y(t) ) : t = 1, . . . , T } where the
points x and y are either a class or a numerical response.
We denote by Nx the dimension of the input vector x,
and Ny the dimension of the output vector y. We suppose
that the mapping between the input x and the output y is
given by certain unknown function F (·). The goal consists
in learning a parametric function F̂ (x(t) , L) such that certain
error distance between F̂ (x(t) , L) and y(t) is minimized for all
t. The problem is called regression problem when the learning
set has output numerical variables. Otherwise, it is named
classification problem. In the case of regression problems, it
is recommended to use the a quadratic distance [14]. Even
though we can also use a quadratic distance in classification
problems, it is recommendable to use the Kullback-Leibler
distance in this domain [14].
An ESN model is mainly used for solving supervised learning tasks, wherein the data set presents temporal dependencies.
Although, it can be also used for non-temporal supervised
learning problems [9]. In this article we will concentrate
only in temporal learning tasks with real output variables
(y(t) ∈ RNY , for all t). In this work, we perform the models
using a standard discrete time. We want to forecast some
aspect of the output feature y at time t + k, using some aspect
of the information available at current time t, that is given the
collection ((x(t) , y(t) ), (x(t−1) , y(t−1) ), (x(t−2) , y(t−2) ), . . .)
we would like to predict the value y(t+k) (k > 0) [15]. In
this case, the goal consists in estimating a mapping F̂ (·) for
predicting y(t+k) for some k > 0, such that some distance
between F̂ (·) and y is minimized.
B. Additive Models
In [4] was analyzed the Boosting model under the form of an
Additive model. Given a set of functions f (m) : RNX → RNY ,
m = 1 . . . , M characterized by a set of parameters θ and
expansion coefficients β,
f (m) (x) = β (m) h(x, θ(m) ),
an additive model has the following form
F (x) =
M
X
f (m) (x).
(1)
m=1
The functions {h(x; θ)}M
1 are named basis functions. They
are not fixed a priori and are selected depending of the cost
function used and the data set. An important parameter of the
model is the number of basis functions (M ) considered in
the expression (1). This parameter controls the generalization
error of the model. Since the main goal in a learning task is to
find a predictor with low generalization error, the parameter
M has an important role in the accuracy of an additive model.
C. The L2 -Boost Procedure
A relationship between the gradient descent technique and
stage-wise additive expansions was introduced at the beginning
of the 2000s [16]. The introduction of the gradient descent
algorithm using a boosting approach was an essential contribution in the field of ensemble learning methods [16]. It
allowed to start to use boosting in regression problems [5].
A Boost method for regression problems with quadratic error
distance was introduced under the name of L2 -Boost in [5]. We
present the L2 -Boost technique in Algorithm 1. Other boosting
variants were presented for other kind of distances, some of
them are described in [4], [5], [16].
We refer by epoch to the iteration of the training algorithm
through all the patterns in the training set [17]. At each epoch
m + 1, the basis function h(m+1) (·, θ) is fitted to the current
residuals: y(i) − F̂ (m) (x(i) ), for all i. Unlike other boosting
techniques such as Adaboost, L2 -Boost does not present any
re-weighting. Another difference between L2 -Boost and other
boosting methods is that L2 -Boost presents a tendency to overfit the data [5]. The model with contracting linear learners
converge to the fully saturated model [5]. Each boosting
epoch contributes to additional overfitting, thus the selection
of the weak learners and the parameter M is an essential
task for this device. In practice, few boosting iterations are
enough to achieve good performances avoiding the overfitting
phenomena.
Algorithm 1 The L2 -Boost algorithm.
Require: L, M , h(x, θ)
Fit an initial model using a least squares fit (see [18]):
F̂ (0) (·) = h(·, θ);
for (m = 1, . . . M ) do
Compute
the
residuals
for
all
pattern
i:
e(i) = y(i) − F̂ (m) (x(i) );
Fit
the
model
fˆ(m+1) (·)
parametrized
as
fˆ(m+1) (x) = h(x, θ) to the current residuals e using the
least squares fit;
Update: F̂ (m+1) (·) = F̂ (m) (·) + fˆ(m+1) (·);
end for
Return the F̂ (M ) (·) function;
to the opposite, when the gradient norm large increases during
the training process [21]. Recently, an effective algorithm
to train RNN was introduced [19], the algorithm uses the
Hessian-free Optimization for setting the network parameters.
Reservoir Computing (RC) models appear as a good alternative for RNNs. The two pioneering RC models are Echo State
Network (ESN) [7] and Liquid State Machine (LSM) [22]. This
computational paradigm covers the main limitations related to
learning processes in RNNs obtaining acceptable performance
in practical applications [9]. In a RC model there are at least
two well-differentiated structures: a dynamical system called
reservoir and another one called readout. The readout is a
supervised learning tool for training with non-temporal data.
For example: feedforward neural network, linear regression,
decision trees, etc. A main characteristic of a RC model is
that the weights involved in circuits are deemed fixed during
the learning process. Thus, the matrix with the weight between
reservoir units (reservoir matrix) is initialized in an arbitrary
way and it remains unchanged during the learning process.
The training algorithm is restricted to update the weights in
the readout structure. Over the last years several kinds of
dynamical systems have been used for generating the reservoir
state, models include: Backpropagation-decorrelation Recurrent Learning [23], Leaky Integrator Echo State Networks
studied [24], Evolino [8], Intrinsic Plasticity [25], Echo State
Queueing Networks [26], Reservoir Computing and Extreme
Learning [27], and so on.
A. Formalization of the Echo State Network Model
III. M ODELING T IME - SERIES WITH E CHO S TATE
N ETWORKS
The Recurrent Neural Networks (RNNs) are powerful tools
for solving time-series benchmarks. They are computational
methods that operate in time. Considering terminology of
graphs, in a RNN at least one circuit is presented in its
topology. The circuits of the network enable to store temporal
information, in order to learn and memorize the input history [9]. Each circuit creates an internal state which makes the
recurrent network a discrete time state-space model. At each
time, the RNN receives an input pattern. Next, the network
updates its hidden state via a non-linear activation function
using the input pattern and the network state at the precedent
time [19]. There are a general consensus in the community
that considers the RNN as powerful tool for forecasting and
time-series prediction.
In spite of that, in practice the model presents some drawbacks. The most important is that is hard to train a RNN
using gradient descent methods [20]. The training methods
that use the first differential information have often stability
problems and high numerical complexity. As a consequence,
much longer training times are necessary to adjust the network
weights. In [20] is analyzed the main limitations of the algorithms of the gradient descent type for training RNNs. These
drawbacks are identified under the names of vanishing and
the exploding gradient problems [20]. The vanishing gradient
phenomena occurs when the norm of the gradient decreases
arbitrarily fast to 0. The exploding gradient phenomena refers
In this work related to the L2 -Boost technique and the RC
methods, we only study the L2 -Boost with the ESN model.
An ESN reservoir is a RNN from an input space RNx into
a larger space RNs with Nx Ns . The connection between
input and hidden neurons are collected in a Ns × Nx weight
matrix win . The connections among the hidden neurons are
represented by a Ns ×Ns weight matrix wr . A Ny ×Ns weight
matrix wout represents the readout weights. At any time t, the
information from the input pattern and the past is represented
in a state vector
s(t) = tanh(win x(t) + wr s(t−1) ).
(2)
At any time t, the output prediction y(t) ∈ RNy is generated
using the input pattern and the reservoir state information.
Most often is computed using a linear regression:
y(t) = wout [x(t) |s(t) ],
(3)
where ·|· is the vertical concatenation of the vectors. For
the sake of the notation simplicity, we omit the bias term,
a constant term is included in all the regressions.
In [7] was analyzed the stability of the reservoir dynamics
in the ESN model. Under certain algebraic conditions the
reservoir state only depends (asymptotically) of the inputs
and the network topology. It becomes independent of its
initial conditions [9]. These conditions were summarized in
the Echo State Property (ESP) [7]. In practice, the stability
of the ESN is almost always ensured when the spectral
radius of the reservoir matrix is less than 1 [9], [28]. As a
consequence, the reservoir weights are appropriately scaled
in order to have a spectral radius less than 1. To scale the
parameters is necessary to compute the spectral radius of the
reservoir matrix. The computation of the spectra requires an
important computational effort [11]. Some attempts to generate
a procedure for initializing the RC models were introduced
in [11], [25], [29]–[31].
A. Description of the Benchmark Problems
We use the following range of time-series benchmarks:
• Fixed kth order NARMA. This data set presents a high
non-linearity and is widely used in the RC literature.
We generate the NARMA serie following the description
in [28], [32],
k−1
X
b(t + 1) = α1 (t) + α2 b(t)
b(t − i)
i=0
B. L2 -Boost Using the ESN Model for Time-series Processing
Information
In this article, we investigate the performance of using L2 Boost in temporal learning tasks, and we consider as weak
learner predictors a set of ESNs with random initialization.
Given an arbitrary parameter M the procedure is as follows.
We initialize an ESN in a random way. The initialization
consists in selecting the size of the network as well as the
pattern of connectivity. We consider a reservoir with fixed
sparse connections. We do not control the spectrum norm of
the reservoir weight. A guide about the initialization procedure
can be seen from [10]. We expand the input information
using the ESN reservoir given by the expression (2), thus
we obtain s(t) , ∀t. Next, we apply Algorithm 1. Finally, we
obtain predictor F (M ) (·). The approach is summarized in
Algorithm 2. In our experiments we use ridge linear regression
for computing the readout weights wout [28].
Algorithm 2 The L2 -Boost with the ESN model.
+α3 s(t − (k − 1))s(t) + α4 ,
where s(t) ∼ U nif [0, 0.5] and the constants values are
shown in Table I. In order to evaluate the memorization
ability of the model, we consider two simulated series
when k = 10 and k = 30. The task consists to predict
k
10
30
IV. E MPIRICAL R ESULTS
We begin this section describing the benchmark problems.
Next, we specify the experimental setup. We concludes this
section with an analysis of our empirical results.
α2
0.05
0.004
α3
1.5
1.5
α4
0.1
0.001
Table I: Parameters considered for the fixed kth order NARMA
serie with k = 10 and k = 30.
•
Initialize an ESN following the comments in Subsection III-B;
Compute the temporal expansion of L using (2);
Generate the set {(s(t) , y(t) ), ∀t};
Apply the Algorithm 1;
Return F̂ (M ) (·);
In order to evaluate the performance of this procedure,
we compare the reached accuracy of the L2 -Boost technique
with a simple baseline approach [17]. The baseline approach
consists in combining K single predictors (in our case the
learning predictors are ESNs). We consider random initialized
reservoirs, without control of the reservoir spectrum norm. In
the baseline method, we train independently each of these
single ESNs. The final prediction is the average among the
single predictions.
For statistical comparisons between the methods we consider K = 30. We remark again that we do not scale the
reservoir weights for obtaining the ESP. Even though some
ESN models can present good accuracy, other ones can be
weak predictors. Additionally, we compare our performances
with the performance obtained when single ESNs with “good”
tuning of the reservoir parameters are used. For that, we use
the results presented in the RC literature.
α1
0.3
0.2
•
the value y(t + 1) based on the history of y(t) up to time
t. We used the first 200 samples as initial washout in
both the training and test procedure. The regularization
parameter used was 0.00001.
The Santa Fe Laser data set [33]. It is an experimental
data that contains the intensity pulsations of a real laser
recorded by a LeCroy oscilloscope. The data is a crosscut through periodic to chaotic intensity laser pulsations.
These pulsations more or less follow the theoretical
Lorenz model of a two level system [33]. In this problem,
the task consists to predict the next measure y(t + 1),
given the precedent values up to t. The original data only
consists of 1000 measurements, we used for training 499
samples and for testing 500 samples. We used a washout
of 10 samples. The regularization parameter γ was 0.001.
Henon Map data set. It is a prototypical invertible map
with chaotic solutions proposed in [34]. The data is
generated by
y(t + 1) = 1 − 1.4(y(t))2 + 0.3y(t − 1) + z(t + 1),
•
where the noise is z(t) ∼ N(0, 0.05). The data is
normalized in [0, 1]. The goal is to predict the next value
y(t + 1) with the past information up to t. We considered
a training data with 3995 samples and a test data with 795
samples. The regularization parameter γ used was 0.001.
We use an initial washout composed by 100 samples. The
network topology has 3 input units set with the last two
precedent y(t) values and the noise at current time.
Freedman’s non linear time data set [35]. The data is
generated by
y(t + 1) = g(y(t)),
where:
g(x) =
2x,
2 − 2x,
ifx ≤ 0.5,
x > 0.5.
We consider a very short data set. The length of the
training data was 30 and the test size was 19. The initial
value is y(0) = 0.23719. The initial washout considered
only was of 3 samples. The network topology has only
one input unit, several reservoir units and one output unit.
The regularization parameter γ used was 0.001.
B. Experimental Setup
DATA
10th NARMA
30th NARMA
Santa Fe Laser
Henon Map
Freedman’s
Initial
Washout
200
200
10
100
3
γ
0.00001
0.00001
0.001
0.001
0.001
Train
samples
1400
1600
499
3995
30
Test
samples
2400
2600
500
795
19
Table II: Parameter setting of the benchmark problems. In all
cases, we initialize the input weights (win ) using Uniform
distribution in [−0.2, 0.2], and we initialize the reservoir
weights with an Uniform distribution in [−0.8, 0.8].
We summarize the setting of the main parameters related
to the benchmark problems in Table II. The table presents
the initial washout period, the regularization parameter (γ) of
the linear ridge regression, and the number of train and test
samples for each benchmark problem.
The benchmarks selected have been widely used in the
RC literature [7], [28], [31], [36]. In all cases, we use
the Normalized Mean Square Error (NMSE) as measure of
accuracy model [9]. The learning method used for computing
the output weight matrix wout was the offline ridge regression.
This algorithm has a regularization parameter γ that we adjust
it for each benchmark problem. The pre-processing data step
consisted in normalizing the patterns in the interval [0, 1] We
investigated the algorithm performance for several reservoir
sizes. The range of the reservoir size values is specified for
each benchmark problem. The connection between the input
and reservoir layer is fully connected with random weights in
[−0.2, 0.2]. The reservoir matrix is initialized using Uniform
distribution in [−0.8, 0.8].
columns 3 and 4 of Table IV are very similar. The model
with larger M performs better over the train data, but it has
problems of generalization. We found this characteristic in
all benchmarks. As a consequence, we can affirm that the
parameter M has a relevant impact in the control of the
overfitting phenomenon. We can found a similar remarks for
the L2 -Boost technique in non-temporal learning tasks [5].
Figure 2 illustrates the NMSE reached according the reservoir size for different M values for the 30th NARMA data set.
This figure shows the training error, we can see the evolution
of the NMSE versus the size of the reservoir. We present few
values of reservoir size between 6 till 11. Figures 1 and 3
show the NMSE of the test data versus the reservoir size for
the 10th and 30th order NARMA data set, respectively. These
figures show 4 curves, the black one (with points represented
by dots) corresponds to the baseline method which combines
several single ESNs. The other curves correspond to the L2 Boost-ESN with different number of epochs M = 6, 8 and
M = 10. We can not affirm that the procedure of L2 -Boost
with single weak ESNs performs better than optimal single
ESNs. The accuracy it is also of the same order that results
presented in the RC literature using a single well-initialized
ESN [28], [31].
Figure 4 illustrates the evolution of the NMSE for the
reservoir size of the test data of the Santa Fe Laser benchmark.
The error was computed for the L2 -Boost with ESNs for
M = 4, 6 and M = 8 and the baseline approach averaging 30
ESNs. Figure 5 shows the accuracy reached for the models on
the Freedman test data set. The graphic shows the evolution
of the L2 -Boost with ESNs for M = 4, 5 and M = 5 and
the baseline approach. In all graphics, we can see that when
the reservoir increases its size the procedure L2 -Boost with
ESNs and the baseline approach decrease their test error. This
behavior about the impact of the reservoir size on the accuracy
of the model, also happens with single ESNs [10], [28], [31].
DATA
10TH NARMA
30TH NARMA
S ANTA F E L ASER
H ENON M AP
C. Result Analysis
Table III shows results reported in the RC literature for these
benchmarks when a single ESN model was used as model
predictor. Table IV presents the train set accuracy reached
on the Henon Map data set. The columns 2, 3 and 4 show
the NMSE reached with L2 -Boost with ESNs for M epochs
(M = 3, 4 and 5), respectively. Column 5 of Table III shows
the accuracy of the baseline approach, that it averaging the
prediction of 30 ESNs. The columns of the table are written
using a scientific notation.
Table IV illustrates the accuracy of the models during the
training. The NMSE corresponds to the training data of the
Henon Map data set. We present this table in order to illustrate
the tendency of overfitting of L2 -Boost with ESN. The additive
model converge very fast to the solution, for this reason the
F REEDMAN ’ S
ACCURACY
0.166 (NMSE)
0.0425 (NMSE)
0.4542 (NRMSE)
0.0184 (NMSE)
0.00819 (NMSE)
0.00975 (NMSE)
0.00868 (NMSE)
0.0004302 (MSE)
Nx
50
200
100
50
200
50
200
40
R EF.
[28]
[28]
[30]
[28]
[28]
[28]
[28]
[31]
Table III: Accuracy of the ESN model for the benchmark
problems. Second column shows the accuracy reached by the
single ESN, third column refers the reservoir size and the last
column shows a bibliographic reference. In the case of the
Freedman’s non linear time data, the reservoir initialization
was done using the Scale Invariant Map method [31], and
the Mean Square Error (MSE) was the error measure. In the
case of 30th NARMA, the authors initialize the reservoir using
permutation matrices. The error measure was the Normalized
Root Square Error (NRMSE) [30]. In the other benchmarks
problems, the authors control some reservoir parameters such
as: spectral radius and reservoir matrix density.
M =4
(1.0 E -13)
0.131036
0.131036
0.131036
0.131036
0.131036
0.131036
0.131036
M =5
(1.0 E -13)
0.131036
0.131036
0.131036
0.131036
0.131036
0.131035
0.131036
BAS . ESN
(1.0 E -12)
0.288720
0.064412
0.572728
0.182942
0.278237
0.524017
0.459002
Table IV: Train set performance of the Henon Map data set.
First column indicates the number of neurons in the reservoir.
The columns 2, 3 and 4 show the NMSE obtained with L2 Boost with M epochs (M = 3, 4 and 5), respectively. Column
5 shows the accuracy of the baseline approach. That is the
accuracy average among 30 ESN predictions. The columns
are written using a scientific notation.
10 order Narma test data
0.18
Baseline ESN
Boost: M=4
Boost: M=5
Boost: M=6
0.17
30 order Narma Train data
Baseline ESN
Boost: M=6
Boost: M=8
Boost: M=10
0.63
0.62
0.61
NMSE
M =3
(1.0 E -12)
0.297367
0.191883
0.142289
0.160910
0.203845
0.272649
0.488265
Nx
6
7
8
9
10
11
12
0.6
0.59
0.58
0.57
6
7
8
9
Reservoir size
10
11
Figure 2: The accuracy reached on the training set of the 30th
order NARMA data. The vertical axis of the graph shows the
NMSE accuracy, and the horizontal axis presents some values
of the reservoir size. We compare the accuracy of L2 -Boost:
M = 6, 8 and 10 with the baseline approach averaging 30
ESNs.
NMSE
0.16
0.15
30 order Narma Test data
0.6
0.14
0.59
0.13
0.12
0.58
6
7
8
9
Reservoir size
10
11
Figure 1: Test set accuracy reached for the 10th order NARMA
data set. The vertical axis of the graph shows the NMSE
accuracy, and the horizontal axis presents some values of
the reservoir size. We compare the accuracy of L2 -Boost:
M = 4, 5 and 6 with the baseline approach averaging 30
ESNs.
V. C ONCLUSIONS AND F UTURE W ORK
At the beginning of the 2000s, an efficient technique to
train and design a RNN was developed under the name
of Echo State Network (ESN). This approach overcome the
limitations to train RNN using the gradient descent method.
The performance of an ESN is highly dependent on its
parameters and pattern of connectivity of the hidden-hidden
weights Besides, the network setting can be computational
expensive, in particular to compute the spectral radius of the
hidden-hidden weight matrix.
In this article, we investigated boosting ideas with ESNs, in
order to built a robust new learning tool. In particular, we
studied the utilization of L2 -Boost with random initialized
ESNs. We merge a set of weak single ESNs. We call weak
ESNs because they are random initialized, and we do not use
extra computational effort for tuning the initial hidden-hidden
weights.
NMSE
0.11
Baseline ESN
Boost: M=6
Boost: M=8
Boost: M=10
0.57
0.56
0.55
0.54
6
7
8
9
Reservoir size
10
11
Figure 3: The accuracy reached for the 30th order NARMA
data set. The vertical axis of the graph shows the NMSE
accuracy, and the horizontal axis presents some values of
the reservoir size. We compare the accuracy of L2 -Boost:
M = 6, 8 and 10 with the baseline approach averaging 30
ESNs.
In spite of the realization of numerous tests, we can not
affirm that L2 -Boost with ESNs performs better than a single
well-initialized ESN (according the results presented in the RC
literature). However, the main advantage of the L2 -Boost with
weak ESNs is that the procedure is automatic and does not
require the computational effort of computing the spectra of
the hidden-hidden weight matrix. Additionally, the procedure
has a control parameter for the overfitting phenomena.
In a future work we will test the model using another
technique for decrease the generalization error, as well as on
a more number of benchmark problems. Additionally, we can
R EFERENCES
Laser test data
0.3
Baseline ESN
Boost: M=4
Boost: M=6
Boost: M=8
0.28
0.26
0.24
NMSE
0.22
0.2
0.18
0.16
0.14
0.12
0.1
6
7
8
9
Reservoir size
10
11
Figure 4: Test set accuracy of the Santa Fe Laser data. The
vertical axis of the graph shows the NMSE accuracy, and the
horizontal axis presents some values of the reservoir size. We
compare the accuracy of L2 -Boost: M = 4, 6 and 8 with a
baseline approach averaging 30 ESNs.
Freedman test data
0.24
Baseline ESN
Boost: M=4
Boost: M=5
Boost: M=6
0.22
0.2
NMSE
0.18
0.16
0.14
0.12
0.1
0.08
0.06
6
7
8
9
Reservoir size
10
11
Figure 5: Test set performance of the Freedman data set.
Size of the reservoir versus the NMSE accuracy. We compare
the accuracy of L2 -Boost: M = 4, 5 and 6 with a baseline
approach averaging 30 ESNs.
test the approach using another supervised learning tool for
the readout structure.
ACKNOWLEDGMENT
This work was supported within the framework of
the IT4Innovations Centre of Excellence project, reg. no.
CZ.1.05/1.1.00/02.0070 supported by Operational Programme
’Research and Development for Innovations’ funded by Structural Funds of the European Union and state budget of the
Czech Republic. Additionally, this article has been elaborated
in the framework of the project New creative teams in priorities
of scientific research, reg. no. CZ.1.07/2.3.00/30.0055.
[1] R. E. Shapire, “The strength of weak learnability,” Machine Learning,
vol. 5, no. 2, 1990.
[2] Y. Freund, “Boosting a weak learning algorithm by majority,” AT&T
Bell Laboratories, New Jersey, USA, Tech. Rep., January 1995.
[3] Y. Freund and R. E. Shapire, “Experiments with a new Boosting
Algorithm,” Machine Learning: Proceedings of Thirteenth International
Conference, 1996.
[4] J. Friedman, T. Hastie, and R. Tibshirani, “Additive Logistic Regression:
a Statistical View of Boosting,” The Annals of Statistics, vol. 28, no. 2,
pp. 337–407, 2000.
[5] P. Bühlmann and B. Yu, “Boosting with the L2 loss: Regression and
Classification,” Journal of the American Statistical Association, vol. 98,
2003.
[6] W. Maass, “Liquid State Machines: Motivation, Theory, and Applications,” in In Computability in Context: Computation and Logic in the
Real World. Imperial College Press, 2010, pp. 275–296.
[7] H. Jaeger, “The “echo state” approach to analysing and training recurrent
neural networks,” German National Research Center for Information
Technology, Tech. Rep. 148, 2001.
[8] J. Schmidhuber, D. Wierstra, M. Gagliolo, and F. Gomez, “Training
Recurrent Networks by Evolino,” Neural Networks, vol. 19, pp. 757–
779, 2007.
[9] M. Lukos̆evic̆ius and H. Jaeger, “Reservoir Computing Approaches to
Recurrent Neural Network Training,” Computer Science Review, pp.
127–149, 2009.
[10] M. Lukoševičius, “A Practical Guide to Applying Echo State Networks,”
in Neural Networks: Tricks of the Trade, ser. Lecture Notes in Computer
Science, G. Montavon, G. Orr, and K.-R. Müller, Eds. Springer
Berlin Heidelberg, 2012, vol. 7700, pp. 659–686. [Online]. Available:
http://dx.doi.org/10.1007/978-3-642-35289-8_36
[11] A. A. Ferreira, T. B. Ludermir, and R. R. B. De Aquino, “An
Approach to Reservoir Computing Design and Training,” Expert Syst.
Appl., vol. 40, no. 10, pp. 4172–4182, Aug. 2013. [Online]. Available:
http://dx.doi.org/10.1016/j.eswa.2013.01.029
[12] D. Opitz and R. Maclin, “Popular Ensemble Methods: an Empirical
study,” Journal of Artificial Intelligence Research, vol. 11, pp. 169–198,
1999.
[13] S. Basterrech, “An Empirical Study of L2-Boost with Echo State
Networks,” IEEE Intelligent Systems Design and Applications (ISDA),
December 2013.
[14] T. Hastie, R. Tibshirani, and J. Friedman, The elements of Statistical
Learning, ser. Spring series in statistics. New York, USA: SpringerVerlag, 2001.
[15] Y. Bengio and N. Chapados, “Extensions to metric-based model selection,” Journal of Machine Learning Research, vol. 3, pp. 1209–1227,
2003.
[16] J. H. Friedman, “Greedy Function Approximation: A Gradient Boosting
Machine,” Annals of Statistics, vol. 29, pp. 1189–1232, 2000.
[17] H. Schwenk and Y. Bengio, “Boosting Neural Networks,” Neural
Computation, vol. 12, no. 8, pp. 1869–1887, Aug. 2000.
[18] W. Press, S. Teukolsky, W. Vetterling, and B. Flannery, Numerical
Recipes in C, 2nd ed. Cambridge, UK: Cambridge University Press,
1992.
[19] J. Martens and I. Sutskever, “Learning Recurrent Neural Networks with
Hessian-Free Optimization,” in Proceeding of the 28th International
Conference on Machine Learning, 2011, pp. 1033–1040.
[20] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies
with gradient descent is difficult,” Neural Networks, IEEE Transactions
on, vol. 5, no. 2, pp. 157–166, 1994.
[21] R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty of training
recurrent neural networks,” Proceedings of the 30th International Conference on Machine Learning, vol. 28, pp. 37–48, 2013.
[22] W. Maass, T. Natschläger, and H. Markram, “Real-time computing
without stable states: a new framework for a neural computation based
on perturbations,” Neural Computation, pp. 2531–2560, november 2002.
[23] J. J. Steil, “Backpropagation-Decorrelation: online recurrent learning
with O(N) complexity,” In Proceedings of IJCNN’04, vol. 1, 2004.
[24] H. Jaeger, M. Lukos̆evic̆ius, D. Popovici, and U. Siewert, “Optimization
and applications of Echo State Networks with leaky-integrator neurons,”
Neural Networks, no. 3, pp. 335–352, 2007.
[25] B. Schrauwen, M. Wardermann, D. Verstraeten, J. J. Steil, and
D. Stroobandt, “Improving Reservoirs using Intrinsic Plasticity,” Neurocomputing, vol. 71, pp. 1159–1171, March 2007.
[26] S. Basterrech and G. Rubino, “Echo State Queueing Network: a new
Reservoir Computing learning tool,” IEEE Consumer Comunications &
Networking Conference (CCNC’13), January 2013.
[27] J. B. Butcher, D. Verstraeten, B. Schrauwen, C. R. Day, and P. W.
Haycock, “Reservoir Computing and Extreme Learning Machines for
Non-linear Time-series Data Analysis,” Neural Networks, vol. 38, pp.
76–89, feb 2013.
[28] A. Rodan and P. Tin̆o, “Minimum Complexity Echo State Network,”
IEEE Transactions on Neural Networks, pp. 131–144, 2011.
[29] M. Lukos̆evic̆ius, “On self-organizing reservoirs and their hierarchies,”
Jacobs University, Bremen, Tech. Rep. 25, 2010.
[30] J. Boedecker, O. Obst, N. M. Mayer, and M. Asada, “Studies on
Reservoir Initialization and Dynamics Shaping in Echo State Networks,”
in Proceedings of the 17th European Symposium On Artificial Neural
Networks (ESANN’09), Evere, Belgium, Apr. 2009, pp. 227–232.
[31] S. Basterrech, C. Fyfe, and G. Rubino, “Self-Organizing Maps and
Scale-Invariant Maps in Echo State Networks,” in Intelligent Systems
Design and Applications (ISDA), 2011 11th International Conference
on, nov. 2011, pp. 94–99.
[32] S. Dasgupta, F. Worgotter, and P. Manoonpong, “Information dynamics
based self-adaptive reservoir for delay temporal memory tasks,” Evolving
Systems, pp. 1–15, 2013.
[33] U. Huebner, N. B. Abraham, and C. O. Weiss, “The Santa Fe Time Series
Competition Data,” available at: http://goo.gl/6IKBb9, date of access: 12
September 2013.
[34] M. Hénon, “A two dimensional mapping with a strange attractor,”
Commun. Math. Phys., vol. 50, pp. 69–77, 1976.
[35] R. Hyndman, “Time series data library,” available at:
http://goo.gl/PZzReR, date of access: 12 September 2013.
[36] D. Verstraeten, B. Schrauwen, M. D’Haene, and D. Stroobandt, “An
experimental unification of reservoir computing methods,” Neural Networks, no. 3, pp. 287–289, 2007.
| 9cs.NE
|
Towards More Practical Linear Programming-based
Techniques for Algorithmic Mechanism Design⋆
Khaled Elbassioni1 , Kurt Mehlhorn2 , and Fahimeh Ramezani3
1
Masdar Institute of Science and Technology, Abu Dhabi, UAE
Max Planck Institute for Informatics, Campus E1 4, 66123, Saarbrucken, Germany
3
Department of Mathematics, University of Isfahan, Isfahan 81746-73441, Iran
[email protected], [email protected], [email protected]
arXiv:1408.1577v4 [cs.GT] 14 Jun 2016
2
Abstract. R. Lavi and C. Swamy (FOCS 2005, J. ACM 2011) introduced a general method
for obtaining truthful-in-expectation mechanisms from linear programming based approximation algorithms. Due to the use of the Ellipsoid method, a direct implementation of the
method is unlikely to be efficient in practice. We propose to use the much simpler and usually faster multiplicative weights update method instead. The simplification comes at the
cost of slightly weaker approximation and truthfulness guarantees.
1
Introduction
Algorithmic mechanism design is the art of designing and implementing the rules of a game to
achieve a desired outcome from a set of possible outcomes. Each player (agent) has a valuation that
assigns a value to each possible outcome. The desired outcome is the one that maximizes the sum
of the valuations; this sum is usually called social welfare. The players are assumed to be selfish:
they report valuations to the mechanism, which may differ from the true valuations. Players may
lie about their valuations in order to direct the mechanism into an outcome favorable to them.
The mechanism computes an outcome and payments for the players. The utility of a player is
her/his value of the outcome computed by the mechanism minus her/his payment charged by
the mechanism. The agents are interested in optimizing their personal utility. Social welfare and
personal utilities are determined with respect to the true valuations of the players, although they
are not public knowledge. The purpose of the payments is to incentivize the players to report their
true valuations. A mechanism is truthful if reporting the truth is a best strategy for each player
irrespective of the inputs provided by the other players. A mechanism is efficient if the outcome
and the payments can be computed in polynomial time. The underlying optimization problem is
the computation of an outcome maximizing social welfare given the valutions of the players.
If the underlying optimization problem can be efficiently solved to optimality, the celebrated
VCG mechanism (see, e.g., [NRTV07]) achieves truthfulness, social welfare optimization, and
polynomial running time. The computation of the outcome and the computation of the payments
requires to solve the underlying optimization problem to optimality.
Many optimization problems are NP-hard and hence are unlikely to have an exact algorithm
with polynomial running time. However, it might be possible to solve the problem approximately
in polynomial running time.
An example is the combinatorial auction problem. There is a set of m items to be sold to a set
of n players. The (reported) value of a set S of items to the i-th player is vi (S) with vi (∅) = 0 and
vi (S) ≤
is given
P to player i.
Pvi (T ) whenever S ⊆ T . Let xi,S be a 0-1 variable indicating that set S P
Then S xi,S ≤ 1 for every player i as at most one set can be given to i, and i PS;j∈S xi,S ≤ 1
for every item j as any item can be given away only once. The social welfare is i,S vi (S)xi,S .
The linear programming relaxation is obtained by replacing the integrality constraints for xi,s by
0 ≤ xi,S ≤ 1. Note that the number d of variables is exponential in the number of items, namely
d = n2m . The linear program is of the packing type,
√ i.e., if x is feasible and y ≤ x, then y is
feasible. For the combinatorial auction problem, O( n)-approximation algorithms exist and these
⋆
A preliminary version of these results was presented at SAGT 2015 [EMR15].
algorithms also
√ provide the corresponding integrality-gap-verifier (the definition is given below)
with α = 1/ n ([BKV05,KS98,Rag88]).
For many integer linear programming problems, approximation algorithms are known that
first solve the corresponding linear programming relaxation and then construct an integral solution
either by rounding or by primal-dual methods. Lavi and Swamy ([LS05,LS11]) showed that certain
linear programming based approximation algorithms for the social welfare problem can be turned
into randomized mechanisms that are truthful-in-expectation, i.e., reporting the truth maximizes
the expected utility of a player. The LS-mechanism is powerful (see [LS05,LS11,CEF10,HKV11]
for applications), but unlikely to be efficient in practice because of its use of the Ellipsoid method.
We show how to use the multiplicative weights update method instead. This results in simpler
algorithms at the cost of somewhat weaker approximation and truthfulness guarantees.
We next review the LS-mechanism. It applies to integer linear programming problems of the
packing type for which the linear programming relaxation can be solved exactly and for which an
α-integrality gap verifier is available. More precisely:
1. Let Q ⊆ Rd≥0 be a packing polytope, i.e., Q is a bounded convex polytope contained in the
non-negative orthant of d-dimensional space with the property that if y ∈ Q and x ≤ y then
x ∈ Q. The linear programming problem for Q asks to find for a given d-dimensional vector v
a point x∗ = argmaxx∈Q v T x.
2. We use QI := Q ∩ Zd for the set of integral points in Q. The integer linear programming
problem for QI asks to find for a given d-dimensional vector v a point x∗ = argmaxx∈QI v T x.
We use x1 , x2 , . . . , xj , . . . to denote the elements of QI and N for the index set of all elements
in QI .
3. An α-integrality-gap-verifier for QI for some α ∈ (0, 1] is an efficient algorithm that on input
v ∈ Rd and x∗ ∈ Q, returns an x ∈ QI such that
v T x ≥ αv T x∗ .
The mechanism consists of three main steps:
P
1. Let vi ∈ Rd≥0 , 1 ≤ i ≤ n, be the reported valuation of the i-th player and let v =
i vi
be the accumulated reported valuation. Solve the LP-relaxation, i.e., find a maximizer x∗ =
argmaxx∈Q v T x for the social welfare of the fractional problem, and determine the VCG prices4
p1 , . . . , pn . The allocation x∗ and the VCG-prices are a truthful mechanism for the fractional
problem.
P
j
2. Write α · x∗ as a convex combination of integral solutions in Q, i.e., α · x∗ =
j∈N λj x ,
P
λj ≥ 0, j∈N λj = 1, and xj ∈ QI . This step requires the α-integrality-gap-verifier.
3. Pick the integral solution xj with probability λj , and charge the i-th player the price pi ·
(viT xj /viT x∗ ). If viT x∗ = 0, charge zero.
The LS-mechanism approximates social welfare with factor α (is α-socially efficient ) and guarantees truthfulness-in-expectation, i.e., it converts a truthful fractional mechanism into an αapproximate truthful-in-expectation integral mechanism. With respect to practical applicability,
steps 1 and 2 are the two major bottlenecks. Step 1 requires solving n + 1 linear programs, one
for the fractional solution and one for each price; an exact solution requires the use of the Ellipsoid method (see e.g. [GLS88]), if the dimension is exponential. Furthermore, up to recently, the
only method known to perform the decomposition in Step 2 is through the Ellipsoid method. An
alternative method avoiding the use of the Ellipsoid method was recently given by Kraft, Fadaei,
and Bichler [KFB14]. We comment on their result in the next section.
1.1
Our Results
Our result concerns the design and analysis of a practical algorithm for the LS-scheme. We first
consider the case where the LP-relaxation of SWM (social welfare maximization) in Step 1 of the
4
pi =
P
j6=i
vjT (x̂ − x∗ ), where x̂ = argmaxx∈Q
P
j6=i
2
vjT x.
LS-scheme can be solved exactly and efficiently and then our problem reduces to the design of a
practical algorithm for Step 2. Afterwards, we consider the more general problem where only an
FPTAS for the LP-relaxation is available.
Convex Decomposition. Over the past 15 years, simple and fast methods have been developed
for solving packing and covering linear programs [BI06,GK95,GK07,Kha04,KY07,PST91,You01]
within an arbitrarily small error guarantee ε. These methods are based on the multiplicative
weights update (MWU) method [AHK12], in which a very simple update rule is repeatedly performed until a near-optimal solution is obtained. We show how to replace the use of the Ellipsoid
method in Step 2 by an approximation algorithm for covering linear programs. This result is the
topic of Section 2.
Theorem 1. Let ε > 0 be arbitrary. Given a fractional point x∗ ∈ Q, and an α-integrality-gap
verifier for QI , we can find a convex decomposition
X
α
· x∗ =
λj xj .
1 + 4ε
j∈N
The convex decomposition has size (= number of nonzero λj ) at most s(1 + ⌈ε−2 ln s⌉), where s
is the size of the support of x∗ (= number of nonzero components). The algorithm makes at most
s⌈ε−2 ln s⌉ calls to the integrality-gap-verifier.
Kraft, Fadaei, and Bichler [KFB14] obtained a related result independently. However, their construction is less efficient in two aspects. First, it requires O(s2 ε−2 ) calls of the integrality-gapverifyer. Second, the size of their convex decomposition might be as large as O(s3 ε−2 ). In the
combinatorial auction problem, s = n + m. Theorem 1 together with Steps 1 and 3 of the LS
scheme implies a mechanism that is truthful-in-expectation and has (α/(1 + 4ε))-social efficiency.
We leave it as an open problem whether the quadratic dependency of the size of the decomposition on ε can be improved5 .
Approximately Truthful-in-Expectation Mechanism. We drop the assumption that the fractional
SWM-problem can be solved optimally and assume instead that we have an FPTAS for it. We
assume further that the problem is separable, which means that the variables can be partitioned
into disjoint groups, one for each player, such that the value of an allocation for a player depends
only on the variables in his group, i.e,
vi (x) = vi (xi ),
where xi is the set of variables associated6 with player i. Formally, any outcome x ∈ Q ⊆ Rd can
be written as x = (x1 , . . . , xn ) where xi ∈ Rdi and d = d1 + . . . + dn . We further assume that for
each player i ∈ [n], there is a dominating allocation ui ∈ Q that maximizes his value for every
valuation vi , i.e.,
vi (ui ) = max vi (z),
(1)
z∈Q
for every vi ∈ Vi , where Vi denotes the possible valuations of player i. For the case of a combinatorial auction, the allocation ui allocates all items to player i.
ε5
Theorem 2. Let ε0 ∈ (0, 1/2]. Define ε = Θ( n04 ). Assuming that the fractional SWM-problem
has an FPTAS, is separable, and has a dominant allocation for every player i, and that there is an
α-integrality gap verifier for QI , there is a polynomial time randomized integral mechanism with
the following properties:
5
6
We remark that recent progress [ZO15,WRM15] on solving LPs of the packing/covering type has resulted
in an almost linear dependence of the running time on 1ε . However, the current methods do not work
in the oracle model and hence cannot be directly applied in our setting.
In the combinatorialPauction problem, xi comprises all variables xi,S . The value of an allocation x for
player i is given by S vi (S)xi,S .
3
(C1) No positive transfer, i.e., prices are nonnegative.
(C2) Individually rational with probability 1−ε0 , i.e., i.e., the utility of any truth-telling player
is non-negative with probability at least 1 − ε0 .
(C3) (1 − ε0 )-truthful-in-expectation, i.e., reporting the truth maximizes the expected utility of
a player up to a factor 1 − ε0 .
(C4) γ-socially efficient, where γ = α(1 − ε)(1 − ε0 )/(1 + 4ε).
Our mechanism is based on constructing a randomized fractional mechanism with properties
(C1) to (C3) and being (1 − ε)(1 − ε0)-socially efficient and then converting the mechanism into an
integral mechanism with the properties above. The conversion is simple. Let us assume that x is a
fractional allocation obtained from the fractional mechanism. We apply our convex decomposition
technique and Step 3 of the Lavi-Swamy mechanism to obtain an integral randomized mechanism
that satisfies (C1) to (C4).We show this result in Section 3.
Our fractional mechanism refines the one given in [DRVY11], where the dependency of ε on n
and ε0 is as ε = Θ(ε0 /n9 ). A recent experimental study of our mechanism on Display Ad Auctions
[EJ15] shows the applicability of our techniques in practice.
We leave it as an open problem whether the dependency of ε on ε0 and n can be improved.
On the Existence of an FPTAS for the Fractional SWM-Problem. We close the survey of our
results with a comment on the existence of an FPTAS for the fractional SWM-problem. Consider
a packing linear program
max cT x
subject to Ax ≤ b,
x ≥ 0,
is an m × n matrix with non-negative entries and c ∈ Rn>0 , b ∈ Rm
where A ∈ Rm×n
>0 are
≥0
positive vectors. We may assume that each column of A contains a non-zero entry as otherwise
the problem is trivially unbounded. For every κ ≥ 1 and weight vector z ∈ Rm
≥0 , let Oκ (z) denote
a κ-approximation oracle that returns a j such that
m
m
1 X zi aij ′
1 X zi aij
≤ κ · min
.
cj i=1 bi
bi
j ′ ∈[n] cj ′
i=1
Garg and Könemann [GK07] presented an algorithm that uses the oracle Oκ to construct an
approximation with a factor arbitrarily close to 1/κ. For κ = 1, their algorithm is an FPTAS.
What is the approximation oracle in case of the combinatorial auction problem? In this problem,
we have one constraint for each player and one constraint for each item. Let yi ≥ 0 be the weight
for agent i and zj ≥ 0 be the weight for item j. Then oracle O1 (y, z) requires to find the pair
X
1
(i, S) := argmin(k,T )
yk +
zj .
vk (T )
j∈T
P
In other words, for each k, one needs to find the set T which minimizes (yk + j∈T zj )/vk (T ). If
yk is interpreted as a fixed cost incurred by agent k and zj as the cost of item j, then T is the set
that minimizes the ratio of cost relative to value. For a simple-minded bidder who is interested
in the items in a subset T0 and no other item, i.e., vk (T ) = vk (T0 ) if T0 ⊆ T and vkP
(T ) = 0,
otherwise, T0 is the minimizer. Another simple case is additive valuations, i.e., vk (T ) = j∈T akj ,
P
where akj ≥ 0 is the value of item j for agent k. In this situation, vk 1(T ) yk + j∈T zj ≤ β for a
P
set T and a positive real β if and only if j∈T (βakj − zj ) ≥ yk and hence the minimal β for which
such a set T exists is readily determined by binary search on β.
2
A Fast Algorithm for Convex Decompositions
Let x∗ ∈ Q be arbitrary. Carr and Vempala [CV02] showed how to construct a convex combination
of points in QI dominating αx∗ using a polynomial number of calls to an α-integrality-gap-verifier
4
for QI . P
Lavi and Swamy [LS11] modified the construction to get an exact convex decomposition
αx∗ = i∈N λi xi for the case of packing linear programs. The construction uses the Ellipsoid
method. We show an approximate version that replaces the use of the Ellipsoid method by the
multiplicative weights update (MWU) method. For any ε > 0, we show how to obtain a convex
decomposition of αx∗ /(1 + ε). Let s be the number of non-zero components of x∗ . The size of the
decomposition and the number of calls to the α-integrality gap verifier are O(sε−2 ln s).
This section is structured as follows. We first review Khandekar’s FPTAS for covering linear
programs (Subsection 2.1). We then use it and the α-integrality gap verifier to construct, on input
x∗ ∈ Q, a dominating convex combination for αx∗ /(1 + 4ε) (Subsection 2.2). In Subsection 2.3,
we show how to convert a dominating convex combination into an exact convex decomposition.
Finally, in Subsection 2.4, we put the pieces together.
2.1
Khandekar’s Algorithm for Covering Linear Programs
Consider a covering linear program:
min cT x subject to
Ax ≥ b, x ≥ 0,
(2)
where A ∈ Rm×n
is an m × n matrix with non-negative entries and c ∈ Rn≥0 and b ∈ Rm
≥0 are
≥0
non-negative vectors. We assume the availability of a κ-approximation oracle for some κ ∈ (0, 1].
1 Pm zi aij
Oκ (z): Given z ∈ Rm
within a
≥0 , the oracle finds a column j of A that maximizes cj
i=1 bi
factor of κ:
m
m
1 X zi aij ′
1 X zi aij
≥ κ · max
cj i=1 bi
bi
j ′ ∈[n] cj ′
i=1
For an exact oracle κ = 1, Khandekar [Kha04] gave an algorithm which computes a feasible
solution x̂ to (2) such that cT x̂ ≤ (1 + 4ε)z ∗ where z ∗ is the value of an optimal solution. The
algorithm makes O(mε−2 log m) calls to the oracle, where m is the number of rows in A. There
are algorithms predating Khandekar’s work, see, for example, [K9̈7, Chapter 4]
Theorem 3 (Generalization of Khandekar’s algorithm to arbitrary κ ≤ 1). Let ε ∈ (0, 12 ]
and let z ∗ be the value of an optimum solution to (2). Procedure Covering(Oκ ) (see Algorithm 3
in Appendix I) terminates in at most m⌈ε−2 ln m⌉ iterations with a feasible solution x̂ of (2) of at
most m⌈ε−2 ln m⌉ positive components. At termination, it holds that
cT x̂ ≤
(1 + 4ε) ∗
z .
κ
(3)
For completeness, we give a proof of Khandekar’s result in Appendix I. The proof of Theorem
3 can be modified to give (see Appendix I):
Corollary 1. Suppose b = 1, c = 1, and we use the following oracle O′ instead of O in Algorithm
3:
T
T
O′ (z): Given z ∈ Rm
≥0 , such that 1 z = 1, the oracle finds a column j of A such that z A1j ≥ 1.
Then the algorithm terminates in at most m⌈ε−2 ln m⌉ iterations with a feasible solution x̂ having
at most m⌈ε−2 ln m⌉ positive components, such that 1T x̂ ≤ 1 + 4ε.
2.2
Finding a Dominating Convex Combination
Recall that we use N to index the elements in QI . We assume the availability of an α-integralitygap-verifier F for QI . We will use the results of the preceding section and show how to obtain for
any x∗ ∈ Q and any positive ε a convex composition of points in QI that covers αx∗ /(1 + 4ε).
Our algorithm requires O(sε−2 ln s) calls to the oracle, where s is size of the support of x∗ .
5
Theorem 4. Let ε > 0 be arbitrary. Given a fractional point x∗ ∈ Q and an α-integrality-gap
verifier F for QI , we can find a convex combination x̄ of integral points in QI such that
X
α
· x∗ ≤ x̄ =
λi xi .
1 + 4ε
i∈N
−2
The convex decomposition has size at most s⌈ε ln s⌉, where s is the number of positive entries
of x∗ . The algorithm makes at most s⌈ε−2 ln s⌉ calls to the integrality-gap verifier.
Proof. The task of finding the multipliers λi is naturally formulated as a covering LP ([CV02]),
namely,
X
min
λi
(4)
i∈N
s.t.
X
i∈N
X
i∈N
λi xij ≥ α · x∗j
for all j,
λi ≥ 1, λi ≥ 0.
λi ≥ 0.
Clearly, we can restrict our attention to the j ∈ S + := {j : x∗j > 0} and rewrite the constraint for
P
j ∈ S + as i∈N λi xij /(α ·x∗j ) ≥ 1. For simplicity of notation, we assume S + = [1..s]. We thus have
a covering linear program as in (2) with m := s + 1 constraints, n := |N | variables λi , right-hand
side b := 1, cost vector c := 1, and constraint matrix A = (aj,i ) (note that we use j for the row
index and i for the column index), where
i
xj /(αx∗j ) 1 ≤ j ≤ s, i ∈ N
aj,i :=
1
j = s + 1, i ∈ N
Thus we can apply Corollary 1 of Section 2.1, provided we can efficiently implement the required
oracle O′ . We do so using F .
Oracle O′ has is given a z̃) such that 1T z̃ = 1. Let us conveniently write z̃ = (w, z), where
Pj=1
w ∈ Rs≥0 , z ∈ R≥0 , and j=s wj + z = 1. Oracle O′ needs to find a column i such that z̃ T A1i ≥ 1.
Ps
In our case z̃ T A1i = j=1 wj xij /αx∗j + z, and we need to find a column i for which this expression
is at least one. Since z does not depend on i, we concentrate on the first term. Define
( w
j
for j ∈ S +
∗
Vj := αxj
0
otherwise.
Call algorithm F with x∗ ∈ Q and V := (V1 , . . .). F returns an integer solution xi ∈ QI such that
X wj
X
xij = V T xi ≥ α · V T x∗ =
wj ,
∗
αxj
+
+
j∈S
and hence,
j∈S
X
X wj
xij + z ≥
wj + z = 1.
∗
αxj
+
+
j∈S
j∈S
Thus i is the desired column of A.
|N |
It follows by Corollary 1 that Algorithm 3 finds a feasible solution λ′ ∈ R≥0 to the covering
′
LP (4), and a set Q′I ⊆ QI of vectors (returned by F ), such that λ′i > 0 only
P for i ′∈ N , where
λ
≤
(1
+ 4ε).
N ′ is the index set returned by oracle O′ and |N ′ | ≤ s⌈ε−2 ln s⌉. Also Λ := i∈N
P′ i
Scaling λ′i by Λ, we obtain a set of multipliers {λi = λ′i /Λ : i ∈ N ′ }, such that i∈N ′ λi = 1 and
X
α
x∗ .
λi xi ≥
1
+
4ε
′
i∈N
xij
+
We may assume
= 0 for all j ∈
/ S whenever λi > 0; otherwise simply replace xi by a vector in
+
which all components not in S are set to zero. By the packing property this is possible.
⊓
⊔
6
Algorithm 1 Changing a dominating convex decomposition into an exact decomposition
P
i
Require: A packing convex set Q and point x∗ ∈ Q and a convex combination
i∈N λi x of integral
∗
points in QI dominating x .
P
Ensure: A convex decomposition x∗ = i∈N ′ λi xi with xi ∈ QI .
P
1: while ∆j := i∈N λi xi − x∗j > 0 for some j do
2:
let i be such that λi xij > 0 and ∆j > 0Pfor some j.
3:
if there is a j such that λi xij > 0 and h∈N λh xh − λi 1j ≥ x∗ then
4:
replace xi by xi − 1j .
5:
else
6:
Among the indices j with xij > 0 and ∆j > 0, let k minimize ∆k /xik .
7:
let y be such that yj = xij , if ∆j = 0, and yj = 0, if ∆j > 0.
8:
change the lefthand side of (5) as follows: replace λi by λi − ∆k /xik and increase the coefficient
of y by ∆k /xik .
9:
end if
10: end while
2.3
From Dominating Convex Combination to Exact Convex Decomposition
We will show how to turn a dominating convex combination into an exact decomposition. The
construction is general and uses only the packing property. Such a construction seems to have
been observed in [LS05], but was not made explicit. Kraft, Fadaei, and Bichler [KFB14] describe
an alternative construction. Their construction may increase the size of the convex decomposition
(= number of non-zero λi ) by a multiplicative factor s and an additive factor s2 . In contrast, our
construction increases the size only by an additive factor s.
Theorem 5. Let x∗ ∈ Q be dominated by a convex combination
QI , i.e.,
X
λi xi ≥ x∗ .
P
i∈N
λi xi of integral points in
(5)
i∈N
Then Algorithm 1 achieves equality in (5). It increases the size of the convex combination by at
most s, where s is the number of positive components of x∗ .
Proof. Let S + = {j : x∗j > 0}. We may assume xij = 0 for all j 6∈ S + and all i ∈ N with λi > 0.
P
For j ∈ S + , let ∆j = i∈N λi xij − x∗j be the gap in the j-th component. If ∆j = 0 for all
j ∈ S + , we are done. Otherwise, choose j and i ∈ N such that ∆j > 0 and λi xij > 0.
Let 1j be the j-th unit vector. If, for some j with xij > 0 and ∆j > 0, replacing xi by
xi − 1j maintains feasibility, i.e., satisfies constraint (5), we perform this replacement. Since xi is
an integer vector in QI , the vector xi − 1j is nonnegative and, by the packing property, in QI .
The replacement decreases ∆j by λi and does not increase the number of nonzero λi .
Otherwise, ∆j < λi for all j with ∆j > 0 and xij > 0. Since xi is integral, we also have
∆j ≤ λi xij for all such j. Among the indices j with ∆j > 0 and xij > 0, let k minimize ∆k /xik .
Let y be such that yj = xij if ∆j = 0 and yj = 0 if ∆j > 0. Then y ∈ QI since Q is a packing
polytope. In the convex combination, replace
λi xi
by (λi −
∆k
∆k
) · xi + i · y.
i
xk
xk
k
≥ 0. Let ∆′j be the new gaps. Then clearly ∆′j = ∆j , if ∆j = 0. Consider any
Notice that λi − ∆
xik
j with ∆j > 0. Then
∆′j
∆k
= ∆j − i · xij =
xk
(
0
≥ (∆j −
7
∆j
)
xij
·
xij
if j = k
= 0 if j =
6 k.
The inequality in the second case holds since ∆k /xik ≤ ∆j /xij . We have decreased the number of
nonzero ∆j by one at the cost one additional nonzero λi . Thus the total number of vectors added
to the convex decomposition is at most s.
⊓
⊔
2.4
Fast Convex Decomposition
We are now ready to prove Theorem 1.
Proof of Theorem 1. Theorem 4 yields a convex combination of integer points of QI dominating
αx∗ /(1 + 4ε). The convex decomposition has size at most s⌈ǫ−2 ln s⌉, where s is the number of
positive entries of x∗ . The algorithm makes at most s⌈ǫ−2 ln s⌉ calls to the integrality-gap verifier.
Theorem 5 turns this dominating convex combination into an exact combination. It adds up to s
additional vectors to the convex combination.
⊓
⊔
3
Approximatly Truthful-in-Expectation Mechanisms
The goal of this section is to derive an approximate VCG-mechanism. We do not longer assume
that the fractional SWM-problem can be solved exactly, but instead assume that we have an
FPTAS for it. We will first design a randomized fractional algorithm (Theorem 6 in Subsection 3.1)
and then convert the fractional mechanism into an integral mechanism and prove Theorem 2 in
Subsection 3.2.
3.1
Approximately Truthful-in-Expectation Fractional Mechanisms
ε5
Theorem 6. Let ε0 ∈ (0, 1/2]. Define ε = Θ( n04 ). Assuming that the fractional SWM-problem has
an FPTAS, is separable, and has a dominant allocation for every player i, there is a polynomial
time randomized fractional mechanism (Algorithm 2) with the following properties:
(D1) No positive transfer, i.e., prices are nonnegative.
(D2) Individually rational with probability 1 − ε0 , i.e., the utility of any truth-telling player is
non-negative with probability at least 1 − ε0 .
(D3) (1 − ε0 )-truthful-in-expectation, i.e., reporting the truth maximizes the expected utility of
a player up to a factor 1 − ε0 .
(D4) γ-socially efficient, where γ = (1 − ε)(1 − ε0 ).
In order to present Algorithm 2 and prove Theorem 6, we introduce some notation and prove
some preliminary Lemmas. Let
X
Li :=
vj (uj ) and βi := εLi .
(6)
j6=i
Note that Li does not depend on the valuation of player i. Let A be an ε-approximation algorithm
for the LP relaxation of SWM. Note that A is polynomial time since the running time of an
FPTAS is polynomial in 1ε . We use A(v) to denote the outcome of A on input v; A(v) is a
fractional allocation in Q. In the following, we will apply A to different valuations which we
denote by v = (vi , v−i ), v̄ = (v̄i , v−i ), and v ′ = (0, v−i ). Here vi is the reported valuation of player
i, v̄i is his true valuation and vi′ = 0. We denote the allocation returned by A on input v (resp.,
v̄, v ′ ) by x (resp., x̄, x′ ). Note that x, x̄, x′ are fractional allocations.
We first bound the maximal change in social welfare induced by a change of the valuation of
the i-th player.
Lemma 1. Let ε ≥ 0 and let A be an ε-approximation algorithm which returns allocation x on
input vector v. Let x̂ ∈ Q be an arbitrary point, then
v(x) ≥ v(x̂) − βi − ε · vi (x̂)
for every i.
8
(7)
Algorithm 2 The mechanism M of Theorem 6. The vectors ui are defined as in (1) and the
quantities Li are defined in (6). The definitions of q0 , qj , active and inactive player are given in
the proof of Theorem 6.
Require: A valuation vector v, a packing convex set Q and an ε-approximation algorithm, where ε is as
Theorem 6.
Ensure: An allocation x ∈ Q and a payment p ∈ Rn satisfying (D1) to (D4).
1: Choose an index j ∈ {0, 1, . . . , n}, where 0 is chosen with probability q0 and j ∈ {1, . . . , n} is chosen
with probability qj = (1 − q0 )/n.
2: if j = 0 then
3:
Use ε-approximation algorithm A to compute an allocation x = (x1 , . . . , xn ) ∈ Q and compute
payments with payment rule (8). For all inactive i, change xi and pi to zero.
4: else
5:
For every 1 ≤ i ≤ n, set
i
′
xi = u , pi = η Li if i = j and i is active,
xi = ui , pi = 0
if i = j and i is inactive,
xi = 0, pi = 0
if i 6= j.
6: end if
7: return (x, p)
Proof. We have
v(x) ≥ (1 − ε) max v(x)
x∈Q
≥ (1 − ε)v(x̂)
X
= v(x̂) − ε ·
vj (x̂) − ε · vi (x̂)
j6=i
≥ v(x̂) − βi − ε · vi (x̂),
where the first inequality follows
P that A is an ε-approximation algorithm, and the
P from the fact
⊓
⊔
last inequality follows from ε j6=i vj (x̂) ≤ ε j6=i vj (uj ) = βi .
We use the following payment rule:
pi (v) := max{pVCG
(v) − βi , 0}
i
(8)
where
P
pVCG
(v) := v−i (x′ ) − v−i (x).
i
v−i (x) = j6=i vj (x), x = A(v) and x′ = A(0, v−i ). Observe the similarity in the definition of
pVCG
(v) to the VCG payment rule. In both cases, the payment is defined as the difference of
i
the total value of two allocations to the players different from i. The first allocation ignores the
influence of player i (x′ = A(0, v−i )) and the second allocation takes it into account (x = A(v)).
The difference to the VCG rule is that x′ and x are not true maximizers but are computed by an
ε-approximation algorithm.
Let Ui (v) = v̄i (x) − pi (v) be the utility of player i for bid vector v. Note that the value of the
allocation x = A(v) is evaluated with the true valuation v̄i of player i. Let Ui (v̄) = v̄i (x̄) − pi (v̄)
be the utility of player i for valuation vector v̄ = (v̄i , v−i ).
Lemma 2. Let ε ≥ 0 and let A be an ε-approximation algorithm. Let M0 be the mechanism with
allocation function A(v) and the payment rule (8). M0 is an individually rational mechanism with
no positive transfer, such that for all i,
Ui (v̄) ≥ Ui (v) − ε · v̄i (x) − 3βi .
9
(9)
Proof. By definition, pi (v) ≥ 0 for all v and all x; so the mechanism has no positive transfer. We
next address individual rationality. Assume pi (v̄) = pVCG
(v̄) − βi > 0, as otherwise Ui (v̄) ≥ 0. We
i
have
Ui (v̄) = v̄i (x̄) − pi (v̄)
= v̄i (x̄) − pVCG
(v̄) + βi
i
= v̄i (x̄) + v̄−i (x̄) − v̄−i (x′ ) + βi
= v̄(x̄) − v̄(x′ ) + v̄i (x′ ) + βi
≥ (1 − ε)v̄i (x′ ) ≥ 0,
where the first inequality follows from Lemma 1 with v = v̄ and x̂ = x′ .
Finally, we prove (9). We have v ′ (x′ ) = v−i (x′ ), v ′ (x) = v−i (x), and vi′ (x) = 0. Thus,
pVCG
(v) = v−i (x′ ) − v−i (x) = v ′ (x′ ) − v ′ (x) + ε · vi′ (x)
i
Applying Lemma 1 for v = v ′ and x̂ = x, we obtain
v ′ (x′ ) − v ′ (x) + ε · vi′ (x) ≥ −βi
Therefore,
pVCG
(v) + βi ≥ 0.
i
(10)
To see (9), we consider two cases:
Case 1: pi (v) = 0. Then using (10)
Ui (v̄) = v̄i (x̄) − 0 ≥ v̄i (x̄) − pVCG
(v̄) − βi .
i
Case 2: pi (v) = pVCG
(v) − βi .
i
Ui (v̄) = v̄i (x̄) − pi (v̄) = v̄i (x̄) − pVCG
(v̄) + βi ≥ v̄i (x̄) − pVCG
(v̄) − βi ,
i
i
where the last inequality follows from βi ≥ 0. Therefore, in both cases we have:
Ui (v̄) ≥ v̄i (x̄) − pVCG
(v̄) − βi .
i
and Lemma 1, we get
Now by using the definition of pVCG
i
Ui (v̄) ≥ v̄i (x̄) − pVCG
(v̄) − βi
i
= v̄i (x̄) + v̄−i (x̄) − v̄−i (x′ ) − βi
= v̄(x̄) − v̄−i (x′ ) − βi
≥ v̄(x) − βi − εv̄i (x) − v̄−i (x′ ) − βi
= v̄i (x) − pVCG
(v) − εv̄i (x) − 2βi
i
≥ v̄i (x) − pi (v) − βi − εv̄i (x) − 2βi
= Ui (v) − εv̄i (x) − 3βi .
⊓
⊔
In what follows we prove Theorem 6.
Proof of Theorem 6. Define q0 = (1 − εn0 )n , ε̄ = ε0 /2, and qj = (1 − q0 )/n for 1 ≤ j ≤ n. Let
η = ε̄(1 − q0 )2 /n3 , η ′ = η/qj , and ε = ηε̄(1 − q0 )/(8n). Then using7 q0 = (1 − εn0 )n ≥ 1 − ε0 and
q0 = (1 − εn0 )n ≤ 1 − ε0 /2, we get
ε50
ε̄2 (ε0 /2)3
ε̄2 (1 − q0 )3
ε̄2 ε30
ε50
=
≤
ε
=
ηε̄(1
−
q
)/(8n)
=
≤
=
,
0
128n4
8n4
8n4
8n4
16n4
7
Let f (x) = (1 − x/n)n − 1 + x. Then f ′ (x) = (1 − x/n)n−1 + 1 ≥ 0. Hence, for 0 ≤ x ≤ 1 and n ≥ 1,
the function is increasing and f (x) ≥ f (0) = 0.
10
as stated in the Theorem. Let Ui (v) = v̄i (x) − pi (v) be the utility of player i obtained by the
bi (v) = vi (x) − pi (v). Following [DRVY11], we call player
mechanism M0 of Lemma 2. Let further U
i active if the following two conditions hold:
bi (v) + ε̄qi vi (ui ) ≥ qi η ′ Li ,
U
q0
q0
i
vi (u ) ≥ ηLi .
(11)
(12)
Note that these conditions do not depend on the true valuation v̄. We denote by T = T (v) the set
of active players when the valuation is v = (v1 , . . . , vn ). Note that Li does not depend on vi . Thus
when we refer to conditions (11) and (12) for v̄, we replace v and x by v̄ and x̄ on the left side
and keep the right side unchanged. Non-negativity of payments is immediate from the definition
of mechanism M and Lemma 2. Moreover, the utility of a truth-telling bidder i can be negative
only if he/she is allocated in step 5, i.e., at most withP
probability qi . It follows that the mechanism
n
is individually rational with probability at least 1 − i=1 qi = q0 = (1 − εn0 )n ≥ 1 − ε0 .
Now we address truthfulness. Let us denote the expected utility of player i obtained from the
mechanism in Algorithm 2 on input v ∈ V by E[Ui′ (v)]. Assume j = 0 in Algorithm 2. We run
ε-approximation algorithm A on v to compute allocation x = (x1 , . . . , xn ). Then we change xi and
pi to zero for all inactive i. Let x
e be the allocation obtained in this way. The value for player i is
vi (e
x). When the i-th player is active, this value is equal to vi (x) because vi depends only on the
valuation in the i-th group (separability property). Therefore in this case his utility is Ui (v). So
we have that
q0 · Ui (v) + qi (v̄i (ui ) − η ′ Li ) if i ∈ T (v),
′
E[Ui (v)] =
(13)
qi v̄i (ui )
if i 6∈ T (v).
We first observe
E[Ui′ (v̄)] ≥ (1 − ε̄)qi · v̄i (ui ).
(14)
Indeed, the inequality is trivially satisfied if i 6∈ T (v̄). On the other hand, if i ∈ T (v̄), then (11)
bi (v̄) ≥ qi η ′ Li − ε̄v̄i (ui ) , therefore
implies Ui (v̄) = U
q0
E[Ui′ (v̄)] = q0 · Ui (v̄) + qi (v̄i (ui ) − η ′ Li )
qi ′
≥ q0 ·
η Li − ε̄v̄i (ui ) + qi (v̄i (ui ) − η ′ Li )
q0
= (1 − ε̄)qi · v̄i (ui ).
We now consider four cases:
Case 1: i ∈ T (v̄) ∩ T (v). Note that (12) for v̄ implies βi = εLi ≤
using assumption (1) that v̄i (x) ≤ v̄i (ui ), we have
εv̄i (ui )
.
η
Thus, by Lemma 2, and
4ε
3
Ui (v̄) ≥ Ui (v) − ε(1 + )v̄i (ui ) ≥ Ui (v) − v̄i (ui ).
η
η
Hence by using (13) and (15), we have
E[Ui′ (v)] = q0 · Ui (v) + qi (v̄i (ui ) − η ′ Li )
4ε
≤ q0 (Ui (v̄) + v̄i (ui )) + qi (v̄i (ui ) − η ′ Li )
η
4ε
= q0 Ui (v̄) + qi (v̄i (ui ) − η ′ Li ) +q0 v̄i (ui )
|
{z
}
η
E[Ui′ (v̄)]
= E[Ui′ (v̄)] + q0
4ε
v̄i (ui ).
η
11
(15)
Now applying (14) in the above inequality, we get
4ε
E[Ui′ (v)] ≤ E[Ui′ (v̄)] + q0 v̄i (ui )
η
q0
4ε
≤ 1+
E[Ui′ (v̄)]
(1 − ε̄)qi η
≤ (1 + ε̄)E[Ui′ (v̄)],
where the last inequality follows from the definition of ε. Note that (since q0 ≤ 1 and ε̄ ≤ 1/2)
ε
q0
4
4
1
ηε̄(1 − q0 ) 8
≤ε
≤
= ε̄.
(1 − ε̄)qi η
(1 − ε̄)qi η
8n
qi η
Case 2: i 6∈ T (v). By (14), we have
E[Ui′ (v)] = qi v̄i (ui ) ≤
Since,
1
1−ε̄
1
E[Ui′ (v̄)] ≤ (1 + ε0 )E[Ui′ (v̄)].
1 − ε̄
= 1 + ε̄(1 + ε̄ + ε̄2 + . . .) ≤ 1 + 2ε̄ = 1 + ε0 .
Case 3: i ∈ T (v) \ T (v̄) and (12) does not hold for v̄. Since Ui (v) ≤ v̄i (ui ), we have
E[Ui′ (v)] = q0 · Ui (v) + qi (v̄i (ui ) − η ′ Li ) ≤ (q0 + qi )v̄i (ui ) − qi η ′ Li < (q0 + qi − 1)v̄i (ui )
≤ 0 ≤ qi v̄i (ui ) = E[Ui′ (v̄)],
where the second inequality holds because (12) does not hold for v̄ and qi η ′ /η = 1.
Case 4: i ∈ T (v) \ T (v̄) and (12) holds for v̄. Then (11) does not hold for v̄ and hence
bi (v̄) <
Ui (v̄) = U
qi ′
η Li − ε̄v̄i (ui ) .
q0
(16)
Since (12) holds for v̄, we have (15). Hence by (14), (15) and (16) we have
E[Ui′ (v)] = q0 · Ui (v) + qi (v̄i (ui ) − η ′ Li )
4ε
≤ q0 Ui (v̄) + v̄i (ui ) + qi (v̄i (ui ) − η ′ Li )
η
4ε
≤ qi η ′ Li − qi ε̄v̄i (ui ) + v̄i (ui ) + qi (v̄i (ui ) − η ′ Li )
η
4ε
= (1 − ε̄)qi · v̄i (ui ) + v̄i (ui )
η
1
4ε
≤ 1+
E[Ui′ (v̄)]
(1 − ε̄)qi η
≤ (1 + ε̄)E[Ui′ (v̄)],
where the last inequality follows from the definition of ε (see Case 1).
We finally argue about the approximation ratio. Note that for i 6∈ T (v), one of the inequalities (11) or (12) does not hold. Also, Ui (v) ≥ 0 in this case since pi = 0, and hence
vi (ui ) < max{η, η ′ /ε̄}Li = η ′ Li /ε̄. Since A returns allocation x that is (1 − ε)-social efficiency
12
′
and8 qi − q0 n ηε̄ ≥ 0, it follows that for any v ∈ V, (recall x = A(v))
E[v(e
x)] = q0
X
vi (x) +
= q0
i∈[n]
qi vi (ui )
i∈[n]
i∈T (v)
X
X
vi (x) − q0
= q0 v(x) − q0
X
X
vi (x) +
i∈T
/ (v)
i
vi (u ) +
i∈T
/ (v)
X
X
qi vi (ui )
i∈[n]
qi vi (ui )
i∈[n]
X
η′ X
> q0 v(x) − q0
qi vi (ui )
Li +
ε̄
i6∈T (v)
′
= q0 v(x) − q0
η
ε̄
X X
i∈[n]
vj (uj ) +
i6∈T (v) j6=i
X
qi vi (ui )
i∈[n]
X
η′ X
qi vi (ui )
vj (uj ) +
≥ q0 v(x) − q0 n
ε̄
i∈[n]
j∈[n]
′
X
η
qi − q0 n
≥ q0 v(x) +
vi (ui )
ε̄
i∈[n]
≥ q0 (1 − ε) · max v(z)
z∈Q
≥ (1 − ε0 )(1 − ε) · max v(z).
z∈Q
⊓
⊔
3.2
Approximately Truthful-in-Expectation Integral Mechanisms
In this subsection, we derive a randomized mechanism M ′ which returns an integral allocation.
Let ε > 0 be arbitrary. First run Algorithm
P 2 to obtain x and p(v). Then compute a convex
α
α
x, which is 1+4ε
x = j∈N λxj xj . Finally with probability λxj (we use the
decomposition of 1+4ε
superscript x to distinguish the convex decompositions of different x) return the allocation xj
j
)
and charge the i-th player the price pi (v) vvii(x
(x) , if vi (x) > 0, and zero otherwise. We now prove
Theorem 2.
Proof of Theorem 2. Let M be a fractional randomized mechanism obtained in Theorem 6. Since
M has no positive transfer, M ′ does neither. M is individually rational with probability 1 − ε0 ,
therefore for any allocation x̄, we have v̄i (x̄) − pi (v̄) ≥ 0 with probability 1 − ε0 . So
v̄i (xl ) − pi (v̄)
v̄i (xl )
v̄i (xl )
= (v̄i (x̄) − pi (v̄))
≥ 0,
v̄i (x̄)
v̄i (x̄)
′′
hence M ′ is individually rational with probability 1 − ε0. Now we prove truthfulness. Let E[Ui (v̄)]
′′
be the expected utility of player i when she inputs her true valuation and let E[Ui (v)] be her
8
′
q0 n ηε̄ ≤
nη
qi ε̄
=
n2 ε̄(1−q0 )2
(1−q0 )ε̄n3
=
1−q0
n
= qi .
13
′′
expected utility when she inputs vi . Then by definition of E[Ui (v̄)], we have
"
#
X
′′
v̄i (xl )
x̄
l
E[Ui (v̄)] = Ex̄∼M(v̄)
λl v̄i (x ) − pi (v̄)
v̄i (x̄)
l∈N
!#
"
P
X
v̄i ( l∈N λx̄l xl )
x̄ l
= Ex̄∼M(v̄)
v̄i (
λl x ) − pi (v̄)
v̄i (x̄)
l∈N
α
α
= Ex̄∼M(v̄) [
v̄i (x̄) −
pi (v̄)]
1 + 4ε
1 + 4ε
α
Ex̄∼M(v̄) [v̄i (x̄) − pi (v̄)]
=
1 + 4ε
α
=
E[U ′ (v̄)]
1 + 4ε
α
E[U ′ (v)]
≥ (1 − ε0 )
1 + 4ε
α
Ex∼M(v) [v̄(x) − pi (v)]
= (1 − ε0 )
1 + 4ε
α
α
vi (x)
= (1 − ε0 )Ex∼M(v) [
v̄(x) − pi (v)
·
]
1 + 4ε
1 + 4ε vi (x)
"
#
X
vi (xl )
l
x
= (1 − ε0 )E
λl v̄i (x ) − pi (v)
vi (x)
l∈N
′
= (1 − ε0 )E[Ui (v)].
0 )(1−ε)
-socially efficient.
Taking expectation with respect to x shows that the mechanism is α(1−ε
1+4ε
"
#
"
#
X
X
α
x
l
x l
E[v (x )] = Ex∼M(v)
x)
λl v(x ) = Ex∼M(v) v(
λl x ) = Ex∼M(v) v(
1 + 4ε
l∈N
l∈N
α
α
Ex∼M(v) [v(x)] ≥
(1 − ε0 )(1 − ε) max v(z).
=
z∈Q
1 + 4ε
1 + 4ε
⊓
⊔
This completes the proof of Theorem 2.
References
[AHK12] Sanjeev Arora, Elad Hazan, and Satyen Kale. The multiplicative weights update method: a
meta-algorithm and applications. Theory of Computing, 8(6):121–164, 2012.
[BI06]
D. Bienstock and G. Iyengar. Approximating fractional packings and coverings in O(1/epsilon)
iterations. SIAM J. Comput., 35(4):825–854, 2006.
[BKV05] P. Briest, P. Krysta, and B. Vöcking. Approximation techniques for utilitarian mechanism
design. In STOC, pages 39–48, 2005.
[CEF10] G. Christodoulou, K. Elbassioni, and M. Fouz. Truthful mechanisms for exhibitions. In WINE,
pages 170–181, 2010.
[CV02]
Robert D. Carr and Santosh Vempala. Randomized metarounding. Random Struct. Algorithms,
20(3):343–352, 2002.
[DRVY11] Shaddin Dughmi, Tim Roughgarden, Jan Vondrák, and Qiqi Yan. An approximately truthfulin-expectation mechanism for combinatorial auctions using value queries. CoRR, abs/1109.1053,
2011.
[EJ15]
K. Elbassioni and M. Jha. On the power of combinatorial bidding in web display ads. In ICIW,
pages 46–55, 2015.
[EMR15] Khaled Elbassioni, Kurt Mehlhorn, and Fahimeh Ramezani. Towards More Practical Linear
Programming-Based Techniques for Algorithmic Mechanism Design. In Martin Hoefer, editor,
Algorithmic Game Theory, volume 9347 of Lecture Notes in Computer Science, pages 98–109.
Springer Berlin Heidelberg, 2015.
14
[GK95]
M.D. Grigoriadis and L.G. Khachiyan. A sublinear-time randomized approximation algorithm
for matrix games. Operations Research Letters, 18(2):53 – 58, 1995.
[GK07]
N. Garg and J. Könemann. Faster and simpler algorithms for multicommodity flow and other
fractional packing problems. SIAM J. Comput., 37(2):630–652, 2007. a preliminary version
appeared in FOCS 1998.
[GLS88] M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial Optimization. Springer, New York, 1988.
[HKV11] M. Hoefer, T. Kesselheim, and B. Vöcking. Approximation algorithms for secondary spectrum
auctions. In SPAA, pages 177–186, 2011.
[K9̈7]
J. Könemann. Fast combinatorial algorithms for packing and covering problems. Master’s
thesis, Fachbereich Informatik, Universität des Saarlandes, 1997.
[KFB14] D. Kraft, S. Fadaei, and M. Bichler. Fast convex decomposition for thruthful social welfare
approximation. In WINE, pages 120–132, 2014.
[Kha04] R. Khandekar. Lagrangian Relaxation Based Algorithms for convex Programming Problems.
PhD thesis, Indian Institute of Technology, Delhi, India, 2004.
[KS98]
S. Kolliopoulos and C. Stein. Approximating disjoint-path problems using greedy algorithms
and packing integer programs. In IPCO, pages 153–168, 1998.
[KY07]
C. Koufogiannakis and N.E. Young. Beating simplex for fractional packing and covering linear
programs. In FOCS, pages 494–504, 2007.
[LS05]
R. Lavi and C. Swamy. Truthful and near-optimal mechanism design via linear programming.
In FOCS, pages 595–604, 2005.
[LS11]
Ron Lavi and Chaitanya Swamy. Truthful and near-optimal mechanism design via linear programming. J. ACM, 58(6):25, 2011.
[NRTV07] N. Nisan, T. Roughgarden, E. Tardos, and V. V. Vazirani. Algorithmic Game Theory. Cambridge University Press, 2007.
[PST91] S.A. Plotkin, D.B. Shmoys, and É. Tardos. Fast approximation algorithms for fractional packing
and covering problems. In FOCS, pages 495–504, 1991.
[Rag88] P. Raghavan. Probabilistic construction of deterministic algorithms: Approximating packing
integer programs. J. Comput. Syst. Sci., 37(2):130–143, 1988.
[WRM15] Di Wang, Satish Rao, and Michael W. Mahoney. Unified acceleration method for packing and
covering problems via diameter reduction. CoRR, abs/1508.02439, 2015.
[You01] N.E. Young. Sequential and parallel algorithms for mixed packing and covering. In FOCS,
pages 538–546, 2001.
[ZO15]
Z. Allen Zhu and L. Orecchia. Nearly-linear time positive LP solver with faster convergence
rate. In STOC 2015, Portland, OR, USA, June 14-17, 2015, pages 229–236, 2015.
Appendix I: Khandekar’s Algorithm for Covering Linear Programs
Consider a covering linear program:
min cT x subject to
Ax ≥ b, x ≥ 0,
(17)
where A ∈ Rm×n
is an m × n matrix with non-negative entries and c ∈ Rn≥0 and b ∈ Rm
≥0 are
≥0
non-negative vectors. We assume the availability of a κ-approximation oracle for some κ ∈ (0, 1].
1 Pm zi aij
Oκ (z): Given z ∈ Rm
within a
≥0 , the oracle finds a column j of A that maximizes cj
i=1 bi
factor of κ:
m
m
1 X zi aij
1 X zi aij ′
≥ κ · max
(18)
cj i=1 bi
bi
j ′ ∈[n] cj ′
i=1
We use Ai to denote the i-th row of A. Algorithm 3 constructs vectors x(t) ∈ Rn≥0 , for t =
0, 1, . . . , until M (t) := mini∈[m] Ai x(t)/bi becomes at least T := lnε2m . Define the active list at time
t by L(t) := {i ∈ [m] : Ai x(t − 1)/bi < T }. For i ∈ L(t), define
pi (t) := (1 − ε)Ai x(t−1)/bi ,
15
(19)
Algorithm 3 Covering(Oκ)
m
Require: a covering system (A, b, c) given by a κ−approximation oracle Oκ , where A ∈ Rm×n
≥0 , b ∈ R>0 ,
c ∈ Rn
,
and
an
accuracy
parameter
ε
∈
(0,
1/2]
>0
(1+4ε) ∗
T
Ensure: A feasible solution x̂ ∈ Rn
z
≥0 to (2) s.t. c x̂ ≤
κ
ln m
1: x(0) := 0; t := 0; and T := ε2
2: while M (t) < T do
3:
t := t + 1
4:
Let j(t) := Oκ (p(t)/kp(t)k1 )
5:
xj(t) (t) := xj(t) (t − 1) + δ(t) and xj (t) = xj (t − 1) for j 6= j(t)
6: end while
x(t)
7: return x̂ = M
(t)
and set pi (t) = 0 for i 6∈ L(t). At each time t, the algorithm calls the oracle with the vector
zt = p(t)/kp(t)k1 , and increases the variable xj(t) by
δ(t) :=
bi
min
i∈L(t) and ai,j(t) 6=0
ai,j(t)
,
(20)
where j(t) is the index returned by the oracle.
Proof of Theorem 3. Note that the RHS
P of (18) is positive for our choice of zt since every row
of A contains a non-zero entry and hence i∈L(t) pi (t)aij(t) /(bi cj(t) ) > 0. This conclude that there
exist at least one i ∈ L(t) which ai,j(t) is non zero and thus δ(t) > 0 always. In each iteration, some
entry of x is increased and hence the values Ai x(t)/bi are non-decreasing. Thus L(t + 1) ⊆ L(t)
for all t. At the end, we scale x(t) by M (t) to guarantee feasibility.
Let 1j denote the j-th unit vector of dimension n and B ∈ Rm×m be a diagonal matrix
with entries bii = bi . Feasibility is obvious since we scale by M (t). The bound on the number of
iterations is also obvious since in each iteration at least one of the Ai x/bi increases by one and we
remove i from the active list once Ai x/bi reaches T . We conclude that the number of iterations
is bounded by m⌈T ⌉. Let t0 be the number of iterations, i.e., vectors x(0), x(1), . . . , x(t0 ) are
defined and M (t0 − 1) < TP≤ M (t0 ). In the t-th iteration exactly one entry of x is increased by
δ(t) and hence 1T x(t0 ) = 1≤t≤t0 δ(t) and Ai x(t)/bi ≤ Ai x(t − 1)/bi + 1 for i ∈ L(t). To show
(3), we analyze the decrease of kp(t)k1 . Let t ≤ t0 . Then
X
i∈L(t)
(1 − ε)Ai x(t)/bi =
=
X
i∈L(t)
X
i∈L(t)
≤
X
i∈L(t)
(1 − ε)Ai x(t−1)/bi +δ(t)Ai 1j(t) /bi
pi (t)(1 − ε)δ(t)Ai 1j(t) /bi
pi (t)(1 − εδ(t)Ai 1j(t) /bi )
(using (20) conclude that δ(t)Ai 1j(t) /bi ≤ 1 and
(1 − ε)x ≤ 1 − εx for all ε ∈ [0, 1), x ∈ [0, 1])
!
εδ(t)p(t)T B −1 A1j(t)
= kp(t)k1 1 −
kp(t)k1
≤ kp(t)k1 e
p(t)T
−εδ(t) kp(t)k B −1 A1j(t)
1
16
since 1 − x ≤ e−x .
(21)
By using L(t + 1) ⊆ L(t), we have
kp(t + 1)k1 =
≤
X
i∈L(t+1)
X
i∈L(t)
(1 − ε)Ai x(t)/bi
(1 − ε)Ai x(t)/bi
(22)
and hence applying inequalities (21) and (22) we get,
kp(t + 1)k1 ≤
X
i∈L(t)
T
(1 − ε)Ai x(t)/bi ≤ kp(t)k1 e
Let i0 ∈ L(t0 ) be arbitrary. Then
(1 − ε)Ai0 x(t0 )/bi0 ≤
X
i∈L(t0 )
p(t)
B −1 A1j(t)
−εδ(t) kp(t)k
1
.
(23)
(1 − ε)Ai x(t0 )/bi0
≤ kp(t0 )k1 e
≤ kp(0)k1 e
p(t )T
−εδ(t0 ) kp(t0 )k B −1 A1j(t0 )
−ε
0
P
1
p(t)T
1≤t≤t0
δ(t) kp(t)k B −1 A1j(t)
1
,
where the second inequality uses (21) for t = t0 and the third inequality uses (23) for 0 ≤ t < t0 .
Taking logs and using kp(0)k1 = m, we conclude that
Ai0 x(t0 )/bi0 · ln(1 − ε) ≤ ln m − ε
X
δ(t)
1≤t≤t0
p(t)T
B −1 A1j(t)
kp(t)k1
(24)
P
We next relate the objective value cT x(t0 ) = 1≤t≤t0 cj(t) δ(t) at time t0 to the optimal value z ∗
by the following claim.
T
P
p(t)T
0)
−1
Claim.
A1j(t) ≥ κ·c zx(t
.
∗
1≤t≤t0 δ(t) kp(t)k1 B
Proof. Let x∗ ∈ Rn≥0 be an optimal solution to (2). Since x∗ is feasible, B −1 Ax∗ ≥ 1, and thus for
any t,
p(t)T B −1 Ax∗ ≥ p(t)T 1 = kp(t)k1 .
By the choice of the index j(t), we have that
∗
T
1
T −1
A1j(t)
cj(t) p(t) B
∗
j ∈ [n]. Since z = c x , we conclude further
X
cj x∗j p(t)T B −1 A1j(t)
z ∗ p(t)T B −1 A1j(t) =
≥
1
T −1
A1j
cj κp(t) B
for all
j∈[n]
=
X
j∈[n]
≥
X
j∈[n]
cj ·
cj(t) ∗
x p(t)T B −1 A1j(t)
cj(t) j
cj ·
cj(t) ∗
x κp(t)T B −1 A1j
cj j
= κcj(t) p(t)T B −1 Ax∗
≥ κcj(t) kp(t)k1 .
Multiplying both sides of this inequality by δ(t)/kp(t)k1 and summing up over 1 ≤ t ≤ t0 finishes
the proof of the claim.
⊓
⊔
Using the claim above, we deduce from (24)
Ai0 x(t0 )/bi0 · ln(1 − ε) ≤ ln m − ε ·
17
κ · cT x(t0 )
z∗
Dividing both sides by M (t0 ), arranging, and using M (t0 ) ≥ T = (ln m)/ε2 , Ai0 x(t0 )/bi0 ≤
Ai0 x(t0 − 1)/bi0 + 1 ≤ T + 1, and
ln
1
1−ε
ε
≤ 1 + 2ε, valid for all ε ∈ (0, 12 ], we obtain
1
ln 1−ε
κ · cT x̂
κ · cT x(t0 )
Ai x(t0 )/bi0
ln m
=
≤
· 0
+
z∗
M (t0 )z ∗
ε
M (t0 )
ε · M (t0 )
T +1
+ ε ≤ 1 + 4ε.
≤ (1 + 2ε)
T
⊓
⊔
Proof of Corollary 1. Recall (24):
Ai0 x(t0 )/bi0 · ln(1 − ε) ≤ ln m − ε
X
δ(t)
p(t)T −1
B A1j(t) .
kp(t)k1
X
δ(t)
p(t)T
A1j(t) .
kp(t)k1
1≤t≤t0
With assumption b = 1, we have,
Ai0 x(t0 ) · ln(1 − ε) ≤ ln m − ε
1≤t≤t0
The vector zt = p(t)/kp(t)k1 satisfies 1T zt = 1. Apply oracle O′ with input vector zt , it returns
p(t)T
A1j(t) ≥ 1. Thus, we have
index j(t) such that we have kp(t)k
1
Ai0 x(t0 ) · ln(1 − ε) ≤ ln m − ε · 1T x(t0 ).
Proceeding as in the proof of Theorem 3, we get the result.
18
⊓
⊔
| 8cs.DS
|
arXiv:1802.06173v1 [math.ST] 17 Feb 2018
Matrix variate Birnbaum-Saunders distribution under
elliptical models
José A. Dı́az-Garcı́a
∗
Universidad Autónoma de Chihuahua
Facultad de Zootecnia y Ecologı́a
Periférico Francisco R. Almada Km 1, Zootecnia
33820 Chihuahua, Chihuahua, México
E-mail: [email protected]
Francisco J. Caro-Lopera
Departament of Basic Sciences
Universidad de Medellı́n
Medellı́n, Colombia
E-mail: [email protected]
Abstract
This paper solves the open problem for an elliptical matrix variate version of the well
known univariate Birnbaum and Saunders distribution. A generalisation based on
a matrix transformation, instead of the independent element by element representation, was elusive in literature since the Gaussian univariate version was proposed
in 1969. New results on linear structures were needed to derived the matrix variate
distribution. A number of special cases are studied and some basic properties are
found.
1
Introduction
Some restricted situations in statistics accepts that the hypothesis for an experimental or
observational data can be based on univariate tests. But the complex reality involves multivariate or matrix variate decision problems with several dependent variables that must be
considered simultaneously.
This is the source of motivation to generalise the univariate probability distributions
into the multivariate or matrix variate cases. However, the few known successful generalisations have required the creation of advanced mathematics, usually out of the scope of
popular books and journals of high impact in decision sciences. Moreover, sometimes the
leading techniques and the representations are not unique, then the associated theoretical
relations enlarge the problem. For example, the extension of the univariate chi-squared into
the so termed matrix variate Wishart distribution required the construction in the 50’s of
the theory of zonal polynomials of matrix arguments. Three different methods constructed
∗ Corresponding
author
Key words. Matrix multivariate distributions, linear structures, random matrices, elliptical distributions,
Birnbaum-Saunders distribution.
2000 Mathematical Subject Classification. 62E15; 60E05; 15B52
1
the non singular central distribution: the singular value decomposition (SVD), the QR decomposition and the polar factorisation; see for example James (1954), Roy (1957) and Herz
(1955), respectively. But their use in the computation of the joint latent roots distribution
in the central case took more than 50 years after their apparition, and the relations among
the densities are still unclear today. In fact, the theory for the extension to the non central
Wishart was so advanced that the created invariant polynomials of several matrix arguments
of Davis (1979) cannot be calculated even in this time of super computers.
Now, there are two ways to generalise a univariate random variable into a random vector
or a random matrix:
i) Define the random vector or random matrix element by element.
ii) Propose a matrix transformation equivalent to the univariate function that defines the
random variable Y .
For example, suppose a random variable Y with a chi-square distribution of n degrees
of freedom, that is, Y ∼ χ2 (n). Now, assume that the random vector Z ∈ ℜn follows an ndimensional normal distribution, with vector mean E(Z) = 0n ∈ ℜn and covariance matrix
Cov(Z) = In ; where 0n is a vector of zeros and In is the n × n identity matrix. In notation,
d
d
Z ∼ Nn (0n , In ). Then, we know that Y = kZk2 ; where = holds for equally distributed and
kZk denotes the Euclidiana norm of the vector Z. So, we ask for the multivariate version of
the random variable Y .
Applying the first method (element-to-element) we can proceed as follows: let Z ∼
Nn (0n , In ), such that n = n1 + n2 and Z′ = (Z′1 , Z′2 ), Z1 ∈ ℜn1 and Z2 ∈ ℜn2 . Then,
define the random variables Yi = kZi k2 , i = 1, 2 and the vector random Y′ = (Y1 , Y2 )′ .
Thus Y is said to have a bidimensional χ2 distribution, such that, Yi ∼ χ2 (ni ), i = 1, 2; see
Libby and Novick (1982). Using the same technique we can get the multivariate version
of the random variable Y . Sometimes the matrix case can be obtained directly from multivariate (vector) version: let Y ∈ ℜn×m and define v = vec(Y), where vec(Y) denotes de
vectorisation of the matrix Y, then the distribution of Y is obtained from the distribution
of the random vector v.
Alternatively, the matrix variate extension of the χ2 -distribution became more popular that the addressed multivariate case. Assume n independent Zi ∼ Nm (0, Σ), with
Cov(Zi ) = Σ and i = 1, . . . , n. Define the random matrix
V=
n
X
Zi Z′i .
i=1
If n ≥ m, then V is positive definite (V > 0) and V is said to have a Wishart distribution.
Otherwise, if n < m, then V is positive semidefinite, (V ≥ 0) and V is said to have a pseudoWishart distribution. These facts are denoted as V ∼ Wm (n, Σ) and V ∼ PW m (n, Σ),
respectively, see Srivastava and Khatri (1979) and Muirhead (2005), among many others.
Note that, if m = 1, Σ is an scalar, say σ 2 , then nV/σ 2 ≡ nY /σ 2 ∼ χ2 (n). However,
note that it is impossible to obtain a vector version of the distribution of V from the
Wishart distribution. In addition, not all elements vij (in V) follow a χ2 -distribution. Note
that, if the univariate random variable is a function of square or square root operators,
the corresponding matrix variate version via a matrix transformation, must be a random
square matrix; moreover, in general, it must be a random symmetric matrix, see Cadet
(1996), Olkin and Rubin (1964), Muirhead (2005), and references therein. Then, the matrix
version includes the univariate case, but the vector version cannot be derived, moreover, the
elements of the matrix does not follow the original univariate distribution.
2
However, a matrix variate version via element-to-element has not order constraint. The
vectorial and the univariate cases can be derived directly from the matrix case, and all the
elements of the matrix have as marginal distribution, the original univariate distribution,
see Chen and Novick (1984) and Libby and Novick (1982).
Extreme unusual cases allows equivalence among the vector version and the elementto-element representation and the matrix transformation. This occurs in the multivariate t-distribution; which is a consequence of a property for the t-distribution family, see
(Kotz and Nadarajah, 2004, p. 2, 4). A random p-dimensional vector with distribution t
can be defined in two ways:
y1 /S −1 + µ1
y2 /S −1 + µ2
−1
S
Y
+
µ
=
,
..
.
−1
y
/S
+
µ
p
p
t=
νS 2
∼ χ2 (ν) and Y ∼ Np (0, Σ);
with
2
σ
−1/2
Y + µ,
W
with W ∼ Wp (ν + p − 1, Σ) and Y ∼ Np (0, νIp ).
with (W1/2 )2 = W and µ : p × 1 a constant vector.
Nevertheless, this unusual property is not fulfilled in the matrix case. Consider the
sample t1 , . . . , tn of a multivariate population with t distribution, and consider the matrix
T = (t1 · · · tn ) : p × n, then
−1 T
T
S Y1 + µT1
..
−1
= S Y + M,
.
S −1 YnT + µTn
νS 2
with
∼ χ2 (ν) and Y ∼ Np×n (0, Σ ⊗ In )
σ2
T=
or
T
T −1/2
Y1 W
+ µT1
..
−1/2
Y + M,
=W
.
T
−1/2
T
Yn W
+ µn
with W ∼ Wp (ν + p − 1, Σ) and Y ∼ Np×n (0, ν(Ip ⊗ In ))
where M = (µ1 · · · µn ) : p × n, and Y = (Y1 · · · Yn ). But the random matrix T does not
have the same distribution under the above two representations, even when their rows have
the same distribution. In the first representation, T has a matrix multivariate t−distribution
and under the second one it has a matricvariate T −distribution, see (Kotz and Nadarajah,
2004, p. 2, 4). Also, note that the matricvariate T −distribution cannot be obtained from
the matrix-variate t−distribution, and vice versa.
Now we focus on the distribution of this work. An important lifetime model was introduced by Birnbaum and Saunders (1969) in the context of a problem of material fatigue.
The so called Birbaum-Saunders distribution is a lifetime model for fatigue failure caused
under cyclic loading and assumed that the failure is due to the development and growth of
a dominant crack. A more general derivation was provided by Desmond (1985) based on a
biological model.
The original univariate random variable was supported by a normal distribution, then
the so termed Gaussian Birnbaum-Saunders random variable T is the distribution of
!2
r
α
α 2
T =β
Z+
Z +1 ,
(1)
2
2
3
where Z ∼ N (0, 1). We shall denotes this fact as T ∼ BS(α, β), where α > 0 is the shape
parameter, and β > 0 is both scale parameter an the median value of the distribution. Then,
the inverse relation establishes that if T ∼ BS(α, β), then
s
r !
1
T
β
Z=
∼ N (0, 1)
(2)
−
α
β
T
Dı́az-Garcı́a and Leiva-Sánchez (2005, 2006) propose a generalization of the Birnbaum-Saunders
distribution, replacing the Gaussian hypothesis in (2) by a symmetric distribution, i.e. they
assume that Z ∼ E(0, 1, h). We recall that the density function of Z ∼ E(0, 1, h) is defined
as fZ (z) = h(z 2 ), for z ∈ ℜ. Therefore, (1) defines the so called generalized BirnbaumSaunders distribution, which shall be denoted by T ∼ GBS(α, β; h). Note the long delay
to appear the elliptical univariate version. In fact, the element-to-element elliptical matrix variate version of Birnbaum and Saunders (1969) was published in Caro-Lopera et al.
(2012), it demanded the develop of some theory to connect the Hadamard product and the
usual matrix product. In the same direction, Caro-Lopera and Dı́az-Garcı́a (2016) studied
the so termed diagonalization matrix and applied it in another matrix representation of the
element-to-element generalized matrix variate Birnbaum-Saunders distribution. Moreover,
Sánchez et al. (2015) performed estimation for the matrix parameters of that type matrix
case. But a matrix transformation has been so elusive in literature and no clue to derive such
transformation can be inferred or proposed from the existing extensions of another families
matrix variate distributions. This paper develops a particular theory of linear structures in
order to derive the matrix variate Birnbaum-Saunders distribution under elliptical models.
Some basic properties are studied and the expected corollaries are derived.
2
Preliminary results
Some properties and definitions in matrix variate elliptical theory are summarised below.
A detailed study of this family of distributions is presented in Fang and Zhang (1990)
and Gupta and Varga (1993), among many others authors. This section also presents the
published element-to-element representations of the Birnbaum-Saunders distribution. First,
some results and notations about the new required matrix algebra are considered, see Magnus
(1988).
2.1
Notation
For our purposes: if A ∈ ℜn×m denotes a matrix, this is, A have n rows and m columns,
then A′ ∈ ℜm×n denotes its transpose matrix, and if A ∈ ℜn×n has an inverse, it shall be
denoted by A−1 ∈ ℜn×n . An identity matrix shall be denoted by I ∈ ℜn×n , to specified the
size of the identity, we will use In . A null matrix shall be denoted as 0 ≡ 0n×m ∈ ℜn×m .
For all matrix A ∈ ℜn×m exist A− ∈ ℜm×n which is termed Moore-Penrose inverse. The
eigenvalues of A ∈ ℜn×n are the roots of the equation |λIn − A| = 0. A ∈ ℜn×n is a
symmetric matrix if A = A′ and if all their eigenvalues are positive then A is positive definite
matrix, which shall be denoted as A > 0. Given a definite positive matrix A ∈ ℜm×m ,
2
there exist a positive definite matrix A1/2 ∈ ℜm×m such that A = A1/2 , which is
termed positive definite root matrix. If A ∈ ℜn×m is writing in terms of its m columns,
A = (A1 , A2 , . . . , Am ), Aj ∈ ℜn , j = 1, 2 . . . , m, vec(A) ∈ ℜnm denotes the vectorisation
of A, moreover, vec′ (A) = (vec(A))′ = (A′1 , A′2 , . . . , A′m ). Let A ∈ ℜr×s and B ∈ ℜn×m ,
then A ⊗ B ∈ ℜsn×rm denotes its Kronecker product. For A, B, and C, matrices of suitable
matrices orders, we have
vec(ABC) = (C′ ⊗ A) vec B.
(3)
4
The commutative matrix Knm ∈ ℜnm×nm is the matrix with the property that Knm vec A =
vec A′ , for every matrix A ∈ ℜn×m . In addition for A ∈ ℜm×m , and B ∈ ℜp×q ,
Kpm (A ⊗ B) = (B ⊗ A)Kqn .
(4)
Let A ∈ ℜn×m with rank r. Denote the r nonzero (hence positive) eigenvalues of A′ A as
λ1 , . . . , λr . Then the eigenvalues of the matrix Knm (A′ ⊗ A) are given by
p
λi , i = 1, . . . , r and ± λi λj , i < j.
(5)
2.2
Matrix variate distribution.
Definition 2.1. Is said that Y ∈ ℜn×m has a matrix variate elliptically contoured distribution if its density with respect to the Lebesgue measure is given by:
dFY (Y) =
1
|Σ|n/2 |Θ|m/2
h tr (Y − µ)′ Θ−1 (Y − µ)Σ−1 (dY),
where µ ∈ ℜn×m , Σ ∈ ℜm×m , Θ ∈ ℜn×n , Σ > 0 and Θ > 0 and (dY) is the Lebesgue
The function h : ℜ → [0, ∞) is termed the generator function and satisfies
Rmeasure.
∞ mn−1
2
u
h(u
)du < ∞. Such a distribution is denoted by Y ∼ En×m (µ, Θ ⊗ Σ, h), see
0
Gupta and Varga (1993).
When µ = 0n×m , Σ = Im and Θ = In , such distribution is termed matrix variate
symmetric distribution and shall be denoted as Y ∼ En×m (0, Inm , h).
Finally, observe that this class of matrix variate distributions includes normal, contaminated normal, Pearson type II and VI, Kotz, logistic, power exponential, and so on; these
distributions have tails that are weighted more or less, and/or they have greater or smaller
degree of kurtosis than the normal distribution.
Now, from Dı́az-Garcı́a and Leiva-Sánchez (2005, 2006) if T ∼ GBS(α, β, h), then
1
t
β
t−3/2 (t + β)
√
h 2
+ − 2 dt, t > 0.
(6)
dFT (t) =
α
β
t
2α β
√
Alternatively, let V = T , with dt = 2vdv, then under a symmetric distribution, (2) can be
rewrite as
√
V
1
β
√ −
.
(7)
Z=
α
V
β
Moreover
1 + βv −2
1 v2
β
√
h 2
dFV (v) =
(8)
+ 2 − 2 dv, v > 0.
α
β
v
α β
Among other authors, Dı́az-Garcı́a and Domı́nguez Molina (2006) proposed a multivariate
version (vector version) defined element-to-element of the density function (6), this is, they
assumed that z ∼ En (0n , In ; h) and define the change of variable
2
s
2
1
1
αi zi + 1 , αi > 0, βi > 0, i = 1, . . . , n;
ti = βi αi zi +
2
2
Then, the distribution DFt (t1 , . . . , tn ) of t = (t1 , . . . , tn )′ ∈ ℜn+ , termed multivariate generalised Birnbaum-Saunders distribution, is given by
!
! " n
# ^
n −3/2
n
X 1 ti
(ti + βi )
1 Y ti
βi
√
(9)
dti
= n
h
+
−2
2
α2 βi
ti
αi βi
i=1
i=1 i
i=1
5
V
where
denotes the exterior product, see (Muirhead, 2005, Section 2.1.1, p. 50). This
fact is denoted as t ∼ GBS n (α, β; h), with α = (α1 , . . . , αn )′ and β = (β1 , . . . , βn )′ . This
distribution was studied in detail by Dı́az-Garcı́a and Domı́nguez Molina (2007) when β1 =
· · · = βn = β and α1 = · · · = αn = α.
As we mentioned above, the matrix variate generalised Birbaum-Saunders distribution
can be obtained from the multivariate case by defining the vector r = vec T, where T ∈
ℜn×m and
2
s
2
1
1
tij = βij αij zij +
+ 1 , αij > 0, βij > 0,
αij zij
2
2
with i= 1, . . . , n;,
j = 1, . . . , m. Then, assuming that Z ∼ En×m (0, Inm , h) the distribution
i=1,...,n
DFT (tij )j=1,...,m , tij > 0 is given by
=
1
2nm
−3/2
n Y
m
Y
tij (tij + βij )
p
αij βij
i=1 j=1
#
n X
m
X
βij
tij
1
h
+
−2
α2
βij
tij
i=1 j=1 ij
! "
n ^
m
^
i=1 j=1
dtij
!
(10)
which is denoted as T ∼ GBS n×m (A, B; h), with A = (αij ), and B = (βij ), i = 1, . . . , n;,
j = 1, . . . , m.
This distribution was found and studied by Caro-Lopera et al. (2012). Their main
goal was to construct a matrix representation of the matrix variate generalised BirnbaumSaunders distribution. Using the diagonalisation operator, the Hadamard product and partition theory, they propose two matrix representations of the density function (10). In terms
of the diagonalisation matrix, an alternative matrix representation of the generalised matrix variate Birnbaum-Saunders distribution was proposed by Caro-Lopera and Dı́az-Garcı́a
(2016).
3
Generalised matrix variate Birnbaum-Saunders distribution
This section derives the main result of the paper, the so termed generalised matrix variate
Birnbaum-Saunders distribution via a matrix transformation. First we find the distribution
of a random matrix V ∈ ℜn×m , such that T = V′ V has a generalised matrix variate
Birnbaum-Saunders distribution; i.e. we shall get the matrix variate version of the density
function defined by (8). Then, some special cases are found and, finally the generalised
matrix variate Birnbaum-Saunders distribution is obtained.
Theorem 3.1. Assume that Z ∼ En×m (0n×m , Inm , h) and consider the following matrix
version of (7)
′
Z = V∆−1 − V − ∆ Ξ−1 ,
(11)
where Ξ ∈ ℜm×m , Ξ > 0 is the shape parameter matrix; ∆ ∈ ℜm×m , ∆ > 0 is the scale
parameter matrix, such that ∆ is the positive definite square root of β ( ∆2 = β); and
V ∈ ℜn×m such that T = V′ V, with rank(V) = m ≤ n. Then the density function of V is
′
∆−1 ⊗ In + Kmn V − ⊗ ∆V−
dFV (V) =
|Ξ|n
i
h
−1
(12)
× h tr Ξ−2 ∆−1 V′ V∆−1 + ∆ (V′ V) ∆ − 2Im (dV).
6
Proof. Write
′
Z = V∆−1 − V − ∆ Ξ−1 .
The density of Z is
(13)
dFZ (Z) = h(tr Z′ Z)(dZ).
To determine the explicit form of the volume element (dZ) under the change of variable (13),
we will proceed using the theory developed by Magnus (1988). Hence, taking differentials
in (13) we have
dZ = dV∆−1 − (−V− dVV− )′ ∆ Ξ−1
=
′
′
dV∆−1 Ξ−1 + V − dV′ V − ∆Ξ−1 .
By vectorisation, we get
d vec Z =
=
=
′
′
′
V − ∆Ξ−1 ⊗ V − d vec V′
⊗ In d vec V +
′
′
′
′
Ξ −1 ∆ −1 ⊗ In + Ξ −1 ∆′ V− ⊗ V − Knm d vec V
′
′
′
Ξ −1 ⊗ In ∆ −1 ⊗ In + Kmn V − ⊗ ∆′ V− d vec V.
∆−1 Ξ−1
′
Therefore, given that Ξ and ∆ are symmetric matrices,
′
′
′
(dV)
(dZ) =
Ξ −1 ⊗ In ∆ −1 ⊗ In + Kmn V − ⊗ ∆′ V−
′
= |Ξ|−n ∆−1 ⊗ In + Kmn V − ⊗ ∆V− (dV).
Thus,
dFV (V) =
′
∆−1 ⊗ In + Kmn V − ⊗ ∆V−
|Ξ|n
h
i
i′ h
′
′
V∆−1 − V − ∆ Ξ−1 (dV).
×h tr V∆−1 − V − ∆ Ξ−1
The required result is obtained by noting that rank (V− V) = rank(V− ) = rank(V) = m =
′
′
′
rank(V′ V) and V− V ∈ ℜm×m and V V ∈ ℜm×m then, V′ V − = (V− V) = V− V = Im
′
−
′
−1
and (V V) = (V V) .
Now, some special cases of the density function (12) are found.
Corollary 3.1. Under the hypotheses of Theorem 3.1 we obtain the following results:
i) Assuming that Ξ = αIm , α > 0, the density function (12) turns into
′
∆−1 ⊗ In + Kmn V − ⊗ ∆V−
dFV (V) =
αnm
1
−1
×h 2 tr ∆−1 V′ V∆−1 + ∆ (V′ V) ∆ − 2Im (dV).
α
7
√
ii) If we take β = ∆′ ∆ = βIm , that is ∆ = βIm , β > 0, then we have
m
m
Y
Y
p
(1 + βλi ) abs
1 ± β λi λj
dFV (V) =
i=1
i<j
β nm/2 |Ξ|n
1 ′
−1
′
−2
V V + β (V V) − 2Im (dV).
×h tr Ξ
β
Where λi , i = 1, . . . , m, are the eigenvalues of the matrix (V′ V)−1 and abs(·) denotes
de absolute value, furthermore,
m
m
m
Y
Y
Y
p
p
p
abs
1 ± β λi λj =
1 + β λi λj abs
1 − β λi λj .
i<j
i<j
i<j
√
iii) Analogously, taking Ξ = αIm and ∆ = βIm , we obtain
m
m
Y
Y
p
(1 + βλi ) abs
1 ± β λi λj
dFV (V) =
i=1
i<j
(α2 β)nm/2
1 ′
−1
′
V V + β (V V) − 2Im (dV).
β
√
iv) Finally, defining m = n = 1, then Ξ ≡ α, ∆ ≡ β, and (V′ V)−1 ≡ v −2
1 + βv −2
1 v2
β
√
h 2
dFV (v) =
+ 2 − 2 dv
α
β
v
α β
1
×h 2 tr
α
Proof. i) The result follows by using tr(aI)B = a tr B and |aIm | = am .
ii) First recall that: if the eigenvalues of the matrix A are λ1 , . . . , λm the
√ eigenvalues of the
matrix (I ± A) are 1 ± λ1 , . . . , 1 ± λm . Also, observe that if ∆ = βIm ,
′
′
∆−1 ⊗ In + Kmn V − ⊗ ∆V− = β −1/2 Inm + βKmn V − ⊗ V−
′ −1
Thus, by (5), if λi are the eigenvalues of V V
, we have that the eigenvalues of
′
p
−
−
are βλi , i = 1, . . . , m and ±β λi λj , i < j. Hence,
βKmn V ⊗ V
′
Inm + βKmn V − ⊗ V−
=
m
Y
i=1
m
Y
p
(1 + βλi ) abs
1 ± β λi λj ,
i<j
where abs(·) denotes de absolute value, and
m
m
m
Y
Y
Y
p
p
p
1 + β λi λj abs
1 − β λi λj .
abs
1 ± β λi λj =
i<j
i<j
i<j
The result now is evident.
8
iii) The result follows from i) and ii).
iv) The result is a consequence of iii), taking m = n = 1 and observing that (V′ V)−1 ≡
v −2 .
The next result considers β = βIm as a particular case of generalised matrix variate
Birnbaum-Saunders distribution via matrix transformation . This fact shall be denoted as
T ∼ GBS m (n, Ξ, β, h), where Ξ ∈ ℜm×m , Ξ > 0 is the shape parameter matrix, ∆ ∈ ℜm×m ,
∆ > 0 such that ∆ is the positive definite square root of the scale parameter matrix β, i.e.
∆2 = β.
Theorem 3.2. We say that T ∼ GBS m (n, Ξ, βIm , h) if its density function is given by
m
m
Y
Y
p
(1 + βλi ) abs
π nm/2
1 ± β λi λj
i=1
dFT (T) =
i<j
2m Γ
m
[n/2]β nm/2 |Ξ|n
|T|(n−m−1)/2
1
−2
−1
× h tr Ξ
T + βT − 2Im (dT).
β
(14)
Where λi , i = 1, . . . , m are the eigenvalues of T−1 , and Γm [·] denotes de multivariate gamma
function, see (Muirhead, 2005, Definition 2.1.10, p.61),
Γm [a] = π m(m−1)/4
m
Y
i=1
Γ[a − (i − 1)/2], [Re(a) > (m − 1)/2]
and Re(·) denotes de real part of the argument.
Proof. By analogy with the univariate case, Equations (6), (7) and (8), starting from (11),
we shall say that the positive definite matrix T = V′ V have a generalised matrix variate
Birnbaum-Saunders distribution. First consider that β = βIm then the density function of
the corresponding random matrix V is given in Corollary 3.1 ii), moreover,
m
m
Y
Y
p
(1 + βλi ) abs
1 ± β λi λj
dFV (V) =
i<j
i=1
β nm/2 |Ξ|n
1 ′
−1
−2
′
×h tr Ξ
V V + β (V V) − 2Im (dV).
β
Define T = V′ V with V = H1 R, where H1 ∈ ℜn×m is a semi-orthogonal matrix, that is
H′1 H1 = Im and R ∈ ℜm×m is a real upper triangular matrix. Then T = V′ V = R′ R. Note
that in the considered QR factorisation (V = H1 R), the matrices H1 and R are defined in
(Mathai, 1997, p. 100), see Theorem 2.9 and the preceding discussion for the unique choice
of H1 and R. Then by (Muirhead, 2005, Theorem 2.1.14, p. 66)
(dV) = 2−m |T|(n−m−1)/2 (dT)(H′1 dH1 )
Thus, the joint density function of T and H1 is
m
m
Y
Y
p
(1 + βλi ) abs
1 ± β λi λj
dFT,H1 (T, H1 ) =
i=1
i<j
2m β nm/2 |Ξ|n
9
1
×h tr Ξ−2
T + βT−1 − 2Im |T|(n−m−1)/2 (dT)(H′1 dH1 ),
β
where λi , i = 1, . . . , m are the eigenvalues of T−1 . In this case, see (Mathai, 1997, p. 117),
Z
(H′1 dH1 ) =
H1
π mn/2
.
Γm [n/2]
Then the required result is obtained.
Corollary 3.2. A particular case of Theorem 3.2 is given next:
i) If Ξ = αIm , then we have that
π nm/2
dFT (T) =
m
Y
i=1
m
Y
p
(1 + βλi ) abs
1 ± β λi λj
i<j
2m Γ
nm/2
2
m [n/2] (α β)
×h
1
tr
α2
|T|(n−m−1)/2
1
T + βT−1 − 2Im
β
(dT).
ii) If m = n = 1 then Ξ ≡ α and β ≡ β and
dFT (t) =
t−3/2 (t + β)
t
β
1
√
+ − 2 dt.
h 2
α
β
t
2α β
Proof. This follows directly from Theorem 3.2, just recall that Γ[1/2] =
√
π.
Next, we propose a general expression for the density function of T, when T ∼ GBS m (n, Ξ, β, h).
Theorem 3.3. Suppose that T ∼ GBS m (n, Ξ, β, h), T ∈ ℜm×m , T > 0, Ξ ∈ ℜm×m , Ξ > 0
and β ∈ ℜm×m , β > 0; where β = (∆)2 , ∆ is the positive definite square root of β. Then
m
m
Y
Y
p
(1 + δi ) abs
1 ± δi δj
π nm/2
dFT (T) =
i=1
i<j
|T|(n−m−1)/2
2m Γm [n/2]|β|n |Ξ|n
×h tr Ξ−2 ∆−1 T∆−1 + ∆T−1 ∆ − 2Im (dT),
where δi are the eigenvalues of βT−1 .
Proof. Assume that U ∼ GBS m (Ξ, βIm , h). By (14), taking β = 1 and considering the
change of variable T = ∆U∆, such that (dU) = |∆|−(m+1) (dT) = |β|−(m+1)/2 (dT). Then
we have that U = ∆−1 T∆−1 , with
|∆−1 T∆−1 |(n−m−1)/2 = |β|−(n−m−1)/2 |T|(n−m−1)/2
and δi , i = 1, . . . , m are the eigenvalues of (∆−1 T∆−1 )−1 = ∆T−1 ∆ or βT−1 . And the
required result follows.
The Gaussian case is obtained by taking Z as a matrix variate normal distribution in
Theorem 3.1. Hence, from Theorem 3.3 we obtain the matrix variate Birnbaum-Saunders
distribution, which shall be denoted as T ∼ BS m (n, Ξ, β)
10
Corollary 3.3. Suppose that T ∼ BS m (n, Ξ, β), T ∈ ℜm×m , T > 0, Ξ ∈ ℜm×m , Ξ > 0
and β ∈ ℜm×m , β > 0; where β = (∆)2 , ∆ is the positive definite square root of β. Then
m
m
Y
Y
p
(1 + δi ) abs
1 ± δi δj
dFT (T) =
i=1
i<j
2m(n+2)/2 Γ
m [n/2]|β|
n |Ξ|n
|T|(n−m−1)/2
1
× etr − Ξ−2 ∆−1 T∆−1 + ∆T−1 ∆ − 2Im (dT),
2
where δi are the eigenvalues of βT−1 .
Proof. In the Gaussian case we just take h(z) = (2π)−nm/2 etr(−z/2). Then the proof
follows straightforwardly from Theorem 3.3.
Some basic properties of the generalised matrix variate Birnbaum-Saunders distribution
are summarised in the next result.
Theorem 3.4. Assume that T ∼ GBS m (n, Ξ, β, h), then
i) if S = T−1 , its density function is
π nm/2
m
Y
i=1
dFS (S) =
m
Y
√
(1 + ρi ) abs
1 ± ρi ρj
i<j
|S|−(n+m+1)/2
2m Γm [n/2]|β|n |Ξ|n
×h tr Ξ−2 ∆−1 S−1 ∆−1 + ∆S∆ − 2Im (dS),
where ρi are the eigenvalues of βS.
ii) The density function of Y = C′ TC is,
m
m
Y
Y
√
(1 + ̺i ) abs
1 ± ̺i ̺j
π nm/2
i=1
i<j
|Y|−(n+m+1)/2
2m Γm [n/2]|β|n |C|n |Ξ|n
i
h
′
×h tr Ξ−2 (∆B) −1 Y(∆B)−1 + ∆BY−1 (∆B)′ − 2Im (dY),
dFY (Y) =
where ̺i are the eigenvalues of C′ βCY−1 .
Proof. The corresponding proofs are obtained by considering the following Jacobians, see
(Muirhead, 2005, Section 2.1.1).
i) Let S = T−1 , then (dT) = |S|−(m+1) (dS) and
ii) Let Y = C′ TC, then (dT) = |C|−(m+1) (dY),
respectively.
We note that the general and particular distributions here derived can be easily computed. A number of challenges for future work involve the parameter estimation, the moments, the associated Wishart type distributions, etc.. An interesting application refers to
their use in shape theory under skew distributions. We must say that at present, the statistical shape theory has been supported only by symmetric distributions. Scatters of certain
classical landmark data suggest multi modal clusters and skew elliptical distribution. Some
special transformation such as QR, SVD and polar are of interest for applications, but at
first sight the kernel of the transformed matrix distribution ask for a new theory of integrals
and properties of zonal polynomials.
11
4
Conclusions
This paper revised the literature and techniques for extending the univariate Gaussian
Birnbaum-Saunders distribution into a matrix version. The addressed distribution was
defined more than 50 years ago and its generalisation in the matrix case was of interest
for applications in complex data model by skew and multimodal distributions. Literature
tells that a number of efforts were given to obtained a matrix version, but only the restricted element-to-element representation was achieved. This paper obtains a consistent
matrix representation via a matrix transformation which allows to express the density in
terms of the random matrix instead of their elements. The new representation demanded
the construction of some aspects in linear structures. This work also sets the matrix variate
Birnbaum-Saunders distribution into a general class of elliptical distribution, which allows
a number dependent relations and models. From the algebraical point of view, the new
family of generalised matrix variate Birnbaum-Saunders distribution is set into the classical
matrix theory and allows a number of possible transformations to research in future. Finally,
the work establishes several particular matrix variate distributions and particular low order
cases which allows posterior works for inference and related aspects.
References
Birnbaum, Z.W., Saunders, S.C. 1969. A New Family of Life Distributions. J. Appl. Probab.
6, 637-52.
Cadet, A., 1996. Polar coordinates in Rnp ; Application to the computation of the Wishart
and beta laws. Sankhyā A 58, 101–113.
Caro-Lopera, F. J., Leiva, V., Balakrishnan, N. 2012. Connection between the Hadamard
and matrix products with an application to matrix-variate Birnbaum-Saunders distributions. J. Multivariate Anal. 104 (1), 126-139.
Caro-Lopera, F. J., Dı́az-Garcı́a, J. A. 2016. Diagonalization matrix and its application in
distribution theory. Statistics, 50 (4), 870-880.
Chen, J. J., Novick, M. R., 1984. Bayesian analysis for binomial models with generalized
beta prior distributions. J. Educational Statist. 9, 163–175.
Davis, A. W. 1979. Invariant polynomials with two matrix arguments, extending the zonal
polynomials: Applications to multivariate distribution theory. Annals of the Institute of
Statistical Mathematics A31, 465–485.
Desmond, A., 1985. Stochastic models of failure in random enviorments. Canad. J. Statist.
13, 171–183.
Dı́az-Garcı́a, J. A., Domı́nguez Molina, J. R. 2006. Some generalisations of BirnbaumSaunders and sinh-normal distributions. Int. Math. Forum. 1 (35), 1709-1727.
Dı́az-Garcı́a, J. A., Domı́nguez Molina, J. R. 2007. A new family of life distributions for
dependent data: Estimation. Comp. Statist. Data Anal. 51(12), 5927-5939.
Dı́az-Garcı́a, J. A., Leiva-Sánchez, V., 2005. A new family of life distributions based on
elliptically contoured distributions. J. Statist. Plan. Inf. 128(2), 445-457.
Dı́az-Garcı́a, J. A., Leiva-Sánchez, V., 2006. Erratum to ”A new family of life distributions
based on the elliptically contoured distributions”: [J. Statist. Plann. Inference 128(2)
(2005) 445-457]. J. Statist. Plan. Inf. 137(4), 1512-1513.
12
Fang, K. T., Zhang, Y. T., 1990. Generalized Multivariate Analysis, Science Press, SpringerVerlag, Beijing.
Gupta A. K., Varga, T., 1993 Elliptically Contoured Models in Statistics, Kluwer Academic
Publishers, Dordrecht.
Herz, C. S., 1955. Bessel Functions of Matrix Argument. Annals of Mathematics 61(3):
474–523.
James, A. T. 1954. Normal Multivariate Analysis and the Orthogonal Group. Ann Math
Statist 25(1): 40–75.
Kotz, S., and Nadarajah, S., 2004. Multivariate t Distributions and Their Applications.
Cambridge University Press, United Kingdom.
Libby, D. L., Novick, M. R., 1982. Multivariate generalized beta distributions with applications to utility assessment. J. Educational Statist. 7, 271–294.
Magnus, J. R, 1988. Linear Structures. Charles Griffin & Company Ltd., London.
Mathai, A. M, 1997. Jacobian of Matrix Transformations and Functions of Matrix Argument.
World Scinentific, Singapore.
Muirhead, R. J., 2005. Aspects of Multivariate Statistical Theory. John Wiley & Sons, New
York.
Olkin, I., Rubin, H., 1964. Multivariate beta distributions and independence properties of
Wishart distribution. Ann. Math. Statist. 35, 261–269. Correction 1966, 37(1), 297.
Roy, S. N. 1957). Some aspects of multivariate analysis. John Wiley & Sons, Inc., New York.
Sánchez, L., Leiva, V., Caro-Lopera, F., Cysneiros, F.J 2015. On matrix-variate BirnbaumSaunders distributions and their estimation and application. Brazilian J. Probab. Statist.
29(4), 790–812
Srivastava, M. S., Khatri, C. G., 1979. An Introduction to Multivariate Analysis. NorthHolland Publ., Amsterdam.
13
| 10math.ST
|
PRACTICAL SKETCHING ALGORITHMS
FOR LOW-RANK MATRIX APPROXIMATION∗
arXiv:1609.00048v2 [cs.NA] 2 Jan 2018
JOEL A. TROPP† , ALP YURTSEVER‡ , MADELEINE UDELL§ , AND VOLKAN CEVHER¶
Abstract. This paper describes a suite of algorithms for constructing low-rank approximations
of an input matrix from a random linear image of the matrix, called a sketch. These methods can
preserve structural properties of the input matrix, such as positive-semidefiniteness, and they can
produce approximations with a user-specified rank. The algorithms are simple, accurate, numerically
stable, and provably correct. Moreover, each method is accompanied by an informative error bound
that allows users to select parameters a priori to achieve a given approximation quality. These claims
are supported by numerical experiments with real and synthetic data.
Key words. Dimension reduction; matrix approximation; numerical linear algebra; randomized
algorithm; single-pass algorithm; sketching; streaming algorithm; subspace embedding.
AMS subject classifications. Primary, 65F30; Secondary, 68W20.
1. Motivation. This paper presents a framework for computing structured lowrank approximations of a matrix from a sketch, which is a random low-dimensional
linear image of the matrix. Our goal is to develop simple, practical algorithms that
can serve as reliable modules in other applications. The methods apply for the real
field (F = R) and for the complex field (F = C).
1.1. Low-Rank Matrix Approximation. Suppose that A ∈ Fm×n is an arbitrary matrix. Let r be a target rank parameter where r min{m, n}. The computational problem is to produce a low-rank approximation  of A whose error is
comparable to a best rank-r approximation:
(1.1)
kA − ÂkF ≈
min
rank(B)≤r
kA − BkF .
The notation k · kF refers to the Frobenius norm. We explicitly allow the rank of Â
to exceed r because we can obtain more accurate approximations of this form, and
the precise rank of  is unimportant in many applications. There has been extensive
research on randomized algorithms for (1.1); see Halko et al. [19].
1.2. Sketching. Here is the twist. Imagine that our interactions with the matrix
A are severely constrained in the following way. We construct a linear map L :
Fm×n → Fd that does not depend on the matrix A. Our only mechanism for collecting
data S about A is to apply the linear map L :
(1.2)
S := L (A) ∈ Fd .
We refer to S as a sketch of the matrix, and L is called a sketching map. The number
d is called the dimension or size of the sketch.
∗ Dated
30 August 2016. Revised 13 January 2017 and 6 June 2017 and 4 September 2017.
Funding: JAT and MU were supported in part by ONR Award N00014-11-1002 and the Gordon
& Betty Moore Foundation. MU was also supported in part by DARPA Award FA8750-17-2-0101.
AY and VC were supported in part by the European Commission under Grant ERC Future Proof,
SNF 200021-146750, and SNF CRSII2-147633.
† California Institute of Technology, Pasadena, CA ([email protected]).
‡ École Polytechnique Fédéral de Lausanne, Lausanne, Switzerland ([email protected]).
§ Cornell University, Ithaca, NY ([email protected]).
¶ École Polytechnique Fédéral de Lausanne, Lausanne, Switzerland ([email protected]).
1
2
JOEL A. TROPP ET AL.
The challenge is to make the sketch as small as possible while collecting enough
information to approximate the matrix accurately. In particular, we want the sketch
dimension d to be much smaller than the total dimension mn of the matrix A. As
a consequence, the sketching map L has a substantial null space. Therefore, it is
natural to draw the sketching map at random so that we are likely to extract useful
information from any fixed input matrix.
1.3. Why Sketch? There are a number of situations where the sketching model
(1.2) is a natural mechanism for acquiring data about an input matrix.
First, imagine that A is a huge matrix that can only be stored outside of core
memory. The cost of data transfer may be substantial enough that we can only afford
to read the matrix into core memory once [19, Sec. 5.5]. We can build a sketch as we
scan through the matrix. Other types of algorithms for this problem appear in [15, 16].
Second, there are applications where the columns of the matrix A are revealed
one at a time, and we must be able to compute an approximation at any instant. One
approach is to maintain a sketch that is updated when a new column arrives. Other
types of algorithms for this problem appear in [4, 21].
Third, we may encounter a setting where the matrix A is presented as a sum of
ordered updates:
(1.3)
A = H1 + H2 + H3 + H4 + · · · .
We must discard each innovation Hi after it is processed [9, 34]. In this case, the
random linear sketch (1.2) is more or less the only way to maintain a representation
of A through an arbitrary sequence of updates [23]. Our research was motivated by
a variant [36] of the model (1.3); see subsection 3.8.
1.4. Overview of Algorithms. Let us summarize our basic approach to sketching and low-rank approximation of a matrix. Fix a target rank r and an input matrix
A ∈ Fm×n . Select sketch size parameters k and `. Draw and fix independent standard normal matrices Ω ∈ Fn×k and Ψ ∈ F`×m ; see Definition 2.1. We realize the
randomized linear sketch (1.2) via left and right matrix multiplication:
(1.4)
Y := AΩ and W := ΨA
We can store the random matrices and the sketch using (k + `)(m + n) scalars. The
arithmetic cost of forming the sketch is Θ((k + `)mn) floating-point operations (flops)
for a general matrix A.
Given the random matrices (Ω, Ψ) and the sketch (Y , W ), we compute an approximation  in three steps:
1. Form an orthogonal–triangular factorization Y =: QR where Q ∈ Fm×k .
2. Solve a least-squares problem to obtain X := (ΨQ)† W ∈ Fk×n .
3. Construct the rank-k approximation  := QX.
The total cost of this computation is Θ(kl(m + n)) flops. See subsection 4.2 for the
intuition behind this approach.
Now, suppose that we set the sketch size parameters k = 2r + 1 and ` = 4r + 2.
For this choice, Theorem 4.3 yields the error bound
E kA − ÂkF ≤ 2 ·
min
rank(B)≤r
kA − BkF .
In other words, we typically obtain an approximation with rank ≈ 2r whose error
lies within twice the optimal rank-r error! Moreover, the total storage cost is about
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
3
6r(m + n), which is comparable with the number of degrees of freedom in an m × n
matrix with rank r, so the sketch size cannot be reduced substantially.
1.5. Our Contributions. This paper presents a systematic treatment of sketching algorithms for low-rank approximation of a matrix. All of the methods rely on
the simple sketch (1.4) of the input matrix (subsection 3.5). The main algorithm uses
this sketch to compute a high-quality low-rank approximation  of the input matrix
(Algorithm 4). We prove that this method automatically takes advantage of spectral
decay in the input matrix (Theorem 4.3); this result is new.
We also explain how to compute approximations with additional structure—
such as symmetry, positive semidefiniteness, or fixed rank—by projecting the initial
low-rank approximation onto the family of structured matrices (sections 5 and 6).
This approach ensures that the structured approximations also exploit spectral decay
(Fact 5.1 and Proposition 6.1). In the sketching context, this idea is new.
Each algorithm is accompanied by an informative error bound that provides a
good description of its actual behavior. As a consequence, we can offer the first
concrete guidance on algorithm parameters for various types of input matrices (subsection 4.5), and we can implement the methods with confidence. We also include
pseudocode and an accounting of computational costs.
The paper includes a collection of numerical experiments (section 7). This work
demonstrates that the recommended algorithms can significantly outperform alternative methods, especially when the input matrix has spectral decay. The empirical
work also confirms our guidance on parameter choices.
Our technical report [32] contains some more error bounds for the reconstruction
algorithms. It also documents additional numerical experiments.
1.6. Limitations. The algorithms in this paper are not designed for all lowrank matrix approximation problems. They are specifically intended for environments
where we can only make a single pass over the input matrix or where the data matrix
is presented as a stream of linear updates. When it is possible to make multiple passes
over the input matrix, we recommend the low-rank approximation algorithms documented in [19]. Multi-pass methods are significantly more accurate because they drive
the error of the low-rank approximation down to the optimal low-rank approximation
error exponentially fast in the number of passes.
1.7. Overview of Related Work. Randomized algorithms for matrix approximation date back to research [30, 17] in theoretical computer science (TCS) in the
late 1990s. Starting around 2004, this work inspired numerical analysts to develop
practical algorithms for matrix approximation and related problems [26]. See the paper [19, Sec. 2] for a comprehensive historical discussion. The surveys [25, 34] provide
more details about the development of these ideas within the TCS literature.
1.7.1. Sketching Algorithms for Matrix Approximation. To the best of
our knowledge, the first sketching algorithm for low-rank matrix approximation appears in Woolfe et al. [35, Sec. 5.2]. Their primary motivation was to compute a
low-rank matrix approximation faster than any classical algorithm, rather than to
work under the constraints of a sketching model. A variant of their approach is
outlined in [19, Sec. 5.5].
Clarkson & Woodruff [9] explicitly frame the question of how to perform numerical linear algebra tasks under the sketching model (1.2). Among other things, they
develop algorithms and lower bounds for low-rank matrix approximation. Some of
the methods that we recommend are algebraically—but not numerically—equivalent
4
JOEL A. TROPP ET AL.
to formulas [9, Thm. 4.7, 4.8] that they propose. Their work focuses on obtaining a
priori error bounds. In contrast, we also aim to help users implement the methods,
choose parameters, and obtain good empirical performance in practice. Additional
details appear throughout our presentation.
There are many subsequent theoretical papers on sketching algorithms for lowrank matrix approximation, including [34, 12, 6]. This line of research exploits a
variety of tricks to obtain algorithms that, theoretically, attain better asymptotic upper bounds on computational resource usage. Subsection 7.3 contains a representative
selection of these methods and their guarantees.
1.7.2. Added in Press. When we wrote this paper, the literature did not contain sketching methods tailored for symmetric or positive-semidefinite matrix approximation. A theoretical paper [8] on algorithms for low-rank approximation of a sparse
psd matrix was released after our work appeared.
1.7.3. Error Bounds. Almost all previous papers in this area have centered on
the following problem. Let A ∈ Fm×n be an input matrix, let r be a target rank, and
let ε > 0 be an error tolerance. Given a randomized linear sketch (1.2) of the input
matrix, produce a rank-r approximation Âeps that satisfies
(1.5)
kA − Âeps k2F ≤ (1 + ε) ·
min
rank(B)≤r
kA − Bk2F
with high probability.
To achieve (1.5) for a general input, the sketch must have dimension Ω(r(m+n)/ε) [9,
Thm. 4.10]. Furthermore, the analogous error bound for the spectral norm cannot be
achieved for all input matrices under the sketching model [34, Ch. 6.2]. Nevertheless,
Gu [18, Thm. 3.4] has observed that (1.5) implies a weak error bound in the spectral
norm.
Li et al. [22, App.] caution that the guarantee (1.5) is often vacuous. For example,
we frequently encounter matrices for which the Frobenius-norm error of an optimal
rank-r approximation is larger than the Frobenius norm of the approximation itself.
In other settings, it may be necessary to compute an approximation with very high
accuracy. Either way, ε must be tiny before the bound (1.5) sufficiently constrains
the approximation error. For a general input matrix, to achieve a small value of ε,
the sketch size must be exorbitant. We tackle this issue by providing alternative error
estimates (e.g., Theorem 4.3) that yield big improvements for most examples.
1.7.4. Questions... Our aim is to address questions that arise when one attempts to use sketching algorithms in practice. For instance, how do we implement
these methods? Are they numerically stable? How should algorithm parameters depend on the input matrix? What is the right way to preserve structural properties?
Which methods produce the best approximations in practice? How small an approximation error can we actually achieve? Does existing theoretical analysis predict
performance? Can we obtain error bounds that are more illuminating than (1.5)?
These questions have often been neglected in the literature.
Our empirical study (section 7) highlights the importance of this inquiry. Surprisingly, numerical experiments reveal that the pursuit of theoretical metrics has been
counterproductive. More recent algorithms often perform worse in practice, even
though—in principle—they offer better performance guarantees.
2. Background. In this section, we collect notation and conventions, as well as
some background on random matrices.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
5
2.1. Notation and Conventions. We write F for the scalar field, which is
either R or C. The letter I signifies the identity matrix; its dimensions are determined
by context. The star ∗ refers to the (conjugate) transpose operation on vectors and
matrices. The dagger † is the Moore–Penrose pseudoinverse. The symbol k·kF denotes
the Frobenius norm.
The expression “M has rank r” and its variants mean that the rank of M does
not exceed r. The symbol JM Kr represents an optimal rank-r approximation of M
with respect to Frobenius norm; this approximation need not be unique [20, Sec. 6].
It is valuable to introduce notation for the error incurred by a best rank-r approximation in the Frobenius norm. For each natural number j, we define the jth tail
energy
X
(2.1)
τj2 (A) := min kA − Bk2F =
σi2 (A).
rank(B)<j
i≥j
We have written σi (A) for the ith largest singular value of A. The equality follows
from the Eckart–Young Theorem; for example, see [20, Sec. 6].
The symbol E denotes expectation with respect to all random variables. For
a given random variable Z, we write EZ to denote expectation with respect to the
randomness in Z only. Nonlinear functions bind before the expectation.
In the description of algorithms in the text, we primarily use standard mathematical notation. In the pseudocode, we rely on some Matlab R2017a functions in
an effort to make the presentation more concise.
We use the computer science interpretation of Θ(·) to refer to the class of functions
whose growth is bounded above and below up to a constant.
2.2. Standard Normal Matrices. Let us define an ensemble of random matrices that plays a central role in this work.
Definition 2.1 (Standard Normal Matrix). A matrix G ∈ Rm×n has the real
standard normal distribution if the entries form an independent family of standard
normal random variables (i.e., Gaussian with mean zero and variance one).
A matrix G ∈ Cm×n has the complex standard normal distribution if it has the
form G = G1 +iG2 where G1 and G2 are independent, real standard normal matrices.
Standard normal matrices are also known as Gaussian matrices.
We introduce numbers α and β that reflect the field over which the random matrix
is defined:
(
(
1, F = R
1, F = R
(2.2)
α := α(F) :=
and β := β(F) :=
.
0, F = C
2, F = C
This notation allows us to treat the real and complex case simultaneously. The number
β is a standard parameter in random matrix theory.
Last, we introduce notation to help make our theorem statements more succinct:
s
for integers that satisfy t > s + α > α.
(2.3)
f (s, t) :=
t−s−α
Observe that the function f (s, ·) is decreasing, with range (0, s].
3. Sketching the Input Matrix. First, we discuss how to collect enough data
about an input matrix to compute a low-rank approximation. We summarize the
matrix by multiplying it on the right and the left by random test matrices. The
dimension and distribution of these random test matrices together determine the
potential accuracy of the approximation.
6
JOEL A. TROPP ET AL.
3.1. The Input Matrix. Let A ∈ Fm×n be a matrix that we wish to approximate. Our algorithms work regardless of the relative dimensions of A, but there may
sometimes be small benefits if we apply them to A∗ instead.
3.2. The Target Rank. Let r be a target rank parameter with 1 ≤ r ≤
min{m, n}. We aim to construct a low-rank approximation of A whose error is close
to the optimal rank-r error. We explicitly allow approximations with rank somewhat
larger than r because they may be significantly more accurate.
Under the sketching model (1.2), the practitioner must use prior knowledge about
the input matrix A to determine a target rank r that will result in satisfactory error
guarantees. This decision is outside the scope of our work.
3.3. Parameters for the Sketch. The sketch consists of two parts: a summary
of the range of A and a summary of the co-range. The parameter k controls the size
of the range sketch, and the parameter ` controls the size of the co-range sketch. They
should satisfy the conditions
(3.1)
r ≤ k ≤ ` and k ≤ n
and ` ≤ m.
We often choose k ≈ r and ` ≈ k. See (4.6) and subsection 4.5 below.
The parameters k and ` do not play symmetrical roles. We need ` ≥ k to ensure
that a certain ` × k matrix has full column rank. Larger values of both k and ` result
in better approximations at the cost of more storage and arithmetic. These tradeoffs
are quantified in the sequel.
3.4. The Test Matrices. To form the sketch of the input matrix, we draw and
fix two (random) test matrices:
(3.2)
Ω ∈ Fn×k
and Ψ ∈ F`×m .
This paper contains a detailed analysis of the case where the test matrices are statistically independent and follow the standard normal distribution. Subsection 3.9
describes other potential distributions for the test matrices. We always state when
we are making distributional assumptions on the test matrices.
3.5. The Sketch. The sketch of A ∈ Fm×n consists of two matrices:
(3.3)
Y := AΩ ∈ Fm×k
and W := ΨA ∈ F`×n .
The matrix Y collects information about the action of A, while the matrix W collects
information about the action of A∗ . Both parts are necessary.
Remark 3.1 (Prior Work). The matrix sketching algorithms that appear in [35,
Sec. 5.2] and [9, Thm. 4.9] and [19, Sec. 5.5] and [34, Thm. 4.3] all involve a sketch
of the form (3.3). In contrast, the most recent approaches ([6, Sec. 6.1.2] and [33,
Sec. 3]) use more complicated sketches; see subsection 7.3.2.
3.6. The Sketch as an Abstract Data Type. We present the sketch as an
abstract data type using ideas from object-oriented programming. Sketch is an
object that contains information about a specific matrix A. The test matrices (Ω, Ψ)
and the sketch matrices (Y , W ) are private variables that are only accessible to the
Sketch methods. A user interacts with the Sketch object by initializing it with a
specific matrix and by applying linear updates. The user can query the Sketch object
to obtain an approximation of the matrix A with specific properties. The individual
algorithms described in this paper are all methods that belong to the Sketch object.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
7
Algorithm 1 Sketch for Low-Rank Approximation. Implements (3.2) and (3.3).
Require: Input matrix A ∈ Fm×n ; sketch size parameters k ≤ `
Ensure: Constructs test matrices Ω ∈ Fn×k and Ψ ∈ F`×m , range sketch Y = AΩ ∈
Fm×k , and co-range sketch W = ΨA ∈ F`×n as private variables
1
private: Ω, Ψ, Y , W
2
function Sketch(A; k, `)
. Constructor
if F = R then
Ω ← randn(n, k)
Ψ ← randn(`, m)
if F = C then
Ω ← randn(n, k) + i randn(n, k)
Ψ ← randn(`, m) + i randn(`, m)
Ω ← orth(Ω)
. (optional) Improve numerical stability
Ψ∗ ← orth(Ψ∗ )
. (optional) Improve numerical stability
Y ← AΩ
W ← ΨA
3
4
5
6
7
8
9
10
11
12
. Internal variables for Sketch object
. Accessible to all Sketch methods
Algorithm 2 Linear Update to Sketch. Implements (3.4).
Require: Update matrix H ∈ Fm×n ; scalars θ, η ∈ F
Ensure: Modifies sketch (Y , W ) to reflect linear update A ← θA + ηH
1
2
3
function Sketch.LinearUpdate(H; θ, η)
Y ← θY + ηHΩ
. Linear update to range sketch
W ← θW + ηΨH
. Linear update to co-range sketch
3.7. Initializing the Sketch and its Costs. See Algorithm 1 for pseudocode
that implements the sketching procedure (3.2) and (3.3) with either standard normal
test matrices (default) or random orthonormal test matrices (optional steps). Note
that the orthogonalization step requires additional arithmetic and communication.
The storage cost for the sketch (Y , W ) is mk + `n floating-point numbers in the
field F. The storage cost for two standard normal test matrices is nk + `m floating
point numbers in F. Some other types of test matrices (Ω, Ψ) have lower storage
costs, but the sketch (Y , W ) remains the same size.
For standard normal test matrices, the arithmetic cost of forming the sketch (3.3)
is Θ((k + `)mn) flops when A is dense. If A is sparse, the cost is proportional to
the number nnz(A) of nonzero entries: Θ((k + `) nnz(A)) flops. Other types of test
matrices sometimes yield lower arithmetic costs.
3.8. Processing Linear Updates. The sketching model (3.3) supports a linear
update that is more general than (1.3). Suppose the input matrix A is modified as
A ← θA + ηH
where θ, η ∈ F.
Then we update the sketch (3.3) via the rule
(3.4)
Y ← θY + ηHΩ and W ← θW + ηΨH.
8
JOEL A. TROPP ET AL.
The precise cost of the computation depends on the structure of H. See Algorithm 2
for pseudocode. This type of update is crucial for certain applications [36].
3.9. Choosing the Distribution of the Test Matrices. Our analysis is specialized to the case where the test matrices Ω and Ψ are standard normal so that we
can obtain highly informative error bounds.
But there are potential benefits from implementing the sketch using test matrices
drawn from another distribution. The choice of distribution leads to some tradeoffs
in the range of permissible parameters; the costs of randomness, arithmetic, and
communication to generate the test matrices; the storage costs for the test matrices
and the sketch; the arithmetic costs for sketching and updates; the numerical stability
of matrix approximation algorithms; and the quality of a priori error bounds.
Let us list some of the contending distributions along with background references.
We have ranked these in decreasing order of reliability.
• Orthonormal. The optional steps in Algorithm 1 generate matrices Ω and
Ψ∗ with orthonormal columns that span uniformly random subspaces of dimension k and `. When k and ` are very large, these matrices result in smaller
errors and better numerical stability than Gaussians [14, 19].
• Gaussian. Following [26, 19], this paper focuses on test matrices with the
standard normal distribution. Benefits include excellent practical performance and accurate a priori error bounds.
• Rademacher. These test matrices have independent Rademacher1 entries.
Their behavior is similar to Gaussian test matrices, but there are minor improvements in the cost of storage and arithmetic, as well as the amount of
randomness required. For example, see [9].
• Subsampled Randomized Fourier Transform (SRFT). These test matrices take the form
(3.5)
Ω = D1 F1 P1
and Ψ = P2 F2∗ D2
where D1 ∈ Fn×n and D2 ∈ Fm×m are diagonal matrices with independent
Rademacher entries; F1 ∈ Fn×n and F2 ∈ Fm×m are discrete cosine transform
(F = R) or discrete Fourier transform (F = C) matrices; and P1 ∈ Fn×k and
P2 ∈ F`×m are restrictions onto k and ` coordinates, chosen uniformly at
random. These matrices work well in practice; they require a modest amount
of storage; and they support fast arithmetic. See [1, 35, 2, 19, 31, 5, 13].
• Ultra-Sparse Rademacher. Let s be a sparsity parameter. In each row
of Ω and column of Ψ, we place independent Rademacher random variables
in s uniformly random locations; the remaining entries of the test matrices
are zero. These matrices help control storage, arithmetic, and randomness
costs. On the other hand, they are somewhat less reliable. For more details,
see [10, 28, 27, 29, 34, 3, 11].
Except for ultra-sparse Rademacher matrices, these distributions often behave quite
like a Gaussian distribution in practice [19, Sec. 7.4]. An exhaustive comparison of
distributions for the test matrices is outside the scope of this paper; see [24].
1A
Rademacher random variable takes the values ±1 with equal probability.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
9
4. Low-Rank Approximation from the Sketch. Suppose that we have acquired a sketch (Y , W ) of the input matrix A, as in (3.2) and (3.3). This section
presents the most basic algorithm for computing a low-rank approximation of A from
the data in the sketch. This simple approach is similar to earlier proposals; see [35,
Sec. 5.2], [9, Thm. 4.7], [19, Sec. 5.5], [34, Thm. 4.3, display 1].
We have obtained the first accurate error bound for this method. Our result
shows how the spectrum of the input matrix affects the approximation quality. This
analysis allows us to make parameter recommendations for specific input matrices.
In section 5, we explain how to refine this algorithm to obtain approximations
with additional structure. In section 6, we describe modifications of the procedures
that produce approximations with fixed rank and additional structure. Throughout,
we maintain the notation of section 3.
4.1. The Main Algorithm. Our goal is to produce a low-rank approximation
of the input matrix A using only the knowledge of the test matrices (Ω, Ψ) and the
sketch (Y , W ). Here is the basic method.
The first step in the procedure is to compute an orthobasis for the range of Y by
means of an orthogonal–triangular factorization:
(4.1)
Y =: QR
where
Q ∈ Fm×k .
The matrix Q has orthonormal columns; we discard the triangular matrix R. The
second step uses the co-range sketch W to form the matrix
(4.2)
X := (ΨQ)† W ∈ Fk×n .
The random matrix ΨQ ∈ F`×k is very well-conditioned when ` k, so we can
perform this computation accurately by solving a least-squares problem. We report
the rank-k approximation
(4.3)
 := QX ∈ Fm×n
where
Q ∈ Fm×k
and X ∈ Fk×n .
The factors Q and X are defined in (4.1) and (4.2).
Remark 4.1 (Prior Work). The approximation  is algebraically, but not numerically, equivalent with the approximation that appears in Clarkson & Woodruff [9,
Thm. 4.7]; see also [34, Thm. 4.3, display 1]. Our formulation improves on theirs by
avoiding a badly conditioned least-squares problem.
4.2. Intuition. To motivate the algorithm, we recall a familiar heuristic [19,
Sec. 1] from randomized linear algebra, which states that
(4.4)
A ≈ QQ∗ A.
Although we would like to form the rank-k approximation Q(Q∗ A), we cannot compute the factor Q∗ A without revisiting the input matrix A. Instead, we exploit the
information in the co-range sketch W = ΨA. Notice that
W = Ψ(QQ∗ A) + Ψ(A − QQ∗ A) ≈ (ΨQ)(Q∗ A).
The heuristic (4.4) justifies dropping the second term. Multiplying on the left by the
pseudoinverse (ΨQ)† , we arrive at the relation
X = (ΨQ)† W ≈ Q∗ A.
10
JOEL A. TROPP ET AL.
Algorithm 3 Simplest Low-Rank Approximation. Implements (4.3).
Ensure: For some q ≤ k, returns factors Q ∈ Fm×q with orthonormal columns and
X ∈ Fq×n that form a rank-q approximation Âout = QX of the sketched matrix
1
2
3
4
function Sketch.SimpleLowRankApprox( )
Q ← orth(Y )
. Orthobasis for range of Y
X ← (ΨQ)\W
. Multiply (ΨQ)† on left side of W
return (Q, X)
Algorithm 4 Low-Rank Approximation. Implements (4.3).
Ensure: Returns factors Q ∈ Fm×k with orthonormal columns and X ∈ Fk×n that
form a rank-k approximation Âout = QX of the sketched matrix
1
2
3
4
5
function Sketch.LowRankApprox( )
(Q, ∼) ← qr(Y , 0)
(U , T ) ← qr(ΨQ, 0)
X ← T † (U ∗ W )
return (Q, X)
. Orthobasis for range of Y
. Orthogonal–triangular factorization
. Apply inverse by back-substitution
These considerations suggest that
 = QX ≈ QQ∗ A ≈ A.
One of our contributions is to give substance to these nebulae.
Remark 4.2 (Prior Work). This intuition is inspired by the discussion in [19,
Sec. 5.5], and it allows us to obtain sharp error bounds. Our approach is quite
different from that of [9, Thm. 4.7] or [34, Thm. 4.3].
4.3. Algorithm and Costs. Algorithms 3 and 4 give pseudocode for computing
the approximation (4.3). The first presentation uses Matlab functions to abbreviate
some of the steps, while the second includes more implementation details. Note that
the use of the orth command may result in an approximation with rank q for some
q ≤ k, but the quality of the approximation does not change.
Let us summarize the costs of the approximation procedure (4.1)–(4.3), as implemented in Algorithm 4. The algorithm has working storage of O (k(m + n)) floating
point numbers. The arithmetic cost is Θ(k`(m + n)) flops, which is dominated by the
matrix–matrix multiplications. The orthogonalization step and the back-substitution
require Θ(k 2 (m + n)) flops, which is almost as significant.
4.4. A Bound for the Frobenius-Norm Error. We have established a very
accurate error bound for the approximation (4.3) that is implemented in Algorithms 3
and 4. This analysis is one of the key contributions of this paper.
Theorem 4.3 (Low-Rank Approximation: Frobenius Error). Assume that the
sketch size parameters satisfy ` > k + α. Draw random test matrices Ω ∈ Fn×k and
Ψ ∈ F`×m independently from the standard normal distribution. Then the rank-k
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
11
approximation  obtained from formula (4.3) satisfies
2
E kA − Âk2F ≤ (1 + f (k, `)) · min (1 + f (%, k)) · τ%+1
(A)
%<k−α
(4.5)
k
k
· min
· τ 2 (A).
=
` − k − α %<k−α k − % − α %+1
The index % ranges over natural numbers. The quantity α(R) := 1 and α(C) := 0; the
function f (s, t) := s/(t − s − α); the tail energy τj2 is defined in (2.1).
The proof of Theorem 4.3 appears below in Appendix A.3.
To begin to understand Theorem 4.3, it is helpful to consider a specific parameter
choice. Let r be the target rank of the approximation, and select
(4.6)
k = 2r + α
and ` = 2k + α.
For these sketch size parameters, with % = r, Theorem 4.3 implies that
2
E kA − Âk2F ≤ 4 · τr+1
(A).
In other words, for k ≈ 2r, we can construct a rank-k approximation of A that
has almost the same quality as a best rank-r approximation. This parameter choice
balances the sketch size against the quality of approximation.
But the true meaning of Theorem 4.3 lies deeper. The minimum in (4.5) reveals
that the approximation (4.3) automatically takes advantage of decay in the tail energy.
This fundamental fact explains the strong empirical performance of (4.3) and other
approximations derived from it. Our analysis is the first to identify this feature.
Remark 4.4 (Prior Work). The analysis in [9, Thm. 3.7] shows that  achieves
a bound of the form (1.5) when the sketch size parameters scale as k = Θ(r/ε) and
` = Θ(k/ε). A precise variant of the same statement follows from Theorem 4.3.
Remark 4.5 (High-Probability Error Bound). The expectation bound presented
in Theorem 4.3 also describes the typical behavior of the approximation (4.3) because
of measure concentration effects. It is possible to develop a high-probability bound
using the methods from [19, Sec. 10.3].
Remark 4.6 (Spectral-Norm Error Bound). It is also possible to develop bounds
for the spectral-norm error incurred by the approximation (4.3). These results depend
on the decay of both the singular values and the tail energies. See [32, Thm. 4.2].
4.5. Theoretical Guidance on the Sketch Size. Theorem 4.3 is precise
enough to predict the performance of the approximation (4.3) for many types of
input matrices. As a consequence, we can offer concrete guidance on the best sketch
size parameters (k, `) for various applications.
Observe that the storage cost of the sketch (3.3) is directly proportional to the
sum T := k + ` of the sketch size parameters k and `. In this section, we investigate
the best way to apportion k and ` when we fix the target rank r and the total sketch
size T . Throughout this discussion, we assume that T ≥ 2r + 3α + 3. See Table 1 for
a summary of these rules; see subsection 7.5 for an empirical evaluation.
4.5.1. Flat Spectrum. First, suppose that the singular values σj (A) of the
input matrix A do not decay significantly for j > r. This situation occurs, for
example, when the input is a rank-r matrix plus white noise.
12
JOEL A. TROPP ET AL.
Table 1: Theoretical Sketch Size Parameters. This table summarizes how
to choose the sketch size parameters (k, `) to exploit prior information about the
spectrum of the input matrix A.
Problem Regime
General purpose
Flat spectrum
Decaying spectrum
Rapidly decaying spectrum
Notation
(k\ , `\ )
(k[ , `[ )
(k\ , `\ )
(k] , `] )
Equation
(4.9)
(4.7) and (4.8)
(4.9)
(4.10)
In this setting, the minimum in (4.5) is likely to occur when % ≈ r. It is natural
to set % = r and to minimize the resulting bound subject to the constraints k + ` = T
and k > r + α and ` > k + α. For F = C, we obtain the parameter recommendations
%)
(
$
p
r(T − r) − r
and `[ := T − k[ .
(4.7)
k[ := max r + 1, T ·
T − 2r
In case F = R, we modify the formula (4.7) so that
(
$
%)
p
r(T − r − 2)(1 − 2/(T − 1)) − (r − 1)
(4.8) k[ := max r + 2, (T − 1) ·
.
T − 2r − 1
We omit the routine details behind these calculations.
4.5.2. Decaying Spectrum or Spectral Gap. Suppose that the singular values σj (A) decay at a slow to moderate rate for j > r. Alternatively, we may suppose
that there is a gap in the singular value spectrum at an index j > r.
In this setting, we want to exploit decay in the tail energy by setting k r, but
we need to ensure that the term f (k, `) in (4.5) remains small by setting ` ≈ 2k + α.
This intuition leads to the parameter recommendations
(4.9)
k\ := max{r + α + 1, b(T − α)/3c}
and `\ := T − k\ .
This is the best single choice for handling a range of examples. The parameter recommendation (4.6) is an instance of (4.9) with a minimal value of T .
4.5.3. Rapidly Decaying Spectrum. Last, assume that the singular values
σj (A) decay very quickly for j > r. This situation occurs in the application [36] that
motivated us to write this paper.
In this setting, we want to exploit decay in the tail energy fully by setting k as large
as possible; the benefit outweighs the increase in f (k, `) from choosing ` = k + α + 1,
the minimum possible value. This intuition leads to the parameter recommendations
(4.10)
k] := b(T − α − 1)/2c
and `] := T − k] .
Note that the choice (4.10) is unwise unless the input matrix has sharp spectral decay.
5. Low-Rank Approximations with Convex Structure. In many instances,
we need to reconstruct an input matrix that has additional structure, such as symmetry or positive-semidefiniteness. The approximation formula (4.3) from section 4
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
13
produces an approximation with no special properties aside from a bound on its rank.
Therefore, we may have to reform our approximation to instill additional virtues.
In this section, we consider a class of problems where the input matrix belongs
to a convex set and we seek an approximation that belongs to the same set. To
accomplish this goal, we replace our initial approximation with the closest point in
the convex set. This procedure always improves the Frobenius-norm error.
We address two specific examples: (i) the case where the input matrix is conjugate
symmetric and (ii) the case where the input matrix is positive semidefinite. In both
situations, we must design the algorithm carefully to avoid forming large matrices.
5.1. Projection onto a Convex Set. Let C be a closed and convex set of
matrices in Fm×n . Define the projector ΠC onto the set C to be the map
ΠC : Fm×n → C where ΠC (M ) := arg min kC − M k2F : C ∈ C .
The arg min operator returns the matrix C? ∈ C that solves the optimization problem.
The solution C? is uniquely determined because the squared Frobenius norm is strictly
convex and the constraint set C is closed and convex.
5.2. Structure via Convex Projection. Suppose that the input matrix A
belongs to the closed, convex set C ⊂ Fm×n . Let Âin ∈ Fm×n be an initial approximation of A. We can produce a new approximation ΠC (Âin ) by projecting the
initial approximation onto the constraint set. This procedure always improves the
approximation quality in Frobenius norm.
Fact 5.1 (Convex Structure Reduces Error). Let C ∈ Fm×n be a closed convex
set, and suppose that A ∈ C. For any initial approximation Âin ∈ Fm×n ,
(5.1)
kA − ΠC (Âin )kF ≤ kA − Âin kF .
This result is well known in convex analysis. It follows directly from the first-order
optimality conditions [7, Sec. 4.2.3] for the Frobenius-norm projection of a matrix onto
the set C. We omit the details.
Warning 5.2 (Spectral Norm). Fact 5.1 does not hold if we replace the Frobenius
norm by the spectral norm.
5.3. Low-Rank Approximation with Conjugate Symmetry. When the
input matrix is conjugate symmetric, it is often critical to produce a conjugate symmetric approximation. We can do so by combining the simple approximation from
section 4 with the projection step outlined in subsection 5.1.
5.3.1. Conjugate Symmetric Projection. Define the set Hn (F) of conjugate
symmetric matrices with dimension n over the field F:
Hn := Hn (F) := {C ∈ Fn×n : C = C ∗ }.
The set Hn (F) is convex because it forms a real-linear subspace in Fn×n . In the sequel,
we omit the field F from the notation unless there is a possibility of confusion.
The projection Msym of a matrix M ∈ Fn×n onto the set Hn takes the form
(5.2)
Msym := ΠHn (M ) =
For example, see [20, Sec. 2].
1
(M + M ∗ ).
2
14
JOEL A. TROPP ET AL.
Algorithm 5 Low-Rank Symmetric Approximation. Implements (5.6).
Require: Matrix dimensions m = n
Ensure: For q = 2k, returns factors U ∈ Fn×q with orthonormal columns and S ∈ Hq
that form a rank-q conjugate symmetric approximation Âout = U SU ∗ of the
sketched matrix
1
2
3
4
5
6
function Sketch.LowRankSymApprox( )
(Q, X) ← LowRankApprox( )
. Get Âin = QX
(U , T ) ← qr([Q, X ∗ ], 0)
. Orthogonal factorization of concatenation
T1 ← T (:, 1:k) and T2 ← T (:, (k + 1):(2k))
. Extract submatrices
S ← (T1 T2∗ + T2 T1∗ )/2
. Symmetrize
return (U , S)
. Return factors
5.3.2. Computing a Conjugate Symmetric Approximation. Assume that
the input matrix A ∈ Hn is conjugate symmetric. Let  := QX be an initial rank-k
approximation of A obtained from the approximation procedure (4.3). We can form
a better Frobenius-norm approximation Âsym by projecting  onto Hn :
(5.3)
Âsym := ΠHn (Â) =
1
1
(Â + Â∗ ) = (QX + X ∗ Q∗ ).
2
2
The second relation follows from (5.2).
In most cases, it is preferable to present the approximation (5.3) in factored form.
To do so, we observe that
0 I
∗
1
1
Q X∗
Q X∗ .
(QX + X ∗ Q∗ ) =
I 0
2
2
Concatenate Q and X ∗ , and compute the orthogonal–triangular factorization
Q X ∗ =: U T1 T2
(5.4)
where U ∈ Fn×2k and T1 ∈ F2k×k .
Of course, we only need to orthogonalize the k columns of X ∗ , which permits some
computational efficiencies. Next, introduce the matrix
0 I
∗
1
1
T1 T2
T1 T2 = (T1 T2∗ + T2 T1∗ ) ∈ F2k×2k .
(5.5)
S :=
I
0
2
2
Combine the last four displays to obtain the rank-(2k) conjugate symmetric approximation
(5.6)
Âsym = U SU ∗ .
From this expression, it is easy to obtain other types of factorizations, such as an
eigenvalue decomposition, by further processing.
5.3.3. Algorithm, Costs, and Error. Algorithm 5 contains pseudocode for
producing a conjugate symmetric approximation of the form (5.6) from a sketch of
the input matrix. One can make this algorithm slightly more efficient by taking
advantage of the fact that Q already has orthogonal columns; we omit the details.
For Algorithm 5, the total working storage is Θ(kn) and the arithmetic cost is
Θ(k`n). These costs are dominated by the call to Sketch.LowRankApprox.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
15
Combining Theorem 4.3 with Fact 5.1, we have the following bound on the error of
the symmetric approximation (5.6), implemented in Algorithm 5. As a consequence,
the parameter recommendations from subsection 4.5 are also valid here.
Corollary 5.3 (Low-Rank Symmetric Approximation). Assume that the input
matrix A ∈ Hn (F) is conjugate symmetric, and assume that the sketch size parameters
satisfy ` > k + α. Draw random test matrices Ω ∈ Fn×k and Ψ ∈ F`×n independently
from the standard normal distribution. Then the rank-(2k) conjugate symmetric approximation Âsym produced by (5.3) or (5.6) satisfies
2
E kA − Âsym k2F ≤ (1 + f (k, `)) · min (1 + f (%, k)) · τ%+1
(A).
%<k−α
The index % ranges over natural numbers. The quantity α(R) := 1 and α(C) := 0; the
function f (s, t) := s/(t − s − α); the tail energy τj2 is defined in (2.1).
5.4. Low-Rank Positive-Semidefinite Approximation. We often encounter
the problem of approximating a positive-semidefinite (psd) matrix. In many situations, it is important to produce an approximation that maintains positivity. Our
approach combines the approximation (4.3) from section 4 with the projection step
from subsection 5.1.
5.4.1. PSD Projection. We introduce the set Hn+ (F) of psd matrices with dimension n over the field F:
Hn+ := Hn+ (F) := C ∈ Hn : z ∗ Cz ≥ 0 for each z ∈ Fn .
The set Hn+ (F) is convex because it is an intersection of halfspaces. In the sequel, we
omit the field F from the notation unless there is a possibility for confusion.
Given a matrix M ∈ Fn×n , we construct its projection onto the set Hn+ in three
steps. First, form the projection Msym := ΠHn (M ) onto the conjugate symmetric
matrices, as in (5.2). Second, compute an eigenvalue decomposition Msym =: V DV ∗ .
Third, form D+ by zeroing out the negative entries of D. Then the projection M+
of the matrix M onto Hn+ takes the form
M+ := ΠHn+ (M ) = V D+ V ∗ .
For example, see [20, Sec. 3].
5.4.2. Computing a PSD Approximation. Assume that the input matrix
A ∈ Hn+ is psd. Let  := QX be an initial approximation of A obtained from the
approximation procedure (4.3). We can form a psd approximation Â+ by projecting
 onto the set Hn+ .
To do so, we repeat the computations (5.4) and (5.5) to obtain the symmetric
approximation Âsym presented in (5.6). Next, form an eigenvalue decomposition of
the matrix S given by (5.5):
S =: V DV ∗ .
In view of (5.6), we obtain an eigenvalue decomposition of Âsym :
Âsym = (U V )D(U V )∗ .
To obtain the psd approximation Â+ , we simply replace D by its nonnegative part
D+ to arrive at the rank-(2k) psd approximation
(5.7)
Â+ := ΠHn+ (Â) = (U V )D+ (U V )∗ .
This formula delivers an approximate eigenvalue decomposition of the input matrix.
16
JOEL A. TROPP ET AL.
Algorithm 6 Low-Rank PSD Approximation. Implements (5.7).
Require: Matrix dimensions m = n
Ensure: For q = 2k, returns factors U ∈ Fn×q with orthonormal columns and nonnegative, diagonal D ∈ Hq+ that form a rank-q psd approximation Âout = U DU ∗
of the sketched matrix
1
2
3
4
5
6
function Sketch.LowRankPSDApprox( )
(U , S) ← LowRankSymApprox( )
. Get Âin = U SU ∗
(V , D) ← eig(S)
. Form eigendecomposition
U ← UV
. Consolidate orthonormal factors
D ← max(D, 0)
. Zero out negative eigenvalues
return (U , D)
5.4.3. Algorithm, Costs, and Error. Algorithm 6 contains pseudocode for
producing a psd approximation of the form (5.7) from a sketch of the input matrix.
As in Algorithm 5, some additional efficiencies are possible
The costs of Algorithm 6 are similar with the symmetric approximation method,
Algorithm 5. The working storage cost is Θ(kn), and the arithmetic cost is Θ(k`n).
Combining Theorem 4.3 and Fact 5.1, we obtain a bound on the approximation
error identical to Corollary 5.3. We omit the details.
6. Fixed-Rank Approximations from the Sketch. The algorithms in sections 4 and 5 produce high-quality approximations with rank k, but we sometimes
need to reduce the rank to match the target rank r. At the same time, we may have
to impose additional structure. This section explains how to develop algorithms that
produce a rank-r structured approximation.
The technique is conceptually similar to the approach in section 5. We project
an initial high-quality approximation onto the set of rank-r matrices. This procedure
preserves both conjugate symmetry and the psd property. The analysis in subsection 5.1 does not apply because the set of matrices with fixed rank is not convex. We
present a general argument to show that the cost is negligible.
6.1. A General Error Bound for Fixed-Rank Approximation. If we have
a good initial approximation of the input matrix, we can replace this initial approximation by a fixed-rank matrix without increasing the error significantly.
Proposition 6.1 (Error for Fixed-Rank Approximation). Let A ∈ Fm×n be a
input matrix, and let Âin ∈ Fm×n be an approximation. For any rank parameter r,
(6.1)
kA − JÂin Kr kF ≤ τr+1 (A) + 2kA − Âin kF .
Recall that J·Kr returns a best rank-r approximation with respect to Frobenius norm.
Proof. Calculate that
kA − JÂin Kr kF ≤ kA − Âin kF + kÂin − JÂin Kr kF
≤ kA − Âin kF + kÂin − JAKr kF
≤ 2kA − Âin kF + kA − JAKr kF .
The first and last relations are triangle inequalities. To reach the second line, note
that JÂin Kr is a best rank-r approximation of Âin , while JAKr is an undistinguished
rank-r matrix. Finally, identify the tail energy (2.1).
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
17
Algorithm 7 Fixed-Rank Approximation. Implements (6.3).
Require: Target rank r ≤ k
Ensure: Returns factors Q ∈ Fm×r and V ∈ Fn×r with orthonormal columns and
nonnegative diagonal Σ ∈ Fr×r that form a rank-r approximation Âout = QΣV ∗
of the sketched matrix
1
2
3
4
5
function Sketch.FixedRankApprox(r)
(Q, X) ← LowRankApprox( )
(U , Σ, V ) ← svds(X, r)
Q ← QU
return (Q, Σ, V )
. Get Âin = QX
. Form full SVD and truncate
. Consolidate orthonormal factors
Remark 6.2 (Spectral Norm). A result analogous to Proposition 6.1 also holds
with respect to the spectral norm. The proof is the same.
6.2. Fixed-Rank Approximation from the Sketch. Suppose that we wish
to compute a rank-r approximation of the input matrix A ∈ Fm×n . First, we form
an initial approximation  := QX using the procedure (4.3). Then we obtain a
rank-r approximation JÂKr of the input matrix by replacing  with its best rank-r
approximation in Frobenius norm:
(6.2)
JÂKr = JQXKr .
We can complete this operation by working directly with the factors. Indeed,
suppose that X = U ΣV ∗ is an SVD of X. Then QX has an SVD of the form
QX = (QU )ΣV ∗ .
As such, there is also a best rank-r approximation of QX that satisfies
JQXKr = (QU )JΣKr V ∗ = QJXKr .
Therefore, the desired rank-r approximation (6.2) can also be expressed as
(6.3)
JÂKr = QJXKr .
The formula (6.3) is more computationally efficient than (6.2) because the factor
X ∈ Fk×n is much smaller than the approximation  ∈ Fm×n .
Remark 6.3 (Prior Work). The approximation JÂKr is algebraically, but not numerically, equivalent to a formula proposed by Clarkson & Woodruff [9, Thm. 4.8].
As above, our formulation improves on theirs by avoiding a badly conditioned leastsquares problem.
6.2.1. Algorithm and Costs. Algorithm 7 contains pseudocode for computing
the fixed-rank approximation (6.3).
The fixed-rank approximation in Algorithm 7 has storage and arithmetic costs
on the same order as the simple low-rank approximation (Algorithm 3). Indeed,
to compute the truncated SVD and perform the matrix–matrix multiplication, we
expend only Θ(k 2 n) additional flops. Thus, the total working storage is Θ(k(m + n))
numbers and the arithmetic cost is Θ(k`(m + n)) flops.
18
JOEL A. TROPP ET AL.
6.2.2. A Bound for the Error. We can obtain an error bound for the rank-r
approximation (6.3) by combining Theorem 4.3 and Proposition 6.1.
Corollary 6.4 (Fixed-Rank Approximation: Frobenius-Norm Error). Assume
the sketch size parameters satisfy k > r +α and ` > k +α. Draw random test matrices
Ω ∈ Fn×k and Ψ ∈ F`×m independently from the standard normal distribution. Then
the rank-r approximation JÂKr obtained from the formula (6.3) satisfies
(6.4)
E kA − JÂKr kF ≤ τr+1 (A) + 2
p
1 + f (k, `) · min
%<k−α
p
1 + f (%, k) · τ%+1 (A).
The index % ranges over natural numbers. The quantity α(R) := 1 and α(C) := 0; the
function f (s, t) := s/(t − s − α); the tail energy τj2 is defined in (2.1).
This result indicates that the fixed-rank approximation JÂKr automatically exploits spectral decay in the input matrix A. Moreover, we can still rely on the
parameter recommendations from subsection 4.5. Ours is the first theory to provide
these benefits.
Remark 6.5 (Prior Work). The analysis [9, Thm. 4.8] of Clarkson & Woodruff
implies that the approximation (6.3) can achieve the bound (1.5) for any ε > 0,
provided that k = Θ(r/ε2 ) and ` = Θ(k/ε2 ). It is possible to improve this scaling;
see [32, Thm. 5.1].
Remark 6.6 (Spectral-Norm Error Bound). It is possible to obtain an error bound
for the rank-r approximation (6.3) with respect to the spectral norm by combining
[32, Thm. 4.2] and Remark 6.2.
6.3. Fixed-Rank Conjugate Symmetric Approximation. Assume that the
input matrix A ∈ Hn is conjugate symmetric and we wish to compute a rank-r
conjugate symmetric approximation. First, form an initial approximation Âsym using
the procedure (5.6) in subsection 5.3.2. Then compute an r-truncated eigenvalue
decomposition of the matrix S defined in (5.5):
S =: V JDKr V ∗ + approximation error.
In view of the representation (5.6),
(6.5)
JÂsym Kr = (U V )JDKr (U V )∗ .
Algorithm 8 contains pseudocode for the fixed-rank approximation (6.5). The total
working storage is Θ(kn), and the arithmetic cost is Θ(k`n).
If A is conjugate symmetric, then Corollary 5.3 and Proposition 6.1 shows that
JÂsym Kr admits an error bound identical to Corollary 6.4. We omit the details.
6.4. Fixed-Rank PSD Approximation. Assume that the input matrix A ∈
Hn+ is psd, and we wish to compute a rank-r psd approximation JÂ+ Kr . First, form an
initial approximation Â+ using the procedure (5.7) in subsection 5.4.2. Then compute
an r-truncated positive eigenvalue decomposition of the matrix S defined in (5.5):
S =: V JD+ Kr V ∗ + approximation error.
In view of the representation (5.7),
(6.6)
JÂ+ Kr = (U V )JD+ Kr (U V )∗ .
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
19
Algorithm 8 Fixed-Rank Symmetric Approximation. Implements (6.5).
Require: Matrix dimensions m = n; target rank r ≤ k
Ensure: Returns factors U ∈ Fn×r with orthonormal columns and diagonal D ∈ Hr
that form a rank-r conjugate symmetric approximation Âout = U DU ∗ of the
sketched matrix
1
2
3
4
5
function Sketch.FixedRankSymApprox(r)
(U , S) ← LowRankSymApprox( )
. Get Âin = U SU ∗
(V , D) ← eigs(S, r, ’lm’)
. Truncate full eigendecomposition
U ← UV
. Consolidate orthonormal factors
return (U , D)
Algorithm 9 Fixed-Rank PSD Approximation. Implements (6.6).
Require: Matrix dimensions m = n; target rank r ≤ k
Ensure: Returns factors U ∈ Fn×r with orthonormal columns and nonnegative,
diagonal D ∈ Hr+ that form a rank-r psd approximation Âout = U DU ∗ of the
sketched matrix
1
2
3
4
5
6
function Sketch.FixedRankPSDApprox(r)
(U , S) ← LowRankSymApprox( )
. Get Âin = U SU ∗
(V , D) ← eigs(S, r, ’lr’)
. Truncate full eigendecomposition
U ← UV
. Consolidate orthonormal factors
D ← max(D, 0)
. Zero out negative eigenvalues
return (U , D)
Algorithm 9 contains pseudocode for the fixed-rank psd approximation (6.6). The
working storage is Θ(kn), and the arithmetic cost is Θ(k`n). If A is psd, then Corollary 5.3 and Proposition 6.1 show that JÂpsd Kr satisfies an error bound identical to
Corollary 6.4; we omit the details.
7. Computational Experiments. This section presents the results of some
numerical tests designed to evaluate the empirical performance of our sketching algorithms for low-rank matrix approximation. We demonstrate that the approximation
quality improves when we impose structure, and we show that our theoretical parameter choices are effective. The presentation also includes comparisons with several
other algorithms from the literature.
7.1. Overview of Experimental Setup. For our numerical assessment, we
work over the complex field (F = C). Results for the real field (F = R) are similar.
Let us summarize the procedure for studying the behavior of a specified approximation method on a given input matrix. Fix the input matrix A and the target rank
r. Then select a pair (k, `) of sketch size parameters where k ≥ r and ` ≥ r.
Each trial has the following form. We draw (complex) standard normal test
matrices (Ω, Ψ) to form the sketch (Y , W ) of the input matrix. [We do not use the
optional orthogonalization steps in Algorithm 1.] Next compute an approximation
Âout of the matrix A by means of a specified approximation algorithm. Then calculate
j th singular value
20
JOEL A. TROPP ET AL.
10 0
10 0
10 0
10 -1
10 -1
10 -2
10 -2
10 -6
10 -3
10 -3
10 -9
10 -4
10 -12
LowRank
LowRankMed
LowRankHigh
10 -4
10 -5
PolySlow
PolyFast
ExpSlow
ExpFast
10 -5
10 0
10 1
10 2
10 3
10 0
10 1
j
10 2
10 3
Data
10 -3
10 -15
10 0
10 1
j
10 2
10 3
j
Fig. 7.1: Spectra of input matrices. These plots display the singular value
spectrum for an input matrix from each of the classes (LowRank, LowRankMedNoise,
LowRankHiNoise, PolyDecaySlow, PolyDecayFast, ExpDecaySlow, ExpDecayFast,
Data) described in subsection 7.2.
the error relative to the best rank-r approximation:
(7.1)
relative error
:=
kA − Âout kF
− 1.
τr+1 (A)
The tail energy τj is defined in (2.1). If Âout is a rank-r approximation of A, the
relative error is always nonnegative. To facilitate comparisons, our experiments only
examine fixed-rank approximation methods.
To obtain each data point, we repeat the procedure from the last paragraph 20
times, each time with the same input matrix A and an independent draw of the test
matrices (Ω, Ψ). Then we report the average relative error over the 20 trials.
We include our Matlab implementations in the supplementary materials for
readers who seek more details on the methodology.
7.2. Classes of Input Matrices. We perform our numerical tests using several
types of complex-valued input matrices. Figure 7.1 illustrates the singular spectrum
of a matrix from each of the categories.
7.2.1. Synthetic Examples. We fix a dimension parameter n = 103 and a
parameter R = 10 that controls the rank of the “significant part” of the matrix. In
our experiments, we compute approximations with target rank r = 5. Similar results
hold when the parameter R = 5 and when n = 104 .
We construct the following synthetic input matrices:
1. Low-Rank + Noise: These matrices take the form
r
γR
IR 0
+
(G + G∗ ) ∈ Cn×n .
0 0
2n2
The matrix G is complex standard normal. The quantity γ −2 can be interpreted as the signal-to-noise ratio (SNR). We consider three cases:
(a) No noise (LowRank): γ = 0.
(b) Medium noise (LowRankMedNoise): γ = 10−2 .
(c) High noise (LowRankHiNoise): γ = 1.
For these models, all the experiments are performed on a single exemplar that
is drawn at random and then fixed.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
21
2. Polynomially Decaying Spectrum: These matrices take the form
diag 1, . . . , 1, 2−p , 3−p , 4−p , . . . , (n − R + 1)−p
∈ Cn×n ,
| {z }
R
where p > 0 controls the rate of decay. We consider two cases:
(a) Slow polynomial decay (PolyDecaySlow): p = 1.
(b) Fast polynomial decay (PolyDecayFast): p = 2.
3. Exponentially Decaying Spectrum: These matrices take the form
∈ Cn×n ,
diag 1, . . . , 1, 10−q , 10−2q , 10−3q , . . . , 10−(n−R)q
| {z }
R
where q > 0 controls the rate of decay. We consider two cases:
(a) Slow exponential decay (ExpDecaySlow): q = 0.25.
(b) Fast exponential decay (ExpDecayFast): q = 1.
We can focus on diagonal matrices because of the rotational invariance of the test
matrices (Ω, Ψ). Results for dense matrices are similar.
7.2.2. A Matrix from an Application in Optimization. We also consider
a dense, complex psd matrix (Data) obtained from a real-world phase retrieval application. This matrix has dimension n = 25, 921 and exact rank 250. The first five
singular values decrease from 1 to around 0.1; there is a large gap between the fifth
and sixth singular value; the remaining nonzero singular values decay very fast. See
our paper [36] for more details about the role of sketching in this context.
7.3. Alternative Sketching Algorithms for Matrix Approximation. In
addition to the algorithms we have presented, our numerical study comprises other
methods that have appeared in the literature. We have modified all of these algorithms
to improve their numerical stability and to streamline the computations. To the extent
possible, we adopt the sketch (3.3) for all the algorithms to make their performance
more comparable.
7.3.1. Methods Based on the Sketch (3.3). We begin with two additional
methods that use the same sketch (3.3) as our algorithms.
First, let us describe a variant of a fixed-rank approximation scheme that was
proposed by Woodruff [34, Thm. 4.3, display 2]. First, form a matrix product and
compute its orthogonal–triangular factorization: ΨQ =: U T where U ∈ F`×k has
orthonormal columns. Then construct the rank-r approximation
(7.2)
Âwoo := QT † JU ∗ W Kr .
Woodruff shows that Âwoo satisfies (1.5) when the sketch size scales as k = Θ(r/ε)
and ` = Θ(k/ε2 ). Compare this result with Remark 6.5.
Second, we outline a fixed-rank approximation method that is implicit in Cohen
et al. [12, Sec. 10.1]. First, compute the r dominant left singular vectors of the
range sketch: (V , ∼, ∼) := svds(Y , r). Form a matrix product and compute its
orthogonal–triangular factorization: ΨV =: U T where U ∈ F`×r . Then form the
rank-r approximation
(7.3)
Âcemmp := V T † JU ∗ W Kr .
22
JOEL A. TROPP ET AL.
The results in Cohen et al. imply that Âcemmp satisfies (1.5) when the sketch size
scales as k = Θ(r/ε2 ) and ` = Θ(r/ε2 ).
The approximations (7.2) and (7.3) both appear similar to our fixed-rank approximation, Algorithm 7. Nevertheless, they are derived from other principles, and their
behavior is noticeably different.
7.3.2. A Method Based on an Extended Sketch. Next, we present a variant of a recent approach that requires a more complicated sketch and more elaborate
computations. The following procedure is adapted from [6, Thm. 12], using simplifications suggested in [33, Sec. 3].
Let A ∈ Fm×n be an input matrix, and let r be a target rank. Choose integer
parameters k and s that satisfy r ≤ k ≤ s ≤ min{m, n}. For consistent notation, we
also introduce a redundant parameter ` = k. Draw and fix four test matrices:
(7.4)
Ψ ∈ Fk×m ;
Ω ∈ Fn×` ;
Φ ∈ Fs×m ;
and Ξ ∈ Fn×s .
The matrices (Ψ, Ω) are standard normal, while (Φ, Ξ) are SRFTs; see subsection 3.9.
The sketch now has three components:
(7.5)
W := ΨA;
Y := AΩ;
and Z := ΦAΞ.
To store the test matrices and the sketch, we need (2k + 1)(m + n) + s(s + 2) numbers.
To obtain a rank-r approximation of the input matrix A, first compute four thin
orthogonal–triangular factorizations:
Y =: Q1 R1
ΦQ1 =: U1 T1
and
W =: R2∗ Q∗2 ;
and
Q∗2 Ξ =: T2∗ U2∗ .
Then construct the rank-r approximation
(7.6)
Âbwz := Q1 T1† JU1∗ ZU2 Kr (T2∗ )† Q∗2 .
By adapting and correcting [6, Thm. 12], one can show that Âbwz achieves (1.5)
for sketch size parameters that satisfy k = Θ(r/ε) and s = Θ((r log(1 + r))2 /ε6 ).
With this scaling, the total storage cost for the random matrices and the sketch is
Θ((m + n)r/ε + (r log(1 + r))2 /ε6 ).
The authors of [6] refer to their method as “optimal” because the scaling of the
term (m + n)r/ε in the storage cost cannot be improved [9, Thm. 4.10]. Nevertheless,
because of the ε−6 term, the bound is incomparable with the storage costs achieved
by other algorithms.
7.4. Performance with Oracle Parameter Choices. It is challenging to
compare the relative performance of sketching algorithms for matrix approximation
because of the theoretical nature of previous research. In particular, earlier work does
not offer any practical guidance for selecting the sketch size parameters.
The only way to make a fair comparison is to study the oracle performance of the
algorithms. That is, for each method, we fix the total storage, and we determine the
minimum relative error that the algorithm can achieve. This approach allows us to
see which techniques are most promising for further development. Nevertheless, we
must emphasize that the oracle performance is not achievable in practice.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
23
7.4.1. Computing the Oracle Error. It is straightforward to compare our
fixed-rank approximation methods, Algorithms 7 to 9, with the alternatives (7.2) and
(7.3) from the literature. In each case, the sketch (3.3) requires storage of (k+`)(m+n)
numbers, so we can parameterize the cost by T := k + `. For a given choice of T , we
obtain the oracle performance by minimizing the empirical approximation error for
each algorithm over all pairs (k, `) where the sum k + ` = T .
It is trickier to include the Boutsidis et al. [6, Thm. 12] method (7.6). For a given
T , we obtain the oracle performance of (7.6) by minimizing the empirical approximation error over pairs (k, s) for which the storage cost of the sketch (7.5) matches the
cost of the simple sketch (3.3). That is, (2k + 1)(m + n) + s(s + 2) ≈ T (m + n).
7.4.2. Numerical Comparison with Prior Work. For each input matrix
described in subsection 7.2, Figure 7.2 compares the oracle performance of our fixedrank approximation, Algorithm 7, against several alternative methods (7.2), (7.3),
and (7.6) from the literature. We make the following observations:
• For matrices that are well-approximated by a low-rank matrix (LowRank,
PolyDecayFast, ExpDecaySlow, ExpDecayFast, Data), our fixed-rank approximation, Algorithm 7, dominates all other methods when the storage
budget is adequate. In particular, for the rank-1 approximation of the matrix
Data, our approach achieves relative errors 3–6 orders of magnitude better
than any competitor.
• When we consider matrices that are poorly approximated by a low-rank
matrix (LowRankMedNoise, LowRankHiNoise, PolyDecaySlow), the recent
method (7.6) of Boutsidis et al. [6, Thm. 12] has the best performance, especially when the storage budget is small. But see subsection 7.4.3 for more
texture.
• Our method, Algorithm 7, performs reliably for all of the input matrices,
and it is the only method that can achieve high accuracy for the matrix
Data. Its behavior is less impressive for matrices that have poor low-rank
approximations (LowRankMedNoise, LowRankHiNoise, PolyDecaySlow), but
it is still competitive for these examples.
• The method (7.6) of Boutsidis et al. [6, Thm. 12] offers mediocre performance
for matrices with good low-rank approximations (LowRank, ExpDecaySlow,
ExpDecayFast, Data). Strikingly, this approach fails to produce a highaccuracy rank-5 approximation of the rank-10 matrix LowRank, even with
a large storage budget.
• The method (7.2) of Woodruff [34, Thm. 4.3, display 2] is competitive for
most synthetic examples, but it performs rather poorly on the matrix Data.
• The method (7.3) of Cohen et al. [12, Sec. 10.1] has the worst performance
for almost all the examples.
In summary, Algorithm 7 has the best all-around behavior, while the Boutsidis
et al. [6, Thm. 12] method (7.6) works best for matrices that have a poor low-rank
approximation. See subsection 7.6 for more discussion.
7.4.3. Structured Approximations. In this section, we investigate the effect
of imposing structure on the low-rank approximations. Figure 7.3 compares the oracle
performance of our fixed-rank approximation methods, Algorithms 7 to 9. We make
24
JOEL A. TROPP ET AL.
the following observations:
• The symmetric approximation method, Algorithm 8, and the psd approximation method, Algorithm 9, are very similar to each other for all examples.
• The structured approximations, Algorithms 8 and 9, always improve on the
unstructured approximation, Algorithm 7. The benefit is most significant
for matrices that have a poor low-rank approximation (LowRankMedNoise,
LowRankHiNoise, PolyDecaySlow).
• Algorithms 8 and 9 match or exceed the performance of the Boutsidis et al. [6,
Thm. 12] method (7.6) for all examples.
In summary, if we know that the input matrix has structure, we can achieve a
decisive advantage by enforcing the structure in the approximation.
7.5. Performance with Theoretical Parameter Choices. It remains to understand how closely we can match the oracle performance of Algorithms 7 to 9 in
practice. To that end, we must choose the sketch size parameters a priori using only
the knowledge of the target rank r and the total sketch size T . In some instances, we
may also have insight about the spectral decay of the input matrix. Figure 7.4 shows
how the fixed-rank approximation method, Algorithm 7, performs with the theoretical
parameter choices outlined in subsection 4.5. We make the following observations:
• The parameter recommendation (4.7), designed for a matrix with a flat spectral tail, works well for the matrices LowRankMedNoise, LowRankHiNoise, and
PolyDecaySlow. We also learn that this parameter choice should not be used
for matrices with spectral decay.
• The parameter recommendation (4.9), for a matrix with a slowly decaying spectrum, is suited to the examples LowRankMedNoise, LowRankHiNoise,
PolyDecaySlow, and PolyDecayFast. This parameter choice is effective for
the remaining examples as well.
• The parameter recommendation (4.10), for a matrix with a rapidly decaying
spectrum, is appropriate for the examples PolyDecayFast, ExpDecaySlow,
ExpDecayFast, and Data. This choice must not be used unless the spectrum
decays quickly.
• We have observed that the same parameter recommendations allow us to
achieve near-oracle performance for the structured matrix approximations,
Algorithms 8 and 9. As in the unstructured case, it helps if we tune the
parameter choice to the type of input matrix.
In summary, we always achieve reasonably good performance using the parameter
choice (4.9). Furthermore, if we match the parameter selection (4.7), (4.9), and
(4.10) to the spectral properties of the input matrix, we can almost achieve the oracle
performance in practice.
7.6. Recommendations. Among the fixed-rank approximation methods that
we studied, the most effective are Algorithms 7 to 9 and the Boutsidis et al. [6,
Thm. 12] method (7.6). Let us make some final observations based on our numerical
experience.
Algorithms 7 to 9 are superior to methods from the literature for input matrices
that have good low-rank approximations. Although Algorithm 7 suffers when the in-
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
25
put matrix has a poor low-rank approximation, the structured variants, Algorithms 8
and 9, match or exceed other algorithms for all the examples we tested. We have
also established that we can attain near-oracle performance for our methods using
the a priori parameter recommendations from subsection 4.5. Finally, our methods
are simple and easy to implement.
The Boutsidis et al. [6, Thm. 12] method (7.6) exhibits the best performance for
matrices that have very poor low-rank approximations when the storage budget is very
small. This benefit is diminished by its mediocre performance for matrices that do
admit good low-rank approximations. The method (7.6) requires more complicated
sketches and additional computation. Unfortunately, the analysis in [6] does not
provide guidance on implementation.
In conclusion, we recommend the sketching methods, Algorithms 7 to 9, for computing structured low-rank approximations. In future research, we will try to design
new methods that simultaneously dominate our algorithms and (7.6).
Appendix A. Analysis of the Low-Rank Approximation. In this appendix,
we develop theoretical results on the performance of the basic low-rank approximation (4.3) implemented in Algorithms 3 and 4.
A.1. Facts about Random Matrices. Our arguments require classical formulae for the expectations of functions of a standard normal matrix. In the real case,
these results are [19, Prop. A.1 and A.6]. The complex case follows from the same
principles, so we omit the details.
Fact A.1. Let G ∈ Ft×s be a standard normal matrix. For all matrices B and
C with conforming dimensions,
E kBGCk2F = βkBk2F kCk2F .
(A.1)
Furthermore, if t > s + α,
(A.2)
E kG† k2F =
s
1
1
·
= · f (s, t).
β t−s−α
β
The numbers α and β are given by (2.2); the function f is introduced in (2.3).
A.2. Results from Randomized Linear Algebra. Our arguments also depend heavily on the analysis of randomized low-rank approximation developed in [19,
Sec. 10]. We state these results using the familiar notation from sections 3 and 4.
Fact A.2 (Halko et al. 2011). Fix A ∈ Fm×n . Let % be a natural number such
that % < k − α. Draw the random test matrix Ω ∈ Fk×n from the standard normal
distribution. Then the matrix Q computed by (4.1) satisfies
2
EΩ kA − QQ∗ Ak2F ≤ (1 + f (%, k)) · τ%+1
(A).
The number α is given by (2.2); the function f is introduced in (2.3).
This result follows immediately from the proof of [19, Thm. 10.5] using Fact A.1 to
handle both the real and complex case simultaneously.
A.3. Proof of Theorem 4.3: Frobenius Error Bound. In this section, we
establish a second Frobenius-norm error bound for the low-rank approximation (4.3).
We maintain the notation from sections 3 and 4, and we state explicitly when we are
making distributional assumptions on the test matrices.
26
JOEL A. TROPP ET AL.
10 0
10 0
relative error
10 -2
10 0
10 -4
10 -1
10 -6
10 -1
10 -8
10 -2
10 -10
13
26
52
104
208
13
26
52
104
208
13
26
T
T
(a) LowRank
52
104
208
T
(b) LowRankMedNoise
(c) LowRankHiNoise
10 0
10 0
10 0
relative error
10 -2
10 -2
10 -1
10 -4
10 -4
10 -6
10 -2
10 -6
10 -8
10 -3
10 -10
13
26
52
104
208
13
26
T
104
208
13
26
T
(d) PolyDecaySlow
52
104
208
T
(e) PolyDecayFast
(f) ExpDecaySlow
10 0
10 0
relative error
52
10 0
10
-2
10 -2
10
-4
10 -4
10
-6
-6
10 -2
10
10 -4
10 -6
10 -8
10 -8
10 -10
10 -10
13
26
52
104
T
(g) ExpDecayFast
208
Woo (7.2)
Cemmp (7.3)
Bwz (7.6)
Algorithm 7
10 -8
10 -10
5
10
20
40
T
(h) Data (r = 1)
80
13
26
52
104
T
(i) Data (r = 5)
Fig. 7.2: Oracle performance of sketching algorithms for fixed-rank matrix approximation as a function of storage cost. For each of the input
matrices described in subsection 7.2, we compare the oracle performance of our
fixed-rank approximation, Algorithm 7, against alternative methods (7.2), (7.3),
and (7.6) from the literature. The matrix dimensions are m = n = 103 for the
synthetic examples and m = n = 25, 921 for the matrix Data from the phase retrieval application. Each approximation has rank r = 5, unless otherwise stated.
The variable T on the horizontal axis is (proportional to) the total storage used by
each sketching method. Each data series displays the best relative error (7.1) that
the specified algorithm can achieve with storage T . See subsection 7.4.1 for details.
208
27
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
10 0
10 0
10 0
relative error
10 -2
10 -1
10 -4
10 -6
10 -1
10 -2
10 -8
10 -10
13
26
52
104
208
13
26
52
(a) LowRank
208
13
26
104
208
(c) LowRankHiNoise
10 0
10 0
10 -2
10 -2
10 -1
52
T
(b) LowRankMedNoise
10 0
relative error
104
T
T
10 -4
10
10 -4
-2
10 -6
10 -6
10 -3
10 -8
10 -10
13
26
52
104
208
13
26
T
52
104
208
13
26
T
(d) PolyDecaySlow
52
104
208
T
(e) PolyDecayFast
(f) ExpDecaySlow
10 0
10 0
relative error
10 -2
10
-2
10 -2
-4
10 -4
10 -6
10
-6
10 -8
10 -8
10
10 -10
10 -4
10 -6
10 -10
13
26
52
104
T
(g) ExpDecayFast
208
Algorithm 7
Algorithm 8
Algorithm 9
10 -8
10 -10
5
10
20
40
T
(h) Data (r = 1)
80
13
26
52
104
T
(i) Data (r = 5)
Fig. 7.3: Oracle performance of sketching algorithms for structured fixedrank matrix approximation as a function of storage cost. For each of the
input matrices described in subsection 7.2, we compare the oracle performance of the
unstructured approximation (Algorithm 7), the conjugate symmetric approximation
(Algorithm 8), and the positive-semidefinite approximation (Algorithm 9). The
matrix dimensions are m = n = 103 for the synthetic examples and m = n = 25, 921
for the matrix Data from the phase retrieval application. Each approximation has
rank r = 5, unless otherwise stated. The variable T on the horizontal axis is
(proportional to) the total storage used by each sketching method. Each data series
displays the best relative error (7.1) that the specified algorithm can achieve with
storage T . See subsection 7.4.1 for details.
208
28
JOEL A. TROPP ET AL.
relative error
10 0
10 0
10 -1
10 -2
10 -1
13
26
52
104
208
13
26
52
(a) LowRankMedNoise
208
(b) LowRankHiNoise
10 0
10 0
104
T
T
10 0
relative error
10 -2
10 -2
10 -1
10 -4
10 -4
10 -2
10 -6
10 -6
10 -8
10 -3
10 -10
13
26
52
104
208
13
26
T
52
104
208
13
26
52
T
(c) PolyDecaySlow
104
208
T
(d) PolyDecayFast
(e) ExpDecaySlow
10 0
10 0
relative error
10
-2
10
-2
10 -2
10 -4
10 -4
-6
-6
10
10
10 -8
10 -8
-10
-10
10
10
13
26
52
104
T
(f) ExpDecayFast
208
10 -4
10 -6
10
ORACLE
FLAT
DECAY
RAPID
-8
10 -10
5
10
20
40
T
(g) Data (r = 1)
80
13
26
52
104
T
(h) Data (r = 5)
Fig. 7.4: Performance of a sketching algorithm for fixed-rank matrix approximation with a priori parameter choices. For each of the input matrices
described in subsection 7.2, we compare the oracle performance of the fixed-rank
approximation, Algorithm 7, against its performance at theoretically motivated parameter choices. The matrix dimensions are m = n = 103 for the synthetic examples
and m = n = 25, 921 for the matrix Data from the phase retrieval application. Each
approximation has rank r = 5, unless otherwise stated. The variable T on the horizontal axis is (proportional to) the total storage used by each sketching method.
The oracle performance is drawn from Figure 7.2. Each data series displays the
relative error (7.1) that Algorithm 7 achieves for a specific parameter selection.
The parameter choice FLAT (4.7) is designed for matrices with a flat spectral tail;
DECAY (4.9) is for a slowly decaying spectrum; RAPID (4.10) is for a rapidly decaying
spectrum. See subsection 7.5 for details.
208
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
29
A.3.1. Decomposition of the Approximation Error. Fact A.2 formalizes
the intuition that A ≈ Q(Q∗ A). The main object of the proof is to demonstrate that
X ≈ Q∗ A. The first step in the argument is to break down the approximation error
into these two parts.
Lemma A.3. Let A be an input matrix, and let  = QX be the approximation
defined in (4.3). The approximation error decomposes as
kA − Âk2F = kA − QQ∗ Ak2F + kX − Q∗ Ak2F .
We omit the proof, which is essentially just the Pythagorean theorem.
A.3.2. Approximating the Second Factor. Next, we develop an explicit expression for the error in the approximation X ≈ Q∗ A. It is convenient to construct
a matrix P ∈ Fn×(n−k) with orthonormal columns that satisfies
P P ∗ = I − QQ∗ .
(A.3)
Introduce the matrices
Ψ1 := ΨP ∈ F`×(n−k)
(A.4)
and Ψ2 := ΨQ ∈ F`×k .
We are now prepared to state the result.
Lemma A.4. Assume that the matrix Ψ2 has full column-rank. Then
X − Q∗ A = Ψ†2 Ψ1 (P ∗ A).
(A.5)
The matrices Ψ1 and Ψ2 are defined in (A.4).
Proof. Recall that W = ΨA, and calculate that
W = ΨA = ΨP P ∗ A + ΨQQ∗ A = Ψ1 (P ∗ A) + Ψ2 (Q∗ A).
The second relation holds because P P ∗ +QQ∗ = I. Then we use (A.4) to identify Ψ1
and Ψ2 . By hypothesis, the matrix Ψ2 has full column-rank, so we can left-multiply
the last display by Ψ†2 to attain
Ψ†2 W = Ψ†2 Ψ1 (P ∗ A) + Q∗ A.
Turning back to (4.2), we identify X = Ψ†2 W .
A.3.3. The Expected Frobenius-Norm Error in the Second Factor. We
are now prepared to compute the average Frobenius-norm error in approximating
Q∗ A by means of the matrix X. In contrast to the previous steps, this part of the
argument relies on distributional assumptions on the test matrix Ψ. Remarkably, for
a Gaussian test matrix, X is even an unbiased estimator of the factor Q∗ A.
Lemma A.5. Assume that Ψ ∈ F`×n is a standard normal matrix that is independent from Ω. Then
EΨ [X − Q∗ A] = 0.
Furthermore,
EΨ kX − Q∗ Ak2F = f (k, `) · kA − QQ∗ Ak2F .
30
JOEL A. TROPP ET AL.
Proof. Observe that P and Q are partial isometries with orthogonal ranges. Owing to the marginal property of the standard normal distribution, the random matrices Ψ1 and Ψ2 are statistically independent standard normal matrices. In particular,
Ψ2 ∈ F`×k almost surely has full column-rank because (3.1) requires that ` ≥ k.
First, take the expectation of the identity (A.5) to see that
EΨ [X − Q∗ A] = EΨ2 EΨ1 [Ψ†2 Ψ1 P ∗ A] = 0.
In the first relation, we use the statistical independence of Ψ1 and Ψ2 to write the
expectation as an iterated expectation. Then we observe that Ψ1 is a matrix with
zero mean.
Next, take the expected squared Frobenius norm of (A.5) to see that
EΨ kX − Q∗ Ak2F = EΨ2 EΨ1 kΨ†2 Ψ1 (P ∗ A)k2F
= β · EΨ2 kΨ†2 k2F · kP ∗ Ak2F = f (k, `) · kP ∗ Ak2F .
The last two identities follow from (A.1) and (A.2) respectively, where we use the fact
that Ψ2 ∈ F`×k . To conclude, note that
kP ∗ Ak2F = kP P ∗ Ak2F = kA − QQ∗ Ak2F .
The first relation holds because P is a partial isometry and the Frobenius norm is
unitarily invariant. Last, we apply the definition (A.3) of P .
A.3.4. Proof of Theorem 4.3. We are now prepared to complete the proof
of the Frobenius-norm error bound stated in Theorem 4.3. For this argument, we
assume that the test matrices Ω ∈ Fn×k and Ψ ∈ F`×m are drawn independently
from the standard normal distribution.
According to Lemma A.3,
kA − Âk2F = kA − QQ∗ Ak2F + kX − Q∗ Ak2F .
Take the expectation of the last display to reach
E kA − Âk2F = EΩ kA − QQ∗ Ak2F + EΩ EΨ kX − Q∗ Ak2F
= (1 + f (k, `)) · EΩ kA − QQ∗ Ak2F
2
≤ (1 + f (k, `)) · (1 + f (%, k)) · τ%+1
(A).
In the first line, we use the independence of the two random matrices to write the
expectation as an iterated expectation. To reach the second line, we apply Lemma A.5
to the second term. Invoke the randomized linear algebra result, Fact A.2. Finally,
minimize over eligible indices % < k − α.
REFERENCES
[1] N. Ailon and B. Chazelle, Approximate nearest neighbors and the fast JohnsonLindenstrauss transform, in STOC’06:
Proceedings of the 38th Annual ACM
Symposium on Theory of Computing, ACM, New York, 2006, pp. 557–563,
doi:10.1145/1132516.1132597, http://dx.doi.org/10.1145/1132516.1132597.
[2] N. Ailon and B. Chazelle, The fast Johnson-Lindenstrauss transform and approximate
nearest neighbors, SIAM J. Comput., 39 (2009), pp. 302–322, doi:10.1137/060673096,
http://dx.doi.org/10.1137/060673096.
SKETCHING ALGORITHMS FOR MATRIX APPROXIMATION
31
[3] J. Bourgain, S. Dirksen, and J. Nelson, Toward a unified theory of sparse dimensionality reduction in Euclidean space, Geom. Funct. Anal., 25 (2015), pp. 1009–1088,
doi:10.1007/s00039-015-0332-9, http://dx.doi.org/10.1007/s00039-015-0332-9.
[4] C. Boutsidis, D. Garber, Z. Karnin, and E. Liberty, Online principal components analysis, in Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SIAM, 2015, pp. 887–901.
[5] C. Boutsidis and A. Gittens, Improved matrix algorithms via the subsampled randomized Hadamard transform, SIAM J. Matrix Anal. Appl., 34 (2013), pp. 1301–1340,
doi:10.1137/120874540, http://dx.doi.org/10.1137/120874540.
[6] C. Boutsidis, D. Woodruff, and P. Zhong, Optimal principal component analysis in distributed and streaming models, in Proc. 48th Ann. Symp. Theory of Computing (STOC
2016), Cambridge, MA, 2016.
[7] S. Boyd and L. Vandenberghe, Convex optimization, Cambridge University Press,
Cambridge,
2004,
doi:10.1017/CBO9780511804441,
http://dx.doi.org/10.1017/
CBO9780511804441.
[8] K. Clarkson and D. Woodruff, Low-rank PSD approximation in input-sparsity time. Unpublished, Jan. 2017.
[9] K. L. Clarkson and D. P. Woodruff, Numerical linear algebra in the streaming model, in
Proc. 41st ACM Symposium on Theory of Computing (STOC), Bethesda, 2009.
[10] K. L. Clarkson and D. P. Woodruff, Low rank approximation and regression in input sparsity time, in STOC’13—Proceedings of the 2013 ACM Symposium on Theory of Computing, ACM, New York, 2013, pp. 81–90, doi:10.1145/2488608.2488620,
http://dx.doi.org/10.1145/2488608.2488620.
[11] M. Cohen, Nearly tight oblivious subspace embeddings by trace inequalities, in Proc. 27th Ann.
ACM-SIAM Symp. Discrete Algorithms (SODA), Arlington, Jan. 2016, pp. 278–287.
[12] M. B. Cohen, S. Elder, C. Musco, C. Musco, and M. Persu, Dimensionality reduction
for k-means clustering and low rank approximation, in Proceedings of the Forty-Seventh
Annual ACM on Symposium on Theory of Computing, ACM, 2015, pp. 163–172.
[13] M. B. Cohen, J. Nelson, and D. P. Woodruff, Optimal Approximate Matrix Product in Terms of Stable Rank, in 43rd International Colloquium on Automata, Languages, and Programming (ICALP 2016), I. Chatzigiannakis, M. Mitzenmacher, Y. Rabani, and D. Sangiorgi, eds., vol. 55 of Leibniz International Proceedings in Informatics (LIPIcs), Dagstuhl, Germany, 2016, Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, pp. 11:1–11:14, doi:http://dx.doi.org/10.4230/LIPIcs.ICALP.2016.11, http://drops.
dagstuhl.de/opus/volltexte/2016/6278.
[14] J. Demmel, I. Dumitriu, and O. Holtz, Fast linear algebra is stable, Numer.
Math., 108 (2007), pp. 59–91, doi:10.1007/s00211-007-0114-x, http://dx.doi.org/10.1007/
s00211-007-0114-x.
[15] D. Feldman, M. Schmidt, and C. Sohler, Turning big data into tiny data: constant-size
coresets for k-means, PCA and projective clustering, in Proceedings of the Twenty-Fourth
Annual ACM-SIAM Symposium on Discrete Algorithms, SIAM, Philadelphia, PA, 2012,
pp. 1434–1453.
[16] D. Feldman, M. Volkov, and D. Rus, Dimensionality reduction of massive sparse datasets
using coresets, in Advances in Neural Information Processing Systems 29 (NIPS 2016),
2016.
[17] A. Frieze, R. Kannan, and S. Vempala, Fast Monte-Carlo algorithms for finding low-rank
approximations, J. ACM, 51 (2004), pp. 1025–1041, doi:10.1145/1039488.1039494, http:
//dx.doi.org/10.1145/1039488.1039494.
[18] M. Gu, Subspace iteration randomization and singular value problems, SIAM J. Sci. Comput., 37 (2015), pp. A1139–A1173, doi:10.1137/130938700, http://dx.doi.org/10.1137/
130938700.
[19] N. Halko, P. G. Martinsson, and J. A. Tropp, Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions, SIAM Rev., 53
(2011), pp. 217–288.
[20] N. J. Higham, Matrix nearness problems and applications, in Applications of matrix theory
(Bradford, 1988), Oxford Univ. Press, New York, 1989, pp. 1–27.
[21] P. Jain, C. Jin, S. M. Kakade, P. Netrapalli, and A. Sidford, Streaming PCA: Matching
matrix Bernstein and near-optimal finite sample guarantees for Oja’s algorithm, in 29th
Annual Conference on Learning Theory, 2016, pp. 1147–1164.
[22] H. Li, G. C. Linderman, A. Szlam, K. P. Stanton, Y. Kluger, and M. Tygert, Algorithm
971: An implementation of a randomized algorithm for principal component analysis,
ACM Trans. Math. Softw., 43 (2017), pp. 28:1–28:14, doi:10.1145/3004053, http://doi.
32
JOEL A. TROPP ET AL.
acm.org/10.1145/3004053.
[23] Y. Li, H. L. Nguyen, and D. P. Woodruff, Turnstile streaming algorithms might as well
be linear sketches, in STOC’14—Proceedings of the 2014 ACM Symposium on Theory of
Computing, ACM, New York, 2014, pp. 174–183.
[24] E. Liberty, Accelerated dense random projections, PhD thesis, Yale Univ., New Haven, 2009.
[25] M. W. Mahoney, Randomized algorithms for matrices and data, Foundations and Trends R
in Machine Learning, 3 (2011), pp. 123–224.
[26] P.-G. Martinsson, V. Rokhlin, and M. Tygert, A randomized algorithm for the
decomposition of matrices, Appl. Comput. Harmon. Anal., 30 (2011), pp. 47–68,
doi:10.1016/j.acha.2010.02.003, http://dx.doi.org/10.1016/j.acha.2010.02.003.
[27] X. Meng and M. W. Mahoney, Low-distortion subspace embeddings in input-sparsity
time and applications to robust linear regression, in STOC’13—Proceedings of the
2013 ACM Symposium on Theory of Computing, ACM, New York, 2013, pp. 91–100,
doi:10.1145/2488608.2488621, http://dx.doi.org/10.1145/2488608.2488621.
[28] J. Nelson and H. L. Nguyen, OSNAP: faster numerical linear algebra algorithms via sparser
subspace embeddings, in 2013 IEEE 54th Annual Symposium on Foundations of Computer Science—FOCS 2013, IEEE Computer Soc., Los Alamitos, CA, 2013, pp. 117–126,
doi:10.1109/FOCS.2013.21, http://dx.doi.org/10.1109/FOCS.2013.21.
[29] J. Nelson and H. L. Nguyen, Lower bounds for oblivious subspace embeddings, in Automata, languages, and programming. Part I, vol. 8572 of Lecture Notes in Comput.
Sci., Springer, Heidelberg, 2014, pp. 883–894, doi:10.1007/978-3-662-43948-7 73, http:
//dx.doi.org/10.1007/978-3-662-43948-7 73.
[30] C. H. Papadimitriou, P. Raghavan, H. Tamaki, and S. Vempala, Latent semantic indexing: a probabilistic analysis, J. Comput. System Sci., 61 (2000), pp. 217–235,
doi:10.1006/jcss.2000.1711, http://dx.doi.org/10.1006/jcss.2000.1711. Special issue on the
Seventeenth ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems (Seattle, WA, 1998).
[31] J. A. Tropp, Improved analysis of the subsampled randomized Hadamard transform, Adv.
Adapt. Data Anal., 3 (2011), pp. 115–126, doi:10.1142/S1793536911000787, http://dx.doi.
org/10.1142/S1793536911000787.
[32] J. A. Tropp, A. Yurtsever, M. Udell, and V. Cevher, Randomized single-view algorithms
for low-rank matrix approximation, ACM Report 2017-01, Caltech, Pasadena, Jan. 2017.
Available at http://arXiv.org/abs/1609.00048, v1.
[33] J. Upadhyay, Fast and space-optimal low-rank factorization in the streaming model with application in differential privacy. Available at http://arXiv.org/abs/1604.01429, Apr. 2016.
[34] D. P. Woodruff, Sketching as a tool for numerical linear algebra, Found. Trends Theor.
Comput. Sci., 10 (2014), pp. iv+157.
[35] F. Woolfe, E. Liberty, V. Rokhlin, and M. Tygert, A fast randomized algorithm for the
approximation of matrices, Appl. Comput. Harmon. Anal., 25 (2008), pp. 335–366.
[36] A. Yurtsever, M. Udell, J. A. Tropp, and V. Cevher, Sketchy decisions: Convex low-rank
matrix optimization with optimal storage. Manuscript, May 2016.
| 8cs.DS
|
arXiv:1706.00378v1 [math.ST] 1 Jun 2017
New goodness-of-fit diagnostics for conditional
discrete response models
Igor Kheifets∗ and Carlos Velasco†
June 2, 2017
Abstract
This paper proposes new specification tests for conditional models with discrete responses, which are key to apply efficient maximum likelihood methods, to
obtain consistent estimates of partial effects and to get appropriate predictions of
the probability of future events. In particular, we test the static and dynamic ordered choice model specifications and can cover infinite support distributions for
e.g. count data. The traditional approach for specification testing of discrete response models is based on probability integral transforms of a jittered discrete data
which leads to continuous uniform iid series under the true conditional distribution.
Then, standard specification testing techniques for continuous variables could be
applied to the transformed series, but the extra randomness from jitters affects
the power properties of these methods. We investigate in this paper an alternative
transformation based only on original discrete data that avoids any randomization. We analyze the asymptotic properties of goodness-of-fit tests based on this
new transformation and explore the properties in finite samples of a bootstrap algorithm to approximate the critical values of test statistics which are model and
parameter dependent. We show analytically and in simulations that our approach
dominates the methods based on randomization in terms of power. We apply the
new tests to models of the monetary policy conducted by the Federal Reserve.
Keywords: Specification tests, count data, dynamic discrete choice models, conditional probability integral transform.
JEL classification: C12, C22, C52.
∗
†
ITAM, Mexico. Email: [email protected]
Department of Economics, Universidad Carlos III de Madrid. Email: [email protected]
1
1
INTRODUCTION
Many statistical models specify the conditional distribution of a discrete response variable given some explanatory variables, including the description of binary, multinomial,
ordered choice and count data. In this paper we analyze goodness-of-fit tests for both
static models with covariates as well as dynamic ordered choice and count data models,
where the conditioning information set may also include past information on the discrete
variable and a set of (contemporaneous) explanatory variables which frequently appear in
the social sciences, see Kedem and Fokianos (2002) and Greene and Hensher (2010). For
example, dynamic models are popular in macroeconomic applications, see for instance
Hamilton and Jordá (2002), Dolado and Maria-Dolores (2002) and Basu and de Jong
(2007) for modeling central banks decisions or Kauppi and Saikkonen (2008) and Startz
(2008) for predicting US recessions; in finance, see e.g. Rydberg and Shephard (2003) for
modeling the size of asset price movements and Fokianos et al. (2009) for the number of
transactions per minute of a particular stock.
Suppose we observe the random variables {Yt , Xt′}Tt=1 and consider the information sets
Ωt = {Xt , Yt−1 , Xt−1 , Yt−2 , Xt−2 , . . .} for each period t = 1, 2, . . . , T . We are interested
in testing the null hypothesis that the distribution of Yt conditional on Ωt is in the
parametric family Ft,θ (· | Ωt ), i.e.
H0 : Yt | Ωt ∼ Ft,θ0 (· | Ωt ) for some θ0 ∈ Θ, t = 1, 2, . . . , T,
where Θ ⊂ Rm is the parameter space, while the alternative hypothesis (H1 ) for the
omnibus test would be the negation of H0 .
We consider a class M of discrete conditional distributions defined on K = {1, 2, . . . , K},
for integer K > 1 or on K = {1, 2, . . . , ∞} such that for all F ∈ M it holds that F (0) = 0,
P
f (k) := F (k)−F (k − 1) > 0 for all k = 1, 2, . . . and k∈K f (k) = 1. This setup includes
numerous models that have been used extensively in applied work both for dynamic and
for iid data, here we describe briefly two of them.
Example 1 (Dynamic multinomial ordered choice model). The discrete responses Yt are
2
assumed to be generated by the rule
1
2
Yt =
K
Vt∗ ≤ τ 1
if
if τ 1 < Vt∗ ≤ τ 2
..
.
Vt∗ > τ K−1 ,
if
where Vt∗ is a continuous latent variable and τ 1 , . . . , τ K−1 are threshold parameters that
define K intervals in R. In a simple model, e.g. Basu and de Jong (2007), the latent
variable is determined through the linear equation
Vt∗ = Xt′ β + ρYt−1 + εt ,
where Xt is a vector of stationary exogenous regressors, β a vector of regression parameters, εt is the shock in each period, and Yt−1 could be replaced by any function of the
past {Yt−1 , . . . , Yt−n } for some finite n. The cdf of εt , Fε , is going to determine the class
of multinomial model, i.e. ordered multinomial probit (if εt is standard normal) or logit
(if εt is logistic), since Ft,θ0 is defined at once from
Pr (Yt = k | Ωt ) = Pr (τ k−1 < Vt∗ ≤ τ k | Ωt )
= Fε (τ k − Xt′ β − ρYt−1 ) − Fε (τ k−1 − Xt′ β − ρYt−1 ) ,
′
with τ 0 = −∞ and τ K = ∞ and θ 0 = (β ′ , ρ, τ 1 , . . . , τ K−1) .
Example 2 (Poisson Model). The variate Yt = Yt∗ + 1 is defined on the counts Yt∗ =
0, 1, 2, . . . which are assumed to follow a conditional Poisson distribution
Yt∗ | Ωt ∼ Poisson(λt ),
where the conditional mean can depend on covariates through an exponential link as
λt = exp(Xt′ β) or on previous observations through an identity link as λt = α0 + α1 λt−1 +
∗
ρYt−1
, e.g. Fokianos et al. (2009), or through the logarithmic canonical link as log(λt ) =
Xt′ β + ρet−1 , where et = (Yt∗ − λt ) /λt are scaled and centered errors, e.g. Davis et al.
(2003).
3
Despite that a correct specification is key to apply efficient maximum likelihood methods, to obtain consistent estimates of partial effects and to get appropriate predictions of
the probability of future events, empirical researchers typically do not perform goodness
of fit testing of such models as they would do in a continuous case. In general, there
are only a few specification tests available for discrete data, see Mora and Moro-Egido
(2007). Two of them, the test of the Generalized Linear Model (GLM) of Stute and Zhu
(2002) and the conditional Kolmogorov test of Andrews (1997), based on the specification of the conditional mean for binary data, can be adapted for this purpose and we
discuss this possibility and compare it to our approach in Section 6. A related test to
Andrews derived for time series by Corradi and Swanson (2006) could be adapted also
for discrete data, but this is testing a different null hypothesis concerning a distribution
given a finite conditioning set not characterizing the complete dynamics of the process.
There are also tests designed specifically for Poisson models (see e.g. Neumann 2011;
Fokianos and Neumann, 2013).
In what follows we propose conditional, dynamic discrete analogs of the KolmogorovSmirnov goodness of fit measure that can exploit different restrictions derived from the
martingale difference property of a particular transformation of the data under the null
hypothesis. This property is derived from the specification of a complete dynamic model
given the information set generated by all the past observations of the discrete response
and other explanatory variables and is used to build the asymptotic theory for our tests.
Under i.i.d. assumptions this martingale difference property leads to an exact independence of the transformation sequence under the null and a much simpler parallel asymptotic theory.
When the fitted distribution is continuous, the relative distribution of Yt compared
to Ft,θ0 defined as the cdf of the Rosenblatt’s (1952) transforms, also called conditional
Probability Integral Transforms (PIT),
Ut (θ0 ) := Ft,θ0 (Yt | Ωt ) ,
t = 1, 2, . . . , T
is standard uniform and Ut (θ 0 ) are distributed as independent [0, 1] uniform random
4
variables under H0 . This serves as a basis for several specification tests of H0 , see e.g.
Bai (2003) and Kheifets (2015) for dynamic models and Delgado and Stute (2008) for
independent and identical distributed (iid) data. However Rosenblatt transformation is
not appropriate for discrete support random variables, producing non-iid pseudo residuals
even under the null of correct specification. To solve the limitations of PIT-based testing
techniques for discrete data, several alternative transforms have been proposed, see Jung,
Kukuk and Liesenfeld (2006), Czado, Gneiting and Held (2009) and references therein.
An easy and popular way is to randomize, i.e. to interpolate the discrete values of Yt with
independent noise in [0, 1], recent references include Kheifets and Velasco (2013) and Lee
(2014). Unfortunately the additional simulated noise affects the power of the tests and
may lead to different conclusions depending on the simulation outcome.
In this paper instead, we consider a nonrandomized transform Yt 7→ It,θ0 (u) for u ∈
[0, 1],
It,θ0 (u) :=
0,
u − Ut− (θ 0 )
,
Ut (θ0 ) − Ut− (θ0 )
1,
u ≤ Ut− (θ0 ) ;
Ut− (θ0 ) ≤ u ≤ Ut (θ0 ) ;
(1)
Ut (θ0 ) ≤ u,
where Ut− (θ0 ) := Ft,θ0 (Yt − 1 | Ωt ). This transform, conditional on data, is nonrandomized in the sense that it does not depend on extra sources of randomness, as opposed to
interpolation transforms discussed in the next section. The unconditional version of this
transform appears in Handcock and Morris (1999) and more recently in Czado, Gneiting
and Held (2009) where it is used for calibration, but no formal tests are proposed there.
This transformation can also be seem as a particular case of the multilinear extension
as defined in Genest, Nešlehová and Rémillard (2014). As we show below, for every
u ∈ [0, 1], It,θ0 (u) − u constitute a martingale difference sequence (MDS) with respect to
Ωt under H0 and can be used for testing H0 as It,θ0 (u) loses this property when the model
is misspecified. For instance, we can compute the pseudo empirical relative distribution
of Yt compared to Ft,θ0
T
1X
Feθ0 (u) :=
It,θ0 (u) ,
T t=1
5
u ∈ [0, 1] ,
which can be contrasted with the uniform cdf using the following empirical process
S1T (u) :=
T
1 X
T 1/2
t=1
{It,θ0 (u) − u} = T 1/2 Feθ0 (u) − u ,
which converges weakly to a Gaussian process. In addition, in order to control dynamics
in It,θ0 (u), we can compare the joint pseudo empirical cdf with the uniform on a square
using the biparameter process
S2T
T
X
1
(u) :=
{It,θ0 (u1 ) It−1,θ0 (u2 ) − u1 u2 } ,
(T − 1)1/2 t=2
(2)
where u = (u1 , u2). To obtain feasible tests we need to consider norms of SjT for j = 1, 2.
R
We use the Cramer-von Mises SjT (u)2 dϕ (u) for some absolute continuous measure ϕ
in [0, 1]j , or Kolmogorov-Smirnov supu∈[0,1]j |SjT (u)| norms.
When the parameter θ0 is unknown under the null, we use an estimate b
θT and ac-
count for the parameter estimation effect in the p-value computation with a parametric
bootstrap method. It might be possible also to derive, e.g. martingale, distribution-free
transforms, but since they typically need to be programmed on a case by case basis for
each model, so can be impractical, and are beyond the scope of this paper. As far as we
know, our proposal is the first formal specification test of ordered discrete choice models
which accounts properly for parameter uncertainty and is based on a nonrandomized
transform, which makes it attractive in terms of power against a wide set of alternative
hypotheses.
The rest of the paper is organized as follows. In the next section, we describe different
alternatives to the PIT. In Sections 3 and 4, we provide the main asymptotic properties
of the nonrandomized transforms and of the resulting univariate and bivariate empirical
processes using martingale theory. In particular, we establish weak limits under fixed
and local alternatives accounting for parameter estimation effect. Section 5 discusses
the implementation of new tests with a simple bootstrap algorithm. Section 6 provides a
small simulation exercise and an application exploring the properties of specification tests
based on both randomized and non randomized transformations. Then we conclude. All
proofs are contained in the Appendix.
6
2
ALTERNATIVES TO PIT FOR DISCRETE DATA
In order to further motivate the nonrandomized transform It,θ0 defined in (1), we introduce the randomized PIT,
Utr (θ 0 ) := Ut− (θ0 ) + ZtU Ut (θ0 ) − Ut− (θ 0 ) ,
(3)
where {ZtU }Tt=1 are independent standard uniform random variables, and independent of
Yt . Alternatively, Utr can be obtained by applying the standard continuous PIT to the
continuous random variable Yt† := Yt − 1 + Zt , where {Zt }Tt=1 are iid with any continuous
cdf FZ on [0, 1]. Indeed, we can construct the cdf of Yt† ,
†
Ft,θ
(y | Ωt ) = Ft,θ0 (⌊y⌋ | Ωt ) + FZ (y − ⌊y⌋) (Ft,θ 0 (⌊y + 1⌋ | Ωt ) − Ft,θ0 (⌊y⌋ | Ωt )) ,
0
where ⌊y⌋ is the floor function, i.e. the maximum integer not exceeding y, and find that
†
†
Utr (θ0 ) = Ft,θ
Y
|
Ω
t ,
t
0
for ZtU = FZ (Zt ) and any choice of FZ , see Kheifets and Velasco (2013). Note that the
cdf of Yt† conditional on Ωt and {Ωt , Zt−1 , Zt−2 , . . . , Z1 } coincide. Under H0 , Utr (θ0 ) are
iid U [0, 1] variables as under any continuous distribution specification, while Ut (θ0 ) and
Ut− (θ0 ) are not independent nor U [0, 1]. Using the typical discrepancy measures, the
empirical cdf of Utr (θ0 ), estimated using the randomized transform Yt 7→ 1{Utr (θ0 ) ≤ u},
T
1X
r
b
Fθ0 (u) :=
1{Utr (θ0 ) ≤ u} ,
T t=1
u ∈ [0, 1] ,
can be compared to the uniform cdf. Kheifets and Velasco (2013) then test H0 using
empirical process based on the randomized transform
R1T (u) := T
1/2
n
o
Fbθr0 (u) − u =
T
1 X
T 1/2
t=1
[1{Utr (θ0 ) ≤ u} − u] ,
u ∈ [0, 1] .
We can also consider reducing the dependence on a particular outcome of the noise
ZtU in (3) and in the randomized transform by taking averages over M replications of
{ZtU }Tt=1 , conditional on the original data, similar to “average-jittering” of Machado and
7
Santos Silva (2005). Suppose that for each t we have M independent sequences of uniform
U
r
U[0, 1] noises Zt,m
, m = 1, 2, . . . , M, which generate Ut,m
(θ 0 ) according to (3). Define
the M-random transform Yt 7→ It,θ0 ,M (Yt , u),
M
1 X r
It,θ0 ,M (Yt , u) :=
1 Ut,m (θ0 ) ≤ u ,
M m=1
which takes values on the set {0, 1/M, 2/M, . . . , 1} and has mean u under H0 . Then the
cdf of Utr (θ0 ) is estimated by
T
1X
r
b
It,θ0 ,M (Yt , u) ,
Fθ0 ,M (u) :=
T t=1
u ∈ [0, 1] .
Note that with M = 1 we are back to Fbθr0 (u), and therefore, we can generalize R1T to
n
o
R1T,M (u) := T 1/2 Fbθr0 ,M (u) − u ,
u ∈ [0, 1] .
In order to propose specification tests, following Handcock and Morris (1999), we define the discrete relative distribution of Yt compared to Ft,θ0 as the cdf of Utr (θ0 ). Under
H0 , the discrete relative distribution is the uniform U [0, 1]. As we show in the next
section, three consistent estimators of the discrete relative distribution of Yt compared
to Ft,θ 0 can be ordered in terms of efficiency in the following way: Feθ0 (u) (the most
efficient), Fbθr0 ,M (u) and Fbθr0 (u). This ordering is determined by the amount of noise
introduced in the definitions of the transforms: i.e. in nonrandomized, M-randomized
and (1-)randomized transforms. The nonrandomized transform can be equivalently obtained by integrating out the extra noise in the randomized transform It,θ0 (Yt , u) =
R
1{Utr (θ0 ) ≤ u} dFZ or by taking the number of replications M to infinity, thus com-
pletely removing the noise from the estimate of the discrete relative distribution and other
functionals of the transforms. The efficiency of the nonrandomized transform translates
into the increased power of the specification tests based on this transform, whose properties we study next.
8
3
PROPERTIES OF EMPIRICAL PROCESSES BASED
ON THE NONRANDOMIZED TRANSFORM
As shown in the next lemma, the building blocks of Feθ0 (u) , It,θ0 (u) − u, constitute a
martingale difference sequence (MDS) with respect to Ωt , and therefore Feθ0 (u) is an
unbiased and consistent estimate of the uniform cdf under the null, a reasonable basis
for developing tests of H0 . Moreover, the MDS property will allow us to establish the
asymptotic properties of our test without imposing any additional restrictions. Let for
u, v ∈ [0, 1]
γ t,θ0 (u, v) :=
(Fk − u ∨ v) (u ∧ v − Fk−1 ) −1
−1
(v | Ωt ) ,
1 Ft,θ0 (u | Ωt ) = Ft,θ
0
Fk − Fk−1
−1
−1
where k = k (u) = Ft,θ
(u | Ωt ), with Ft,θ
(u | Ωt ) := min{y : Ft,θ 0 (y | Ωt ) ≥ u} being
0
0
the conditional quantile function and Fk := Ft,θ0 (k | Ωt ).
Lemma 1. Under H0 , It,θ0 (u) − u is a martingale difference sequence with respect to Ωt ,
i.e.
E [It,θ0 (u) | Ωt ] = u,
a.s.,
with conditional covariance
E [It,θ0 (u) It,θ0 (v) | Ωt ] = u ∧ v − uv − γ t,θ0 (u, v) ,
a.s.
Note that It,θ0 (u) are not necessarily independent across t despite the fact that by
the martingale difference property, It,θ0 (u) and It−j,θ0 (v) are serially uncorrelated for
all j 6= 0 and all u, v ∈ [0, 1] , see the Appendix. On the other hand, the It,θ0 (u) are
(conditionally) heteroskedastic, therefore the variance of S1T is model and parameter
dependent, but its distribution can be simulated conditional on exogenous information
in Ωt .
Let V1T (u, v) := Cov [S1T (u) , S1T (v)], then since 0 ≤ γ t,θ0 (u, v) < 1 a.s.,
V1T
T
1X
(u, v) = u ∧ v − uv −
E γ t,θ0 (u, v) ≤ u ∧ v − uv,
T t=1
9
i.e. the covariance and variance of S1T are not larger than those of the randomized
transformation-based process R1T or its weak limit, the Brownian sheet, see Corollary 4
in Kheifets and Velasco (2013).
i
h
e
Due to Lemma 1, E Fθ0 (u) = u under H0 and the natural empirical process for
performing tests on H0 is then S1T . This process, being based on a nonrandomized
transform, does not involve the extra noise that appears in the randomized transform
based empirical process R1T for testing Utr ∼ U[0, 1], proposed by Kheifets and Velasco
(2013), or in its modification R1T,M based on the M-randomized transform. The next
lemma is the key to understand the improvement of the M-randomized over the randomized and of the nonrandomized, advocated in this paper, over the M-randomized
transform approaches.
Lemma 2. Suppose that the uniform law of large numbers holds for Fbθr0 ,M (u) and Feθ0 (u).
Independently of whether H0 holds or not, Fbθr0 ,M (u) and Feθ0 (u) consistently and uniformly in u estimate the relative distribution, i.e. the cdf of Utr (θ0 ). Feθ0 (u) is more
efficient, but the difference in efficiency goes to 0 as M → ∞. In particular, under H0 ,
1
1
E [R1T,M (u) R1T,M (v)] =
E [S1T (u) S1T (v)] .
E [R1T (u) R1T (v)] + 1 −
M
M
From Lemma 2, it follows that S1T has the smallest variance, the variance of R1T,M is a
weighted sum of those of S1T and R1T , see also Equation (5) in Machado and Santos Silva
(2005). Other advantages of S1T over R1T,M , are 1) computational, as there is no need
to simulate M paths of transformations and 2) theoretical, since the weak convergence
is easier to prove for processes which are piece-wise linear in parameters. Therefore we
concentrate on studying the properties of tests based on the nonrandomized transform,
for which we introduce the following assumption.
Assumption 1. Ft,θ0 (· | Ωt ) (k) ∈ M a.s. for all t. Moreover, there exists a finite funcP
tion γ ∞ (u, v), such that uniformly in (u, v) ∈ [0, 1]2 , T −1 Tt=1 γ t,θ0 (u, v) →p γ ∞ (u, v).
This assumption implicitly restrict dynamics such that a uniform law of large numbers
(LLN) holds for the averaged conditional covariance function. In the case of stationary
10
and ergodic data, γ ∞ (u, v) = E γ 1,θ0 (u, v) . Sufficient conditions for the stationarity
and ergodicity of dynamic multinomial ordered choice models are given in Basu and de
Jong (2007) and for autoregressive Poisson are given in Davis et al. (2003), Fokianos
et al. (2009) and Doukhan et al. (2012). Then it is possible to show the uniformity of
the convergence from a point-wise result, since the summands are continuous, piece-wise
polynomials in u and v. As an illustration, in Section 8.5 in Appendix we discuss the
assumptions for the Poisson model.
The next result describes the asymptotic distribution of S1T under the null hypothesis.
Let ⇒ denote weak convergence in ℓ∞ [0, 1], see e.g. van der Vaart and Wellner (1996).
In fact, our empirical processes are continuous, which simplifies tightness verification.
Let V1∞ (u, v) := u ∧ v − uv − γ ∞ (u, v).
Lemma 3. Suppose Assumption 1 holds. Under H0 ,
S1T ⇒ S1∞ ,
where S1∞ is a Gaussian process in [0, 1] with zero mean and covariance function V1∞ .
The asymptotic distribution of S1T is model and parameter dependent, and the practical implementation of tests when θ0 is unknown is discussed in Section 3.2 after presenting
a general class of local alternatives to the null of correct specification of the conditional
distribution.
3.1
Local Alternatives
We next discuss the asymptotic properties of the empirical process S1T under a class of
alternative hypothesis, that will lead to consistency of the specification tests based on
S1T for a wide class of alternatives. We consider the following class of local alternatives
to H0 ,
H1T : Yt | Ωt ∼ GT,t,θ0 (· | Ωt ) for some θ0 ∈ Θ,
where
δ
δ
GT,t,θ0 (y | Ωt ) = 1 − 1/2 Ft,θ0 (y | Ωt ) + 1/2 Ht (y | Ωt ),
T
T
11
for some 0 < δ < T 1/2 and for all t, Ht (· | Ωt ) ∈ M. When δ = 0 then H1T nests H0 .
Following Kheifets and Velasco (2013), for any discrete distributions G and F in M,
with probability functions g and f , define
d (G, F, u) = G F −1 (u) − F F −1 (u)
−
F (F −1 (u)) − u
−1
−1
g
F
(u)
−
f
F
(u)
.
f (F −1 (u))
Note, that d (G, F, u) = EG [IF (Y, u)]−EF [IF (Y, u)] = EG [IF (Y, u)]−u and d (G, F, u) ≡ 0
if and only if G ≡ F . Under any Gt (· | Ωt ) ∈ M,
1
T 1/2
E [S1T
T
1X
(u)] =
E [d (Gt (· | Ωt ) , Ft,θ0 (· | Ωt ) , u)] .
T t=1
The next assumption guarantees that a LLN can be applied to the empirical discrepancy
between Ht and Ft,θ0 .
Assumption 2. Under H1T , there exists a finite function D1 (u), such that uniformly in
P
u ∈ [0, 1] , T1 Tt=1 d (Ht (· | Ωt ) , Ft,θ0 (· | Ωt ) , u) →p D1 (u).
Then the following lemma shows that the departure of H0 in the direction of H1T
introduces a drift in the asymptotic distribution of S1T that will render consistency of
hypothesis tests based on functionals of H1T .
Lemma 4. Suppose Assumptions 1-2 hold. Under H1T ,
S1T ⇒ S1∞ + δD1 ,
where S1∞ is as in Lemma 3.
3.2
Parameter Estimation Effect
In practice, tests based on S1T are unfeasible since θ0 is unknown, and has to be estimated
by b
θT , say. We assume that we have available an estimate b
θT so that under H1T
T 1/2 b
θT − θ0 = Op (1) ,
12
and define the process with estimated parameters
Sb1T (u) :=
T
1 Xn
T 1/2
t=1
o
It,bθT (u) − u .
We next analyze the consequences of replacing θ0 by b
θT in Sb1T .
p
Let k · k be Euclidean norm, i.e. for matrix A, kAk = tr (AA′ ), where A′ is a
transpose of A. For ε > 0, B(a, ε) is an open ball in Rm with the center at point a and
radius ε. For a cdf Fθ in M define
∇ (Fθ , u) := Ḟθ
−1
F
F
(u)
−u ˙
θ
θ
fθ Fθ−1 (u) ,
Fθ−1 (u) −
−1
fθ Fθ (u)
where Ḟθ := (∂/∂θ) Fθ and f˙θ := (∂/∂θ) fθ . We need the following assumptions to analyze
the asymptotic properties of Sb1T .
Assumption 3 (Parametric family). (A) The parameter space Θ is a compact set in
a finite-dimensional Euclidean space, θ ∈ Θ ⊂ Rm .
(B) There exists δ > 0, such that Ft,θ (· | Ωt ) ∈ M, for all t, Ωt , T and θ ∈ B(θ0 , δ).
(C) Ft,θ (k | Ωt ) is differentiable with respect to θ ∈ B(θ0 , δ) and under H1T
i
h
maxt E maxk supθ∈B(θ 0 ,δ) Ḟt,θ (k | Ωt ) ≤ MF < ∞.
(D) Under H1T , there exists a finite L1 (u) := plimT →∞ T −1
PT
t=1
∇ (Ft,θ0 (· | Ωt ) , u).
Conditions (A)-(C) about the parametric family of distribution are standard, see e.g.
Bai (2003, Assumptions A1-A2). For dynamic ordered choice and Poisson models the
differentiability of the conditional distribution with respect to the parameter is equivalent
to the differentiability of the link function. Part (D) guarantees a nice limit behaviour
of the average generalized derivative of It,θ . Conditions for no effect of information
truncation can be provided in a similar way to Bai (2003, Assumption A4).
The following lemma provides an expansion of the empirical process with estimated
parameters as the sum of the process with known parameters and a random drift describing parameter estimation.
13
Lemma 5. Suppose Assumptions 1-3 hold and T
1/2
b
θT − θ0 = Op (1). Under H1T ,
T
′ 1 X
Sb1T (u) = S1T (u) + T 1/2 b
θT − θ0
∇ (Ft,θ0 (· | Ωt ) , u) + op (1) ,
T t=1
(4)
uniformly in u.
Then, continuous functionals of Sb1T no longer converge to those of S1 + δD1 under
H1T , but the estimation effect also has to be taken into account using the following
assumption. Let Z (Ψ) be a normal vector with zero mean and covariance matrix Ψ.
Assumption 4 (Parameter estimation). Under H1T , the estimator b
θT admits the asymptotic linear expansion
T
1/2
T
1 X
b
ℓt (Yt , Ωt ) + op (1) ,
θT − θ0 = δξ 0 + 1/2
T
t=1
(5)
where ξ 0 is a m×1 vector and the summands ℓt constitute a martingale difference sequence
with respect to Ωt , such that
(A) E [ℓt (Yt , Ωt ) | Ωt ] = 0 and T −1
(B) Lindeberg condition T −1
holds.
p
′
ℓ
(Y
,
Ω
)
ℓ
(Y
,
Ω
)
|
Ω
→ Ψ.
E
t
t
t
t
t
t
t
t=1
PT
p
2 −1/2
kℓ
(Y
,
Ω
)k
T
kℓ
(Y
,
Ω
)k
>
ε
|
Ω
→0
1
E
t
t
t
t
t
t
t
t=1
PT
(C) There exists a finite function W1 (u), such that T −1
W1 (u) uniformly in u.
PT
t=1
E [It,θ0 (u) ℓt (Yt , Ωt ) | Ωt ] →p
In particular, under H0 , δξ 0 = 0, the estimate b
θT is centered and T 1/2 b
θT − θ0
converges in distribution to Z (Ψ).
Assumption 4(A) and 4(B) hold for the MLE of many popular discrete models, including dynamic probit and logit and general discrete choice models. As an example consider
14
estimates b
θT , which are asymptotically equivalent to the (conditional) maximum likelihood estimates, i.e.,
T
1/2
T
B0−1 X
b
θT − θ0 = − 1/2
st (Yt , Ωt ) + op (1) ,
T
t=1
where st (k, Ωt ) := f˙t,θ0 (k | Ωt ) /ft,θ0 (k | Ωt ) is the score function and B0 is a symmetric
m × m positive definite matrix given by the limit of the Hessian,
T
K
1 XX
st (k, Ωt ) f˙t,θ0 (k | Ωt )′ .
T
T →∞
t=1 k=1
B0 := plim
P
Under H1T , E [st (Yt , Ωt ) | Ωt ] = δT −1/2 K
k=1 st (k, Ωt ) ht (k | Ωt ). Then equation (5)
P
P
holds with ξ 0 = − plimT →∞ B0−1 T −1 Tt=1 K
k=1 st (k, Ωt ) ht (k | Ωt ) and
P
ℓt (Yt , Ωt ) = −B0−1 st (Yt , Ωt ) + B0−1 K
k=1 st (k, Ωt ) ht (k | Ωt ).
We can derive the covariance matrix between the process S1T (u) and T 1/2 b
θT − θ0
and obtain joint convergence results, so under H1T
S1T , T 1/2 b
θT − θ0
⇒ (S1∞ + δD1 , Z (Ψ) + δξ 0 ) ,
(6)
where the covariance function between S1∞ and Z (Ψ) is W1 (u).
We can state now the result on the asymptotic distribution of the empirical process Sb1T
under local alternatives, whose drift is different with respect to the case without estimated
parameters.
Theorem 1. Suppose Assumptions 1-4 hold. Under H1T ,
Sb1T ⇒ Sb1∞ + δ {D1 + ξ ′0 L1 } ,
where Sb1∞ := S1∞ +Z (Ψ)′ L1 is a Gaussian process with zero mean and variance function
V1 (u, v) + L1 (u)′ ΨL1 (v) + W1 (u)′ L1 (v) + W1 (v)′ L1 (u).
15
4
EMPIRICAL PROCESSES FOR DYNAMIC SPECIFICATION
Test statistics based on S1T , R1T and R1T,M verify that the conditional distribution of Yt
is right on average across all possible Ωt , so these tests might not capture all sources of
misspecification. This issue is raised in Corradi and Swanson (2006), Delgado and Stute
(2008) and Kheifets (2015) in relation to testing continuous distributions. However, it is
not possible to develop specification tests conditioned on infinite dimensional values of Ωt .
Instead of truncating Ωt or restricting the class of models, we consider S2T , a biparameter
analog of S1T to control the possible dynamic misspecification. From Lemma 1, since
under H0 , It,θ0 (u1 ) − u1 is a MDS, It,θ0 (u1 ) It−1,θ0 (u2 ) − u1 u2 is centered around zero,
and moreover
E [It,θ0 (u1 ) It−1,θ0 (u2 ) | Ωt−1 ] = u1 u2 ,
a.s.
This motivates us to develop tests based on S2T defined in (2). This process also has zero
mean under the null and identifies not only departures from the null derived from deviations of the unconditional expectation of It,θ0 (u) from u, but also from a possible failure
of the martingale property, so that It,θ0 (u1 ) and It−1,θ0 (u2 ) would become correlated.
This idea is similar to that exploited in Kheifets’ (2015) in the context of conditional
distribution testing for continuous distributions, where different methods of checking the
independence property of the PIT are proposed. Alternative statistics exploiting the lack
of correlations with any other lag could be proposed, but we expect that low lags are
typically more useful for detecting general forms of misspecification.
One could also consider a biparameter analog of R1T,M , i.e. for some M = 1, 2, . . . ,
R2T,M (u) :=
T X
M
X
r
r
1
1
U
(θ
)
≤
u
1
U
(θ
)
≤
u
−
u
u
,
0
1
0
2
1
2
t,m
t−1,m
(T − 1)1/2 M t=2 m=1
where u = (u1 , u2 ) ∈ [0, 1]2 . In particular, a bivariate analog of R1T , R2T (u) := R2T,1 (u),
is introduced in Kheifets and Velasco (2013). Tests based on R2T and R2T,M involve
randomized transforms and therefore suffer from power loss compared to tests based on
the nonrandomized transform.
16
Note, that S2T (u) − u1 S1T −1 (u2 ) is a martingale. This observation will allow us to
derive weak convergence of S2T by employing limiting theorems for MDS. Properties of
R2T were established in Kheifets and Velasco (2013) and could be extended to R2T,M .
Here we discuss the properties of S2T when we estimate θ0 .
In practice we use the process
Sb2T (u) :=
T n
o
X
1
(u
)
−
u
u
(u
)
I
I
b
2
1 2 ,
1
t,b
θT
(T − 1)1/2 t=2 t,θT
where we can write under H1T
T
′ 1 X
∇2,t (u) + op (1) ,
Sb2T (u) = S2T (u) + T 1/2 b
θT − θ0
T t=2
(7)
uniformly in u, where
∇2,t (u) := It−1,θ0 (u2 ) ∇ (Ft,θ0 (· | Ωt ) , u1 )+u1 ∇ (Ft−1,θ0 (· | Ωt−1 ) , u2 ) and the asymptotic
covariance function is W2 (u) := ACov S2T (u) , T 1/2 b
θT − θ0 . To study the asymptotic properties of the biparameter process we introduce the next assumption, which
extends Assumption 2.
Assumption 5. Under H1T , there exist finite functions D2 (u) and L2 (u), such that
uniformly in u
(A) T −1
PT
{It−1,θ 0 (u2 ) d (Ht (· | Ωt ) , Ft,θ0 (· | Ωt ) , u1 )
PT
∇2,t (u) →p L2 (u).
t=2
+ u1 d (Ht (· | Ωt ) , Ft,θ0 (· | Ωt ) , u2 )} →p D2 (u).
(B) T −1
t=2
Note that the second terms in the definitions of D2 and L2 correspond to u1 D1 (u2 ) and
u1 L1 (u2) respectively, the equivalent for the single parameter process S1T , but the first
ones are new. To state the next result, we need to assume existence of probabilistic limits
of several random functions. For the sake of presentation, we defer precise statements to
the Appendix, see Assumption A.
17
Theorem 2. Suppose that in addition to the conditions of Theorem 1, Assumption 5 and
Assumption A from the Appendix hold. Under H1T ,
S2T (u) ⇒ S2∞ + δD2 ,
where S2∞ is a Gaussian process in [0, 1] with mean zero and covariance function V2∞ (u, v)
defined in the Appendix. Under H1T , if parameters are estimated,
Sb2T ⇒ Sb2∞ + δ {D2 + ξ ′0 L2 } ,
where Sb2∞ := S2∞ +Z (Ψ)′ L2 is a Gaussian process with zero mean and variance function
V2∞ (u, v) + L2 (u)′ ΨL2 (v) + W2 (u)′ L2 (v) + W2 (v)′ L2 (u).
When Gt (· | Ωt ) is different from Ft,θ 0 (· | Ωt ) such that D2 is non-zero, the test based
on Sb2T has nontrivial power in the direction of H1T . In contrast to the univariate case
with S1T , the first term in the definition of D2 contains correlation with the past information and can therefore capture dynamic misspecification when this induces in such a
correlation, even if the unconditional expectation of d, which appears in the second term
u1 D1 (u2 ), is zero. This fact is crucial if misspecification occurs in the dynamics and not
only in the link function or other static aspects of the model.
5
BOOTSTRAP TESTS
To test H0 we consider Cramer-von Mises, Kolmogorov-Smirnov or any other continuous
functionals of SbjT , j = 1, 2, η SbjT . Then consistency properties of specification tests
based on SbjT can be derived using the discussion in the previous sections by applying the
continuous mapping theorem, so we omit the proof of the following result.
Theorem 3. Suppose that conditions of Theorem 2 hold. Under H1T ,
b
b
η SjT →d η Sj∞ , j = 1, 2.
Since the asymptotic distributions of SjT (u) are model dependent, and those of SbjT (u)
further depend on the estimation effect, we need to resort to bootstrap methods to
18
implement our tests in practice. In the literature, there are several resampling methods
suitable for dependent data, but since under H0 the parametric conditional distribution is
fully specified, we apply a conditional parametric bootstrap algorithm that only requires
to make draws from Ft,bθ (· | Ωt ) to mimic the null distribution of the test statistics. For a
discussion of the parametric bootstrap see Stute et al. (1993) and Andrews (1997), which
can be adapted to the complications with information truncation and initialization arising
in the dynamic case using the discussion in Bai (2003).
To estimate the true 1 − α quantiles cj (θ0 ) of the null asymptotic distribution of the
test statistics, given by some continuous functional η applied to Sbj∞ with δ = 0, we
implement the following steps.
1. Estimate the model with data (Yt , Xt′ ), t = 1, 2, ..., T , get parameter estimator b
θT
and compute test statistics η(SbjT ).
2. Simulate Yt∗ with FθcT (· | Ω∗t ) recursively for t = 1, 2, ..., T , where the bootstrap
∗
∗
, Xt−2 , ... .
, Xt−1 , Yt−2
information set is Ω∗t = Xt , Yt−1
∗
3. Estimate the model with simulated data Yt∗ , get b
θT using the same method as for
∗
b
θT , get bootstrapped test statistics η SbjT
.
4. Repeat 2-3 B times, compute the percentiles of the empirical distribution of the B
bootstrapped test statistics.
5. Reject H0 if η SbjT is greater than the (1 − α)th percentile of the empirical dis
∗
tribution of the B bootstrapped test statistics denoted by b
cjB b
θT .
To analyze the properties of our parametric bootstrap, we need to assume that the
same conditions on the estimation method hold for both for original and resampled data.
More formally, we have
Assumption 6. (A) The conditional distribution of Yt conditional on Ωt coincides with
the conditional distribution of Yt conditional on Ωt ∪ {Xk′ }Tk=t+1 .
19
(B) Suppose that the sample is generated by FθT , for some nonrandom sequence θT
converging to θ0 , i.e. we have a triangular array of random variables {YT t : t =
1, 2, . . . , T } with (T, t) element generated by FθT (· | ΩT t ), where
ΩT t = {Xt , YT t−1 , Xt−1 , YT t−2 , Xt−2 , . . .}. Then the estimator b
θT of θT admits an
asymptotic linear expansion as in Assumption 4. Moreover, assume that under the
alternative H1 , there exists some θ 1 ∈ Θ so that θ1 = plimT →∞ b
θT .
This assumption insures that by simulating from the conditional distribution FθT
we obtain the correct joint distribution of SjT and T 1/2 b
θ T − θT in parallel to those
required in Theorems 1-2. Assumption 6 (A) says that Yt and future Xt are independent
conditionally on past information, i.e. that there is no direct feedback effect. For example,
in a latent variable form of the ordered probit model, this assumption translates to strict
exogeneity, i.e. that innovations are independent of future Xt . Dependence between Yt
and future Xt is still allowed through serial dependence in Xt and Yt . Assumption 6 (B)
is similar to Condition (5.5) in Burke et al. (1979), Assumption (A1) in Stute et al. (1993)
and Assumption E2 in Andrews (1997), and introduces a triangular array version of the
expansion and central limit theorem for parameter estimates, see also the discussion in
Section 4.1 in Andrews (1997).
We obtain the following result.
Theorem 4. Suppose that in addition to conditions of Theorem 2, Assumption 6 holds.
Under H1T , as B, T → ∞,
in probability, so
b
c∗jB
∗
η SbjT
b
→d η Sj∞ ,
j = 1, 2,
∗
b
b
θ T →p cj (θ0 ), and therefore, under H0 , Pr η SjT > b
cjB b
θT
→
α. Suppose also that the conditions of Theorem 2 hold for any θ0 ∈ Θ. Under H1 , as
B, T → ∞, b
c∗jB b
θT = Op (1).
This theorem shows that the bootstrap test statistic has the same limit distribution as
the original one under local alternatives, so that under the null we get the right asymptotic
size using bootstrap estimated critical values and that under local alternatives we get non
20
trivial power when the drifts of the stochastic processes Sb1T and Sb2T are non negligible.
Similarly, under fixed alternatives we are able to get a bootstrap consistent test when the
asymptotic test is consistent itself, i.e. limT →∞ Pr η SbjT > b
c∗jB b
θT
= 1 if η SbjT
diverges asymptotically.
6
APPLICATION AND SIMULATIONS
In this section we use a Monte Carlo simulation exercise to investigate the finite sample
properties of the tests proposed in this paper. We take as reference the dynamic ordered
discrete choice models investigated in Basu and de Jong (2007) for the modeling of the
monetary policy conducted by the Federal Reserve (FED). The dependent variable uses
the following codification of the changes in the reference interest rate in US, the federal
funds rate it ,
1
2
Yt =
3
4
if
∆it < −0.25
if −0.25 ≤ ∆it < 0
if
0 ≤ ∆it < 0.25
if
∆it ≥ 0.25.
Data is monthly and spans January 1990 to December 2006, leading to T = 204
complete observations. The explanatory variables that Basu and de Jong (2007) used to
explain the decisions of the FED on ∆it are the current value and 4 lags of inflation (inf),
the current value and a lag of four different measures of output gap (out) and a series of
dummies that describe the decision of the FED in the previous period, dum1t = I(∆it−1 <
0), dum2t = I(∆it−1 > 0), dum3t = I(∆it−1 < −0.25), dum4t = I(∆it−1 > 0.25).
Instead of these four dummies, we implement an AR(1), ’dynamic’ version with one lag
of the discrete Yt as explanatory variable (and a version without lags that we refer to as
’static’ to serve as a benchmark to the inclusion of lagged endogenous variables in Ωt ).
We consider both the Logit and Probit versions of the models. We fit four versions of
the basic model based on different definitions of the output gap and conditional on the
series of inflation and output gap and on the parameter estimates obtained, we simulate
21
Table 1: Scenarios for Monte Carlo simulations.
Scenario
Null and Alternative
Size 1
H0 : static probit
Size 2
H0 : static logit
Power 1
H0 : static probit vs H1 : static logit
Power 2
H0 : static probit vs H1 : dynamic probit
Power 3
H0 : static probit vs H1 : dynamic logit
series Yt and conduct our tests on these (see Monte Carlo scenarios in Table 1).
The four choices of output gap lead to Models I-IV. The output gap is the percentage
deviation of the actual from the potential output, which is interpolated to obtain a
series of monthly frequency by replicating the GDP observation for any quarter to all
the months in that quarter. Then two different measures of potential output are used:
the potential output series provided by the Congressional Budget Office and a potential
output series constructed in a real-time setting using the HP filter, leading to Models I
and II. Apart from output gap, other measures of economic activity are used, such as
unemployment rate and capacity utilization, leading to Models III and IV. Data sources
are described in Basu and de Jong (2007).
We compare the performance of our tests with an alternative test which is also omnibus and does not require smoothing (and choice of smoothing parameters). Two general
approaches can be adapted to our setup: the test of the Generalized Linear Model (GLM)
of Stute and Zhu (2002) and the Conditional Kolmogorov test of Andrews (1997), as
discussed in Mora and Moro-Egido (2007). The first one is a test based on a marked em
h
i
e , where me (·)
e = x = me x′ β
pirical process for testing the null H0′ : E Y | X
02
β 01
β 01
e ,β
e are finite dimensional parameters. In the cases
is a parametric link function and β
01
02
where Y takes only two values {0, 1}, the conditional mean coincides with the conditional
probability of Y = 1 and the null is similar to our H0 if we were considering an i.i.d setup.
22
′
e
e
e
To test Yt | Xt ∼ Pβe01 · | Xt β 20 define the process
ZT (y) :=
T
T
i
oh
1 X n e′e
′
e
e
1 Xt β 20 ≤ y Yt − Pβe01 Yt = 1 | Xt β 20 ,
1/2
t=1
y ∈ R.
n
n
o
o
e ≤ y with 1 X
e′β
et ≤ x
e
The second test by Andrews is obtained by substituting 1 X
t
et ) in ZT , but since it always underperforms
(where x
e is a real vector of dimension of X
according to simulations of Mora and Moro-Egido (2007), it is not considered here. If Y
takes values {1, . . . , K}, Mora and Moro-Egido (2007) substitute testing H0 by K tests of
e , with corresponding processes Zj,T , where
et′ β
e t ∼ P e Yt | X
the hypotheses Yjt | X
20
j,β 01
Yjt = 1{Yt = j} and j = 1, 2, . . . , K, then the resulting pooled test statistics are
η CvM
= T −1
Z
K X
T
X
j=1 t=1
and
−1
η KS
max
Z = T
j=1,...,K
′ 2
e
eβ
Zj,T X
t 20
T
X
t=1
′ 2
e
eβ
Zj,T X
,
t 20
which we call the CvM and KS tests respectively. To apply these tests to our model, let
′
e = β ′ , ρ, −τ 1 and take the corresponding link functions.
et = X ′ , Yt−1 , 1 ′ and β
X
t
We analyze tests based on S1T , R1T,M , R1T and S2T , R2T,M , R2T and ZT . In all cases
we use Kolmogorov-Smirnov (KS) and Cramer-von Mises (CvM) measures. We only
b1T,M , etc, where we replace
consider feasible bootstrap versions of tests based on Sb1T , R
θ0 by root-T consistent estimates b
θT , the ML estimator in our case. We are not aware
of any theoretical results for bootstrap assisted tests based on ZbT in our setup, although
Mora and Moro-Egido (2007) provide some simulations.
Parameter estimates for real data are reported in Tables 2 and 3. The main question
is whether the static Probit or Logit models are appropriate for changes in the interest
rates, and we check this with our tests. The p-values in Tables 4 and 5 say that all these
models are rejected even at the 1% significance level by biparameter nonrandomized
b1T , Sb1T ) cannot
transform based tests. Note that single parameter static tests (e.g. R
reject any proposed model with the sole exception of Sb1T which rejects at 5% Model II
with Cramer – von Misses test statistics.
23
Table 2: ML estimates and standard errors of Models I-IV with static and dynamic
specifications and Probit link function applied to the real US data, T = 204.
I-static I-dynamic II-static II-dynamic III-static III-dynamic IV-static IV-dynamic
τ 1 −4.81 −2.07 −3.31
−1.05
−3.15
−1.17
−3.41
−1.48
(0.51) (0.66)
(0.35)
(0.47)
(0.36)
(0.48)
(0.37)
(0.50)
τ 2 −4.05 −1.14 −2.64
−0.19
−2.34
−0.20
−2.57
−0.50
(0.47) (0.64)
(0.31)
(0.46)
(0.32)
(0.47)
(0.32)
(0.48)
τ 3 −1.72
1.66
−0.39
2.60
0.09
2.62
−0.11
2.29
(0.40) (0.63)
(0.26)
(0.48)
(0.28)
(0.48)
(0.27)
(0.49)
inf −1.39 −1.36 −1.51
−1.60
−1.83
−1.82
−1.70
−1.70
(0.68) (0.72)
(0.67)
(0.71)
(0.69)
(0.73)
(0.69)
(0.73)
inf−1 1.86
2.90
1.94
3.05
2.05
3.07
2.14
3.01
(0.99) (1.06)
(0.98)
(1.06)
(1.00)
(1.07)
(1.01)
(1.07)
inf−2 −1.30 −2.81 −1.27
−2.80
−1.60
−2.92
−2.12
−3.11
(0.98) (1.07)
(0.97)
(1.06)
(0.99)
(1.07)
(1.02)
(1.09)
inf−3 1.39
2.44
1.60
2.74
1.79
2.79
1.27
2.33
(0.99) (1.06)
(0.98)
(1.06)
(1.00)
(1.08)
(1.03)
(1.09)
inf−4 0.43
−0.53 −0.23
−1.05
−0.00
−0.85
0.88
−0.20
(0.68) (0.73)
(0.66)
(0.71)
(0.67)
(0.73)
(0.71)
(0.76)
out −1.02 −1.02
0.36
0.40
3.35
2.54
−0.98
−0.62
(0.30) (0.33)
(0.59)
(0.63)
(0.68)
(0.74)
(0.22)
(0.23)
out−1 0.81
0.90
0.84
0.65
2.48
0.95
−1.03
−0.65
(0.29) (0.32)
(0.59)
(0.64)
(0.67)
(0.73)
(0.22)
(0.23)
Y−1
—
−1.08
—
−1.12
—
−1.03
—
−0.94
(0.15)
(0.15)
(0.16)
(0.16)
24
Table 3: ML estimates and standard errors of Models I-IV with static and dynamic
specifications and Logit link function applied to the real US data, T = 204.
I-static I-dynamic II-static II-dynamic III-static III-dynamic IV-static IV-dynamic
τ 1 −8.46 −3.77 −6.01
−2.12
−5.61
−2.15
−6.15
−2.82
(0.98) (1.20)
(0.68)
(0.83)
(0.69)
(0.85)
(0.72)
(0.89)
τ 2 −7.03 −1.96 −4.71
−0.46
−4.12
−0.31
−4.56
−0.90
(0.90) (1.17)
(0.60)
(0.81)
(0.59)
(0.83)
(0.61)
(0.86)
τ 3 −3.00
3.02
−0.85
4.52
0.07
4.60
−0.24
4.04
(0.72) (1.12)
(0.47)
(0.84)
(0.49)
(0.86)
(0.49)
(0.87)
inf −2.44 −2.29 −2.53
−2.89
−3.17
−3.28
−2.81
−3.06
(1.21) (1.30)
(1.21)
(1.29)
(1.21)
(1.32)
(1.22)
(1.32)
inf−1 3.28
4.95
3.22
5.46
3.59
5.43
3.41
5.31
(1.78) (1.92)
(1.77)
(1.92)
(1.76)
(1.93)
(1.82)
(1.95)
inf−2 −2.48 −5.02 −2.17
−5.22
−2.97
−5.21
−3.52
−5.40
(1.74) (1.95)
(1.73)
(1.94)
(1.76)
(1.95)
(1.86)
(1.99)
inf−3 2.42
4.36
2.61
5.20
2.94
5.11
1.65
4.02
(1.75) (1.92)
(1.75)
(1.93)
(1.77)
(1.95)
(1.86)
(1.99)
inf−4 0.93
−0.87 −0.17
−1.88
0.32
−1.54
2.11
−0.28
(1.20) (1.32)
(1.18)
(1.28)
(1.19)
(1.30)
(1.27)
(1.36)
out −1.78 −1.79
0.43
0.63
5.87
4.12
−1.83
−1.15
(0.54) (0.60)
(1.04)
(1.14)
(1.24)
(1.34)
(0.40)
(0.42)
out−1 1.43
1.59
1.61
1.29
4.21
1.50
−1.88
−1.14
(0.52) (0.59)
(1.04)
(1.15)
(1.20)
(1.33)
(0.40)
(0.42)
Y−1
—
−1.98
—
−2.04
—
−1.86
—
−1.71
(0.28)
(0.27)
(0.28)
(0.28)
25
Table 4: P-values of Cramer – von Misses tests for static Probit and Logit link function
applied to the real US data, T = 204.
b2T,25
R
b2T
R
Sb1T
b1T,50
R
b1T,25
R
b1T
R
bT
Z
0.001
0.001
0.237
0.009
0.026
0.078
0.516
0.244
0.001
0.001
0.001
0.166
0.077
0.057
0.229
0.167
0.022
Model III
0.001
0.001
0.001
0.307
0.492
0.632
0.616
0.731
0.109
Model IV
0.001
0.002
0.002
0.496
0.721
0.509
0.582
0.668
0.268
Sb2T
b2T,50
R
Model I
0.001
Model II
H0 : static probit
H0 : static logit
Model I
0.001
0.001
0.001
0.152
0.021
0.079
0.221
0.793
0.199
Model II
0.001
0.001
0.001
0.112
0.113
0.155
0.459
0.240
0.032
Model III
0.001
0.001
0.001
0.360
0.314
0.493
0.541
0.745
0.171
Model IV
0.001
0.001
0.001
0.448
0.890
0.804
0.899
0.634
0.272
Table 5: P-values of Kolmogorov – Smirnov tests for static Probit and Logit link function
applied to the real US data, T = 204.
b2T,25
R
b2T
R
Sb1T
b1T,50
R
b1T,25
R
b1T
R
bT
Z
0.002
0.002
0.082
0.047
0.193
0.372
0.354
0.392
0.001
0.001
0.002
0.586
0.351
0.426
0.626
0.450
0.107
Model III
0.001
0.001
0.001
0.155
0.454
0.435
0.244
0.742
0.124
Model IV
0.001
0.002
0.002
0.799
0.936
0.913
0.801
0.355
0.230
Sb2T
b2T,50
R
Model I
0.003
Model II
H0 : static probit
H0 : static logit
Model I
0.001
0.001
0.001
0.133
0.010
0.050
0.212
0.684
0.220
Model II
0.001
0.001
0.001
0.354
0.114
0.201
0.319
0.416
0.058
Model III
0.001
0.001
0.001
0.149
0.511
0.472
0.350
0.642
0.173
Model IV
0.002
0.002
0.001
0.769
0.975
0.968
0.867
0.411
0.207
26
Table 6: Simulated size/power rates for the nominal 5% level of Cramer – von Misses
tests of Models I-IV with static and dynamic specifications applied to simulated data,
T = 100.
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
b1T,25
b1T,50 R
b2T Sb1T R
b2T,25 R
b2T,50 R
R
Size 1 H0 : static probit
5.5
6.0
5.5
4.5
6.6
6.3
5.7
5.3
6.7
5.0
5.5
6.3
5.2
4.2
7.7
7.0
6.5
5.4
6.0
3.7
3.3
5.2
6.7
5.6
3.9
5.1
4.6
4.9
Size 2 H0 : static logit
6.5
6.5
4.9
4.1
7.2
5.6
6.0
5.6
6.7
7.6
4.0
4.6
5.3
4.6
7.3
9.0
6.4
3.3
6.4
7.8
5.2
6.6
6.3
5.0
4.5
6.5
4.6
4.7
Power 1 H0 : static probit vs H1 : static logit
8.5
7.7
6.6
4.9
8.4
6.5
6.0
5.1
5.0
4.4
4.0
6.4
6.9
5.3
9.1
9.4
7.9
4.7
9.0
8.3
7.7
6.3
6.2
5.3
4.5 10.2
8.6
7.5
Power 2 H0 : static probit vs H1 : dynamic probit
89.2
85.2
82.7
25.7 13.2
12.0
11.7
92.8
92.3
91.1
34.2 10.5
8.1
8.8
90.7
88.4
86.1
22.5 9.2
9.8
8.5
88.1
84.1
83.0
27.7 10.3
7.8
7.4
Power 3 H0 : static probit vs H1 : dynamic logit
90.1
89.3
86.0
22.9 12.1
10.0
8.5
94.2
93.0
90.6
29.8 9.6
9.1
7.1
93.5
91.9
90.9
30.3 10.0
8.0
7.8
91.1
88.4
85.9
26.0 11.1
12.3
11.4
Sb2T
b1T
R
bT
Z
4.4
4.8
3.3
4.7
7.2
5.6
8.5
9.1
3.6
4.0
4.6
3.8
7.1
8.7
8.2
8.3
4.6
3.0
5.0
4.4
18.4
17.2
9.4
12.5
5.0
3.9
4.4
4.7
12.6
14.6
10.9
14.7
5.4
3.3
4.5
2.8
7.8
6.5
6.4
6.4
Table 7: Simulated size/power rates for the nominal 5% level of Kolmogorov – Smirnov
tests of Models I-IV with static and dynamic specifications applied to simulated data,
T = 100.
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
b1T,25
b1T,50 R
b2T Sb1T R
b2T,25 R
b2T,50 R
R
Size 1 H0 : static probit
5.1
6.4
5.2
3.9
7.8
6.3
6.8
5.5
6.5
3.9
4.9
5.9
5.1
4.1
7.7
7.8
6.8
5.1
6.1
7.0
6.0
6.5
5.4
5.3
3.4
5.3
5.3
4.8
Size 2 H0 : static logit
7.0
6.4
6.1
5.4
9.1
6.4
6.3
4.7
4.9
4.6
3.5
5.6
3.8
4.0
8.3
8.3
6.7
3.2
6.2
5.7
3.5
6.2
6.5
5.1
4.7
6.6
5.8
5.3
Power 1 H0 : static probit vs H1 : static logit
7.0
6.2
5.4
3.7
5.2
3.3
3.9
4.3
3.8
4.5
3.7
4.1
3.9
3.6
10.2
7.3
7.1
3.9
7.1
5.7
5.7
5.6
6.6
4.3
3.2
6.4
5.1
6.2
Power 2 H0 : static probit vs H1 : dynamic probit
82.8
79.0
74.5
13.6 10.3
9.1
7.1
87.9
85.5
83.3
17.7 12.1
11.2
9.3
85.7
83.2
79.4
13.8 7.1
6.4
7.2
81.7
78.5
74.6
13.8 7.7
7.8
6.7
Power 3 H0 : static probit vs H1 : dynamic logit
86.2
82.7
79.0
14.2 7.7
4.9
3.8
90.0
86.2
82.2
15.9 9.3
7.9
8.1
89.0
86.4
83.7
15.9 5.6
5.1
4.4
87.5
83.8
79.3
16.1 9.4
7.5
7.7
Sb2T
27
b1T
R
bT
Z
3.7
4.8
4.0
4.0
6.7
5.8
10.0
8.1
3.2
3.9
4.5
3.4
7.7
8.9
9.2
6.8
3.5
3.3
3.9
4.9
16.9
14.0
9.2
11.3
4.2
4.1
4.6
5.9
11.8
14.2
10.5
12.9
4.9
4.8
4.9
3.6
7.9
6.2
5.6
7.2
Table 8: Simulated size/power rates for the nominal 5% level of Cramer – von Misses
tests of Models I-IV with static and dynamic specifications applied to simulated data,
T = 200.
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
b1T,25
b1T,50 R
b2T Sb1T R
b2T,25 R
b2T,50 R
R
Size 1 H0 : static probit
4.0
5.4
5.7
6.2
4.2
4.6
4.9
4.5
4.4
3.5
2.4
6.3
4.7
5.9
4.6
4.4
3.4
4.2
5.4
5.5
5.2
5.3
6.1
6.3
4.4
4.8
4.6
7.0
Size 2 H0 : static logit
7.2
8.2
6.7
5.8
5.8
6.7
6.4
5.4
6.4
6.1
4.7
4.8
5.6
5.3
5.3
5.2
3.9
4.0
5.6
5.8
6.7
5.4
6.8
5.0
4.0
5.6
5.2
5.0
Power 1 H0 : static probit vs H1 : static logit
7.2
8.2
6.6
6.9 10.9
10.3
10.9
4.5
4.9
5.6
6.3
7.5
6.4
7.3
6.0
5.2
6.1
5.9
6.9
6.8
7.9
6.5
6.6
6.6
4.3
7.1
6.1
7.5
Power 2 H0 : static probit vs H1 : dynamic probit
98.5
97.3
95.2
33.2 13.6
11.6
9.8
99.5
99.3
98.5
41.5 16.0
14.8
12.6
98.5
97.0
95.9
30.7 13.0
11.8
9.8
95.8
93.6
91.6
22.9 10.2
9.5
7.6
Power 3 H0 : static probit vs H1 : dynamic logit
98.6
97.5
95.6
34.5 15.2
14.0
14.0
99.5
98.9
98.6
39.1 16.9
16.1
13.7
98.8
98.1
96.4
31.2 14.8
13.7
11.6
95.8
94.2
91.8
23.9 11.4
10.1
8.6
Sb2T
b1T
R
bT
Z
3.8
5.2
4.4
4.1
5.2
6.1
6.9
8.3
6.9
6.7
6.6
5.8
9.2
6.5
7.0
5.4
7.5
7.1
7.9
5.3
16.2
18.2
13.9
13.7
5.4
7.4
6.6
5.3
16.7
20.8
17.9
11.4
5.8
4.4
3.3
4.9
5.2
7.0
5.4
6.9
Table 9: Simulated size/power rates for the nominal 5% level of Kolmogorov – Smirnov
tests of Models I-IV with static and dynamic specifications applied to simulated data,
T = 200.
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
Model I
Model II
Model III
Model IV
b1T,25
b1T,50 R
b2T Sb1T R
b2T,25 R
b2T,50 R
R
Size 1 H0 : static probit
5.1
5.0
6.0
4.8
4.5
5.9
3.9
3.7
3.9
3.9
2.9
6.3
5.6
6.4
4.5
5.2
4.3
3.9
4.2
4.5
4.9
5.0
6.4
7.0
4.6
4.6
4.8
6.4
Size 2 H0 : static logit
5.7
5.7
6.3
4.9
6.3
5.9
6.3
5.5
5.1
5.9
3.4
5.4
4.6
5.9
3.6
5.4
4.6
3.6
6.4
4.3
5.2
6.4
7.3
5.6
4.7
6.6
6.4
4.7
Power 1 H0 : static probit vs H1 : static logit
6.3
6.5
4.2
6.6
7.2
5.9
5.0
4.6
5.0
6.4
6.3
5.2
4.9
5.7
5.0
6.2
5.7
5.2
3.7
4.1
5.0
5.7
7.0
5.6
4.5
5.4
3.4
4.6
Power 2 H0 : static probit vs H1 : dynamic probit
94.3
92.0
86.5
22.8 11.4
10.6
9.7
98.1
96.5
94.4
26.3 15.5
13.1
13.5
94.3
91.0
87.9
21.0 14.7
13.1
12.7
90.5
85.0
82.0
17.9 11.0
9.5
9.4
Power 3 H0 : static probit vs H1 : dynamic logit
97.1
93.8
91.8
24.7 12.4
12.8
11.1
98.9
97.6
96.5
29.5 16.9
17.1
14.6
96.1
93.8
91.8
26.0 14.6
14.4
11.9
93.0
89.2
86.9
14.1 13.0
12.5
8.5
Sb2T
28
b1T
R
bT
Z
3.5
4.9
5.3
4.5
4.8
5.3
7.8
8.5
6.7
6.5
6.3
6.1
8.7
6.1
7.1
5.1
5.9
7.3
7.3
6.3
14.0
13.1
13.8
11.4
5.5
7.4
8.0
5.8
13.4
16.9
15.4
10.4
5.1
4.6
4.3
6.9
5.5
5.7
5.3
6.8
To study the reliability of these results we conduct a Monte Carlo experiment using
the estimated models with the real data as data generating processes and obtain the
simulations for the discrete response conditional on the covariates time series. In Tables 6
and 7 we provide the empirical size and power results of our tests across simulations for
sample size T = 100 and static Probit and Logit and output gap choices (Models I
to IV). To speed up the simulation procedure, we use the warp bootstrap algorithm of
Giacomini, Politis and White (2013). We see that all bootstrap tests provide reasonable
size accuracy, tests based on single parameter empirical processes underrejecting slightly,
while ones based on bivariate processes tend to overreject moderately. KolmogorovSmirnov and Cramer-von Mises tests perform similarly in all cases, and the choice of the
output gap series does not make large differences either, nor does the introduction of
lagged endogenous (discrete) variables in the information set.
The power of the tests for the static Probit model is analyzed against three different
alternatives: static Logit, dynamic Probit and dynamic Logit. We see that the tests
without randomization, Sb1T and Sb2T always perform better than random continuous
b1T,M and R
b2T,M , which in turn dominate R
b1T and R
b2T , thus confirming our
processes R
theoretical findings. When we compare Probit and Logit specifications while letting the
dynamic aspect of the model be well specified, static in both cases, we observe that with
this sample size and these specifications, it is almost impossible to distinguish Probit from
Logit models. The power against a dynamic Probit and Logit alternatives is very high.
Since the nature of misspecification is dynamic, once again bivariate processes should
have more power compared to single parameter counterparts, as it is confirmed in our
simulation results. It can also be observed that for these alternatives, the Cramer-von
Mises criterium provides more power than Kolmogorov-Smirnov tests. As for alternative
tests based on ZbT , they have power comparable to Sb1T , sometimes slightly better, and
are always outperformed by any bivariate test. This is not surprising, since ZbT has more
structure, i.e. it assumes a single-index model for covariates, but averages across points,
thus suffering the same problems as other single parameter tests considered here.
In Tables 8 and 9 we provide the empirical size and power results of our tests for the
29
larger sample size T = 200. Here the size properties are similar, while power rejections
rates are noticeably higher for the dynamic alternatives.
7
CONCLUSIONS
In this paper we have proposed new specification tests for the conditional distribution of
discrete data with possibly infinite support. The new tests are functionals of empirical
processes based on a nonrandomized transform that solves the implementation problem
of the usual PIT for discrete distributions and achieves consistency against a wide class
of alternatives. We show the validity of a bootstrap algorithm for approximating the
null distribution of the test statistics, which are model and parameter dependent. In
our simulation study, we show that our method compares favorably in many relevant
situations with other methods available in the literature and have illustrated the new
method in a small application.
8
APPENDIX
8.1
Properties of the nonrandomized transform
In this section we derive the basic properties of the nonrandomized transform, which
are required prior to proving the weak convergence results for our empirical process.
Without loss of generality and in order to make the exposition more transparent, we
omit subscripts t, θ0 and conditioning set Ωt , and use shortcuts IF (Y, u) = It,θ0 (Yt , u)
and IF,M (Y, u) = It,θ0 ,M (Yt , u).
For F ∈ M, F (F −1 (u)) ≥ u > F (F −1 (u) − 1) and equality holds iff u = F (k)
for some integer k. For a random variable Y ∼ G ∈ M we find PrG (F (Y ) < u) =
G (F −1 (u) − 1) and g (F −1 (u)) := PrG (Y = F −1 (u)) = G (F −1 (u)) − G (F −1 (u) − 1).
For G = F , we have that PrF (F (Y ) < u) = F (F −1 (u) − 1) < u, i.e. F (Y ) is not
uniform and the expectation of the indicator function I (F (Y ) < u) is never u as it is for
30
continuous F .
The nonrandomized transform can be written as
where
IF (Y, u) = (1 − δ F (u)) 1 Y = F −1 (u) + 1 Y < F −1 (u) ,
F (F −1 (u)) − u
.
f (F −1 (u))
Note that δ F (u) ∈ [0, 1). We see that IF (Y, u) is a piecewise linear (continuous) function
δ F (u) :=
increasing in u. Let
δ F (u, v) := (δ F (u ∨ v) − δ F (u) δ F (v)) f F −1 (u ∧ v)
× 1 F −1 (u) = F −1 (v) ∈ 0, u ∧ v ∧ f F −1 (u ∧ v) ,
d (G, F, u, v) := d (G, F, u ∧ v)
− (δ F (u ∨ v) − δ F (u) δ F (v)) 1 F −1 (u) = F −1 (v)
× g F −1 (u) − f F −1 (u) ,
In Table 10 and Lemma A we list the properties of this transform.
Lemma A. For 0 ≤ v ≤ u ≤ 1 and F, G, H ∈ M,
R
(i) EG [IF (Y, u)] = u + d (G, F, u), where EG [·] = (·)dG and d (G, F, u) ∈ [−u, 1 − u].
When G = F , the expectation is u.
(ii) IF (Y, u) IF (Y, v) = IF (Y, u ∧ v) −
(δ F (u ∨ v) − δF (u) δ F (v)) × 1{Y = F −1 (u) = F −1 (v)} .
(iii) EG [IF (Y, u) IF (Y, v)] = u ∧ v − δ F (u, v) + d (G, F, u, v).
(Y −1)−H(Y −1)|
(iv) |IF (Y, u) − IH (Y, u)| ≤ 1 ∧ |F (Y )−H(Yf)|∨|F
(Y )∨h(Y )
Moreover, EF |IF (Y, u) − IH (Y, u)|2 ≤ 9 supk |F (k) − H (k)|.
(v) |IF (Y, u) − u − IF (Y, v) + v| ≤ |u−v|∨(1−f (Y )) and |IF (Y, u) − u − IF (Y, v) + v| =
|u − v| if u, v ≤ F (Y − 1) or u, v ≥ F (Y ).
Moreover, EF supu,v∈Ψ(ε) |IF (Y, u) − u − IF (Y, v) + v|2 ≤ 4ε2 , for any interval
Ψ(ε) ⊂ [0, 1] of length ε2 .
31
Table 10: Values of functionals of the new nonrandomized transform I (·, ·) for all possible
values of Y relative to inverted cdfs at points u and v. For instance, IF (Y, u)−IF (Y, v) =
0 if Y < F −1 (u) and Y < F −1 (v), while IF (Y, u) − IF (Y, v) = −δ F (u) if Y = F −1 (u) <
F −1 (v).
Y < F −1 (u)
Y = F −1 (u)
Y > F −1 (u)
The value of IF (Y, u)
1
1 − δ F (u)
0
The value of 1{IF (Y, u) ≤ v}
v=0
0
0
1
v ∈ (0, 1)
0
1{1 − δ F (u) ≤ v}
1
v=1
1
1
1
Y < F −1 (v)
The value of IF (Y, u) − IF (Y, v)
0
Y = F −1 (v)
−δ F (u)
−1
δ F (v)
Y > F −1 (v)
δ F (v) − δ F (u)
−1 + δ F (v)
1
1 − δ F (u)
0
The value of IF (Y, u) IF (Y, v)
Y < F −1 (v)
1
0
Y = F −1 (v)
1 − δ F (u)
1 − δ F (v)
(1 − δ F (u)) (1 − δ F (v))
0
0
0
0
Y > F −1 (v)
Y < H −1 (u)
The value of IF (Y, u) − IH (Y, u)
0
Y = H −1 (u)
−δ F (u)
−1
δ H (u)
Y > H −1 (u)
δ H (u) − δ F (u)
−1 + δ H (u)
1
1 − δ F (u)
0
The value of IF (Y, u) IH (Y, u)
Y < H −1 (u)
1
0
Y = H −1 (u)
1 − δ F (u)
1 − δ H (u)
(1 − δ F (u)) (1 − δ H (u))
0
0
0
0
Y > H −1 (u)
32
(vi) EFz 1 F † Y † < u = IF (Y, u).
(vii) EFz [IF,M (Y, u) IF,M (Y, v)] =
8.2
1
I
M F
(Y, u ∧ v) + 1 −
1
M
IF (Y, u) IF (Y, v).
Functional weak convergence of discrete martingales
In this section we present Lindeberg-Feller-type sufficient conditions for functional weak
convergence of discrete martingales. In general, to establish the weak convergence one
needs to check tightness and finite-dimensional convergence. In case of martingales, both
parts can be verified without imposing restrictive conditions. Here we state a result of
Nishiyama (2000), which extends Theorem 2.11.9 of van der Vaart and Wellner (1996)
to martingales, see also Theorem A.1 in Delgado and Escanciano (2007). Further details
on notation and definitions can be found in books Van der Vaart and Wellner (1996) for
empirical processes and row-independent triangular arrays and in Jacod and Shiryaev
(2003) for finite-dimensional semimartingales. For every T , let ΩT , F T , {FtT }, P T be
a discrete stochastic basis, where ΩT , F T , P T is a probability space equipped with a
filtration FtT . For nonempty set Ψ, let {ξ Tt }t=1,2,... be a ℓ∞ (Ψ)-valued martingale
difference array with respect to filtration FtT , i.e. for every t, ξ Tt maps ΩT to ℓ∞ (Ψ),
the space of bounded, R-valued functions on Ψ with sup-norm k · k = k · k∞ and for
each u ∈ Ψ, ξ Tt (u) is a R-valued martingale difference array: ξ Tt (u) is FtT -measurable
and E ξ Tt (u) | FtT = 0. We are interested in studying the weak convergence of discrete
P
martingales Tt=1 ξ Tt . Denote a decreasing series of finite partitions (DFP) of Ψ as Π =
SNΠ (ε)
{Π(ε)}ε∈(0,1)∩Q , where Π(ε) = {Ψ(ε; k)}1≤k≤NΠ (ε) such that Ψ = k=1
Ψ(ε; k), NΠ (1) =
1 and limε→0 NΠ (ε) = ∞ monotonically in ε. The ε-entropy of the DFP Π is HΠ (ε) =
p
log NΠ (ε). The quadratic Π-modulus of ξ Tt is R+ ∪{∞}-valued process
v
"
#
u T
uX
1
2
T
T
T
max t
ξ t Π,T = sup
sup ξ t (u) − ξ t (v) | FtT .
(8)
E
ε∈(0,1)∩Q ε 1≤k≤NΠ (ε)
u,v∈Ψ(ε;k)
t=1
Theorem A. Let {ξ Tt }t=1,2,... be a ℓ∞ (Ψ)-valued martingale difference array and
P
N1) (conditional variance convergence) Tt=1 E ξ Tt (u)ξ Tt (v) | FtT →P T V (u, v) for every
u, v ∈ Ψ;
33
N2) (Lindeberg condition)
PT
t=1
E
h
ξ Tt
2
1
ξ Tt
>ε |
FtT
i
→P T 0 for every ε > 0;
N3) (partitioning entropy condition) there exist a DFP Π of Ψ such that ξ Tt Π,T =
R1
OP T (1) and 0 HΠ (ε)dε < ∞.
P
Then Tt=1 ξ Tt ⇒ S, where S has normal marginals (S (v1 ) , S (v2 ) , . . . , S (vd ))) ∼d N(0, Σ)
with covariance Σ = {V (vi , vj )}ij .
8.3
Additional technical assumptions
To establish the asymptotic properties of the biparameter process S2T we need the following assumption for uniform convergence of different empirical quantities.
Assumption A. Under H1T , the following uniform limits to continuous functions exist
1. plimT →∞
1
T
2. plimT →∞
1
T
3. plimT →∞
1
T
4. plimT →∞
1
T
5. plimT →∞
1
T
PT
t=2
γ t−1,θ0 (u2 , v2 ) γ t,θ0 (u1 , v1 ),
PT
t=2 It−1,θ 0
(v2 ) γ t,θ0 (u1 , v1 ),
PT
(u2 ) d (Ht (· | Ωt ) , Ft,θ0 (· | Ωt ) , u1 ),
PT
(u2 ) ∇ (Ft,θ0 (· | Ωt ) , u1 ).
t=2 It−1,θ 0
PT
t=2 It−1,θ 0
t=2 It−1,θ 0
(u2 ) E [It,θ0 (u1 ) ℓt (Yt , Ωt ) | Ωt ],
As it is discussed in the text, these conditions restrict the dynamics of the data process
such that some LLN holds, which is the case, e.g., for stationary and ergodic processes.
8.4
Proofs
Proof of Lemma A. (i) By definition of IF (Y, u), EG [IF (Y, u)] = (1−δ F (u))g(F −1(u))+
G(F −1 (u)) − g(F −1(u)) = d (G, F, u) − δ F (u)f (F −1(u)) + F (F −1 (u)) = d (G, F, u) + u.
Similarly, by direct calculation we obtain (ii), (iii), (vi) and (vii). We now provide a
detailed proof of (iv) and (v).
34
(iv) We prove a stronger result that for G ∈ M, such that supk |F (k) − G (k)| ∨
|H (k) − G (k)| ≤ supk |F (k) − H (k)| the expectation with respect to G is bounded:
2
EG (IF (Y, u) − IH (Y, u)) ≤ 9 supk |F (k) − H (k)|. Then, the required bound is obtained by setting G ≡ F .
Since |IF (Y, u) − IH (Y, u)| never exceeds 1, we have that EG (IF (Y, u) − IH (Y, u))2 ≤
EG [|IF (Y, u) − IH (Y, u)|], therefore we bound the latter expectation.
Suppose that F −1 (u) = H −1(u). Then IF (Y, u) − IH (Y, u) = δ H (u) − δ F (u) for
Y = F −1 (u), i.e. with probability g (F −1 (u)), and is zero for other Y . Therefore,
−1
EG [|IF (Y, u) − IH (Y, u)|] = |δ H (u) − δ F (u)| g F (u)
≤ F F −1 (u) − H F −1 (u) + f F −1 (u) − g F −1 (u) δ F (u) + h F −1 (u) − g F −1 (u) δ H (
≤ sup |F (k) − H (k)| + sup |f (k) − g (k)| + sup |h (k) − g (k)|
k
k
k
≤ 5 sup |F (k) − H (k)| ,
k
since δ F (u), δH (u) ∈ [0, 1) and supk |h (k) − g (k)| ≤ 2 supk |F (k) − G (k)|.
Suppose that F −1 (u) < H −1 (u).
Note that IF (Y, u) − IH (Y, u) = 0 for Y 6∈
[F −1 (u), H −1(u)]. We separately bound each term in
−1
EG [|IF (Y, u) − IH (Y, u)|] = EG |IF (Y, u) − IH (Y, u)| 1 Y = F (u)
+ EG |IF (Y, u) − IH (Y, u)| 1 Y = H −1 (u)
+ EG |IF (Y, u) − IH (Y, u)| 1 F −1 (u) < Y < H −1(u) .
For Y = F −1 (u), IF (Y, u) − IH (Y, u) = −δ F (u). Then
−1
−1
−1
−1
EG (IF (Y, u) − IH (Y, u)) 1 Y = F (u) = IF F (u), u − IH F (u), u g(F (u))
= δ F (u)g F −1 (u)
= F F −1 (u) − u + δ F (u) g F −1 (u) − f F −1 (u)
≤ sup |F (k) − H (k)| + sup |f (k) − g (k)| ≤ 3 sup |F (k) − H (k)| ,
k
k
k
since δ F (u) ∈ [0, 1) and for u ∈ [H (F −1 (u)) , F (F −1 (u))] we have that F (F −1 (u)) − u ≤
F (F −1 (u)) − H (F −1 (u)).
35
For Y = H −1 (u), IF (Y, u) − IH (Y, u) = −1 + δ H (u).
Then
−1
EG |IF (Y, u) − IH (Y, u)| 1 Y = H (u)
= IF H −1(u), u − IH H −1 (u), u g(H −1(u))
= (1 − δ H (u)) g H −1(u)
= u − H H −1(u) − 1 + (1 − δ H (u)) g H −1 (u) − h H −1 (u)
≤ sup |F (k) − H (k)| + sup |h (k) − g (k)| ≤ 3 sup |F (k) − H (k)| ,
k
k
k
since δ H (u) ∈ [0, 1) and for u ∈ [H (H −1 (u) − 1) , F (H −1 (u) − 1)] we have that u −
H (H −1 (u) − 1) ≤ F (H −1(u) − 1) − H (H −1 (u) − 1).
For F −1 (u) < Y < H −1 (u), IF (Y, u) − IH (Y, u) = −1. Then
EF |IF (Y, u) − IH (Y, u)| 1 F −1 (u) < Y < H −1(u)
H −1 (u)−1
=
X
k=F −1 (u)+1
g(k) = G(H −1 (u) − 1) − G(F −1 (u))
≤ F (H −1(u) − 1) − F (F −1(u)) + 2 sup |G(k) − F (k)|
k
≤ F (H −1(u) − 1) − H(H −1(u) − 1) + 2 sup |G(k) − F (k)| ≤ 3 sup |F (k) − H(k)|
k
k
since H(H −1(u) − 1) < u < F (F −1 (u)) < F (H −1 (u) − 1).
Adding everything together, get that EG |IF (Y, u) − IH (Y, u)|2 ≤ 9 supk |F (k) − H (k)|
for F −1 (u) < H −1 (u). This equation is symmetric with respect to F and H; therefore, it
holds also for F −1 (u) > H −1 (u).
(v) Let [a, b] denote the interval Ψ(ε) of length ε2 , sup ξ 2 denote the supremum of ξ 2
over u, v ∈ [a, b], where ξ := IF (Y, u) − u − IF (Y, v) + v.
Note that |ξ| ≤ 1; moreover, if [F (Y − 1), F (Y )] ∩ [a, b] = ∅, then sup |ξ| = ε2 and if
[a, b] ⊂ [F (Y − 1), F (Y )], then sup |ξ| =
1−f (Y ) 2
ε.
f (Y )
Suppose that F −1 (a) = F −1 (b), i.e. [a, b] ⊂ [F (F −1 (a) − 1), F (F −1(a))]. Then
P
1−f (F −1 (a))
2
2
ε2 f (F −1(a)) = 2(1 −
≤
[sup
|ξ|]
=
ε
sup
ξ
f
(k)
+
EF
EF
k6=F −1 (a)
f (F −1 (a))
f (F −1(a)))ε2 ≤ 2ε2 .
36
Suppose that F −1 (a) < F −1 (b), i.e. [a, b] contains at least one point F (k) or even
intervals [F (k − 1), F (k)] ⊂ [a, b]. On such intervals, |ξ| goes up to 1 − f (k), but the
probability of Y taking all such k is bounded by b − a. More precisely,
2
≤ EF [sup |ξ|]
EF sup ξ
X
1 − f (F −1 (a)) 2
2
f (k) +
= ε
ε f (F −1 (a))
−1
f (F (a))
k<F −1 (a)−1
X
1 − f (F −1 (b)) 2
ε f (F −1 (b)) + ε2
+
f (k) +
−1
f (F (b))
−1
−1
k∈[F
(a)+1,F
(b)−1]
X
f (k)
k>F −1 (b)
< 4ε2 ,
since the sum of the first and the last terms is below ε2 , the second and the fourth terms
P
each is bounded by ε2 and the third term is k∈[F −1 (a)+1,F −1 (b)−1] f (k) = F (F −1 (b) − 1)−
F (F −1 (a) + 1) ≤ b − a = ε2 .
Proof of Lemma 1. Substitute G = F = Fθ 0 (· | Ωt ) in Lemma A(i) to demonstrate
that E [It,θ0 (u) | Ωt ] = E [It,θ0 (u)] = u, therefore It,θ0 (u)−u is a martingale difference sequence for every u ∈ [0, 1]. The conditional variance expression follows from Lemma A(iii)
by taking G = F = Fθ0 (· | Ωt ).
However the It,θ0 (u) are not independent in general. To show that, note that bivariate
independence requires that
Pr (It,θ0 (u) ≤ u1 , It−1,θ0 (u) ≤ u2 ) = Pr (It,θ0 (u) ≤ u1 ) Pr (It−1,θ0 (u) ≤ u2 )
for all u, u1 and u2 ∈ [0, 1]. Now we see that the lhs is
E [1{It,θ0 (u) ≤ u1 } 1{It−1,θ0 (u) ≤ u2 }] = E [E [1{It,θ0 (u) ≤ u1 } 1{It−1,θ0 (u) ≤ u2 } | Ωt ]]
= E [1{It−1,θ0 (u) ≤ u2 } E [1{It,θ0 (u) ≤ u1 } | Ωt ]]
and now, for u1 , u ∈ (0, 1) and under H0 ,
−1
E [1{It,θ0 (u) ≤ u1 } | Ωt ] = 1 − Fθ0 Fθ0 (u | Ωt ) | Ωt
n
o
+ 1 1 − δFθ0 (·|Ωt ) (u) ≤ u1 f θ0 (· | Ωt ) Fθ−1
(u | Ωt ) ,
0
which depends on Ωt , and therefore E (1{It,θ0 (u) ≤ u1} | Ωt ) 6= E (1{It,θ0 (u) ≤ u1 }) with
positive probability, and independence does not follow in general.
37
Proof of Lemma 2. Because Utr (θ0 ) are continuous, Fbθr0 (u) is a (uniform) consistent
estimate of cdf of Utr (θ0 ). Then by Lemma A(vi) and A(vii) and ULLN we get the uniform
consistency of Fbθr0 ,M (u) and Feθr0 (u). The efficiency gain comes from Lemma A(ii).
Proof of Lemma 3. We need to verify conditions N1-N3 of Theorem A. Fix ε > 0 and
take Ψ = [0, 1] with usual norm and equidistant partition 0 = u0 < u1 < . . . < uNΠ (ε) =
1, i.e. partition of [0, 1] in NΠ (ε) = [ε−2 ] + 1 equal intervals of length ε2 (the last
√
interval maybe even smaller), Ψ(ε; k) = [uk−1 , uk ] and ξ Tt = (IF (Yt , u) − u) / T , which
is a square integrable martingale difference by Lemma 1. Then Condition N1 follows
from Lemma 1 and Assumption 1. Condition N2 is satisfied because for T > 1 + [ε−2 ],
n
o
√
the indicator 1 supu∈[0,1] |IF (Yt , u) − u| / T > ε = 0. Condition N3 follows from the
R1
bound in Lemma A(v). Indeed, 0 HΠ (ε)dε < ∞ and
ξ Tt
Π,k
≤
√
1
max
ε2 ≤ 1 a.s.
1≤k≤N
(ε)
ε
Π
ε∈(0,1)∩Q
sup
Proof of Lemma 4. Apply weak convergence result from Lemma 3 under GT,θ0 (· | Ωt )
√
T
with ξ t := IFθ0 (·|Ωt ) (Yt , u) − u − d (GT,θ0 (· | Ωt ) , Fθ0 (· | Ωt ) , u) / T , which is a
square integrable martingale difference because of Lemma A(i) with G = GT,θ0 (· | Ωt )
and F = Fθ0 (· | Ωt ). Then Condition N1 follows from Lemma A(iii) and the fact that
d (G, F, u, v) are bounded in absolute value by T −1/2 a.s. Condition N2 is satisfied because
for T > 1 + [ε−2 ], the indicator is 0. Condition N3 follows from the bound in Lemma A(v)
and the fact that (EG [·] − EF [·]) applied to a.s. bounded r.v. are bounded in absolute
P
value by T −1/2 a.s. We obtain that Tt=1 ξ Tt ⇒ S, the same limit as in Lemma 3. Finally,
P
use additivity of d (·, ·, ·) in the first argument and apply ULLN to ST − Tt=1 ξ Tt =
√
PT
PT
t=1 d (GT,θ0 (· | Ωt ) , Fθ0 (· | Ωt ) , u) / T = δ
t=1 d (H (· | Ωt ) , Fθ 0 (· | Ωt ) , u) /T .
Proof of Lemma 5. Under H1T , i.e. under GT,θ0 , Equation (4) can be established
using standard methods, applying Doob and Rosenthal inequalities for MDS (Hall and
√ T
Heyde, 1980) T ξ t := IFbθ (·|Ωt ) (Yt , u)−IFθ0 (·|Ωt ) (Yt , u)−d GT,θ0 (· | Ωt ) , FbθT (· | Ωt ) , u
T
38
+d (GT,θ0 (· | Ωt ) , Fθ0 (· | Ωt ) , u) . Define zT :=
PT
t=1
ξ Tt . When it is necessary, we will
write explicitly arguments: zT (u, b
θT ). We show that supu |zT | = op (1). Since
√
T b
θT − θ0 = OP (1), it is sufficient to establish that for some γ < 1/2
sup
u,kη−θ0 k≤T −γ
|zT (u, η)| = op (1).
Note that for T > δ 2 /ν 21 , by Assumption 3C,
Pr sup max |GT,t,θ0 (y | Ωt ) − Ft,η (y | Ωt )| > ν 1 ≤ MF T −γ /ν 1 .
η,t
y
(9)
First, we will show that ∀ η, u |zT | = op (1). Since ξ Tt are bounded by 2 in absolute
value and form a martingale difference sequence with respect to Ωt , by the Doob inequality
∀p ≥ 1 and ∀ε > 0
P
max |zt | > ε
t=1,...,T
≤ E |zT |p /εp ,
and by Rosenthal inequality, ∀p ≥ 2 ∃C1
nX
op/2 X
p
T 2
T p
E |zT | ≤ C1 E
E ξ t | Ωt
+
E ξt
.
Take p = 4. The first term is small because of bounds in Lemma A(iv) and (9). Because
√ P
p
E ξ Tt
≤ 2T 1−p/2 . Therefore we have a pointwise bound. Uniforξ Tt ≤ 2/ T ,
mity in u, η can be established using monotonicity of IFθ (·|Ωt ) (Yt , u) and continuity of
d GT,θ0 (· | Ωt ) , FbθT (· | Ωt ) , u by employing bounds in Lemma A(iv) and (9).
Finally, use that uniformly in u
1 X
√
d GT,θ0 (· | Ωt ) , FbθT (· | Ωt ) , u − d (GT,θ0 (· | Ωt ) , Fθ0 (· | Ωt ) , u)
T t
1X
√
θT − θ0
= T b
∇ (Fθ0 (· | Ωt ) , u) + op (1).
T t
Proof of Theorem 1. The joint weak convergence (6) follows from finite-dimensional
convergence by CLT for MDS, while tightness was established in the proof of Lemma 4.
39
Proof of Theorem 2. Note that
S2T =
T
X
ξ Tt +
t=2
where
ξ Tt :=
1
{(Iθ0 ,T (u1 ) − u1 ) Iθ0 ,T −1 (u2 ) + u1 (Iθ0 ,1 (u2 ) − u2 )} ,
(T − 1)1/2
1
{(It,θ0 (u1 ) − u1 ) It−1,θ0 (u2 ) + u1 (It,θ0 (u2 ) − u2 )}
(T − 1)1/2
is a square integrable martingale difference by Lemma 1. The rest is similar to the
proof of Theorem 1. To obtain S2T (u) ⇒ S2∞ (u) under H0 , verify conditions N1-N3
of Theorem A for ξ Tt as it is done in the proof of Lemma 3. The covariance function of
S2∞ (u) is
V2 (u, v) := (u1 ∧ v1 ) (u2 ∧ v2 ) − 3u1v1 u2 v2
T
1X
+ (u1 ∧ v1 ) plim
γ t−1,θ0 (u2 , v2 )
T →∞ T t=2
T
1X
− plim
γ t,θ0 (u1 , v1 ) It−1,θ 0 (u2 ∧ v2 ) − γ t−1,θ0 (u2 , v2 )
T →∞ T
t=2
T
1X
γ t,θ0 (u1 , v1 ) It−1,θ0 (v2 )
T →∞ T t=2
+ (u2 ∧ v1 ) u1 v2 − u1 plim
T
1X
γ t,θ0 (u1 , v1 ) It−1,θ 0 (u2 ) .
T →∞ T t=2
+ (u1 ∧ v2 ) u2 v1 − v1 plim
Under H1T , apply the same weak convergence result under GT,t,θ0 (· | Ωt ) with
√
ζ Tt := ξ Tt − It−1,θ0 (u2 ) d (GT,t,θ0 (· | Ωt ) , Ft,θ0 (· | Ωt ) , u1) / T − 1
√
+ u1 d (GT,t,θ0 (· | Ωt ) , Ft,θ0 (· | Ωt ) , u2) / T − 1,
which is a square integrable martingale difference because of Lemma A(i) with G =
GT,t,θ0 (· | Ωt ) and F = Fθ0 (· | Ωt ). Then proceed as in the proof of Lemma 4.
T
T
ζt ,
In order to establish (7), repeat the steps of the proof of Lemma 5 for e
ζ t := ζ Tt − b
T
where b
ζ t is ζ Tt with Ft,bθT in place of Ft,θ0 .
Proof of Theorem 4. Repeat the arguments of the proofs of Theorems 1 and 2 for
sample generated by Fθ T , defined in Assumption 6, to obtain conditional convergence.
Then follow as in Andrews (1997) proof of Corollary 1.
40
8.5
Checking assumptions for the Poisson model
Here we write Yt for Yt⋆ . For Poisson model Yt | Ωt ∼ Poisson(λt ) the probability
distribution is Pr(Yt = k | Ωt ) = Pλt (k) =
λkt exp(−λt )
k!
and the cumulative distribution
function is
Ft,θ (k | Ωt ) =
k
X
j=0
Pr(Yt = j | Ωt ) =
k
X
λj exp(−λt )
t
j=0
j!
= Q(k, λt ),
where Q(·, ·) is the regularized gamma function, and λt = λt (β) = exp(Xt′ β), t = 1, 2, . . ..
If covariates Xt are iid or stationary and ergodic, and Ωt omits lags of the dependent
variable Yt , then the LLN applies both under the null and local alternatives (like, e.g.,
the local alternative considered in Eq. (2.12) in Cameron and Trivedi, 1990) to justify
Assumptions 2-6 and Assumption A, which involve functions of Ωt that are uniformly
continuous in u. However, it can also be interesting to allow the intensity to depend on
lags of the dependent variable. For simplicity we consider AR(1) dynamics. AR(p) can
be treated similarly but is more lengthy. The parameters enter through λt = λt (θ) =
α0 + α1 λt−1 + ρYt−1 , t = 1, 2, . . ., and are gathered in θ = (α0 , α1 , ρ)′ . We assume that
α0 , α1 , ρ are positive, λ0 and Y0 are fixed and α1 + ρ < 1. Under these conditions, there
exist a unique stationary and ergodic solution to this model (Fokianos et al., 2009). Such
data generating processes allow to use results on (generic, uniform) LLN, which facilitate
the checking of assumptions in the paper. Conditions for stationarity and ergodicity
for nonlinear λt (θ) can be found in Neumann (2011) and are directly applicable to the
analysis under the null hypothesis. However, we are not aware of LLN results for these
models under local alternatives despite Fokianos and Neumann (2013, Proposition 2.3(ii))
use related arguments.
Let λt,0 = λt (θ0 ) and the null hypothesis is Yt | Ωt ∼ Poisson(λt,0 ) for some θ0 ∈ Θ.
Then Ut = Q(Yt + 1, λt,0 ) and Ut− = Q(Yt , λt,0 ), and the nonrandomized transform Yt 7→
41
It,θ0 (u) for u ∈ [0, 1] is
0,
u − Q(Y , λ )
t
t,0
Yt !,
It,θ0 (u) =
Yt
λt,0 exp(−λt,0 )
1,
u ≤ Q(Yt , λt,0 );
Q(Yt , λt,0 ) ≤ u ≤ Q(Yt + 1, λt,0 );
Q(Yt + 1, λt,0 ) ≤ u,
from where one obtains the empirical processes and the test statistics defined in Sections 1-2.
Now consider Assumption 1. For Poisson model
γ t,θ0 (u, v) =
(Q(k + 1, λt,0 ) − u ∨ v) (u ∧ v − Q(k, λt,0 ))
k!1{k(u) = k(v)} ,
λkt,0 exp(−λt,0 )
where k = k (u) = min{y : Q(y, λt,0 ) ≥ u}. For the Poisson DGP described above, Yt is
stationary and ergodic, γ ∞ (u, v) := E γ 1,θ0 (u, v) satisfies Assumption 1. By the same
argument Assumptions 2, 3D, 4C, 5 are fulfilled.
Assumption 3A and 3B are trivial. For Assumption 3C note that
!
k
k−1 j
X
λk
λt X λjt
exp(−λt )λ̇t = − t exp(−λt )λ̇t ,
−
Ḟt,θ (k | Ωt ) =
j!
j!
k!
j=0
j=0
where
λ̇t =
∂λt−1
∂λt−1
∂λt−1
1 + α1
, λt−1 + α1
, Yt−1 + α1
∂α0
∂α1
∂ρ
′
.
The last expression can be iterated from t − 1 to t = 1 and because α1 < 1 the arithmetic
progression sum of mean squares is bounded, as in the proof of Lemma 3.2 of Fokianos
et al. (2009).
Assumption 4A, 4B and 6B are standard, see e.g. Andrews (1997) which adapts to
Poisson model using Theorem 3.1 of Fokianos et al. (2009).
Assumption 6A is trivial, because there is no explanatory variables other than own
past values.
42
9
Acknowledgements
We thank Juan Mora for useful comments. Support from the Ministerio Economia y
Competitividad (Spain), grants ECO2012-31748, ECO2014-57007p, MDM 2014-0431,
Comunidad de Madrid, MadEco-CM (S2015/HUM-3444), and Fundación Ramón Areces
is gratefully acknowledged.
References
[1] Andrews, D.W.K. (1997) A conditional Kolmogorov test, Econometrica 65, 10971128.
[2] Bai, J. (2003) Testing Parametric Conditional Distributions of Dynamic Models,
Review of Economics and Statistics 85, 531-549.
[3] Basu, D. and R. de Jong (2007). Dynamic Multinomial Ordered Choice with an
Application to the Estimation of Monetary Policy Rules. Studies in Nonlinear Dynamics and Econometrics, 11, 1507-1507.
[4] Burke, M. D., Csorgo M., Csorgo S. and P. Revesz (1978). Approximaiton of the
empirical process whe parameters are estimated. Annals of Probability, 7, 790-810.
[5] Cameron A.C. and P.K. Trivedi (1990) Regression-based tests for overdispersion in
the Poisson model, Journal of Econometrics 46, 347-364.
[6] Corradi, V. and R. Swanson (2006) Bootstrap conditional distribution test in the
presence of dynamic misspecification, Journal of Econometrics 133, 779-806.
[7] Czado, C., T. Gneiting and L. Held (2009). Predictive model assessment for count
data. Biometrics, 65, 1254-1261.
[8] Davis R. A., W. T. M. Dunsmuir and S. B. Streett (2003) Observation-Driven Models
for Poisson Counts. Biometrika 90, 777-790.
[9] Delgado, M. and Escanciano, J. C. (2007) Nonparametric tests for conditional symmetry in dynamic models. Journal of Econometrics 141, 652-682.
[10] Delgado, M. and W. Stute (2008) Distribution-free specification tests of conditional
models. Journal of Econometrics 143, 37-55.
[11] Dolado, J. J. and Maria-Dolores, R. (2002). Evaluating Changes in the Bank of
Spain’s Interest Rate Target: An Alternative Approach Using Marked Point Processes, Oxford Bulletin of Economics and Statistics, 64, 159-182.
43
[12] Doukhan P., K. Fokianos and D. Tjostheim (2012) On weak dependence conditions
for Poisson autoregressions, Statistics and Probability Letters 82, 942-948.
[13] Fokianos K. and M. Neumann (2013) A goodness-of-fit test for Poisson count processes, Electronic Journal of Statistics 7, 793-819.
[14] Fokianos K., A. Rahbek and D. Tjostheim (2009) Poisson Autoregression, Journal
of the American Statistical Association 104, 1430-1439.
[15] Genest, C., J. Nešlehová and B. Rémillard (2014). On the empirical multilinear
copula process for count data. Bernoulli, 20, 1344-1371.
[16] Giacomini R., Politis D. and H. White (2013) A Warp-Speed Method for Conducting
Monte Carlo Experiments Involving Bootstrap Estimators, Econometric Theory 29,
567-589.
[17] Greene, W. and Hensher D., (2010). Modeling Ordered Choices. A Primer. Cambridge University Press.
[18] Hamilton, J. and O. Jordá (2002). A model of the Federal Funds rate target. The
Journal of Political Economy 110, 1135-1167.
[19] Handcock, M. S. and Morris, M. (1999). Relative Distribution Methods in the Social
Sciences. Springer, New York.
[20] Jacod, J. and Shiryaev, A. N. (2003). Limit Theorems for Stochastic Processes.
Springer, 2d Ed. Berlin.
[21] Jung, R.C., M. Kukuk and R. Liesenfeld (2006). Time series of count data: modeling,
estimation and diagnostics. Computational Statistics and Data Analysis, 51, 23502364.
[22] Kauppi, H. and Saikkonen, P. (2008). Predicting U.S. recessions with dynamic binary
response models. Review of Economics and Statistics 90, 777-791.
[23] Kedem and Fokianos (2002). Regression Models for Time Series Analysis. Willey,
New Jersey.
[24] Kheifets, I. (2015). Specification Tests for Nonlinear Time Series Models. Econometrics Journal 18, 67-94.
[25] Kheifets, I., and C. Velasco (2013). Model Adequacy Checks for Discrete Choice
Dynamic Models. In Recent Advances and Future Directions in Causality, Prediction,
and Specification Analysis Essays in Honor of Halbert L. White Jr. Chen, Xiaohong;
Swanson, Norman R. (Eds.), 363-382.
[26] Lee, S. (2014). Goodness of fit test for discrete random variables, Computational
Statistics and Data Analysis 69, 92-100.
44
[27] Machado, J. A. F., and Santos Silva, J. M. C. (2005). Quantiles for Counts, Journal
of the American Statistical Association 100, 1226-1237.
[28] Mora, J. and A.I. Moro-Egido (2007). On specification testing of ordered discrete
choice models. Journal of Econometrics 143, 191-205.
[29] Neumann M. (2011) Absolute regularity and ergodicity of Poisson count processes,
Bernoulli 17, 1268-1284.
[30] Nishiyama Y. (2000). Weak convergence of some classes of martingales with jumps.
Annals of Probability 28, 685-712.
[31] Rydberg and Shephard (2003). Dynamics of Trade-By-Trade Price Movements: Decomposition and Models. Journal of Financial Econometrics 1, 2-25.
[32] Rosenblatt M. (1952). Remarks on a Multivariate Transformation. Annals of Mathematical Statistics 23, 470-72.
[33] Startz R. (2008). Binomial Autoregressive Moving Average Models with an Application to U.S. Recessions. Journal of Business and Economic Statistics 26, 1-8.
[34] Stute W., Gonzalez Manteiga W., and M. Presedo Quindimil (2002). Bootstrap
Based Goodness-Of-Fit-Tests. Metrika 40, 243-256.
[35] Stute W. and Zhu, L.-X. (2002). Model checks for generalized linear models. Scandinavian Journal of Statistics 29, 535-545.
[36] van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and Empirical
Processes: With Applications to Statistics. Springer, New York.
45
| 10math.ST
|
1
Throughput Maximization for Delay-Sensitive
Random Access Communication
arXiv:1711.02056v1 [cs.IT] 6 Nov 2017
Derya Malak, Howard Huang, and Jeffrey G. Andrews
Abstract
We consider a single cell uplink in which many devices randomly transmit a data payload to the base
station. Given a fixed latency constraint per device, we propose a time and frequency slotted random access
scheme with retransmissions, which when necessary, are Chase combined at the receiver. We analyze the
proposed setting at constant SNR. We characterize the scaling of random access throughput versus the latency,
by optimizing the number of retransmissions and the number of frequency bins. For infinite block length
(IBL), we conclude that at low SNR devices should completely share the time and frequency resources. For
high SNR, however, the number of frequency bins should scale with altered load, and the slot duration for each
retransmission is determined by the outage tolerance. Since infinite packet sizes are not possible, we extend
our model to the finite block length (FBL) regime and characterize the gap versus the IBL regime. We also
provide some new results for FBL capacity to bound the probability of outage. The proposed random access
model gives an upper bound for the total uplink traffic that can be successfully decoded for a single receive
antenna given the latency constraint, and provides insights for 5G cellular system design.
I. I NTRODUCTION
Machine-type communication (MTC) applications are rapidly growing, and will become an increasingly important source of traffic and revenue in cellular networks [2]. Unlike most human-generated
Part of the manuscript has been presented at the 2017 IEEE ICC Wireless Communications Symposium [1].
D. Malak is with the Research Laboratory of Electronics (RLE), at The Massachusetts Institute of Technology, Cambridge, MA 02139
USA (email: [email protected]).
H. Huang is with Bell Labs, Nokia, Murray Hill, NJ 07974, USA (e-mail: [email protected]).
J. G. Andrews is with the Wireless Networking and Communications Group (WNCG), The University of Texas at Austin, Austin,
TX 78701, USA (email: [email protected]).
This work was done when D. Malak was with WNCG, The University of Texas at Austin, TX, USA.
2017.
Last revised: November 7,
2
traffic, MTC is often characterized by a very large number of small transactions. The LTE air interface
design for high-data-rate applications may not effectively support reliable and low-latency MTC with
a vast number of devices. Emerging Internet of Things (IoT) examples include sensor [3] and time
of arrival measurements [4], smart grids for power distribution automation, industrial manufacturing,
control and automation applications [5], traffic safety, and monitoring of materials in industrial plants
[6].
Random access (RA) is a critical challenge in IoT networks, because of the massive number
of devices that wish to access the channel and their frequently very short payloads [7], [8], [9].
Furthermore, there is overhead for scheduled access and it doesn’t make sense to invest if the payload
size is small. In such scenarios it can be envisioned that the devices never move into a scheduled
access phase [7], [8], instead just using the RA process [10], to send a small data payload. However,
a large number of RA transmissions can overload the random access channel (RACH), resulting in
a significant access delay. Thus, MTC will require sophisticated access management and resource
allocation to prevent RACH congestion [3].
A. Related Work and Motivation
Use cases for MTC and requirements for massive and ultra-reliable MTC have been outlined in
[11]. Guidelines for context-aware functionality for MTC were provided in [12]. Random access for
MTC and scheduling have been extensively studied in [13], [14], [15], and access for ultra-reliable and
low-latency communication was investigated in [5]. Random access overhead and congestion in MTC
networks were analyzed in [16], and access with collision resolution for massive MTC was proposed
in [17]. Various aspects such as power control [18], spectrum allocation [19], and energy-efficient
uplink resource allocation [20], and load control [21] for MTC traffic, have been studied.
There are complex tradeoffs between the energy efficiency [7], [22], reliability (using short packets)
[2], latency [9], [23], signaling overhead [24], [25], data payload size, and number of devices that can
successfully complete the RA process within a time slot [26]. An information theoretic upper bound on
the random access transmission rate was derived in [27], where a large number of devices with sporadic
traffic communicate in the uplink. Effective capacity of retransmission schemes was studied in [28].
Throughput of hybrid ARQ (HARQ) with incremental redundancy and the effect of the transmission
3
rate, queuing and hard-deadline constraints on the throughput were investigated in [29]. The power
and energy optimal uplink design for various access strategies was studied in [7], while an optimal
uncoordinated strategy to maximize the average throughput for a time-slotted RACH was developed in
[8]. A strategy that transmits both identity and data over the RACH was shown to support significantly
more devices compared to the conventional approach, where transmissions are scheduled after an initial
RA stage. An online delay-optimal proportionally fair scheduler for delay-sensitive MTC uplink traffic
was proposed in [30]. There is also an ongoing effort to optimize cellular networks to support ultra-low
complexity, power, and throughput for IoT, currently under standardization for LTE-A [31], [32].
Recently, some results on finite block length (FBL) have also been developed for some scenarios
[2], [9], [33], [34], [35], [36]. FBL for interference-limited regimes was analyzed in [33], and spectrum
sharing in the FBL regime using rate adaptation was proposed in [34].
Variants of slotted ALOHA have been studied to mitigate channel contention and achieve high
throughput values. Contention resolution diversity is an improvement of slotted ALOHA that relies
on MAC burst repetition and on interference cancellation (IC) for resolving collisions [37]. Irregular
repetition slotted ALOHA techniques can achieve high-throughput performance for small MAC frame
sizes when IC methods are adopted [38].
The optimal strategy in [7] for minimizing the total power consumed to transmit a fixed size
payload is not achievable in practice because it assumes ideal channel knowledge and ideal successive
interference cancellation (SIC). Therefore, we consider a suboptimal but practical random access
strategy with retransmissions and many orthogonal frequency bins. We initially studied a random
access structure with slotted ALOHA in [1] for an idealized regime that does not capture the finiteness
of the number of resources. In this paper, we characterize the theoretical performance of this strategy
in terms of the maximum arrival rate that can be supported for a fixed maximum delay for a given
number of resource symbols. We then optimize the resource utilization (retransmissions and number
of frequency bins) for different SNR regimes.
B. Contributions
We consider a single cell model in which a set of transmitters transmit to a common base station (BS).
For each transmit opportunity, there are K users wishing to communicate, where K is a realization of a
4
Poisson distribution with mean λ. The payload size of each user is L bits. We characterize maximum
Poisson arrival rate such that each device meets a fixed delay constraint, with probability at least
1 − δ. Our proposed RA policy is a generalization of slotted ALOHA, and has flexible slot length and
memory. Each frame is of duration T , which is the delay constraint, and is evenly partitioned into
M ≥ 1 slots. The total bandwidth W is equally divided into B ≥ 1 subbands. To prevent decoding
failure, i.e. outage, each device needs to complete the RA phase within M retransmission attempts
(not exceeding total duration T ) by selecting one bin at random at each attempt. The resulting SINR
of each transmission attempt is combined at the receiver; similar to HARQ with Chase combining
[39].
We analyze the proposed setting in the infinite block length (IBL) and finite block length (FBL)
regimes under low and high SNR conditions. We maximize the throughput over multiple retransmissions with respect to the deadline constraint T for a given bandwidth and outage constraint by
optimizing M and B. Exploiting the scaling results for the IBL regime, we then determine the scaling
of the throughput in the FBL regime. We also provide some new results for FBL block error rate
hence the maximal rate achievable to bound the probability of outage.
The key design insights for the proposed strategy are as follows.
•
At constant SNR, i.e. no fading, for the high SNR regime, the resources should be split sufficiently
such that devices do not experience any interference provided that the total number of resources
is sufficiently large. Hence, the number of frequency bins should approximately be the same as
the number of device arrivals, i.e. B ≈ K.
•
At constant SNR, for the low SNR regime we show that the devices should share the resources.
•
For the case of Rayleigh fading, although the variability of the channel causes a drop in the
number of arrivals that can successfully complete the random access phase, we briefly discuss
that similar conclusions extend to that case.
•
FBL regime is the practical case for short packet sizes. Although there is a gap in the throughput
of the FBL model from the IBL model, scaling results for both regimes are similar.
The insights could be applied to 5G cellular system design where delay-constrained communications
will be an important use case.
5
II. S YSTEM M ODEL
We consider a single cell uplink in which a set of devices transmits a fixed payload to the BS. We
propose a slotted ALOHA-like scheme with retransmissions and multiple frequency bins. Below, we
summarize the key notation and the random access model parameters used in the paper:
•
There is a fixed size payload of L bits per device.
•
Each device has a fixed latency constraint T .
•
The total available bandwidth is W .
•
The total symbol resources N = T W .
•
The total bandwidth W is evenly partitioned into B frequency bins.
•
Each payload is granted up to a maximum of M − 1 retransmission attempts, with M being the
total number of transmissions.
•
Each time frame of duration T is evenly split into M slots of duration T /M each.
•
The N resources are partitioned into N/MB resources each bin per retransmission.
•
The encoded block length per transmission should satisfy n ≤ N/MB.
•
The transmit rate is L/n bits per transmission attempt.
•
At a transmission attempt, a given device selects one of the B bins at random and transmits.
•
There is precise synchronization between the beginning and ending of transmissions and the slots,
and the transmission attempts occur at the beginning of each slot.
•
If a device can successfully transmit its payload in 1 ≤ m ≤ M consecutive attempts, it does not
use the remaining M − m slots.
•
The effective channel gains are constant. We later relax this assumption to the case of Rayleigh
fading (see Sect. III-B), and discuss how the constant gain results are modified by incorporating
the channel power distributions.
•
Devices perform open loop power control under which the received power at the BS from any
device, Pr , is constant.
•
•
The noise power at the base station is additive and constant with value σ 2 .
The received SNR per device at the BS is ρ = Pr /σ 2 .
In this section, we optimistically assume that the encoded block length of n symbols is sufficiently
6
large that we can exploit the IBL1 limit to characterize the system performance. An outage occurs
when a given device fails to communicate its payload on M consecutive attempts, i.e. a decoding error
occurs if the signal-to-noise-plus-interference ratio (SINR) across multiple transmissions is below the
L
threshold Γ. This threshold is given by Γ = 2 n − 1 for a given block length n per transmission.
Chase combining is used to aggregate the received signals across multiple transmissions at the BS,
resulting in maximal ratio combining of the desired signal. After the device transmits its payload, the
BS either sends ACK to the sender (device) to indicate the successful decoding of the payload if the
received payload is error free, or NACK if the decoding of the received payload was unsuccessful.
For tractability, we assume there are no errors or delay in the feedback (ACK/NACK), so there is
immediate retransmission on the next slot after a failure. In practice, there would be a gap of at least
one slot while feedback from the BS was sent and processed, which would increase the latency.
Proposition 1. Chase combiner output SNR. If M > 1, a device is allowed to retransmit if the
preceding one fails, for a total of M transmissions. In general, if the received signal vector during
transmission i = 1, . . . , M is ri = ai s + ni , where s ∈ Cn is the desired signal, ai is the complex
amplitude, and ni is an n-dimensional complex Gaussian noise vector with zero mean and covariance
σi2 In , then the SNR at the output of the Chase combiner is [1]
M
X
i=1
|ai |
2
!2 ,
M
X
i=1
|ai |2 σi2 .
(1)
Given M, denote the set of retransmission indices by M = {1, . . . , M}. For ease of notation, the
number of arrivals on the mth attempt where m ≤ M is denoted by km , and the set of arrivals up to
and including mth attempt is denoted by Km := {k1 , . . . , km }. If there are km ≥ 1 devices arriving
during transmission m ∈ M, each with receiver SNR ρ, in a given resource bin, then the SINR during
transmission m is ρ/(1 + (km − 1)ρ). Hence, from Prop. 1, the Chase combiner output SINR from
(1) as a result of m ∈ M transmissions is
SINR(Km ) =
ρm2
m
,
P
m+ρ
ki − m
i=1
1
We relax this assumption in Sect. V to the finite block length (FBL) regime.
m ∈ M.
(2)
7
A device fails on the mth attempt if the combined output SINR up to and including the mth attempt,
i.e. SINR(Km ), is below the threshold Γ. Given a target SINR outage rate δ per device, outage occurs
if more than a certain number of users share the same resources.
We define PFail (m) to be the probability of outage up to and including the mth attempt for a given
SINR threshold Γ, which is given as
PFail (m) = P[SINR(K1 ) < Γ, SINR(K2 ) < Γ, . . . , SINR(Km ) < Γ].
(3)
In other words, the joint probability in (3) is the intersection of the decoding failure events up to
and including the mth attempt. We emphasize that these m events are dependent, since the Chase
combiner output on the mth attempt, i.e. SINR(Km ), is a function of all previous receptions as well,
i.e. Km = {k1 , . . . , km }.
For the proposed setting, our objective is to optimize the random access strategy by calculating
the values of the number of bins B and the number of retransmission attempts M that maximize the
average user arrival rate λ that can be supported per slot, denoted by λopt . At the same time, for a
given number of symbol resources N (T sec, W Hz), the probability of outage given up to a maximum
of M transmission attempts should be less than or equal to δ. We then characterize the tradeoffs for
random access between throughput in terms of the average number of devices supported and latency.
The system model for M = 5 retransmissions is shown in Fig. 1. In this example, a device arriving
during slot 1 has until slot 5 to transmit. Another device arriving during slot 2 has until slot 6 to
transmit. This emphasizes the fact that devices can arrive during each slot. In the same figure, we
also illustrate a typical device that uses only m = 3 time slots out of 5 to retransmit its payload to
the BS by choosing frequency bins at random at each time slot. The first 2 attempts combined at the
BS cannot be decoded and the BS sends NACK back to the device after each failure. The payload is
successfully decoded after 2 retransmission attempts and the BS sends an ACK to the device.
Given the number of frequency bins B, the average number of arrivals per bin per transmission
interval is Poisson distributed with an average arrival rate of λB = λ/B. Hence, the probability of
k ≥ 1 devices choosing a given resource bin out of B bins for transmission, given that there is at
8
Fig. 1: An example RA scenario. The bandwidth is partitioned into B(= 4) frequency bins and time is partitioned into
M (= 5) transmit opportunities. The average arrival rate of devices for each slot is λM /M . We illustrate a typical device
that is unsuccessful on the first transmission attempt, and retransmits 2 times by choosing frequency bins at random at
each time slot. The transmission attempts are combined at the BS, yielding success in slot m = 3.
least one arrival, is given by the following conditional probability:
D(k, λB ) =
P(k device arrivals per bin, k ≥ 1)
λkB exp(−λB )
=
,
P(k ≥ 1)
k!(1 − exp(−λB ))
k ≥ 1,
(4)
where we note that the probability of at least one device arrival is P(k ≥ 1) = 1 − exp(−λB ).
Denote the average aggregate device arrival rate with up to M total transmissions per frame by λM ,
which is the sum of the rates of the original arrivals per slot, i.e. λ, and the arrivals occurring as a
result of failed transmissions up to a maximum of M − 1 consecutive attempts. It is given by
"
#
M
−1
X
λM = λ 1 +
PFail (m) ,
(5)
m=1
where PFail (m) as given in (3) denotes the fraction of devices that fail up to and including the mth
attempt. Hence, given maximum number of retransmission attempts M per time frame, the number of
arrivals at any time slot 1 ≤ m ≤ M is Poisson distributed2 with an average arrival rate of
2
λM
.
M
For tractability, we inherently have the Poisson distribution assumption for the composite arrival process. From [8] and [40], this
assumption is justifiable when the number of retransmissions is not too large.
9
The optimization problem is
λopt = max + λ
B, M ∈Z
s.t.
PFail (M) ≤ δ,
L
Γ = 2 n − 1,
C(SINR(Km )) ≥
(6)
TW
L
, m ∈ M, n ≤
,
n
MB
where PFail (M) is given in (3), and C(SINR(Km )) is the channel capacity of a typical device in the
uplink as a function of (2).
III. I NFINITE B LOCK L ENGTH
Shannon’s channel capacity is achievable at an arbitrarily low error rate when coding is performed
in the infinite block length (IBL) regime, i.e. using a code block of infinite length. However, since the
number of resources N is finite, the ratio L/N is always finite. Therefore, given L, the IBL scheme
gives an upper bound on the achievable transmit rate, hence a lower bound on n.
In this section, we assume that the encoded block length n symbols is sufficiently large so that we
can exploit the Shannon limit. The capacity of an additive white Gaussian noise (AWGN) channel
where interference is treated as noise in that case is
C(SINR) = log2 (1 + SINR).
(7)
For some number of retransmission attempts, combining the capacity constraint in (6) with the Chase
combiner output (2), we investigate the outage probability for both constant SNR and Rayleigh fading.
A. Constant SNR
At constant SNR, the Chase combiner output SINR as a result of m ∈ M transmissions, i.e.
SINR(Km ) as a function of Km , is given in (2). Hence, incorporating (7) in the capacity constraint in
(6), the maximum number of arrivals that can be supported at time slot m ∈ M can be determined
∗ :=
as a function of Km−1 . We denote this maximum by km
km (Km−1 ).
For a given maximum number of retransmission attempts M, an outage occurs if there are more
∗
than km
devices sharing the same bin on attempt m ∈ M, due to limited number of symbol resources
N/MB. Using (4), the outage probability at slot m ∈ M is calculated adding up the probabilities
10
D(km , λB ) over the number of arrivals that cannot be supported. Given B and N, Prop. 2 gives the
outage probability at constant SNR up to a maximum of M transmissions.
Proposition 2. Given B, M, symbol resources N, the outage probability in the IBL regime in the case
of constant SNR is given by the following expression:
X
PFail (M) =
∗ (K
km >km
m−1 )
λM
,
D km ,
B
m∈M
Y
(8)
where the lower limit of the summation in (8) is
nj
m−1
m2
m X k o
m+ L
−
−
ki , 1 ,
ρ
2n − 1
i=1
∗
km
(Km−1 )
= max
m ∈ M,
n ≤ N/(BM).
(9)
The aggregate arrival rate λM with up to M total transmissions is given by (5).
Proof. See Appendix A.
B. Small Scale (Rayleigh) Fading
For the case of small scale Rayleigh fading with mean 1/µ, let km arrivals choose a given frequency
bin, each having SNR ρ, during transmission m ∈ M. Using (1) and incorporating the channel power
distributions, the Chase combiner output SINR as a result of M transmissions is obtained as
ρ
(a)
SINR(KM ) = P
m∈M
P
m∈M
hm
2
hm 1 + Ikm
(b)
=
M 2 ρhM
P
,
M+
Ik m
(10)
m∈M
where hm , gi,m ∼ exp(µ), m ∈ M are independent and identically distributed (i.i.d.) channel power
distributions of the desired device and the interferers, respectively, where im ∈ {1, . . . , km − 1} is
P m −1
ρgim denote
the interferer index at retransmission attempt m, (a) follows from letting Ikm = ki=1
the total interference seen at transmission attempt m, and (b) is based on the assumption that hM is
unchanged within a time slot. This is consistent with the Rayleigh block fading model [26] in which
the power fading coefficients remain static over each time slot, and are temporally (and spatially)
independent with exponential distribution with parameter µ = 1.
11
The outage probability in the case of small scale Rayleigh fading is derived next.
Proposition 3. In the case of Rayleigh fading with mean 1/µ, the outage probability as a function of
B, M, L, ρ, N, and SINR threshold Γ is characterized as
PFail (M) =
X
(−1)
lM
lm ∈{lm−1 , lm−1 +1}
Y λM
, ln Γ ,
f µ,
B
n∈M
(11)
where l0 = 0 and the term f (µ, α, Γ) is expressed as
−1
f (µ, α, Γ) = e−µΓρ (Γ + 1)
eα/(Γ+1) − 1
,
eα − 1
(12)
and the relation between λopt and λM is given by (5).
Proof. See Appendix B.
Corollary 1. For the case of no retransmissions, the outage probability can be derived as
λopt
,Γ .
PFail (1) = 1 − f µ,
B
(13)
In the next section, we investigate throughput scaling laws for different SNR regimes exploiting
the outage probability given in (8). We only focus on the constant SNR case, for which the analytical derivations are more tractable than the Rayleigh fading case. However, we do not present the
technical details of the Rayleigh fading results since they complicate the analysis without providing
any additional insights. We discussed in detail how to derive scaling results for the Rayleigh fading
case exploiting the constant SNR results in [1]. In Sect. VI, we illustrate scaling results both for the
constant SNR and Rayleigh fading cases and provide numerical comparisons.
IV. O PTIMAL D ESIGNS
FOR
H IGH
AND
L OW SNR
The goal of this section is to determine what the optimal choices of the number of bins B and
the number of retransmissions M, and the corresponding maximum average throughput λopt . For the
proposed random access setting, we solve the throughput optimization problem in (6) to maximize λ
with respect to the deadline constraint T for high and low SNR regimes, since this is not tractable for
12
general SNR settings. The total number of resources N = T W is evenly split into M retransmissions
and B bins.
Let Bopt and Mopt denote the optimal number of bins and the optimal number of retransmissions,
respectively, and λopt be the maximum allowed arrival rate for a fixed deadline constraint T . We
investigate how Bopt , Mopt and λopt scale with the number of resource symbols N at high and low
SNR. Resource allocation is significantly different for these regimes. At one extreme, the combined
SINR at the BS can be made high by splitting the resources into many resource bins and at the same
time satisfying the outage and capacity constraints in (6). At another extreme, however, the combined
SINR will be very low no matter how the resources are split, and the resources have to be shared in
order not to violate the constraints of (6).
A. High SNR Regime
At high SNR, resources can be utilized more efficiently because we have more freedom in selecting
B and M in solving the optimization formulation in (6). From (8), PFail (M) can be made arbitrarily
∗
small by increasing B. We can also observe from (9) that given 1 ≤ km ≤ km
(Km−1 ) arrivals choose a
∗
given frequency bin, each having SNR ρ, the value of km
(Km−1 ) for m ∈ M can be made arbitrarily
small by decreasing the block length n, i.e. increasing B and/or increasing M. However, B or M may
not be increased arbitrarily because the capacity constraint of (6) may not be satisfied for small n.
∗
We first study the SNR threshold ρT for which ρ ≥ ρT such that km
(Km−1 ) = 1 during transmission
m ∈ M, i.e. no other interferer is allowed in the same frequency bin. Using (9) we have
∗
km
(Km−1 )
= 1 = max
nj
m−1
nj m2 m
k o
m2 m X k o
m+
−
−
ki , 1 = max
−
+ 1 ,1 ,
Γ
ρ
Γ
ρ
i=1
L
where Γ = 2 n − 1. Hence, for ρT =
Γ
,
m
∗
the maximum number of devices per bin is km
(Km−1 ) = 1.
In this case, devices experience no interference and the outage constraint is eliminated from (6).
We next aim to provide an upper bound on the number of frequency bins. The following condition
is required in the IBL regime for a typical device to successfully transmit L bits, given Km :
min {C(SINR(Km ))} ≥
m∈M
N
L
, n≤
.
n
MB
(14)
13
At high SNR, we have km = 1 for m ∈ M, and from (2) the Chase combiner output SINR linearly
scales with M. Hence, from (14) we require C(SINR(Km )) = log2 (1 + ρm) ≥
L
L
,
n
for all m ∈ M.
Equivalently, we need to satisfy ρ ≥ Γ = 2 n − 1 where n ≤ N/(MB). This yields the following
upper bound on B:
B≤
N
log2 (1 + ρ).
ML
(15)
The outage constraint is satisfied by choosing B sufficiently large so that there is no other interferer
in the same resource bin. In order to maximize λ, the number of bins B should be maximized given
the resource constraints.
Proposition 4. Given δ, L, ρ, N, the optimal number of bins, Bopt , for the high SNR regime equals
Bopt =
N
log2 (1 + ρ),
Mopt L
(16)
and the maximum average user arrival rate per frame is λMopt = αMopt (δ)Bopt , where αM (δ) satisfies
1
1
1
αM (δ) = −W (δ M − 1)eδ M −1 + δ M − 1,
(17)
where W is the Lambert function and Mopt is given as
Mopt = arg max Bopt αM (δ)(1 − δ 1/M ).
(18)
M ≥1
Proof. See Appendix C.
At high SNR, from Prop. 4, we can infer that the optimal number of bins Bopt ≥ 1 linearly
scales with the number of resource symbols N. Therefore, resources should be split into as many
frequency bins as possible such that the devices do not experience any interference. The value of
the optimal number of retransmission attempts Mopt ≥ 1 however, depends on the desired outage δ,
should decrease as the outage tolerance decreases, and can be determined from (18) as a function
of the outage constraint δ. In this regime, the maximum average arrival rate scales with the optimal
number of bins Bopt and hence with N.
14
B. Low SNR Regime
At low SNR, the Chase combiner output SINR will also be low and it might not be possible to
improve the output SINR via Chase combining. Due to low SINR, the capacity constraint of the
optimization formulation in (6) may only be satisfied by choosing a sufficiently large block length n.
In that case, Bopt = Mopt = 1. In addition to that, due to the interference we always have SINR ≤ ρ.
2L
Given the total number of resources N, when SNR ρ is such that ρ < 2 N −1, using the rate constraint
in (6)
N
L
log2 (1 + SINR) ≥ MB, we have Bopt = Mopt = 1. Therefore, the resource utilization will
be mainly constrained by the capacity requirement. The maximum average rate of device arrivals can
be computed using the outage probability in (8).
Given a target SINR outage rate δ per device, we evaluate the outage probability for Bopt = Mopt = 1
∗
using (8) and km
(Km−1 ) using (9). Then, we want to determine the optimal value of λ that satisfies
λ X
D k,
δ = PFail (1) =
D k, λ ,
=
Bopt
k>k ∗
k>k ∗
X
where λ is increasing in k ∗ . For sufficiently large values of λ, (λ > 1000), the normal distribution with
√
mean λ and standard deviation λ is an excellent approximation to the Poisson distribution. Using
the Gaussian approximation for the Poisson arrivals, we can approximate the maximum device arrival
rate λopt using
δ ≈ Q (k ∗ − λopt )
where Q(x) =
L
√1
2π
R∞
x
p
λopt ,
(19)
exp(−u2 /2)du is the tail probability of the standard normal distribution, and
k
p
1
− + 1 ≈ Q−1 (δ) λopt + λopt ,
k =
Γ ρ
j1
∗
(20)
where Γ = 2 N − 1. Thus, at low SNR, if ρ is in a range such that
−1
−1
1
1
∗
∗
1+ −k
−k
≤ρ<
,
Γ
Γ
it yields Bopt = 1 and λopt that satisfies (20). Using (19), the maximum average arrival rate that can
be supported at low SNR can be determined using Prop. 5.
Proposition 5. At low SNR, Bopt = 1, Mopt = 1, and the maximum arrival rate λopt can be
15
Low SNR/Constant SNR
q
2
−1
−1
2
λopt ≈
k1∗ + (Q 4(δ)) − Q 2 (δ)
Mopt = 1∗ , Bopt = 1
k1∗ ≈ 1 + L1 − ρ1
2 N −1
High SNR/Constant SNR
1/M
opt
λopt = λMopt 1−δ1−δ
Mopt = arg max λM (1 − δ 1/M )
M ≥1
Bopt =
N
Mopt L
log2 (1 + ρ)
Low SNR/Rayleigh fading
, Γopt }
λopt = {λopt |δ = 1 − f µ, Bλopt
opt
Mopt = 1∗ , Bopt = 1
Γopt = 2
Mopt L
N
−1
High SNR/Rayleigh fading
i
h
MP
opt −1
PFail (m)
λopt = λMopt / 1 +
m=1
Mopt = arg max
M ≥1
Bopt ≈
N
Mopt L
1+
MP
−1
λM
m!(µΓopt ρ−1 )m
m=1
log2 (1 + Γopt )
Γopt = −Mopt log(1 − δ)ρ
TABLE I: Scaling results for different SNR regimes for constant SNR and with Rayleigh fading [1] in the IBL regime.
approximated by
λopt ≈
2
q
k ∗ + (Q−1 (δ))2 /4 − Q−1 (δ)/2 .
(21)
Proof. See Appendix D.
At low SNR, from Prop. 5, we can infer that all devices should share the resources in order to
maximize the average arrival rate that can be supported. This is because the combined SINR at the
BS will also be low, and for successful decoding the block length should be made as large as possible
to decrease the threshold Γ. Hence, we require Bopt = Mopt = 1.
We summarize the scaling of parameters λopt , Bopt and Mopt in the IBL regime in Table I.
V. F INITE B LOCK L ENGTHS
In this section, we assume that the encoded block length n symbols is finite, which is the practical
case for short packet sizes. We exploit the finite block length (FBL) model of Polyanskiy et al. [35] to
characterize the performance of the FBL regime by optimizing the number of retransmission attempts
and the required number of bins.
16
The maximum achievable rate with error probability ε, as a function of the block length n, and the
Chase combiner output SINR given in (2), is characterized by [35]
r
0.5 log2 (n)
1
V (SINR) −1
,
Q (ε) +
+o
CFBL (n, SINR, ε) = C(SINR) −
n
n
n
(22)
where V (SINR) is the channel dispersion given by
V (SINR) = 1 −
1
log22 (e).
2
(1 + SINR)
(23)
Note that the maximal rate achievable with error probability ε is closely approximated by (22) for
block lengths n as short as 100 bits [35].
Based on an asymptotic expansion of the maximum achievable rate in (22), we can calculate the
block error rate for the FBL regime given the payload size L bits and encoded block length n symbols,
as
ǫFBL (n, L, SINR) ≈ Q(f (n, L, SINR)),
(24)
which is shown to be valid for packet sizes as small as L = 50 bits [35], where
f (n, L, SINR) =
which is non-decreasing in SINR.
nC(SINR) + 0.5 log2 n − L
p
,
nV (SINR)
(25)
We define GFBL = C − CFBL , which is the gap between the maximum achievable rate for the FBL
and the Shannon capacity. Note from (23) that the channel dispersion V (SINR) increases with SINR.
From (22), we can infer that as the block length n approaches ∞, the gap GFBL decreases to 0. We
next study throughput scaling as function of the delay constraint in the FBL regime.
The throughput – characterized by Poisson arrival rate – optimization problem in the FBL regime
can be written as
λopt = max + λ
B, M ∈Z
s.t.
PFail,FBL (M) ≤ δ,
CFBL (n, SINR(Km ), ε) ≥
(26)
L
TW
, m ∈ M, n ≤
,
n
MB
17
where PFail,FBL (M) is the probability of outage up to and including the M th retransmission attempt
for the FBL regime, and is given by Prop. 6. Since we have a retransmission based model, the block
error rate ε can be larger than the target SINR outage rate δ, and the outage requirement may be met
after Chase combining for general M. When M = 1, the block error rate has to satisfy ε ≤ δ.
Proposition 6. Given B, M, symbol resources N, the outage probability in the FBL regime is given
by
PFail,FBL (M) =
λM
ǫFBL n, L, SINR(Km ) ,
D km ,
B
m∈M
X Y
km ≥1
(27)
where
"
λM = λ 1 +
M
−1
X
m=1
#
PFail,FBL (m) .
(28)
Proof. See Appendix G.
Note that (28) is a fixed point equation in the form of λM = g(λM ), where g(λM ) is determined
by (27). The solution λM is unique because (27) is a monotonically decreasing function of λM .
Before we characterize the FBL regime at low and high SNR, we provide the following Lemma.
Lemma 1. The function f (n, L, X) is monotonically increasing, positive, and concave in X for X > 0.
Proof. See Appendix E.
Proposition 7. Given a random variable X, the function f (n, L, X) satisfies the following relation:
E[Q(f (n, L, X))] ≥ Q(E[f (n, L, X)]) ≥ Q(f (n, L, E[X])).
(29)
Proof. See Appendix F.
Prop. 7 will be used in Sect. V-B to derive bounds for the probability of outage for the FBL regime.
The goal of the rest of this section is to figure out what the optimal choices of the number of bins
B and the number of retransmissions M should be, and what this gives for the maximum average
arrival rate λ for the FBL regime. For the proposed random access setting, we solve the throughput
optimization problem in (26) to maximize λ with respect to the deadline constraint T for high and low
18
SNR regimes, since this is not tractable for general SNR settings as encountered in the IBL regime.
The total number of resources N = T W is evenly split into M retransmissions and B bins.
FBL
FBL
Let Bopt
and Mopt
denote the optimal number of bins and the optimal number of retransmissions,
respectively, and λFBL
opt be the maximum allowed arrival rate for a fixed deadline constraint T . We next
FBL
FBL
investigate how Bopt
, Mopt
and λFBL
opt scale with the number of resource symbols N at high and low
SNR for the FBL regime.
A. High SNR Regime
At high SNR, similar to the IBL regime, resource utilization can be optimized by splitting the
FBL
resources, hence using multiple frequency bins, i.e. Bopt
> 1, and Chase combining to aggregate the
FBL
FBL
FBL
received signals across multiple transmissions, i.e. Mopt
> 1. We need to determine Bopt
and Mopt
in this regime by carefully incorporating the constraints of the optimization formulation for the FBL
model in (26).
Proposition 8. At high SNR, for a given device transmission, the probability of outage for the FBL
regime can be approximated as
λ M Y
M
Q(f (n, L, ρm)).
PFail,FBL (M) ≈ D 1,
B
m∈M
(30)
Proof. See Appendix I.
Comparing the FBL regime to the IBL regime, we can observe that there will a decrease in the
throughput of the multiuser channel which is mainly due to the capacity of the FBL regime (see the
second constraint in (26)). For the FBL regime, the achievable rate is lower than the IBL regime when
the SNR at the BS is fixed. Hence, to achieve the same rate, it is required to increase the received
SNR ρ per device at the BS.
B. Low SNR Regime
At low SNR, since the Chase combiner output SINR will be low, the capacity constraint of the
optimization formulation in (26) may only be satisfied by choosing a sufficiently large block length
FBL
FBL
n. Hence, the devices should share the resources such that Bopt
= Mopt
= 1. In addition to that,
19
λFBL
opt
Low SNR/Constant SNR
o
P
= λδ=
D k, Bλ ǫFBL (n, L, SINR(K))
n
k≥1
FBL
FBL
Mopt
= 1∗ , Bopt
=1
High SNR/Constant SNR
FBL −1
Mopt
i
h
P
FBL
FBL
PFail (m)
λopt = λMopt / 1 +
m=1
h
i
M
−1
P
FBL
FBL
Mopt = arg max λM / 1 +
PFail (m)
M ≥1
n
FBL
Bopt
= B
max
B|n=
N
FBL
BMopt
m=1
FBL
Mopt
T
m=1
L
n
o
FBL
N
≤ CFBL mρ, Bm
, δ 1/Mopt
TABLE II: Scaling results for constant SNR in the FBL regime.
due to the interference we have SINR ≤ ρ. Given the total number of resources N, when SNR ρ is
2L
such that ρ < 2 N − 1 (see Sect. IV-B), since the rate constraint in (26) is stricter than the constraint
N
L
FBL
FBL
log2 (1 + SINR) ≥ MB of (6), it is guaranteed that we have Bopt
= Mopt
= 1.
Given N, Prop. 9 gives an approximation for the outage probability for the FBL regime at high
SNR up to a maximum of M transmissions.
Proposition 9. At low SNR, for a given device transmission, the probability of outage for the FBL
model can be approximated as
PFail,FBL (1) ≈ EK Q(f (n, L, SINR(K))) ,
(31)
where K is a random variable that denotes the number of arrivals on the 1st attempt given that there
is at least one arrival. Its distribution is given by (4).
Proof. At low SNR, for the FBL regime, the outage probability can be derived as
PFail,FBL (1) =
∞
X
λ
D k,
ǫFBL (n, L, SINR(K)) = EK ǫFBL (n, L, SINR(K)) ,
B
k=1
where the final result follows from the approximation in (24).
20
We next provide a lower and upper bound on the probability of outage at low SNR for FBL.
Proposition 10. At low SNR, for a given device transmission, the probability of outage for the FBL
model can be lower and upper bounded as follows
PFail,FBL (1) ≥ Q(f (n, L, EK [SINR(K)])),
∞
X
λ
exp (−f (n, L, SINR(K))2 ).
PFail,FBL (1) ≤
D k,
B
k=1
(32)
Proof. See Appendix H.
At low SNR, similar to the IBL regime, all devices should share the resources in order to maximize
the average number of arrivals that can be supported. We illustrate the scaling of parameters λFBL
opt ,
FBL
FBL
Bopt
and Mopt
for constant SNR at high and low SNR in the FBL regime in Table II. These results
can be extended to characterize the scaling of throughput versus latency for Rayleigh fading case,
which is left as future work.
VI. P ERFORMANCE C OMPARISONS
We investigate the scaling of the throughput λopt in (6), with respect to W , ρ and T for constant SNR
and Rayleigh fading cases. In the plots, solid and dashed curves denote the analytical approximations
for the scaling results, while simulation results are squares.
Throughput scaling with respect to bandwidth W . At high SNR, there is a linear relationship
between λopt and W because from Prop. 4, λopt /B is fixed for given δ and B linearly scales with W .
Given the fixed parameters L, T , and δ, the slope λopt /W for the high SNR regime as a function of
ρ can be found using Table I, which yields
λopt (a)
T 1 − δ 1/Mopt
0.332 · ρ (dB),
≈ αMopt (δ)
W
Mopt L
1−δ
where (a) follows from the high SNR approximation.
For the low SNR regime, we approximate λopt using (21). Since we have Bopt = 1, and using
2x − 1 ≈ x log(2), we can observe that k ∗ = 1 + 1/(2L/N − 1) − 1/ρ ≈ ⌊1 + T W/(L log(2)) − 1/ρ⌋,
hence λopt scales sublinearly in W , and k ∗ changes when W > L log(2)/T , implying that there is
a minimum W amount that yields a change in k ∗ hence λopt due to the granularity of k ∗ , and as T
21
10 4 L=100 bits, =0.1,
W=10 KHz
=20
1500
Simulation
High SNR scaling
Low SNR scaling
(# arrivals/T)
(# arrivals/T)
2000
=10
=0
1000
=-10
10 3
T=10 sec
10 2
T=1 sec
T=0.1 sec
10 1
500
L=100 bits, T=0.1 s, =0.1
0
0
2
4
6
W (Bandwidth, KHz)
8
10
T=0.01 sec
0
10
-10
0
10
10 5
Fig. 2: λopt versus bandwidth W (kHz) for different
ρ (dB) with constant SNR, in the IBL regime.
Simulation
Scaling
20
30
40
(SNR)
Fig. 3: λopt versus SNR ρ (dB) for different delay constraints
T (s) with constant SNR, in the IBL regime.
becomes smaller or L becomes larger, a linear scaling between λopt and W is not possible to achieve.
p
2
From (21) of Prop. 5, we have λopt ≈
k ∗ + (Q−1 (δ))2 /4 − Q−1 (δ)/2 . At low SNR, when δ
is sufficiently large (i.e. Q−1 (δ) is very small), λopt ≈ k ∗ . The trend of λopt with respect to W is
illustrated in Fig. 2.
Throughput scaling with respect to SNR. For high SNR, the capacity formula can be approximated
as C(SINR) ≈ 0.332 · SINR (dB), where SINR (dB) = 10 log10 (SINR). Using this approximation, and
1/M
)
. Thus, we have
the expression for λopt in Table I, we have λopt ≈ αM (δ)3.32 MNL log10 (ρ) (1−δ
(1−δ)
λopt
10 log10 (ρ)
1/M
T
≈ αM (δ) W
0.332 1−δ
. For different latency constraints ranging between Tmin = 10 msec
ML
1−δ
and Tmax = 10 sec, we illustrate the trend of λopt with respect to ρ in Fig. 3.
Throughput scaling with respect to outage constraint δ. From (6), we can observe the tradeoff
between the throughput and delay such that T is lower bounded by T ≥
LBM
.
W log2 (1+SINR(KM ))
Note that
λopt increases in δ, i.e. as the outage constraint becomes loose. Large δ implies a rate constrained system
and there is a lot of devices per bin. Small δ implies an SINR constrained model with fewer number
of devices per bin. In Fig. 4, we illustrate the trend of λopt with respect to δ for ρ = [−10, 0, 10, 20]
dB. At low SNR, Bopt = Mopt = 1, and k ∗ increases with λopt . However, it is constrained by the
outage requirement in (6) (see Table I), using which λopt can be obtained. For the high SNR regime,
Bopt , Mopt ≥ 1, and k ∗ = 1, λopt changes linearly in B, and it can be computed using the rate constraint
in (6).
22
100
L=100 bits,
50 T=0.1 s,
W=10 KHz
40
Simulation
High SNR scaling
Low SNR scaling
=[-10, 0, 10, 20]
30
20
Simulation
High SNR scaling
Low SNR scaling
60
=[-10, 0, 10, 20]
40
20
10
0
0.05
=0.1,
T=0.1 s,
W=10 KHz
80
(# arrivals/T)
(# arrivals/T)
60
0
0.1
0.15
0
100
Outage,
200
300
L (bits)
Fig. 4: λopt versus outage constraint δ for different ρ (dB).
Fig. 5: λopt versus payload L (bits) for different ρ (dB).
Throughput scaling with respect to packet size L. In Fig. 5, we illustrate the trend with respect
to L for ρ = [−10, 0, 10, 20] dB, and for fixed δ = 0.1, deadline T = 0.1 sec, and bandwidth W = 10
KHz. It is intuitive that the maximum average supportable rate λopt decays in payload size L.
At low SNR, since Γ increases in L, to satisfy a given outage constraint in (6), λopt decays in L.
Similarly, at high SNR, from Table I, Bopt is inversely proportional to L and from
λopt ≈ αMopt (δ)Bopt (1 − δ 1/Mopt ) (1 − δ),
which justifies the inversely proportional relation between λopt and L. We can also observe that Bopt
and Mopt decrease in L due to the capacity constraint in (6).
Throughput scaling with respect to delay constraint T . We consider the SNR range ρ = [−10, 20]
dB. At high SNR, the trend of λopt versus T (or N) is similar to trend between λopt and W . For the
low SNR regime, we can observe that k ∗ changes when T is a multiple of L log(2)/W . Hence, it
can be approximated as linear for sufficiently large W . The variation of λopt with T is shown in Fig.
6-(a). In the same figure, for a few points along the curves, the optimal values of B and M are also
illustrated.
Although the variability of the channel causes a drop in the number of arrivals that can successfully
complete the random access phase, we showed that similar conclusions in the constant SNR case
extend to the case of Rayleigh fading in Fig. 6-(b).
23
L=100 bits
30 W=10 KHz
=0.1
25
Simulation
High SNR scaling
Low SNR scaling
B
20
opt
=25, M
opt
(# arrivals/T)
(# arrivals/T)
35
=2
=20
15
B
10
opt
=8, M
opt
=10
=2
B
opt
=1, M
opt
=1
=0
5
5 L=100 bits
W=10 KHz
4 =0.1
=40
Simulation
High SNR scaling
Low SNR scaling
=30
3
=20
2
=10
1
=0
=-10
0
0
0
0.02
0.04
0.06
0.08
0.1
=-10
0
0.02
0.04
0.06
0.08
0.1
T (Delay, s)
T (Delay, s)
(a) Constant SNR.
(b) Rayleigh fading with µ = 1 [1].
Fig. 6: λopt versus delay constraint T (s) for different ρ (dB) in the IBL regime.
(# arrivals/T)
15
L=100 bits
W=10 KHz
=0.1
10
Simulation
High SNR scaling
Low SNR scaling
=20
=10
5
=0
=-10
0
0
0.02
0.04
0.06
0.08
0.1
T (Delay, s)
Fig. 7: λopt versus delay constraint T (s) at constant SNR for different ρ (dB) in the FBL regime.
Finally, we consider the λopt versus T trend of the FBL, which is illustrated in Fig. 7 at constant
SNR, and compare it to the IBL case. For a given L, T , and W , the IBL throughput is an upper
bound to the FBL throughput. This is mainly due to the stricter capacity constraint of the FBL model
as given in (26). Note from (23) that V (SINR) increases with SINR. Therefore, from (22), we can
infer that this gap GFBL increases as ρ increases for a given block length n and block error rate ε.
24
VII. C ONCLUSIONS
We proposed a random access model for a single cell uplink where the objective is to characterize
the fundamental limits of supported arrival rate by optimizing the number of retransmissions and
resource allocation under a maximum latency constraint. We evaluate the performance of the model
with respect to total bandwidth, latency constraint, payload, outage constraint, and SNR. We obtain
the following design insights for the infinite block length (IBL) and finite block length (FBL) models
for different SNR, with different packet sizes and outage constraints:
•
Independent of the payload size L, at low SNR, the resources should be shared, implying small
Bopt and Mopt . As SNR increases, the resources should be split, yielding large Bopt and Mopt .
•
As the outage constraint δ decreases, Bopt and Mopt increase.
The insights can be applied to the design of narrowband IoT systems for 5G with ultra-reliability,
ultra-low latency, and a large number of connected devices. Possible extensions include the modeling
and analysis of the random access in a cellular network setting, by capturing the scaling of throughput
using realistic power control and fading models, and characterizing the uplink interference. They
also include the analysis of heterogeneous random access strategies to accommodate a wide range of
device-chosen strategies with different capabilities and requirements. The overhead caused by channelestimation is significant and different diversity combining or multiplexing techniques can improve the
system performance, especially when there is deep fading. Another direction would be to incorporate
the FBL model in a cellular network setting, and characterize the finite block-error probability by
capturing the effect of the uplink interference at low and high contention.
A PPENDIX
A. Proof of Proposition 2
The probability of outage given up to a maximum of M transmission attempts is given as
PFail (M) =
X
X
k1 >k1∗ k2 >k2∗ (k1 )
···
X
∗ >k ∗ (K
kM
M −1 )
M
λM
D km ,
,
B
m∈M
Y
∗
where km
is a function of Km−1 = {k1 , . . . , km−1 } for 2 ≤ m ≤ M to be calculated based on the rate
requirement. Note that PFail (m) monotonically decreases in m.
25
The aggregate arrival rate λM with up to M total transmissions is given by
XM −1
PFail (m) .
λM = λopt 1 +
m=1
For successful transmission, we require that
L
n
∗
≤ C(SINR(Km )) for 1 ≤ km ≤ km
(Km−1 ) for m ∈ M.
Using (2), we derive
m2 m
ki ≤ m +
, m ∈ M,
−
i=1
Γ
ρ
Xm
L
where Γ = 2 n − 1, yielding the lower limit of the summation in (8), as given by (9).
B. Proof of Proposition 3
The outage probability given up to a maximum of M transmissions with Rayleigh fading with mean
1/µ equals
i
h
L
PFail (M) = E P max {C(SINR(Km ))} < KM
m∈M
n
m
ii
h
h
X
Γ
(a)
m
+
I
I
,
K
= EKM EIKM P h <
k
K
M
i
M
ρm2
i=1
m
P
h YM
h
ii
− µΓ2 m+
I ki
(b)
i=1
1 − e ρm
= EKM EIKM
IK M , K M
m=1
i
h
i
h
−1
−1
−2µΓρ−1
−1
−µΓρ−1
EK2 LIk1 (µΓρ )LIk2 (µΓρ ) K2
=1−e
Ek1 LIk1 (µΓρ ) k1 − e
i
h
−1
+ e−3µΓρ EK2 LIk1 (2µΓρ−1 )LIk2 (µΓρ−1 ) K2 + . . .
(c)
=
1 lX
1 +1
X
l1 =0 l2 =l1
lM −1 +1
X
...
(−1)
lM
lM =lM −1
Y λM
f µ,
, ln Γ ,
B
n∈M
(33)
where we used the shorthand notation for interference as IKm := {Ik1 , . . . , Ikm }, and (a) follows
from the definition of SINR in (10) and (b) from that h ∼ exp(µ), and (c) from the definition of
D k, λBM , independence of Iki ’s, and LIki (s) = LIk1 (s) for i ∈ M. Using the Laplace transform of
the interference given k arrivals, which is denoted by LIk (s) and expressed as
−sIk
LIk (s) = E[e
"
−sρ
]=E e
k−1
P
i=1
gim
#
=E
"k−1
Y
i=1
−sρgi
e
#
µ k−1
.
=
µ + sρ
26
Averaging LIk (s) using the distribution D(k, λB ), we obtain
LICI (s) = Ek [LIk (s)] =
∞
X
k=1
−1
eλB /(sµ ρ+1) − 1
µ k−1
.
= (sµ−1 ρ + 1)
D(k, λB )
µ + sρ
eλB − 1
−1
To simplify the notation, we let s = µΓρ−1 and define f (µ, α, Γ) = e−µΓρ (Γ + 1) e
α/(Γ+1) −1
eα −1
(34)
, using
which the final expression (c) is evaluated.
C. Proof of Proposition 4
At high SNR regime, since k ∗ = 1, the SINR becomes SINR = ρ. Since the total number of
resources N is split into M retransmissions and B bins to successfully transmit L bits, in the IBL
regime, we have B ≈
N
ML
log2 (1 + ρ).
For the high SNR regime, the relation between the target outage δ and αM (δ) = λM /B is given as
PFail (M) = δ = 1 −
M
αM (δ) M
= PFail (m) m , m ∈ M,
α
(δ)
M
e
−1
where αM (δ) = α1 (δ 1/M ) = α(δ 1/M ). The relation between λ and the aggregate arrival rate is
M
−1
i
.h
X
m
M
= λM (1 − δ 1/M )/(1 − δ).
λ = λM 1 +
δ
m=1
The optimal number of transmissions can be found solving
Mopt = arg max λM (1 − δ 1/M ),
M ≥1
which depends on δ at high SNR, and B ≈
N
ML
log2 (1 + ρ).
D. Proof of Proposition 5
We have B ≤
as kmax → ∞,
λ
B
λ
B
N
L
log2 (1 + SINRk ), and k ∗ ≤ 1 +
1
LB
2 N
−1
− 1ρ , and δ =
P∞
k>k ∗
D k, Bλ . At low SNR,
→ ∞. However, we also require B = 1 because we cannot achieve arbitrarily high
picking B arbitrarily small since B ∈ Z+ .
The proof follows from approximating D k, Bλ by a Gaussian distribution that has the same mean
∗
and variance as the Poisson distribution, i.e. N Bλ , Bλ for high Bλ . Hence, δ ≈ Q k√−λ/B . Therefore,
λ/B
q
q
−1
−1
2
(Q
(δ))
Q
(δ)
λ
λ
∗
≈ k∗ +
− 2 . Thus, we can infer that B scales linearly with k . However, since B ≤
B
4
27
N
L
log2 1 +
ρ
1+ρ(k ∗ −1)
, the number of bins, B, decays sub-linearly with k ∗ . Therefore, their product
λ increases as k ∗ increases. The above approximation with B = 1 gives the maximum supportable
rate λopt for the low SNR regime.
E. Proof of Lemma 1
We rewrite f (n, L, X) as f (n, L, X) = √C(X)+b , and let a =
V (X)/n
√
n
,
log2 e
b=
0.5 log2 n−L
,
n
f1 (x) = √(1+x)
x(x+2)
and f2 (x) = log2 (1 + x). Then, we can express f (n, L, x) as f (n, L, x) = af1 (x)[f2 (x) + b]. Then,
we can derive the identities
∂2
f (x)
∂x2 2
∂
f (x)
∂x 1
1
= − (x(x+2))
3/2 ,
∂2
f (x)
∂x2 1
=
3(1+x)
, ∂ f (x)
(x(x+2))5/2 ∂x 2
1
1
= − log(2)
. We also assume that b ≤ 0. From b =
(1+x)2
0.5 log2 n−L
n
=
1
1
log(2) 1+x
and
≤ 0 that is true when
n ≤ 22L , which is a realistic assumption for the FBL regime.
Monotonically increasing. To prove that f (n, L, x) is increasing in x, we take its partial derivative:
∂
∂
∂f (n, L, x)
=a
f1 (x)[f2 (x) + b] + f1 (x) f2 (x)
∂x
∂x
∂x
√
log2 (1 + x) + b (a)
1
= n 2
1−
> 0, x ≥ 0.
(x + 2x)1/2
log2 (e)(x2 + 2x)
where (a) follows from that x ≥ 0 and 1 −
log(1+x)
x2 +2x
(35)
> 0 and the assumption that b ≤ 0.
Note that f1 (x) and f2 (x) are increasing in x ≥ 0. Hence, f (n, L, x) is also increasing in x. Let
x ≤ y. Then, f (n, L, x) = af1 (x)[f2 (x) + b] ≤ af1 (y)[f2 (y) + b] = f (n, L, y). Hence, f (n, L, x) is a
monotonically increasing function of x.
Positive. From definition of CFBL , it is required to satisfy
r
V (SINR) −1
L
0.5 log2 (n)
1
CFBL (n, SINR, ε) = C(SINR) −
≥ ,
Q (ε) +
+o
n
n
n
n
(36)
which yields the condition that f (n, L, SINR) = √C(X)+b ≥ Q−1 (ε). Note that since Q−1 (ε) > 0 as
V (X)/n
long as ε < 0.5. This condition guaranties the positivity of f (n, L, x).
Concave. To prove that f (n, L, x) is concave, the second partial derivative of f (n, L, x) with respect
to x is derived as
2
∂
∂
∂2
∂
∂2
f (n, L, x) = a
f1 (x)[f2 (x) + b] + 2 f1 (x) f2 (x) + f1 (x) 2 f2 (x)
∂x2
∂x2
∂x
∂x
∂x
x+1
3[log2 (1 + x) + b]
1
(a) √
−1−
< 0, x ≥ 0,
= n 2
3/2
2
(x + 2x)
log2 (e)(x + 2x)
(1 + x)2
28
where (a) follows from the fact that
3 log2 (1+x)
log2 (e)(x2 +2x)
−1−
1
(1+x)2
< 0 for x ≥ 0 and from b ≤ 0.
F. Proof of Proposition 7
Since the second derivative of Q(x) is always greater than or equal to zero, Q(x) is convex and
increasing on x ≥ 0. This implies that for a random variable X, we have that
E[Q(X)] ≥ Q(E[X]).
Note also that from Lemma 1, f (n, L, X) is monotonically increasing, positive, and concave for
X ≥ 0 and Q(f (n, L, X)) is decreasing in X. Then, E[f (n, L, X)] ≤ f (n, L, E[X]). Note also that
Q(X) is decreasing, monotone, and convex for X > 0. Thus, E[Q(X)] ≥ Q(E[X]). Then,
(a)
(b)
E[Q(f (n, L, X))] ≥ Q(E[f (n, L, X)]) ≥ Q(f (n, L, E[X])),
(37)
where (a) follows from the convexity of Q and (b) follows from that f (n, L, X) is concave and Q is
decreasing in X.
G. Proof of Proposition 6
The outage probability for the FBL model at constant SNR with M retransmissions equals
L
PFail (M) = P max {CFBL (n, SINRm , ε)} <
KM
m∈M
n
h
i
L
= E P max {CFBL (n, SINR(Km ), ε)} <
KM
m∈M
n
i
h nC (n, SINR(K ), ε) − L
(a)
FBL
m
p
< 0, m ∈ M KM
=E P
nV (SINR(Km ))
i
h
(b)
= E P Z < −f (n, L, SINR(Km )), m ∈ M KM
X Y
(c) X
λM
···
D km ,
≈
f (n, L, SINR(Km ))
B
km ≥1 m∈M
k1 ≥1
X Y
(d) X
λM
ǫFBL (n, L, SINR(Km )),
≈
···
D km ,
B
k ≥1
k ≥1 m∈M
1
m
where (a) follows from the definition of CFBL using (22) and (23), (b) from the normal distribution
approximation such that Z is standard normal random variable N (0, 1), and (c) from the approximation
based on standard Gaussian ccdf Q(x) and (d) from Q(x) = 1 − Q(−x) and (24).
29
H. Proof of Proposition 10
At low SNR, for the FBL regime, the outage probability can be bounded as
(a)
PFail,FBL (1) ≥ Q(f (n, L, EK [SINR(K)])),
(38)
where (a) follows from (29) and (31), and the mean SINR is given by
∞
X
λ
ρ
D k,
B 1 + ρ(k − 1)
k=1
λ 1−ρ−1
λ i
1
1 h
−1
−1
1+ −
Γ
(ρ
)
−
Γ
ρ
,
−
.
= −1
o
o
ρ −1
B
eλ/B − 1
B
EK [SINR(K)] =
Using the Chernoff bound of the Q-function, which is Q(x) ≤ e−x
2 /2
for x > 0, we can obtain an
upper bound for the probability of outage as follows:
1
2
PFail,FBL (1) ≤ EK exp − f (n, L, SINR(K))
2
∞
X
λ
exp (−f (n, L, SINR(K))2 ).
D k,
=
B
k=1
I. Proof of Proposition 8
At high SNR, the resource symbols have to be shared such that km = 1, for all m ∈ M, and hence,
from (27) we have
PFail,FBL (M) =
λ
λ M Y
M
M
D 1,
ǫFBL (n, L, ρm) ≈ D 1,
Q(f (n, L, ρm)).
B
B
m∈M
m∈M
Y
(39)
R EFERENCES
[1] D. Malak, H. C. Huang, and J. G. Andrews, “Fundamental limits of random access communication with retransmissions,” in Proc.,
IEEE ICC, May 2017.
[2] G. Durisi, T. Koch, J. Östman, Y. Polyanskiy, and W. Yang, “Short-packet communications over multiple-antenna Rayleigh-fading
channels,” IEEE Trans. Commun., vol. 64, no. 2, pp. 618–629, Feb. 2016.
[3] H. S. Dhillon, H. Huang, and H. Viswanathan, “Wide-area wireless communication challenges for the Internet of Things,” IEEE
Communications Magazine, vol. 55, no. 2, pp. 168–174, Feb. 2017.
[4] A. Laya, L. Alonso, and J. Alonso-Zarate, “Is the random access channel of LTE and LTE-A suitable for M2M communications?
A survey of alternatives.” IEEE Commun. Surveys Tuts., vol. 16, no. 1, pp. 4–16, Jan. 2014.
[5] N. A. Johansson, Y.-P. E. Wang, E. Eriksson, and M. Hessler, “Radio access for ultra-reliable and low-latency 5G communications,”
in Proc., IEEE Int. Conf. Commun., London, U.K., Jun. 2015.
30
[6] M. Fallgren and B. Timus, “Scenarios, requirements and KPIs for 5G mobile and wireless system,” METIS project, Deliverable
D1.1, Tech. Rep., Apr. 2013.
[7] H. S. Dhillon, H. C. Huang, H. Viswanathan, and R. A. Valenzuela, “Power-efficient system design for cellular-based machineto-machine communications,” IEEE Trans. Wireless Commun., vol. 12, no. 11, pp. 5740–5753, Nov. 2013.
[8] ——, “Fundamentals of throughput maximization with random arrivals for M2M communications,” IEEE Trans. Commun., vol. 62,
no. 11, pp. 4094–4109, Nov. 2014.
[9] G. Durisi, T. Koch, and P. Popovski, “Towards massive, ultra-reliable, and low-latency wireless communication with short packets,”
Proc. IEEE, vol. 104, no. 9, pp. 1711–1726, Sep. 2016.
[10] N. Pratas and P. Popovski, “Underlay of low-rate machine-type D2D links on downlink cellular links,” in Proc., IEEE ICC
Workshops, Jun. 2014, pp. 423–428.
[11] R. Ratasuk, A. Prasad, Z. Li, A. Ghosh, and M. A. Uusitalo, “Recent advancements in M2M communications in 4G networks
and evolution towards 5G,” in Proc., IEEE. Intelligence in Next Generation Networks (ICIN), Feb. 2015, pp. 52–57.
[12] C. Perera, C. H. Liu, S. Jayawardena, and M. Chen, “A survey on internet of things from industrial market perspective,” IEEE
Access, vol. 2, pp. 1660–79, 2014.
[13] M. Hasan, E. Hossain, and D. Niyato, “Random access for machine-to-machine communication in LTE-advanced networks: issues
and approaches,” IEEE Commun. Mag., vol. 51, no. 6, pp. 86–93, Jun. 2013.
[14] G. C. Madueno, C. Stefanovic, and P. Popovski, “Reliable and efficient access for alarm-initiated and regular M2M traffic in ieee
802.11ah systems,” IEEE Internet Things J., vol. 3, no. 5, pp. 673–682, Oct. 2016.
[15] A. Gotsis, A. Lioumpas, and A. Alexiou, “M2M scheduling over LTE: Challenges and new perspectives,” IEEE Veh. Technol.
Mag., vol. 7, no. 3, pp. 34–39, Sep. 2012.
[16] U. Phuyal, A. T. Koc, M. H. Fong, and R. Vannithamby, “Controlling access overload and signaling congestion in M2M networks,”
in Proc., IEEE ASILOMAR, vol. 4, Nov. 2012, pp. 591–595.
[17] G. C. Madueno, C̆edomir Stefanović, and P. Popovski, “Efficient LTE access with collision resolution for massive M2M
communications,” in Proc., IEEE Globecom Workshops, Dec. 2014.
[18] V. Chandrasekhar, J. G. Andrews, Z. Shen, T. Muharemovic, and A. Gatherer, “Power control in two-tier femtocell networks,”
IEEE Transactions on Wireless Communications, vol. 8, no. 8, pp. 4316–28, Aug. 2009.
[19] V. Chandrasekhar and J. G. Andrews, “Spectrum allocation in two-tier networks,” IEEE Trans. on Commun., vol. 57, no. 10, pp.
3059–3068, Oct. 2009.
[20] A. Biral, H. Huang, A. Zanella, and M. Zorzi, “Uplink resource allocation in cellular systems: An energy-efficiency perspective,”
in Proc., IEEE Globecom, Dec. 2015.
[21] M. Koseoglu, “Pricing-based Load Control of M2M Traffic for the LTE-A Random Access Channel,” IEEE Trans. Commun.,
vol. 65, no. 3, pp. 1353–65, Mar. 2017.
[22] D. Malak, H. S. Dhillon, and J. G. Andrews, “Optimizing data aggregation for uplink machine-to-machine communication
networks,” IEEE Trans. Commun., vol. 64, no. 3, pp. 1274–1290, Mar. 2016.
[23] M. Koseoglu, “Lower Bounds on the LTE-A Average Random Access Delay Under Massive M2M Arrivals,” IEEE Trans. Commun.,
vol. 64, no. 5, pp. 2104–2115, May 2016.
[24] M. Jaber, N. Kouzayha, Z. Dawy, and A. Kayssi, “On cellular network planning and operation with M2M signalling and security
considerations,” in Proc., IEEE ICC 2014 Workshops, Jun. 2014.
31
[25] N. Kouzayha, M. Jaber, and Z. Dawy, “M2M data aggregation over cellular networks: Signaling-delay tradeoffs,” in Proc., IEEE
Globecom 2014 Workshops, Dec. 2014.
[26] Y. Zhong, M. Haenggi, T. Q. Quek, and W. Zhang, “On the stability of static Poisson networks under random access,” IEEE Trans.
Commun., vol. 64, no. 7, pp. 2985–2998, Jul. 2016.
[27] W. Yu, “On the fundamental limits of massive connectivity,” in Proc., IEEE ITA Workshop, Feb. 2017.
[28] P. Larsson, J. Gross, H. Al-Zubaidy, L. K. Rasmussen, and M. Skoglund, “Effective capacity of retransmission schemes: A
recurrence relation approach,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4817–35, Nov. 2016.
[29] Y. Li, M. C. Gursoy, and S. Velipasalar, “On the throughput of hybrid-ARQ under statistical queuing constraints,” IEEE Trans.
Veh. Technol., vol. 64, no. 6, pp. 2725–32, Jun. 2015.
[30] A. Kumar, A. Abdelhadi, and C. Clancy, “A delay-optimal packet scheduler for M2M uplink,” in Proc., IEEE MILCOM, Nov.
2016, pp. 295–300.
[31] 3GPP, “Cellular system support for ultra low complexity and low throughput Internet of Things, (Release 13), 3GPP Technical
Specification Group GSM/EDGE Radio Access Network,” Jun. 2015.
[32] Y. P. E. Wang, X. Lin, A. Adhikary, A. Grovlen, Y. Sui, Y. Blankenship, J. Bergman, and H. S. Razaghi, “A primer on 3GPP
narrowband internet of things,” IEEE Commun. Mag., vol. 55, no. 3, pp. 117–123, Mar. 2017.
[33] B. Makki, T. Svensson, and M. Zorzi, “Finite block-length analysis of spectrum sharing networks: Interference-constrained
scenario,” IEEE Wireless Commun. Letters, vol. 4, no. 4, pp. 433–436, Aug. 2015.
[34] ——, “Finite block-length analysis of spectrum sharing networks using rate adaptation,” IEEE Trans. Commun., vol. 63, no. 8,
pp. 2823–2835, Aug. 2015.
[35] Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 56,
no. 5, pp. 2307–2359, May 2010.
[36] W. Yang, G. Caire, G. Durisi, and Y. Polyanskiy, “Optimum power control at finite blocklength,” IEEE Trans. Inf. Theory, vol. 61,
no. 9, pp. 4598–615, Sep. 2015.
[37] G. Liva, “Graph-based analysis and optimization of contention resolution diversity slotted ALOHA,” IEEE Trans. Commun.,
vol. 59, no. 2, pp. 477–487, Feb. 2011.
[38] K. Dovelos, L. Toni, and P. Frossard, “Finite length performance of random slotted ALOHA strategies,” arXiv preprint
arXiv:1607.02385, Jul. 2016.
[39] E. Dahlman, S. Parkvall, J. Sköld, and P. Beming, 3G Evolution – HSPA and LTE for Mobile Broadband (2 ed.). Academic Press.
Elsevier Science Publishing Co Inc, 2008.
[40] N. Abramson, “The ALOHA system - another alternative for computer communications,” in Proc, Fall AFIPS Computer Conference,
Nov. 1970, pp. 281–285.
| 7cs.IT
|
arXiv:1510.09115v1 [cs.MA] 30 Oct 2015
Continuous Time Gathering of
Agents with Limited Visibility
and Bearing-Only Sensing
Levi-Itzhak Bellaiche and Alfred Bruckstein
Center for Intelligent Systems - MARS Lab
Autonomous Systems Program
Technion Israel Institute of Technology
32000 Haifa, Israel
Abstract
A group of mobile agents, identical, anonymous, and oblivious (memoryless),
having the capacity to sense only the relative direction (bearing) to neighborhing
agents within a finite visibility range, are shown to gather to a meeting point in
finite time by applying a very simple rule of motion. The agents’ rule of motion
is : set your velocity vector to be the sum of the two unit vectors in R2 pointing
to your “extremal” neighbours determining the smallest visibility disc sector in
which all your visible neighbors reside, provided it spans an angle smaller than
π, otherwise, since you are “surrounded” by visible neighbors, simply stay put
(set your velocity to 0). Of course, the initial constellation of agents must have
a visibility graph that is connected, and provided this we prove that the agents
gather to a common meeting point in finite time, while the distances between
agents that initially see each other monotonically decreases.
1
1
Introduction
This paper studies the problem of mobile agent convergence, or robot gathering under severe limitations on the capabilities of the agent-robots. We assume
that the agents move in the environment (the plane R2 ) according to what they
currently “see”, or sense in their neighborhood. All agents are identical and
indistinguishable (i.e. they are anonimous having no i.d’s) and, all of them are
performing the same “reactive” rule of motion in response to what they see.
Our assumption will be that the agents have a finite visibility range V , a distance beyond which they cannot sense the presence of other agents. The agents
within the “visibility disk” of radius V around each agent are defined as his
neighbors, and we further assume that the agent can only sense the direction
to its neighbors, i.e. it performs a “bearing only” measurement yielding unit
vectors pointing toward its neighbor.
Therefore, in our setting, each agent senses its neighbors within the visibility
disk and sets its motion only according to the distribution of unit vectors pointing to its current neighbors. Figure 1 shows a constellation of agents in the
plane (R2 ), their “visibility graph” and the visibility disks of some of them,
each agent moves based on the set of unit vectors pointing to its neighbors.
Figure 1: A constellation of agents in the plane displaying the “visibility disks”
of agents Ak , Al , Ai , Aj , Ap and the visibility graph that they define, having
edges connecting pairs of agents that can see each other.
2
In this paper we shall prove that continuous time limited visibility sensing of directions only and continuous adjustment of agents’ velocities according to what
they see is enough to ensure the gathering of the agents in finite time to a point
of encounter.
The literature of robotic gathering is vast and the problem was addressed under various assumptions on the sensing and motion capabilities of the agents.
Here we shall only mention papers that deal with gathering assuming continuous time motion and limited visibility sensing, since these are most relevant
to our work reported herein. The paper [1] by Olfati-Saber, Fox, and Murray,
nicely surveys the work on the topic of gathering (also called consensus) for
networked multi agent systems, where the connections between agents are not
necessarily determined by their relative position or distance. This approach to
multi-agent systems was indeed the subject of much investigation and some of
the results, involving “switching connection topologies” are useful in dealing
with constellation-defined visibility-based interaction dynamics too. A lot of
work was invested in the analysis of “potential functions” based multi-agent
dynamics, where agents are sensing each other through a “distance-based” influence field, a prime example here being the very influential work of Gazi and
Passino [2] which analyses beautifully the stability of a clustering process. Interactions involving hard limits on the “visibility distance” in sensing neighbors
were analysed in not too many works. Ji and Eggerstedt in [3] analysed such
problems using potential functions that are “visibility-distance based barrier
functions” and proved connectedness-preservation properties at the expense of
making some agents temporarily “identifiable” and “traceable” via a hysteresis process. Ando, Oasa, Suzuki and Yamashita in [4] were the first to deal
with hard constraints of limited visibility and analysed the “point convergence”
or gathering issue in a discrete-time synchronized setting, assuming agents can
see and measure both distances and bearing to neighbors withing the visibility
range.
Subsequently, in a series of papers, Gordon, Wagner, and Bruckstein, in [5],
[6], [7], analysed gathering with limited visibility and bearing only sensing constraints imposed on the agents. Their work proved gathering or clustering results
in discrete-time settings, and also proposed dynamics for the continuous-time
settings. In the sequel we shall mention the continuous time motion model they
analysed and compare it to our dynamic rule of motion.
In our work, as well as most of the papers mentioned above one assumed that
the agents can directly control their velocity with no acceleration constraints.
We note that the literature of multi-agent systems is replete with papers assuming more complex and realistic dynamics for the agents, like unicycle motions,
second order systems and double integration models relating the location to the
controls, and seek sensor based local control-laws that ensure gathering or the
achievement of some desired configuration. However we feel that it is still worthwhile exploring systems with agents directly controlling their velocity based on
very primitive sensing, in order to test the limits on what can be achieved by
agents with such simple, reactive behaviours.
3
2
The gathering problem
We consider N agents located in the plane (R2 ) whose positions are given by
{Pk = (xk , yk )T }k=1,2,...,N , in some absolute coordinate frame which is unknown
to the agents.We define the vectors
(
uij =
Pj −Pi
kPj −Pi k
0
0 < kPj − Pi k ≤ V
kPj − Pi k = 0 or kPj − Pi k > V
hence uij are, if not zero, the unit vectors from Pi to all Pj ’s which are neighbors
of Pi in the sense of being at a distance less than V from Pi , i.e. Pj ’s obeying
:
kPj − Pi k , [(Pj − Pi )T (Pj − Pi )]1/2 ≤ V
Note that we have uij = −uji , ∀(i, j). For each agent Pi , let us define the special
−
vectors, u+
i and ui (from among the vectors uij defined above). Consider the
nonzero vectors from the set {uij }j=1,2,...,N . Anchor a moving unit vector η̄(θ)
at Pi poiting at some arbitrary neighbor, i.e. at uik 6= 0, η̄(0) = uik and rotate
it clockwise, sweeping a full circle about Pi . As η̄(θ) goes from η(0) to η(2π)
it will encounter all the possible uij ’s and these encounters define a sequence
of angles α1 , α2 , . . . , αr that add to 2π = α1 + . . . + αr (αk = angle from k-th
to (k+1)-th encounter with a uij , αr = angle from last encounter to first one
again, see Figure 2). If none of the angles {α1 , . . . , αr } is bigger than π, set
−
+
−
u+
i = ui = 0. Otherwise define ui = ui(m) and ui = ui(n) the unit vectors
encountered when entering and exiting the angle αb > π bigger than π.
4
Figure 2: Leftmost and rightmost visible agents of agent located at Pi .
+
One might call u−
i the pointer to the “leftmost visible agent” from Pi and ui
the pointer to the “rightmost visible agent” among the neighbors of Pi . If Pi has
nonzero right and leftmost visible agents it means that all its visible neighbors
belong to a disk sector defined by an angle less than π, and Pi will be movable.
Otherwise we call him “surrounded” by neighbors and, in this case, it will stay
in place while it remains surrounded (see Appendix 2 for an alternative way of
defining the leftmost and rightmost agents).
The dynamics of the multi-agent system will be defined as follows.
dPi
−
= v0 (u+
i + ui ) for i = 1, . . . , N
dt
(1)
Note that the speed of each agent is in the span of [0, 2v0 ].
With this we have defined a local, distributed, reactive law of motion based on
the information gathered by each agent. Notice that the agents do not communicate directly, are all identical, and have limited sensing capabilities, yet we
shall show that, under the defined reactive law of motion, in response to what
they can “see” (which is the bearings to their neighboring agents), the agents
will all come together while decreasing the distance between all pairs of visible
agents.
Assume that we are given an initial configuration of N agents placed in the
plane in such a way that their visibility graph is connected. This just means
5
that there is a path (or a chain) of mutually visible neighbors from each agent
to any other agent.
Our first result is that while agents move according to the above described rule
of motion, the visibility graph will only be supplemented with new edges and
old “visibility connections” will never be lost.
2.1
Connectivity is never lost
We shall show that
Theorem 2.1.1. A multi agent systems under the dynamics
−
{Ṗi = v0 (u+
i + ui )}i=1,...,N )
ensures that pairs of neighboring agents at t = 0 (i.e. agents at a distance less
than V ) will remain neighbors forever.
Proof. To prove this result we shall consider the dynamics of distances between
pairs of agents.
We have that the distance ∆ij between Pi and Pj is
∆ij = kPj − Pi k = [(Pj − Pi )T (Pj − Pi )]1/2
hence
or
1
d (t)
∆ij =
(Pj − Pi )T (P˙j − Ṗi )
dt
kPj − Pi k
(t)
d
dt ∆ij
=
uTij (P˙j − Ṗi )
= −uTij Ṗi + uTij P˙j
= −uTij Ṗi − uTji P˙j
But we know that the dynamics (1) is
Ṗi
P˙j
−
= v0 (u+
i + ui )
+
= v0 (uj + u−
j )
Therefore
d
−
+
−
T
∆ij = −v0 uTij (u+
i + ui ) − v0 uji (uj + uj )
dt
−
However for every agent Pi we have either u+
i +uj , 0 if agent is surrounded, or
−
u+
i + ui is in the direction of the center of the disk sector in which all neighbors
(including Pj ) reside (see Figure 3).
6
−
Figure 3: uTij (u+
i + ui ) ≥ 0
+
−
−
Therefore the inner product uTij (u+
i + ui ) =< uij , (ui + ui ) > will necessary
be positive (see Appendix 3 for a formal proof), hence
d (t)
∆ = −(v0 ∗ positive + v0 ∗ positive) ≤ 0
dt ij
This shows that distances between neighbors can only decrease (or remain the
same). Hence agents never lose neighbors under the assumed dynamics.
2.2
Finite-time gathering
We have seen that the dynamics of the system (1) ensures that agents that are
neighbors at t = 0 will forever remain neighbors. We shall next prove that,
as time passes, agents acquire new neighbors and in fact will all converge to a
common point of encounter. We prove the following.
Theorem 2.2.1. A multi-agent system with dynamics given by (1) gathers all
agents to a point in R2 , in finite time.
Proof. We shall rely on a Lyapunov function L(P1 , . . . , PN ), a positive function
defined on the geometry of agent constellations which becomes zero if and only
if all agents’ locations are identical. We shall show that, due to the dynamics
7
of the system, the function L(P1 , . . . , PN ) decreases to zero at a rate bounded
away from zero, ensuring finite time convergence.
The function L will be defined as the perimeter of the convex hull of all agents’
locations, CH{Pi (t)}i=1,...,N . Indeed, consider the set of agents that are, at a
given time t, the vertices of the convex hull of the set {Pi (t)}i=1,...,N . Let us call
these agents {P̃k (t)} for k = 1, . . . , K ≤ N . For every agent P̃k on the convex
hull (i.e. for every agent that is a corner of the convex polygon defining the
convex hull), we have that all other agents, are in a region (wedge) determined
by the half lines from P̃k in the directions P̃k P̃k−1 and P̃k P̃k+1 , a wedge with
an opening angle say θk (see Figure 4).
Figure 4: Angle at a vertex of the convex hull
Since clearly θk ≤ π for all k we must have that agent P̃k has all its visible
neighbors in a wedge of his visibility disk with an angle αk ≤ θk ≤ π hence
−
his u+
k and uk vectors will not be zero, causing the motion of P̃k towards the
interior of the convex hull. This will ensure the shrinking of the convex hull,
while it exists, and the rate of this shrinking will be determined by the evolution
of the constellation of agents’ locations. Let us formally prove that indeed, the
convex hull will shrink to a point in finite time.
Consider the perimeter L(t) of CH{Pi (t)}i=1,...,N
K(t)
L(t) =
X
k=1
K(t)
∆k,k+1 =
X
[(P̃k+1 )(t) − P̃k (t))T (P̃k+1 (t) − P̃k (t))]1/2
k=1
where the indices are considered modulo K(t).
We have, assuming that K remains the same for a while,
8
K
K
X
X
d
d
−
+
−
T
L(t) =
∆k = −
v0 ũTk,k+1 (u+
k + uk ) + v0 ũk,k+1 (uk+1 + uk+1 )
dt
dt
k=1
k=1
−
but note that ũk,k+1 does not necessarily lie between u+
k and uk anymore, since,
in fact, P̃k and P̃k+1 might not even be neighbors.
Now let us consider
d
dt L(t)
and rewrite it as follows
K
K
k=1
k=1
X
X
d
−
−
L(t) = −v0
ũTk,k+1 (u+
ũTk+1,k (u+
k + uk ) − v0
k+1 + uk+1 )
dt
Rewriting the second term above, by moving the indices k by -1 we get
K
K
k=1
k=1
X
X
d
−
−
L(t) = −v0
ũTk,k+1 (u+
ũTk,k−1 (u+
k + u k ) − v0
k + uk )
dt
This yields
K
K
k=1
k=1
X
X
d
L(t) = −v0
< u+
< u−
k , ũk,k+1 + ũk,k−1 > −v0
k , ũk,k+1 + ũk,k−1 >
dt
Note that we have here inner products between unit vectors, yielding the cosines
of the angles between them. Therefore, defining θk = the angle between ũk,k−1
and ũk,k+1 (i.e. the interior angle of the convex hull at the vertex k, see Figure
5), and the angles :
αk+
βk+
αk−
βk−
, γ(u+
k , ũk,k+1 )
, γ(ũk,k−1 , u+
k)
, γ(ũk,k−1 , u−
k)
, γ(u−
,
ũ
k,k+1 )
k
9
Figure 5: Angles at a vertex of the convex hull.
we have αk+ + βk+ = αk− + βk− = θk and all these angles are between 0 and π.
Using these angles we can rewrite
K
K
k=1
k=1
X
X
d
L(t) = −
v0 (cos αk+ + cos βk+ ) −
v0 (cos αk− + cos βk− )
dt
Now, using the inequality (proved in Apenedix 1)
cos α + cos β ≥ 1 + cos(α + β)
0 ≤ α, β, α + β ≤ π
(2)
K
X
d
L(t) ≥ 2v0
(1 + cos θi )
dt
i=1
(3)
we obtain that
−
For any convex polygon we have the following result (see the detailed proof in
Appendix 1) :
Lemma 1. For any convex polygon with K vertices and interior angles θ1 , . . . , θK ,
with (θ1 + . . . + θK ) = (K − 2)π we have that
K
1 + (K − 1) cos (K−2)π
X
2≤K≤6
K−1
(4)
cos(θi ) ≥
(K−2)π
K cos
K≥7
k=1
K
10
Therefore, we obtain from (3) and (4) that
−
d
L(t) ≥ µ(K)
dt
(5)
where
1 + (K − 1) cos (K−2)π
2
≤
K
≤
6
K−1
2v0 K +
K cos (K−2)π
K≥7
K
n
o
K−1
π
1
2v0 K 1 − max cos 2π
,
cos
−
K
K
K−1
K
µ(K)
=
=
Note here that, since (1 − max{. . .}) > 0 we have that the rate of decrease in
the perimeter of the configuration is srictly positive while the convex hull of the
agent location is not a single point.
The argument outlined so far assumed that the number of agents determining
the convex hull of their constellation is a constant K. Suppose however that
in the course of evolution some agents collide and/or some agents become “exposed” as vertices of the convex hull, and hence K may jump to some different
integer value. At a collision between two agents we assume that they merge and
thereafter continue to move as a single agent. Since irrespective of the value
of K the perimeter decreases at a rate which is strictly positive and bounded
away from zero we have effectively proved that in finite time the perimeter of
the convex hull will necessarily reach 0. This concludes the proof of Theorem
2.2.1.
Figure 6 shows the bound as a function of K assuming v0 = 1. Note that we
always have K ≤ N , and µ(K) is a decreasing function of K, hence for any finite
number of agents there will be a strictly positive constant µ(N ) so that
−
d
L(t) ≥ µ(N )
dt
ensuring that after a finite time of Tub given by
µ(N )Tub = L(0) ⇒ Tub =
L(0)
µ(N )
we shall have that L(Tub ) = 0.
Hence we have an upper bound on the time of convergence for any configuration
11
of N agents given by
L(0)
µ(N ) .
Note from (3) that the rate of decrease does not depend on the perimeter of the
convex hull but only on the number of agents forming it. This was an expected
result, since the dynamics does not rely on Euclidian distances. This bound is
decreasing with K, so that the more agents form the convex hull, the smaller
will be the rate of decrease. For K = 2 and K = 3 the bound is −8v0 , for
K = 4, it is −7v0 , and then it keeps increasing for higher values of K, slowly
converging towards 0 form below. Note the change of curve between K = 6 and
K = 7, due to the “interresting” discontinuity in the geometric result exhibited
in equation (1).
Figure 6: Graph of the bound µ(K) of (3). The graph on the right is a zoom
on small values of K.
The inequalities of (2) and of (4) become equalities for particular configurations
of the agents (for example a regular polygon in which each pair of adjacent
neighbors are visible to each other, if K ≥ 7). In this case, the bound in (3)
will yield the exact rate of convergence of the convex-hull perimeter as long as
K remains the same.
3
Generalizations
All the above analysis can be generalized for dynamics of the form
dPi
−
= f (P (i) )(u+
i + ui ) for i = 1, . . . , N
dt
(6)
f (P (i) ) ≥ 0 is some positive function of the configuration of the neighbours seen
by agent i. This generalization also guarantees that the rule of motion is locally
defined and reactive, and defined in the same way for all agents.
The dynamics (1) correspond to a particular case of (6), with f (P (i) ) = v0 = constant
12
for all agents.
It is easy to slightly change the proofs above in order to show that Theorem
2.1.1 (ensuring that connectivity is not lost) is still valid as long as f (P (i) ) ≥ 0
for all i, and that Theorem 2.2.1 (ensuring finite time gathering) is also valid as
long as f (P (i) ) ≥ > 0 for all i.
Note that in the work of Gordon et al [5], a constant speed for the agents was
1
for a mobile
considered, and this corresponds to setting f (P (i) ) = ||u+ +u
−
||
i
i
agent i, rather than v0 . Given that in this case f (P (i) ) ≥ 12 , the conditions
for Theorem 2.1.1 and Theorem 2.2.1 are verified, and hence the dynamics
with constant speed also ensures convergence to a single point without pairs
of initially visible agents losing connectivity. We therefore also have a proof
for the convergence of the algorithm that was proposed in the above-mentioned
paper.
However, it was pointed out in the above-mentioned paper that in the model
with constant speed agents of one has to deal with quite unpleasant chattering
effects and “Zeno”-ness in order to effectively modulate the speed of motion of
some of the agents. In fact, in certain configurations, an agent may “oscillate”
infinitely often between a position in which it is “surrounded” to a position
in which it is not. This implies alternating its speed infinitely often between
zero and a constant value. In contrast, in our model, the speed is defined to
vary smoothly in the range [0, 2v0 ]. Therefore our model presents some clear
advantages and natural modulation for the speed of the agents.
4
Simulations
Let us consider some simulations of the multi-agent dynamics discussed in this
paper. We will start with a randomly generated swarm and then we shall have
a look at some interesting particular configurations.
4.1
Some random generated swarm of 15 agents
First, we randomly generate a swarm of 15 agents, with a connected visibility
graph as initial configuration. We set v0 to 1 and visibility to 200. The configuration of the swarm at different times during the evolution is given in Figure
7.
13
Figure 7: Configuration of the swarm at different times. The convex hull of the
set of agents is also shown.
We also plot some properties of the swarm during the time of its evolution.
Figure 8 represents the perimeter of the convex hull of the set of agents. Figure
9 plots the count of indistinguishable agents (i.e. collided agents count as one)
in the convex hull. Figure 10 represents the time derivative of the convex hull
perimeter and the theoretical bound given by equation (5), as function of the
number of indistinguishable agents in the convex hull.
14
Figure 8: Perimeter of the convex hull, decreasing until it reaches zero.
Figure 9: Number of agents forming the convex hull of the set. Decreases when
two or more agents of the hull “merge” and increases when one or more agents
are “collected” by the hull and added to it.
15
Figure 10: Derivative of the convex hull perimeter (blue) and theoretical bound
(green), function of number of agents forming the convex hull, given by (5).
For a fixed number of agents forming the convex hull, one can see in Figure 10
that the derivative of the perimeter goes towards the theoretical bound. This
can be intuitively explained by the facts that far away agents evolve towards the
inside more rapidly, making the convex hull shape more regular, approaching to
the shape that yields the theoretical bound.
The discontinuity of the derivative of the perimeter of the convex hull that
occurs when there is no change in the number of agents forming the hull, for
example around t = 35 in Figure 10, is due to change in the connectivity graph
(which has not been printed for clarity). In this particular case, two agents on
the top left became visible to each other at this time and their directions and
speed changed, slowing down slightly the rate of perimeter decrease.
4.2
Regular polygon of 10 agents
The initial configuration is a regular polygon with 10 agents. Again, v0 is set
to 1 and visibility is 200. As expected, the decreasing rate of the perimeter of
the convex hull is constant, and all the agents contribute to the convex hull all
along, until the very end where they collide and merge. Simulations and the
converging parameters’ plots are seen in Figures 11 and 12.
16
Figure 11: Starting as regular polygon, the swarm keeps its regular shape until
the final collision. The leftmost and rightmost agents of each agent are its two
neighbors in the polygon.
Figure 12: From left to right : the perimeter of the convex hull decreasing at a
constant rate, the number of agents forming the convex hull being constant, and
the derivative of the perimeter of the convex hull being constant and equaling
its theoretical bound all along the evolution.
4.3
Close-to-minimum-configuration with n = 4
We start at a configuration close to the configuration that reaches the minimum
of the sum of the cosinuses of the interior angles of the polygon. Results are
represented in Figures 13, 14, 15. One can see that this configuration provides
a smaller decreasing rate than the configuration of a regular polygon, in conformity with our analysis, where the bound is not realized for a regular polygon
configuration for small values of n.
17
Figure 13: The cone configuration approaches the theoretical bound configuration. In this case, the gathering can be divided in two stages. First, the 3
agents of one side merge, then the resulting composite agent merges with the
single agent of the other side.
Figure 14: From top to bottom : The perimeter of the convex hull, the number
of agents forming it, and the derivative of it.
18
Figure 15: A zoom on the the derivative of the perimeter of the convex hull
(in blue) at the beginning of the dynamics. A regular polygon (a square in this
case) would give a constant rate of -8, lower than the current rate approaching
the theoretical bound of -7 (in green).
5
Concluding remarks
We have shown that a very simple local control on the velocity of agents in
the plane, based on limited visibility and bearing only sensing of neighbors
ensures their finite time gathering. The motion rule is simply to set the agents’
velocity to equal the vector sum of unit vector pointers to two external neighbors
if all visible neighbors reside inside a halfplane (a half-disk) about the agent,
otherwise set the velocity to zero. This very simple rule of behavior is slightly
different from the one assumed by Gordon, Wagner, and Bruckstein in [5], where
the motion was set to have a constant velocity in the direction bisecting the disk
sector where visible neighbors reside, or zero if the agent was “surrounded”.
As we showed in this paper, that model also ensures gathering. However, it
was pointed out there that their proposed model had to deal with some quite
unpleasant chattering or zeno-ness effects in order to effectively modulate the
speed of motion of agents.
In this paper, in conjunction with our model, and some generalizations too,
including the model of [5], we provided a very simple geometric proof that finite
time gathering is achieved, and provided precise bounds on the rate of decrease
of the perimeter of the agent configuration’s convex hull. These bounds are
based on a geometric lower bound on the sum of cosines of the interior angles
of an arbitrary convex planar polygon, that is interesting on its own right (a
curious breakpoint occurring in the bound at 7 vertices). Our result may be
regarded as a convergence proof for a highly nonlinear autonomous dynamic
system, naturally handling dynamic changes in its dimension (the events when
two agents meet and merge).
19
References
[1] Olfati Saber R., Fox J.A., and Murray R.M. Consensus and cooperation
in networked multi-agent systems. Proceedings of the IEEE, 95-1:215–233,
2007.
[2] Gazi V. and Passimo K.M. Stability analysis of swarms. IEEE Transactions
on Automatic Control, 51-6:692–697, 2006.
[3] Ji M. and Eggerstedt M. Distributed coordination control of multiagent
systems while preserving connectedness. IEEE Transactions on Robotics,
23-4:693–703, 2007.
[4] Ando H.and Oasa Y., Suzuki I., and Yamashita. Distributed memoryless
point convergence algorithm for mobile robots with limited visibility. IEEE
Transactions on Robotics and Automation, 15-5:818–828, 1999.
[5] Gordon N., Wagner I.A., and Bruckstein A.M. Gathering multiple robotic
a(ge)nts with limited sensing capabilities. Lecture Notes in Computer Science, 3172:142–153, 2004.
[6] Gordon N., Wagner I.A., and Bruckstein A.M. A randomized gathering
algorithm for multipe robots with limited sensing capabilities. MARS 2005
Workshop Proceedings (ICINCO 2005), Barcelona, Spain, 2005.
[7] Gordon N., Elor Y., and Bruckstein A.M. Gathering multiple robotic agents
with crude distance sensing capabilities. Lecture Notes in Computer Science,
5217:72–83, 2008.
20
Appendix 1 : Proof of Lemma 1
We will first prove the following facts :
Fact a. Let 0 ≤ a ≤ b ≤ π and 0 ≤ a + b ≤ π. Then we have
p
2(1 + cos(a + b)) = 2 cos
a+b
2
≥ cos(a)+cos(b) ≥ 2 cos2
a+b
2
Proof. The function cosine is decreasing in [0, π], and given that
1 ≥ cos
a+b
2
Multiplying by 2 cos
2 cos
2 cos
a+b
2
a+b
2
≥
≥
b−a
2
≥ cos
a+b
2
a+b
2
= 1+cos(a+b)
≥
b−a
2
:
≥0:
2 cos a+b
cos
2
cos(a) + cos(b)
b−a
2
≥ 2 cos2 a+b
2
≥ 1 + cos(a + b)
A direct consequence is the following lemma.
Fact b. Let 0 ≤ a, b ≤ π. Then
cos(a) + cos(b) ≥
1 + cos(a +
b) : a + b ≤ π
2 cos a+b
:a+b≥π
2
Proof. The first line is already part of Lemma a. The second line can be proven
by using the left inequality of Lemma a with π − a and π − b, noticing that
0 ≤ π − a ≤ π, 0 ≤ π − b ≤ π, and π − a + π − b ≤ π for a + b ≥ π.
Now we can prove Lemma 1. Suppose any given initial configuration of the polygon with interior angles 0 ≤ x1 , . . . , xn ≤ π. We then have x1 +. . .+xn = (n−2)π.
Now repeat the following step : Go through all the pairs of non-zero values
(xi , xj ). As long as there is still a pair verifying xi + xj ≤ π, transform it from
(xi , xj ) to (0, xi + xj ). When there are no such pairs, then among all the nonzero values, take the the minimum value and the maximum value, say xi and
xj (they must verify xi + xj ≥ π due
applied process) , and
to the previously
transform the pair from (xi , xj ) to
xi +xj xi +xj
2 ,
2
.
Repeat the above process until convergence. We prove that the process converges and that we can get as close as wanted to a configuration where all
non-zero values are equal. Note that after each step, the sum of the values is
21
unchanged, (n − 2)π, and that the values of all xi ’s remain between 0 and π.
The number of values that the above process set to zero must be less or equal to
2 in order to have the sum of the n positive values equal to (n − 2)π. Therefore
we can be sure that after a finite number of iterations, there will be no pairs of
nonzero values whose sum will be less than π (otherwise this would allow us to
add a zero value without changing the sum).
Once in this situation,all we do is replacing
pairs of “farthest” non-zero values
xi +xj xi +xj
. Let us show that all the nonzero values
(xi , xj ) with the pair
2 ,
2
will converge to the same value, specifically to their mean.
Let k be the number of remaining non-zero values after the iteration t0 which
sets the “last value” to zero. Denote these values at the i-th iteration by
(i)
(i)
(x1 , . . . , xk ). Define :
(i)
m=
(i)
x1 + . . . + xk
(n − 2)π
=
k
k
(i)
(i)
Ei = (x1 − m)2 + . . . + (xk − m)2
Suppose, without loss of generality, that at the i-th iteration
the extreme values
(i)
(i)
(i+1)
were x1 and x2 and so we transformed (x1 , x2 ) into x1
(i)
=
(i)
x1 +x2
2
So we have :
(i)
Ei+1 − Ei
(i)
= 2(
=
(i)
x1 +x2
2
(i)
(i)
− m)2 − (x1 − m)2 − (x2 − m)2 )
(i)
(i)
− 21 (x1 − x2 )2
(i)
But x1 and x2 being the extreme values, we have for any 1 ≤ l ≤ k :
(i)
(i)
(i)
(x1 − x2 )2 ≥ (xl − m)2
and by summing over l we get that :
(i)
(i)
k(x1 − x2 )2 ≥ Ei
Hence
22
(i+1)
, x2
=
(i)
(i)
x1 +x2
2
.
Ei+1 − Ei
Ei+1
≤
Ei
0
(i)
(i)
i
= − 12 (x1 − x2 )2 ≤ − E
2k
1
≤
1 − 2k Ei
1 i−t0
Et0
≤
1 − 2k
proving that Ei converges to zero, i.e. all the non-zero values converge to m.
At each step of the above described process, according to fact b, the sum of
cosines can only decrease. Therefore from any given configuration we can get
as close as possible to a configuration in which all non-zero values are equal,
without increasing the sum of the cosines. Hence, the minimum value must be
reached in a configuration in which all non-zero values are equal.
Remebering that there can be at most only two zero values, the minimum value
of the sum of the cosines is the minimum of the following :
= −(n − 4) (case with 2 zeros)
• 2 + (n − 2) cos (n−2)π
n−2
(case with 1 zero)
• 1 + (n − 1) cos (n−2)π
n−1
• n cos (n−2)π
(case with no zero)
n
Now let us compare these values. In order to do so define the function e(x) = cos(x)+x sin(x)
0
for 0 ≤ x ≤ π. Basic
calculations give us
e (x)
= x cos(x) and therefore e
is increasing in 0, π2 and decreasing in π2 , π . Therefore for 0 ≤ x ≤ π2 ,
e(x) ≥ e(0) = 1
In order to compare the case with 2 zeros to the case with 1 zero, define for
n≥2
f (n)
= 1 + (n − 1) cos
(n−2)π
n−1
+ (n − 4)
π
n − 3 − (n − 1) cos n−1
π
π
π
f (n) = 1 − cos n−1
− n−1
sin n−1
π
= 1 − e n−1
=
0
0
Therefore f (n) ≤ 0 for n ≥ 4 because
π
n−1
f (2) = 0 ≤ 0
f (3) = 0 ≤ 0
f (4) = − 12 ≤ 0
f (n) ≤ f (4) ≤ 0 for n ≥ 4
23
≤
π
2
Therefore f (n) ≤ 0 and the case with 2 zeros is never exclusively the optimal solution (since the case with 1 zero always has a smaller or an equal value).
In order to compare the two remaining values, define for n ≥ 2
h(n) = n cos
2π
n
− (n − 1) cos
π
n−1
+1
The derivative according to n is
0
h (n)
2π
n
sin
+ 2π
n
2π
π
= e n − e n−1
=
cos
2π
n
− cos
π
n−1
0
−
π
n−1
sin
π
n−1
π
π
For n ≥ 4, we have n−1
≤ 2π
n ≤ 2 and therefore h (n) ≥ 0. One can check
that h(n) ≤ 0 for 2 ≤ n ≤ 6, and h(7) > 0, therefore h(n) > 0 for n ≥ 7. This
allows us to conclude that for 2 ≤ n ≤ 6 the minimal configuration is the one
corresponding to 1 zero, whereas for n ≥ 7, it is the one with no zeros.
24
Appendix 2
Let us understand the geometric Lemma 1 by some illustrations. Consider a
convex polygon of n vertices. The sum of its interior angles is equal to (n − 2)π
and each angle is between 0 and π. Denote by Cn the bound given by Lemma
1.
First let us notice that sometimes the minimum value of Cn corresponds to a
set of interior angles that cannot be realized by a polygon in the plane (for
example we can show that in a convex polygon, if one of the interior angles is
0, then there are exactly two interior angles of 0 and (n − 2) interior angles
of π, but for example the configuration that realizes the minimum of C4 with
n = 4 does not correspond to such a realizable configuration, see figure 19 for
an illustration). However, we can get arbitrarily close to this value of Cn with
changing the angles with value of 0 by some and substracting in other angles
(that are not 0) the added values (see figures 17 and 20 for examples).
Let us explain the intuition behind Lemma 1. If ones tries to minimize the cosinus of an angle, he would open it at maximum. But the constraint of forming a
convex polygon forces one to close the loop of the polygon. This is what limits
how much one can open the angles. It is intuitive that the more agents we have,
the more freedom we have to open the angles and use the numerous agents at
our disposal to close the loop.
The abstract case when there is an infinity of agents corresponds to a circle
where all angles can be open at maximum, i.e. with an angle of π.
What this Lemma also infers is that the configuration of the polygon that reaches
the minimum value is a regular polygon for n ≥ 7, and that for n ≤ 6, the minimum value of the sum of the cosinuses of the interior angles can be arbitrarily
closely approached by a polygon in a shape of a cone. Figures 16, 17, and 18
illustrate this for the case n = 3, while figures 19, 20, and 21 illustrate it for
the case n = 4. The cases n = 5 and n = 6 are similar, and for n = 7 and
above, the configuration
of the minimum is a regular polygon leading the value
(n−2)π
(see examples in figures 22,23,24 ).
of Cn = n cos
n
25
Figure 16: Theoretical configuration corresponding the the minimum value of
C3 = 1
Figure 17: Practical configuration for n = 3 provinding a value that can get
potentially arbitrarily close to the theoretical minimum, with the help of .
Figure 18: Regular
polygon with n = 3, the sum of cosinuses of the interior
angles is 3 cos π3 = 32 . This value is greater than the minimum possible value.
26
Figure 19: Theoretical configuration corresponding the the minimum value of
C4 = − 12
Figure 20: Practical configuration for n = 4 provinding a value that can get
potentially arbitrarily close to the theoretical minimum, with the help of .
Figure 21: Regular
polygon with n = 4, the sum of cosinuses of the interior
= 0. This value is greater than the minimum possible value.
angles is 4 cos 2π
4
27
Figure 22: n = 7 is the first value of n for which the configuration of minimum
value of the sum of the cosinuses of the interior
angles corresponds to a regular
≈ −4.36.
polygon. In this case, this value is 7 cos 5π
7
Figure 23: Regular
n = 10, with
polygon and configuration of minimum value for8π
value 10 cos 8π
≈
−8.09,
each
angle
contributing
a
value
of
cos
10
10 ≈ −0.81.
Figure 24: Regular polygon
and configuration of minimum value for n = 30,
28π
with value
30
cos
≈
−29.34, each angle contributing a value of
30
28π
cos 30 ≈ −0.98. This last value gets closer to −1 with higher values of
n as explained.
28
Appendix 3
The dynamics of the sytem described by (1) can be defined in the following
alternative way.
We define the two following functions :
h+ (x) =
1
:x>0
:x=0
:x<0
1
2
0
and h− (x) = h+ (−x).
Let ez be a unitary vector orthogonal to the plane (in any direction). Then we
define :
sijk = (uij × uik ) · ez
p±
ij =
Y
h± (sijk )
k
p±
ij
±
wij
=P ±
j pij
±
where wij
= 0 if
P
j
p±
ij = 0.
+
−
Finally, we define wij = wij
+ wij
and the equations of movement are given by
:
ẋi = v0
X
wij uij
j
−
And the vectors u+
i and ui of system (1) are given by:
u±
i =
X
j
29
±
wij
uij
Appendix 4
Using the fact that u−
i and uij are, by the way they are defined, on the same
half-plane of u+
i
+
−
(u+
i × uij ) · (ui × ui ) ≥ 0
where × is the cross product of vectors. In the same way, uij and u+
i are on the
same half-plane of u−
:
i
−
+
(u−
i × uij ) · (ui × ui ) ≥ 0
But, using the fact that these vectors are unit vectors,
+
−
−
+
−
+
(u+
i × uij ) · (ui × ui ) = uij · ui − (ui · ui )(uij · ui )
−
−
+
+
−
+
−
(ui × uij ) · (ui × ui ) = uij · ui − (ui · ui )(uij · ui )
Therefore
+
−
−
−
+
+
−
+
−
(u+
i × uij ) · (ui × ui ) + (ui × uij ) · (ui × ui ) = (1 − ui · ui )uij · (ui + ui ) (7)
−
+
−
+
−
Now u+
i ·ui = 1 implies that ui = ui = uij , and in this case (ui +ui )·uij = 2 > 0.
−
In any other case, 1 − u+
i · ui > 0, and given that the left hand side of (7) is
positive, we must have
−
(u+
i + ui ) · uij ≥ 0
30
| 3cs.SY
|
EXPLOITING CHORDAL STRUCTURE IN POLYNOMIAL IDEALS:
A GRÖBNER BASES APPROACH
arXiv:1411.1745v2 [cs.SC] 9 May 2016
DIEGO CIFUENTES AND PABLO A. PARRILO
Abstract. Chordal structure and bounded treewidth allow for efficient computation in numerical linear algebra, graphical models, constraint satisfaction and many
other areas. In this paper, we begin the study of how to exploit chordal structure in
computational algebraic geometry, and in particular, for solving polynomial systems.
The structure of a system of polynomial equations can be described in terms of a
graph. By carefully exploiting the properties of this graph (in particular, its chordal
completions), more efficient algorithms can be developed. To this end, we develop
a new technique, which we refer to as chordal elimination, that relies on elimination
theory and Gröbner bases. By maintaining graph structure throughout the process,
chordal elimination can outperform standard Gröbner bases algorithms in many cases.
The reason is that all computations are done on “smaller” rings, of size equal to the
treewidth of the graph (instead of the total number of variables). In particular, for
a restricted class of ideals, the computational complexity is linear in the number of
variables. Chordal structure arises in many relevant applications. We demonstrate
the suitability of our methods in examples from graph colorings, cryptography, sensor
localization and differential equations.
1. Introduction
Systems of polynomial equations can be used to model a large variety of applications.
In most cases the systems arising have a particular sparsity structure, and exploiting
such structure can greatly improve their efficiency. When all polynomials have degree
one, we have the special case of systems of linear equations, which are often represented
using matrices. In such case, it is well known that under a chordal structure many
matrix algorithms can be done efficiently [36,37,39]. Similarly, many hard combinatorial
problems can be solved efficiently for chordal graphs [26]. Chordal graphs are also a
keystone in constraint satisfaction, graphical models and database theory [4,16,31]. We
address the question of whether chordality might also help solve nonlinear equations.
It is natural to expect that the complexity of “solving” a system of polynomials
should depend on the underlying graph structure of the equations. In particular, a
parameter of the graph called the treewidth determines the complexity of solving the
problems described above, and it should influence polynomial equations as well. For
instance, several combinatorial problems (e.g., Hamiltonian circuit, vertex colorings,
vertex cover) are NP-hard in general, but are tractable if the treewidth is bounded [6].
Date: May 11, 2016.
Key words and phrases. Chordal graph, Elimination theory, Gröbner bases, Structured polynomials, Treewidth.
1
2
DIEGO CIFUENTES AND PABLO A. PARRILO
Nevertheless, standard algebraic geometry techniques typically do not make use of this
graph. This paper links Gröbner bases with this graph structure of the system.
It should be mentioned that, unlike classical graph problems, the ubiquity of systems
of polynomials makes them hard to solve in the general case, even for small treewidth.
Indeed, solving zero dimensional quadratic equations of treewidth 1 is already NPcomplete, as seen in the following example.
Example 1.1 (Polynomials on trees are hard). Let a1 , . . . , an and S be given integers.
The Subset Sum problem asks for a subset A ⊆ {a1 , . . . , an } whose sum is equal to S.
Let si denote the sum of A ∩ {a1 , . . . , ai } and note that we can recover the subset A
from the values of s1 , . . . , sn . We can thus formulate the problem as:
0 = s0
0 = (si − si−1 )(si − si−1 − ai ),
S = sn
for 1 ≤ i ≤ n
Observe that the structure of these equations can be represented with the path graph
s0 —s1 —s2 · · · sn−1 —sn , which is a tree. However, it is well-known that the Subset Sum
problem is NP-complete.
Despite this hardness result, it is still desirable to take advantage of this chordal
structure. In this paper, we introduce a new method that exploits this structure. We
refer to it as chordal elimination. Chordal elimination is based on ideas used in sparse
linear algebra. In particular, if the equations are linear chordal elimination defaults to
sparse Gaussian elimination.
We proceed to formalize our statements now. We consider the polynomial ring R =
K[x0 , x1 , . . . , xn−1 ] over some algebraically closed field K. We fix once and for all the
lexicographic term order with x0 > x1 > · · · > xn−1 1. Given a system of polynomials
F = {f1 , f2 , . . . , fs } in the ring R, we associate to it a graph G(F ) with vertex set
V = {x0 , . . . , xn−1 }. Note that the vertices of G(F ) inherit the order from R. Such
graph is given by a union of cliques: for each fi we form a clique in all its variables.
Equivalently, there is an edge between xi and xj if and only if there is some polynomial
that contains both variables. We say that G(F ) constitutes the sparsity structure of
F . In constraint satisfaction problems, G(F ) is usually called the primal constraint
graph [16].
Throughout this document we fix an ideal I ⊆ R with a given set of generators F .
We assume that the associated graph G = G(F ) is chordal. Even more, we assume
that x0 > · · · > xn−1 is a perfect elimination ordering (see Definition 2.1) of the graph.
In the event that G is not chordal, the same reasoning applies by considering a chordal
completion. We want to compute the elimination ideals of I, denoted as eliml (I), while
preserving the sparsity structure. As we are mainly interested in the zero set of I
rather than finding the exact elimination ideals, we attempt to find some Il such that
V(Il ) = V(eliml (I)).
Question. Consider an ideal I ⊆ R with generators F , and fix the lex order x0 >
x1 > · · · > xn−1 . Assume that such order is a perfect elimination ordering of its
1Observe
that smaller indices correspond to larger variables.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
3
associated graph G(F ). Can we find ideals Il , with some generators Fl , such that
V(Il ) = V(eliml (I)) and the sparsity structure is preserved, i.e., G(Fl ) ⊆ G(F )?
We could also ask a stronger question: Does there exist a Gröbner basis gb that
preserves the sparsity structure, i.e., G(gb) ⊆ G(F )? It turns out that it is not generally
possible to find a Gröbner basis that preserves chordality, as seen in the next example.
Example 1.2 (Gröbner bases may destroy chordality). Let I = hx0 x2 − 1, x1 x2 − 1i,
whose associated graph is the path x0 —x2 —x1 . Note that any Gröbner basis must
contain the polynomial p = x0 − x1 , breaking the sparsity structure. Nevertheless,
we can find some generators for its first elimination ideal elim1 (I) = hx1 x2 − 1i, that
preserve such structure.
As evidenced in Example 1.2, a Gröbner basis with the same graph structure might
not exist, but we might still be able to find its elimination ideals. Our main method,
chordal elimination, attempts to find ideals Il as proposed above. It generalizes the
ideas of sparse linear algebra. As opposed to Gaussian elimination, in the general case
chordal elimination may not lead to the right elimination ideals. Nevertheless, we can
certify when the ideals found are correct. This allows us to prove that for a large family
of problems, which includes the case of linear equations, chordal elimination succeeds
in finding the elimination ideals.
The aim of chordal elimination is to obtain a good description of the ideal (e.g.,
a Gröbner basis), while preserving at the same time the underlying graph structure.
However, as illustrated above, there may not be a Gröbner basis that preserves the
structure. For larger systems, Gröbner bases can be extremely big and thus they may
not be practical. Nonetheless, we can ask for some sparse generators of the ideal that are
the closest to such Gröbner basis. We argue that one such representation can be given
by finding the elimination ideals of all maximal cliques of the graph. We extend chordal
elimination to compute these ideals in Algorithm 3. In case I is zero dimensional, it is
straightforward to obtain the roots from such representation.
Chordal elimination shares many of the limitations of other elimination methods. In
particular, if V(I) is finite, the complexity depends intrinsically on the size of the projection |πl (V(I))|. As such, it performs much better if such set is small. In Theorem 29
we show complexity bounds for certain family of ideals where this condition is met.
Specifically, we show that chordal elimination is O(n) if the treewidth is bounded.
Chordal structure arises in many different applications and we believe that algebraic
geometry algorithms should take advantage of it. The last part of this paper evaluates
our methods on some of such applications, including cryptography, sensor localization
and differential equations.
We now summarize our contributions.
• We present a new elimination algorithm that exploits chordal structure in systems of polynomial equations. This method is presented in Algorithm 2. To our
knowledge, this is the first work that exploits chordal structure in computational
algebraic geometry, as we will argue in the “Related work” section below.
• We prove that the chordal elimination algorithm computes the correct elimination ideals for a large family of problems, although (as explained in Section 3) in
4
DIEGO CIFUENTES AND PABLO A. PARRILO
general it may fail to do so. In particular, Lemma 8 specifies conditions under
which chordal elimination succeeds. We show in Theorem 14 that these conditions are met for a large class of problems. Among others, this class includes
linear equations and generic dense ideals.
• We present a recursive method (Algorithm 3) to compute the elimination ideals
of all maximal cliques of the graph. These ideals provide a good sparse description from which we can easily find all solutions, as seen in Section 5.2. We
show in Corollary 21 that this algorithm succeeds under the same conditions of
chordal elimination.
• We show in Theorem 29 and Corollary 30 that the complexity of our methods is
linear in the number of variables and exponential in the treewidth for a restricted
class of problems.
• Section 7 provides experimental evaluation of our methods in problems from
graph colorings, cryptography, sensor localization and differential equations. In
all these cases we show the advantages of chordal elimination over standard
Gröbner bases algorithms. In some cases, we show that we can also find a lex
Gröbner basis faster than with degrevlex ordering, by making use of chordal
elimination. This subverts the heuristic of preferring degrevlex.
The document is structured as follows. In Section 2 we provide a brief introduction
to chordal graphs and we recall some ideas from algebraic geometry. In Section 3 we
present our main method, chordal elimination. Section 4 presents some types of systems
under which chordal elimination succeeds. In Section 5, we present a method to find
the elimination ideals of all maximal cliques of the graph. In Section 6 we analyze the
computational complexity of the algorithms proposed for a certain class of problems.
Finally, Section 7 presents an experimental evaluation of our methods.
Related work. Even though there is a broad literature regarding chordality/treewidth
and also polynomial system solving, their interaction has received almost no attention.
A meeting point between these two areas is the case of linear equations, for which graph
modelling methods have been very successful. There is also a lot of research studying
connections between graph theory and computational/commutative algebra. We now
proceed to review previous works in these areas, comparing them with our methods.
Chordality and bounded treewidth. The concepts of chordality and bounded treewidth
are pervasive in many different research areas. In fact, several hard graph problems (e.g.,
vertex colorings, vertex covers, weighted independent set) can be solved efficiently in
chordal graphs and in graphs of bounded treewidth [6,26]. In a similar way, many problems in constraint satisfaction and graphical models become polynomial-time solvable
under bounded treewidth assumptions [11,16]. In other words, some hard problems are
fixed-parameter-tractable when they are parametrized by the treewidth.
The logic community has also studied families of graph problems which are fixedparameter-tractable with respect to the treewidth [9]. Makowsky and Meer applied
these methods to algebraic problems such as evaluation, feasibility and positivity of
polynomials [33]. They show that these problems are tractable under bounded treewidth
and finite domain conditions. On the contrary, our methods do not require a discrete
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
5
domain as they rely on well studied tools from computational algebraic geometry. Moreover, their methods are not implementable due to the large underlying constants.
Chordality in linear algebra. The use of graph theory methods in sparse linear algebra
goes back at least to the work of Parter [35]. It was soon realized that symmetric
Gaussian elimination (Cholesky factorization) does not introduce additional nonzero
entries (i.e., no fill-in), only in the case where the adjacency graph of the matrix is
chordal [39]. Current numerical linear algebra methods exploit this property by first
finding a small chordal completion of this adjacency graph [37]. The nonsymmetric
case is quite more complicated, but a standard approach is to use instead a chordal
completion of the adjacency graph of AT A [12, 37]. In this paper we generalize these
ideas to the case of nonlinear equations. We note that chordality is also used in several
sparse matrix problems from optimization, such as matrix inversion, positive semidefinite matrix completion and Hessian evaluation [36, 43].
Structured polynomials. Solving structured systems of polynomial equations is a wellstudied problem in computational algebraic geometry. Many past techniques make use
of different types of structure. In particular, properties such as symmetry [20, 25] and
multi-homogeneous structure [21] have been exploited within the Gröbner basis framework. Symmetry and multi-homogeneous structure have also been used in homotopy
continuation methods; see e.g., [40].
Sparsity in the equations has also been exploited by making use of polytopal abstractions of the system [42]. This idea has led to faster algorithms based on homotopy
methods [29, 32], sparse resultants [18] and more recently Gröbner bases [22]. All these
methods will perform efficiently provided that certain measure of complexity of the
system, known as the BKK bound, is small. Nonetheless, these type of methods do not
take advantage of the chordal structure we study in this paper. Indeed, we will see that
our methods may perform efficiently even when the number of solutions, and thus the
BKK bound, is very large.
A different body of methods come from the algebraic cryptanalysis community, which
considers very sparse equations over small finite fields. One popular approach is converting the problem into a SAT problem and use SAT solvers [2]. A different idea is seen
in [38], where they represent each equation with its zero set and treat it as a constraint
satisfaction problem (CSP). These methods implicitly exploit the graph structure of the
system as both SAT and CSP solvers can take advantage of it. Our work, on the other
hand, directly relates the graph structure with the algebraic properties of the ideal. In
addition, our methods apply to positive dimensional systems and arbitrary fields.
Graphs in computer algebra. There is a long standing interaction between graph theory
and computational algebra. Indeed, several polynomial ideals have been associated
to graphs in the past years [3, 13, 27, 44]. One of the earliest such ideals was the
coloring ideal [3, 13, 28], which was recently considered for the special case of chordal
graphs [14]. The edge ideal is perhaps the most widely studied example [44, 45], given
its tight connections with simplicial complexes, and the many graph properties that
can be inferred from the ideal (e.g., connectedness, acyclicity, colorability, chordality).
More recently, the related binomial edge ideals have also attracted a lot of research [27].
6
DIEGO CIFUENTES AND PABLO A. PARRILO
These graph ideals allow to infer combinatorial properties by means of commutative/computational algebra, and they are also crucial in understanding the complexity
of computational algebra problems. However, previous work has mostly focused on
structural properties of these specific families of ideals, as opposed to effective methods
for general polynomials, such as the ones from sparse linear algebra. In contrast, our
paper uses graph theoretic methods as a constructive guide to perform computations
on arbitrary sparse polynomial systems.
2. Preliminaries
2.1. Chordal graphs. Chordal graphs, also known as triangulated graphs, have many
equivalent characterizations. A good presentation is found in [5]. For our purposes, we
use the following definition.
Definition 2.1. Let G be a graph with vertices x0 , . . . , xn−1 . An ordering of its vertices
x0 > x1 > · · · > xn−1 is a perfect elimination ordering if for each xl the set
(1)
Xl := {xl } ∪ {xm : xm is adjacent to xl , xm < xl }
is such that the restriction G|Xl is a clique. A graph G is chordal if it has a perfect
elimination ordering.
Remark. Observe that lower indices correspond to larger vertices.
Chordal graphs have many interesting properties. Observe, for instance, that the
number of maximal cliques is at most n. The reason is that any clique should be
contained in some Xl . It is easy to see that trees are chordal graphs: by successively
pruning a leaf from the tree we get a perfect elimination ordering.
Given a chordal graph G, a perfect elimination ordering can be found in linear time.
A classic and simple algorithm to do so is Maximum Cardinality Search (MCS) [5].
This algorithm successively selects a vertex with maximal number of neighbors among
previously chosen vertices, as shown in Algorithm 1. The ordering obtained is a reversed
perfect elimination ordering.
Algorithm 1 Maximum Cardinality Search [5]
Input: A chordal graph G = (V, E) and an optional initial clique
Output: A reversed perfect elimination ordering σ
1: procedure MCS(G, start = ∅)
2:
σ := start
3:
while |σ| < n do
4:
choose v ∈ V − σ, that maximizes |adj (v) ∩ σ|
5:
append v to σ
6:
return σ
Definition 2.2. Let G be an arbitrary graph. We say that G is a chordal completion
of G, if it is chordal and G is a subgraph of G. The clique number of G is the size of
its largest clique. The treewidth of G is the minimum clique number of G (minus one)
among all possible chordal completions.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
7
Observe that given any ordering x0 > · · · > xn−1 of the vertices of G, there is a
natural chordal completion G, i.e., we add edges to G in such a way that each G|Xl
is a clique. In general, we want to find a chordal completion with a small clique
number. However, there are n! possible orderings of the vertices and thus finding the
best chordal completion is not simple. Indeed, this problem is NP-hard [1], but there
are good heuristics and approximation algorithms [6].
0
6
9
1
7
8
4
3
5
2
Figure 1. 10-vertex graph (blue, solid) and a chordal completion (green, dashed).
Example 2.1. Let G be the blue/solid graph in Figure 1. This graph is not chordal
but if we add the three green/dashed edges shown in the figure we obtain a chordal
completion G. In fact, the ordering x0 > · · · > x9 is a perfect elimination ordering of
the chordal completion. The clique number of G is four and the treewidth of G is three.
As mentioned in the introduction, we will assume throughout this document that
the graph G = G(F ) is chordal and the ordering of its vertices (inherited from the
polynomial ring) is a perfect elimination ordering. However, for a non-chordal graph G
the same results hold by considering a chordal completion.
Remark (The linear case). We finalize this section by explaining how for linear equations
finding a chordal completion of G(F ) agrees with standard methods from numerical
linear algebra. A linear set of equations F can be written in matrix form as Ax = b.
This is equivalent to (AT A)x = AT b, which can be solved with a Cholesky factorization.
As mentioned earlier, to minimize the fill-in (nonzero entries) we need to find a small
chordal completion of the adjacency graph G of matrix AT A. It can be seen that this
adjacency graph G coincides with the graph G(F ) that we associate to the equations.
Alternatively, we can directly perform an LU decomposition (Gaussian elimination) on
matrix A, and it turns out that the adjacency graph G also bounds the fill-in of the LU
factors [12, 37].
2.2. Algebraic geometry. We use standard tools from computational algebraic geometry, following the notation from [10]. In particular, we assume familiarity with
Gröbner bases, elimination ideals and resultants.
We let eliml (I) be the l-th elimination ideal, i.e.,
eliml (I) := I ∩ K[xl , . . . , xn−1 ].
We will denote by Il the “approximation” that we will compute to this elimination
ideal, defined in Section 3. We also denote πl : Kn → Kn−l the projection onto the last
8
DIEGO CIFUENTES AND PABLO A. PARRILO
n − l coordinates.
We recall the correspondence between elimination and projection given by
V(eliml (I)) = πl (V(I))
where S denotes the closure of S with respect to the Zariski topology.
Remark. In order for the above equation to hold, we require that K is algebraically
closed, as we assume throughout this paper. However, if the coefficients of the equations
F are contained in a smaller field (e.g., K = C but the coefficients are in Q) then all
computations in our algorithms will stay within such field.
3. Chordal elimination
In this section, we present our main method: chordal elimination. As mentioned
before, we attempt to compute some generators for the elimination ideals with the
same structure G. The approach we follow mimics the Gaussian elimination process
by isolating the polynomials that do not involve the variables that we are eliminating.
The output of chordal elimination is an “approximate” elimination ideal that preserves
chordality. We call it approximate in the sense that, in general, it might not be the
exact elimination ideal, but hopefully it will be close to it. In fact, we will find inner
and outer approximations to the ideal, as will be seen later. In the case that both
approximations are the same we will be sure that the elimination ideal was computed
correctly.
3.1. Incremental elimination. We follow an incremental approach to compute the
elimination ideals, in a similar way as in Gaussian elimination. We illustrate the basic
methodology through the following example.
Example 3.1. Consider the following ideal
I = hx40 − 1, x20 + x2 , x21 + x2 , x22 + x3 i.
The associated graph is the tree in Figure 2. We incrementally eliminate each of the
0
1
2
3
Figure 2. Simple 3-vertex tree.
variables, considering at each step only the equations involving it. First, we consider
only the polynomials involving x0 ; there are two: x40 − 1, x20 + x2 . If we eliminate x0
from these equations we obtain x22 − 1. This leads to the following elimination ideal
I1 = hx21 + x2 , x22 − 1, x22 + x3 i.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
9
We now consider the polynomials involving x1 ; there is only one: x21 + x2 . Thus, we
cannot eliminate x1 , so our second elimination ideal is
I2 = hx22 − 1, x22 + x3 i.
Finally, we eliminate x2 from the remaining equations obtaining
I3 = hx3 + 1i.
For this example all elimination ideals found are correct, as can be seen from the lex
Gröbner basis gb = {x20 + x2 , x21 + x2 , x22 − 1, x3 + 1}.
Example 3.1 shows the basic idea we follow. Namely, to eliminate a variable xi
we only consider a subset of the equations. In the above example, these equations
only involved two variables at each step. In general, to eliminate xi we only take into
account its neighboring variables in the graph. Therefore, if the neighborhood of each
xi is small, we should require less computation. The chordality property will imply that
these neighborhoods (cliques) are never expanded in the process.
This successive elimination process is simple, but it is not clear whether it always
leads to the correct elimination ideals. The following example illustrates that this is
not always the case.
Example 3.2 (Incremental elimination may fail). Consider the ideal
I = hx0 x1 + 1, x1 + x2 , x1 x2 i.
The associated graph is the path x0 —x1 —x2 . We proceed in an incremental way as
before. First, we consider only the polynomials involving x0 , there is only one: x0 x1 +1.
Thus, we cannot eliminate x0 , and we are left with the ideal
I1 = hx1 + x2 , x1 x2 i.
Eliminating x1 from the two equations above, we obtain
I2 = hx22 i.
Observe that the original ideal I is infeasible, i.e., I = h1i, but the ideals I1 , I2 found
are feasible. Thus the elimination ideals found are not correct.
Example 3.1 and Example 3.2 show this incremental approach to obtain elimination
ideals. In the first case the elimination process was correct, but in the second case
it was not correct. The problem in the second example can be seen in the following
equation:
elim1 (hx0 x1 + 1, x1 + x2 , x1 x2 i) 6= elim1 (hx0 x1 + 1i) + hx1 + x2 , x1 x2 i.
The goal now is to understand why these ideals are different, and when can we ensure
that we successfully found the elimination ideals.
10
DIEGO CIFUENTES AND PABLO A. PARRILO
3.2. Bounding the first elimination ideal. We just introduced an incremental approach to compute elimination ideals and we observed that it might not be correct. As
will be shown next, the result of this process is always an inner approximation to the
actual elimination ideal. Even more, we will see that we can also find an outer approximation to it. By comparing these approximations (or bounds) we can certify the cases
where the elimination is correct. We now analyze the case of the first elimination ideal,
and we will later proceed to further elimination ideals.
We formalize the elimination procedure presented in Section 3.1. Let I1 be our
estimation to the first elimination ideal as described before. Recall that to compute
the ideal I1 we want to use only a subset of the equations; in the examples above the
ones containing variable x0 . Let’s denote as J the ideal of these equations, and K the
ideal of the remaining equations. Then I = J + K, and our estimation to the first
elimination ideal is given by I1 = elim1 (J) + K. Note that the equations of I involving
x0 must certainly be part of J.
In this way, to compute I1 we only to do operations on the generators of J; we never
deal with K. As a result, the computation of I1 can be done on a smaller ring, whose
variables correspond to a neighborhood, or clique, of the chordal graph. Chordality
will ensure that graphical structure of I is preserved, i.e., the graph associated to (the
generators of) I1 is a subgraph of G. We elaborate more on this later.
We want to show the relationship between our estimate I1 and the actual elimination
ideal elim1 (I). To do so, the key will be the Closure Theorem [10, Chapter 3].
Definition 3.1. Let 1 ≤ l < n and let I = hf1 , . . . , fs i ⊆ K[xl−1 , . . . , xn−1 ] be an ideal
with a fixed set of generators. For each 1 ≤ t ≤ s assume that ft is of the form
t
ft = ut (xl , . . . , xn−1 )xdl−1
+ (terms with smaller degree in xl−1 )
for some dt ≥ 0 and ut ∈ K[xl , . . . , xn−1 ]. We define the coefficient ideal of I to be
coeff l (I) := hut : 1 ≤ t ≤ si ⊆ K[xl , . . . , xn−1 ].
Theorem 1 (Closure Theorem). Let I = hf1 , . . . , fs i ⊆ K[x0 , . . . , xn−1 ]. Let W :=
coeff 1 (I) be the coefficient ideal, let elim1 (I) be the first elimination ideal, and let π :
Kn → Kn−1 be the projection onto the last factor. Then,
V(elim1 (I)) = π(V(I))
V(elim1 (I)) − V(W ) ⊆ π(V(I)).
The next lemma tells us that I1 is an inner approximation to the actual elimination
ideal elim1 (I). It also describes an outer approximation to it, which depends on I1 and
some ideal W . If the two bounds are equal, this implies that we successfully found the
elimination ideal.
Lemma 2. Let J = hf1 , . . . , fs i ⊆ K[x0 , . . . , xn−1 ], let K = hg1 , . . . , gr i ⊆ K[x1 , . . . , xn−1 ]
and let
I := J + K = hf1 , . . . , fs , g1 , . . . , gr i.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
11
Let the ideals I1 , W ⊆ K[x1 , . . . , xn−1 ] be
I1 := elim1 (J) + K
W := coeff 1 (J) + K.
Then the following equations hold:
V(elim1 (I)) = π(V(I)) ⊆ V(I1 )
(2)
V(I1 ) − V(W ) ⊆ π(V(I)).
(3)
Proof. We first show Equation (2). The Closure Theorem says that V(elim1 (I)) =
π(V(I)). We will show that π(V(I)) ⊆ V(I1 ), from which Equation (2) follows because
V(I1 ) is closed.
In the following equations sometimes we will consider the varieties in Kn and sometimes in Kn−1 . To specify, we will denote them as Vn and Vn−1 respectively. Notice
that
π(Vn (I)) = π(Vn (J + K)) = π(Vn (J) ∩ Vn (K)).
Now observe that
π(Vn (J) ∩ Vn (K)) = π(Vn (J)) ∩ Vn−1 (K).
The reason is the fact that if S ⊆ Kn , T ⊆ Kn−1 are arbitrary sets, then
π(S ∩ (K × T )) = π(S) ∩ T.
Finally, note that π(Vn (J)) = V(elim1 (J)). Combining everything we conclude:
π(V(I)) = π(Vn (J)) ∩ Vn−1 (K)
⊆ π(Vn (J)) ∩ Vn−1 (K)
= Vn−1 (elim1 (J)) ∩ Vn−1 (K)
= Vn−1 (elim1 (J) + K)
= V(I1 ).
We now show Equation (3). The Closure Theorem states that
V(elim1 (J)) − V(coeff 1 (J)) ⊆ π(V(J)).
Then,
V(I1 ) − V(W ) = [Vn−1 (elim1 (J)) ∩ Vn−1 (K)] − [Vn−1 (coeff 1 (J)) ∩ Vn−1 (K)]
= [Vn−1 (elim1 (J)) − Vn−1 (coeff 1 (J))] ∩ Vn−1 (K)
⊆ π(Vn (J)) ∩ Vn−1 (K)
= π(V(I)).
This concludes the proof.
12
DIEGO CIFUENTES AND PABLO A. PARRILO
Note that the lemma above implies the following equations:
V(I1 ) − V(W ) ⊆ V(elim1 (I)) ⊆ V(I1 )
p
p
p
√
I1 : W ⊇ elim1 (I) ⊇ I1
where we used that set difference of varieties corresponds to ideal quotient. Thus, the
ideal W bounds the approximation error of our estimation I1 to the ideal elim1 (I). In
particular, if V(W ) is empty then I1 and elim1 (I) determine the same variety.
3.3. Bounding all elimination ideals. Lemma 2 gave us the relationship between
our estimation I1 and the actual elimination ideal elim1 (I). We generalize this now to
further elimination ideals.
We denote by Il to our estimation to the l-th elimination ideal. As before, to estimate
eliml+1 (Il ) we only use a subset of the equations of Il , which we denote as Jl . The
remaining equations are denoted as Kl+1 . Then Il+1 = eliml+1 (Il )+Kl+1 . The following
theorem establishes the relationship between Il+1 and eliml+1 (I).
Theorem 3. Let I ⊆ K[x0 , . . . , xn−1 ] be an ideal. Consider ideals Il ⊆ K[xl , . . . , xn−1 ]
for 0 ≤ l < n, with I0 := I, which are constructed recursively as follows:
(i) Given Il , let Jl ⊆ K[xl , . . . , xn−1 ], Kl+1 ⊆ K[xl+1 , . . . , xn−1 ] be2 such that Il =
Jl + Kl+1 .
(ii) Let Il+1 := eliml+1 (Jl ) + Kl+1 .
(iii) Also denote Wl+1 := coeff l+1 (Jl ) + Kl+1 .
Then for each l the following equations hold:
V(eliml (I)) = πl (V(I)) ⊆ V(Il )
(4)
V(Il ) − [πl (V(W1 )) ∪ · · · ∪ πl (V(Wl ))] ⊆ πl (V(I)).
(5)
Proof. It follows from Lemma 2 by induction. See Appendix A.1.
The lemma above implies the following equations:
V(IL ) − V(W ) ⊆ V(elimL (I)) ⊆ V(IL )
p
p
p
√
IL : W ⊇ elimL (I) ⊇ IL
(6a)
(6b)
where the ideal W is
W := elimL (W1 ) ∩ · · · ∩ elimL (WL ).
(7)
Note also that by construction we always have that if xm < xl then Im ⊆ Il .
3.4. Chordal elimination algorithm. The recursive construction given in Theorem 3
is not fully specified (see item (i)). In particular, it is not clear which decomposition
Il = Jl + Kl+1 to use. We now describe the specific decomposition we use to obtain the
chordal elimination algorithm.
We recall the definition of the cliques Xl from Equation (1). Equivalently, Xl is
the largest clique containing xl in G|{xl ,...,xn−1 } . Let fj be a generator of Il . If all the
variables in fj are contained in Xl , we put fj in Jl . Otherwise, if some variable of fj is
not in Xl , we put fj in Kl+1 . We refer to this procedure as clique decomposition.
2Note
that this decomposition is not unique, since we are not fully specifying the ideals Jl , Kl+1 .
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
13
Example 3.3. Let I = hf, g, hi where f = x20 +x1 x2 , g = x31 +x2 and h = x1 +x3 . Note
that the associated graph consists of a triangle x0 , x1 , x2 and the edge x1 , x3 . Thus, we
have X0 = {x0 , x1 , x2 }. The clique decomposition sets J0 = hf, gi, K1 = hhi.
We should mention that this decomposition method is reminiscent to the bucket
elimination algorithm from constraint satisfaction [15]. However, we do not place an
equation f in its largest variable, but rather in the largest variable xl such that f ∈
K[Xl ]. The reason for doing this is to shrink the variety V(Jl ) further. This leads
to a tighter approximation of the elimination ideals and simplifies the Gröbner basis
computation.
It is easy to see that the procedure in Theorem 3, using this clique decomposition,
preserves chordality. We state that now.
Proposition 4. Let I be an ideal with chordal graph G. If we follow the procedure in
Theorem 3 using the clique decomposition, then the graph associated to Il is a subgraph
of G.
Proof. Observe that we do not modify the generators of Kl+1 , and thus the only part
where we may alter the sparsity pattern is when we compute eliml+1 (Jl ) and coeff l+1 (Jl ).
However, the variables involved in Jl are contained in the clique Xl and thus, independent of which operations we apply to its generators, we will not alter the structure.
After these comments, we solved the ambiguity problem of step (i) in Theorem 3.
However, there is still an issue regarding the “error ideal” W of Equation (7). We recall
that Wl+1 depends on the coefficient ideal of Jl . Thus, Wl+1 does not only depend on the
ideal Jl , but it depends on the specific set of generators that we are using. In particular,
some set of generators might lead to a larger/worse variety V(Wl+1 ) than others. This
problem is inherent to the Closure theorem, and it is discussed in [10, Chapter 3]. It
turns out that a lex Gröbner basis of Jl is an optimal set of generators, as shown in [10].
Therefore, it is convenient to find this Gröbner basis before computing the coefficient
ideal.
Algorithm 2 presents the chordal elimination algorithm. The output of the algorithm
is the inner approximation IL to the L-th elimination ideal and the ideals W1 , . . . , WL ,
that satisfy (6). In the event that V(Wl ) = ∅ for all l, the elimination was correct.
This is the case that we focus on for the rest of the paper.
Remark. Observe that in line 14 of Algorithm 2 we append a Gröbner basis to Jl , so
that we do not remove the old generators. There are two reasons to compute this lex
Gröbner basis: it allows to find the eliml+1 (Jl ) easily and we obtain a tighter Wl+1 as
discussed above. However, we do not replace the old set of generators but instead we
append to them this Gröbner basis. We will explain the reason to do that in Section 3.5.
3.5. Elimination tree. We now introduce the concept of elimination tree, and show its
connection with chordal elimination. This concept will help us to analyze our methods.
Definition 3.2. Let G be an ordered graph with vertex set x0 > · · · > xn−1 . We
associate to G the following directed spanning tree T that we refer to as the elimination
tree: For each xl > xn−1 there is an arc from xl towards the largest xp that is adjacent
14
DIEGO CIFUENTES AND PABLO A. PARRILO
Algorithm 2 Chordal Elimination to find the L-th Elimination Ideal
Input: An ideal I, given by generators with chordal graph G, and an integer L
Output: Ideals IL and W1 , . . . , WL approximating elimL (I) as in (6)
1: procedure ChordElim(I, G, L)
2:
I0 = I
3:
for l = 0 : L − 1 do
4:
get clique Xl of G
5:
Jl , Kl+1 = SplitGens(Il , Xl )
6:
FindElim&Coeff(Jl )
7:
Il+1 = eliml+1 (Jl ) + Kl+1
8:
Wl+1 = coeff l+1 (Jl ) + Kl+1
9:
return IL , W1 , . . . , WL
10:
11:
12:
procedure SplitGens(Il , Xl )
Jl = hf : f generator of Il and f ∈ K[Xl ]i
Kl+1 = hf : f generator of Il and f ∈
/ K[Xl ]i
. Partition generators of Il
13:
14:
15:
16:
procedure FindElim&Coeff(Jl )
. Eliminate xl in the ring K[Xl ]
append to Jl its lex Gröbner basis
eliml+1 (Jl ) = hf : f generator of Jl with no xl i
coeff l+1 (Jl ) = hleading coefficient of f : f generator of Jl i
to xl and xp < xl . We will say that xp is the parent of xl and xl is a descendant of xp .
Note that T is rooted at xn−1 .
0
1
0
0
6
6
7
6
2
4
3
5
7
7
8
9
9
1
8
4
3
5
9
2
1
8
4
3
5
2
Figure 3. Chordal graph G and its elimination tree T .
Figure 3 shows an example of the elimination tree of a given graph. It is easy to see
that eliminating a variable xl corresponds to pruning one of the leaves of the elimination
tree. We now present a simple property of such tree.
Lemma 5. Let G be a chordal graph, let xl be some vertex and let xp be its parent in
the elimination tree T . Then,
Xl \ {xl } ⊆ Xp
where Xi is as in Equation (1).
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
15
Proof. Let C = Xl \ {xl }. Note that C is a clique that contains xp . Even more, xp is
its largest variable because of the definition of T . As Xp is the unique largest clique
satisfying such property, we must have C ⊆ Xp .
A consequence of the lemma above is the following relation:
(8)
eliml+1 (I ∩ K[Xl ]) ⊆ I ∩ K[Xp ]
where I ∩ K[Xl ] is the set of all polynomials in I that involve only variables in Xl . The
reason of the relation above is that
eliml+1 (I ∩ K[Xl ]) = (I ∩ K[Xl ]) ∩ K[xl+1 , . . . , xn−1 ] = I ∩ K[Xl \ {xl }].
There is a simple geometric interpretation of Equation (8). The variety V(I ∩ K[Xl ])
can be interpreted as the set of partial solutions restricted to the set Xl . Thus, Equation (8) is telling us that any partial solution on Xp extends to a partial solution on Xl
(the inclusion is reversed). Even though this equation is very simple, this is a property
that we would like to keep in chordal elimination.
Clearly, we do not have a representation of the clique elimination ideal I ∩ K[Xl ].
However, the natural relaxation to consider is the ideal Jl ⊆ K[Xl ] that we compute in
Algorithm 2. To preserve the property above (i.e., every partial solution of Xp extends
to Xl ), we would like to have the following relation:
(9)
eliml+1 (Jl ) ⊆ Jp .
It turns out that there is a very simple way to ensure this property: we preserve the
old generators of the ideal during the elimination process. This is precisely the reason
why in line 14 of Algorithm 2 we append a Gröbner basis to Jl .
We now prove that Equation (9) holds. We need one lemma before.
Lemma 6. In Algorithm 2, let f ∈ Il be one of its generators. If xm is such that
xm ≤ xl and f ∈ K[Xm ], then f is a generator of Jm . In particular, this holds if xm is
the largest variable in f .
Proof. For a fixed xm , we will show this by induction on xl .
The base case is l = m. In such case, by construction of Jm we have that f ∈ Jm .
Assume now that the assertion holds for any x0l with xm ≤ xl0 < xl and let f be a
generator of Il . There are two cases: either f ∈ Jl or f ∈ Kl+1 . In the second case,
f is a generator of Il+1 and using of the induction hypothesis we get f ∈ Jm . In the
first case, as f ∈ K[Xm ] then all variables of f are less or equal to xm , and thus strictly
smaller than xl . Following Algorithm 2, we see that f is a generator of eliml+1 (Jl ).
Thus, f is again a generator of Il+1 and we conclude by induction.
We now prove the second part, i.e., it holds if xm is the largest variable. We just
need to show that f ∈ K[Xm ]. Let Xlm := Xl ∩ {xm , . . . , xn−1 }, then f ∈ K[Xlm ] as xm
is the largest variable. Note that as f ∈ K[Xl ] and f involves xm , then xm ∈ Xl . Thus,
Xlm is a clique of G|{xm ,...,xn−1 } and it contains xm . However, Xm is the unique largest
clique that satisfies this property. Then Xlm ⊆ Xm so that f ∈ K[Xlm ] ⊆ K[Xm ].
Corollary 7. Let xl be arbitrary and let xp be its parent in the elimination tree T .
Then Equation (9) holds for Algorithm 2.
16
DIEGO CIFUENTES AND PABLO A. PARRILO
Proof. Let f ∈ eliml+1 (Jl ) be one of its generators. Then f is also one of the generators
of Il+1 , by construction. It is clear that the variables of f are contained in Xl \{xl } ⊆ Xp ,
where we used Lemma 5. From Lemma 6 we get that f ∈ Jp , concluding the proof.
The reader may believe that preserving the old set of generators is not necessary.
The following example shows that it is necessary in order to have the relation in (9).
Example 3.4. Consider the ideal
I = hx0 − x2 , x0 − x3 , x1 − x3 , x1 − x4 , x2 − x3 , x3 − x4 , x22 i,
whose associated graph consists of two triangles {x0 , x2 , x3 } and {x1 , x3 , x4 }. Note
that the parent of x0 is x2 . If we preserve the old generators (as in Algorithm 2),
we get I2 = hx2 − x3 , x3 − x4 , x22 , x23 , x24 i. If we do not preserve them, we get instead
Iˆ2 = hx2 − x3 , x3 − x4 , x24 i. In the last case we have Jˆ2 = hx2 − x3 i so that x22 ∈
/ Jˆ2 ,
even though x22 ∈ J0 . Moreover, the ideal J0 is zero dimensional, but Jˆ2 has positive
dimension. Thus, Equation (9) does not hold.
4. Successful elimination
rad
Notation. We will write I = J whenever we have V(I) = V(J).
In Section 3 we showed an algorithm that gives us an approximate elimination ideal.
In this section we are interested in finding conditions under which such algorithm returns
the actual elimination ideal. We will say that chordal elimination, i.e., Algorithm 2,
succeeds if we have V(Il ) = V(eliml (I)). Following the convention above, we write
rad
Il = eliml (I).
4.1. The domination condition. Theorem 3 gives us lower and upper bounds on the
actual elimination ideals. We use these bounds to obtain a condition that guarantees
that chordal elimination succeeds.
Definition 4.1. We say that a polynomial f is xi -dominated if its leading monomial
has the form xdi for some d. We say that an ideal J is xi -dominated if there is some
f ∈ J that is xi -dominated. Equivalently, J is xi -dominated if its initial ideal in(J)
contains a pure power of xi .
Definition 4.2 (Domination condition). Let I be an ideal and use Algorithm 2. We
say that the domination condition holds if Jl is xl -dominated for each l.
The domination condition implies that chordal elimination succeeds, as shown now.
rad
Lemma 8 (Domination implies Success). If the domination condition holds then Il =
eliml (I) and the corresponding variety is πl (V(I)), for all l.
Proof. As Jl is xl -dominated, then its initial ideal in(Jl ) contains a pure power of xl .
Thus, there must be a g that is part of the Gröbner basis of Jl and is xl -dominated. The
coefficient ut that corresponds to such g is ut = 1, and therefore 1 ∈ Wl and V(Wl ) = ∅.
Thus the two bounds in Theorem 3 are the same and the result follows.
We will now show some classes of ideals in which the domination condition holds.
Using the previous lemma, this guarantees that chordal elimination succeeds.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
17
Corollary 9. Let I be an ideal and assume that for each l such that Xl is a maximal
clique of G, the ideal Jl ⊆ K[Xl ] is zero dimensional. Then the domination condition
holds and chordal elimination succeeds.
Proof. Let xm be arbitrary, and let xl ≥ xm be such that Xm ⊆ Xl and Xl is a maximal
clique. As Jl ⊆ K[Xl ] is zero dimensional, then it is xj -dominated for all xj ∈ Xl .
Thus, there is a g that is part of the Gröbner basis of Jl and is xm -dominated. From
Lemma 6 we obtain that g ∈ Jm , and thus the domination condition holds.
Corollary 10. Let I be an ideal and assume that for each l there is a generator fl of
I that is xl -dominated. Then the domination condition holds and chordal elimination
succeeds.
Proof. It follows from Lemma 6 that fl ∈ Jl , so that Jl is xl -dominated and the domination condition holds.
The previous corollary presents a first class of ideals for which we are guaranteed to
have successful elimination. Note that when we solve equations over a finite field Fq ,
usually we include equations of the form xql − xl , so the corollary holds. In particular,
it holds for 0/1 problems.
4.2. Simplicial equations. The assumptions of Corollary 10 are too strong for many
cases. In particular, if l = n − 2 the only way that such assumption holds is if there is
a polynomial that only involves xn−2 , xn−1 . We will show now a bigger class of ideals
for which the domination condition also holds and thus chordal elimination succeeds.
The following concept is the basis for this class.
Definition 4.3. Let f ∈ K[x0 , . . . , xn−1 ] be such that for each variable xl of positive
degree, the monomial ml of f with largest degree in xl is unique and has the form
ml = xdl l for some dl > 0. We say that f is simplicial.
Example 4.1. Consider the polynomials of Example 3.2:
f1 = x0 x1 + 1,
f2 = x1 + x2 ,
f3 = x1 x2 .
Then f2 is simplicial, as for both x1 , x2 the monomials of largest degree in these variables
are pure powers. In general, linear equations are always simplicial. On the other hand,
f1 , f3 are not simplicial. This makes sense, as we will see that if all polynomials are
simplicial then chordal elimination succeeds, which was not the case of Example 3.2.
On the contrary, all the polynomials of Example 3.1 are simplicial.
Note that the definition of simplicial is independent of the monomial ordering used,
as opposed to xi -domination. The reason for the term simplicial is that the (scaled)
standard simplex
X
∆ = {x : x ≥ 0,
xl /dl = |Xf |}
xl ∈Xf
where Xf are the variables of f , is a face of the Newton polytope of f and they are the
same if f is homogeneous.
18
DIEGO CIFUENTES AND PABLO A. PARRILO
We will make an additional genericity assumption on the simplicial polynomials.
Concretely, we assume that the coefficients of ml = xdl l are generic, in a sense that will
be clear in the next lemma.
Lemma 11. Let q1 , q2 be generic simplicial polynomials. Let X1 , X2 denote their sets
of variables and let x ∈ X1 ∩ X2 . Then h = Resx (q1 , q2 ) is generic simplicial and its set
of variables is X1 ∪ X2 \ x.
Proof. Let q1 , q2 be of degree m1 , m2 as univariate polynomials in x. As q2 is simplicial,
for each xi ∈ X2 \ x the monomial with largest degree in xi has the form xdi 2 . It is easy
to see that the largest monomial of h, as a function of xi , that comes from q2 will be
xdi 2 m1 . Such monomial arises from the product of the main diagonal of the Sylvester
matrix. In the same way, the largest monomial that comes from q1 has the form xdi 1 m2 .
If d2 m1 = d1 m2 , the genericity guarantees that such monomials do not cancel each other
max{d2 m1 ,d1 m2 }
and then h is
out. Thus, the leading monomial of h in xi has the form xi
simplicial. The coefficients of the extreme monomials are polynomials in the coefficients
of q1 , q2 , so if they were not generic (they satisfy certain polynomial equation), then
q1 , q2 would not be generic either.
Observe that in the lemma above we required the coefficients to be generic in order to avoid cancellations in the resultant. This is the only part where we need this
assumption.
We recall that elimination can be viewed as pruning the elimination tree T of G
(Definition 3.2). We attach each of the generators of I to some node of T . More
precisely, we attach a generator f to the largest variable it contains, which we denote
as x(f ). The following lemma tells us that if there are many simplicial polynomials
attached to the subtree of T rooted in xl , then Jl is xl -dominated.
Lemma 12. Let I = hf1 , . . . , fs i and let 1 ≤ l < n. Let Tl be a subtree of T with
t vertices and minimal vertex xl . Assume that there are fi1 , . . . , fit generic simplicial
with largest variable x(fij ) ∈ Tl for 1 ≤ j ≤ t. Then Jl is xl -dominated.
Proof. We will show that we can find a simplicial polynomial fl ∈ Jl that contains xl ,
which implies the desired result. Let’s ignore all ft such that its largest variable is not
in Tl . By doing this, we get smaller ideals Jl , so it does not help to prove the statement.
Let’s also ignore all vertices which do not involve one of the remaining equations. Let
S be the set of variables which are not in Tl . As in any of the remaining equations the
largest variable should be in Tl , then for any xi ∈ S there is some xj ∈ Tl with xj > xi .
We will show that for any xi ∈ S we have xl > xi .
Assume by contradiction that it is not true, and let xi be the smallest counterexample.
Let xp be the parent of xi . Note that xp ∈
/ S because of the minimality of xi , and thus
xp ∈ Tl . As mentioned earlier, there is some xj ∈ Tl with xj > xi . As xj > xi and xp is
the parent of xi , this means that xi is in the path of T that joins xj and xp . However
xj , xp ∈ Tl and xi ∈
/ Tl so this contradicts that Tl is connected.
Thus, for any xi ∈ S, we have that xi < xl . This says that to obtain Jl we don’t
need to eliminate any of the variables in S. Therefore, we can ignore all variables in S.
Thus, we can assume that l = n − 1 and Tl = T . This reduces the problem the specific
case considered in the following lemma.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
19
Lemma 13. Let I = hf1 , . . . , fn i such that fj is generic simplicial for all j. Then there
is a simplicial polynomial f ∈ In−1 = Jn−1 .
l
∈ Il
Proof. We will prove the more general result: for each l there exist f1l , f2l , . . . , fn−l
which are all simplicial and generic. Moreover, we will show that if xj denotes the
largest variable of some fil , then fil ∈ Jj . Note that as xj ≤ xl then Jj ⊆ Ij ⊆ Il . We
will explicitly construct such polynomials.
Such construction is very similar to the chordal elimination algorithm. The only
difference is that instead of elimination ideals we use resultants.
Initially, we assign fi0 = fi for 1 ≤ i ≤ n. Inductively, we construct the next
polynomials:
(
l
l
Resxl (f0l , fi+1
) if fi+1
involves xl
fil+1 =
l
l
if fi+1 does not involve xl
fi+1
for 1 ≤ i ≤ n − l, where we assume that f0l involves xl , possibly after rearranging
them. In the event that no fil involves xl , then we can ignore such variable. Notice that
Lemma 11 tells us that fil are all generic and simplicial.
We need to show that fil ∈ Jj , where xj is the largest variable of fil . We will prove
this by induction on l.
The base case is l = 0, where fi0 = fi are generators of I, and thus Lemma 6 says
that fi ∈ Jj .
Assume that the hypothesis holds for some l and consider some f := fil+1 . Let xj be
l
. By the induction hypothesis,
its largest variable. Consider first the case where f = fi+1
f ∈ Jj and we are done.
l
). In this case the largest variable of
Now consider the case that f = Resxl (f0l , fi+1
l
both f0l , fi+1
is xl and thus, using the induction hypothesis, both of them lie in Jl . Let
xp be the parent of xl . Using Equation (9) we get f ∈ eliml+1 (Jl ) ⊆ Jp . Let’s see
now that xj ≤ xp . The reason is that xj ∈ Xp , as f ∈ K[Xp ] and xj is its largest
variable. Thus we found an xp with xj ≤ xp < xl and f ∈ Jp . If xj = xp , we are
done. Otherwise, if xj < xp , let xr be the parent of xp . As f does not involve xp , then
f ∈ elimp+1 (Jp ) ⊆ Jr . In the same way as before we get that xj ≤ xr < xp and f ∈ Jr .
Note that we can repeat this argument again, until we get that f ∈ Jj . This concludes
the induction.
Lemma 12 can be used to show the domination condition and thus to certify that
chordal elimination succeeds. In particular, we can do this in the special case that all
polynomials are simplicial, as we show now.
Theorem 14. Let I = hf1 , . . . , fs i be an ideal such that for each 1 ≤ i ≤ s, fi is generic
simplicial. Then chordal elimination succeeds.
Proof. For each l, let Tl be the largest subtree of T with minimal vertex xl . Equivalently,
Tl consists of all the descendants of xl . Let tl := |Tl | and let x(fj ) denote the largest
variable of fj . If for all xl there are at least tl generators fj with x(fj ) ∈ Tl then
Lemma 12 implies the domination condition and we are done. Otherwise, let xl be the
largest where this fails. The maximality of xl guarantees that elimination succeeds up
20
DIEGO CIFUENTES AND PABLO A. PARRILO
to such point, i.e., Im = elimm (I) for all xm ≥ xl . We claim that no equation of Il
involves xl and thus we can ignore it. Proving this claim will conclude the proof.
If xl is a leaf of T , then tl = 1, which means that no generator of I involves xl .
Otherwise, let xs1 , . . . , xsr be its children. Note that Tl = {xl } ∪ Ts1 ∪ . . . ∪ Tsr . We
know that there are at least tsi generators with x(fj ) ∈ Tsi for each si , and such bound
has to be exact as xl does not have such property. Thus for each si there are exactly
tsi generators with x(fj ) ∈ Tsi , and there is no generator with x(fj ) = xl . Then, for
each si , when we eliminate all the tsi variables in Tsi in the corresponding tsi equations
we must get the zero ideal, i.e., elimsi +1 (Jsi ) = 0. On the other hand, as there is no
generator with x(fj ) = xl , then all generators that involve xl are in some Tsi . But we
observed that the l-th elimination ideal in each Tsi is zero, so that Il does not involve
xl , as we wanted.
5. Elimination ideals of cliques
rad
Notation. We will write I = J whenever we have V(I) = V(J).
Algorithm 2 allows us to compute (or bound) the elimination ideals I∩K[xl , . . . , xn−1 ].
In this section we will show that once we compute such ideals, we can also compute
many other elimination ideals. In particular, we will compute the elimination ideals of
the maximal cliques of G.
We recall the definition of the cliques Xl from Equation (1). Let Hl := I ∩ K[Xl ]
be the corresponding elimination ideal. As any clique is contained in some Xl , we can
restrict our attention to computing Hl . In particular, all maximal cliques of the graph
are of the form Xl for some l.
The motivation behind these clique elimination ideals is to find sparse generators of
the ideal that are the closest to a Gröbner basis. Lex Gröbner bases can be very large,
and thus finding a sparse approximation to them might be much faster as will be seen
in Section 7. We attempt to find such “optimal” sparse representation by using chordal
elimination.
Specifically, S
let gbHl denote a lex Gröbner basis of each Hl . We argue that the
concatenation l gbHl constitutes such closest sparse representation. In particular, the
following proposition
S says that if there exists a lex Gröbner basis of I that preserves
the structure, then l gbHl is also one.
Proposition 15. Let I be an ideal with graph G and let gb be a lex Gröbner basis. Let
Hl denote the clique elimination ideals, and let gbHl be the corresponding
lex Gröbner
S
bases. If gb preserves the graph structure, i.e., G(gb) ⊆ G, then l gbHl is a lex Gröbner
basis of I.
Proof. It is clear that S
gbHl ⊆ Hl ⊆ I. Let m ∈ in(I) be some monomial, we just need
to show that m ∈ in( l gHl ). As in(I) = in(gb), we can restrict m to be the leading
monomial m = lm(p) of some p ∈ gb. By the assumption on gb, the variables of p are
in some clique Xl of G. Thus, p ∈ Hl so that m = lm(p) ∈ in(Hl ) = in(gbl ). This
concludes the proof.
Before computing Hl , we will show how to obtain elimination ideals of simpler sets.
These sets are determined by the elimination tree of the graph, and we will find the
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
21
corresponding elimination ideals in Section 5.1. After that we will come back to computing the clique elimination ideals in Section 5.2. Finally, we will elaborate more on
the relationship between lex Gröbner bases and clique elimination ideals in Section 5.3.
5.1. Elimination ideals of lower sets. We will show now how to find elimination
ideals of some simple sets of the graph, which depend on the elimination tree. To do
so, we recall that in chordal elimination we decompose Il = Jl + Kl+1 which allows us
to compute next Il+1 = eliml+1 (Jl ) + Kl+1 . Observe that
Il = Jl + Kl+1
= Jl + eliml+1 (Jl ) + Kl+1
= Jl + Il+1
= Jl + Jl+1 + Kl+2
= Jl + Jl+1 + eliml+2 (Jl+1 ) + Kl+2 .
Continuing this way we conclude:
(10)
Il = Jl + Jl+1 + · · · + Jn−1 .
We will obtain a similar summation formula for other elimination ideals apart from Il .
Consider again the elimination tree T . We present another characterization of it.
Proposition 16. Consider the directed acyclic graph (DAG) obtained by orienting the
edges of G with the order of its vertices. Then the elimination tree T corresponds to
the transitive reduction of such DAG. Equivalently, T is the Hasse diagram of the poset
associated to the DAG.
Proof. As T is a tree, it is reduced, and thus we just need to show that any arc from
the DAG corresponds to a path of T . Let xi → xj be an arc in the DAG, and observe
that being an arc is equivalent to xj ∈ Xi . Let xp be the parent of xi . Then Lemma 5
implies xj ∈ Xp , and thus xp → xj is in the DAG. Similarly if xr is the parent of xp
then xr → xj is another arc. By continuing this way we find a path xi , xp , xr , . . . in T
that connects xi → xj , proving that T is indeed the transitive reduction.
Definition 5.1. We say a set of variables Λ is a lower set if T |Λ is also a tree rooted in
xn−1 . Equivalently, Λ is a lower set of the poset associated to the DAG of Proposition 16.
Observe that {xl , xl+1 , . . . , xn−1 } is a lower set, as when we remove x0 , x1 , . . . we are
pruning some leaf of T . The following lemma gives a simple property of these lower
sets.
Lemma 17. If X is a set of variables such that G|X is a clique, then T |X is contained in
some branch of T . In particular, if xl > xm are adjacent, then any lower set containing
xl must also contain xm .
Proof. For the first part, note that the DAG induces a poset on the vertices, and
restricted to X we get a linear order. Thus, in the Hasse diagram X must be part of
a chain (branch). The second part follows by considering the clique X = {xl , xm } and
using the previous result.
22
DIEGO CIFUENTES AND PABLO A. PARRILO
The next lemma tells us how to obtain the elimination ideals of any lower set.
Lemma 18. Let I be an ideal, let V = V(I) and assume that the domination condition
holds for chordal elimination. Let Λ ⊆ {x0 , . . . , xn−1 } be a lower set. Then,
X
rad
I ∩ K[Λ] =
Ji
xi ∈Λ
and the corresponding variety is πΛ (V ), where πΛ : Kn → KΛ is the projection onto Λ.
Proof. See Appendix A.2.
5.2. Cliques elimination algorithm. Lemma 18 tells us that we can very easily obtain the elimination ideal of any lower set. We return now to the problem of computing
the elimination ideals of the cliques Xl , which we denoted as Hl . Before showing how
to get them, we need a simple lemma.
Lemma 19. Let G be a chordal graph and let X be a clique of G. Then there is a
perfect elimination ordering v0 , . . . , vn−1 of G such that the last vertices of the ordering
correspond to X, i.e., X = {vn−1 , vn−2 , . . . , vn−|X| } .
Proof. We can apply Maximum Cardinality Search (Algorithm 1) to the graph, choosing
at the beginning all the vertices of clique X. As the graph is chordal, this gives a reversed
perfect elimination ordering.
Theorem 20. Let I be a zero dimensional ideal with chordal graph G. Assume that
the domination condition holds for chordal elimination. Then we can further compute
rad
ideals Hl ∈ K[Xl ] such that Hl = I ∩ K[Xl ], preserving the structure.
Proof. We will further prove that the corresponding variety is πXl (V ), where V = V(I)
and πXl : Kn → KXl is the projection onto Xl . We proceed by induction on l.
rad
The base case is l = n−1. As chordal elimination is successful then In−1 = elimn−1 (I)
and the variety is πn−1 (V ), so we can set Hn−1 = In−1 .
Assume that we found Hm for all xm < xl . Let Λ be a lower set with largest element
xl . By Lemma 18, we can compute an ideal IΛ with V(IΛ ) = πΛ (V ). Note that Xl ⊆ Λ
because of Lemma 17. Thus, we should use as Hl the ideal IΛ ∩ K[Xl ]. Naturally, we
will use chordal elimination to approximate this ideal. For a reason that will be clear
later, we modify IΛ , appending to it the ideals Hr for all xr ∈ Λ \ {xl }. Observe that
this does not change the variety.
Consider the induced graph G|Λ , which is also chordal as G is chordal. Thus,
Lemma 19 implies that there is a perfect elimination ordering σ of G|Λ where the
last clique is Xl . We can now use Algorithm 2 in the ideal IΛ using such ordering of
the variables to find an ideal Hl that approximates IΛ ∩ K[Xl ]. We will show now that
rad
this elimination is successful and thus Hl = IΛ ∩ K[Xl ].
Let Xjσ ⊆ G|Λ denote the cliques as defined in (1) but using the new ordering σ in
σ
G|Λ . Similarly, let Ijσ = Jjσ + Kj+1
denote the clique decompositions used in chordal
elimination with such ordering. Let xm be one variable that we need to eliminate to
σ
obtain Hl , i.e., xm ∈ Λ \ Xl . Let’s assume that xm is such that Xm
is a maximal clique
σ
of G|Λ . As the maximal cliques do not depend on the ordering, it means that Xm
= Xr
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
23
rad
σ
for some xr < xl , and thus we already found an Hr with Hr = I ∩ K[Xm
]. Observe that
σ
Hr ⊆ Jm by recalling that we appended Hr to IΛ and using Lemma 6. As Hr is zero
σ
is also zero dimensional for all such xm . Therefore, Corollary 9
dimensional, then Jm
says that the domination condition holds and chordal elimination succeeds.
rad
Finally, let’s prove that Hl = I ∩ K[Xl ]. Observe that as the domination condition
holds in the elimination above (to get Hl ), then
V(Hl ) = V(IΛ ∩ K[Xl ]) = πXl (V(IΛ )).
As V(IΛ ) = πΛ (V ), we obtain that V(Hl ) = πXl (V ). On the other hand, we also have
Hl ⊆ I ∩ K[Xl ], so that
V(Hl ) ⊇ V(I ∩ K[Xl ]) ⊇ πXl (V ).
Therefore, the three terms above must be equal.
Observe that the above proof hints to an algorithm to compute Hl . However, the
proof depends on the choice of some lower set Λ for each xl . To avoid eliminations we
want to use a lower set Λ as small as possible. By making a good choice we can greatly
simplify the procedure and we get, after some observations made in Corollary 21, the
Algorithm 3. Note that this procedure recursively computes the clique elimination
ideals: for a given node xl it only requires Jl and the clique elimination ideal of its
parent xp .
Algorithm 3 Compute Elimination Ideals of Cliques
Input: An ideal I, given by generators with chordal graph G
rad
Output: Ideals Hl such that Hl = I ∩ K[Xl ]
1: procedure CliquesElim(I, G)
2:
get cliques X0 , . . . , Xn−1 of G
3:
get J0 , . . . , Jn−1 from ChordElim(I, G)
4:
Hn−1 = Jn−1
5:
for l = n − 2 : 0 do
6:
xp = parent of xl
7:
C = Xp ∪ {xl }
8:
IC = Hp + Jl
9:
order = MCS(G|C , start = Xl )
10:
Hl = ChordElim(ICorder , G|order
)
C
11:
return H0 , . . . , Hn−1
Corollary 21. Let I be a zero dimensional ideal with chordal graph G. Assume that
the domination condition holds for chordal elimination. Then Algorithm 3 correctly
rad
computes the clique elimination ideals, i.e., Hl = I ∩ K[Xl ].
Proof. We refer to the proof of Theorem 20. For a given xl , let xp be its parent and
let Pl denote the directed path in T from xl to the root xn−1 . It is easy to see that Pl
is a lower set, and that Pl = Pp ∪ {xl }. We will see that Algorithm 3 corresponds to
24
DIEGO CIFUENTES AND PABLO A. PARRILO
selecting the lower set Λ to be this Pl and reusing the eliminations performed to get Hp
when we compute Hl .
In the proof of Theorem 20, to get Hl we need a perfect elimination ordering (PEO)
σl of G|Λ that ends in Xl . This order σl determines the eliminations performed in IΛ .
Let σp be a PEO of G|Pp , whose last vertices are Xp . Let’s see that we can extend σp
to obtain the PEO σl of G|Pl . Let C := Xp ∪ {xl } and observe that Xl ⊆ C due to
Lemma 5, and thus Pl = Pp ∪ C. Let σC be a PEO of G|C whose last vertices are Xp
(using Lemma19). We will argue that the following ordering works:
σl := (σp \ Xp ) + σC .
By construction, the last vertices of σl are Xl , so we just need to show that it is
indeed a PEO of G|Pl . Let v ∈ Pl , and let Xvσl be the vertices adjacent to it that
follow v in σl . We need to show that Xvσl is a clique. There are two cases: v ∈ C or
v∈
/ C. If v ∈ C, then Xvσl is the same as with σC , so that it is a clique because σC is
a PEO. If v ∈
/ C, we will see that Xvσl is the same as with σp , and thus it is a clique.
Consider the partition Xvσl = (Xvσl \ Xp ) ∪ (Xvσl ∩ Xp ), and note the part that is not
in Xp only depends on σp . The part in Xp is just adj (v) ∩ Xp , i.e., its neighbors in
Xp , given that we put σC to the end of σl . Observe that the same happens for σp , i.e.,
σ
σ
Xv p ∩ Xp = adj (v) ∩ Xp , by construction of σp . Thus Xvσl = Xv p , as wanted.
The argument above shows that given any PEO of Pp and any PEO of C, we can
combine them into a PEO of Pl . This implies that the eliminations performed to obtain
Hp can be reused to obtain Hl , and the remaining eliminations correspond to G|C . Thus,
we can obtain this clique elimination ideals recursively, as it is done in Algorithm 3.
Computing a Gröbner basis for all maximal cliques in the graph might be useful as it
decomposes the system of equations into simpler ones. We can extract the solutions of
the system by solving the subsystems in each clique independently and “glueing” them.
We elaborate on this now.
Lemma 22. Let I be an ideal and let Hj = I ∩ K[Xj ] be the cliques elimination ideals.
Then,
I = H0 + H1 + · · · + Hn−1 .
Proof. As Hj ⊆ I for any xj , then H0 + · · · + Hn−1 ⊆ Il . On the other hand, let f ∈ I
be one of its generators. By definition of G, the variables of f must be contained in
some Xj , so we have f ∈ Hj . This implies I ⊆ H0 + · · · + Hn−1 .
Lemma 22 gives us a strategy to solve zero dimensional ideals. Note that Hj is also
zero dimensional. Thus, we can compute the elimination ideals of the maximal cliques,
we solve each Hj independently, and finally we can merge the solutions. We illustrate
that now.
Example 5.1. Let G be the blue/solid graph in Figure 1, and let I be given by:
x3i − 1 = 0,
x9 − 1 = 0
0≤i≤8
x2i + xi xj + x2j = 0,
(i, j) blue/solid edge
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
25
Note that the graph associated to the above equations is precisely G. However, to use
chordal elimination we need to consider the chordal completion G, which includes the
three green/dashed edges of Figure 1. In such completion, we identify seven maximal
cliques:
X0 = {x0 ,x6 , x7 }, X1 = {x1 , x4 , x9 }, X2 = {x2 , x3 , x5 }
X3 = {x3 , x5 , x7 , x8 }, X4 = {x4 , x5 , x8 , x9 }
X5 = {x5 , x7 , x8 , x9 }, X6 = {x6 , x7 , x8 , x9 }.
With Algorithm 3 we can find the associated elimination ideals. Some of them are:
H0 = hx0 + x6 + 1, x26 + x6 + 1, x7 − 1i
H5 = hx5 − 1, x7 − 1, x28 + x8 + 1, x9 − 1i
H6 = hx6 + x8 + 1, x7 − 1, x28 + x8 + 1, x9 − 1i.
Denoting ζ = e2πi/3 , the corresponding varieties are:
H0 : {x0 , x6 , x7 }
→ ζ, ζ 2 , 1 , ζ 2 , ζ, 1
H5 : {x5 , x7 , x8 , x9 } → {1, 1, ζ, 1} , 1, 1, ζ 2 , 1
H6 : {x6 , x7 , x8 , x9 } → ζ 2 , 1, ζ, 1 , ζ, 1, ζ 2 , 1 .
There are only two solutions to the whole system, one of them corresponds to the values
on the left and the other to the values on the right.
From the example above we can see that to obtain a solution of I we have to match
solutions from different cliques Hl . We can do this matching iteratively following the
elimination tree. Any partial solution is guaranteed to extend as the elimination was
successful. Let’s see now an example were this matching gets a bit more complex.
Example 5.2. Consider again the blue/solid graph in Figure 1, and let I be given by:
x4i − 1 = 0,
x9 − 1 = 0
0≤i≤8
x3i + x2i xj + xi x2j + x3j = 0,
(i, j) blue/solid edge
The graph (and cliques) is the same as in Example 5.1, but the variety this time is
larger. This time we have |V(H0 )| = 18, |V(H5 )| = 27, |V(H6 )| = 12. This numbers
are still small. However, when we merge all partial solutions we obtain |V(I)| = 528.
5.3. Lex Gröbner bases and chordal elimination. To finalize this section, we will
show the relationship between lex Gröbner bases of I and lex Gröbner bases of the clique
elimination ideals Hl . We will see that both of them share many structural properties.
This justifies our claim that these polynomials are the closest sparse representation of
I to a lex Gröbner basis. In some cases, the concatenation of the clique Gröbner bases
might already be a lex Gröbner basis of I. This was already seen in Proposition 15, and
we will see now another situation where this holds. In other cases, a lex Gröbner basis
can be much larger than the concatenation of the clique Gröbner bases. As we can find
26
DIEGO CIFUENTES AND PABLO A. PARRILO
Hl while preserving sparsity, we can outperform standard Gröbner bases algorithms in
many cases, as will be seen in Section 7.
We focus on radical zero dimensional ideals I. Note that this radicality assumption
is
√ not restrictive, as we have always been concerned with V(I), and we can compute
Hl for each l. We recall now that in many cases (e.g., generic coordinates) a radical
zero dimensional has a very special type of Gröbner bases. We say that I is in shape
position if the reduced lex Gröbner basis has the structure:
x0 − g0 (xn−1 ), x1 − g1 (xn−1 ), . . . , xn−2 − gn−2 (xn−1 ), gn−1 (xn−1 ).
We will prove later the following result for ideals in shape position.
Proposition 23. Let I be a radicalSzero dimensional ideal in shape position. Let gbHl
be a lex Gröbner basis of Hl . Then l gbHl is a lex Gröbner basis of I.
If the ideal is not in shape position, then the concatenation of such smaller Gröbner
bases might not be a Gröbner basis for I. Indeed, in many cases any Gröbner basis for
I is extremely large, while the concatenated polynomials gbHl are relatively small as
they preserve the structure. This will be seen
S in the application studied of Section 7.1,
where we will show how much simpler can l gbHl be compared to a full Gröbner basis.
Even when the ideal is not in shape position, the concatenated polynomials already
have some of the structure of a lex Gröbner basis of I, as we will show. Therefore, it is
usually simpler to find such Gröbner basis starting from such concatenated polynomials.
In fact, in Section 7.1 we show that by doing this we can compute a lex Gröbner basis
faster than a degrevlex Gröbner basis.
Theorem 24. Let I be a radical zero dimensional ideal. For each xl let gbIl and gbHl be
minimal lex Gröbner bases for the elimination ideals Il = eliml (I) and Hl = I ∩ K[Xl ].
Denoting deg to the degree, the following sets are equal:
DIl = {degxl (p) : p ∈ gbIl }
DHl = {degxl (p) : p ∈ gbHl }.
Proof. See Appendix A.2.
Corollary 25. Let I be a radical zero dimensional ideal, then for each xl we have that
xdl ∈ in(I) if and only if xdl ∈ in(Hl ), using lex ordering.
Proof. Let gbIl , gbHl be minimal lex Gröbner bases of Il , Hl . As I is zero dimensional
then there are dl , dH such that xldl is the leading monomial of some polynomial in gbIl
and xdl H is the leading monomial of some polynomial in gbHl . All we need to show
is that dl = dH . This follows by noting that dl = max{DIl } and dH = max{DHl },
following the notation from Theorem 24.
Proof of Proposition 23. As I is in shape position, then its initial ideal has the form
in(I) = hx0 , x1 , . . . , xn−2 , xdn−1 i
for some d. For each xl > xn−1 , Corollary 25 implies that gbHl contains some fl with
leading monomial xl . For xn−1 , the corollary says that there is a fn−1 ∈ gbHn−1 with
leading monomial xdn−1 . Then in(I) = hlm(f0 ), . . . , lm(fn−1 )i and as fl ∈ Hl ⊆ I, these
polynomials form a Gröbner basis of I.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
27
6. Complexity analysis
Solving systems of polynomials in the general case is hard even for small treewidth, as
it was shown in Example 1.1. Therefore, we need some additional assumptions to ensure
tractable computation. In this section we study the complexity of chordal elimination
for a special type of ideals where we can prove such tractability.
Chordal elimination shares the same limitations of other elimination methods. In
particular, for zero dimensional ideals its complexity is intrinsically related to the size
of the projection |πl (V(I))|. Thus, we will make certain assumptions on the ideal that
allow us to bound the size of this projection. The following concept will be the key.
Definition 6.1 (q-domination). We say that a polynomial f is (xi , q)-dominated if its
leading monomial has the form xdi for some d ≤ q. Let I = hf1 , . . . , fs i, we say that I
is q-dominated if for each xi there is a generator fj that is (xi , q)-dominated.
We will assume that I satisfies this q-dominated condition. Observe that Corollary 10
holds, and thus chordal elimination succeeds. Note that this condition also implies that
I is zero dimensional.
It should be mentioned that the q-dominated condition applies to finite fields. Let Fq
denote the finite field of size q. If we are interested in solving a system of equations in
Fq (as opposed to its algebraic closure) we can add the equations xqi − xi . Even more,
by adding such equations we obtain the radical ideal I(VFq (I)) [23].
We need to know the complexity of computing a lex Gröbner basis. To simplify the
analysis, we assume from now that the generators of the ideal have been preprocessed
to avoid redundancies. Specifically, we make the assumption that the polynomials have
been pseudo reduced so that no two of them have the same leading monomial and no
. Note that the latter assumption can be made because the
monomial is divisible by xq+1
i
ideal is q-dominated. These conditions allow us to bound the number of polynomials.
Lemma 26. Let I = hf1 , . . . , fs i be a preprocessed q-dominated ideal. Then s = O(q n ).
Proof. As I is q-dominated, for each 0 ≤ i < n there is a generator gi with leading
monomial xdi i with di ≤ q. The leading monomial of all generators, other than the
gi ’s, are not divisible by xqi . There are only q n monomials with degrees less that q
in any variable. As the leading monomials of the generators are different, the result
follows.
The complexity of computing a Gröbner basis for a zero-dimensional ideal is known
to be single exponential in n [30]. This motivates the following definition.
Definition 6.2. Let α be the smallest constant such that the complexity of computing
e αn ) for any (preprocessed) q-dominated ideal. Here O
e ignores
a Gröbner basis is O(q
polynomial factors in n.
A rough estimate of α is stated next. The proof in [23] is for the case of Fq , but the
only property that they use is that the ideal is q-dominated.
Proposition 27 ( [23]). Buchberger’s algorithm in a q-dominated ideal requires O(q 6n )
field operations.
28
DIEGO CIFUENTES AND PABLO A. PARRILO
We should mention that the complexity of Gröbner bases has been actively studied
and different estimates are available. For instance, Faugère et al. [19] show that for
e ω ), where D is the number of solutions and 2 <
generic ideals the complexity is O(D
ω < 3 is the exponent of matrix multiplication. Thus, if we only considered generic
polynomials we could interpret such condition as saying that α ≤ ω. However, even if
the generators of I are generic, our intermediate calculations are not generic and thus
we cannot make such an assumption.
Nevertheless, to obtain good bounds for chordal elimination we need a slightly
stronger condition than I being q-dominated. Let X1 , . . . , Xr denote the maximal
cliques of the graph G, and let
(11)
Ĥj = hf : f generator of I, f ∈ K[Xj ]i.
Note that Ĥj ⊆ I ∩ K[Xj ]. We assume that each (maximal) Ĥj is q-dominated. Note
that such condition is also satisfied in the case of finite fields. The following lemma
shows the reason why we need this assumption.
Lemma 28. Let I be such that for each maximal clique Xj the ideal Ĥj (as in (11)) is
q-dominated. Then in Algorithm 2 we have that Jl is q-dominated for any xl .
Proof. See Appendix A.3.
It should be mentioned that whenever we have a zero dimensional ideal I such that
each Ĥj is also zero dimensional, then the same results apply by letting q be the largest
degree in a Gröbner basis of any Ĥj .
We derive now complexity bounds, in terms of field operations, for chordal elimination
under the assumptions of Lemma 28. We use the following parameters: n is the number
of variables, s is the number of equations, κ is the clique number (or treewidth), i.e.,
the size of the largest clique of G.
Theorem 29.
the complexity
e ακ ). Here
O(nq
Let I be such that each (maximal) Ĥj is q-dominated. In Algorithm 2,
e + lq ακ ). We can find all elimination ideals in
of computing Il is O(s
e ignores polynomial factors in κ.
O
Proof. In each iteration there are essentially only two relevant operations: decomposing
Il = Jl + Kl+1 , and finding a Gröbner basis for Jl .
For each xl , Lemma 28 tells us that Jl is q-dominated. Thus, we can compute a
e ακ ). Here we assume that the initial s equations were
lex Gröbner basis of Jl in O(q
preprocessed, and note that the following equations are also preprocessed as they are
obtained from minimal Gröbner bases. To obtain Il we compute at most l Gröbner
e ακ ).
bases, which we do in O(lq
It just remains to bound the time of decomposing Il = Jl + Kl+1 . Note that if we do
this decomposition in a naive way we will need Θ(ls) operations. But we can improve
such bound easily. For instance, assume that in the first iteration we compute for every
generator fj the largest xl such that fj ∈ Jl . Thus fj will be assigned to Km+1 for all
e
xm > xl , and then it will be assigned to Jl . We can do this computation in O(s).
We
can repeat the same process for all polynomials p that we get throughout the algorithm.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
29
Let sl be the number of generators of eliml+1 (Jl ). Then we can do all decompositions
e + s0 + s1 + . . . + sl−1 ). We just need to bound sl .
in O(s
It follows from Lemma 26 that for each clique Xl , the size of any minimal Gröbner
basis of arbitrary polynomials in Xl is at most q κ + κ. As the number of generators of
e κ ).
eliml+1 (Jl ) is bounded by the size of the Gröbner basis of Jl ⊆ K[Xl ], then sl = O(q
e + lq κ ).
Thus, we can do all decompositions in O(s
Thus, the total cost to compute Il is
e + lq κ + lq ακ ) = O(s
e + lq ακ ).
O(s
e + nq ακ ). Note that as each of the original s
In particular, we can compute In−1 in O(s
equations is in some Xl , then Lemma 26 implies that s = O(nq κ ). Thus, we can find
e ακ ).
all elimination ideals in O(nq
Remark. Note that to compute the bound W of (7) we need to use chordal elimination
l times, so the complexity is O(ls + l2 q ακ ).
Corollary 30. Let I be such that each (maximal) Ĥj is q-dominated. The complexity of
e ακ ). Thus, we can also describe V(I) in O(nq
e ακ ). Here O
e ignores
Algorithm 3 is O(nq
polynomial factors in κ.
Proof. The first part of the algorithm is chordal elimination, which we can do in
O(nq ακ ), as shown above. Observe also that Maximum Cardinality Search runs in
linear time, so we can ignore it. The only missing part is to compute the elimination
ideas of IC , where C = Xp ∪ {xl }. As |C| ≤ κ + 1, then the cost of chordal elimination
e ακ ) = O(q
e ακ ). Thus the complexity of Algorithm 3 is still O(nq
e ακ ).
is O(κq
We now prove the second part. As mentioned before, the Corollary 10 applies for qdominated ideals, so all eliminations are successful. From Lemma 22 and the following
remarks we know that the elimination ideals Hl , found with Algorithm 3, give a natural
description of V(I).
The bounds above tell us that for a fixed κ, we can find all clique elimination ideals,
and thus describe the variety, in O(n). This is reminiscent to many graph problems
(e.g., Hamiltonian circuit, vertex colorings, vertex cover) which are NP-hard in general,
but are linear for fixed treewidth [6]. Similar results hold for some types of constraint
satisfaction problems [16]. These type of problems are said to be fixed parameter
tractable (FPT) with treewidth as the parameter.
Our methods provide an algebraic solution to some classical graph problems. In
particular, we show now an application of the bounds above for finding graph colorings. It is known that the coloring problem can be solved in linear time for bounded
treewidth [6]. We can prove the same result by encoding colorings into polynomials.
Corollary 31. Let G be a graph and Ḡ a chordal completion with largest clique of
e ακ ).
size κ. We can describe all q-colorings of G in O(nq
30
DIEGO CIFUENTES AND PABLO A. PARRILO
Proof. It is known that graph q-colorings can be encoded with the following system of
polynomials:
(12a)
xqi − 1 = 0,
i∈V
(12b)
xq−1
+ xq−2
xj + · · · + xi xq−2
+ xq−1
= 0,
i
i
j
j
(i, j) ∈ E
where V, E denote the vertices and edges, and where each color corresponds to a different
square root of unity [3,28]. Note that the ideal IG given by these equations satisfies the
q-dominated condition stated before. The chordal graph associated to such ideal is Ḡ.
The result follows from Corollary 30.
To conclude, we emphasize the differences between our results to similar methods in
graph theory and constraint satisfaction. First, note that for systems of polynomials we
do not know a priori a discrete set of possible solutions. And even if the variety is finite,
the solutions may not have a rational (or radical) representation. In addition, by using
Gröbner bases methods we take advantage of many well studied algebraic techniques.
Finally, even though our analysis here assumes zero dimensionality, we can use our
methods in underconstrained systems and, if they are close to satisfy the q-dominated
condition, they should perform well. Indeed, in Section 7.3 we test our methods on
underconstrained systems.
7. Applications
In this section we show numerical evaluations of the approach proposed in some
concrete applications. Our algorithms were implemented using Sage [41]. Gröbner
bases are computed with Singular’s interface [17], except when K = F2 for which we
use PolyBoRi’s interface [7]. Chordal completions of small graphs (n < 32) are found
using Sage’s vertex separation algorithm. The experiments are performed on an i7 PC
with 3.40GHz, 15.6 GB RAM, running Ubuntu 12.04.
We will show the performance of chordal elimination compared to the Gröbner bases
algorithms from Singular and PolyBoRi. In all the applications we give here chordal
elimination is successful because of the results of Section 4. It can be seen below that
in all the applications our methods perform better, as the problem gets bigger, than
the algorithms from Singular and PolyBoRi.
As mentioned before, chordal elimination shares some of the limitations as other
elimination methods and it performs the best under the conditions studied in Section 6.
We show two examples that meet such conditions in Sections 7.1 and 7.2. The first
case relates to the coloring problem, which was already mentioned in Corollary 31. The
second case is an application to cryptography, where we solve equations over the finite
field F2 .
After that, Sections 7.3 and 7.4 show cases where the conditions from Section 6 are
not satisfied. We use two of the examples from [34], where they study a similar chordal
approach for semidefinite programming relaxations (SDP). Gröbner bases are not as
fast as SDP relaxations, as they contain more information, so we use smaller scale
problems. The first example is the sensor localization problem and the second one is
given by discretizations of differential equations.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
31
7.1. Graph colorings. We consider Equations (12) for q-colorings of a graph, over
the field K = Q. We fix the graph G of Figure 4 and vary the number of colors q.
Such graph was considered in [28] to illustrate a characterization of uniquely colorable
graphs using Gröbner bases. We use a different ordering of the vertices that determines
a simpler chordal completion (the clique number is 5).
9
10
7
12
19
5
22
15
6
1
21
23
3
11
16
17
2
8
14
4
0
20
13
18
Figure 4. Graph with a unique 3-coloring [28].
Table 1 shows the performance of Algorithm 2 and Algorithm 3, compared to Singular’s default Gröbner basis algorithm using degrevlex order (lex order takes much
longer). It can be seen how the cost of finding a Gröbner basis increases very rapidly as
we increase q, as opposed to our approach. In particular, for q = 4 we could not find a
Gröbner basis after 60000 seconds (16.7 hours), but our algorithms ran in less than one
second. The underlying reason for such long time is the large size of the solution set
(number of 4-colorings), which is |V(I)| = 572656008. Therefore, it is expected that
the size of its Gröbner basis is also very large. On the other hand, the projection on
each clique is much smaller, |V(Hl )| ≤ 576, and thus the corresponding Gröbner bases
(found with Algorithm 3) are also much simpler.
Table 1. Performance (in seconds) on Equations (12) (graph of Figure 4) for: Algorithm 2, Algorithm 3, computing a degrevlex Gröbner
basis with the original equations (Singular). One experiment was interrupted after 60000 seconds.
q
Variables
Equations
Monomials
ChordElim
CliquesElim
DegrevlexGB
2
24
69
49
0.058
0.288
0.001
3
24
69
94
0.141
0.516
5.236
4
24
69
139
0.143
0.615
> 60000
5
24
69
184
0.150
0.614
-
6
24
69
229
0.151
0.638
-
We repeat the same experiments, this time with the blue/solid graph of Figure 1.
Table 2 shows the results obtained. This time we also show the cost of computing a
32
DIEGO CIFUENTES AND PABLO A. PARRILO
lex Gröbner basis, using as input the clique elimination ideals Hl . Again, we observe
that chordal elimination is much faster than finding a Gröbner basis. We also see that
we can find faster a lex Gröbner basis than for degrevlex, by making use of the output
from chordal elimination.
Table 2. Performance (in seconds) on Equations (12) (blue/solid graph
of Figure 1) for: Algorithm 2, Algorithm 3, computing a lex Gröbner
basis with input Hl , and computing a degrevlex Gröbner basis with the
original equations (Singular).
q
Vars
Eqs
Mons
ChordElim
CliquesElim
LexGB from Hl
DegrevlexGB
5
10
28
75
0.035
0.112
0.003
0.003
10
10
28
165
0.044
0.130
0.064
0.202
15
10
28
255
0.065
0.188
4.539
8.373
20
10
28
345
0.115
0.300
73.225
105.526
7.2. Cryptography. We consider the parametric family SR(n, r, c, e) of AES variants
from [8]. Such cypher can be embedded into a structured system of polynomials equations over K = F2 as shown in [8]. Note that as the field is finite the analysis from
Section 6 holds.
We compare the performance of Algorithm 2 to PolyBoRi’s default Gröbner bases
algorithm, using both lex and degrevlex order. As the input to the cipher is probabilistic, for the experiments we seed the pseudorandom generator in fixed values of 0, 1, 2.
We fix the values r = 1, c = 2, e = 4 for the experiments and we vary the parameter n,
which corresponds to the number of identical blocks used for the encryption.
Table 3 shows the results of the experiments. We observe that for small problems
standard Gröbner bases outperform chordal elimination, particularly using degrevlex
order. Nevertheless, chordal elimination scales better, being faster than both methods
for n = 10. In addition, standard Gröbner bases have higher memory requirements,
which is reflected in the many experiments that aborted for this reason.
7.3. Sensor Network Localization. We consider the sensor network localization
problem, also called graph realization problem, given by the equations:
(13a)
kxi − xj k2 = d2ij
(i, j) ∈ A
(13b)
kxi − ak k2 = e2ik
(i, k) ∈ B
where x1 , . . . , xn are unknown sensor positions, a1 , . . . , am are some fixed anchors, and
A, B are some sets of pairs which correspond to sensors that are close enough. We
consider the problem over the field K = Q. Observe that the set A determines the graph
structure of the system of equations. Note also that the equations are simplicial (see
Definition 4.3) and thus Theorem 14 says that chordal elimination succeeds. However,
the conditions from Section 6 are not satisfied.
We generate random test problems in a similar way as in [34]. First we generate
n = 20 random sensor locations x∗i from the unit square [0, 1]2 . The m = 4 fixed
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
33
Table 3. Performance (in seconds) on the equations of SR(n, 1, 2, 4)
for: Algorithm 2, and computing (lex/degrevlex) Gröbner bases (PolyBoRi). Three different experiments (seeds) are considered for each n.
Some experiments aborted due to insufficient memory.
n
Variables
Equations
4
120
216
6
8
10
176
232
288
320
424
528
Seed
ChordElim
LexGB
DegrevlexGB
0
517.018
217.319
71.223
1
481.052
315.625
69.574
2
507.451
248.843
69.733
0
575.516
402.255
256.253
1
609.529
284.216
144.316
2
649.408
258.965
133.367
0
774.067
1234.094
349.562
1
771.927
> 1500, aborted
369.445
2
773.359
1528.899
357.200
0
941.068
> 1100, aborted
1279.879
1
784.709
> 1400, aborted
1150.332
2
1124.942
> 3600, aborted
> 2500, aborted
anchors are (1/2 ± 1/4, 1/2 ± 1/4). We fix a proximity threshold D which we set to
either D = 1/4 or D = 1/3. Set A is such that every sensor is adjacent to at most
3 more sensors and kxi − xj k ≤ D. Set B is such that every anchor is related to all
sensors with kxi − ak k ≤ D. For every (i, j) ∈ A and (i, k) ∈ B we compute dij , eik .
We compare the performance of Algorithm 2 and Singular’s algorithms. We consider
Singular’s default Gröbner bases algorithms with both degrevlex and lex orderings, and
FGLM algorithm if the ideal is zero dimensional.
We use two different values for the proximity threshold D = 1/4 and D = 1/3.
For D = 1/4 the system of equations is underconstrained (positive dimensional), and
for D = 1/3 the system is overconstrained (zero dimensional). We will observe that
in both cases chordal elimination performs well. Degrevlex Gröbner bases perform
slightly better in the overconstrained case, and poorly in the underconstrained case.
Lex Gröbner bases do not compete with chordal elimination in either case.
Table 4 summarizes the results obtained. We used 50 random instances for the
underconstrained case (D = 1/4) and 100 for the overconstrained case (D = 1/3). We
can see that in the underconstrained case neither lex or degrevelex Gröbner bases ever
finished within 1000 seconds. On the other hand, chordal elimination completes more
than half of the instances. For the overconstrained case, lex Gröbner basis algorithm
continues to perform poorly. On the other hand, degrevlex Gröbner bases and the
FGLM algorithm have slightly better statistics than chordal elimination.
Despite the better statistics of degrevlex and FGLM in the overconstrained case, one
can identify that for several of such instances chordal elimination performs much better.
34
DIEGO CIFUENTES AND PABLO A. PARRILO
Table 4. Statistics of experiments performed on random instances of
Equations (13). We consider two situations: 50 cases of underconstrained
systems (D = 1/4) and 100 cases of overconstrained systems (D = 1/3).
Experiments are interrupted after 1000 seconds.
D
1/4
1/3
Repet.
50
100
Vars
Eqs
40
39 ± 5
48 ± 6
40
ChordElim
LexGB
DegrevlexGB
LexFGLM
478.520
1000
1000
-
Mean time (s)
56%
0%
0%
-
Completed
298.686
1000
219.622
253.565
Mean time (s)
73%
0%
81%
77%
Completed
This can be seen in Figure 5, where we observe the scatter plot of the performance of
both FGLM and Algorithm 2. In about half of the cases (48) both algorithms are within
one second and for the rest: in 29 cases FGLM is better, in 23 chordal elimination is
better. To understand the difference between these two groups, we can look at the clique
number of the chordal completions. Indeed, the 23 cases where chordal elimination is
better have a mean clique number of 5.48, compared to 6.97 of the 29 cases where
FGLM was better. This confirms that chordal elimination is a suitable method for
cases with chordal structure, even in the overconstrained case.
Time in seconds of ChordElim (cubic scale)
1000
729
512
FGLM better
343
216
125
Within 1s
64
27
ChordElim better
8
1
0
0
1
8
64 125 216 343 512 729
27
Time in seconds of Singular's FGLM (cubic scale)
1000
Figure 5. Scatter plot of the time used by Singular’s FGLM and Algorithm 2 on 100 random overconstrained (D = 1/3) instances of Equations (13). Darker points indicate overlap.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
35
7.4. Differential Equations. We consider now the following equations over the field
K = Q:
1
(14a)
0 = 2x1 − x2 + h2 (x1 + t1 )3
2
1
0 = 2xi − xi−1 − xi+1 + h2 (xi + ti )3
(14b)
for i = 2, . . . , n − 1
2
1
0 = 2xn − xn−1 + h2 (xn + tn )3
(14c)
2
with h = 1/(n + 1) and ti = i/(n + 1). Such equations were used in [34], and arise from
discretizing the following differential equation with boundary conditions:
1
x00 + (x + t)3 = 0,
x(0) = x(1) = 0.
2
Note that these polynomials are simplicial (see Definition 4.3) and thus chordal elimination succeeds because of Theorem 14. Even more, the equations Jl obtained in
chordal elimination form a lex Gröbner basis. However, the results from Section 6
do not hold. Nevertheless, we compare the performance of chordal elimination with
Singular’s default Gröbner basis algorithm with lex order. We also consider Singular’s
FGLM implementation.
Table 5. Performance (in seconds) on Equations (14) for: Algorithm 2,
and computing a lex Gröbner basis with two standard methods (Singular’s default and FGLM).
n Variables
Equations
ChordElim
LexGB
LexFGLM
3
3
3
0.008
0.003
0.007
4
4
4
0.049
0.044
0.216
5
5
5
1.373
1.583
8.626
6
6
6
76.553
91.155
737.989
7
7
7
7858.926
12298.636
43241.926
Table 5 shows the results of the experiments. The fast increase in the timings observed is common to all methods. Nevertheless, it can be seen that chordal elimination
performs faster and scales better than standard Gröbner bases algorithms. Even though
the degrevlex term order is much simpler in this case, FGLM algorithm is not efficient
to obtain a lex Gröbner basis.
References
[1] S. Arnborg, D. G. Corneil, and A. Proskurowski. Complexity of finding embeddings in a k-tree.
SIAM Journal on Algebraic Discrete Methods, 8(2):277–284, 1987.
[2] G. V. Bard, N. T. Courtois, and C. Jefferson. Efficient methods for conversion and solution of
sparse systems of low-degree multivariate polynomials over GF(2) via SAT-solvers. Cryptology
ePrint Archive, Report 2007/024, 2007.
36
DIEGO CIFUENTES AND PABLO A. PARRILO
[3] D. A. Bayer. The Division Algorithm and the Hilbert Scheme. PhD thesis, Harvard University,
Cambridge, MA, USA, 1982. AAI8222588.
[4] C. Beeri, R. Fagin, D. Maier, and M. Yannakakis. On the desirability of acyclic database schemes.
Journal of the ACM (JACM), 30(3):479–513, 1983.
[5] J. Blair and B. Peyton. An introduction to chordal graphs and clique trees. In Graph theory and
sparse matrix computation, pages 1–29. Springer, 1993.
[6] H. L. Bodlaender and A. Koster. Combinatorial optimization on graphs of bounded treewidth.
The Computer Journal, 51(3):255–269, 2008.
[7] M. Brickenstein and A. Dreyer. PolyBoRi: A framework for Gröbner-basis computations with
boolean polynomials. Journal of Symbolic Computation, 44(9):1326 – 1345, 2009. Effective Methods in Algebraic Geometry.
[8] C. Cid, S. Murphy, and M. Robshaw. Small scale variants of the AES. In Fast Software Encryption,
pages 145–162. Springer, 2005.
[9] B. Courcelle and J. Engelfriet. Graph structure and monadic second-order logic: a languagetheoretic approach, volume 138. Cambridge University Press, 2012.
[10] D. A. Cox, J. Little, and D. O’Shea. Ideals, varieties, and algorithms: an introduction to computational algebraic geometry and commutative algebra. Springer, 2007.
[11] V. Dalmau, P. G. Kolaitis, and M. Y. Vardi. Constraint satisfaction, bounded treewidth, and
finite-variable logics. In Principles and Practice of Constraint Programming-CP 2002, pages 310–
326. Springer, 2002.
[12] T. A. Davis, J. R. Gilbert, S. I. Larimore, and E. G. Ng. A column approximate minimum degree
ordering algorithm. ACM Transactions on Mathematical Software (TOMS), 30(3):353–376, 2004.
[13] J. A. De Loera, J. Lee, S. Margulies, and S. Onn. Expressing combinatorial problems by systems of
polynomial equations and Hilbert’s nullstellensatz. Comb. Probab. Comput., 18(4):551–582, July
2009.
[14] J. A. De Loera, S. Margulies, M. Pernpeintner, E. Riedl, D. Rolnick, G. Spencer, D. Stasi, and
J. Swenson. Graph-coloring ideals: Nullstellensatz certificates, Gröbner bases for chordal graphs,
and hardness of Gröbner bases. In Proceedings of the 40th International Symposium on Symbolic
and Algebraic Computation, ISSAC’15, pages 133–140. ACM, 2015.
[15] R. Dechter. Bucket elimination: A unifying framework for probabilistic inference. In Learning in
graphical models, pages 75–104. Springer, 1998.
[16] R. Dechter. Constraint processing. Morgan Kaufmann, 2003.
[17] W. Decker, G. M. Greuel, G. Pfister, and H. Schönemann. Singular 3-1-6 — A computer algebra
system for polynomial computations. http://www.singular.uni-kl.de, 2012.
[18] I. Z. Emiris and J. F. Canny. Efficient incremental algorithms for the sparse resultant and the
mixed volume. Journal of Symbolic Computation, 20(2):117–149, 1995.
[19] J. C. Faugère, P. Gaudry, L. Huot, and G. Renault. Polynomial systems solving by fast linear
algebra. arXiv preprint arXiv:1304.6039, 2013.
[20] J. C. Faugère and S. Rahmany. Solving systems of polynomial equations with symmetries using SAGBI-Gröbner bases. In Proceedings of the 34th International Symposium on Symbolic and
Algebraic Computation, ISSAC’09, pages 151–158. ACM, 2009.
[21] J. C. Faugère, M. Safey El Din, and P. J. Spaenlehauer. Gröbner bases of bihomogeneous ideals
generated by polynomials of bidegree (1, 1): Algorithms and complexity. Journal of Symbolic
Computation, 46(4):406–437, 2011.
[22] J. C. Faugère, P. J. Spaenlehauer, and J. Svartz. Sparse Gröbner bases: The unmixed case.
In Proceedings of the 39th International Symposium on Symbolic and Algebraic Computation,
ISSAC’14, pages 178–185. ACM, 2014.
[23] S. Gao. Counting Zeros over Finite Fields Using Gröbner Bases. PhD thesis, MS Thesis in Logic
and Computation, 2009.
[24] S. Gao, V. M. Rodrigues, and J. Stroomer. Gröbner basis structure of finite sets of points. preprint,
2003.
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
37
[25] K. Gatermann. Symbolic solution of polynomial equation systems with symmetry. In Proceedings
of the International Symposium on Symbolic and Algebraic Computation, ISSAC’90, pages 112–
119. ACM, 1990.
[26] M. C. Golumbic. Algorithmic graph theory and perfect graphs, volume 57. Elsevier, 2004.
[27] J. Herzog, T. Hibi, F. Hreinsdóttir, T. Kahle, and J. Rauh. Binomial edge ideals and conditional
independence statements. Advances in Applied Mathematics, 45(3):317 – 333, 2010.
[28] C. J. Hillar and T. Windfeldt. Algebraic characterization of uniquely vertex colorable graphs.
Journal of Combinatorial Theory, Series B, 98(2):400–414, 2008.
[29] B. Huber and B. Sturmfels. A polyhedral method for solving sparse polynomial systems. Mathematics of computation, 64(212):1541–1555, 1995.
[30] Y. N. Lakshman. On the complexity of computing a Gröbner basis for the radical of a zero
dimensional ideal. In Proceedings of the 22nd annual ACM Symposium on Theory of Computing,
pages 555–563. ACM, 1990.
[31] S. L. Lauritzen and D. J. Spiegelhalter. Local computations with probabilities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society. Series B
(Methodological), pages 157–224, 1988.
[32] T. Y. Li. Numerical solution of multivariate polynomial systems by homotopy continuation methods. Acta Numerica, 6:399–436, 1997.
[33] J. A. Makowsky and K. Meer. Polynomials of bounded treewidth. In F. Cucker and M. Rojas,
editors, Foundations of Computational Mathematics, Proceedings of the Smalefest 2000, pages
211–250. World Scientific, 2002.
[34] J. Nie and J. Demmel. Sparse SOS relaxations for minimizing functions that are summations of
small polynomials. SIAM Journal on Optimization, 19(4):1534–1558, 2008.
[35] S. Parter. The use of linear graphs in Gauss elimination. SIAM review, 3(2):119–130, 1961.
[36] V. I. Paulsen, S. C. Power, and R. R. Smith. Schur products and matrix completions. Journal of
functional analysis, 85(1):151–178, 1989.
[37] A. Pothen and S. Toledo. Elimination structures in scientific computing. In D. Mehta and S. Sahni,
editors, Handbook on Data Structures and Applications, pages 59–1–59–29. CRC Press, 2004.
[38] H. Raddum and I. Semaev. New technique for solving sparse equation systems. IACR Cryptology
ePrint Archive, 2006:475, 2006.
[39] D. J. Rose. Triangulated graphs and the elimination process. Journal of Mathematical Analysis
and Applications, 32(3):597–609, 1970.
[40] A. J. Sommese and C. W. Wampler. The Numerical solution of systems of polynomials arising in
engineering and science, volume 99. World Scientific, 2005.
[41] W. A. Stein et al. Sage Mathematics Software. The Sage Development Team, 2014.
http://www.sagemath.org.
[42] B. Sturmfels. Gröbner bases and convex polytopes, volume 8. American Mathematical Soc., 1996.
[43] L. Vandenberghe and M. S. Andersen. Chordal graphs and semidefinite optimization. Foundations
and Trends in Optimization, 1(4):241–433, 2014.
[44] R. Villarreal. Cohen-Macaulay graphs. Manuscripta Mathematica, 66(1):277–293, 1990.
[45] R. Villarreal. Monomial algebras, volume 8. CRC Press, 2015.
Appendix A. Additional proofs
A.1. Proofs from Section 3.
Proof of Theorem 3. We prove it by induction on l. The base case is Lemma 2. Assume
that the result holds for some l and let’s show it for l + 1.
38
DIEGO CIFUENTES AND PABLO A. PARRILO
By induction hypothesis Il , W1 , . . . , Wl satisfy Equation (4) and Equation (5). Lemma 2
with Il as input tell us that Il+1 , Wl+1 satisfy:
π(V(Il )) ⊆ V(Il+1 )
(15)
V(Il+1 ) − V(Wl+1 ) ⊆ π(V(Il ))
(16)
where π : Kn−l → Kn−l−1 is the projection onto the last factor. Then,
πl+1 (V(I)) = π(πl (V(I))) ⊆ π(V(Il )) ⊆ V(Il+1 )
and as V(Il+1 ) is closed, we can take the closure. This shows Equation (4).
We also have
πl+1 (V(I)) = π(πl (V(I))) ⊇ π(V(Il ) − [πl (V(W1 )) ∪ · · · ∪ πl (V(Wl ))])
⊇ π(V(Il )) − π[πl (V(W1 )) ∪ · · · ∪ πl (V(Wl ))]
= π(V(Il )) − [πl+1 (V(W1 )) ∪ · · · ∪ πl+1 (V(Wl ))]
⊇ (V(Il+1 ) − V(Wl+1 )) − [πl+1 (V(W1 )) ∪ · · · ∪ πl+1 (V(Wl ))]
= V(Il+1 ) − [πl+1 (V(W1 )) ∪ · · · ∪ πl+1 (V(Wl+1 ))]
which proves Equation (5).
A.2. Proofs from Section 5.
P
Proof of Lemma 18. Let HΛ := I ∩ K[Λ] and JΛ := xi ∈Λ Ji . Let xl ∈ Λ be its largest
element. For a fixed xl , we will show by induction on |Λ| that V(HΛ ) = V(JΛ ) = πΛ (V ).
The base case is when Λ = {xl , . . . , xn−1 }. Note that as xl is fixed, such Λ is indeed
the largest possible lower set. In such case JΛ = Il as seen in Equation (10), and as we
are assuming that the domination condition holds, then V(HΛ ) = V(Il ) = πl (V ).
Assume that the result holds for k + 1 and let’s show it for some Λ with |Λ| = k.
Consider the subtree Tl = T |{xl ,...,xn−1 } of T . As Tl |Λ is a proper subtree of Tl with
the same root, there must be an xm < xl with xm ∈
/ Λ and such that xm is a leaf in
Tl |Λ0 , where Λ0 = Λ ∪ {xm }. We apply the induction hypothesis in Λ0 , obtaining that
V(HΛ0 ) = V(JΛ0 ) = πΛ0 (V ).
Note now that Jm is is a subset of both HΛ0 , JΛ0 . Observe also that we want to
eliminate xm from these ideals to obtain HΛ , JΛ . To do so, let’s change the term order
to xm > xl > xl+1 > · · · > xn−1 . Note that such change has no effect inside Xm , and
thus the term ordering for Jm remains the same. As the domination condition holds,
then Jm is xm -dominated, and thus HΛ0 , JΛ0 are also xm -dominated. This means that
Lemma 8 holds for HΛ0 , JΛ0 when we eliminate xm , and then
V(elimm+1 (HΛ0 )) = πm+1 (V(HΛ0 )) = πΛ (V )
V(elimm+1 (JΛ0 )) = πm+1 (V(JΛ0 )) = πΛ (V ).
rad
Notice that HΛ = elimm+1 (HΛ0 ), so all we have to do now is to show that JΛ =
elimm+1 (JΛ0 ). Note that
elimm+1 (JΛ0 ) = elimm+1 (Jm + JΛ ).
CHORDAL STRUCTURE IN POLYNOMIAL IDEALS
39
Observe that the last expression is reminiscent of Lemma 2, but in this case we are
eliminating xm . As mentioned before, Jm is xm -dominated, so elimination succeeds.
Therefore, we have
rad
elimm+1 (Jm + JΛ ) = elimm+1 (Jm ) + JΛ .
Let xp be the parent of xm in T . Then Equation (9) says that elimm+1 (Jm ) ⊆ Jp , where
we are using that the term order change maintains Jm . Observe that xp ∈ Λ by the
construction of xm , and then Jp ⊆ JΛ . Then,
elimm+1 (Jm ) + JΛ = JΛ .
Combining the last three equations we complete the proof.
For the proof of Theorem 24, we will need two results before.
Lemma 32. Let I be a zero dimensional ideal, let Hj = I ∩ K[Xj ] and let Il = eliml (I).
Then,
rad
Il = Hl + Hl+1 + · · · + Hn−1 .
S
Proof. For each xj let gbHj be a lex Gröbner basis of Hj . Let F = xj gbHj be the
concatenation of all gbHj ’s. Then the decomposition of I from Lemma 22 says that
I = hF i. Observe now that if we use chordal elimination on F , at each step we only
remove the polynomials involving some variable; we never generate a new S
polynomial.
Therefore our approximation of the l-th elimination ideal is given by Fl = xj ≤xl gbHj .
Note now that as Hj is zero dimensional it is also xj -dominated, and thus Corollary 10
P
rad
says that elimination succeeds. Thus Il = hFl i = xj ≤xl Hj .
Theorem 33 ( [24]). Let I be a radical zero dimensional ideal and V = V(I). Let gb
be a minimal Gröbner basis with respect to an elimination order for x0 . Then the set
D = {degx0 (p) : p ∈ gb}
where deg denotes the degree, is the same as
F = {|π −1 (z) ∩ V | : z ∈ π(V )}
where π : Kn → Kn−1 is the projection eliminating x0 .
Proof of Theorem 24. If xl = xn−1 , then Il = Hl and the assertion holds. Otherwise,
note that Il , Hl are also radical zero dimensional so we can use Theorem 33. Let
FIl = {|πI−1
(z) ∩ V(Il )| : z ∈ πIl (V(Il ))}
l
−1
FHl = {|πH
(z) ∩ V(Hl )| : z ∈ πHl (V(Hl ))}
l
where πIl : Kn−l → Kn−l−1 and πHl : K|Xl | → K|Xl |−1 are projections eliminating xl .
Then we know that DIl = FIl and DHl = FHl , so we need to show that FIl = FHl .
For some z ∈ Kn−l , let’s denote z =: (zl , zH , zI ) where zl is the xl coordinate, zH are
the coordinates of Xl \ xl , and zI are the coordinates of {xl , . . . , xn−1 } \ Xl . Thus, we
have πIl (z) = (zH , zI ) and πHl (zl , zH ) = zH .
40
DIEGO CIFUENTES AND PABLO A. PARRILO
rad
As I is zero dimensional, then Lemma 32 implies that Il = Hl + Il+1 . Note also that
V(Il+1 ) = πIl (V(Il )) as it is zero dimensional. Then,
z ∈ V(Il ) ⇐⇒ (zl , zH ) ∈ V(Hl ) and (zH , zI ) ∈ πIl (V(Il )).
Thus, for any (zH , zI ) ∈ πIl (V(Il )) we have
(zl , zH , zI ) ∈ V(Il ) ⇐⇒ (zl , zH ) ∈ V(Hl ).
Equivalently, for any (zH , zI ) ∈ πIl (V(Il )) we have
(17)
−1
(zH ) ∩ V(Hl )
(zH , zI ) ∩ V(Il ) ⇐⇒ ρ(z) ∈ πH
z ∈ πI−1
l
l
where ρ(zl , zH , zI ) := (zl , zH ). Therefore, FIl ⊆ FHl .
On the other hand, note that if zH ∈ πHl (V(Hl )), then there is some zI such that
(zH , zI ) ∈ πIl (V(Il )). Thus, for any zH ∈ πHl (V(Hl )) there is some zI such that (17)
holds. This says that FHl ⊆ FIl , completing the proof.
A.3. Proofs from Section 6.
Proof of Lemma 28. Let xl be arbitrary and let xm ∈ Xl . We want to find a generator
of Jl that is (xm , q)-dominated. Let xj ≥ xl be such that Xl ⊆ Xj and Xj is a maximal
clique. Note that xm ∈ Xj . Observe that Ĥj ⊆ Jj because of Lemma 6, and thus Jj is
q-dominated. Then there must be a generator f ∈ Jj that is (xm , q)-dominated.
Let’s see that f is a generator of Jl , which would complete the proof. To prove this
we will show that f ∈ K[Xl ], and then the result follows from Lemma 6. As the largest
variable of f is xm , then all its variables are in Xj \{xm+1 , . . . , xj } ⊆ Xj \{xl+1 , . . . , xj }.
Thus, it is enough to show that
Xj \ {xl+1 , xl+2 , . . . , xj } ⊆ Xl .
The equation above follows by iterated application of Lemma 5, as we will see. Let
xp be the parent of xj in T , and observe that xl ∈ Xp as xl ≤ xp and both are in clique
Xj . Then Lemma 5 implies that Xj \ {xp+1 , . . . , xj } ⊆ Xp . If xp = xl , we are done.
Otherwise, let xr be the parent of xp , and observe that xl ∈ Xr as before. Then,
Xj \ {xr+1 , . . . , xj } ⊆ Xp \ {xr+1 , . . . , xp } ⊆ Xr .
If xr = xl , we are done. Otherwise, we can continue this process that has to eventually
terminate. This completes the proof.
Laboratory for Information and Decision Systems (LIDS), Massachusetts Institute
of Technology, Cambridge MA 02139, USA
E-mail address: [email protected]
Laboratory for Information and Decision Systems (LIDS), Massachusetts Institute
of Technology, Cambridge MA 02139, USA
E-mail address: [email protected]
| 0math.AC
|
Accepted for publication in IEEE International Symposium on Biomedical Imaging (ISBI) 2018
DEEP MULTI-MODAL CLASSIFICATION OF INTRADUCTAL PAPILLARY MUCINOUS
NEOPLASMS (IPMN) WITH CANONICAL CORRELATION ANALYSIS
Sarfaraz Hussein1 , Pujan Kandel2 , Juan E. Corral2 , Candice W. Bolan2 ,
Michael B. Wallace2 and Ulas Bagci1
arXiv:1710.09779v2 [cs.CV] 9 Jan 2018
1
Center for Research in Computer Vision (CRCV), University of Central Florida, Orlando, FL.
2
Mayo Clinic, Jacksonville, FL.
ABSTRACT
Pancreatic cancer has the poorest prognosis among all cancer
types. Intraductal Papillary Mucinous Neoplasms (IPMNs)
are radiographically identifiable precursors to pancreatic cancer; hence, early detection and precise risk assessment of
IPMN are vital. In this work, we propose a Convolutional
Neural Network (CNN) based computer aided diagnosis
(CAD) system to perform IPMN diagnosis and risk assessment by utilizing multi-modal MRI. In our proposed approach, we use minimum and maximum intensity projections
to ease the annotation variations among different slices and
type of MRIs. Then, we present a CNN to obtain deep
feature representation corresponding to each MRI modality (T1-weighted and T2-weighted). At the final step, we
employ canonical correlation analysis (CCA) to perform a
fusion operation at the feature level, leading to discriminative
canonical correlation features. Extracted features are used for
classification. Our results indicate significant improvements
over other potential approaches to solve this important problem. The proposed approach doesn’t require explicit sample
balancing in cases of imbalance between positive and negative examples. To the best of our knowledge, our study is
the first to automatically diagnose IPMN using multi-modal
MRI.
Index Terms— Pancreatic cancer, IPMN, Magnetic Resonance Imaging (MRI), Deep learning, Canonical Correlation
Analysis (CCA)
1. INTRODUCTION
Cancer is one of the main causes of death in the world with
a mortality rate of 171.2 per 100,000 people per year (based
on 2008-2012 stats) [1]. Among all cancers, pancreatic cancer has the poorest prognosis with a 5-year survival rate of
just 7% in the United States [1]. To address the problem of
automatic diagnosis of pancreatic cancer, we propose a new
CAD framework for Intraductal Papillary Mucinous Neoplasms (IPMN). IPMN is a mucin-producing neoplasm found
in the main and branch pancreatic ducts. They are radiographically identifiable precursors to pancreatic cancer [2, 3].
If left untreated, they can progress into invasive cancer. For
instance, around one-third of resected IPMNs are found to be
associated with invasive carcinoma [4]. In 2012, Tanaka et
al. [5] published the International consensus guidelines for the
preoperative management of IPMN using radiographic and
clinical criteria. These guidelines can be used in the development of CAD approaches for the separation of IPMNs from
normal pancreas. The CAD approaches can yield to identify
important imaging bio-markers that may assist radiologists
for improved diagnosis, staging, and treatment planning.
In the literature, there are a limited number of studies
addressing the automatic diagnosis of IPMN using radiology
images. Hanania et al. [6] studied the contribution of numerous low-level imaging features such as texture, intensity, and
shape to perform low and high-grade IPMN classification.
In the approach by Gazit et al. [7] texture and component
enhancing features were extracted from the segmented cysts.
The process is then followed by a feature selection and classification framework. Both of these approaches [6, 7], however,
are evaluated on CT images and require the segmentation of
cysts or pancreas. In contrast to these methods, our approach
doesn’t require prior segmentation of cysts or pancreas and is
evaluated on multi-modal MRI scans rather than CT. In this
work, we hypothesize and evaluate the influence of complementary information in T1-weighted and T2-weighted scans
that can be utilized to perform improved diagnosis of IPMN.
Our Contributions:
• To the best of our knowledge, this is the first study to use
deep learning for the classification of IPMN.
• We employ multiple imaging modalities of MRI (T1 and
T2) and fuse the feature representation using Canonical
Correlation Analysis (CCA) to obtain better discrimination
between normal and subjects with IPMN. We also perform
further stratification of IPMN in low-grade and high-grade
categories.
• Extensive experimental evaluations are performed on a
dataset comprising 139 subjects, the largest study of IPMN
to date.
4096-D
T1-scan
Minimum Intensity
Projection
Convolutional Neural Network
…
Convolution
T2-scan
Canonical Correlation
Analysis
−1
−1
�Ф = 𝛬2 𝑊
�Ф
𝐶ФФ
𝐶Ф𝛹 𝐶𝛹𝛹
𝐶𝛹Ф 𝑊
4096-D
SVM
Classifier
−1
−1
�𝛹 = 𝛬2 𝑊
�𝛹
𝐶𝛹Ф 𝐶ФФ
𝐶Ф𝛹 𝑊
𝐶𝛹𝛹
Max-pooling
Maximum Intensity
Projection
Fig. 1. An overview of the proposed method. First, the minimum and maximum intensity projections are computed corresponding to T1 and T2 scans respectively. The intensity projections are then fed into a pre-trained Convolutional Neural Network
(CNN) to obtain feature representation. Canonical Correlation Analysis (CCA) based feature fusion is performed in order to
obtain discriminative and transformed feature representation. Finally, an SVM based classifier is employed to obtain the final
label (normal or IPMN).
2. MATERIALS
We evaluated our proposed approach for the classification of
IPMN on a dataset comprising post-contrast volumetric T1
and T2 MRI scans from 139 subjects. The scans were labeled
as normal or IPMN using pathology report obtained after
surgery. Out of 139 scans, 108 were from subjects diagnosed
with IPMN, whereas the rest of 31 subjects were normal.
The in-plane spacing (xy-plane) of T1-weighted scans was
ranging from 0.66 mm to 1.48 mm and that of T2-weighted
scans from 0.47 mm to 1.41 mm. For pre-processing, we first
apply N4 bias field correction [8] to each scan in order to
minimize intensity inhomogeneity. Next, we use a curvature
anisotropic image filter to smooth images while preserving
edge information. For each image, a single slice which has a
significant portion of the pancreas is annotated to be normal
or IPMN.
the other slices. In order to address these issues, we sample
k = 2 consecutive slices before and after Iu and Jv . Since the
input to the deep network is a 2D image, we use Maximum
and Minimum Intensity Projections to combine information
across various slices into a single slice. We employ minimum
intensity projection for T1 scans since IPMN and pancreatic
cysts are hypo-intensity regions in T1 scans. In contrast,
we use maximum intensity projection for T2 scans, because
IPMN and pancreatic cysts correspond to hyper-intensity regions in these scans. The intensity projections corresponding
to T1 and T2 scans can be represented as:
I 0 = min [I(X1 , Y1 , Z1 )],
Z1
0
J = max [J (X2 , Y2 , Z2 )],
(1)
Z2
3.1. CNN for Multi-modal Feature Representation:
where Z1 and Z2 consists of k slices around Iu and Jv
respectively. Moreover, I 0 and J 0 represent the intensity projections from T1 and T2 scans, respectively. The overview of
the proposed approach is shown in Figure 1
Problem Formulation:
Our proposed approach consists of inputs from two different
MRI image modalities T1 and T2. Let IT 1 = [I1 , I2 . . . IN1 ] ∈
RX1 ×Y1 ×N1 be the T1 scan whereas the corresponding T2
scan is represented as JT 2 = [J1 , J2 . . . JN2 ] ∈ RX2 ×Y2 ×N2 ,
with N1 and N2 number of slices.
Consider Iu be the slice with pancreas from T1 scan and
Jv be the slice from T2. Predicting the label from a single
slice, however, may yield to hypersensitivity in annotation
labels as well as miss important contextual information from
Network Architecture:
In order to obtain deep feature representation for our proposed
IPMN classification approach, we use (fast) CNN-F architecture trained on ImageNet [9]. The architecture consists of 5
convolutional and 3 fully-connected layers. The input 2D image is resized to 224 × 224. The first convolutional layer contains 64 filters with stride 4 and there are 256 filters with stride
1 in the other 4 convolutional layers. Our input to the network are the 2D intensity projections I 0 and J 0 , whereas the
features are extracted from the second fully connected layer
3. METHODS
without applying non-linearities such as ReLU (Rectified Linear Units). The features are `2 normalized to obtain the final
representation.
3.2. Feature Fusion with Canonical Correlation Analysis:
The next step is to combine information from the two imaging
modalities so as to improve the classification performance. As
these two imaging modalities (T1 and T2) have complementary information, the fusion of features from these modalities
can help improve IPMN diagnosis. Assume that Φ ∈ Rn×p
and Ψ ∈ Rn×q comprise the deep features from the intensity projections of n training images from T1 and T2 scans
respectively. Each sample has a corresponding binary label
given by Y = [y1 , y2 . . . yn ], where Y ∈ {0, 1}n×1 . Consider
CΦΦ ∈ Rp×p and CΨΨ ∈ Rq×q represent the within sets covariance matrices of Φ and Ψ respectively. Additionally, the
between set covariance matrix is referred as CΦΨ ∈ Rp×q
T
. The covariance matrix C can theresuch that CΨΦ = CΦΨ
fore be written as:
C=
cov(Φ)
cov(Φ, Ψ)
CΦΦ
=
cov(Ψ, Φ)
cov(Ψ)
CΨΦ
CΦΨ
CΨΨ
(2)
In this regard, CCA is employed to find the linear combinations, Φ∗ = WΦT Φ and Ψ∗ = WΨT Ψ such that the pair-wise
correlation between the two sets is maximized [10]. CCA is a
method that can help explore the relationship between the two
multi-variate variables. The pairwise correlation between the
two sets can be modeled as:
corr(Φ∗ , Ψ∗ ) =
cov(Φ∗ , Ψ∗ )
,
var(Φ∗ ).var(Ψ∗ )
(3)
where cov(Φ∗ , Ψ∗ ) = WΦT CΦΨ WΨ , var(Φ∗ ) = WΦT CΦΦ WΦ
and var(Ψ∗ ) = WΨT CΨΨ WΨ . The covariances are then used
to find the transformation matrices WΦ and WΨ using the
following equations:
−1
−1
fΦ = Λ2 W
fΦ ,
CΦΦ
CΦΨ CΨΨ
CΨΦ W
−1
−1
fΨ = Λ2 W
fΨ .
CΨΨ
CΨΦ CΦΦ
CΦΨ W
(4)
fΦ and W
fΨ are the eigenvectors and
In the above equation, W
Λ2 is the eigenvalue diagonal matrix.
Lastly, the final feature matrix can be represented as the
sum of the transformed feature matrices from the two modalities:
F =
WΦT Φ
+
WΨT Ψ
=
WΦ
WΨ
T
Φ
.
Ψ
Table 1. Results for accuracy, sensitivity and specificity of the
proposed multi-modal fusion approach along with standard
error of the mean (SEM) in comparison with single modality
and feature concatenation based approaches.
Methods
T1-weighted
T2-weighted
Concat. of T1 & T2
Feature Fusion (Proposed)
Accuracy
(SEM %)
Sensitivity
(SEM %)
Specificity
(SEM %)
84.23 (1.10)
61.04 (1.35)
82.09 (1.01)
82.80 (1.17)
89.16 (0.88)
59.59 (2.25)
88.49 (0.90)
83.55 (1.13)
55.00 (4.16)
57.67 (2.95)
49.33 (3.48)
81.67 (2.53)
4. EXPERIMENTS AND RESULTS
In order to account for the mis-alignment between T1weighted and T2-weighted scans, we performed Multiresolution image registration using image pyramids [11].
The registration results were examined and images with misregistration were removed from the final evaluation set. Our
final evaluation set comprised 139 scans from each modality
and we performed 10 fold cross validation over the dataset.
The minimum (maximum) intensity projection images from
T1 (T2) scans were fed into the deep CNN-F network and
feature representation from each of these images was used
to obtain the final CCA based discriminative representation
(Eq. 5). We then employed Support Vector Machine (SVM)
classifier to obtain the final classification labels i.e. normal vs
IPMN.
We compared our proposed multi-modal feature fusion
based approach with single modality and feature concatenation based approaches. Since there exists an imbalance between the number of positive and negative examples, we performed Adaptive Synthetic Sampling (ADASYN) [12] to generate synthetic samples. This sampling enabled to generate
synthetic feature examples from the minority class (normal).
Table 2. Classification accuracy along with standard error
of the mean (SEM) for three class classification (normal, low
grade IPMN and high grade IPMN) of proposed approach in
comparison with other approaches.
Methods
T1-weighted
T2-weighted
Concat. of T1 and T2
Feature Fusion (Proposed)
Accuracy % (SEM %)
58.30 (1.52)
45.93 (1.60)
56.81 (1.51)
64.67 (0.83)
(5)
The learned transformation is also applied to the features
from test images in order to obtain the final transformed testing features.
Table 1 shows the results of our proposed approach in
varying conditions. It can be observed that the performance
of our proposed approach significantly outperforms the single
vasive carcinoma. Using the features obtained from the CCA
based fusion, we train a 3-class SVM classifier with classes
normal, low-grade IPMN and high-grade IPMN. These diagnostic labels were obtained using the pathology report after
surgery. Table 2 shows the performance of our proposed
approach for normal, low-grade and high-grade IPMN classification. The proposed CCA based classification approach
outperforms single modality and feature concatenation based
approaches. The CCA based approach reports around 8%
improvement in comparison to the feature concatenation
based approach.
5. DISCUSSION AND CONCLUSION
Fig. 2. Qualitative results of our proposed approach, showing minimum and maximum intensity projected images for
T1 and T2 scans on left and right respectively. Each row represents a different case where the images correctly classified
as IPMN are shown in green, whereas the misclassification of
IPMN as normal are shown in red.
modality and feature concatenation based approaches. The T1
based classification yielded the highest sensitivity, but with
very low specificity. For IPMN classification, low specificity
can be a serious problem as that can lead to unwarranted
surgery and resection. In this regard, our proposed approach
reports more than 30% improvement in specificity in comparison with the feature concatenation based approach.
It is important to note that since our proposed approach is
based on the correlation and covariance in the data, it doesn’t
require explicit sample balancing using ADASYN. Moreover,
for experiments, we also tried features from various layers
of CNN-F as well as features from deeper residual networks
such as ResNet-50, ResNet-101, and ResNet-152 [13]. The
best classification results, however, were obtained using the
second fully-connected layer of CNN-F. Figure 2 shows the
qualitative results of our proposed approach with intensity
projections from T1 and T2 scans. The cases shown in green
are the correctly classified as IPMN whereas those shown in
red are incorrectly classified as normal.
Low and High grade IPMN classification:
We also performed 3-class classification using our proposed
approach. Out of 108 IPMN subjects, 48 had low-grade
IPMN whereas the remaining 60 had high-grade IPMN or in-
Pancreatic cancer is projected to become the second leading
cause of cancer-related deaths before 2030 [14]. IPMNs are
the radiographically identifiable precursor to pancreatic cancer. In this paper, we proposed a multi-modal feature fusion
framework to perform the classification of IPMN. Inspired by
the clinical need to identify subjects with IPMN, our proposed
approach can help radiologists in diagnosing invasive pancreatic carcinoma. In contrast to previous studies, this is the first
approach to use deep CNN feature representation for IPMN
diagnosis. Moreover, we empirically show the importance of
feature level fusion of two different MRI imaging modalities
i.e. T1 and T2 scans.
Another advantage of our proposed approach is that it
doesn’t require manual segmentation of pancreas or cysts as
in other approaches. We only need to identify a single slice
where pancreatic tissues can be prominently observed. Additionally, by using the intensity projections across a consecutive set of slices, we can develop robustness to the manual
selection of a single slice. As the CCA is used to learn the
transformation, its use also circumvents the need to have explicit data balancing in the cases of imbalance between positive and negative examples.
As an extension to this study, our future work will involve
joint detection and diagnosis of IPMN in MRI scans. As the
number of subjects undergoing screening for IPMN increases,
we can get sufficient data to perform an end-to-end training or
fine-tuning of a 3D convolutional neural network. The use of
Generative Adversarial Networks (GANs) [15] can assist in
data augmentation by generating realistic examples to further
improve the training of the network.
Furthermore, the segmentation of pancreas and IPMN
cysts can help in localizing the regions of interest. These
regions can be used not only to extract discriminative imaging features, but also to extract important measurements such
as the diameters of main pancreatic duct and cysts [5]. The
inclusion of additional imaging modalities such as CT scans
along with demographic and clinical characteristics, including age, gender, family history, symptoms and body fat can
help further improve diagnostic decision making in the future.
6. REFERENCES
[1] American Cancer Society, “Cancer Facts & Figures,”
American Cancer Society, 2016.
[2] Chanjuan Shi and Ralph H Hruban, “Intraductal papillary mucinous neoplasm,” Human pathology, vol. 43,
no. 1, pp. 1–16, 2012.
[3] Eran Sadot, Olca Basturk, David S Klimstra, Mithat
Gönen, Lokshin Anna, Richard Kinh Gian Do, et al.,
“Tumor-associated neutrophils and malignant progression in intraductal papillary mucinous neoplasms: an
opportunity for identification of high-risk disease,” Annals of surgery, vol. 262, no. 6, pp. 1102, 2015.
[4] Hanno Matthaei, Richard D Schulick, Ralph H Hruban,
and Anirban Maitra, “Cystic precursors to invasive pancreatic cancer,” Nature Reviews Gastroenterology and
Hepatology, vol. 8, no. 3, pp. 141–150, 2011.
[5] Masao Tanaka, Carlos Fernández-del Castillo, Volkan
Adsay, Suresh Chari, Massimo Falconi, Jin-Young Jang,
Wataru Kimura, Philippe Levy, Martha Bishop Pitman,
C Max Schmidt, et al., “International consensus guidelines 2012 for the management of ipmn and mcn of the
pancreas,” Pancreatology, vol. 12, no. 3, pp. 183–197,
2012.
[6] Alexander N Hanania, Leonidas E Bantis, Ziding Feng,
Huamin Wang, Eric P Tamm, Matthew H Katz, Anirban
Maitra, and Eugene J Koay, “Quantitative imaging to
evaluate malignant potential of IPMNs,” Oncotarget,
vol. 7, no. 52, pp. 85776, 2016.
[7] Lior Gazit, Jayasree Chakraborty, Marc Attiyeh, Liana
Langdon-Embry, Peter J Allen, Richard KG Do, and
Amber L Simpson, “Quantification of CT Images for the
Classification of High-and Low-Risk Pancreatic Cysts,”
in SPIE Medical Imaging. International Society for Optics and Photonics, 2017, pp. 101340X–101340X.
[8] Nicholas J Tustison, Brian B Avants, Philip A Cook,
Yuanjie Zheng, Alexander Egan, Paul A Yushkevich,
and James C Gee, “N4ITK: Improved N3 bias correction,” IEEE Transactions on Medical Imaging, vol. 29,
no. 6, pp. 1310–1320, 2010.
[9] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return of the devil in the details: Delving deep
into convolutional nets,” in British Machine Vision Conference, 2014.
[10] Mohammad Haghighat, Mohamed Abdel-Mottaleb, and
Wadee Alhalabi, “Fully automatic face normalization
and single sample face recognition in unconstrained environments,” Expert Systems with Applications, vol. 47,
pp. 23–34, 2016.
[11] Hans J. Johnson, M. McCormick, L. Ibáñez, and The
Insight Software Consortium, The ITK Software Guide,
Kitware, Inc., third edition, 2013.
[12] Haibo He, Yang Bai, Edwardo A Garcia, and Shutao
Li, “ADASYN: Adaptive synthetic sampling approach
for imbalanced learning,” in Neural Networks, 2008.
IJCNN 2008.(IEEE World Congress on Computational
Intelligence). IEEE International Joint Conference on.
IEEE, 2008, pp. 1322–1328.
[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian
Sun, “Deep residual learning for image recognition,” in
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778.
[14] Lola Rahib, Benjamin D Smith, Rhonda Aizenberg, Allison B Rosenzweig, Julie M Fleshman, and Lynn M
Matrisian, “Projecting cancer incidence and deaths to
2030: The unexpected burden of thyroid, liver, and pancreas cancers in the United States,” Cancer research,
vol. 74, no. 11, pp. 2913–2921, 2014.
[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza,
Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron
Courville, and Yoshua Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing
Systems, 2014, pp. 2672–2680.
| 1cs.CV
|
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
arXiv:1711.02777v1 [math.AC] 8 Nov 2017
MICHAEL PERLMAN
Abstract. Let S be the coordinate ring of the space of n × n complex skew-symmetric matrices. This ring has
an action of the group GLn (C) induced by the action on the space of matrices. For every invariant ideal I ⊆ S,
we provide an explicit description of the modules Ext•S (S/I, S) in terms of irreducible representations. This allows
us to give formulas for the regularity of basic invariant ideals and (symbolic) powers of ideals of Pfaffians, as
well as to characterize when these ideals have a linear free resolution. In addition, given an inclusion of invariant
ideals I ⊇ J, we compute the (co)kernel of the induced map ExtjS (S/I, S) → ExtjS (S/J, S) for all j ≥ 0. As a
consequence, we show that if an invariant ideal I is unmixed, then the induced maps ExtjS (S/I, S) → HIj (S) are
injective, answering a question of Eisenbud-Mustaţă-Stillman in the case of Pfaffian thickenings. Finally, using our
Ext computations and local duality, we verify an instance of Kodaira vanishing in the sense described in the recent
work of Bhatt-Blickle-Lyubeznik-Singh-Zhang.
1. Introduction
Let W be a vector space of dimension n over the complex numbers, and consider the polynomial ring S =
Sym(∧2 W ) as a representation of GL = GL(W ). The ideals I ⊆ S invariant under this GL-action have been
classified [ADF80]. Identifying S with the ring C[xi,j ]1≤i<j≤n , the ideal I2k of 2k × 2k Pfaffians of (xi,j ) defines
the Pfaffian variety of matrices of rank < 2k, while all other GL-invariant ideals of S define non-reduced scheme
structures (or thickenings) of the Pfaffian variety. In this paper we study the Castelnuovo-Mumford regularity
of equivariant Pfaffian thickenings I ⊆ S by determining the GL-structure of ExtjS (S/I, S) for all j ≥ 0. This
project was inspired by the work of Raicu, who solved the similar problem in the case of generic m × n matrices
with a GLm (C) × GLn (C) action [Rai16]. In addition to analogues of Raicu’s theorems, we characterize when basic
thickenings have linear free resolution.
Our theorem on regularity of powers is stated below and can be found in Section 6. We write I sat for the
saturation of I with respect to the irrelevant ideal, and I (d) for the d-th symbolic power of I.
Regularity of Powers of Pfaffians. Let 2 < 2k ≤ n − 2. If one of the following holds: (1) n is even and
d ≥ n − 2, (2) n is odd and d ≥ n − 3, then
(
if k is even
k k2 − 1
(d)
d sat
d
= dk + 1
reg I2k = dk, and reg I2k = reg I2k
2
if k is odd
2 (k − 1)
Further, if neither of these conditions hold, then
(d)
Finally, if d ≤ n − 4 then reg(I2k ) > dk.
d sat
d
> dk.
≥ reg I2k
reg I2k
When I = I2k is a Pfaffian ideal, the syzygies (and thus the regularity) are known by the work of Józefiakd
Pragacz-Weyman [JPW81]. When I = In−1
is any power of the ideal of sub-maximal Pfaffians, the syzygies were
known by Boffi-Sánchez [BS92] and independently by Kustin-Ulrich [KU92]. In all cases, we recover the previous
results on regularity using our Ext computations. The above theorem yields:
d
Linear Resolutions of Powers of Pfaffians. Consider an integer d ≥ 1 and let 2 ≤ 2k ≤ n. Then I2k
has linear
minimal free resolution if and only if one of the following holds:
(1) 2k = 2 or 2k = n, and d ≥ 1,
(2) 2k = n − 1, and d is even or d ≥ n − 3,
(3) n is even, 2k = 4, and d ≥ n − 2,
(4) n is odd, 2k = 4, and d ≥ n − 3.
d
d
When 2k = 2, I2k
is a power of the irrelevant ideal, and when 2k = n, I2k
is the ideal of a hypersurface. Thus, case
d
(1) is classical, and case (2) follows from the previous work described above. Our contribution is showing that I2k
has linear minimal free resolution in cases (3) and (4), as well as showing that this list is exhaustive.
1
2
MICHAEL PERLMAN
Every GL-invariant ideal in S is a sum of basic invariant ideals Ix , where x is a partition with at most ⌊n/2⌋
parts. In Section 2 we recall the definition of these ideals and their properties. We compute the regularity of basic
invariant ideals. In the statement of the theorems we write x′ for the conjugate partition of the partition x.
Regularity of Basic Thickenings. Let x be a partition with at most ⌊n/2⌋ parts. If n is even then
(n − 2x′c+1 + 1)(x′c+1 − 1) + cn + 1 .
reg(Ix ) = max
0≤c≤x1 −1
If n is odd then
reg(Ix ) =
max
0≤c≤x1 −1
b(x′c+1 − 1, n, c) + 1 ,
where the function b : Z3 → Z is defined by
(
(⌊n/2⌋ − l)(2l + c) + l(c − 1)
b(l, n, c) =
(⌊n/2⌋ − l)(2l + c) + c(l − 1/2)
if c > 2l, or c < 2l and c is odd
if c ≤ 2l and c is even.
(1.1)
These computations are the content of Section 4. The above formula allows us to characterize when the basic
invariant ideals have linear minimal free resolution.
Linear Resolutions of Basic Thickenings. If n is even, then Ix has linear minimal free resolution if and only
if x′ = ((n/2)k , 1l ), where k ≥ 0 and l ∈ {0, 1}. If n is odd, then Ix has linear minimal free resolution if and only
if one of the following holds:
(1) x′ = (⌊n/2⌋k ) for k ≥ n − 2, or k is even and k ≤ n − 3,
(2) x′ = (⌊n/2⌋k , 1) for k ≥ n − 5, or k is even and k ≤ n − 7.
In ([EMS00], Example 6.3), the authors show that in order for the induced maps ExtiS (S/I, S) → HIj (S) to be
injective for all j ≥ 0, the ideal I needs to be unmixed. Similar to Raicu’s result on determinantal thickenings
([Rai16], Corollary 5.10) we show that the converse holds when I is a GL-invariant ideal in S, providing an answer
to ([EMS00], Question 6.2) in the case of Pfaffian thickenings. This is a consequence of
(d)
d
Theorem on Growth of Ext Modules. For 2 ≤ 2k ≤ n and d ≥ 1 consider the inclusions I2k
⊆ I2k and
(d+1)
(d)
I2k
⊆ I2k . For each j ≥ 0, the induced maps of Ext modules
(d)
(d)
(d+1)
d
ExtjS (S/I2k , S) → ExtjS (S/I2k
, S) and ExtjS (S/I2k , S) → ExtjS (S/I2k
, S)
are injective.
In another direction, we obtain the following vanishing theorem analogous to the recent results of Raicu [Rai16].
Kodaira Vanishing for Pfaffian Thickenings. Let I ⊆ S be a GL-invariant ideal and Y ⊂ P( 2 )−1 the thickening
that it defines. Then
H q (Y, OY (−j)) = 0 for q < 2n − 4 and j > 0.
(1.2)
n
In particular, if we let Yred denote the underlying Pfaffian variety, and if we make the convention that codim(Sing(Yred )) =
dim(Yred ) when Yred is non-singular, then
H q (Y, OY (−j)) = 0 for q < codim(Sing(Yred )) and j > 0.
The above results are all a consequence of the following theorem, recalling that for any graded S-module M :
n
o
reg(M ) = max −r − j | ExtjS (M, S)r 6= 0 ,
and reg(I) = reg(S/I) + 1 for any graded ideal I ⊆ S.
Main Theorem. To any GL-invariant ideal I ⊆ S we can associate a finite set M(I) of GL-equivariant S-modules
with the property that for each j ≥ 0
M
ExtjS (M, S),
(1.3)
ExtjS (S/I, S) ∼
=
M∈M(I)
where the above isomorphism is GL-equivariant and degree preserving, but is not necessarily an isomorphism of
S-modules. It follows that
reg(S/I) = max reg(M ).
M∈M(I)
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
3
The sets M(I) and the modules ExtjS (M, S) for M ∈ M(I) can be computed explicitly. Furthermore, the
association I 7→ M(I) has the property that whenever I ⊇ J are GL-invariant ideals, the (co)kernels and images
of the induced maps ExtjS (S/I, S) → ExtjS (S/J, S) are computed as follows.
M
ExtjS (M, S),
(1.4)
ker ExtjS (S/I, S) → ExtjS (S/J, S) ∼
=
M∈M(I)\M(J)
Im ExtjS (S/I, S) → ExtjS (S/J, S) ∼
=
M
ExtjS (M, S),
(1.5)
M∈M(I)∩M(J)
coker ExtjS (S/I, S) → ExtjS (S/J, S) ∼
=
M
ExtjS (M, S).
(1.6)
M∈M(J)\M(I)
We make a couple remarks about the statment of the Main Theorem. There is a GL-equivarant filtration of S/I
by S-modules, and the M appearing in M(I) are the factors of this filtration. These M are among the quotients
Jz,l of GL-invariant ideals defined in (2.7) below and were first studied in [RW16]. The sets of pairs (z, l) for which
M = Jz,l belongs to M(I) for a given GL-invariant ideal is described in Definition 3.4.
Organization. In Section 3 we compute Ext modules of the subquotients Jz,l , as well as describe the differences
between the proof of the Main Theorem and the proof of ([Rai16], Theorem 3.3). In Section 4 we study the basic
thickenings, and in Section 5 we solve an optimization problem in order to compute the regularity of (symbolic)
powers of Pfaffian ideals in Section 6. In Section 7 we prove Kodaira vanishing for Pfaffian thickenings.
2. Preliminaries
Let W be an n-dimensional complex vector space. The irreducible representations of GL = GL(W ) are classified
by Zndom , the set of dominant weights λ = V
(λ1 ≥ · · · ≥ λn ) ∈ Zn . We write Sλ W for the irreducible representation
n
W for the irreducible Sλ W corresponding to λ = (1, . . . , 1) = (1n ).
corresponding to λ, and write det(W ) =
The size of a weight λ is |λ| = λ1 + · · · + λn . A partition is a dominant weight whose entries are non-negative
integers, and we will use underlined roman letters for partitions to distinguish them from the arbitrary dominant
weights. We will identify a partition x with the associated Young diagram:
x = (4, 3, 1, 0) ←→
When we refer to a row or column of x, we mean a row or column of the associated Young diagram. Given a
partition x, we construct the conjugate partition x′ by transposing the associated Young diagram. In other words,
x′i is the number of boxes in the i-th column of x. Given a positive integer c, we write x(c) for the partition defined
by x(c)i = min(xi , c). We let P(k) denote the set of partitions z = (z1 ≥ · · · ≥ zk ≥ 0), and write Pe (k) for the
partitions with even column lengths in their Young diagrams. Given z ∈ P(k) we write
z (2) = (z1 , z1 , z2 , z2 , · · · ) ∈ Pe (2k).
Whenever convenient, we will identify these sets as subsets of each other, i.e. P(k) ⊂ P(k + 1).
For ease of notation, we will write m = ⌊n/2⌋ throughout. Let S = Sym(∧2 W ) be the ring of polynomial
functions on the space on n × n skew-symmetric matrices. By ([Wey03], Prop 2.3.8), we have the following formula:
M
(2.1)
Sz(2) W.
S=
z∈P(m)
Thinking of S as C[xi,j ]1≤i<j≤n , we define for each even integer 0 ≤ q ≤ n the polynomial Pfaffq = Pfaff(xi,j )1≤i<j≤q .
For z ∈ P(m) we define
z1
Y
Pfaff2zi′ .
(2.2)
Pfaffz =
i=1
Concretely, we realize Sz(2) W as the linear span of the GL-orbit of Pfaffz , the highest weight vector.
For any partition z ∈ P(m) we define
Iz = the ideal in S generated by Sz(2) W .
(2.3)
By (2.2) we see that Iz is generated in degree |z|. This will be used when characterizing the partitions z for which
Iz has linear minimal free resolution.
4
MICHAEL PERLMAN
Since any GL-invariant ideal I ⊂ S is the direct sum of irreducibles in (2.1), it follows that
X
Ix , for X ⊂ P(m) a finite subset.
I = IX =
x∈X
We next recall the descriptions of powers and symbolic powers of Pfaffian ideals from [ADF80]. When z = (1k ),
the ideal Iz is generated by the 2k × 2k Pfaffians of the generic skew-symmetric matrix (xi,j )1≤i<j≤n . When no
confusion may arise, we will denote Iz by I2k . Let
Xkd = {x ∈ P(m) | |x| = kd, x1 ≤ d}.
(2.4)
d
By ([ADF80], Theorem 4.1) the d-th power of I2k is given by I2k
= IXkd .
Let
(d)
Xk = {x ∈ P(m) | x1 = · · · = xk , xk + · · · + xm = d}.
(2.5)
(d)
I2k
Then by ([ADF80], Theorem 5.1) the d-th symbolic power of I2k is given by
= IX (d) .
k
We now recall the definition of the subquotients Jz,l from ([RW16], Lemma 2.5). Given 0 ≤ l ≤ m and z ∈ P(m)
with z1 = · · · = zl+1 , we consider the collection of partitions obtained from z by adding a single box to its Young
diagram in row l + 1 or higher:
succ(z, l) = y ∈ P(m) | y ≥ z and yi > zi for some i > l .
(2.6)
We write
Jz,l = Iz /Isucc(z,l) ,
and note that our notation for these modules is different from [RW16] (we write Jz,l for their Jz(2) ,l ).
(2.7)
3. Ext Modules for the GL-invariant Ideals
Let S be the ring of polynomial functions on the space of n × n skew-symmetric matrices, and continue to write
m = ⌊n/2⌋. In this section we describe the proof of the Main Theorem. We begin by computing Ext•S (Jz,l , S) for
all 0 ≤ l ≤ m − 1 and z ∈ P(m) with z1 = · · · = zl+1 .
3.1. Ext Modules of the Subquotients Jz,l and Regularity. We first prove a statement which will allow us
to compute the Ext modules in question using Bott’s Theorem for Grassmannians ([Wey03], Corollary 4.1.9).
Lemma 3.1. Let Q, R denote the tautological quotient and sub-bundles on the Grassmannian G = G(2l, W ).
Then
M
2l
n
Ext•S (Jz,l , S) =
H ( 2 )−( 2 )−• (G, Sβ R ⊗ Sα Q)∗ ,
α∈Al
where β =
(2)
(z2l+1
+ n − 1, · · ·
(2)
, zn
+ n − 1) and
Al = {α ∈ Z2l
dom | α2i−1 = α2i for 1 ≤ i ≤ l, and α1 ≤ z1 + n − 2l}.
(3.1)
Proof. Define bundles
∨
S = det
2
^
Q
∗
!
⊗ Sym
2
^
Q
∗
!
, and V = Sz2 R ⊗ Sz1 Q,
where z 1 = (z1 , . . . , zl )(2) and z 2 = (zl+1 , . . . , zm )(2) . By ([RW16], Lemma 2.5) we have
!
2
^
2l
n
•
−•
−
∨
∗
∗
Ext (J , S) = H ( 2 ) ( 2 ) (G, S ⊗ V) ⊗ det
.
W
S
(3.2)
z,l
For ease of notation set z1 = d. We make a few observations:
!
!
2
2
^
^
∗ ⊗(n−1)
∗
∗
= det(Q)⊗(−2l+1) .
= det(W )
, (2) det(W ) ⊗ OG = detQ ⊗ detR, (3) det
(1) det
Q
W
Then from (1),(2),(3), and (3.2) we have that
Ext•S (Jz,l , S)
2l
n
= H ( 2 )−( 2 )−•
=
M
y∈P(l)
⊗(d+n−2l)
G, det(Q)
⊗ Sym
2
^
Q
∗
!
⊗(n−1)
⊗ Sz2 R ⊗ det(R)
∗
l
n
H ( 2 )−(2)−• G, Sβ R ⊗ Sy(2) Q∗ ⊗ det(Q)⊗(d+n−2l) .
!∗
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
By ([Wey03], Prop 2.3.8), the result follows.
5
For 0 ≤ l ≤ m and z ∈ P(m), we define
(2)
(2)
| z2l+i − z2l+i+1 ≥ 2ti − 2ti+1 for 1 ≤ i ≤ n − 2l − 1}.
Tl (z) = {t = (l = t1 ≥ t2 ≥ · · · ≥ tn−2l ) ∈ Zn−2l
≥0
(3.3)
Theorem 3.2. Fix 0 ≤ l ≤ m − 1 and assume that z ∈ P(m) with z1 = · · · = zl+1 . For t ∈ Tl (z) we consider the
set W (z, l, t) of dominant weights λ ∈ Zndom satisfying
(2)
λ2l+i−2ti = z2l+i + n − 1 − 2ti for i = 1, · · · , n − 2l
(3.4)
λ2i = λ2i−1 for 0 < 2i < n − 2tn−2l
λn−2i = λn−2i−1 for 0 ≤ i ≤ tn−2l − 1
Then for all j ≥ 0
ExtjS (Jz,l , S) =
M
Sλ W ∗ ,
(3.5)
t∈Tl (z)
P
(n2 )−(2l2 )−2 n−2l
i=1 ti =j
λ∈W (z,l,t)
where Sλ W ∗ appears in degree −|λ|/2.
Proof. In the notation of Lemma 3.1, we have that
M
2l
n
H ( 2 )−( 2 )−• (G, Sβ R ⊗ Sα Q)∗ .
Ext•S (Jz,l , S) =
(3.6)
α∈Al
Write γ = (α1 , · · · , α2l , β1 , · · · , βn−2l ) for α ∈ Al . Let ρ = (n − 1, n − 2, · · · , 0) and consider γ + ρ = (γ1 + n −
1, γ2 + n − 2, · · · , γn ). We write γ̄ = sort(γ + ρ) for the sequence obtained by arranging the entries of γ + ρ in
non-increasing order. If we write q for the number of pairs (x, y) with 1 ≤ x < y ≤ n and γx − x < γy − y, then
Bott’s Theorem ([Wey03] Cor. 4.1.9) yields
(
Sγ̄−ρ W ∗ if γ + ρ has distinct entries and • = q;
•
∗
H (G, Sβ R ⊗ Sα Q) =
(3.7)
0
otherwise.
We start by showing that λ = γ̄ − ρ ∈ W (z, l, t) for some t ∈ Tl (z). Let σ denote the permutation of {1, · · · , n}
that sorts γ + ρ. This permutation is unique if γ + ρ has distinct entries. Notice that since α1 ≥ · · · ≥ α2l and
β1 ≥ · · · ≥ βn−2l it follows that σ(1) < · · · < σ(2l) and σ(2l + 1) < · · · < σ(n). Thus, λ is determined uniquely by
the partition u ∈ P(n − 2l) defined by
ui = 2l + i − σ(2l + i), for i = 1, · · · , n − 2l.
We may view the sorting process as moving β1 + ρ2l+1 left by u1 spaces, then moving β2 + ρ2l+2 left by u2 spaces,
and so on. Note that the condition γx − x < γy − y is equivalent to γ̄σ(x) < γ̄σ(y) . If γ + ρ has distinct entries, we
get that
q = #{(x, y) | x < y, σ(x) > σ(y)} = |u|.
Our proof is based on two claims which we explain at the end:
Claim 1: The integer u1 is maximal, i.e. σ(2l + 1) = 1 and u1 = 2l.
Claim 2: ui is even for all i = 1, · · · , n − 2l, and setting ti = ui /2 for all i gives t ∈ Tl (z).
We now explain the proof of λ ∈ W (z, l, t) based on Claims 1 and 2. For ease of notation, set w = z (2) . Notice
that
(3.8)
λ2l+i−ui = βi − ui = w2l+i + n − 1 − ui ,
as needed to confirm that λ satisfies condition one of (3.4). To see that λ satifies the second and third condition of
(3.4), assume first that n is even. In this case, we need to show that λ2i = λ2i−1 for i = 1, · · · , n/2. If 2i = 2l+j −uj
for some 1 ≤ j ≤ n − 2l, then j is even by Claim 2. Since w2l+j = w2l+j−1 , it follows from Claim 2 that uj = uj−1 .
Thus,
λ2i = λ2l+j−uj = w2l+j + n − 1 − uj = w2l+j−1 + n − 1 − uj−1 = λ2l+j−1−uj−1 = λ2i−1 ,
as needed. Fix 1 ≤ i ≤ n/2 and suppose that there does not exist j = 1, · · · , n − 2l with 2i = 2l + j − uj . Then
there exists 1 ≤ j ≤ l such that λ2i (resp. λ2i−1 ) is obtained by moving α2j + n − 2j (resp. α2j−1 + n − 2j + 1)
2i − 2j spaces to the right when sorting γ + ρ. Thus,
λ2i = (α2j + n − 2j) − (n − 2i) = α2j + 2i − 2j = α2j−1 + 2i − 2j = (α2j−1 + n − 2j + 1) − (n − 2i + 1) = λ2i−1 .
Setting ti = ui /2 for i = 1, · · · , n − 2l we see that λ ∈ W (z, l, t), where t ∈ Tl (z), as required. If n is odd, the proof
is similar except it may not be true that λn−2tn−2l = λn−2tn−2l +1 .
6
MICHAEL PERLMAN
For the reverse containment, assume that λ ∈ W (z, l, t), where t ∈ Tl (z). Its clear that a suitable weight α can
be constructed from λ such that λ = γ̄ − ρ.
We now prove the claims.
Proof of Claim 1: Let α ∈ Al . If l = 0 the claim is clear, so suppose l ≥ 1. Since γ + ρ has distinct entries, it
suffices to show α2 + ρ2 ≤ β1 + ρ2l+1 . By (3.1) we have
α2 + ρ2 ≤ w1 + n − 1 + n − 2l − 1 = β1 + n − 2l − 1 = β1 + ρ2l+1 ,
as desired.
Proof of Claim 2: Since α2i = α2i−1 for i = 1, · · · , l, it follows that (γ + ρ)2i = (γ + ρ)2i−1 − 1 for i = 1, · · · , l.
Thus, it cannot be the case that (γ + ρ)2i−1 − 1 > (γ + ρ)j > (γ + ρ)2i for some i ≤ l and j > 2l. Therefore,
βi + ρ2l+i gets shifted left by an even amount while sorting, for all i > l, i.e. uj even for all j = 1, · · · , n − 2l. We
only need to show that w2l+i − w2l+i+1 ≥ ui − ui+1 for all i = 1, · · · , n − 2l − 1. Since λ is dominant we have
λ2l+i−ui ≥ λ2l+i+1−ui+1 for all i = 1, · · · , n − 2l − 1, so that by (3.8), βi − ui ≥ βi+1 − ui+1 , which is equivalent to
the desired inequality by the definition of β.
Given z ∈ P(m) and t ∈ Tl (z), we define
fl (z, t) =
n−2l−1
X
(2)
(2)
ti+1 ((z2l+i − z2l+i+1 ) − (2ti − 2ti+1 )).
(3.9)
i=1
Theorem 3.3. For 0 ≤ l ≤ m − 1 and z ∈ P(m) satisfying z1 = · · · = zl+1 , we have that
!
m
X
reg Jz,l = max
zi + l(z1 − 1) + |t| − fl (z, t) .
t∈Tl (z)
(3.10)
i=l+1
Proof. By Theorem 3.2, reg(Jz,l ) is given by
n
o
reg(Jz,l ) = max −r − j | ExtjS (Jz,l , S)r 6= 0 =
max
t∈Tl (z)
λ∈W (z,l,t)
n
2l
|λ|
−
+
+ 2|t| .
2
2
2
Given t ∈ Tl (z), let λ be the largest weight in W (z, l, t) with respect to |λ|. Notice that
|λ| =
n−2l−1
X
(2)
(2ti − 2ti+1 + 1)(z2l+i + n − 1 − 2ti ) + (2tn−2l + 1)(zn(2) + n − 1 − 2tn−2l ).
i=1
Therefore, the proof is completed by verifying the following equality, which we leave to the interested reader.
m
X
|λ|
n
2l
zi + l(z1 − 1) + |t| − fl (z, t).
−
+
+ 2|t| =
2
2
2
i=l+1
3.2. The Main Theorem. Let I ⊆ S be a GL-invariant ideal. We begin by defining the pairs (z, l) for which Jz,l
appears as a subquotient of S/I. We continue to write m = ⌊n/2⌋.
Definition 3.4. For a finite subset X ⊂ P(m), we define Z(X ) to be the set consisting of pairs (z, l) where
z ∈ P(m) and 0 ≤ l ≤ m − 1 are such that if we write c = z1 then the following hold:
(1) There exists a partition x ∈ X such that x(c) ≤ z and x′c+1 ≤ l + 1.
(2) For every partition x ∈ X satisfying (1) we have x′c+1 = l + 1.
Remark 3.5. If X = {x} is a singleton, then
o
n
Z(X ) = (z, x′c+1 − 1) | 0 ≤ c ≤ x1 − 1, z1 = · · · = zx′c+1 = c and x(c) ≤ z .
(3.11)
This will be used when computing the regularity of the ideals Ix .
We may now give a precise description of the sets M(I) in the statement of the Main Theorem. Given X ⊂ P(m),
we have M(IX ) is the set of Jz,l , where (z, l) ∈ Z(X ). The proof of the Main Theorem is identical to the proof of
([Rai16], Theorem 3.3), with the two exceptions being Raicu’s Lemma 3.8 and (part of) Lemma 3.11. Proving the
analogues of these requires the computations from our Theorem 3.2. We prove our version of these lemmas below.
We first recall notation from [Rai16]: to every (z, l) with z ∈ P(m) and z1 = · · · = zl+1 , we associate the collection
of rectangular partitions
(3.12)
Yz,l = (z1 + 1)l+1 ∪ (zi + 1)i | i > l + 1 and zi−1 > zi .
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
7
Lemma 3.6. Let Y = Yz,l ∪ {z}. If (y, k) ∈ Z(Y) then there exists no non-zero GL-equivariant maps between
ExtjS (Jz,l , S) and Extj+1
S (Jy,k , S).
Proof. Write z1 = c and let (y, k) ∈ Z(Y) with y1 = d. Then by (3.4) there’s x ∈ Y such that y ≥ x(d) and
x′d+1 = k + 1. By (3.12) the nonzero columns of x have size at least l + 1, so that k + 1 ≥ l + 1, i.e. k ≥ l. Since
x1 ≥ d + 1, and since x1 ≤ c + 1, conclude that d ≤ c.
Now assume that Sλ W ∗ is a subrepresentation of both ExtjS (Jz,l , S) and Extj+1
S (Jy,k , S). Then by (3.5) it follows
that λ1 = c + n − 1 − 2l = d + n − 1 − 2k. Since k ≥ l and d ≤ c, we se that k = l and c = d. Also by (3.5), λ is
parametrized by partitions t ∈ Tl (z) and s ∈ Tl (y) such that
n−2l
n−2l
X
X
n
2l
n
2l
si = j + 1
(3.13)
ti = j and
−
−2
−
−2
2
2
2
2
i=1
i=1
This cannot happen because the left hand sides of both expressions in (3.13) have the same parity. Conclude that
no such λ exists and thus, there are no equivariant maps between ExtjS (Jz,l , S) and Extj+1
S (Jy,k , S)
Lemma 3.7. Let 0 ≤ l ≤ m be and let z ∈ P(m). If
X = ((zi + 1)i ) | i > l + 1 ⊂ P(m),
then the S-modules ExtjS (Jz,l , S) and ExtjS (S/IX , S) share no irreducible GL-representations for all j ≥ 0.
Proof. Every nonzero column of x ∈ X has size bigger than l + 1, so it follows from part (2) of Definition 3.4 that
if (y, u) ∈ Z(X ) then u + 1 > l + 1, i.e. u > l. Every x ∈ X satisfies x1 ≤ c + 1, so if (y, u) ∈ Z(X ) then y1 ≤ c.
If Sλ W ∗ occurs as a subrepresentation of ExtjS (S/IX , S) then there exists (y, u) ∈ Z(X ) such that Sλ W ∗ occurs
inside ExtjS (Jy,u , S). It follows from Theorem 3.2 that λ1 = y1 + n − 1 − 2u. If Sλ W ∗ also occurs in ExtjS (Jz,l , S),
by the same reasoning we have λ1 = z1 + n − 1 − 2l. Thus,
z1 + n − 1 − 2l = y1 + n − 1 − 2u ⇐⇒ z1 − 2l = y1 − 2u,
and since u > l and y1 ≤ c = z1 , this is a contradiction.
4. Regularity of Basic Thickenings
As before, S will denote the ring of polynomial functions on the space of n × n complex skew-symmetric
matrices, and we write m = ⌊n/2⌋. In this section we use the Ext computations in the previous section to compute
the regularity of the basic GL-invariant ideals in S (for reminder of definition see (2.3)). As a consequence, we
determine the partitions x ∈ P(m) for which Ix has linear minimal free resolution. We start by treating the case
where n is even, followed by the case when n is odd.
Lemma 4.1. If n is even and 0 ≤ l ≤ m, then for an integer c ≥ 0 we have
max reg Jz,l | z ∈ P(m), z1 = · · · = zl+1 = c = reg J(cm ),l = (n − 2l − 1)l + cm.
(4.1)
Proof. By Theorem 3.3, if z ∈ P(m) and z1 = · · · = zl+1 = c, we have
reg Jz,l = max
t∈Tl (z)
m
X
!
zi + l(c − 1) + |t| − fl (z, t) .
i=l+1
(4.2)
Pm
m
Note that |t| is maximized for t = (ln−2l ) and
i=l+1 zi is maximized for z = (c ). Since ti = ti+1 for all
i = 1, · · · , n − 2l − 1 and zl+i = zl+i+1 for all i = 1, · · · , m − l − 1, we see that t ∈ Tl (z) and fl (z, t) = 0 for this
choice of z and t. Therefore, the result follows.
Theorem 4.2. If n even and x ∈ P(m), then
reg Ix = max
0≤c≤x1 −1
(n − 2x′c+1 + 1)(x′c+1 − 1) + cm + 1 .
Proof. By the Main Theorem and Remark 3.5, we have
Lemma 4.1
reg Jz,l
reg(S/Ix ) =
=
max
(z,l)∈Z({x})
max
0≤c≤x1 −1
The result follows after recalling that reg(Ix ) = reg(S/Ix ) + 1.
(n − 2x′c+1 + 1)(x′c+1 − 1) + cm .
For the following statement, recall that by (2.2) the ideal Ix is generated in degree |x|.
(4.3)
8
MICHAEL PERLMAN
Corollary 4.3. If n is even, then Ix has linear free resolution if and only if x′ = (mk , 1l ), where k ≥ 0 and
l ∈ {0, 1}.
Proof. ⇐= : Using Theorem 4.2 it is easy to check that reg(Ix ) = |x| if x is of the form stated.
=⇒ : Now suppose x ∈ P(m) and Ix has linear free resolution. Note that if m = 1, there is nothing to show,
so assume m ≥ 2. We start by showing that x has at most one column of length one. Let s = x1 and suppose
for contradiction that x′s = · · · = x′s−t = 1 for t ≥ 1. The largest x can be is x′ = (ms−t−1 , 1t+1 ), in which case
|x| = (s − t − 1)m + (t + 1). By Theorem 4.2, we have that
reg(Ix ) ≥ (n − 2x′s + 1)(x′s − 1) + (s − 1)m + 1 = (s − 1)m + 1.
Thus, it would be a contradiction if m(s − 1) + 1 > (s − t − 1)m + (t + 1) ≥ |x|. Note that
m(s − 1) + 1 ≤ (s − t − 1)m + (t + 1) ⇐⇒ t(1 − m) ≥ 0 ⇐⇒ t = 0 or m ≤ 1,
but t ≥ 1 and m ≥ 2, so the claim is proved.
Now suppose for contradiction that x is not of the form stated and let c be largest such that 1 < x′c+1 < m. For
ease of notation, set x′c+1 = d. By the previous paragraph, the largest x can be is x′ = (mc , d, 1), in which case
|x| = cm + d + 1. Also, the Corollary is proved when m = 2, so we may assume m ≥ 3. By Theorem 4.2 we know
that
reg(Ix ) ≥ (n − 2d + 1)(d − 1) + cm + 1.
Thus, it would be a contradiction if (n − 2d + 1)(d − 1) + cm + 1 > cm + d + 1 ≥ |x|. Note that
(n − 2d + 1)(d − 1) + cm + 1 > cm + d + 1 ⇐⇒ d(n − 2d + 2) − n − 1 > 0.
Since 2 ≤ d ≤ m − 1 and n ≥ 6, its easy to check that this inequality holds, completing the proof.
We now compute the regularity of Ix when n is odd. Recall the definition of the b : Z3 → Z from (1.1).
Lemma 4.4. If n is odd and 0 ≤ l ≤ m − 1, then for an integer c ≥ 0 we have
max reg Jz,l | z ∈ P(m), z1 = · · · = zl+1 = c = reg J(cm ),l = b(l, n, c).
(4.4)
Proof. We start by showing that reg(J(cm ),l ) = b(l, n, c). By Theorem 3.3 we have
reg J(cm ),l = maxm ((m − l)c + l(c − 1) + |t| − fl (cm , t)) .
t∈Tl (c )
m
Since t ∈ Tl (c ) it follows that t1 = · · · = tn−2l−1 = l, so that
reg J(cm ),l = maxm ((m − l)c + l(c − 1) + 2l(m − l) + tn−2l − tn−2l (c − 2l + 2tn−2l ))
t∈Tl (c )
(4.5)
Thus, we need to maximize
tn−2l − tn−2l (c − 2l + 2tn−2l )
(4.6)
with respect to tn−2l . If c > 2l then (4.6) is maximized when tn−2l = 0 since tn−2l ≥ 0. If c ≤ 2l then we find that
(4.6) is maximized
when tn−2l = (2l − c)/2 if c is even, and tn−2l = (2l − c + 1)/2 if c is odd. Using (4.5) we see
that reg J(cm ),l = b(l, n, c), as claimed.
We now show that if w ∈ P(m) and w1 = · · · = wl+1 = c, then reg Jw,l ≤ reg J(cm ),l , completing the proof.
Let w be lexicographically maximal such that: w1 = · · · = wl+1 = c and
reg Jw,l = max reg Jz,l | z ∈ P(m), z1 = · · · = zl+1 = c ,
(4.7)
and let t ∈ Tl (w) be such that
m
X
wi + l(c − 1) + |t| − fl (w, t).
reg Jw,l =
i=l+1
Note that such a t exists by Theorem 3.3.
(2)
(2)
First assume that w2l+i − w2l+i+1 = 2ti − 2ti+1 for all i = 1, · · · , n − 2l − 1. It follows from (3.3) that
(2)
(2)
w2l+i − 2ti = c − 2l for all i = 1, · · · , n − 2l − 1. Since wn = 0 it follows that tn−2l = (2l − c)/2, and since tn−2l ≥ 0
we see that 2l ≥ c and c is even. Since fl (w, t) = 0, we need to show that
m
X
wi + l(c − 1) + |t| ≤ b(l, n, c).
i=l+1
Since |w| is maximized when w = (cm ) and |t| is maximized when t = (ln−2l−1 , (2l − c)/2), the result follows.
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
(2)
9
(2)
Now assume that there exists 1 ≤ j ≤ n − 2l − 1 such that w2l+j − w2l+j+1 > 2tj − 2tj+1 , and notice that
it must be true that j is even. Let j be the minimal index so that this inequality holds. If j < n − 2l − 1 then
(2)
(2)
(2)
(2)
(2)
define x ∈ P(m) via x2l+j+1 = x2l+j+2 = w2l+j+1 + 1, and xi = wi otherwise. Notice that t ∈ Tl (x). Then
x1 = · · · = xl+1 = c and by Theorem 3.3 we know that
m
X
reg Jx,l ≥
xi + l(c − 1) + |t| − fl (x, t) = reg Jw,l + 1 + tj+1 − tj+3 .
i=l+1
Since tj+1 ≥ tj+3 , this contradicts the fact that w is lexicographically maximal such that (4.7) holds.
Now suppose that j = n − 2l − 1.
(2)
(2)
(2)
Case 1: tn−2l = 0. If tn−2l = 0 then wn−1 > 2tn−2l−1 by (3.3) and w2l+i − w2l+i+1 ≥ 2ti − 2ti+1 for
(2)
(2)
(2)
i = 1, · · · , n − 2l − 1 implies that w2l+i − 2ti ≥ w2l+i+1 − 2ti+1 for all i = 1, · · · , n − 2l − 1. Thus, w2l+1 − 2t1 ≥
(2)
wn−1
− 2tn−2l−1 > 0, so that c =
reg Jw,l =
m
X
(2)
w2l+1
> 2t1 = 2l. Then by Theorem 3.3 we have that
wi + l(c − 1) + |t| − fl (w, t) ≤ (m − l)c + l(c − 1) + l(n − 2l − 1),
i=l+1
which is equal to b(l, n, c) since c > 2l, as needed.
(2)
Case 2: tn−2l ≥ 1. If wn−1 = 2tn−2l−1 − 2tn−2l + 1 then by Theorem 3.3 we have
m
X
wi + l(c − 1) + |t| − tn−2l ≤ (m − l)c + l(c − 1) + (n − 2l − 1)l.
reg Jw,l =
i=l+1
If c > 2l, or c ≤ 2l and c is odd, the right side of this inequality is equal to b(l, n, c). If c ≤ 2l and c is even, then
l(c − 1) ≤ c(l − 1/2), so the right hand side of the above inequality is less than or equal to b(l, n, c).
(2)
Now assume that wn−1 > 2tn−2l−1 − 2tn−2l + 1 and let s ∈ P(n − 2l) be defined by sn−2l = tn−2l − 1 and si = ti
otherwise. Its easy to see that s ∈ Tl (w) and by Theorem 3.3 we have that
m
X
wi + l(c − 1) + |s| − fl (w, s)
reg Jw,l ≥
=
i=l+1
m
X
(2)
wi + l(c − 1) + |t| − fl (w, t) + 4tn−2l + wn−1 − 2tn−2l−1 − 3,
i=l+1
(2)
which is greater than reg Jw,l since tn−2l ≥ 1 and wn−1 > 2tn−2l−1 − 2tn−2l + 1. This is a contradiction.
Theorem 4.5. If n is odd and x ∈ P(m), then
reg Ix = max
0≤c≤x1 −1
b(x′c+1 − 1, n, c) + 1 .
Proof. The proof is analogous to that of Theorem 4.2 and uses the Main Theorem and Lemma 4.4.
(4.8)
Corollary 4.6. If n is odd, then Ix has linear free resolution if and only if one of the following holds:
(1) x′ = (mk ) for k ≥ n − 2 or k is even and k ≤ n − 3
(2) x′ = (mk , 1) for k ≥ n − 5 or k is even and k ≤ n − 7.
Proof. ⇐= : Using Theorem 4.5 it is easy to check that reg(Ix ) = |x| if x is of the form stated.
=⇒ : Suppose now that x ∈ P(m) and Ix has a linear free resolution.
Case 1: x′i = 1 for some 1 ≤ i ≤ x1 . In this case, we claim that x has a unique column of length 1, i.e. x′x1 = 1
and x′x1 −1 > 1. The proof of this fact follows by similar reasoning as in the proof of Corollary 4.3 and is left to the
interested reader.
Now suppose for contradiction that x has a column of length not equal to 1 or m, and let c be largest such that
1 < x′c+1 < m. In this case we must have n ≥ 7. For ease of notation, set d = x′c+1 . Then the largest x can be is
x′ = (mc , d, 1), in which case |x| = cm+ d+ 1. We know that reg(Ix ) ≥ b(d− 1, n, c)+ 1. Since Ix has linear minimal
free resolution, it follows that reg(Ix ) = |x|. Thus, it would be a contradiction if b(d − 1, n, c) + 1 > cm + d + 1 ≥ |x|.
Since n ≥ 7, the formula (1.1) gives the desired contradiction. Therefore x′ = (mk , 1) for some k ≥ 0.
We only need to verify that if k is odd and k ≤ n − 6 then reg(Ix ) > km + 1. It suffices to show that if k is odd
and k ≤ n − 6 then b(m − 1, n, k − 1) > km. In this case, b(m − 1, n, k − 1) = k(m − 1/2) + m − 3/2 > km.
Case 2: x has no column of length 1. In this case we have that x has no columns of legnth d < m. Indeed, the
proof of this fact is identical to the proof in the previous case. Conclude that x′ = (mk ) for some k ≥ 1. We need
10
MICHAEL PERLMAN
to show that if k is odd and k ≤ n − 4 that reg(Ix ) > km. It suffices to show that if k is odd and k ≤ n − 4 then
b(m − 2, n, k − 1) + 1 > km. In this case, b(m − 1, n, k − 1) + 1 = k(m − 1/2) + m − 1/2 > km.
5. An Optimization Problem
In this section we solve an optimization problem that will allow us to compute the regularity of large powers
and symbolic powers of ideals of Pfaffians. The main result of the section is Proposition 5.1 below. For q even and
positive integers k, n with 0 ≤ q < 2k ≤ n we consider
o
n
|y|≤d(2k−q)−2, |y|−2y ≥d(2k−q−2)
(5.1)
YU(q, k, n, d) = (y, u) ∈ Pe (n − q) × P(n − q) | ui even, u1 =q, yi −yi+1 ≥ui −ui+11for i = 1, · · · , n − q − 1
For (y, u) ∈ YU (q, k, n, d) we define
gq,k,n,d (y, u) =
and let
n−q−1
X ui+1
|y| |u| q(y1 − 1)
+
+
−
((yi − yi+1 ) − (ui − ui+1 )),
2
2
2
2
i=1
Rq,k,n,d =
max
(y,u)∈YU(q,k,n,d)
gq,k,n,d (y, u),
(5.2)
(5.3)
with the convention that Rq,k,n,d = −∞ when YU(q, k, n, d) is empty.
Proposition 5.1. If one of the following holds:
(1) n is even, q is even, 0 ≤ q < 2k ≤ n − 2, and d ≥ n − 2
(2) n is odd, q is even, 0 ≤ q < 2k ≤ n − 2, and d ≥ n − 3
(3) n is even, 2k = n, q = 2k − 2, and d ≥ 1
(4) n is odd, 2k = n − 1, q = 2k − 2, and d ≥ n − 3
then
q
Rq,k,n,d = dk − 1 + q k − − 1 .
2
The proof of this statement is the content of the remainder of the section. Consider q even and positive integers
k, n with 0 ≤ q < 2k ≤ n. If (y, u) ∈ YU (q, k, n, d) then its easy to check that
y1 ≤ d − 1.
Lemma 5.2. If n is even we have that Rn−2, n2 ,n,d =
nd
2
(5.4)
− 1 for all d ≥ 1, and Rq, n2 ,n,d = −∞ for q ≤ n − 4.
Proof. Let q = n − 2 and k = n/2. The partitions y, u satisfying the conditions in (5.1) have two equal parts,
y = (y1 , y1 ), u = (u1 , u1 ), and they satisfy the conditions y1 ≤ d − 1 and u1 = q = n − 2. Then
(n − 2)(y1 − 1)
,
2
which is maximized when y1 = d − 1. It follows that Rn−2, n2 ,n,d = nd
2 − 1, as claimed.
Assume now that k = n/2 and q ≤ n − 4. Then (n − q − 2)y3 ≥ |y| − 2y1 ≥ d(2k − q − 2) = d(n − q − 2), so
y3 ≥ d, contradicting (5.4). Therefore, YU(q, n2 , n, d) is empty for q ≤ n − 4.
gq,k,n,d (y, u) = y1 + (n − 2) +
Lemma 5.3. If n is odd we have that
Rn−3, n−1 ,n,d =
2
(
d(n−1)
2
d(n−1)
2
+ 21 (n − d − 4)
−1
if d odd and d < n − 2
otherwise
Further, Rq, n−1 ,n,d = −∞ for q ≤ n − 5.
2
Proof. We first show the second assertion. Assume that k = (n − 1)/2 and q ≤ n − 5. Then since yn−q = 0 we have
(n − q − 3)y3 ≥ |y| − 2y1 ≥ d(2k − q − 2) = d(n − q − 3),
so y3 ≥ d, contradicting (5.4). Therefore YU(q, n−1
2 , n, d) is empty for q ≤ n − 5.
d
). By (2.4), it
To verify the first assertion, note that by Corollary 6.3, we have that Rn−3, n−1 ,n,d = reg(S/In−1
2
d
follows that In−1
= Ix where x = (d(n−1)/2 ) = (dm ). By Theorem 4.5 the result follows.
Lemma 5.4. If q is even, 0 ≤ q < 2k ≤ n − 2, and d ≥ n − 3, then
q
Rq,k,n,d ≥ dk − 1 + q k − − 1 .
2
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
11
Proof. Let y1 = · · · = y2k−q = d − 1, y2k−q+1 = y2k−q+2 = k − q/2 − 1, and yi = 0 for i > 2k − q + 2, and let
u1 = · · · = u2k−q = l, ui = 0 for i > 2k − q. Its easy to show that (y, u) ∈ YU(q, k, n, d). Observe that
d(2k − q) − 2 q(2k − q) q(d − 2)
q
gq,k,n,d (y, u) =
+
+
= dk − 1 + q k − − 1 .
2
2
2
2
Since Rq,k,n,d ≥ gq,k,n,d (y, u), the result follows.
Proof of Proposition 5.1. By Lemma 5.2 and Lemma 5.3 we only need to check two cases:
(1) n is even, q is even, 0 ≤ q < 2k ≤ n − 2, and d ≥ n − 2
(2) n is odd, q is even, 0 ≤ q < 2k ≤ n − 2, and d ≥ n − 3
By Lemma 5.4, it suffices to show that
q
Rq,k,n,d ≤ dk − 1 + q k − − 1 ,
(5.5)
2
in these two cases. Among the elements (y, u) ∈ YU (q, k, n, d) for which gq,k,n,d (y, u) = Rq,k,n,d , we consider one
for which y is lexicographically maximal. Its easy to show that
|y| = d(2k − q) − 2.
(5.6)
To prove the Proposition we proceed by induction on n. We divide our analysis into four cases:
Case 1: un−q = yn−q = 0. We can think of y as an element of Pe (n − q − 1) and u as an element of P(n − q − 1).
It follows from (5.1) that (y, u) ∈ YU (q, k, n − 1, d). We get
Rq,k,n,d = gq,k,n,d (y, u) = gq,k,n−1,d (y, u) ≤ Rq,k,n−1,d .
If 2k = n − 2 = (n − 1) − 1 then since YU (q, k, n − 1, d) is nonempty, we must have that q = (n − 1) − 3 = n − 4
by Lemma 5.3. In this case n is even, and since d ≥ n − 2 = (n − 1) − 1, we have
q
Lemma 5.3 d((n − 1) − 1)
=
Rq,k,n−1,d
− 1 = dk − 1 + q k − − 1 .
2
2
If 2k < n − 2, then since d ≥ n − 3 = (n − 1) − 2 (when n is even or odd), it follows from induction on n that
Rq,k,n−1,d = dk − 1 + q(k − q/2 − 1). In both cases, conclude that the Proposition holds.
Case 2: un−q = 0 and yn−q > 0. The proof of this case goes as in the proof of Case 2 in ([Rai16], Proposition
4.1), and is left to the interested reader.
Case 3: un−q = p > 0 and yn−q ≤ (p − n + d + 1) + 2(k − q/2 − 1). Since 0 ≤ p ≤ q is even, there is a bijection
p
{(z, w) ∈ YU(q, k, n, d) | vn−q ≥ p} ←→ YU q − p, k − , n − p, d
2
given by
(z, w) ←→ (z, w − (pn−q ))
Let v = u − (pn−q ). Then (y, v) ∈ YU (q − p, k − p/2, n − p, d). Moreover,
p
gq−p,k−p/2,n−p,d (y, v) = gq,k,n,d (y, u) − (n − p + yn−q − 1).
2
Therefore,
q
p
pd
+ p(k − − 1).
(5.7)
gq,k,n,d (y, u) ≤ Rq−p,k− p2 ,n−p,d + (n − p + yn−q − 1) ≤ Rq−p,k− p2 ,n−p,d +
2
2
2
Since d ≥ n − 3 ≥ (n − p) − 2 and q − p < 2k − p ≤ (n − p) − 2, by induction
q
p
− 1 + (q − p) k − − 1 .
Rq−p,k− p2 ,n−p,d = d k −
2
2
Thus, by (5.7) we have that (5.5) holds in this case.
Case 4: un−q = p > 0, yn−q > (p − n + d + 1) + 2(k − q/2 − 1). Since 0 ≤ q < 2k ≤ n − 2 and d ≥ n − 3 it
follows that
(p − n + d + 1) + 2(k − q/2 − 1) ≥ (2 − n + (n − 3) + 1) = 0.
Thus, yn−q ≥ 1 and since y ∈ Pe (n−q) it follows that n is even. Notice that un−q −yn−q ≤ (n−d−2)+2(q/2−k+1).
By the conditions (5.1), we have that ui −ui+1 ≤ yi −yi+1 for i = 1, · · · , n−q−1. It follows that ui −yi ≤ ui+1 −yi+1
for i = 1, · · · , n − q − 1, so ui − yi ≤ (n − d − 2) + 2 2q − k + 1 for i ≤ n − q. Since u1 = q adding these inequalities
for i = 3, · · · , n − q gives
q
−k+1 .
(5.8)
|u| ≤ (|y| − 2y1 ) + 2q + (n − q − 2) (n − d − 2) + 2
2
The the proof proceeds as in the proof of Case 4 in ([Rai16], Proposition 4.1), and it uses that d ≥ n − 2.
12
MICHAEL PERLMAN
6. Regularity of Powers of Ideals of Pfaffians
In this section we assume as before that 2 ≤ 2k ≤ n, that S = C[xi,j ]1≤i<j≤n , and I2k is the ideal of 2k × 2k
Pfaffians of the generic skew-symmetric matrix (xi,j ). For ease of notation we again write m = ⌊n/2⌋. The main
result of the section is the following:
Theorem 6.1. Let 2 < 2k ≤ n − 2. If one of the following holds: (1) n is even and d ≥ n − 2, (2) n is odd and
d ≥ n − 3, then
q
(d)
d
d sat
(6.1)
= dk + max q k − − 1 .
= reg I2k
reg I2k = dk, and reg I2k
2≤q≤2k−2
2
q even
Further, if neither of these conditions hold, then
Finally, if d ≤ n − 4 then
d sat
d
> dk.
≥ reg I2k
reg I2k
(6.2)
(d)
reg I2k > dk.
(6.3)
Using the notation of (2.4), let Zkd = Z Xkd . For the following lemma, recall that by ([Rai16], Lemma 5.3) we
have
n
o
0≤l≤k−1, z1 =···=zl+1 ≤d−1
Zkd = (z, l) ∈ P(m) × Z | |z|+(d−z1 )l+1≤kd≤|z|+(d−z
.
(6.4)
1 )(l+1)
Lemma 6.2. For each l with 0 ≤ l ≤ k − 1 we have using the notation of (5.3) the equality
max reg Jz,l | z ∈ P(m) and (z, l) ∈ Zkd = R2l,k,n,d .
(6.5)
Proof. The proof is similar to the proof of ([Rai16], Lemma 5.4) and proceeds by showing the equivalence
(z, l) ∈ Zkd and t ∈ Tl (z) ⇐⇒ (y, u) ∈ YU (2l, k, n, d),
where y = (zl+1 , · · · , zm )(2) and ui = 2ti for all i. Under this equivalence we have that
g2l,k,n,d (y, u) =
m
X
zi + l(z1 − 1) + |t| − fl (z, t).
i=l+1
Corollary 6.3. For every 2 ≤ 2k ≤ n and d ≥ 1, we have
d
d sat
= max Rq,k,n,d , reg S/(I2k
= max
reg S/I2k
)
0≤q≤2k−2
q even
2≤q≤2k−2
q even
(d)
Rq,k,n,d , and reg S/I2k = R2k−2,k,n,d .
(d)
d
d sat
In particular, reg I2k
≥ reg (I2k
)
≥ reg I2k .
Proof. The proof uses the Main Theorem and Lemma 6.2, and is identical to the proof of ([Rai16], Corollary
5.5).
Lemma 6.4. If 2 < 2k ≤ n − 2 and 1 ≤ d ≤ n − 4 then
R2k−2,k,n,d ≥ dk.
Proof. We break the proof into 2 cases, when d is even and when d is odd.
Case 1: d is even. Define (y, u) ∈ Pe (n − 2k + 2) × P(n − 2k + 2) via y1 = y2 = d − 1, yi = 0 for i > 2, and
u1 = u2 = 2k − 2, ui = max(2, 2k − d) for i > 2. Then
|y| = 2d − 2 = d(2k − (2k − 2)) − 2, and |y| − 2y1 = 0 = d(2k − (2k − 2) − 2).
Also, y2 − y3 = d − 1 ≥ min(2k − 4, d − 2) = u2 − u3 and yi − yi+1 = ui − ui+1 = 0 for all i 6= 2. Thus,
(y, u) ∈ YU (2k − 2, k, n, d). Conclude that R2k−2,k,n,d ≥ g2k−2,k,n,d (y, u). We will show that g2k−2,k,n,d (y, u) ≥ dk.
Suppose first that 2k − d ≥ 4, in which case ui = 2k − d for i > 2. Then
(n − 2k)(2k − d) − 4 (2k − d)
g2k−2,k,n,d (y, u) = dk + 1 +
−
.
2
2
Thus, we need (n − 2k − 1)(2k − d) − 4 ≥ −2. Since n − 2k − 1 ≥ 0 and d ≤ 2k − 4 we have that this holds. Now
suppose that 2k − d ≤ 2, so that ui = 2 for i > 2. Then
2(n − 2k) − 4
− (d − 2k + 3).
g2k−2,k,n,d (y, u) = dk + 1 +
2
REGULARITY AND COHOMOLOGY OF PFAFFIAN THICKENINGS
13
Thus, we need 2(n − 2k) − 4 − 2(d − 2k + 3) ≥ −2. Since d ≤ n − 4 this is easy to check.
Case 2: d is odd. Define (y, u) ∈ Pe (n − 2k + 2) × P(n − 2k + 2) via y1 = y2 = d − 1, yi = 0 for i > 2, and
u1 = u2 = 2k − 2, ui = max(2, 2k − d − 1) for i > 2. The proof then goes as in the previous case.
Lemma 6.5. If n is even, 2 < 2k ≤ n − 2, and d = n − 3, then
R2k−4,k,n,d ≥ dk.
Proof. Define (y, u) ∈ Pe (n − 2k + 4) × P(n − 2k + 4) via y1 = · · · = y4 = d − 1 = n − 4, y5 = y6 = 1, yi = 0 for
i > 6, and u1 = · · · = u4 = 2k − 4, ui = 0 for i > 4. Then (y, u) ∈ YU (2k − 4, k, n, d) and
R2k−4,k,n,d ≥ g2k−4,k,n,d (y, u) =
1
(4(n − 3) − 2 + (2k − 4)(d + 2)) ≥ (n − 3)k = dk.
2
Proof of Theorem 6.1. The equalities (6.1), (6.2), and (6.3) all follow from Corollary 6.3. The equality (6.1) then
follows from the optimization result Proposition 5.1. The equality (6.2) follows from Lemma 6.4 and Lemma 6.5.
The equality (6.3) follows from Lemma 6.4
Using Corollary 6.3, Lemma 5.2, and Lemma 5.3, we recover the following well-known formulas:
(d)
d
d sat
Theorem 6.6. If 2k = 2 or 2k = n then reg(I2k
) = reg((I2k
) ) = reg(I2k ) = dk for all d ≥ 1. If 2k = n − 1 then
(
dk + 12 (n − d − 4) if d is odd and d < n − 2
(d)
d
d sat
= reg I2k =
= reg I2k
reg I2k
dk
otherwise.
This completes the verification of the Theorem of Linear Resolutions of Powers of Pfaffians.
(d)
(d)
(d)
We now recall a fact which shows that the modules Ext•S (S/I2k , S) grow with d. If we let Zk = Z(Xk ) then
by ([Rai16], Lemma 5.8) we have
(d)
Zk
(d)
= {(z, k − 1) | z ∈ P(m), z1 = · · · = zk , zk + zk+1 + · · · + zm ≤ d − 1} .
(d)
(6.6)
(d+1)
Since Zk ⊂ Zkd and Zk ⊂ Zk
for all d, an immediate application of (1.4) yields the Theorem on Growth
of Ext Modules. As a corollary, we have that ExtjS (S/I, S) → HIj (S) is injective if I is unmixed. The proof of this
fact is identical to the proof of ([Rai16], Corollary 5.10).
7. Kodaira Vanishing for Pfaffian Thickenings
In [BBL+ 16], the authors prove a version of the Kodaira vanishing theorem for the thickenings of local complete
intersections which are defined by a power of the ideal sheaf. While they also show that the statement is false
for more general thickenings, Raicu recently showed that an analogue of their vanishing result holds for arbitrary
GLm (C) × GLn (C)-equivariant thickenings of determinantal varieties of m × n generic matrices [Rai16]. In this
section, we prove a similar result for GL-equivariant Pfaffian thickenings.
Theorem 7.1. If Y ⊂ P( 2 )−1 is a GL-equivariant Pfaffian thickening then
n
H q (Y, OY (−j)) = 0 for q < 2n − 4 and j > 0.
(7.1)
In particular, if we let Yred denote the underlying Pfaffian variety, and if we make the convention that codim(Sing(Yred )) =
dim(Yred ) when Yred is non-singular, then
H q (Y, OY (−j)) = 0 for q < codim(Sing(Yred )) and j > 0.
Proof. Write d = n2 . The beginning of the proof is identical to the proof of ([Rai16], Theorem 6.1): using graded
local duality ([BH98], Theorem 3.6.19) we have isomorphisms of finite dimensional vector spaces
d−1−q
H q (Y, OY (−j)) ∼
(S/I, S)−d+j , for all q.
= ExtS
Thus, in order to prove the desired vanishing statement it is enough to check by the Main Theorem that
Extd−1−q
(Jz,l , S)p = 0 for p > −d and q < 2n − 4.
S
(7.2)
Suppose that λ ∈ Zndom is a dominant integral weight such that Sλ W ∗ appears as a subrepresentation of Extd−1−q
(Jz,l , S)p .
S
Then by Theorem 3.2 we have that p = −|λ|/2, and |λ| < n(n − 1). Also by Theorem 3.2 there exists t ∈ Tl (z)
14
MICHAEL PERLMAN
(2)
(2)
such that λ ∈ W (z, l, t). If tn−2l = 0 or l = 0, then λn = zn + n − 1, so that |λ| ≥ n(n − 1) + nzn ≥ n(n − 1), a
contradiction. Thus, we may assume that l ≥ 1 and ti ≥ 1 for all i = 1, · · · , n − 2l. We know that by Theorem 3.2,
n−2l
X
n
n
2l
ti .
−1−q =
−
−2
2
2
2
i=1
Notice that
n−2l
X
2l
2l
ti ≤ −
− 2(n − 2l) = −2n − 2l2 + 5l ≤ −2n + 3.
−
−2
2
2
i=1
Thus, −1 − q ≤ −2n + 3, i.e. q ≥ 2n − 4, which proves (7.1).
For the second assertion we note that Yred is non-singular if and only if its defining ideal is I4 , in which case it
is isomorphic to the Grassmannian G(2, n). Since dim(G(2, n)) = 2n − 4, we see that the result holds in this case.
If the defining ideal of Yred is I2k with 2k ≥ 6, then codim(Sing(Yred )) = 2n − 4k + 5 < 2n − 4.
Acknowledgments
The author is very grateful to Claudiu Raicu for his guidance while this work was done. Experiments were
conducted using the computer algebra software Macaulay2 [GS]. The author was supported by the National
Science Foundation Graduate Research Fellowship under Grant No. DGE-1313583.
References
[ADF80] S Abeasis and A Del Fra. Young diagrams and ideals of pfaffians. Advances in Mathematics, 35(2):158–178, 1980.
[BBL+ 16] Bhargav Bhatt, Manuel Blickle, Gennady Lyubeznik, Anurag K Singh, and Wenliang Zhang. Stabilization of the cohomology
of thickenings. arXiv preprint arXiv:1605.09492, 2016.
[BH98]
Winfried Bruns and H Jürgen Herzog. Cohen-macaulay rings. Cambridge University Press, 1998.
[BS92]
Giandomenico Boffi and Rafael Sánchez. On the resolutions of the powers of the pfaffian ideal. Journal of Algebra, 152(2):463–
491, 1992.
[EMS00] David Eisenbud, Mircea Mustaţa, and Mike Stillman. Cohomology on toric varieties and local cohomology with monomial
supports. Journal of Symbolic Computation, 29(4-5):583–600, 2000.
[GS]
Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2/ .
[JPW81] Tadeusz Józefiak, Piotr Pragacz, and Jerzy Weyman. Resolutions of determinantal varieties and tensor complexes associated
with symmetric and antisymmetric matrices. Astérisque, (87-8):109–189, 1981.
[KU92]
Andrew R Kustin and Bernd Ulrich. A family of complexes associated to an almost alternating map, with applications to
residual intersections. Number 461. American Mathematical Soc., 1992.
[Rai16]
Claudiu Raicu. Regularity and cohomology of determinantal thickenings. arXiv preprint arXiv:1611.00415, 2016.
[RW16]
Claudiu Raicu and Jerzy Weyman. Local cohomology with support in ideals of symmetric minors and pfaffians. Journal of
the London Mathematical Society, page jdw056, 2016.
[RWW14] Claudiu Raicu, Jerzy Weyman, and Emily E Witt. Local cohomology with support in ideals of maximal minors and submaximal pfaffians. Advances in Mathematics, 250:596–610, 2014.
[Wey03] Jerzy Weyman. Cohomology of vector bundles and syzygies, volume 149. Cambridge University Press, 2003.
Department of Mathematics, University of Notre Dame, Notre Dame, IN 46556
E-mail address: [email protected]
| 0math.AC
|
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
arXiv:1801.09266v1 [math.GR] 28 Jan 2018
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Abstract. Let G be a simple algebraic group over an algebraically closed field. A closed subgroup H of
G is called G-completely reducible (G-cr) if, whenever H is contained in a parabolic subgroup P of G, it
is contained in a Levi factor of P . In this paper we complete the classification of connected reductive G-cr
subgroups when G has exceptional type, by determining the L0 -irreducible connected reductive subgroups
for each simple classical factor L0 of a Levi subgroup of G. As an illustration, we enumerate all reducible, Gcr semisimple subgroups when G has type F4 , as well as various properties of these. This work complements
existing results of Lawther, Liebeck, Seitz and Testerman, and is vital in classifying non-G-cr subgroups, a
project currently being undertaken elsewhere by the authors.
1. Introduction
Let G be a connected reductive algebraic group defined over an algebraically closed field K. Following
Serre [21], a closed subgroup X of G is said to be G-completely reducible (G-cr) if whenever X is contained
in a parabolic subgroup P of G, it is contained in a Levi subgroup of P . When G = GL(V ) or SL(V ), a
subgroup X is G-cr if and only if V is a completely reducible KX-module, and thus G-complete reducibility
generalises the standard notion in representation theory. Similarly, X is called G-irreducible if X is not
contained in any proper parabolic subgroup of G. The G-cr subgroups of G are precisely the L-irreducible
subgroups of L, as L ranges over the Levi subgroups of G [4, Corollary 3.5]; the case L = G yields the
G-irreducible subgroups.
In this paper, we enumerate and study the L-irreducible connected subgroups, as L varies over proper
Levi subgroups of a simple algebraic group G of exceptional type. This forms part of a long-running project,
which ultimately seeks to determine all reductive connected subgroups of all reductive algebraic groups.
Together with work of the second author and many others [1, 11, 12, 14, 24, 23, 26, 27, 25], the results here
complete the classification of G-cr connected subgroups for G of exceptional type. In particular, the cited
papers of the second author determine all G-irreducible connected subgroups for G of exceptional type,
and so it remains only to describe L-irreducible connected subgroups for L involving factors of classical
type.
As well as being a part of the classification of reductive subgroups, the information given here is also vital
in the complementary problem of understanding non-G-cr subgroups. An approach to this latter problem,
pioneered in [24] and built upon by the present authors [17, 16], is as follows. For each parabolic subgroup
P of G, and for a fixed Levi factor L of P , take the (known) list of L-irreducible subgroups of L, up to
L-conjugacy. For each such subgroup X, consider the set of complements to the unipotent radical Ru (P )
in the semidirect product Ru (P )X. As X varies, these complements give rise to all non-G-cr subgroups
of G, modulo the subtle but tractable problem of understanding the difference between ‘G-conjugacy’ and
‘P -conjugacy’.
To state our main theorem, recall that for each G-cr subgroup X ≤ G, there exists a Levi subgroup L,
minimal among Levi subgroups containing X, and the conjugacy class of L in G is uniquely determined
by the conjugacy class of X (cf. Proposition 6.1).
2010 Mathematics Subject Classification. Primary 20G07, 20G41; Secondary 20G10.
The first author is supported by the Alexander von Humboldt Foundation.
1
2
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Theorem 1. Let G be a simple algebraic group of exceptional type over an algebraically closed field, let
X be a G-completely reducible connected subgroup of G, and let L be minimal among Levi subgroups of G
containing X. Then either L = G, or the image of X under projection to each classical simple factor of L
is given by an embedding in Section 5.2.
Remark 1.1. Irreducible subgroups of a fixed simple algebraic group form a poset under inclusion, and
in Section 5.2 we give the poset structure of subgroups in each simple factor. The subgroups are given
up to conjugacy in the simple factor, except for type D7 , where for space reasons we list subgroups up to
conjugacy in the full automorphism group of the factor. A Levi factor of type D7 occurs for exceptional G
only when G = E8 , in which case G induces the full automorphism group of this subgroup. In Section 6.1
we describe how the conjugacy classes in the other factors fuse under outer automorphisms.
Note that when L = G or when L has a factor of exceptional type, the image of X in this exceptional
group is irreducible, and therefore appears in existing work of the second author [26, 27, 25]. Since we
give the poset of irreducible subgroups in each possible Levi factor, one can construct the entire poset of
reducible G-cr subgroups from the information in this paper (cf. Lemma 3.3). We illustrate this in Section
6, by constructing the poset of reducible G-completely reducible subgroups when G = F4 (K), in arbitrary
characteristic. We also give a series of lemmas regarding G-completely reducible subgroups, allowing us to
derive much information on the subgroups which occur. The following summarises this information.
Theorem 2. Let G = F4 (K), where K is algebraically closed of arbitrary characteristic, and let X be a
connected G-completely reducible subgroup of G. Then either X is G-irreducible, or the semisimple part of
X is conjugate to precisely one of the subgroups listed in Table 38, each of which is G-completely reducible.
For each reducible semisimple subgroup X, we give the action of X on the adjoint module L(G), the
connected centraliser CG (X)◦ , as well as whether X is separable in G and whether (G, X) is a reductive
pair.
In addition to its illustrative purpose, Theorem 2 also corrects some omissions in [24, Corollary 5]; more
details are given in Remark 6.4.
2. Notation
We now present the notation used throughout the paper, especially in Sections 5 and 6.4. Since this
paper complements [25] we remain consistent with the notation there, which we recall for the convenience
of the reader.
By a subgroup of an algebraic group, we always mean a Zariski-closed subgroup, and a representation is
taken to mean a rational representation. An algebraic group is called simple if it has no proper, connected
normal subgroups, and it is called semisimple if it is connected and is a commuting product of simple
groups. Let G be a simple algebraic group over an algebraically closed field K. We fix a maximal torus T
of G, and a corresponding set of roots Φ. Let Φ+ be a choice of positive roots, with corresponding simple
roots Π = {α1 , . . . , αl } and fundamental dominant weights {λ1 , . . . , λl }. We use Bourbaki numbering [5,
p. 250]. We sometimes use a1 a2 . . . al to denote a dominant weight a1 λ1 + a2 λ2 + · · · + al λl . For G-modules
V and W we write V + W for the module V ⊕ W and let V ∗ denote the dual module of V . We denote
by VG (λ) (or just λ) the irreducible G-module of dominant high weight λ. Similarly, the Weyl module
and tilting module of high weight λ will respectively be denoted W (λ) and T (λ). Given dominant weights
µ1 , µ2 , . . . , µn , we define T (µ1 ; µ2 ; . . . ; µn ) to be the tilting module with the same composition factors as
W (µ1 )+W (µ2 )+· · · +W (µn ), when this exists and is uniquely determined. For example, when G is of type
A1 we use the notation T (6; 2) for p ≥ 5 only. In this case T (6; 2) = 6 + 2 when p ≥ 7 and T (6; 2) = T (6)
when p = 5. We use L(G) to denote the adjoint module for G. The notation X̄ denotes a subgroup of G
that is generated by long root subgroups of G. If the root system of G has short roots then X̃ denotes a
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
3
subgroup generated by short root subgroups of G. Let J = {αj1 , αj2 , . . . , αjr } ⊆ Π and define ΦJ = Φ∩ZJ.
Then we denote the standard Levi subgroup corresponding to J by Lj1 j2 ...jr = hT, Uα : α ∈ ΦJ i.
Suppose now that G = G1 . . . Gm is semisimple and that char(K) = p > 0. Let F be the standard
Frobenius endomorphism of G, acting on root groups Uα = {uα (c) | c ∈ K} by F (uα (c)) = uα (cp ), and let
V be the G-module afforded by a representation ρ : G → GL(V ). Then for a non-negative integer r, V [r]
def
denotes the ‘twisted’ module afforded by the representation ρ[r] = ρ ◦ F r .
We let (V1 , . . . , Vm ) denote the G-module V1 ⊗ · · · ⊗ Vm where Vi is a Gi -module for each i, and we
often use 0 to denote the trivial module for G. Let M1 , . . . , Mk be G-modules and n1 , . . . , nk be positive
integers. Then M1n1 / . . . /Mknk denotes a G-module having the same composition factors as M1n1 +· · ·+Mknk .
Furthermore, V = M1 | · · · | Mk denotes a G-module with socle series V = V1 ⊇ V2 ⊇ . . . ⊇ Vk+1 = {0},
with Vi−1 /Vi ∼
= Mi for i = 1, . . . , k. The following notation is used heavily within Section 5. Suppose µi is
a dominant weight for each Gi then we define
def
Θ(µ1 ; . . . ; µn ) = 0 | ((VG1 (µ1 ), 0, . . . , 0) + · · · + (0, . . . , 0, VGn (µn ))) | 0.
In other words, Θ(µ1 ; . . . ; µn ) has a unique minimal submodule, which is trivial, and a unique maximal
submodule, with trivial corresponding quotient, and the quotient of the maximal submodule by the minimal
submodule is a direct sum of modules, one for each simple factor of G.
We need a notation for diagonal subgroups of a semisimple group Y = H1 H2 . . . Hk , where the Hi are
simple and all of the same type, with simply-connected cover H. Then the simply-connected cover of Y is
Ŷ = H × H × . . . × H, and Hi is the image of the i-th factor H under the projection Ŷ → Y . A diagonal
subgroup of Ŷ is a subgroup X̂ ∼
= H of the form {(φ1 (h), . . . , φk (h)) | h ∈ H}, where each φi is a surjective
endomorphism of H. A diagonal subgroup of Y is the image of a diagonal subgroup of Ŷ under Ŷ → Y .
To describe such a subgroup up to conjugacy, it suffices to specify the surjective endomorphism φi for each
i. By [10, Section 1.15], φi = αi θi F ri for some inner automorphism αi , a graph automorphism θi , and a
power F ri of the standard Frobenius endomorphism. Since we only distinguish these diagonal subgroups
up to conjugacy, we assume that each αi is trivial and give a (possibly trivial) graph automorphism θi of
H and a non-negative integer ri , for each 1 ≤ i ≤ k. Such a diagonal subgroup X is denoted by
[θ r1 ]
X ֒→ Y via (λ1 1
[θ r2 ]
, λ1 2
[θ rk ]
, . . . , λ1 k
),
where λ1 is the first dominant weight of X. We often abbreviate this to
[θ r1 ]
X via (λ1 1
[θ rk ]
, . . . , λ1 k
)
if the group Y is clear. Unless X is of type Dn (n ≥ 4), a graph automorphism is determined up to inner
automorphism by the image of λ1 (this includes the exceptional graph automorphisms of B2 and F4 when
p = 2 and of G2 when p = 3, which send λ1 to 2λ2 , 2λ4 and 3λ2 , respectively). In these cases, instead
[θ r ]
of writing λ1 i i we write µ[ri ] where µ is the image of λ1 under θi . For clarity, note that the field twists
r, s, t, . . . are not automatically assumed to be distinct. This is consistent with [27, 25] but not with [26].
We extend this notation to describe certain semisimple subgroups of the form X = X1 X2 . . . Xn of
Y = H1 H2 . . . Hk (n < k) where each Xi has the same type as H and the projection of X to each Hi is
surjective. Any such subgroup is a commuting product of diagonal subgroups of distinct subsets of the
Hi . For this reason, we extend our use of the term “diagonal subgroup” to include such subgroups X. For
example, consider diagonal subgroups isomorphic to A21 contained in A41 . They are either a commuting
product of one factor A1 and a diagonal subgroup of the remaining three factors (A31 ), or a commuting
product of a diagonal subgroup of A21 and another diagonal subgroup of the other A21 . Therefore, our
notation needs to distinguish which of the subgroups Hi each simple factor of X maps non-trivially to.
We give the first factor of X the label ‘a’, the second factor of X the label ‘b’ and so on. Then for each i
[θ r ]
such that X1 has non-trivial projection to Hi we give a subscript a to λ1 i i . For each j such that X2 has
4
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
[θ r ]
non-trivial projection to Hj we give a subscript b to λ1 j j and so on. For example, consider X = A1 A1
and Y = A41 with the first A1 factor of X embedded diagonally in the first two factors of Y (with field
twists 0 and r) and the second A1 factor embedded diagonally in the last two factors of Y (with field twists
[r]
[s]
0 and s). Then we write X ֒→ Y via (1a , 1a , 1b , 1b ).
We make another natural extension of this notation. Let X and Y be as above and suppose that
S = S1 . . . Sl is a semisimple group with no factors isomorphic to H. Then a subgroup XS of Y S is
denoted by
[θ r1 ]
XS ֒→ Y S via (λ1 1
[θ r2 ]
, λ1 2
[θ rk ]
, . . . , λ1 k
, ν11 , ν12 , . . . , ν1l ),
where ν1i denotes the first fundamental dominant weight of Si . Again, we will still refer to such subgroups
as “diagonal subgroups”. For example, consider a subgroup XZ = A1 B2 of Y Z = A21 B2 where X = A1
is a simple diagonal subgroup of Y = A21 with field twists 0 and r. Then we write A1 B2 ֒→ A21 B2 via
(1, 1[r] , 10).
Finally, we show how to combine all of these notations together. Suppose that A is a diagonal subgroup
of Y = H1 . . . Hn and B is a diagonal subgroup of Z = J1 . . . Jm , where all the Hi (resp. Ji ) are simple and
of the same type H (resp. J) and H is not isomorphic to J. Suppose also that S is a semisimple group
with no factors of type H or J. Then we naturally concatenate the notations above to give a notation for
the subgroup ABS of Y ZS. For example, consider a subgroup ABS = A31 B22 B3 of Y ZS = A41 B23 B3 where
[t]
[s]
[r]
A is a diagonal subgroup of Y via (1a , 1b , 1b , 1c ) and B is a diagonal subgroup of Z via (10a , 10b , 10b ).
[t]
[s]
[r]
Then we write A31 B22 B3 ֒→ A41 B23 B3 via (1a , 1b , 1b , 1c , 10a , 10b , 10b , 100).
In Tables 2–14 we give an identification number to each of the conjugacy classes of L′ -irreducible connected subgroups. The notation L′ (#a) (or simply a if L is clear from the context) means the L′ -irreducible
subgroup corresponding to the ID number a. We set L′ (#0) to be L′ itself. Sometimes L′ (#a) will refer
to infinitely many conjugacy classes of L′ -irreducible subgroups, depending on some field twists r1 , . . . , rk
and graph automorphisms θ1 , . . . , θk . Sometimes we refer to a subset of the conjugacy classes that L′ (#a)
represents; we only do this when all graph automorphisms are trivial. These are therefore described by an
ordered set of field twists r1 , . . . , rk and are denoted by L′ (#a{r1 ,...,rk } ).
Let us illustrate the above with a concrete example. Consider A6 (#5), the conjugacy classes of diagonal
subgroups A1 ֒→ A1 A1 via (1[r] , 1[s] ) (rs = 0; r 6= s) (see Table 6). Then the notation A6 (#5{r,0} ) refers
to the conjugacy classes with s = 0 and the notation A6 (#1{1,0} ) refers to the single conjugacy class
A1 ֒→ A1 A1 via (1[1] , 1).
In the tables in Section 5 we use a shorthand for n{s1 ,...,sk } in certain situations. The notation n {0 }
simply means that each si is equal to 0. The notation n{δj } means sj = 1 and si = 0 for all i 6= j.
3. Preliminaries
We now give a series of results used in the proof of Theorem 1. In Section 3.1 we give a criterion for
identifying irreducible subgroups of classical simple algebraic groups. In Section 3.2 we enumerate the
maximal connected subgroups of subgroups of type A6 and D7 , which are used in deriving the subgroup
poset structure of subgroups in Section 5. The maximal subgroups of classical groups of other types can
be found in [25, Lemma 3.3].
3.1. Identifying L0 -irreducible subgroups. The following result is central in our proof of Theorem 1,
as it provides a method of identifying L0 -irreducible subgroups when L0 is simple of classical type. The
result is essentially well-known, but is usually stated in the literature as a one-way implication, see for
example [15, Lemma 2.2]. We therefore provide a proof for completeness.
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
5
Proposition 3.1. Let G be a simple algebraic group of classical type in characteristic p ≥ 0, let V =
VG (λ1 ), and let X be a subgroup of G. Then X is G-irreducible if and only if one of the following holds:
(i) G has type An and X acts irreducibly on V ;
(ii) G has type Bn , Cn or Dn and V ↓ X = V1 ⊥ . . . ⊥ Vk with the Vi all non-degenerate, irreducible
and pairwise inequivalent as X-modules;
(iii) p = 2, G has type Dn and X fixes a non-singular vector v ∈ V , such that X is Gv -irreducible and
does not lie in a subgroup Dn−1 of Gv .
Remark 3.2. In part (iii), the subgroup Gv is simple of type Bn−1 , and X lies in a subgroup Dn−1 of
Gv if and only if the non-degenerate bilinear form on the Gv -module hvi⊥ / hvi is the polarisation of an
X-invariant quadratic form.
Proof. It is well-known that for G = SL(V ), Sp(V ) or SO(V ), the parabolic subgroups of G are respectively
the stabilisers of flags of subspaces, flags of totally isotropic subspaces, and flags of totally singular subspaces
[19, Proposition 12.13]. Thus it is clear that if (i) holds then X is G-irreducible. Similarly, if (ii) holds
then every X-invariant subspace of V is a direct sum of some of the factors Vi . Every such subspace being
non-degenerate, it follows that X is G-irreducible.
Next, suppose that (iii) holds, and suppose for a contradiction that X lies in a parabolic subgroup of
G. Then X stabilises a nonzero totally singular subspace W of V . Note that W ∩ hvi = {0} since v is
non-singular, and so W ∩ hvi⊥ maps isomorphically onto its image in hvi⊥ / hvi. This image is a totally
isotropic subspace of hvi⊥ / hvi, and since X is Gv -irreducible we conclude that W ∩ hvi⊥ = {0}. Since
hvi⊥ has codimension 1 in V , this implies that W has dimension at most 1, hence exactly 1 as W is
nonzero. Thus V ↓ X = hvi⊥ ⊕ W as X-modules. Next, (W + hvi)⊥ is an X-submodule contained with
codimension 1 in hvi⊥ . Moreover, if W = hwi then evaluating (av + bw, cv + dw) = (ac + bd)(w, v) = 0 for
all a, b shows that the only combination of v and w orthogonal to both v and w is the zero vector, hence
(W + hvi)⊥ ∩ (W + hvi) = {0}, so (W + hvi)⊥ is non-degenerate. This contradicts the assumption that X
does not lie in a subgroup of type Dn−1 in Gv , and we conclude that X is G-irreducible.
Conversely, suppose X is G-irreducible and that neither (i) nor (iii) hold. If p = 2, G = Dn and X
fixes a non-singular vector v on V , then since (iii) does not hold, X lies either in a subgroup Dn−1 of G,
or in a parabolic subgroup of Gv ; each of these contradicts the hypothesis that X is G-irreducible. So X
does not fix any non-singular vector of V . Thus the bilinear form is non-degenerate on each irreducible
X-submodule, so each irreducible X-submodule has an orthogonal complement, and V ↓ X is completely
reducible. Let U be an irreducible X-submodule of V , and suppose that W ⊆ U ⊥ is an irreducible
X-submodule with U ∼
= W . Since U and W are irreducible, their X-invariant non-degenerate bilinear
or quadratic forms are uniquely determined up to a constant, hence there exists an isometric X-module
isomorphism φ : U → W . For a scalar λ ∈ K, consider the X-invariant subspace {u + λφ(u) : u ∈ U }.
Since U and W are orthogonal, we have (u + λφ(u), u′ + λφ(u′ )) = (1 + λ2 )(u, u′ ). In particular, taking λ
to be a square root of −1 in K, we obtain a totally isotropic submodule. This contradiction shows that
U ⊥ does not contain any submodule isomorphic to U , and thus (ii) holds.
Lemma 3.3. Let L be a Levi subgroup of G, with decomposition L = L1 L2 . . . Lm Z(L), where the Li are
simple and commute pairwise. Let Li denote the image of the projection of L onto each simple factor.
Then a subgroup X < L is L-irreducible if and only if the image of X in each factor Li is Li -irreducible.
Proof. Since L is connected reductive, X is L-irreducible if and only if X does not normalise a nontrivial
connected unipotent subgroup of L. This is the case if and only if, for all i = 1, . . . , m, the image of
X in Li does not normalise any nontrivial connected unipotent subgroup of Li , that is, if this image is
Li -irreducible.
6
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
3.2. Maximal subgroups of classical groups of low rank. Proposition 3.1 lets us identify irreducible
subgroups of classical groups directly through representation theory. However, since we are also interested
in the complete lattice of G-cr subgroups, it will be of use for us to know which subgroups arising in
Proposition 3.1 are maximal. This allows us to work recursively through chains of maximal irreducible
subgroups. For many groups of low rank, this information has been calculated and recorded in [25, Lemma
3.3]; we additionally require the maximal subgroups of a simple algebraic group of type A6 and D7 .
def
def
Lemma 3.4. Let X ≤ S = SO(V ) where dim V is even, and write G = GO(V ) = SO(V ) hτ i, where τ
induces a graph automorphism on SO(V ). Then X G splits into two classes of subgroups of S if and only
if NG (X) ≤ S.
Proof. Consider the transitive action of G on the set X G /S of S-conjugacy classes of G-conjugates of X.
A point stabiliser is conjugate to SNG (X), and therefore |X G /S| = |G : SNG (X)| = 1 or 2, depending on
whether or not NG (X) ≤ S.
Our classification is based on [13], which separates maximal subgroups of classical groups into classes,
of which we require the following. Write G = Cl(V ) to denote that G is one of the groups SL(V ), SO(V )
or Sp(V ).
Class C1 : Subspace stabilisers. Here H ∈ C1 if H = StabG (W ) where W is either a non-degenerate
subspace of V (V is given the zero form if G = SL(V )), or (G, p) = (SO(V ), 2) and W is a non-singular
subspace of dimension 1.
Class C4 : Tensor product subgroups. Suppose that V = V1 ⊗ V2 with dim Vi > 1 for each i. Then
def
H ∈ C4 if H = Cl(V1 ) ◦ Cl(V2 ), acting on V by (g1 , g2 )(v1 ⊗ v2 ) = (g1 v1 ) ⊗ (g2 v2 ). The subgroups occurring
in this way are:
SL ⊗ SL < SL,
Sp ⊗ SO < Sp (p 6= 2),
Sp ⊗ Sp < SO,
SO ⊗ SO < SO (p 6= 2).
The following result can be deduced immediately from [13, Theorem 1] (cf. also [25, Theorem 3.2]).
A ‘restricted’ irreducible module here is a module whose highest weight, when expressed as a sum of
fundamental dominant weights, has coefficients strictly less than p.
Lemma 3.5. Let G = Cl(V ) be a classical simple algebraic group. Suppose that M is a reductive, maximal
connected subgroup of G. Then one of the following holds:
(i) M belongs to C1 ;
(ii) M belongs to C4 ;
(iii) M is a simple algebraic group and V ↓ M is irreducible and restricted.
Proposition 3.6. Let G be a classical simple algebraic group of type A6 or D7 . Then Table 3.2 gives all
G-conjugacy classes of reductive, maximal connected subgroups of G.
Proof. For G of type A6 , the class C1 contains no G-irreducible subgroups, and the class C4 is empty since
the natural module is 7-dimensional. Thus a maximal subgroup M falls into case (iii) of Lemma 3.5. The
only simple algebraic groups with a 7-dimensional irreducible module are those of type A6 , and of type B3
and G2 when p 6= 2. The module is then unique up to taking a Frobenius twist, and so G contains a unique
subgroup of each of these types up to conjugacy. Moreover a subgroup of type B3 contains a subgroup of
type G2 , and it follows that every subgroup of G of type G2 is contained in a subgroup of type B3 .
For G of type D7 , the G-irreducible subgroups in class C1 are of type B6 , A1 B5 , Ā21 D5 , B2 B4 , A3 D4 and
B32 , respectively corresponding to orthogonal decompositions of the 14-dimensional module into spaces of
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
7
Table 1. The maximal subgroups of certain low rank classical algebraic groups.
G Maximal subgroup M
A6 B3 (p 6= 2)
D 7 B6
A1 B5 (p 6= 2)
Ā21 D̄5
B2 B4 (p 6= 2)
Ā3 D̄4
B32 (p 6= 2)
C3 (p 6= 3) (2 classes)
B2 (p 6= 2, 5) (2 classes)
G2 (p ≥ 5) (2 classes)
VG (λ1 ) ↓ M
100
λ1 + 0 (p 6= 2) or 0|λ1 |0 (p = 2)
(2, 0) + (0, λ1 )
(1, 1, 0) + (0, 0, λ1 )
(10, 0) + (00, λ1 )
(010, 0) + (000, λ1 )
(100, 000) + (000, 100)
010
20
01
Comments
classes permuted by graph aut.
classes permuted by graph aut.
classes permuted by graph aut.
dimension i and 14 − i for i ∈ {1, 3, 4, 5, 6, 7} when p 6= 2. When p = 2, the subgroups A1 B5 , B2 B4 and
B32 each fix a non-singular vector, and are therefore contained in a subgroup B6 .
The prime decomposition 14 = 2×7, and the fact that there exists no non-degenerate orthogonal module
of dimension 2 or symplectic module of dimension 7, shows that the class C4 is empty. It remains to consider
the subgroups in case (iii) of Lemma 3.5. By [18] the restricted, irreducible, 14-dimensional modules for
a simple algebraic group are as follows: VB2 (02) (p 6= 2, 5), VG2 (01) (p 6= 3), VA3 (101) (p = 2), VB3 (010)
(p = 2), VC3 (010) (p 6= 3), and VC3 (001) (p 6= 2). Using [22, Lemma 79], it is straightforward to calculate
that if p 6= 2, then all but the last of these modules supports a non-degenerate quadratic form, and therefore
gives rise to a subgroup of G (the final module is symplectic). If p = 2 then inspecting [20, Table 1, pp.
282–283], in particular numbers IV7 , IV8 and S7 there, shows that the relevant modules here still support
V
a non-degenerate quadratic form. Note that, since VB3 (010) is a direct summand of 2 (VB3 (100)), the
representation factors through the morphism B3 → C3 , so the image in D7 is a subgroup of type C3 , and
is thus conjugate to the subgroup given by VC3 (010). When p = 2, a subgroup C3 contains a subgroup A3
acting via VA3 (101) and a subgroup G2 acting via VG2 (01), and so these latter subgroups are not maximal
in G. Thus when p = 2 the only possible maximal subgroups as in part (iii) of Lemma 3.5 are of type C3 .
When p 6= 2, the possible subgroups (of type B2 , G2 and C3 ), are not contained in one another, and so
hence each gives rise to a maximal subgroup of G.
There is a unique subgroup class of each of these three types in GO14 (K), and each has trivial group of
algebraic outer automorphisms. Thus if X is such a subgroup, then either NGO14 (K) (X) ≤ SO14 (K) or X
centralises a non-trivial element of GO14 (K) \ SO14 (K). The latter does not occur, since the centraliser of
such an element is the stabiliser of an orthogonal decomposition of the natural module, and X does not
preserve such a decomposition. Thus by Lemma 3.4, X gives rise to two classes of subgroups of SO14 (K),
hence to two classes of subgroups of G.
4. Proof of Main Theorem
In this section we describe the method used for proving Theorem 1, i.e. for determining the poset of
L0 -irreducible connected subgroups of all simple factors L0 of a Levi subgroup of an exceptional algebraic
group G. It is very similar to, yet more straightforward than, the strategy described in [25, Section 4].
Moreover, the strategy is identical for each type of subgroup L0 . For this reason, once we have described
the method of proof, we give explicit details only for L0 of type A3 .
We iterate the following process for each L0 . Firstly, we find all of the reductive, maximal connected
subgroups of L0 . Since L0 is simple this is immediate from Lemma 3.5 and [25, Lemma 3.3]; these
8
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
subgroups are automatically L0 -irreducible. Let M be the first reductive, maximal connected subgroup of
L0 . We now find the L0 -irreducible maximal connected subgroups M1 of M . To do this, we again find
the reductive, maximal connected subgroups of M , using [25, Lemma 3.3] for classical factors of M and
[2, Corollary 12] for factors of type G2 . Then Proposition 3.1 and Lemma 3.3 allow us to check whether
these maximal connected subgroups are L0 -irreducible. We now repeat this process for each L0 -irreducible
maximal connected subgroup of M . Continuing in this manner will yield all L0 -irreducible subgroups of
L0 . However, we could find the same conjugacy class of L0 -irreducible connected subgroups more than
once. So at each step, we check to see if each subgroup arising has been previously found. If a conjugate
has not been found before, we give it a new unique ID number L0 (#n), and if it has been found before,
we use its existing ID number. To check for these repeats we consider the action of X on VL0 (λ1 ). This
determines the L0 -class of X, unless possibly when L0 has type Dn . In such a case there may be two
conjugacy classes of subgroups acting the same way on VL0 (λ1 ) and we use Lemma 3.4 to decide whether
this is so. At this stage, we now know all of the connected overgroups of each L0 -irreducible subgroup,
and thus understand the poset structure of the irreducible connected subgroups of L0 .
The information obtained from this method is displayed in Tables 2–14. The notation used in these
tables is explained in Section 2, and we explain how to read the tables at the start of Section 5. In
particular, when the identification number n for a subgroup X is written in italics it means that X is
listed elsewhere in the table, hence should not be considered if one wants exactly one conjugate of each
G-irreducible connected subgroup.
We modify the above approach when considering diagonal subgroups. Suppose that X is an L0 irreducible connected subgroup of the form An B for some n ≥ 2 with A and B of different types. Then X
has maximal diagonal subgroups of the form An−1 B. We do not want to then consider all of their maximal
subgroups including those of the form An−2 B. Instead, we list all diagonal subgroups of X immediately
and do not list their subgroups in the table. Doing this significantly reduces the size of the tables, without
missing any L0 -irreducible subgroups. It does however mean that some additional combinatorial work is
required to recover the lattice of overgroups of certain diagonal subgroups. In light of the large number of
diagonal subgroups, we often produce a complete set of non-conjugate classes of diagonal subgroups of an
irreducible subgroup X in a supplementary table. This allows for easier reading of the tables in Section 5.
There is another situation where we deviate slightly from the above approach, but only to alter the order
in which we study the subgroups in question. Let us explain this with an example. Suppose L0 is of type
D4 and for illustration think of this as SO8 . When p 6= 2 there is a maximal connected subgroup SO5 SO3 .
When p = 2, this subgroup still exists and is still SO8 -irreducible, but is now contained in SO7 . In view
of this, we use the same identification number n for both groups but use na for the subgroup when p 6= 2
and nb for the subgroup when p = 2. This allows us to study the subgroups of na and nb together. In the
tables, if we first arrive at the subgroup labelled nb, we postpone listing its subgroups until reaching the
subgroup labelled na.
This generalises to any situation where a subgroup X occurs for p 6= m somewhere in the lattice and
elsewhere when p = m. In particular, we use this only if the composition factors on VL0 (λ1 ) for nb are the
same as those for na when p = m.
4.1. The irreducible subgroups of L0 = A3 . The reductive, maximal connected subgroups of A3 are
B2 and A21 (p 6= 2) acting via 01 and (1, 1) on V = VA3 (λ1 ), respectively. These are L0 -irreducible by
definition. The module (1, 1) yields an irreducible A21 subgroup of A3 , even when p = 2. We will shortly
find this subgroup contained in B2 and this explains why we write 2a for the ID number of A21 (p 6= 2) in
Table 3. Next, we consider the reductive, maximal connected subgroups of B2 = A3 (#1). These are A21 , A21
(p = 2) and A1 (p ≥ 5) acting as (1, 0) + (0, 1), (1, 1) and 3 on V , respectively. The first A21 subgroup is not
L0 -irreducible by Proposition 3.1, since it acts reducibly on V . The second A21 subgroup is A3 -irreducible
and given ID number 2b because of the explanation above. The subgroup A1 (p ≥ 5) is also A3 -irreducible
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
9
and given ID number 3. We now need to consider the A3 -irreducible connected subgroups of A3 (#2b) = A21
(p = 2) and A3 (#3) = A1 . For presentation reasons in the tables (as explained in Section 5), we postpone
studying the irreducible subgroups of A3 (#2b) and instead consider the subgroups of both A3 (#2a) and
A3 (#2b) together later. Since A1 contains no proper reductive, maximal connected subgroups, in this
case we in fact now turn immediately to A3 (#2) = A21 (this means the class of subgroups acting via (1, 1)
with no restriction on the characteristic). The reductive, maximal connected subgroups are diagonal A1
subgroups. These are embedded via (1[r] , 1[s] ) (rs = 0). These then act on V via 1[r] ⊗ 1[s] . Therefore, they
are A3 -irreducible if and only if r 6= s. Furthermore, there is an element of NA3 (A21 ) = (A21 ).2 swapping the
two A1 factors. Therefore, up to A3 -conjugacy we have the L0 -irreducible subgroups A1 ֒→ A21 via (1, 1[r] )
(r 6= 0). These are represented by the ID number 4.
5. Tables of Subgroups
In this section we present the tables referred to in Theorem 1 and we now explain how to obtain all of
the information contained within them.
5.1. Reading the Tables. There is one table for each simple classical factor L0 of a Levi subgroup of
an exceptional algebraic group, except for L0 of type A1 , where no table is required. The subgroups are
given up to L0 -conjugacy, except from in Table 14 where the subgroups of D7 are given up to Aut(D7 )conjugacy. Each table is broken into sections divided by single horizontal lines. There is one section for
each L0 -irreducible connected subgroup X of L0 , such that X contains a proper L0 -irreducible connected
subgroup. This section gives the L0 -irreducible maximal connected subgroups of X, as well as all L0 irreducible diagonal subgroups of X. There is a heading for each section that gives the type of the
subgroup being considered, as well as the identification number and any restrictions on the characteristic
p. There is one more piece of information in the heading which is the “Mi =” for i = 0, 1, . . . (where
M0 is simply written as M ). This is intended to make it easier for the reader to follow the tables, and is
explained below.
We now describe what the entries of a generic row represent. Firstly, there are headers mentioned above.
There are two other types of rows, both of which correspond to L0 -irreducible subgroups Y contained in
X, where X is the subgroup of the current section. In both cases, the entry in the first column gives
the ID number for the conjugacy class or classes of such irreducible subgroups. The second and third
column differ, depending on whether Y is a diagonal subgroup of X or not. If Y is a diagonal subgroup
of X then the second and third columns are merged and the information given is of the form “Y via . . . ”,
denoting the usual notation for an embedding of a diagonal subgroup of X, as well as any restrictions on the
characteristic p. If Y is not a diagonal subgroup of X then the second column gives the isomorphism type of
Y and any restrictions on the characteristic p. The third column contains the restriction of VL0 = VL0 (λ1 )
to Y . For a diagonal subgroup Y of X, it is straightforward to work out VL0 ↓ Y from VL0 ↓ X. We note
that we do not repeat restrictions on p as we list the subgroups of an L0 -irreducible connected subgroup
X. So if, for example, the irreducible connected subgroup X exists only for p 6= 2 then we write X (p 6= 2)
in the heading and it is assumed that any subgroup Y of X inherits this restriction on p without explicit
labelling. However, when we consider the subgroups of Y we do explicitly repeat any restriction inherited
from X in the heading, for clarity.
In Tables 13 and 14 there is one more possibility for a generic row. To make the tables easier to read we
have moved large collections of diagonal subgroups to supplementary tables in Section 5.3. Thus the row
“See Table x” means that all diagonal subgroups of the relevant subgroup Y appear in Table x in Section
5.3.
We now explain how to find the lattice of connected overgroups in L0 of each L0 -irreducible connected
subgroup given in Tables 2–14. Each table starts with the L0 -irreducible subgroups that are maximal
10
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
amongst connected subgroups of L0 , with their identification number listed in the ID column. We then
write “In M = H1 H2 . . . (L0 (#n1 ))” where n1 is the identification number for the first reductive, maximal
connected subgroup of type H1 H2 . . . , and will include any restrictions on the characteristic p. We then list
the L0 -irreducible maximal connected subgroups of M as well as all diagonal connected subgroups of M , if
there are any, not just the maximal ones. Recall that we will not explicitly consider the proper subgroups
of any diagonal connected subgroups, as discussed in Section 4. A horizontal line then indicates the end
of this list. The next row will be a heading “In M1 = X1 X2 . . . (L0 (#n2 ))”, where X1 X2 . . . is the first
L0 -irreducible maximal connected subgroup of M . The “M1 =” tells the reader that we are now listing
the subgroups of a maximal subgroup of a maximal connected subgroup. We then repeat the process,
listing the L0 -irreducible maximal connected subgroups of M1 and all diagonal subgroups of M1 . The
next heading could be “In M2 = Y1 Y2 . . . (L0 (#n2 ))”, where Y1 Y2 . . . is a maximal connected subgroup
of M1 or it could be “In M1 = Z1 Z2 . . . (L0 (#n3 ))”, where Z1 Z2 . . . is the second maximal connected
subgroup of M . This will depend on whether M1 has any proper L0 -irreducible connected subgroups that
need considering or not. Once all L0 -irreducible connected subgroups of M have been listed in this way,
a horizontal lines indicates the end of the subgroups contained in the first reductive, maximal connected
subgroup of G. The next heading will be “In M = K1 K2 . . . (L0 (#n4 ))” and we repeat the process again
for the second reductive, maximal connected subgroup K1 K2 . . . of L0 . We iterate this process until we
have considered all of the L0 -irreducible subgroups contained in the final reductive, maximal connected
subgroup.
There is an important deviation from the process described above. Suppose that X is a representative
for a conjugacy class of L0 -irreducible connected subgroups with more than one conjugacy class of proper
immediate overgroups. Then after one occurrence of X in the table we list all repetitions of X with the
ID number in italics and do not reconsider its subgroups at those points in the table. Moreover, we only
list VL0 ↓ X in the third column once.
There is another important example where we do not immediately consider the subgroups contained
in X, listed as a subgroup of Y1 , say, even if the ID number for X is not in italics nor is X a diagonal
subgroup of Y1 . In this case the ID number will be nb and the subgroup X is defined for all p ≥ k but
only a maximal connected subgroup of Y1 for some prime l ≥ k. At some point later in the table X will
be defined for all p ≥ k except l and given ID number na. At this point we will consider the subgroups
of X for all p ≥ k together. There are instances where the subgroup X occurs again in the table. If X is
listed for p = l, we write nb in the ID column. Similarly, if X is listed for p ≥ k except l, we write na. It
may be that X is listed for all p ≥ k in a later reductive, maximal connected subgroup and in this case we
simply write n in the ID column.
5.2. Tables of embeddings for Theorem 1.
Table 2: Irreducible subgroups of L0 = A2 .
ID Subgroup X VL0 ↓ X
1 A1 (p 6= 2)
2
Table 4: Irreducible subgroups of L0 = A4 .
ID Subgroup X VL0 ↓ X
1 B2 (p 6= 2)
10
In M = B2 (p 6= 2) (A4 (#1))
2 A1 (p ≥ 5)
4
Table 3: Irreducible subgroups of L0 = A3 .
ID Subgroup X VL0 ↓ X
1 B2
01
2
2a A1 (p 6= 2)
(1, 1)
In M = B2 (A3 (#1))
2b A21 (p = 2)
(1, 1)
3 A1 (p ≥ 5)
3
2
In M = A1 (A3 (#2))
4 A1 via (1, 1[r] ) (r 6= 0)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
Table 5: Irreducible subgroups of L0 = A5 .
ID Subgroup X VL0 ↓ X
1 C3
100
2 A1 A2
(1, 10)
3a A3 (p 6= 2)
010
4 A2 (p 6= 2)
20
In M = C3 (A5 (#1))
5 A1 A1 (p 6= 2) (1, 2)
3b Ã3 (p = 2)
010
6 G2 (p = 2)
10
7 A1 (p ≥ 7)
5
In M1 = A1 A1 (p 6= 2) (A5 (#5))
8 A1 via (1[r] , 1[s] ) (rs = 0)
In M = A1 A2 (A5 (#2))
5 A1 A1 (p 6= 2)
Table 8: Irreducible subgroups of L0 = C2 .
ID Subgroup X VL0 ↓ X
1 Ā21
(1, 0) + (0, 1)
2
2 Ã1 (p = 2)
(1, 1)
3a A1 (p ≥ 5)
3
In M1 = Ā21 (C2 (#1))
4 A1 via (1, 1[r] ) (r 6= 0)
In M1 = Ã21 (p = 2) (C2 (#2))
3b A1 via (1, 1[1] )
5 A1 via (1, 1[r] ) (r > 1)
Table 6: Irreducible subgroups of L0 = A6 .
ID Subgroup X VL0 ↓ X
1 B3 (p 6= 2)
100
In M = B3 (p 6= 2) (A6 (#1))
2 G2
10
In M1 = G2 (p 6= 2) (A6 (#2))
3 A1 (p ≥ 7)
6
4 A2 (p = 3)
11
Table 7: Irreducible subgroups of L0 = A7 .
ID Subgroup X VL0 ↓ X
1 C4
λ1
2a D4 (p 6= 2)
λ1
3 A1 A3
(1, 100)
In M = C4 (A7 (#1))
4a A31 (p 6= 2)
(1, 1, 1)
5 A1 (p ≥ 11) 7
2b D4 (p = 2)
λ1
3
In M = A1 (p 6= 2) (A7 (#4))
[r]
6 A1 A1 via (1a , 1a , 1b ) (r 6= 0)
7 A1 via (1, 1[r] , 1[s] ) (0 < r < s)
In M = D4 (A7 (#2))
8 B3
001
9a A1 B2 (p 6= 2) (1, 01)
10 A2 p 6= 3
11
In M = B3 (A7 (#8))
9b A1 B2 (p = 2) (1, 01)
In M = A1 B2 (A7 (#9))
4b A31 (p = 2)
(1, 1, 1)
In M = A1 A3 (A7 (#3))
9 A1 B2
11
12
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Table 9: Irreducible subgroups of L0 = B3 .
ID
1a
2
3
4
In M1
5
6
7
In M1
5 {0 }
In M1
6 {0 }
8
9
In M2
7 {δ2 }
In M1
1b
10
In M1
11
12
Subgroup X VL0 ↓ X
Ā21 Ã1 (p 6= 2) (1, 1, 0) + (0, 0, 2)
Ā3
010 + 000 (p 6= 2) or 010 (p = 2)
G2
10
Ã1 B2 (p = 2) (2, 00) + (0, 10)
= Ā21 Ã1 (B3 (#1))
[r]
A1 Ã1 via (1a , 1a , 1b ) (if p = 2 then r 6= 0)
[r]
[s]
Ā1 A1 via (1a , 1b , 1a ) (rs = 0)
A1 via (1[r] , 1[s] , 1[t] ) (rt = 0; r ≤ s;
if r = s then p 6= 2 and r < t)
= Ā3 (B3 (#2))
Ã21 (p 6= 2)
= G2 (B3 (#3))
Ā1 A1
A2 (p = 3)
11
A1 (p ≥ 7)
6
= A2 (p = 3) (B3 (#8))
A1
= Ã1 B2 (p = 2) (B3 (#4))
(1, 1, 0) + (0, 0, 2)
Ã1 Ā21
3
Ã1
(2, 0, 0) + (0, 2, 0) + (0, 0, 2)
= Ã31 (p = 2) (B3 (#10))
[r]
Ã1 A1 via (1a , 1a , 1b ) (r 6= 0)
A1 via (1, 1[r] , 1[s] ) (0 < r < s)
Table 10: Irreducible subgroups of L0 = C3 .
ID
1
2
3
4
5
In M1
6
7
8
In M2
9
10
11
In M2
12
In M2
13
14
In M1
15
12 {0 }
In M1
9 {δ1 }
Subgroup X VL0 ↓ X
Ā1 C2
(1, 00) + (0, 10)
A1 A1 (p 6= 2) (2, 1)
Ã3 (p = 2)
010
G2 (p = 2)
10
A1 (p ≥ 7)
5
= Ā1 C2 (C3 (#1))
Ā1 Ã21 (p = 2) (1, 0, 0) + (0, 1, 1)
Ā1 A1 (p ≥ 5) (1, 0) + (0, 3)
Ā31
(1, 0, 0) + (0, 1, 0) + (0, 0, 1)
= Ā1 Ã21 (p = 2) (C3 (#6))
[r] [s]
A1 Ã1 via (1a , 1a , 1b ) (rs = 0)
[r]
Ā1 A1 via (1a , 1b , 1b ) (r 6= 0)
A1 via (1[r] , 1[s] , 1[t] ) (rs = 0; s < t)
= Ā1 A1 (p ≥ 5) (C3 (#7))
A1 via (1[r] , 1[s] ) (rs = 0)
= Ā31 (C3 (#8))
[r]
Ā1 A1 via (1a , 1a , 1b ) (r 6= 0)
A1 via (1, 1[r] , 1[s] ) (0 < r < s)
= A1 A1 (p 6= 2) (C3 (#2))
A1 via (1[r] , 1[s] ) (rs = 0; r 6= s)
A1 via (1, 1) (p ≥ 5)
= G2 (p = 2) (C3 (#4))
Ã1 A1
Table 11: Irreducible subgroups of L0 = D4 .
ID
Subgroup X VL0 ↓ X
1
Ā41
(1, 1, 0, 0) + (0, 0, 1, 1)
2
B3
T (λ1 ; 0)
3
B3
001
4
B3
001
5a
A1 B2 (p 6= 2) (2, 00) + (0, 10)
6a
A1 B2 (p 6= 2) (1, 01)
7a
A1 B2 (p 6= 2) (1, 01)
8a
A2 p 6= 3
11
In M = Ā41 (D4 (#1))
[r]
9
A1 Ā21 via (1a , 1a , 1b , 1c )
[r]
10
A1 Ā21 via (1a , 1b , 1a , 1c )
[r]
11
A1 Ā21 via (1a , 1b , 1c , 1a )
[r] [s] [t]
12
A1 Ā1 via (1a , 1a , 1a , 1b ) (rst = 0)
[r] [s] [t]
13
A1 A1 via (1a , 1a , 1b , 1b ) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
[s]
[r]
[t]
14
A1 A1 via (1a , 1b , 1a , 1b ) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
[s] [t] [r]
15
A1 A1 via (1a , 1b , 1b , 1a ) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
16
A1 via (1, 1[r] , 1[s] , 1[t] ) (if r = 0 then s < t; if s = 0 then r < t; if t = 0 then r < s)
In M = B3 (D4 (#2))
9 {0 }
Ā21 A1 (p 6= 2)
17
G2
T (10; 0)
5b
A1 B2 (p = 2) Θ(2; 10)
In M1 = G2 (D4 (#17))
12 {0 } Ā1 A1
8b
A2 (p = 3)
11 + 00
18
A1
6+0
In M = B3 (D4 (#3))
11 {0 } Ā21 A1 (p 6= 2)
17
G2
6b
A1 B2 (p = 2) (1, 01)
In M = B3 (D4 (#4))
10 {0 } Ā21 A1 (p 6= 2)
17
G2
7b
A1 B2 (p = 2) (1, 01)
In M = A1 B2 (D4 (#5))
9 {0 }
A1 Ā21
19
A1 A1 (p ≥ 5) (2, 0) + (0, 4)
20
A31 (p = 2)
Θ(2; 2; 2)
In M1 = A1 A1 (p ≥ 5) (D4 (#19))
21
A1 via (1[r] , 1[s] ) (rs = 0)
In M = A31 (p = 2) (D4 (#20))
[r]
22
A1 A1 via (1a , 1a , 1b ) (r 6= 0)
23
A1 via (1, 1[r] , 1[s] ) (0 < r < s)
In M = A1 B2 (D4 (#6))
11 {0 } A1 Ā21
24
A1 A1 (p ≥ 5) (1, 3)
25
A31 (p = 2)
(1, 1, 1)
In M1 = A1 A1 (p ≥ 5) (D4 (#24))
26
A1 via (1[r] , 1[s] ) (rs = 0)
In M1 = A31 (p = 2) (D4 (#25))
[r]
27
A1 A1 via (1a , 1a , 1b ) (r 6= 0)
28
A1 via (1, 1[r] , 1[s] ) (0 < r < s)
In M = A1 B2 (D4 (#7))
10 {0 } A1 Ā21
29
A1 A1 (p ≥ 5) (1, 3)
30
A31 (p = 2)
(1, 1, 1)
In M1 = A1 A1 (p ≥ 5) (D4 (#29))
31
A1 via (1[r] , 1[s] ) (rs = 0)
In M1 = A31 (p = 2) (D4 (#30))
13
14
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
[r]
32
A1 A1 via (1a , 1a , 1b ) (r 6= 0)
33
A1 via (1, 1[r] , 1[s] ) (0 < r < s)
In M = A2 (D4 (#8))
21 {0 } A1 (p ≥ 5)
16 {δ3 } A1 (p = 3)
Table 12: Irreducible subgroups of L0 = D5 .
ID
Subgroup X VL0 ↓ X
1
Ā21 A3
(1, 1, 000) + (0, 0, 010)
2a
B22 (p 6= 2)
(10, 00) + (00, 10)
3a
A1 B3 (p 6= 2) (2, 000) + (0, 100)
4
B4
T (λ1 ; 0)
5
B2 (p 6= 2)
02
6
B2 (p 6= 2)
02
2
In M = Ā1 A3 (D5 (#1))
7
Ā21 B2
(1, 1, 00) + (0, 0, T (10; 0))
8a
Ā21 A21 (p 6= 2) (1, 1, 0, 0) + (0, 0, 2, 0) + (0, 0, 0, 2)
9
A1 A3 via (1, 1[r], 100)
In M1 = Ā21 B2 (D5 (#2))
10
Ā21 A1 (p ≥ 5) (1, 1, 0) + (0, 0, 4) + 0
8b
Ā21 A21 (p = 2) (1, 1, 0, 0) + (0, 0, Θ(2; 2))
[r]
11
A1 B2 via (1a , 1a , 10) (r 6= 0)
In M2 = Ā21 A1 (p ≥ 5) (D5 (#10))
[r]
12
A1 A1 via (1a , 1a , 1b ) (r 6= 0)
[s]
[r]
13
Ā1 A1 via (1a , 1b , 1a ) (rs = 0)
14
A1 via (1[r] , 1[s] , 1[t] ) (rt = 0; r < s)
In M2 = Ā21 A21 (D5 (#8))
[r]
15
A1 A21 via (1a , 1a , 1b , 1c ) (if p = 2 then r 6= 0)
[r]
[s]
16
A1 Ā1 A1 via (1a , 1b , 1a , 1c ) (rs = 0)
[r]
17
Ā21 A1 via (1a , 1b , 1c , 1c ) (r 6= 0)
[r] [s] [t]
18
A1 A1 via (1a , 1a , 1a , 1b ) (rt = 0; r ≤ s; if r = s then p 6= 2 and r < t)
[r]
[s] [t]
19
A1 Ā1 via (1a , 1b , 1a , 1a ) (rs = 0; s < t)
[r]
[s]
20
A1 A1 via (1a , 1a , 1b , 1b ) (if p = 2 then r 6= 0; s 6= 0)
[r] [t] [s] [u]
21
A1 A1 via (1a , 1b , 1a , 1b ) (rs = tu = 0; r ≤ s; if r = s then t ≤ u)
22
A1 via (1[r] , 1[s] , 1[t] , 1[u] ) (rt = 0; r ≤ s; t < u; if r = s then p 6= 2 and r < t)
In M = B22 (D5 (#2))
7
Ā21 B2
23
A1 B2 (p ≥ 5) (4, 00) + (0, 10)
24
A21 B2 (p = 2) Θ(2; 2; 10)
25
B2 via (10, 10[r]) (r 6= 0)
26
B2 via (10, 02[r]) (p = 2)
In M1 = A1 B2 (p ≥ 5) (D5 (#23))
10
A1 Ā21
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
15
27
A21
(4, 0) + (0, 4)
2
In M2 = A1 (p ≥ 5) (D5 (#27))
28
A1 via (1, 1[r] ) (r 6= 0)
In M1 = A21 B2 (p = 2) (D5 (#24))
8
A21 Ā21
29
A41
Θ(2; 2; 2; 2)
[r]
30
A1 B2 via (1a , 1a , 10) (r 6= 0)
In M2 = A41 (p = 2) (D5 (#29))
[r]
31
A1 A21 via (1a , 1a , 1b , 1c ) (r 6= 0)
[r] [s]
32
A1 A1 via (1a , 1a , 1a , 1b ) (0 < r < s)
[r]
[s]
33
A1 A1 via (1a , 1a , 1b , 1b ) (rs 6= 0; r ≤ s)
34
A1 via (1, 1[r] , 1[s] , 1[t] ) (0 < r < s < t)
In M = A1 B3 (D5 (#3))
9 {0 }
A1 Ā3
8
A21 Ā21 (p 6= 2)
35
A1 G2
(2, 00) + (0, 10) (p 6= 2) or Θ(2; 10) (p = 2)
24
A21 B2 (p = 2)
In M1 = A1 G2 (D5 (#35))
16 {0 } A1 Ā1 A1
36
A1 A2 (p = 3) (2, 00) + (0, 11)
37
A1 A1 (p ≥ 7) (2, 0) + (0, 6)
In M2 = A1 A2 (p = 3) (D5 (#36))
18 {δ2 } A1 A1
In M2 = A1 A1 (p ≥ 7) (D5 (#37))
38
A1 via (1[r] , 1[s] ) (rs = 0)
In M = B4 (D5 (#4))
9 {0 }
A1 Ā3 (p 6= 2)
7
Ā21 B2 (p 6= 2)
39
A21 (p 6= 2)
(2, 2) + 0
40
A1 (p ≥ 11) 8 + 0
3b
B22 (p = 2)
Θ(10; 10)
4b
A1 B3 (p = 2) Θ(2; 100)
In M1 = A21 (p 6= 2) (D5 (#39))
41
A1 via (1, 1[r] ) (r 6= 0)
In M = B2 (p 6= 2) (D5 (#5))
21 {0 } A21
38 {0 } A1 (p ≥ 7)
In M = B2 (p 6= 2) (D5 (#6))
21 {0 } A21
38 {0 } A1 (p ≥ 7)
Subgroups of D6 which are not centralised by the standard graph automorphism can be distinguished by their
actions on the two half-spin modules VD6 (λ5 ) and VD6 (λ6 ). For consistency, in the following table we fix how the
subgroups A1 C3 = D6 (#6), A1 C3 = D6 (#7) and Ā21 A1 B2 = D6 (#12) act on these modules.
16
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
We let A1 C3 = D6 (#6) act on VD6 (λ5 ) with factors (W (3), 000)/(1, W (010)) and therefore act on VD6 (λ6 ) with
factors (W (2), 100)/(0, W (001)). Taking an image under a graph automorphism, it follows that A1 C3 = D6 (#7)
acts on VD6 (λ5 ) with factors (W (2), 100)/(0, W (001)) and on VD6 (λ6 ) with factors (W (3), 000)/(1, W (010)).
We let Ā21 A1 B2 = D6 (#12) act on VD6 (λ5 ) with factors (1, 0, W (2), 00)/(1, 0, 0, W (10))/(0, 1, 1, 01) and on VD6 (λ6 )
with factors (1, 0, 1, 01)/(0, 1, W (2), 00)/(0, 1, 0, W (10)). This also determines the action of the subgroups Ā21 A1 A1 =
D6 (#73) (p ≥ 5) and Ā21 A1 A1 = D6 (#74) (p ≥ 5).
Table 13: Irreducible subgroups of L0 = D6 .
ID
Subgroup X
VL0 ↓ X
2
1
Ā1 D4
(1, 1, 0) + (0, 0, λ1 )
2
Ā23
(010, 000) + (000, 010)
3a
A1 B4 (p 6= 2)
(2, 0) + (0, λ1 )
4a
B2 B3 (p 6= 2)
(10, 000) + (00, 100)
5
B5
T (λ1 ; 0)
6
A1 C3
(1, 100)
7
A1 C3
(1, 100)
2
In M = Ā1 D4 (D6 (#1))
8
Ā61
(1, 1, 0, 0, 0, 0) + (0, 0, 1, 1, 0, 0) + (0, 0, 0, 0, 1, 1)
2
9
Ā1 B3
(1, 1, 000) + (0, 0, T (100; 0))
2
10
Ā1 B3
(1, 1, 000) + (0, 0, 001)
11a
Ā21 A1 B2 (p 6= 2) (1, 1, 0, 00) + (0, 0, 2, 00) + (0, 0, 0, 10)
12a
Ā21 A1 B2 (p 6= 2) (1, 1, 0, 00) + (0, 0, 1, 01)
13a
Ā21 A2 (p 6= 3)
(1, 1, 00) + (0, 0, 11)
[r]
14
A1 D4 via (1, 1 , λ1 )
In M1 = Ā61 (D6 (#8))
See Table 15
In M1 = Ā21 B3 (D6 (#9))
15 {0 }
Ā41 A1 (p 6= 2)
39
Ā21 G2
(1, 1, 00) + (0, 0, T (10; 0))
2
11b
Ā1 A1 B2 (p = 2) (1, 1, 0, 00) + (0, 0, Θ(2; 10))
40
A1 B3 via (1, 1[r], 100) (r 6= 0)
In M2 = Ā21 G2 (D6 (#39))
17 {0 }
Ā31 A1
13b
Ā21 A2 (p = 3)
(1, 1, 00) + (0, 0, 11) + 0
[r]
41
A1 G2 via (1, 1 , 10) (r 6= 0)
In M1 = Ā21 B3 (D6 (#10))
Ā41 A1 (p 6= 2)
16 {0 }
39
Ā21 G2
12b
Ā21 A1 B2 (p = 2) (1, 1, 0, 00) + (0, 0, 1, 01)
42
A1 B3 via (1[r] , 1[s] , 100) (rs = 0)
In M1 = Ā21 A1 B2 (D6 (#11))
Ā41 A1
15 {0 }
43
Ā21 A1 A1 (p ≥ 5) (1, 1, 0, 0) + (0, 0, 2, 0) + (0, 0, 0, 4)
44
Ā21 A31 (p = 2)
(1, 1, 0, 0, 0) + (0, 0, Θ(2; 2; 2))
[r]
45
A1 A1 B2 via (1a , 1a , 1b , 10) (if p = 2 then r 6= 0)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
[r]
[s]
46
A1 Ā1 B2 via (1a , 1b , 1a , 10) (rs = 0)
47
A1 B2 via (1[r] , 1[s] , 1[t] , 10) (rt = 0; r ≤ s; if p = 2 then r < s)
In M2 = Ā21 A1 A1 (p ≥ 5) (D6 (#43))
See Table 17
In M2 = Ā21 A31 (p = 2) (D6 (#44))
See Table 18
In M1 = Ā21 A1 B2 (D6 (#12))
Ā41 A1
16 {0 }
73
Ā21 A1 A1 (p ≥ 5) (1, 1, 0, 0) + (0, 0, 1, 3)
74
Ā21 A31 (p = 2)
(1, 1, 0, 0, 0) + (0, 0, 1, 1, 1)
[r] [s]
75
A1 A1 B2 via (1a , 1a , 1b , 10) (rs = 0)
[s]
[r]
76
A1 Ā1 B2 via (1a , 1b , 1a , 10) (rs = 0)
[r] [s]
77
A1 Ā1 B2 via (1a , 1b , 1b , 10) (rs = 0)
78
A1 B2 via (1[r] , 1[s] , 1[t] , 10) (rst = 0)
In M2 = Ā21 A1 A1 (p ≥ 5) (D6 (#73))
See Table 19
In M2 = Ā21 A31 (p = 2) (D6 (#74))
See Table 20
In M1 = Ā21 A2 (D6 (#13))
Ā21 A1 (p ≥ 5)
51 {0 }
24 {δ3 } Ā21 A1 (p = 3)
113
A1 A2 via (1, 1[r] , 10) (if p = 3 then r 6= 0)
In M = Ā23 (D6 (#2))
(T (10; 0), 0) + (0, 010)
114
B2 Ā3
115a
A21 Ā3 (p 6= 2)
(2, 0, 000) + (0, 2, 000) + (0, 0, 010)
116
A3 via (100, 100[r]) (r 6= 0)
117
A3 via (100, 001[r]) (r 6= 0)
In M1 = B2 Ā3 (D6 (#114))
(4, 000) + (0, 010) + 0
118
A1 Ā3 (p ≥ 5)
115b
A21 Ā3 (p = 2)
(Θ(2; 2), 0) + (0, 0, 010)
{0 }
2
45
B2 A1 (p 6= 2)
In M2 = A1 Ā3 (p ≥ 5) (D6 (#118))
48 {0 }
A1 A21
In M1 = A21 Ā3 (D6 (#115))
45 {0 }
A21 B2 (p 6= 2)
119
A41 (p 6= 2)
(2, 0, 0, 0) + (0, 2, 0, 0) + (0, 0, 2, 0) + (0, 0, 0, 2)
120
A1 Ā3 via (1, 1[r] , 100) (r 6= 0)
In M2 = A41 (p 6= 2) (D6 (#119))
[r]
121
A1 A21 via (1a , 1a , 1b , 1c ) (r 6= 0)
[r] [s]
122
A1 A1 via (1a , 1a , 1a , 1b ) (0 < r < s)
[r]
[s]
123
A1 A1 via (1a , 1a , 1b , 1b ) (rs 6= 0)
124
A1 via (1, 1[r], 1[s] , 1[t] ) (0 < r < s < t)
In M = A1 B4 (D6 (#3))
14 {0 }
A1 D4
17
18
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
A21 Ā3 (p 6= 2)
A1 Ā21 B2 (p 6= 2)
A1 A21 (p 6= 2)
(2, 0, 0) + (0, 2, 2)
A1 A1 (p ≥ 11) (2, 0) + (0, 8)
A21 B3 (p = 2)
Θ(2; 2; 100)
2
A1 B2 (p = 2)
Θ(2; 10; 10)
2
= A1 A1 (p 6= 2) (D6 (#125))
[r] [s]
A1 A1 via (1a , 1a , 1b ) (rs = 0)
[r] [s]
A1 A1 via (1a , 1b , 1b ) (rs = 0; r 6= s)
A1 A1 (p ≥ 5) via (1a , 1b , 1b )
A1 via (1[r] , 1[s] , 1[t] ) (rs = 0; s < t)
= A1 A1 (p ≥ 11) (D6 (#126))
A1 via (1[r] , 1[s] ) (rs = 0)
= A21 B3 (p = 2) (D6 (#127))
A21 Ā3
A21 G2
Θ(2; 2; 10)
3
A1 B2
Θ(2; 2; 2; 10)
A1 B3 via (1, 1[r], 100) (r 6= 0)
= A21 G2 (p = 2) (D6 (#133))
A21 Ā1 A1
A1 G2 via (1, 1[r] , 100) (r 6= 0)
= A31 B2 (p = 2) (D6 (#134))
A31 Ā21
A51
Θ(2; 2; 2; 2; 2)
[r]
A1 A1 B2 via (1a , 1a , 1b , 100) (r 6= 0)
A1 B2 via (1, 1[r], 1[s] , 100) (0 < r < s)
= A51 (p = 2) (D6 (#137))
See Table 21
In M1 = A1 B22 (p = 2) (D6 (#128))
11
A1 Ā21 B2
134
A31 B2
146
A1 B2 via (1, 10, 10[r]) (r 6= 0)
147
A1 B2 via (1, 10, 02[r])
In M = B2 B3 (D6 (#4))
9b
Ā21 B3
148
A1 B3 (p ≥ 5)
(4, 000) + (0, 100)
127
A21 B3 (p = 2)
114
B2 Ā3
11
B2 Ā21 A1 (p 6= 2)
149
B2 G2
(10, 00) + (00, 10) (p 6= 2) or Θ(10; 10) (p = 2)
2
128
A1 B2 (p = 2)
In M1 = A1 B3 (p ≥ 5) (D6 (#148))
118
A1 Ā3
43
A1 Ā21 A1
115
11
125
126
127
128
In M1
129
130
55 {0 }
131
In M1
132
In M1
115
133
134
135
In M2
59 {0 }
136
In M2
44
137
138
139
In M3
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
150
A1 G2
(4, 00) + (0, 10)
In M2 = A1 G2 (p ≥ 5) (D6 (#150))
A1 Ā1 A1
49 {0 }
151
A1 A1 (p ≥ 7)
(4, 0) + (0, 6)
In M3 = A1 A1 (p ≥ 7) (D6 (#151))
152
A1 via (1[r] , 1[s] ) (rs = 0)
In M1 = B2 G2 (p = 2) (D6 (#149))
39
Ā21 G2
133
A21 G2 (p = 2)
46 {0 }
B2 Ā1 A1
In M = B5 (D6 (#5))
14 {0 }
A1 D4 (p 6= 2)
127
Ā21 B3 (p 6= 2)
114
B2 Ā3 (p 6= 2)
153
A1 (p ≥ 11)
10 + 0
3b
A1 B4 (p = 2)
Θ(2; λ1 )
4b
B2 B3 (p = 2)
Θ(10, 100)
In M = A1 C3 (D6 (#6))
A1 Ā1 B2
76 {0 }
154
A1 A1 A1 (p 6= 2) (1, 2, 1)
155
A1 A3 (p = 2)
(1, 010)
156
A1 G2 (p = 2)
(1, 10)
157
A1 A1 (p ≥ 7)
(1, 5)
In M1 = A1 A1 A1 (p 6= 2) (D6 (#154))
[r] [s]
158
A1 A1 via (1a , 1a , 1b ) (rs = 0; r 6= s)
91 {0 }
A1 A1 via (1a , 1a , 1b )
[r]
[s]
159
A1 A1 via (1a , 1b , 1a ) (rs = 0; r 6= s)
129 {0 } A1 A1 via (1a , 1b , 1a )
[r] [s]
160
A1 A1 via (1a , 1b , 1b ) (rs = 0; r 6= s)
90 {0 }
A1 A1 via (1a , 1b , 1b )
161
A1 via (1[r] , 1[s] , 1[t] ) (rst = 0; r, s, t distinct)
In M1 = A1 G2 (p = 2) (D6 (#156))
102 {δ3 } A1 A1 A1
In M1 = A1 A1 (p ≥ 7) (D6 (#157))
162
A1 via (1[r] , 1[s] ) (rs = 0; r 6= s)
152 {0 } A1 via (1, 1)
In M = A1 C3 (D6 (#7))
77 {0 }
A1 Ā1 B2
154
A1 A1 A1 (p 6= 2)
163
A1 A3 (p = 2)
(1, 010)
164
A1 G2 (p = 2)
(1, 10)
165
A1 A1 (p ≥ 7)
(1, 5)
In M1 = A1 G2 (p = 2) (D6 (#164))
102 {δ1 } A1 A1 A1
In M1 = A1 A1 (p ≥ 7) (D6 (#165))
19
20
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
166
152 {0 }
A1 via (1[r] , 1[s] ) (rs = 0; r 6= s)
A1 via (1, 1)
We remind the reader that in the following table, subgroups are listed up to conjugacy in Aut(L0 ).
Table 14: Irreducible subgroups of L0 = D7 .
ID
Subgroup X
VL0 ↓ X
2
1
Ā1 D5
(1, 1, 0) + (0, 0, λ1 )
2
Ā3 D4
(010, 0) + (000, λ1 )
3
B6
T (λ1 ; 0)
4a
B32 (p 6= 2)
(100, 000) + (000, 100)
5a
B2 B4 (p 6= 2)
(10, 0) + (00, λ1 )
6a
A1 B5 (p 6= 2)
(2, 0) + (0, λ1 )
7a
C3 (p 6= 3)
010
8a
B2 (p 6= 2, 5)
20
9a
G2 (p ≥ 5)
01
In M = Ā21 D5 (D7 (#1))
10
Ā41 Ā3
(1, 1, 0, 0, 000) + (0, 0, 1, 1, 000) + (0, 0, 0, 0, 010)
2
11
Ā1 B4
(1, 1, 0) + (0, 0, T (λ1 ; 0))
12a
Ā21 A1 B3 (p 6= 2) (1, 1, 0, 000) + (0, 0, 2, 000) + (0, 0, 0, 100)
13a
Ā21 B22 (p 6= 2)
(1, 1, 00, 00) + (0, 0, 10, 00) + (0, 0, 00, 10)
14
Ā21 B2 (p 6= 2)
(1, 1, 00) + (0, 0, 02)
[r]
15
A1 D5 via (1, 1 , λ1 )
In M1 = Ā41 Ā3 (D7 (#10))
16
Ā41 B2
(1, 1, 0, 0, 00) + (0, 0, 1, 1, 00) + (0, 0, 0, 0, T (10; 0))
4 2
17a
Ā1 A1 (p 6= 2)
(1, 1, 0, 0, 0, 0) + (0, 0, 1, 1, 0, 0) + (0, 0, 0, 0, 2, 0) + (0, 0, 0, 0, 0, 2)
See Table 22
In M2 = Ā41 B2 (D7 (#16))
24
Ā41 A1 (p ≥ 5)
(1, 1, 0, 0, 0) + (0, 0, 1, 1, 0) + (0, 0, 0, 0, 4) + 0
17b
Ā41 A21 (p = 2)
(1, 1, 0, 0, 0, 0) + (0, 0, 1, 1, 0, 0) + (0, 0, 0, 0, Θ(2; 2))
See Table 23
In M3 = Ā41 A1 (p ≥ 5) (D7 (#24))
See Table 24
In M2 = Ā41 A21 (D7 (#17))
See Table 25
In M1 = Ā21 B4 (D7 (#11))
Ā21 A1 Ā3 (p 6= 2)
18 {0 }
16
Ā41 B2 (p 6= 2)
85
Ā21 A21 (p 6= 2)
(1, 1, 0, 0) + (0, 0, 2, 2) + 0
86
Ā21 A1 (p ≥ 11)
(1, 1, 0) + (0, 0, 8) + 0
12b
Ā21 A1 B3 (p = 2) (1, 1, 0, 000) + (0, 0, Θ(2; 100))
13b
Ā21 B22 (p = 2)
(1, 1, 00, 00) + (0, 0, Θ(10; 10))
87
A1 B4 via (1, 1[r] , λ1 ) (r 6= 0)
In M2 = Ā21 A21 (p 6= 2) (D7 (#85))
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
See Table 26
In M2 = Ā21 A1 (p ≥ 11) (D7 (#86))
[r]
96
A1 A1 via (1a , 1a , 1b ) (r 6= 0)
[s]
[r]
97
A1 A1 via (1a , 1b , 1a ) (rs = 0)
98
A1 via (1[r] , 1[s] , 1[t] ) (rt = 0; r < s)
In M1 = Ā21 A1 B3 (D7 (#12))
Ā21 A1 Ā3
18 {0 }
17
Ā41 A21 (p 6= 2)
99
Ā21 A1 G2
(1, 1, 0, 00) + (0, 0, 2, 00) + (0, 0, 0, 10) (p 6= 2) or (1, 1, 0, 00) + (0, 0, Θ(2; 10)) (p = 2)
2 2
100
Ā1 A1 B2 (p = 2) (1, 1, 0, 0, 00) + (0, 2, Θ(2; 2; 10))
[r]
101
A1 A1 B3 via (1a , 1a , 1b , 100) (if p = 2 then r 6= 0)
[r]
[s]
102
A1 Ā1 B3 via (1a , 1b , 1a , 100) (rs = 0)
103
A1 B3 via (1[r] , 1[s] , 1[t] , 100) (rt = 0; r ≤ s; if r = s then r < t and p 6= 2)
In M2 = Ā21 A1 G2 (D7 (#99))
49 {0 }
Ā21 A1 Ā1 A1
104
Ā21 A1 A2 (p = 3) (1, 1, 0, 00) + (0, 0, 2, 00) + (0, 0, 0, 11)
105
Ā21 A1 A1 (p ≥ 7) (1, 1, 0, 0) + (0, 0, 2, 0) + (0, 0, 0, 6)
[r]
106
A1 A1 G2 via (1a , 1a , 1b , 10) (if p = 2 then r 6= 0)
[s]
[r]
107
A1 Ā1 G2 via (1a , 1b , 1a , 10) (rs = 0)
108
A1 G2 via (1[r] , 1[s] , 1[t] , 10) (rt = 0; r ≤ s; if r = s then r < t and p 6= 2)
In M3 = Ā21 A1 A2 (p = 3) (D7 (#104))
52 {δ2 } Ā21 A1 A1
[r]
109
A1 A1 A2 via (1a , 1a , 1b , 10)
[r]
[s]
110
A1 Ā1 A2 via (1a , 1b , 1a , 10) (rs = 0)
111
A1 A2 via (1[r] , 1[s] , 1[t] , 10) (rt = 0; r ≤ s; if r = s then r < t)
In M3 = Ā21 A1 A1 (p ≥ 7) (D7 (#105))
See Table 27
In M2 = Ā21 A21 B2 (p = 2) (D7 (#100))
17
Ā21 A21 Ā21
122
Ā21 A41
(1, 1, 0, 0, 0, 0) + (0, 0, Θ(2; 2; 2; 2))
See Table 28
In M3 = Ā21 A41 (p = 2) (D7 (#122))
See Table 29
In M1 = Ā21 B22 (D7 (#13))
16
Ā41 B2
156
Ā21 A1 B2 (p ≥ 5) (1, 1, 0, 00) + (0, 0, 4, 00) + (0, 0, 0, 10)
100
Ā21 A21 B2 (p = 2)
See Table 30
In M2 = Ā21 A1 B2 (p ≥ 5) (D7 (#156))
24
Ā41 A1
162
Ā21 A21 (p ≥ 5)
(1, 1, 0, 0) + (0, 0, 4, 0) + (0, 0, 0, 4)
[r]
163
A1 A1 B2 via (1a , 1a , 1b , 10)
[r]
[s]
164
A1 Ā1 B2 via (1a , 1b , 1a , 10) (rs = 0)
165
A1 B2 via (1[r] , 1[s] , 1[t] , 10) (rt = 0; r ≤ s)
21
22
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
In M3 = Ā21 A21 (p ≥ 5) (D7 (#162))
See Table 31
In M1 = Ā21 B2 (p 6= 2) (D7 (#14))
61 {0 }
Ā21 A21
115 {0 } Ā21 A1 (p ≥ 5)
174
A1 B2 via (1, 1[r] , 10)
In M = Ā3 D4 (D7 (#2))
175
B2 D 4
(T (10; 0), 0) + (0, λ1 )
176a
A21 D4 (p 6= 2)
(2, 0, 0) + (0, 2, 0) + (0, 0, λ1 )
177
Ā3 B3
(010, 000) + (000, T (100; 0))
178
Ā3 B3
(010, 000) + (000, 001)
179a
Ā3 A1 B2 (p 6= 2) (010, 0, 00) + (000, 2, 00) + (000, 0, 10)
180a
Ā3 A1 B2 (p 6= 2) (010, 0, 00) + (000, 1, 01)
10
Ā3 Ā41
181a
Ā3 A2 (p 6= 3)
(010, 00) + (000, 11)
In M1 = B2 D4 (D7 (#175))
182
A1 D4 (p ≥ 5)
(4, 0) + (0, λ1 ) + 0
2
176b
A1 D4 (p = 2)
(Θ(2; 2), 0) + (0, 0, λ1 )
183
B2 B3
(T (10; 0), 000) + (00, 001)
157
B2 A1 B2 (p 6= 2)
184a
B2 A1 B2 (p 6= 2) (10, 0, 00) + (00, 1, 01) + 0
16
B2 Ā41
185
B2 A2 (p 6= 3)
(T (10; 0), 00) + (00, 11)
In M2 = A1 D4 (p ≥ 5) (D7 (#182))
186
A1 B3
(4, 000) + (0, 001) + 0
{0 }
163
A1 A1 B2
187
A1 A1 B2
(4, 0, 00) + (0, 1, 01) + 0
4
24
A1 Ā1
188
A1 A2
(4, 00) + (0, 11) + 0
In M3 = A1 B3 (p ≥ 5) (D7 (#186))
32 {0 }
A1 Ā21 A1
In M3 = A1 A1 B2 (p ≥ 5) (D7 (#187))
189
A1 A1 A1
(4, 0, 0) + (0, 1, 3) + 0
32 {0 }
A1 A1 Ā21
190
A1 B2 via (1[r] , 1[s] , 10) (rs = 0)
In M4 = A1 A1 A1 (p ≥ 5) (D7 (#189))
See Table 32
In M3 = A1 A2 (p ≥ 5) (D7 (#188))
169 {0 } A1 A1
In M2 = B2 B3 (D7 (#183))
186
A1 B3 (p ≥ 5)
218b
A21 B3 (p = 2)
(Θ(2; 2), 0) + (0, 0, 001)
{0 }
2
26
B2 Ā1 A1 (p 6= 2)
184b
B2 A1 B2 (p = 2) (Θ(10), 0, 00) + (00, 1, 01)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
In M2 = B2 A1 B2 (D7 (#184))
187
A1 A1 B2 (p ≥ 5)
220b
A21 A1 B2 (p = 2) (Θ(2), 0, 00) + (0, 0, 1, 01)
{0 }
26
B2 A1 Ā21
195
B2 A1 A1 (p ≥ 5) (10, 0, 0) + (00, 1, 3) + 0
196
B2 A31 (p = 2)
(Θ(10), 0, 0, 0) + (00, 1, 1, 1)
[r]
197
B2 A1 via (10 , 1, 10[s]) (rs = 0)
198
B2 A1 (p = 2) via (10[r] , 1, 02[s] ) (rs = 0)
In M3 = B2 A1 A1 (p ≥ 5) (D7 (#195))
189
A1 A1 A1
199
B2 A1 via (10, 1[r], 1[s] ) (rs = 0; r 6= s)
169 {0 } B2 A1 via (10, 1, 1)
In M3 = B2 A31 (p = 2) (D7 (#196))
200
A21 A31
(Θ(2; 2), 0, 0, 0) + (0, 0, 1, 1, 1)
[r]
201
B2 A1 A1 via (10, 1a , 1a , 1b ) (r 6= 0)
202
B2 A1 via (10, 1, 1[r], 1[s] ) (0 < r < s)
In M4 = A21 A31 (p = 2) (D7 (#200))
See Table 33
In M2 = B2 A2 (p 6= 3) (D7 (#185))
188
A1 A2 (p ≥ 5)
221
A21 A2 (p = 2)
{0 }
165
B2 A1 (p ≥ 5)
In M1 = A21 D4 (D7 (#176))
101 {0 } A21 B3 (p 6= 2)
218
A21 B3
(2, 0, 000) + (0, 2, 000) + (0, 0, 001) (p 6= 2) or (Θ(2; 2), 000) + (0, 0, 001) (p = 2)
3
219
A1 B2 (p 6= 2)
(2, 0, 0, 00) + (0, 2, 0, 00) + (0, 0, 2, 00) + (0, 0, 0, 10)
220a
A21 A1 B2 (p 6= 2) (2, 0, 0, 00) + (0, 2, 0, 00) + (0, 0, 1, 01)
17
A21 Ā41
221
A21 A2 (p 6= 3)
(2, 0, 00) + (0, 2, 00) + (0, 0, 11) (p ≥ 5) or (Θ(2; 2), 00) + (0, 0, 11) (p = 2)
222
A1 D4 via (1, 1[r] , λ1 ) (r 6= 0)
In M2 = A21 B3 (D7 (#218))
48 {0 }
A21 Ā21 A1 (p 6= 2)
106 {0 } A21 G2 (p 6= 2)
220b
A21 A1 B2 (p = 2) (Θ(2; 2), 0, 00) + (0, 0, 1, 01)
223
A1 B3 via (1, 1[r] , 100) (r 6= 0)
In M2 = A31 B2 (p 6= 2) (D7 (#219))
47 {0 }
A31 Ā21
224
A31 A1 (p ≥ 5)
(2, 0, 0, 0) + (0, 2, 0, 0) + (0, 0, 2, 0) + (0, 0, 0, 4)
[r]
225
A1 A1 B2 via (1a , 1a , 1b , 10) (r 6= 0)
226
A1 B2 via (1, 1[r] , 1[s] , 10) (0 < r < s)
In M3 = A31 A1 (p ≥ 5) (D7 (#224))
See Table 34
In M2 = A21 A1 B2 (D7 (#220))
48 {0 }
A21 A1 Ā21
23
24
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
233
A21 A1 A1 (p ≥ 5) (2, 0, 0, 0) + (0, 2, 0, 0) + (0, 0, 1, 3)
200
A21 A31 (p = 2)
[r]
234
A1 A1 B2 via (1a , 1a , 1b , 10) (r 6= 0)
[r]
[s]
235
A1 A1 B2 via (1a , 1b , 1a , 10) (rs = 0)
236
A1 B2 via (1[r] , 1[s] , 1[t] , 10) (rt = 0; r < s)
In M3 = A21 A1 A1 (p ≥ 5) (D7 (#233))
See Table 35
In M2 = A21 A2 (p 6= 3) (D7 (#221))
228 {0 } A21 A1
247
A1 A2 via (1, 1[r], 10) (r 6= 0)
In M1 = Ā3 B3 (D7 (#177))
101 {0 } A21 B3 (p 6= 2)
Ā3 Ā21 A1 (p 6= 2)
18 {0 }
248
Ā3 G2
(010, 00) + (000, 10) + 0
179b
Ā3 A1 B2 (p = 2) (010, 0, 00) + (000, Θ(2; 10))
In M2 = Ā3 G2 (D7 (#248))
106 {0 } A21 G2 (p 6= 2)
20 {0 }
Ā3 Ā1 A1
181b
Ā3 A2
(010, 00) + (000, 11) + 0
In M1 = Ā3 B3 (D7 (#178))
218
A21 B3 (p 6= 2)
183
B2 B3
{0 }
Ā3 Ā21 A1 (p 6= 2)
19
248
Ā3 G2
180b
Ā3 A1 B2 (p = 2) (010, 0, 00) + (000, 1, 01)
In M1 = Ā3 A1 B2 (D7 (#179))
219
A31 B2 (p 6= 2)
157 {0 } B22 A1 (p 6= 2)
Ā3 A1 Ā21
18 {0 }
249
Ā3 A1 A1 (p ≥ 5) (010, 0, 00) + (000, 2, 0) + (000, 0, 4)
250
Ā3 A31 (p = 2)
(010, 0, 0, 0) + (000, Θ(2; 2; 2))
In M2 = Ā3 A1 A1 (p ≥ 5) (D7 (#249))
224
A31 A1
{0 }
163
B2 A1 A1
251
Ā3 A1 via (100, 1[r], 1[s] ) (rs = 0)
In M2 = Ā3 A31 (p = 2) (D7 (#250))
[r]
252
Ā3 A1 A1 via (100, 1a, 1a , 1b ) (r 6= 0)
253
Ā3 A1 via (100, 1, 1[r], 1[s] ) (0 < r < s)
In M1 = Ā3 A1 B2 (D7 (#180))
220
A21 A1 B2 (p 6= 2)
184
B2 A1 B2
Ā3 A1 Ā21
19 {0 }
254
Ā3 A1 A1 (p ≥ 5) (010, 0, 0) + (000, 1, 3)
(010, 0, 0, 0) + (000, 1, 1, 1)
255
Ā3 A31 (p = 2)
In M2 = Ā3 A1 A1 (p ≥ 5) (D7 (#254))
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
233
A21 A1 A1 (p 6= 2)
195
B2 A1 A1
256
Ā3 A1 via (100, 1[r], 1[s] ) (rs = 0; r 6= s)
In M2 = Ā3 A31 (p = 2) (D7 (#255))
196
B2 A31
[r]
257
Ā3 A1 A1 via (100, 1a, 1a , 1b ) (r 6= 0)
258
Ā3 A1 via (100, 1, 1[r], 1[s] ) (0 < r < s)
In M1 = Ā3 A2 (D7 (#181))
221
A21 A2 (p ≥ 5)
109 {0 } A21 A2 (p = 3)
185
B2 A2 (p 6= 3)
{0 }
Ā3 A1
251
In M = B6 (D7 (#3))
15 {0 }
A1 D̄5 (p 6= 2)
11
Ā21 B4 (p 6= 2)
175
B2 D4 (p 6= 2)
177
Ā3 B3 (p 6= 2)
259
A1 (p ≥ 13)
12 + 0
2
4b
B3 (p = 2)
Θ(100; 100)
5b
B2 B4 (p = 2)
Θ(10; λ1 )
6b
A1 B5 (p = 2)
Θ(2; λ1 )
7b
C3 (p = 3)
010 + 0
8b
B2 (p = 5)
20 + 0
In M = B32 (D7 (#4))
177
Ā3 B3
12
Ā21 A1 B3 (p 6= 2)
260
G2 B3
(10, 000) + (00, 100) (p 6= 2) or Θ(10; 100) (p = 2)
261
A1 B2 B3 (p = 2) Θ(2; 10; 100)
262
B3 via (100, 100[r]) (r 6= 0)
In M1 = G2 B3 (D7 (#260))
102 {0 } Ā1 A1 B3
263
A2 B3 (p = 3)
(11, 000) + (00, 100)
264
A1 B3 (p ≥ 7)
(6, 000) + (0, 100)
248
G2 Ā3
99
G2 Ā21 A1
265
G22
(10, 00) + (00, 10) (p 6= 2) or Θ(10; 10) (p = 2)
266
G2 A1 B2 (p = 2) Θ(10; 2; 10)
In M2 = A2 B3 (p = 3) (D7 (#263))
103 {0 } A1 B3
181
A2 Ā3
104
A2 Ā21 A1
267
A2 G2
(11, 00) + (00, 10)
In M3 = A2 G2 (p = 3) (D7 (#267))
108 {0 } A1 G2
110 {0 } A2 Ā1 A1
25
26
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
268
A22
(11, 00) + (00, 11)
2
In M4 = A2 (p = 3) (D7 (#268))
111 {δ2 } A1 A2
269
A2 via (10, 10[r]) (r 6= 0)
In M2 = A1 B3 (p ≥ 7) (D7 (#264))
264
A1 Ā3
105
A1 Ā21 A1
270
A1 G2
(6, 0) + (0, 10)
In M3 = A1 G2 (p ≥ 7) (D7 (#270))
113 {0 } A1 Ā1 A1
271
A21
(6, 0) + (0, 6)
2
In M4 = A1 (p ≥ 7) (D7 (#271))
272
A1 via (1, 1[r] ) (r 6= 0)
In M2 = G22 (D7 (#265))
107 {0 } Ā1 A1 G2
270
A1 G2 (p ≥ 7)
267
A2 G2 (p = 3)
273
G2 via (10, 10[r]) (r 6= 0)
274
G2 (p = 3) via (10, 03[r])
In M2 = G2 A1 B2 (p = 2) (D7 (#266))
124 {0 } Ā1 A1 A1 B2
99
G2 A1 Ā21
275
G2 A31
Θ(10; 2; 2; 2)
In M3 = G2 A31 (p = 2) (D7 (#275))
132 {0 } Ā1 A1 A31
[r]
276
G2 A1 A1 via (10, 1a , 1a , 1b ) (r 6= 0)
277
G2 A1 via (10, 1, 1[r], 1[s] ) (0 < r < s)
In M1 = A1 B2 B3 (p = 2) (D7 (#261))
12
A1 Ā21 B3
278
A31 B3
Θ(2; 2; 2; 100)
179
A1 B2 Ā3
266
A1 B2 G2
279
A21 B22
Θ(2; 2; 10; 10)
In M2 = A31 B3 (p = 2) (D7 (#278))
250
A31 Ā3
275
A31 G2
280
A41 B2
Θ(2; 2; 2; 2; 10)
[r]
281
A1 A1 B3 via (1a , 1a , 1b , 100) (r 6= 0)
[r] [s]
282
A1 B3 via (1, 1 , 1 , 100) (0 < r < s)
In M3 = A41 B2 (p = 2) (D7 (#280))
122
A41 Ā21
283
A61
Θ(2; 2; 2; 2; 2; 2)
[r]
2
284
A1 A1 B2 via (1a , 1a , 1b , 1c , 10) (r 6= 0)
[r] [s]
285
A1 A1 B2 via (1a , 1a , 1a , 1b , 10) (0 < r < s)
[r]
[s]
286
A1 A1 B2 via (1a , 1a , 1b , 1b , 10) (rs 6= 0; r ≤ s)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
287
A1 B2 via (1, 1[r] , 1[s] , 1[t] , 10) (0 < r < s < t)
In M4 = A61 (p = 2) (D7 (#283))
See Table 36
In M2 = A21 B22 (p = 2) (D7 (#279))
100
A21 Ā21 B2
280
A41 B2
See Table 37
In M = B2 B4 (D7 (#5))
11
Ā21 B4
303
A1 B4 (p ≥ 5)
(4, 0) + (0, λ1 )
304
A21 B4 (p = 2)
Θ(2; 2; λ1 )
175
B2 D 4
179
B2 A1 Ā3 (p 6= 2)
13
B2 Ā21 B2 (p 6= 2)
305
B2 A21 (p 6= 2)
(10, 0, 0) + (00, 2, 2)
306
B2 A1 (p ≥ 11)
(10, 0) + (0, 8)
261
B2 A1 B3 (p = 2)
307
B23 (p = 2)
Θ(10; 10; 10)
In M1 = A1 B4 (p ≥ 5) (D7 (#303))
182
A1 D4
249
A1 A1 Ā3
156
A1 Ā21 B2
308
A1 A21
(4, 0, 0) + (0, 2, 2)
309
A1 A1 (p ≥ 11)
(4, 0) + (0, 8)
2
In M2 = A1 A1 (p ≥ 5) (D7 (#308))
[r] [s]
310
A1 A1 via (1a , 1a , 1b )
[r]
311
A1 A1 via (1a , 1b , 1b ) (r 6= 0)
169 {0 } A1 A1 (p ≥ 5) via (1a , 1b , 1b )
312
A1 via (1[r] , 1[s] , 1[t] ) (rs = 0; s < t)
In M2 = A1 A1 (p ≥ 11) (D7 (#309))
313
A1 via (1[r] , 1[s] )
In M1 = A21 B4 (p = 2) (D7 (#304))
176
A21 D4
278
A31 B3 (p = 2)
279
A21 B22 (p = 2)
314
A1 B4 via (1, 1[r] , λ1 ) (r 6= 0)
In M1 = B2 A21 (p 6= 2) (D7 (#305))
85
Ā21 A21
308
A1 A21 (p ≥ 5)
315
B2 A1 via (10, 1, 1[r]) (r 6= 0)
{0 }
B2 A1 (p ≥ 5) via (10, 1, 1)
165
In M1 = B2 A1 (p ≥ 11) (D7 (#306))
86
Ā21 A1
309
A1 A1
In M1 = B23 (p = 2) (D7 (#307))
27
28
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
13
Ā21 B22
279
A21 B22
[r]
316
B2 B2 via (10a , 10a , 10) (r 6= 0)
[r]
317
B2 B2 via (10a , 02a , 10)
318
B2 via (10, 10[r], 10[s] ) (0 < r < s)
319
B2 via (10[r], 10[s] , 02[t] ) (rt = 0; r < s)
In M = A1 B5 (D7 (#6))
15 {0 }
A1 D5
176
A21 D4 (p 6= 2)
12
A1 Ā21 B3 (p 6= 2)
179
A1 B2 Ā3 (p 6= 2)
320
A1 A1 (p ≥ 11)
(2, 0) + (0, 10)
304
A21 B4 (p = 2)
261
A1 B2 B3 (p = 2)
In M1 = A1 A1 (D7 (#320))
321
A1 via (1[r] , 1[s] ) (rs = 0)
In M = C3 (D7 (#7))
197 {0 } A1 B2
310 {0 } A1 A1 (p ≥ 5)
91 {0 }
A1 A1 (p = 3)
322
A3 (p = 2)
101
9b
G2 (p = 2)
01
313 {0 } A1 (p ≥ 11)
In M = B2 (p 6= 2) (D7 (#8))
94 {0 }
A21
{0 }
A1 (p ≥ 11)
313
{δ1 }
193
A1 (p = 5)
In M = G2 (p 6= 3) (D7 (#9))
245 {0 } A1 A1 (p ≥ 5)
215 {δ3 } A1 A1 (p = 2)
321 {0 } A1 (p ≥ 11)
5.3. Irreducible diagonal subgroups.
In this section we give the tables of diagonal subgroups referred to in Tables 13 and 14. The first
column gives the ID number, as in the previous tables, and the second column gives the embedding of the
diagonal subgroups. To describe the embeddings we use a slightly modified notation, to shorten the tables.
Specifically, we introduce a shorthand for diagonal subgroups of An1 Z, where Z has no simple factor of type
[r] [s]
A1 . For example, instead of writing A21 B2 ֒→ A31 B2 via (1a , 1a , 1b , 10) we just write (a[r] , a[s] , b, 10); from
any such vector it is easy to recover the isomorphism type of the diagonal subgroup. Moreover, in Tables
30 and 37 the usual notation for diagonal subgroups is used but we again omit the isomorphism type of
each diagonal subgroup as they too can be easily recovered from the listed embedding.
Table 16 gives the extra restrictions on the field twists for D6 (#38), a diagonal subgroup of Ā61 = D6 (#8).
The restrictions are given in rows of the tables: the first column lists all permitted equalities amongst the
field twists; the second column lists any further requirements. So an ordered set {r, . . . , w} is permitted
if it satisfies the conditions in the first and second column of a row of the table. We note that a set of
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
29
field twists satisfies the conditions of at most one row. We emphasise that an ordered set may be excluded
either because it yields a D7 -reducible subgroup, or because it yields a repeated diagonal subgroup.
5.3.1. Diagonal subgroups contained in D6 .
Table 15: Irreducible diagonal subgroups of Ā61 = D6 (#8).
ID
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
Diagonal embedding
(a, a[r] , b, c, d, e)
(a, b, a[r] , c, d, e)
(a[r] , a[s] , a[t] , b, c, d) (rt = 0; r ≤ s)
(a, b, a[r] , c, a[s] , d) (r ≤ s)
(a, b, a[r] , c, d, a[s] ) (r ≤ s)
(a, a[r] , b, b[s] , c, d) (r ≤ s; if r = 0 then s 6= 0)
(a[r] , a[s] , b, c, b[t], d) (rs = 0)
(a, b[s] , a[r] , b[t] , c, d) (r ≤ s + t; if r = 0 then t 6= 0)
(a[r] , b[t] , a[s] , c, b[u] , d) (rs = tu = 0)
(a, a[r] , a[s] , a[t] , b, c) (s ≤ t; if r = 0 then s < t; if s = 0 then r < t; if t = 0 then r < s)
(a[r] , a[s] , a[t] , b, a[u] , c) (rst = 0; t ≤ u);
(a[r] , a[s] , a[t] , b[u] , b[v] , c) (rst = uv = 0)
(a[r] , a[s] , a[t] , b, c, c[u] ) (rst = 0; if u = 0 then r < s)
(a[r] , b, a[s] , b[u] , a[t] , c) (rst = 0; if u = 0 then r < s)
(a[r] , b, a[s] , b[u] , c, a[t] ) (rst = 0; if u = 0 then r < s)
(a, a[r] , b, b[s] , c[t] , c[u] ) (tu = 0; r ≤ s; if r = 0 then s 6= 0 and t < u; if s = 0 then t < u)
(a[r] , a[s] , b, c[u] , b[t] , c[v] ) (rs = uv = 0; t ≤ u + v; if t = 0 then v 6= 0)
(a[r] , b[t] , a[s] , c[v] , b[u] , c[w] ) (rs = tu = vw = 0)
(a[r] , a[s] , a[t] , a[u] , a[v] , b) (rv = 0; r ≤ min{s, t, u}; if r = s then t < u; if r = t then s < u; if r = u then s < t)
(a, a[r] , a[s] , a[t] , b, b[u] ) (if u = 0 then r 6= 0 and s < t; if r = 0 then s < t; if s = 0 then r < t; if t = 0 then
r < s)
(a[r] , a[s] , a[t] , b, a[u] , b[v] ) (rstu = 0; if v = 0 then t < u)
(a[r] , a[s] , a[t] , b[u] , b[v] , b[w] ) (rt = uvw = 0; r ≤ s; if r = s then v < w)
(a, b[t] , a[r] , b[u] , a[s] , b[v] ) (tuv = 0; r ≤ s; if r = 0 then t < u; if r = s then u < v; if s = 0 then t < v)
(a[r] , a[s] , a[t] , a[u] , a[v] , a[w] ) (rstuvw = 0 and see Table 16)
Table 16: Conditions on field twists for D6 (#38).
Equalities among r, . . . , w
none
r=s
r=t
r=s=t
r=t=v
r=t=w
r = s; t = v
r = t; s = v
r=s=t=v
Further requirements on r, . . . , w
r < min{s, t}; t < min{u, v, w}; s < u
t < min{u, v}; v < w
s<u
v<w
s<u<w
s<u<v
u<w
none
u<w
30
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
r = s = t; u = v
r = t; s = v; u = w
none
none
Table 17: Irreducible diagonal subgroups of Ā21 A1 A1 = D6 (#43) (p ≥ 5).
ID
48
49
50
51
52
Diagonal embedding
(a, a[r] , b, c)
(a[r] , b, a[s] , c) (rs = 0)
(a[r] , b, c, a[s] ) (rs = 0)
(a, b, c[r] , c[s] ) (rs = 0)
(a[r] , a[s] , a[t] , b) (rt = 0; r ≤ s; if r = s then r < t)
ID
53
54
55
56
57
Diagonal embedding
(a[r] , a[s] , b, a[t] ) (rt = 0; r ≤ s)
(a[r] , b, a[s] , a[t] ) (rst = 0)
(a, a[r] , b[s] , b[t] ) (st = 0)
(a[r] , b[t] , a[s] , b[u] ) (rs = tu = 0)
(a[r] , a[s] , a[t] , a[u] ) (rtu = 0; r ≤ s;
if r = s then r < t)
Table 18: Irreducible diagonal subgroups of Ā21 A31 = D6 (#44) (p = 2).
ID
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
Diagonal embedding
(a, a[r] , b, c, d) (r 6= 0)
(a[r] , b, a[s] , c, d) (rs = 0)
(a, b, c, c[r], d) (r 6= 0)
(a[r] , a[s] , a[t] , b, c) (rt = 0; r < s)
(a[r] , b, a[s] , a[t] , c) (rs = 0; s < t)
(a, b, c, c[r], c[s] ) (0 < r < s)
(a, a[r] , b, b[s] , c) (rs 6= 0)
(a[r] , b[t] , a[s] , b[u] , c) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , b, a[s] , c, c[t] ) (rs = 0 6= t)
(a[r] , a[s] , a[t] , a[u] , b) (rt = 0; r < s; t < u)
(a[r] , b, a[s] , a[t] , a[u] ) (rs = 0; s < t < u)
(a[r] , a[s] , a[t] , b, b[u] ) (rt = 0 6= u; r < s)
(a[r] , b[u] , a[s] , a[t] , b[v] ) (rs = uv = 0; s < t)
(a, a[r] , b, b[s] , b[t] ) (r 6= 0; 0 < s < t)
(a[r] , a[s] , a[t] , a[u] , a[v] ) (rt = 0; r < s; t < u < v)
Table 19: Irreducible diagonal subgroups of Ā21 A1 A1 = D6 (#73) (p ≥ 5).
ID
79
80
81
82
83
84
85
Diagonal embedding
(a[r] , a[s] , b, c) (rs = 0)
(a[r] , b, a[s] , c) (rs = 0)
(a[r] , b, c, a[s] ) (rs = 0)
(a, b[r] , b[s] , c) (rs = 0)
(a, b[r] , c, b[s] ) (rs = 0)
(a, b, c[r] , c[s] ) (rs = 0; r 6= s)
(a[r] , a[s] , a[t] , b) (rst = 0)
ID
86
87
88
89
90
91
92
Diagonal embedding
(a[r] , a[s] , b, a[t] ) (rst = 0)
(a[r] , b, a[s] , a[t] ) (rst = 0; s 6= t)
(a, b[r] , b[s] , b[t] ) (rst = 0; s 6= t)
(a[r] , a[s] , b[t] , b[u] ) (rs = tu = 0; t 6= u)
(a[r] , b[t] , a[s] , b[u] ) (rs = tu = 0)
(a[r] , b[t] , b[u] , a[s] ) (rs = tu = 0)
(a[r] , a[s] , a[t] , a[u] ) (rstu = 0; t 6= u)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
Table 20: Irreducible diagonal subgroups of Ā21 A31 = D6 (#74) (p = 2).
ID
93
94
95
96
97
98
99
100
101
102
Diagonal embedding
(a[r] , a[s] , b, c, d) (rs = 0; r 6= s)
(a[r] , b, a[s] , c, d) (rs = 0)
(a, b[r] , b[s] , c, d) (rs = 0)
(a, b, c, c[r], d) (r 6= 0)
(a[r] , a[s] , a[t] , b, c) (rst = 0; r 6= s)
(a[r] , b, a[s] , a[t] , c) (rs = 0; s < t)
(a, b[r] , b[s] , b[t] , c) (rs = 0; s < t)
(a, b, c, c[r], c[s] ) (0 < r < s)
(a[r] , a[s] , b, b[t] , c) (rs = 0; r 6= s; t 6= 0)
(a[r] , b[t] , a[s] , b[u] , c) (rs = tu = 0)
ID
103
104
105
106
107
108
109
110
111
112
Diagonal embedding
(a[r] , b, a[s] , c, c[t] ) (rs = 0 6= t)
(a, b[r] , b[s] , c, c[t] ) (rs = 0 6= t)
(a[r] , a[s] , a[t] , a[u] , b) (rst = 0; r 6= s; t < u)
(a[r] , b, a[s] , a[t] , a[u] ) (rs = 0; s < t < u)
(a, b[r] , b[s] , b[t] , b[u] ) (rs = 0; s < t < u)
(a[r] , a[s] , a[t] , b, b[u] ) (rst = 0; r 6= s; u 6= 0)
(a[r] , b[u] , a[s] , a[t] , b[v] ) (rs = uv = 0; s < t)
(a[r] , b[t] , b[u] , b[v] , a[s] ) (rs = tu = 0; u < v)
(a[r] , a[s] , b, b[t] , b[u] ) (rs = 0; r 6= s; 0 < t < u)
(a[r] , a[s] , a[t] , a[u] , a[v] ) (rst = 0; r 6= s; t < u < v)
Table 21: Irreducible diagonal subgroups of A51 = D6 (#137) (p = 2).
ID
140
141
142
Diagonal embedding
(a, a[r] , b, c, d) (r 6= 0)
(a, a[r] , a[s] , b, c) (0 < r < s)
(a, a[r] , b, b[s] , c) (rs 6= 0; r ≤ s)
ID
143
144
145
Diagonal embedding
(a, a[r] , a[s] , a[t] , b) (0 < r < s < t)
(a, a[r] , a[s] , b, b[t]) (0 < r < s; t 6= 0)
(a, a[r] , a[s] , a[t] , a[u] ) (0 < r < s < t < u)
5.3.2. Diagonal subgroups contained in D7 .
Table 22: Irreducible diagonal subgroups of Ā41 Ā3 = D7 (#10).
ID
18
19
20
21
22
23
Diagonal embedding
(a, a[r] , b, c, 100)
(a, b, a[r], c, 100)
(a[r] , a[s] , a[t] , b, 100) (rt = 0; r ≤ s)
(a, a[r] , b, b[s] , 100) (r ≤ s; if r = 0 then s 6= 0)
(a, b[s] , a[r] , b[t] , 100) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
(a, a[r] , a[s] , a[t] , 100) (s ≤ t; if r = 0 then s < t; if s = 0 then r < t)
Table 23: Irreducible diagonal subgroups of Ā41 B2 = D7 (#16).
ID
25
26
27
28
29
30
Diagonal embedding
(a, a[r] , b, c, 10) (r 6= 0)
(a, b, a[r] , c, 10)
(a[r] , a[s] , a[t] , b, 10) (rt = 0; r < s)
(a, a[r] , b, b[s] , 10) (rs 6= 0; r ≤ s)
(a, b[s] , a[r] , b[t] , 10) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
(a, a[r] , a[s] , a[t] , 10) (r 6= 0; s < t; if s = 0 then r < t)
31
32
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Table 24: Irreducible diagonal subgroups of Ā41 A1 = D7 (#24) (p ≥ 5).
ID
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
Diagonal embedding
(a, a[r] , b, c, d) (r 6= 0)
(a, b, a[r] , c, d)
(a[r] , b, c, d, a[s] ) (rs = 0)
(a[r] , a[s] , a[t] , b, c) (rt = 0; r < s)
(a[r] , a[s] , c, d, a[t] ) (rt = 0; r < s)
(a[r] , b, a[s] , c, a[t] ) (rt = 0; r ≤ s)
(a, a[r] , b, b[s] , c) (rs 6= 0; r ≤ s)
(a, a[r] , b[s] , c, b[t] ) (r 6= 0; st = 0)
(a, b[s] , a[r] , b[t] , c) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
(a[r] , b[t] , a[s] , c, b[u] ) (rs = tu = 0)
(a, a[r] , a[s] , a[t] , b) (r 6= 0; s < t; if s = 0 then r < t)
(a[r] , a[s] , a[t] , b, a[u] ) (rtu = 0; r < s)
(a, a[r] , b[s] , b[t] , b[u] ) (r 6= 0; su = 0; s < t)
(a, b[s] , a[r] , b[t] , b[u] ) (stu = 0; if r = 0 then s ≤ t)
(a[r] , b[t] , b[u] , b[v] , a[s] ) (rs = tu = 0; u < v)
(a[r] , a[s] , a[t] , a[u] , a[v] ) (rv = 0; r < s; t < u; r ≤ t; if r = t then s < u)
Table 25: Irreducible diagonal subgroups of Ā41 A21 = D7 (#17).
ID
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
Diagonal embedding
(a, a[r] , b, c, d, e) (if p = 2 then r 6= 0)
(a, b, a[r] , c, d, e)
(a[r] , b, c, d, a[s] , d) (rs = 0)
(a, b, c, d, e, e[r]) (r 6= 0)
(a[r] , a[s] , a[t] , b, c, d) (rt = 0; r ≤ s; if p = 2 then r < s)
(a[r] , a[s] , b, c, a[t] , d) (rt = 0; r ≤ s; if r = s then r < t and p 6= 2)
(a[r] , b, a[s] , c, a[t] , d) (rt = 0; r ≤ s)
(a[r] , b, c, d, a[s] , a[t] ) (rs = 0; s < t)
(a, a[r] , b, b[s] , c, d) (r ≤ s; s 6= 0; if p = 2 then r 6= 0)
(a, a[r] , b[s] , c, b[t] , d) (if p = 2 then r 6= 0)
(a, a[r] , b, c, d, d[s] ) (s 6= 0; if p = 2 then r 6= 0)
(a, b[s] , a[r] , b[t] , c, d) (st = 0; r ≤ s + t; if r = 0 then t 6= 0)
(a[r] , b[t] , a[s] , c, b[u] , d) (rs = tu = 0)
(a, b, a[r] , c, d, d[s] ) (s 6= 0)
(a[r] , b[t] , c, d, a[s] , b[u] ) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , b, c[t] , d, a[s] , c[u] ) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a, a[r] , a[s] , a[t] , b, c) (s ≤ t; if r = 0 then p 6= 2 and s < t; if s = 0 then r < t; if s = t then p 6= 2)
(a[r] , a[s] , a[t] , b, a[u] , c) (rtu = 0; r ≤ s; if r = s then r < u and p 6= 2)
(a[r] , a[s] , b, c, a[t] , a[u] ) (rt = 0; r ≤ s; t < u; if r = s then r < t and p 6= 2)
(a[r] , b, a[s] , c, a[t] , a[u] ) (rt = 0; r ≤ s; t < u)
(a[r] , a[s] , a[t] , b[u] , b[v] , c) (rt = uv = 0; r ≤ s; if p = 2 then r < s)
(a[r] , a[s] , a[t] , b, c, c[u] ) (rt = 0 6= u; r ≤ s; if p = 2 then r < s)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
33
(a[r] , a[s] , b, b[u] , a[t] , c) (rt = 0; r ≤ s; if r = s then r < t, u 6= 0 and p 6= 2; if u = 0 then p 6= 2)
(a[r] , a[s] , b[u] , c, a[t] , b[v] ) (rt = uv = 0; r ≤ s; if r = s then r < t and p 6= 2)
(a[r] , b[u] , a[s] , b[v] , a[t] , c) (rt = uv = 0; r ≤ s; if r = s then u < v)
(a[r] , b[u] , a[s] , c, a[t] , b[v] ) (rst = uv = 0)
(a[r] , b[u] , b[v] , c, a[s] , a[t] ) (rs = uv = 0; s < t)
(a[r] , b, c, c[u] , a[s] , a[t] ) (rs = 0; s < t; if p = 2 then u 6= 0)
(a, a[r] , b, b[s] , c, c[u] ) (r ≤ s; su 6= 0; if p = 2 r 6= 0)
(a, a[r] , b[s] , c[u] , b[t] , c[v] ) (uv = 0; s ≤ u; if s = u then t ≤ v; if p = 2 then r 6= 0)
(a, b[s] , a[r] , b[t] , c, c[u] ) (st = 0 6= u; r ≤ s + t; if r = 0 then t 6= 0)
(a, b[s] , a[r] , c[u] , b[t] , c[v] ) (st = uv = 0; if r = 0 then s ≤ u; if r = 0 and s = u then t ≤ v)
(a[r] , a[s] , a[t] , a[u] , a[v] , b) (rv = 0; r ≤ s; r ≤ t ≤ u; if r = s then r < v, t < u and p 6= 2; if r = t then s < u; if
t = u then p 6= 2)
(a[r] , a[s] , a[t] , b, a[u] , a[v] ) (rtu = 0; r ≤ s; u < v; if r = s then r < u and p 6= 2)
(a[r] , a[s] , a[t] , b[u] , b[v] , b[w] ) (rt = uv = 0; r ≤ s; v < w; if p = 2 then r < s)
(a[r] , a[s] , b[u] , b[v] , a[t] , b[w] ) (rt = uw = 0; r ≤ s; u ≤ v; r ≤ u; if r = s then r < t, u < v and p 6= 2; if u = v
then u < w and p 6= 2; if r = u then s ≤ v; if r = u and s = v then t ≤ w)
(a[r] , b[u] , a[s] , b[v] , a[t] , b[w] ) (rt = 0; r ≤ s; r ≤ u; if r = s then u < v; if r = u then s ≤ v; if r = u and s = v
then t ≤ w)
(a[r] , a[s] , a[t] , a[u] , a[v] , a[w] ) (rv = 0; r ≤ s; r ≤ t ≤ u; v < w; if r = s then r < v, t < u and p 6= 2; if r = t then
s < u; if t = u then t < v and p 6= 2)
Table 26: Irreducible diagonal subgroups of Ā21 A21 = D7 (#85) (p 6= 2).
ID
88
89
90
91
Diagonal embedding
(a, a[r] , b, c) (r 6= 0)
(a[r] , b, a[s] , c) (rs = 0)
(a, b, c, c[r] ) (r 6= 0)
(a[r] , a[s] , a[t] , b) (rt = 0; r < s)
ID
92
93
94
95
Diagonal embedding
(a[r] , b, a[s] , a[t] ) (rs = 0; s < t)
(a, a[r] , b, b[s] ) (rs 6= 0)
(a[r] , b[t] , a[s] , b[u] ) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , a[s] , a[t] , a[u] ) (rt = 0; r < s; t < u)
Table 27: Irreducible diagonal subgroups of Ā21 A1 A1 = D7 (#105) (p ≥ 7).
ID
112
113
114
115
116
Diagonal embedding
(a, a[r] , b, c)
(a[r] , b, a[s] , c) (rs = 0)
(a[r] , b, c, a[s] ) (rs = 0)
(a, b, c[r] , c[s] ) (rs = 0)
(a[r] , a[s] , a[t] , b) (rt = 0; r ≤ s; if r = s then r < t)
ID
117
118
119
120
121
Diagonal embedding
(a[r] , a[s] , b, a[t] ) (rst = 0)
(a[r] , b, a[s] , a[t] ) (rst = 0)
(a, a[r] , b[s] , b[t] ) (st = 0)
(a[r] , b[t] , a[s] , b[u] ) (rs = tu = 0)
(a[r] , a[s] , a[t] , a[u] ) (rtu = 0; r ≤ s;
if r = s then r < t)
Table 28: Irreducible diagonal subgroups of Ā21 A21 B2 = D7 (#100) (p = 2).
ID
123
124
125
Diagonal embedding
(a, a[r] , b, c, 10) (r 6= 0)
(a[r] , b, a[s] , c, 10) (rs = 0)
(a, b, c, c[r] , 10) (r 6= 0)
34
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
126
127
128
129
130
(a[r] , a[s] , a[t] , b, 10) (rt = 0; r < s)
(a[r] , b, a[s] , a[t] , 10) (rs = 0; s < t)
(a, a[r] , b, b[s] , 10) (rs 6= 0)
(a[r] , b[t] , a[s] , b[u] , 10) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , a[s] , a[t] , a[u] , 10) (rt = 0; r < s; t < u)
Table 29: Irreducible diagonal subgroups of Ā21 A41 = D7 (#122) (p = 2).
ID
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
Diagonal embedding
(a, a[r] , b, c, d, e) (r 6= 0)
(a[r] , b, a[s] , c, d, e) (rs = 0)
(a, b, c, c[r] , d, e) (r 6= 0)
(a[r] , a[s] , a[t] , b, c, d) (rt = 0; r < s)
(a[r] , b, a[s] , a[t] , c, d) (rs = 0; s < t)
(a, b, c, c[r] , c[s] , d) (0 < r < s)
(a, a[r] , b, b[s] , c, d) (rs 6= 0)
(a[r] , b[t] , a[s] , b[u] , c, d) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , b, a[s] , c, c[t] , d) (rs = 0; t 6= 0)
(a, b, c, c[r] , d, d[s] ) (rs 6= 0; r ≤ s)
(a[r] , a[s] , a[t] , a[u] , b, c) (rt = 0; r < s; t < u)
(a[r] , b, a[s] , a[t] , a[u] , c) (rs = 0; s < t < u)
(a, b, c, c[r] , c[s] , c[t] ) (0 < r < s < t)
(a[r] , a[s] , a[t] , b, b[u] , c) (rt = 0 6= u; r < s)
(a[r] , b[u] , a[s] , a[t] , b[v] , c) (rs = uv = 0; s < t)
(a[r] , b, a[s] , a[t] , c, c[u] ) (rs = 0 6= u; s < t)
(a, a[r] , b, b[s] , b[t] , c) (r 6= 0; 0 < s < t)
(a[r] , b, a[s] , c, c[t] , c[u] ) (rs = 0; 0 < t < u)
(a[r] , a[s] , a[t] , a[u] , a[v] , b) (rt = 0; r < s; t < u < v)
(a[r] , b, a[s] , a[t] , a[u] , a[v] ) (rs = 0; s < t < u < v)
(a[r] , a[s] , a[t] , b, b[v] , b[w]) (rt = 0; r < s)
(a[r] , b[u] , a[s] , a[t] , b[v] , b[w]) (rs = uv = 0; s < t; v < w)
(a, a[r] , b, b[s] , c, c[t] ) (rst 6= 0; s ≤ t)
(a[r] , b[t] , a[s] , b[u] , c, c[v] ) (rs = tu = 0 6= v; r ≤ t; if r = t then s ≤ u)
(a[r] , a[s] , a[t] , a[u] , a[v] , a[w] ) (rt = 0; r < s; t < u < v < w)
Table 30: Irreducible diagonal subgroups of Ā21 B22 = D7 (#13).
ID
Diagonal embedding
157
158
159
[r]
(1a , 1a , 10b , 10c )
[r]
(1a , 1b , 10c , 10c )
[r]
(1a , 1b , 10c , 02c )
ID
Diagonal embedding
[r]
[s]
(if p = 2 then r 6= 0) 160 (1a , 1a , 10b , 10b ) (s 6= 0; if p = 2 then r 6= 0)
[r]
[s]
(r 6= 0)
161 (1a , 1a , 10b , 02b ) (r =
6 0)
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
Table 31: Irreducible diagonal subgroups of Ā21 A21 = D7 (#162) (p ≥ 5).
ID
166
167
168
169
Diagonal embedding
(a, a[r] , b, c)
(a[r] , b, a[s] , c) (rs = 0)
(a, b, c, c[r] ) (r 6= 0)
(a[r] , a[s] , a[t] , b) (rt = 0; r ≤ s)
ID
170
171
172
173
Diagonal embedding
(a[r] , b, a[s] , a[t] ) (rs = 0; s < t)
(a, a[r] , b, b[s] ) (s 6= 0)
(a[r] , b[t] , a[s] , b[u] ) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , a[s] , a[t] , a[u] ) (rt = 0; r < s; t < u)
Table 32: Irreducible diagonal subgroups of A1 A1 A1 = D7 (#189) (p ≥ 5).
ID
191
192
193
Diagonal embedding
ID
Diagonal embedding
{0 }
[r] [s]
(a, b, b)
(a , a , b) (rs = 0)
169
(a[r] , b, a[s] ) (rs = 0)
194
(a[r] , a[s] , a[t] ) (rst = 0; s 6= t)
(a, b[r] , b[s] ) (rs = 0; r 6= s)
Table 33: Irreducible diagonal subgroups of A21 A31 = D7 (#200) (p = 2).
ID
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
Diagonal embedding
(a, a[r] , b, c, d) (r 6= 0)
(a[r] , b, a[s] , c, d) (rs = 0)
(a, b, c, c[r] , d) (r 6= 0)
(a[r] , a[s] , a[t] , b, c) (rt = 0; r < s)
(a[r] , b, a[s] , a[t] , c) (rs = 0; s < t)
(a, b, c, c[r] , c[s] ) (0 < r < s)
(a, a[r] , b, b[s] , c) (rs 6= 0)
(a[r] , b[t] , a[s] , b[u] , c) (rs = tu = 0; r ≤ t; if r = t then s ≤ u)
(a[r] , b, a[s] , c, c[t] ) (rs = 0 6= t)
(a[r] , a[s] , a[t] , a[u] , b) (rt = 0; r < s; t < u)
(a[r] , b, a[s] , a[t] , a[u] ) (rs = 0; s < t < u)
(a[r] , a[s] , a[t] , b, b[u] ) (rt = 0 6= u; r < s)
(a[r] , b[u] , a[s] , a[t] , b[v] ) (rs = uv = 0; s < t)
(a, a[r] , b, b[s] , b[t] ) (r 6= 0; 0 < s < t)
(a[r] , a[s] , a[t] , a[u] , a[v] ) (rt = 0; r < s; t < u < v)
Table 34: Irreducible diagonal subgroups of A31 A1 = D7 (#224) (p ≥ 5).
ID
227
228
229
Diagonal embedding
(a, a[r] , b, c) (r 6= 0)
(a[r] , b, c, a[s] ) (rs = 0)
(a, a[r] , a[s] , b) (0 < r < s)
ID
230
231
232
Diagonal embedding
(a[r] , a[s] , b, a[t] ) (rt = 0; r < s)
(a, a[r] , b[s] , b[t] ) (r 6= 0 = st)
(a[r] , a[s] , a[t] , a[u] ) (ru = 0; r < s < t)
35
36
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Table 35: Irreducible diagonal subgroups of A21 A1 A1 = D7 (#233) (p ≥ 5).
ID
237
238
239
240
228 {0 }
241
Diagonal embedding
(a, a[r] , b, c) (r 6= 0)
(a[r] , b, a[s] , c) (rs = 0)
(a[r] , b, c, a[s] ) (rs = 0)
(a, b, c[r] , c[s] ) (rs = 0; r 6= s)
(a, b, c, c)
(a[r] , a[s] , a[t] , b) (rt = 0; r < s)
242
243
244
245
246
(a[r] , a[s] , b, a[t] ) (rt = 0; r < s)
(a[r] , b, a[s] , a[t] ) (rst = 0; s 6= t)
(a, a[r] , b[s] , b[t] ) (r 6= 0 = st; s 6= t)
(a[r] , b[t] , a[s] , b[u] ) (rs = tu = 0)
(a[r] , a[s] , a[t] , a[u] ) (rtu = 0; r < s; t 6= u)
Table 36: Irreducible diagonal subgroups of A61 = D7 (#283) (p = 2).
ID
288
289
290
291
292
293
Diagonal embedding
(a, a[r] , b, c, d, e) (r 6= 0)
(a, a[r] , a[s] , b, c, d) (0 < r < s)
(a, a[r] , b, b[s] , c, d) (rs 6= 0; r ≤ s)
(a, a[r] , a[s] , a[t] , b, c) (0 < r < s < t)
(a, a[r] , a[s] , b, b[t] , c) (0 < r < s; t 6= 0)
(a, a[r] , b, b[s] , c, c[t] ) (rst 6= 0; r ≤ s ≤ t)
294 (a, a[r] , a[s] , a[t] , a[u] , b) (0 < r < s < t < u)
295 (a, a[r] , a[s] , a[t] , b, b[u] ) (0 < r < s < t; u 6= 0)
296 (a, a[r] , a[s] , b, b[t] , b[u] ) (0 < r < s; 0 < t < u; r ≤ t;
if r = t then s ≤ u)
297 (a, a[r] , a[s] , a[t] , a[u] , a[v] ) (0 < r < s < t < u < v)
Table 37: Irreducible diagonal subgroups of A21 B22 = D7 (#279) (p = 2).
ID
Diagonal embedding
[r]
[r]
[s]
298 (1a , 1b , 10c, 10d ) (r 6= 0) 301 (1a , 1a , 10b , 10b ) (rs 6= 0)
[r]
[r]
[s]
299 (1a , 1b , 10c, 10c ) (r 6= 0) 302 (1a , 1a , 10b , 02b ) (r 6= 0)
[r]
300 (1a , 1b , 10c, 02c )
6. Applying the main theorem
We now describe how Theorem 1 allows one to enumerate all connected G-cr subgroups, and illustrate
with the case G = F4 (K), in arbitrary characteristic. In this section we present a series of additional
lemmas which allow one to derive properties of the G-cr subgroups arising. The subgroups themselves,
and various properties thereof, are given in Table 38.
6.1. G-conjugacy vs. L-conjugacy. Once we know the L′ -irreducible subgroups for each possible Levi
subgroup L of our reductive algebraic group G, it remains to consider how the various classes of subgroups
fuse in G. This is achieved in the following result.
Proposition 6.1. Let X and Y be G-cr subgroups of G.
(i) Suppose X and Y are respectively minimally contained in Levi subgroups L1 and L2 of G. If X
and Y are G-conjugate, then so are L1 and L2 .
(ii) Let X and Y be G-completely reducible subgroups of G, both minimally contained in a Levi subgroup
L. Then X and Y are G-conjugate if and only if they are NG (L)-conjugate.
Proof. (i) By replacing L1 and L2 by G-conjugates (and replacing X and Y accordingly), it does no
harm to assume that L1 and L2 are standard Levi subgroups of G, corresponding to fixed sets of positive
roots. If X = Y g then we can write g = p2 ẇp1 where pi lies in the standard parabolic subgroup Pi
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
37
containing Li , and ẇ ∈ NG (T ) lies in a set of distinguished double-coset representatives [7, §2.8]. Then
X = Y g ≤ P1 ∩ P2g = (P1 ∩ P2ẇ )p1 . The subgroup P1 ∩ P2ẇ is contained in a third standard parabolic
subgroup, which is equal to P1 , by minimality of L1 . From minimality of L2 , it follows that Lẇ
2 = L1 [7,
Corollary 2.8.8], and thus L1 and L2 are conjugate.
(ii) One direction is obvious. Conversely, suppose X and Y are G-conjugate and write Y g = X < L.
Then Z(L)g is a maximal torus of CG (X) and CG (X g ), thus Z(L) and Z(L)g are maximal tori of CG (X)◦ ,
so there exists c ∈ CG (X)◦ such that Z(L)g = Z(L)c . In particular, cg−1 = n ∈ NG (Z(L)), and Y =
−
−1
X g 1 = X c n = X n . Finally, note that NG (L) ≤ NG (Z(L)) as Z(L) is characteristic in L, but also
NG (Z(L)) ≤ NG (CG (Z(L))) = NG (L). So n ∈ NG (Z(L)) = NG (L).
In light of the previous proposition, it is desirable to understand the fusion of conjugacy classes of
irreducible subgroups of simple Levi factors L0 induced by graph automorphisms. This is only necessary
when such a graph automorphism is induced by an element of G. All irreducible subgroups are centralised
by a graph automorphism when L0 is of type An for some n, and there are no graph automorphisms of B2 ,
B3 or C3 induced by F4 . So it remains to consider L0 of type Dn . Now, D7 only occurs as a Levi factor
in E8 and the normaliser of D7 in E8 contains an element inducing a graph automorphism of D7 . For this
reason, the irreducible subgroups in Table 13 are given up to Aut(D7 )-conjugacy. This leaves L0 of type
D4 , D5 and D6 . In these cases, the tables present the irreducible subgroups up to L′ -conjugacy, and we
now give information about how these subgroup classes fuse under a graph automorphism. For D5 and D6
we give a list of conditions of two types. The first type we denote ‘l ∼ m’, which means that Dn (#l) is
Aut(Dn )-conjugate to Dn (#m). The second type is denoted ‘l : condition’, where Dn (#l) refers to a set of
diagonal subgroups and the condition is an extra constraint on the field twists. For D4 we are interested in
the fusion induced by both an involutory graph automorphism and the full outer automorphism group S3 .
For this reason, we present a list of conditions, each with an extra one in brackets. The first condition refers
to the fusion induced by an involutory graph automorphism and the condition inside the brackets is to be
added when considering the full outer automorphism group. For example, 3 ∼ 4(∼ 2) explains that the
two B3 subgroups with ID numbers 3 and 4 are conjugate under an involutory graph automorphism and
that all three B3 subgroups with ID numbers 2, 3 and 4 are conjugate under the full outer automorphism
group.
D4 : 3 ∼ 4 (∼ 2); 6 ∼ 7 (∼ 5); 10 ∼ 11 (∼ 9); 12: r ≤ s (s ≤ t); 14 ∼ 15 (∼ 13); 16: s ≤ t (r ≤ s);
24 ∼ 29 (∼ 19); 25 ∼ 30 (∼ 20); 26 ∼ 31 (∼ 21); 27 ∼ 32 (∼ 22); 28 ∼ 33 (∼ 23).
D5 : 5 ∼ 6
D6 : 6 ∼ 7; 18 ∼ 19; 21: r ≤ s; 23: r ≤ t, if r = t then s ≤ u; 25: r ≤ s; 26: r ≤ s; 27: r ≤ s; 28 ∼ 29;
30: t ≤ u; 31: r ≤ s; 32: r ≤ t ≤ v, if r = t then s ≤ u, if t = v then u ≤ w; 33: t ≤ u; 34: s ≤ t; 35:
r ≤ s; 36: v ≤ w; 37: if t = 0 then r ≤ u, if t = 0 and r = u then s ≤ v; 38: in Table 16 remove line 6 and
add the following to the further requirements column: v < w to lines 1 and 3, u < w to line 8, r < s < u
to line 11; 42: r ≤ s; 75: r ≤ s; 76 ∼ 77; 78: r ≤ s; 79: r ≤ s; 80 ∼ 82; 81 ∼ 83; 85: r ≤ s; 86: r ≤ s;
87 ∼ 88; 89: r ≤ s; 90 ∼ 91; 92: r ≤ s; 93: r < s; 94 ∼ 95; 97: r < s; 98 ∼ 99; 101: r < s; 102: r ≤ t
and if r = t then s ≤ u; 103 ∼ 104; 105: r < s; 106 ∼ 107; 108: r < s; 109 ∼ 110; 111: r < s; 112: r < s;
116 ∼ 117; 155 ∼ 163; 156 ∼ 164; 157 ∼ 165; 158 ∼ 160; 159: r < s; 162 ∼ 166.
6.2. Normaliser Structure. Let X be a G-cr subgroup of G. The following result gives us a method
for calculating the structure of NG (X). In particular, one need only inspect the parabolic subgroups of
G which are minimal subject to containing X, and the normaliser of a Levi subgroup which is minimal
subject to containing X. By Proposition 6.1 there is a unique such Levi subgroup up to conjugacy, and
there there are only finitely many minimal parabolic subgroups to consider, corresponding to the different
standard parabolic subgroups having conjugate Levi factors.
38
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Proposition 6.2. Let X be a G-cr subgroup of G.
(i) For every minimal Levi subgroup L containing X, we have
NG (X) = CG (X)◦ (NG (L) ∩ NG (X))
(ii) There exists a minimal parabolic subgroup P containing X, with Levi decomposition P = QL, such
that CG (X)◦ = QX , (Qop )X , Z(L) .
Proof. (i) If X = X g for some g ∈ G, and if L is minimal among Levi subgroups of G containing X,
then Z(L) and Z(L)g are maximal tori of CG (X)◦ , and thus Z(G)g = Z(G)c for some c ∈ CG (X)◦ . Then
def
n = gc−1 ∈ NG (Z(L)) = NG (L). But also X = X g = X gc
−1
= X n , so n ∈ NG (X).
(ii) Let U be a maximal unipotent subgroup of CG (X)◦ . Then there is a parabolic subgroup P of G
def
containing U X, such that U ≤ Q = Ru (P ). Since X is G-cr, there exists a Levi subgroup L of P containing
X, and moreover we can assume that X is L-irreducible, otherwise there is a smaller parabolic subgroup
of G containing both U X and the unipotent radical of a proper parabolic subgroup of L normalised by X.
Then U ≤ QX , and thus U = QX by maximality of U . Since X is L-irreducible, Z(L) is a maximal
torus of G (X)◦ . Therefore U Z(L), which contains a maximal torus of CG (X)◦ and a maximal unipotent
subgroup of CG (X)◦ , is a Borel subgroup of CG (X)◦ . For the same reason, (Qop )X Z(L) is also a Borel
subgroup of CG (X)◦ , and is clearly opposite to QX Z(L). Thus CG (X)◦ is generated by QX Z(L) and
(Qop )X Z(L), which gives the desired result.
6.3. Action on G-modules, separability and reductive Pairs. Given a reductive subgroup X of G,
it is of interest to know how X acts on various G-modules, particularly the Lie algebra L(G) and the nontrivial G-module(s) of minimal dimension. Such information allows one, for instance, to study the conjugacy
classes of G meeting X. Additionally, recall that X is called separable in G if dim CG (X) = dim CL(G) (X);
equivalently, the scheme-theoretic centraliser of X in G is smooth. Once the normaliser structure and
action of X on L(G) is known, one can see directly whether this equality holds. Recall also that (G, X)
is a reductive pair if L(X) is an X-module direct summand of L(G). Both of these properties are closely
related to complete reducibility, cf. [4, Theorem 3.35].
The following is useful in determining the action of X on L(G). Specifically, one first considers the
action of L, where L is minimal among Levi subgroups of G containing X, and then considers the action
of X on each L-module arising.
Lemma 6.3. Let G be a reductive algebraic group, let P be a parabolic subgroup of G and let L be a Levi
subgroup of P . Then the following hold.
(i)
(ii)
(iii)
(iv)
If V is a tilting module for G then it is tilting for L and for L′ ;
op
op
L(G) ↓ L = L(L)
L ⊕ L(Ru (P )) ⊕ L(Ru (P )), where P is the parabolic opposite to P ;
∼
L(Ru (P )) = S VS , where S ranges over shapes and VS is a shape module, defined in [3];
L(Ru (P op )) ∼
= L(Ru (P ))∗ .
Proof. (i) is [8, Proposition 1.2(ii), Lemma 1.4(i)]. (ii) Let T be a maximal torus of G. Then we may take
L to be a standard Levi subgroup, that is, generated by T and the root subgroups corresponding to a set of
simple roots I, and their negatives. Then L(L) is spanned by L(T ) and the root subspaces corresponding
to the root subsystem spanned by I. The other two summands are respectively spanned by positive and
by negative root subspaces whose roots are not in the span of I; it is clear that L preserves each of these
three summands (cf. [6, p. 64] for the action of root elements on root spaces). (iii) This follows directly
from [3], since T preserves every root subspace, and a root subgroup of L acting on a root vector of shape
S can only produce a sum of other root vectors of shape S. (iv) follows since the weights of T in VS are
precisely the negatives of those in V−S , as explained in [3, p. 561].
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
39
Importantly, the modules VS arising in [3] have a very limited range of possible high weights, enumerated
explicitly in [12, Lemma 3.1]. In most cases, these lie in the tensor algebra generated by the natural modules
for the simple classical factors of L.
6.4. Subgroups of F4 . We now enumerate the G-completely reducible semisimple subgroups of G =
F4 (K). For each such subgroup X, we also give the X-module structure of L(G), as well as the connected
def
centraliser C ◦ = CG (X)◦ . Using these, we also determine whether or not X is a separable subgroup of G
(‘Sep’), and whether (G, X) is a reductive pair (‘RP’).
The process is as follows. For each Levi subgroup L, and each simple factor L0 of L, the image of an
L-irreducible subgroup X in L0 is L0 -conjugate to one of the subgroups in Section 5. Thus each simple
factor of X is either contained in a simple factor of L, or is a diagonal subgroup of a product of two or
more such factors.
Let us illustrate with the case L134 , of type A1 A2 . Then X projects surjectively to the first factor, and
either projects surjectively to the second factor, or to an irreducible subgroup A1 , given by A2 (#1) in
Table 2. Thus one of the following holds:
• X = L′ . We denote this with the ID L134 (#0);
• p 6= 2 and X = A1 A1 , where the second factor is the subgroup A2 (#1) in Table 2. We denote this
by L134 (#0; #1);
• p 6= 2, X = A1 , and X is embedded diagonally in the subgroup L134 (#0; #1) above. Such an embedding is determined by a pair of Frobenius twists, and we therefore denote this by L134 (#0r , #1s ),
where r and s are the powers of p giving the twists. In such a situation we implicitly take rs = 0.
Next Lemma 6.3 and calculating the high weights of L on the shape modules VS gives the composition
factors of X = L134 (#0) = L′134 on L(G), as in the table. When p 6= 2, L(G) is tilting, and thus we obtain
the direct-sum decomposition stated. When p = 2, we consider instead the tilting modules TG (λ4 ) = λ4
(irreducible of dimension 26) and TG (λ1 ) = λ4 | λ1 | λ4 . Each of these restricts to X as a tilting module.
Thus, using for instance Doty’s software [9] for computing the structure of Weyl modules, we deduce that
T (λ1 ) ↓ X = (1, T (20)) + (1, T (02)) + (0, 11)2 + (0, T (20)) + (0, T (02)) + (T (2), 0) + (1, 0)2 ,
λ4 ↓ X = (1, 01) + (1, 10) + (0, 01) + (0, 10) + (0, 11),
where we have T (2) = 0 | 2 | 0 for A1 , and T (20) = 01 | 20 | 01 for A2 . Therefore L(G), a 52-dimensional
submodule of T (λ1 ) with λ4 as a submodule, has structure
L(G) ↓ X = (1, W (20)) + (0, W (20)) + (0, 11) + (1, W (02)) + (0, W (02)) + (T (2), 0) + (1, 0)2
as in the table. For the other possibilities of X, we need only compute the restriction of the modules for
L′134 , which are tensor products of modules for X, to deduce the given module structure.
Finally, each possible subgroup X above clearly centralises the 1-dimensional torus Z(L134 ). Since the
centraliser of a G-completely reducible subgroup is G-completely reducible [4, Corollary 3.17], in particular it is reductive, we deduce that CG (X)◦ is either a 1-dimensional torus (‘T1 ’) or is simple of type
A1 . Using Proposition 6.2, we need to consider the action of X on the unipotent radical of the (unique)
standard parabolic subgroup having L134 as a Levi factor. We find that the proper subgroups of L′134 (and
only these) centralise a vector of Q. Specifically, the shape module VS , where S consists of roots having
α2 -coefficient equal to 2, is an irreducible high weight module (0, 20) for L′134 , and thus restricts to the
diagonal subgroup L134 (#0; #1) as the symmetric square S 2 (2) = 4 + 0. This 1-dimensional subspace gives
rise to a 1-dimensional X-invariant subgroup, hence dim CG (X)◦ ≥ 2, and so CG (X)◦ is simple of type A1 .
This gives all the information for these subgroups in Table 38.
40
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
Remark 6.4. The following semisimple G-cr reducible subgroups were omitted from [24, Table 4.1]:
L124 (#1; #0), L124 (#1[r] ; #0[s] ), L134 (#0; #1), L134 (#0[r] ; #1[s] ), L123 (#5), L123 (#6), L123 (#7) when p =
2, L123 (#11), L123 (#12), L234 (#9), L234 (#10), L234 (#11), L234 (#13), L234 (#14), L234 (#15) when p ≥ 3,
L23 (#3) when p = 2, L23 (#5) and L34 (#1) when p ≥ 3.
Table 38: Reducible, G-cr subgroups of G = F4 (K)
X
B3
ID
L123 (#0)
C3
L234 (#0)
L123 (#2)
L234 (#3)
A1 B2 L123 (#4)
Conditions
p 6= 2
p=2
p 6= 2
p=2
A3
p=2
p=2
L234 (#1)
p 6= 2
p=2
A1 A2 L124 (#0)
p 6= 2
p=2
L134 (#0)
p 6= 2
p=2
A31
L123 (#1)
p 6= 2
p=2
L123 (#10)
p=2
L234 (#6)
p=2
L234 (#8)
p 6= 2
p=2
G2
L123 (#3)
B2
L234 (#4)
L23 (#0)
A2
L12 (#0)
L34 (#0)
p 6= 2
p=2
p=2
p 6= 2
p=2
L(G) ↓ X
010 + 1002 + 0012 + 0
((010 + 0)|100|0) + W (100)2 + 0012
200 + 0012 + 03
(0|200|(010 + 0)) + W (001)2 + 02
T (101; 0) + 0103 + 100 + 001 + 02
((020 + 0)|101) + ((200 + 002)|010)2 + 03
(0|(0, 02)|((2, 10) + 02 )|((2, 0) + (0, 10))|0) +
(((2, 0) + (0, 10))|0)2 + (1, 01)2
(2, 0) + (0, 02) + (0, 01)2 + (1, 01) + (1, 10)2 + 03
(0|(0, 02)|0|((0, 10) + (2, 0))|02 ) + (0, 01)2 + (1, 01) +
(1, W (10))2 + 02
(2, 0) + (0, T (11; 0)) + (0, 10) + (0, 01) + (1, 10) +
(1, 01) + (2, 10) + (2, 01) + (1, 0)2
(T (2), 0) + (0, 11) + (0, 10) + (0, 01) + (1, 10) +
(1, 01) + (W (2), 10) + (W (2), 01) + (1, 0)2
(1, 20) + (1, 02) + (0, 20) + (0, 02) + (0, T (11; 0)) +
(2, 0) + (1, 0)2
(1, W (20)) + (0, W (20)) + (0, 11) + (1, W (02)) +
(0, W (02)) + (T (2), 0) + (1, 0)2
(1, 1, 2) + (1, 1, 0)2 + (0, 1, 1)2 + (1, 0, 1)2 +
(2, 0, 0) + (0, 2, 0) + (0, 0, 2)3 + 0
(1, 1, W (2)) + (1, 1, 0)2 + (0, 1, 1)2 + (1, 0, 1)2 +
(02 |(2, 0, 0) + (0, 2, 0) + (0, 0, 2)|02 ) + (0, 0, W (2))2
(((2, 2, 0) + (2, 0, 2) + (0, 2, 2) + 03 )|((2, 0, 0) +
(0, 2, 0) + (0, 0, 2))|0) + (((2, 0, 0) + (0, 2, 0) +
(0, 0, 2))|0)2 + (1, 1, 1)2
((02 + (0, 2, 2))|((2, 0, 0) + (0, 2, 0) + (0, 0, 2))|02 ) +
(0, 1, 1)2 +(1, 1, 1)+(((1, 2, 0)+(1, 0, 2))|(1, 0, 0))2+02
(2, 0, 0) + (0, 2, 0) + (0, 0, 2) + (0, 1, 1) + (1, 1, 0) +
(1, 0, 1)+(1, 0, 0)2 +(0, 1, 0)2 +(0, 0, 1)2 +(1, 1, 1)2 +03
(0|((2, 0, 0) + (0, 2, 0) + (0, 0, 2))|03 ) + (1, 1, 1)2 +
(1, 1, 0) + (0, 1, 1) + (1, 0, 1) + (1, 0, 0)2 + (0, 1, 0)2 +
(0, 0, 1)2 + 02
W (01) + 105 + 03
01 + W (10)2 + T (10)3 + 03
(0|20|(01 + 0)) + (0|20|0|10)2 + 02
02 + 104 + 014 + 06
(0|10|0|02|0) + W (10)4 + 014 + 05
T (11; 0) + 106 + 016 + 07
T (11; 0) + W (20)3 + W (02)3 + 07
C◦
T1
Ã1
Ā1
Ā1
Ã1
Ā1
Ã1
Sep
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Ā1
Ā1
Yes Yes
No No
T1
Yes p 6= 3
T1
Yes No
T1
Yes p 6= 3
T1
Yes No
T1
Yes Yes
Ã1
No
Ã1
Yes No
Ā1
No
Ā1
Yes Yes
Ā1
No
No
Ã1
Ã1
Ā1
Ā21
B2
Ã2
Ā2
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
No
p 6= 3
p 6= 3
RP
Yes
No
Yes
No
p 6= 2
No
No
No
No
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
A21
L123 (#8)
L13 (#0)
p=3
p 6= 2
p=2
L124 (#1; #0)
L134 (#0; #1)
L23 (#1)
L23 (#2)
p 6= 2
p 6= 2
p 6= 2
p=2
p=2
L123 (#5)
p 6= 2,
p 6= 2,
p = 2,
L123 (#6)
p 6= 2
p=2
L123 (#11)
p=2
L234 (#2)
p≥5
p=3
p≥5
p = 2,
L234 (#7)
L234 (#9)
p = 2,
p = 2,
L234 (#10)
p=2
L234 (#13)
p 6= 2
p=2
A1
L1 (#0)
L3 (#0)
L12 (#1)
L34 (#1)
L13 (#0r ; #0s )
p 6= 2
p=2
p 6= 2
p 6= 2
p 6= 2,
p 6= 2,
((30 + 03 + 0)|11) + 115 + 03
(2, 0) + (0, 2)3 + (1, 0)4 + (0, 1)4 + (1, 1)2 + (1, 2)2 + 04
(T (2), 0) + (1, W (2))2 + (1, 1)2 + (1, 0)4 +
(0, T (2)) + (0, W (2))2 + (0, 1)4 + 02
(0, T (4; 0)) + (2, 0) + (0, 2)3 + (1, 2)2 + (2, 2)2 + (0, 1)2
(1, T (4; 0))2 + (0, T (4; 0))3 + (2, 0) + (0, 2) + (1, 0)2
(2, 0) + (0, 2) + (1, 1)5 + (1, 0)4 + (0, 1)4 + 010
(0|((2, 0) + (0, 2))|02 ) + (1, 1)5 + (1, 0)4 + (0, 1)4 + 09
((02 + (2, 2))|((2, 0) + (0, 2))|0) + (((0, 2) +
(2, 0))|0)4 + (1, 1)4 + 05
r 6= 0
(1 ⊗ 1[r] , 2) + (1 ⊗ 1[r] , 0)2 + (1[r] , 1)2 + (1, 1)2 +
(2, 0) + (2[r] , 0) + (0, 2)3 + 0
r=0
(2, 2) + (2, 0)4 + (0, 2)4 + (1, 1)4 + 03
r 6= 0
(1 ⊗ 1[r] , W (2)) + (1 ⊗ 1[r], 0)2 + (1[r] , 1)2 + (1, 1)2 +
(02 |(2, 0) + (2[r] , 0) + (0, 2)|02 ) + (0, W (2))2
(1[r] ⊗ 2[s] , 1) + (1[r] , 1)2 + (1[s] , 1)2 +
(1[r] ⊗ 1[s] , 0)2 + (2[r] , 0) + (0, 2) + (2[s] , 0)3 + 0
(1[r] ⊗ W (2)[s] , 1) + (1[r] , 1)2 + (1[s] , 1)2 +
(1[r] ⊗ 1[s] , 0)2 + (02 |((2[r] , 0) + (2[s] , 0) +
(0, 2))|02 ) + (W (2)[s] , 0)2
(((2[r] , 2)+(2⊗2[r] , 0)+(2, 2)+03 )|((2, 0)+(2[r] , 0)+
(0, 2))|0)+(((2, 0)+(2[r] , 0)+(0, 2))|0)2 +(1⊗1[r] , 1)2
(4, 2) + (4, 1) + (0, 2) + (0, 3) + (2, 0) + 03
(T (4), 2) + (2, 0) + ((0, 1)|((4, 1) + (0, 3))|(0, 1)) + 03
(2, 0) + (0, T (6; 2)) + (0, 3)2 + (1, 3) + (1, 4)2 + 03
r 6= s, s + 1 (02 + (2[s] , 2)|((2[r] , 0) + (2[s] , 0) + (0, 2))|02 ) +
(1[s] , 1)2 + (1[r] ⊗ 1[s] , 1) + (((1[r] ⊗ 2[s] , 0) +
(1[r] , 2))|(1[r] , 0))2 + 02
r=s
(0 + (2, 2)|((2, 0) + (2, 0))|0) + (0, T (2)[1] ) + (1, 1)2 +
(T (2), 1) + (3, 0)2 + (1, W (2))2 + 02
r =s+1
(02 + (2, 2)|((4, 0) + (2, 0) + (0, 2))|02 ) + (1, 1)2 +
(3, 1) + (((T (2)[1] , 0) + (2, 2))|(2, 0))2 + 02
((02 + (0, 2 ⊗ 2[r] ))|((2, 0) + (0, 2) + (0, 2[r] ))|02 ) +
(0, 1 ⊗ 1[r] )2 + (1, 1 ⊗ 1[r] ) + (((1, 2) +
(1, 2[r] ))|(1, 0))2 + 02
(2, 0) + (2[r] , 0) + (0, 2) + (1 ⊗ 1[r] , 1)2 + (1, 1) +
(1[r] , 1) + (1 ⊗ 1[r], 0) + (1, 0)2 + (1[r] , 0)2 + (0, 1)2 + 03
(0|((2, 0)+(2[r], 0)+(0, 2))|03 )+(1⊗1[r], 1)2 +(1, 1)+
(1[r] , 1) + (1 ⊗ 1[r], 0) + (1, 0)2 + (1[r] , 0)2 + (0, 1)2 + 02
T (2; 0) + 114 + 020
27 + 18 + 015
T (2) + W (2)6 + 18 + 014
T (4; 0) + 213 + 07
T (4; 0)7 + 2 + 07
r 6= s
2[r] + (2[s] )3 + (1[r] )4 + (1[s] )4 + (1[r] ⊗ 1[s] )2 +
(1[r] ⊗ 2[s] )2 + 04
r=s
26 + 18 + T (3; 1)2 + 06
41
A1
Yes No
Ā1 T1 Yes Yes
Ā1 Ã1 Yes No
T1
A1
B2
B2
B2
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
No
No
T1
Yes Yes
Ã1
Ã1
Yes Yes
No No
T1
Yes Yes
Ã1
No
Ã1
Yes No
Ā1
Ā1
Ā1
Ā1
Yes
Yes
Yes
No
Yes
No
p 6= 5
No
Ā1
No
No
Ā1
No
No
Ā1
No
No
Ā1
Yes Yes
Ā1
No
No
C3
Ā3
B3
Ã2
G2
Ā1 T1
Yes
Yes
Yes
Yes
Yes
Yes
p 6= 2
Yes
No
Yes
Yes
Yes
No
Ā1 A1 Yes Yes
42
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
p = 2, r 6= s, s + 1 T (2)[r] + T (2)[s] + (1[r] ⊗ W (2)[s] )2 + (1[r] )4 +
(1[s] )4 + (1[r] ⊗ 1[s] )2 + (W (2)[s] )2 + 02
p = 2, r = s + 1
T (2) + T (2)[1] + (0|4|0|2)2 + 14 + 24 + 32 + W (2)2 + 02
p = 2, r = s
T (2)4 + 110 + 22 + 32 + W (2)2 + 02
L23 (#3)
p≥5
T (6; 2) + 44 + 34 + 06
p=2
((02 + 6)|(2 + 4)|0) + ((2 + 4)|0)4 + 34 + 05
L23 (#4)
p 6= 2
2 + 2[r] + (1 ⊗ 1[r] )5 + 14 + (1[r] )4 + 010
p=2
(0|(2[r] + 2)|02 ) + (1 ⊗ 1[r] )5 + 14 + (1[r] )4 + 09
L23 (#5)
p=2
((02 + (2, 2[r] ))|(2 + 2[r] )|0) + ((2 + 2[r])|0)4 +
(1 ⊗ 1[r] )4 + 05
L134 (#0r ; #1s ) p 6= 2, r 6= s, s + 1 (1[r] ⊗ T (4; 0)[s] )2 + (T (4; 0)[s] )3 + 2[r] + 2[s] + (1[r] )2
p 6= 2, r = s
T (5; 3; 1)2 + 12 + 22 + T (4; 0)3
p = 3, r = s + 1
(1[1] |((1 ⊗ 2[1] ) + 1)|1[1] )2 + T (4; 0)3 + 2 + 2[1] + (1[1] )2
p ≥ 5, r = s + 1
(4 ⊗ 1[1] )2 + (1[1] )4 + 2 + 2[1] + 43 + 03
r
s
L124 (#0 ; #1 ) p 6= 2, r 6= s
T (4; 0)[s] + (2[s] )3 + 2[r] + (1[r] ⊗ 2[s] )2 +
(2[r] ⊗ 2[s] )2 + (1[r] )2
p 6= 2, r = s
T (4; 0)3 + 26 + T (3; 1)2 + 12
L123 (#7)
p 6= 2,
(1[r] ⊗ 1[s] ⊗ 2[t] ) + 2[r] + 2[s] + (2[t] )3 + (1[r] ⊗ 1[s] )2
r, s, t distinct
+(1[s] ⊗ 1[t] )2 + (1[r] ⊗ 1[t] )2 + 0
p 6= 2, r = s < t
(2 ⊗ 2[t] ) + 24 + (2[t] )4 + (1 ⊗ 1[t] )4 + 03
p 6= 2, r = t < s
(T (3; 1) ⊗ 1[s] ) + 26 + 2[s] + (1 ⊗ 1[s] )4 + 03
p 6= 2, r < s = t
2 + (2[s] )6 + (1 ⊗ T (3; 1)[s] ) + (1 ⊗ 1[s] )4 + 03
p = 2,
(1[r] ⊗ 1[s] ⊗ W (2)[t] ) + (1[r] ⊗ 1[s] )2 + (1[s] ⊗ 1[t] )2
r, s, t distinct
+(1[r] ⊗ 1[t])2 + (02 |(2[r] + 2[s] + 2[t] )|02 ) + (W (2)[t] )2
p = 2, r < s = t
(1 ⊗ 3[s] ) + (1 ⊗ 1[s] )5 + (T (2)[s] )3 +
(0|(2 + 2[s] )|0) + (W (2)[s] )2
p = 2, r = t < s
(3⊗1[s] )+(1⊗1[s] )5 +T (2)3 +(0|(2+2[s] )|0)+W (2)2
L123 (#9)
p≥7
T (10; 2) + 65 + 03
L123 (#12)
p=2
(((2[r] ⊗ 2[s] ) + (2 ⊗ 2[r] ) + (2 ⊗ 2[s] ) + 03 )|(2 + 2[r] +
2[s] )|0) + ((2 + 2[r] + 2[s] )|0)2 + (1 ⊗ 1[r] ⊗ 1[s] )2
L234 (#5)
p≥7
T (10; 2) + 6 + T (9; 3)2 + 03
L234 (#11)
p = 2,
((02 + (2[s] ⊗ 2[t] ))|(2[r] + 2[s] + 2[t] )|02 ) +
r 6= s, t, s + 1, t + 1 (1[s] ⊗ 1[t] )2 + (1[r] ⊗ 1[s] ⊗ 1[t] ) + (((1[r] ⊗ 2[s] ) +
(1[r] ⊗ 2[t] ))|1[r] )2 + 02
p = 2, r = s,
((0 + (2 ⊗ 2[t] ))|(2 + 2[t] )|0) + T (2) + (1 ⊗ 1[t] )2 +
t≥2
(T (2) ⊗ 1[t] ) + 32 + ((1 ⊗ 2[t] )|1)2 + 02
p = 2, r = s,
((0 + 6)|(2 + 4)|0) + T (2) + 34 + T (4) + W (5)2 + 02
t=1
p = 2, r = s + 1, ((02 + (2 ⊗ 2[t] ))|(2 + 4 + 2[t] )|02 ) + (1 ⊗ 1[t] )2 +
t≥2
(3 ⊗ 1[t] ) + (0|4|(0 + (2 ⊗ 2[t] ))|2)2 + 02
p = 2, r = t + 1
((02 + (2 ⊗ 4[t] ))|(2 + 2[t] + 4[t] )|02 ) + (1 ⊗ 1[t] )2 +
(1 ⊗ 1[t] ⊗ 2[t] ) + (0|4[t] |((2 ⊗ 2[t] ) + 0)|2[t] )2 + 02
p = 2, r = t = 1
((0 + (2 ⊗ 4))|(2 + 4)|0) + T (2)[1] + 32 +
(T (2)[1] ⊗ 1) + (2 ⊗ 4)2 + ((0|4|0|2)2 + 02
p = 2, r = t > 1
((0 + (2 ⊗ 2[r] ))|(2 + 2[r] )|0) + T (2)[r] + (1 ⊗ 1[r] )2 +
(1 ⊗ T (2)[r] ) + (1[r] ⊗ 2[r] )2 + ((2 ⊗ 1[r] )|1[r] )2 + 02
Ā1 Ã1 Yes No
Ā1 Ã1
Ã2
Ā21
B2
B2
B2
B2
Yes
Yes
Yes
Yes
Yes
No
Yes
No
No
p 6= 5
No
Yes
No
No
A1
A1
A1
A1
T1
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
A1
T1
Yes Yes
Yes Yes
Ã1
Ã1
Ã1
Ã1
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Ã1
No
No
Ã1
A1
Ã1
No No
Yes p 6= 7
Yes No
Ā1
Ā1
Yes p 6= 7
No No
Ā1
No
No
Ā1
No
No
Ā1
No
No
Ā1
No
No
Ā1
No
No
Ā1
No
No
REDUCIBLE SUBGROUPS OF EXCEPTIONAL ALGEBRAIC GROUPS
43
L234 (#12)
p ≥ 5, r 6= s
2[r] +T (6; 2)[s] +(3[r] )2 +(1[r] ⊗ 3[s] )+(1[r] ⊗ 4[s] )2 +03 Ā1
L234 (#14)
p ≥ 5, r = s
p 6= 2
22 + T (6; 2) + 4 + 32 + T (5; 3)2 + 03
2 + 2[r] + 2[s] + (1 ⊗ 1[r] ⊗ 1[s] ) + (1 ⊗ 1[r] ) +
(1 ⊗ 1[s] ) + (1[r] ⊗ 1[s] ) + 12 + (1[r] )2 + (1[s] )2 + 03
(0|(2 + 2[r] + 2[s] )|03 ) + (1 ⊗ 1[r] ⊗ 1[s] )2 + (1 ⊗ 1[r] ) +
(1 ⊗ 1[s] ) + (1[r] ⊗ 1[s] ) + 12 + (1[r] )2 + (1[s] )2 + 02
(4[r] ⊗ 2[s] ) + (4[r] ⊗ 1[s] ) + 2[r] + 2[s] + 3[s] + 03
(T (4)[r] ⊗2[s] )+2[r] +(1[s] |((4[r] ⊗1[s] )+3[s] )|1[s] )+03
T (10) + 2 + (3|(9 + 7 + 1)|3) + 03
p=2
L234 (#15)
p≥5
p = 3, s 6= r + 1
p = 3, s = r + 1
Ā1
Ā1
Yes p 6= 5
or
r<s
Yes Yes
Yes Yes
Ā1
No
Ā1
Ā1
Ā1
Yes Yes
Yes r < s
Yes Yes
No
References
1. Bonnie Amende, G-irreducible subgroups of type A1 , ProQuest LLC, Ann Arbor, MI, 2005, Thesis (Ph.D.)–University of
Oregon. MR 2707891
2. Michael Aschbacher, Chevalley groups of type G2 as the group of a trilinear form, J. Algebra 109 (1987), no. 1, 193–259.
MR 898346 (88g:20089)
3. H. Azad, M. Barry, and G. Seitz, On the structure of parabolic subgroups, Comm. Algebra 18 (1990), no. 2, 551–562.
MR 1047327 (91d:20048)
4. Michael Bate, Benjamin Martin, and Gerhard Röhrle, A geometric approach to complete reducibility, Invent. Math. 161
(2005), no. 1, 177–218. MR 2178661 (2007k:20101)
5. N. Bourbaki, Éléments de mathématique. Fasc. XXXIV. Groupes et algèbres de Lie. Chapitre IV: Groupes de Coxeter
et systèmes de Tits. Chapitre V: Groupes engendrés par des réflexions. Chapitre VI: systèmes de racines, Actualités
Scientifiques et Industrielles, No. 1337, Hermann, Paris, 1968. MR 0240238 (39 #1590)
6. Roger W. Carter, Simple groups of Lie type, John Wiley & Sons, London-New York-Sydney, 1972, Pure and Applied
Mathematics, Vol. 28. MR 0407163 (53 #10946)
7.
, Finite groups of Lie type, Pure and Applied Mathematics (New York), John Wiley & Sons Inc., New York, 1985,
Conjugacy classes and complex characters, A Wiley-Interscience Publication. MR 794307 (87d:20060)
8. Stephen Donkin, On tilting modules for algebraic groups, Mathematische Zeitschrift 212 (1993), no. 1, 39–60 (English).
9. Stephen Doty, Weyl modules, Gap Package, ver. 1.1.
10. Daniel Gorenstein, Richard Lyons, and Ronald Solomon, The classification of the finite simple groups. Number 3. Part
I. Chapter A, Mathematical Surveys and Monographs, vol. 40, American Mathematical Society, Providence, RI, 1998,
Almost simple K-groups. MR 1490581 (98j:20011)
11. R. Lawther and D. M. Testerman, A1 subgroups of exceptional algebraic groups, Mem. Amer. Math. Soc. 141 (1999),
no. 674, viii+131. MR 1466951
12. Martin W. Liebeck and Gary M. Seitz, Reductive subgroups of exceptional algebraic groups, Mem. Amer. Math. Soc. 121
(1996), no. 580, vi+111. MR 1329942 (96i:20059)
13.
, On the subgroup structure of classical groups, Invent. Math. 134 (1998), no. 2, 427–453. MR 1650328 (99h:20074)
, The maximal subgroups of positive dimension in exceptional algebraic groups, Mem. Amer. Math. Soc. 169 (2004),
14.
no. 802, vi+227. MR 2044850 (2005b:20082)
15. Martin W. Liebeck and Donna M. Testerman, Irreducible subgroups of algebraic groups, Q. J. Math. 55 (2004), no. 1,
47–55. MR 2043006 (2005b:20087)
16. Alastair J. Litterick and Adam R. Thomas, Complete reducibility in bad characteristic I, in preparation.
17.
, Complete reducibility in good characteristic, to appear in Trans. Amer. Math. Soc.
18. Frank Lübeck, Small degree representations of finite Chevalley groups in defining characteristic, LMS J. Comput. Math.
4 (2001), 135–169 (electronic). MR 1901354 (2003e:20013)
19. Gunter Malle and Donna Testerman, Linear algebraic groups and finite groups of Lie type, Cambridge Studies in Advanced
Mathematics, vol. 133, Cambridge University Press, Cambridge, 2011. MR 2850737 (2012i:20058)
20. Gary M. Seitz, The maximal subgroups of classical algebraic groups, Mem. Amer. Math. Soc. 67 (1987), no. 365, iv+286.
MR 888704 (88g:20092)
21. Jean-Pierre Serre, Morsund lectures, University of Oregon, 1998.
22. Robert Steinberg, Lecture on Chevalley groups, Notes prepared by J. Faulkner and R. Wilson, Yale University, New Haven,
Conn., 1968.
23. David I. Stewart, The reductive subgroups of G2 , J. Group Theory 13 (2010), no. 1, 117–130. MR 2604850 (2011c:20099)
24.
, The reductive subgroups of F4 , Mem. Amer. Math. Soc. 223 (2013), no. 1049, vi+88. MR 3075783
44
ALASTAIR J. LITTERICK AND ADAM R. THOMAS
25. Adam R. Thomas, The irreducible subgroups of exceptional algebraic groups, to appear in Mem. Amer. Math. Soc.
26.
, Simple irreducible subgroups of exceptional algebraic groups, J. Algebra 423 (2015), 190–238.
27.
, Irreducible A1 subgroups of exceptional algebraic groups, J. Algebra 447 (2016), 240–296.
A.J. Litterick, Fakultät für Mathematik, Ruhr-Universität Bochum, Universitätsstraße 150, D-44780
Bochum, Germany
E-mail address: [email protected]
A.R. Thomas, School of Mathematics, University of Bristol, Bristol, BS8 1TW, UK, and The Heilbronn
Institute for Mathematical Research, Bristol, UK.
E-mail address: [email protected]
| 4math.GR
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.